ZhengPeng7 commited on
Commit
69b7015
1 Parent(s): 3a3b0bf

Upgrade the space codes with deployed HF model.

Browse files
BiRefNet-massive-epoch_240.pth DELETED
@@ -1,3 +0,0 @@
1
- version https://git-lfs.github.com/spec/v1
2
- oid sha256:1ccb3959dfa99b85d7b1f57dcf59616873ff5f45e0a3f1e0b14b07f6ee74b01f
3
- size 885038517
 
 
 
 
app.py CHANGED
@@ -12,13 +12,7 @@ from gradio_imageslider import ImageSlider
12
  torch.set_float32_matmul_precision('high')
13
  torch.jit.script = lambda f: f
14
 
15
- from models.birefnet import BiRefNet
16
- from config import Config
17
-
18
-
19
- config = Config()
20
- device = config.device
21
-
22
 
23
 
24
  def array_to_pil_image(image, size=(1024, 1024)):
@@ -40,17 +34,11 @@ class ImagePreprocessor():
40
  return image
41
 
42
 
43
- model = BiRefNet(bb_pretrained=False)
44
- state_dict = ['BiRefNet-massive-epoch_240.pth'][0]
45
- if os.path.exists(state_dict):
46
- birefnet_dict = torch.load(state_dict, map_location="cpu")
47
- unwanted_prefix = '_orig_mod.'
48
- for k, v in list(birefnet_dict.items()):
49
- if k.startswith(unwanted_prefix):
50
- birefnet_dict[k[len(unwanted_prefix):]] = birefnet_dict.pop(k)
51
- model.load_state_dict(birefnet_dict)
52
- model = model.to(device)
53
- model.eval()
54
 
55
 
56
  # def predict(image_1, image_2):
@@ -71,7 +59,7 @@ def predict(image, resolution):
71
  images_proc = torch.cat([image_proc.unsqueeze(0) for image_proc in images_proc])
72
 
73
  with torch.no_grad():
74
- scaled_preds_tensor = model(images_proc.to(device))[-1].sigmoid() # BiRefNet needs an sigmoid activation outside the forward.
75
  preds = []
76
  for image_shape, pred_tensor in zip(image_shapes, scaled_preds_tensor):
77
  if device == 'cuda':
 
12
  torch.set_float32_matmul_precision('high')
13
  torch.jit.script = lambda f: f
14
 
15
+ device = "cuda" if torch.cuda.is_available() else "cpu"
 
 
 
 
 
 
16
 
17
 
18
  def array_to_pil_image(image, size=(1024, 1024)):
 
34
  return image
35
 
36
 
37
+
38
+ from transformers import AutoModelForImageSegmentation
39
+ birefnet = AutoModelForImageSegmentation.from_pretrained('zhengpeng7/BiRefNet', trust_remote_code=True)
40
+ birefnet.to(device)
41
+ birefnet.eval()
 
 
 
 
 
 
42
 
43
 
44
  # def predict(image_1, image_2):
 
59
  images_proc = torch.cat([image_proc.unsqueeze(0) for image_proc in images_proc])
60
 
61
  with torch.no_grad():
62
+ scaled_preds_tensor = birefnet(images_proc.to(device))[-1].sigmoid() # BiRefNet needs an sigmoid activation outside the forward.
63
  preds = []
64
  for image_shape, pred_tensor in zip(image_shapes, scaled_preds_tensor):
65
  if device == 'cuda':
config.py DELETED
@@ -1,111 +0,0 @@
1
- import os
2
- import math
3
-
4
- import torch
5
-
6
-
7
- class Config():
8
- def __init__(self) -> None:
9
- self.ms_supervision = True
10
- self.out_ref = self.ms_supervision and True
11
- self.dec_ipt = True
12
- self.dec_ipt_split = True
13
- self.locate_head = False
14
- self.cxt_num = [0, 3][1] # multi-scale skip connections from encoder
15
- self.mul_scl_ipt = ['', 'add', 'cat'][2]
16
- self.refine = ['', 'itself', 'RefUNet', 'Refiner', 'RefinerPVTInChannels4'][0]
17
- self.progressive_ref = self.refine and True
18
- self.ender = self.progressive_ref and False
19
- self.scale = self.progressive_ref and 2
20
- self.dec_att = ['', 'ASPP', 'ASPPDeformable'][2]
21
- self.squeeze_block = ['', 'BasicDecBlk_x1', 'ResBlk_x4', 'ASPP_x3', 'ASPPDeformable_x3'][1]
22
- self.dec_blk = ['BasicDecBlk', 'ResBlk', 'HierarAttDecBlk'][0]
23
- self.auxiliary_classification = False
24
- self.refine_iteration = 1
25
- self.freeze_bb = False
26
- self.precisionHigh = True
27
- self.compile = True
28
- self.load_all = True
29
- self.verbose_eval = True
30
-
31
- self.size = 1024
32
- self.batch_size = 2
33
- self.IoU_finetune_last_epochs = [0, -40][1] # choose 0 to skip
34
- if self.dec_blk == 'HierarAttDecBlk':
35
- self.batch_size = 2 ** [0, 1, 2, 3, 4][2]
36
- self.model = [
37
- 'BSL',
38
- ][0]
39
-
40
- # Components
41
- self.lat_blk = ['BasicLatBlk'][0]
42
- self.dec_channels_inter = ['fixed', 'adap'][0]
43
-
44
- # Backbone
45
- self.bb = [
46
- 'vgg16', 'vgg16bn', 'resnet50', # 0, 1, 2
47
- 'pvt_v2_b2', 'pvt_v2_b5', # 3-bs10, 4-bs5
48
- 'swin_v1_b', 'swin_v1_l' # 5-bs9, 6-bs6
49
- ][6]
50
- self.lateral_channels_in_collection = {
51
- 'vgg16': [512, 256, 128, 64], 'vgg16bn': [512, 256, 128, 64], 'resnet50': [1024, 512, 256, 64],
52
- 'pvt_v2_b2': [512, 320, 128, 64], 'pvt_v2_b5': [512, 320, 128, 64],
53
- 'swin_v1_b': [1024, 512, 256, 128], 'swin_v1_l': [1536, 768, 384, 192],
54
- }[self.bb]
55
- if self.mul_scl_ipt == 'cat':
56
- self.lateral_channels_in_collection = [channel * 2 for channel in self.lateral_channels_in_collection]
57
- self.cxt = self.lateral_channels_in_collection[1:][::-1][-self.cxt_num:] if self.cxt_num else []
58
- self.sys_home_dir = '/root/autodl-tmp'
59
- self.weights_root_dir = os.path.join(self.sys_home_dir, 'weights')
60
- self.weights = {
61
- 'pvt_v2_b2': os.path.join(self.weights_root_dir, 'pvt_v2_b2.pth'),
62
- 'pvt_v2_b5': os.path.join(self.weights_root_dir, ['pvt_v2_b5.pth', 'pvt_v2_b5_22k.pth'][0]),
63
- 'swin_v1_b': os.path.join(self.weights_root_dir, ['swin_base_patch4_window12_384_22kto1k.pth', 'swin_base_patch4_window12_384_22k.pth'][0]),
64
- 'swin_v1_l': os.path.join(self.weights_root_dir, ['swin_large_patch4_window12_384_22kto1k.pth', 'swin_large_patch4_window12_384_22k.pth'][0]),
65
- }
66
-
67
- # Training
68
- self.num_workers = 5 # will be decrease to min(it, batch_size) at the initialization of the data_loader
69
- self.optimizer = ['Adam', 'AdamW'][0]
70
- self.lr = 1e-5 * math.sqrt(self.batch_size / 5) # adapt the lr linearly
71
- self.lr_decay_epochs = [1e4] # Set to negative N to decay the lr in the last N-th epoch.
72
- self.lr_decay_rate = 0.5
73
- self.only_S_MAE = False
74
- self.SDPA_enabled = False # Bug. Slower and errors occur in multi-GPUs
75
-
76
- # Data
77
- self.data_root_dir = os.path.join(self.sys_home_dir, 'datasets/dis')
78
- self.dataset = ['DIS5K', 'COD', 'SOD'][0]
79
- self.preproc_methods = ['flip', 'enhance', 'rotate', 'pepper', 'crop'][:4]
80
-
81
- # Loss
82
- self.lambdas_pix_last = {
83
- # not 0 means opening this loss
84
- # original rate -- 1 : 30 : 1.5 : 0.2, bce x 30
85
- 'bce': 30 * 1, # high performance
86
- 'iou': 0.5 * 1, # 0 / 255
87
- 'iou_patch': 0.5 * 0, # 0 / 255, win_size = (64, 64)
88
- 'mse': 150 * 0, # can smooth the saliency map
89
- 'triplet': 3 * 0,
90
- 'reg': 100 * 0,
91
- 'ssim': 10 * 1, # help contours,
92
- 'cnt': 5 * 0, # help contours
93
- }
94
- self.lambdas_cls = {
95
- 'ce': 5.0
96
- }
97
- # Adv
98
- self.lambda_adv_g = 10. * 0 # turn to 0 to avoid adv training
99
- self.lambda_adv_d = 3. * (self.lambda_adv_g > 0)
100
-
101
- # others
102
- self.device = "cuda" if torch.cuda.is_available() else "cpu"
103
-
104
- self.batch_size_valid = 1
105
- self.rand_seed = 7
106
- run_sh_file = [f for f in os.listdir('.') if 'train.sh' == f] + [os.path.join('..', f) for f in os.listdir('..') if 'train.sh' == f]
107
- # with open(run_sh_file[0], 'r') as f:
108
- # lines = f.readlines()
109
- # self.save_last = int([l.strip() for l in lines if 'val_last=' in l][0].split('=')[-1])
110
- # self.save_step = int([l.strip() for l in lines if 'step=' in l][0].split('=')[-1])
111
- # self.val_step = [0, self.save_step][0]
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
dataset.py DELETED
@@ -1,91 +0,0 @@
1
- import os
2
- import cv2
3
- from tqdm import tqdm
4
- from PIL import Image
5
- from torch.utils import data
6
- from torchvision import transforms
7
-
8
- from preproc import preproc
9
- from config import Config
10
-
11
-
12
- Image.MAX_IMAGE_PIXELS = None # remove DecompressionBombWarning
13
- config = Config()
14
- _class_labels_TR_sorted = 'Airplane, Ant, Antenna, Archery, Axe, BabyCarriage, Bag, BalanceBeam, Balcony, Balloon, Basket, BasketballHoop, Beatle, Bed, Bee, Bench, Bicycle, BicycleFrame, BicycleStand, Boat, Bonsai, BoomLift, Bridge, BunkBed, Butterfly, Button, Cable, CableLift, Cage, Camcorder, Cannon, Canoe, Car, CarParkDropArm, Carriage, Cart, Caterpillar, CeilingLamp, Centipede, Chair, Clip, Clock, Clothes, CoatHanger, Comb, ConcretePumpTruck, Crack, Crane, Cup, DentalChair, Desk, DeskChair, Diagram, DishRack, DoorHandle, Dragonfish, Dragonfly, Drum, Earphone, Easel, ElectricIron, Excavator, Eyeglasses, Fan, Fence, Fencing, FerrisWheel, FireExtinguisher, Fishing, Flag, FloorLamp, Forklift, GasStation, Gate, Gear, Goal, Golf, GymEquipment, Hammock, Handcart, Handcraft, Handrail, HangGlider, Harp, Harvester, Headset, Helicopter, Helmet, Hook, HorizontalBar, Hydrovalve, IroningTable, Jewelry, Key, KidsPlayground, Kitchenware, Kite, Knife, Ladder, LaundryRack, Lightning, Lobster, Locust, Machine, MachineGun, MagazineRack, Mantis, Medal, MemorialArchway, Microphone, Missile, MobileHolder, Monitor, Mosquito, Motorcycle, MovingTrolley, Mower, MusicPlayer, MusicStand, ObservationTower, Octopus, OilWell, OlympicLogo, OperatingTable, OutdoorFitnessEquipment, Parachute, Pavilion, Piano, Pipe, PlowHarrow, PoleVault, Punchbag, Rack, Racket, Rifle, Ring, Robot, RockClimbing, Rope, Sailboat, Satellite, Scaffold, Scale, Scissor, Scooter, Sculpture, Seadragon, Seahorse, Seal, SewingMachine, Ship, Shoe, ShoppingCart, ShoppingTrolley, Shower, Shrimp, Signboard, Skateboarding, Skeleton, Skiing, Spade, SpeedBoat, Spider, Spoon, Stair, Stand, Stationary, SteeringWheel, Stethoscope, Stool, Stove, StreetLamp, SweetStand, Swing, Sword, TV, Table, TableChair, TableLamp, TableTennis, Tank, Tapeline, Teapot, Telescope, Tent, TobaccoPipe, Toy, Tractor, TrafficLight, TrafficSign, Trampoline, TransmissionTower, Tree, Tricycle, TrimmerCover, Tripod, Trombone, Truck, Trumpet, Tuba, UAV, Umbrella, UnevenBars, UtilityPole, VacuumCleaner, Violin, Wakesurfing, Watch, WaterTower, WateringPot, Well, WellLid, Wheel, Wheelchair, WindTurbine, Windmill, WineGlass, WireWhisk, Yacht'
15
- class_labels_TR_sorted = _class_labels_TR_sorted.split(', ')
16
-
17
-
18
- class MyData(data.Dataset):
19
- def __init__(self, data_root, image_size, is_train=True):
20
- self.size_train = image_size
21
- self.size_test = image_size
22
- self.keep_size = not config.size
23
- self.data_size = (config.size, config.size)
24
- self.is_train = is_train
25
- self.load_all = config.load_all
26
- self.device = config.device
27
- self.dataset = data_root.replace('\\', '/').split('/')[-1]
28
- if self.is_train and config.auxiliary_classification:
29
- self.cls_name2id = {_name: _id for _id, _name in enumerate(class_labels_TR_sorted)}
30
- self.transform_image = transforms.Compose([
31
- transforms.Resize(self.data_size),
32
- transforms.ToTensor(),
33
- transforms.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225]),
34
- ][self.load_all or self.keep_size:])
35
- self.transform_label = transforms.Compose([
36
- transforms.Resize(self.data_size),
37
- transforms.ToTensor(),
38
- ][self.load_all or self.keep_size:])
39
- ## 'im' and 'gt' need modifying
40
- image_root = os.path.join(data_root, 'im')
41
- self.image_paths = [os.path.join(image_root, p) for p in os.listdir(image_root)]
42
- self.label_paths = [p.replace('/im/', '/gt/').replace('.jpg', '.png') for p in self.image_paths]
43
- if self.load_all:
44
- self.images_loaded, self.labels_loaded = [], []
45
- self.class_labels_loaded = []
46
- # for image_path, label_path in zip(self.image_paths, self.label_paths):
47
- for image_path, label_path in tqdm(zip(self.image_paths, self.label_paths), total=len(self.image_paths)):
48
- _image = cv2.imread(image_path)
49
- _label = cv2.imread(label_path, cv2.IMREAD_GRAYSCALE)
50
- if not self.keep_size:
51
- _image_rs = cv2.resize(_image, (config.size, config.size), interpolation=cv2.INTER_LINEAR)
52
- _label_rs = cv2.resize(_label, (config.size, config.size), interpolation=cv2.INTER_LINEAR)
53
- self.images_loaded.append(
54
- Image.fromarray(cv2.cvtColor(_image_rs, cv2.COLOR_BGR2RGB)).convert('RGB')
55
- )
56
- self.labels_loaded.append(
57
- Image.fromarray(_label_rs).convert('L')
58
- )
59
- self.class_labels_loaded.append(
60
- self.cls_name2id[label_path.split('/')[-1].split('#')[3]] if self.is_train and config.auxiliary_classification else -1
61
- )
62
-
63
-
64
- def __getitem__(self, index):
65
-
66
- if self.load_all:
67
- image = self.images_loaded[index]
68
- label = self.labels_loaded[index]
69
- class_label = self.class_labels_loaded[index] if self.is_train and config.auxiliary_classification else -1
70
- else:
71
- image = Image.open(self.image_paths[index]).convert('RGB')
72
- label = Image.open(self.label_paths[index]).convert('L')
73
- class_label = self.cls_name2id[self.label_paths[index].split('/')[-1].split('#')[3]] if self.is_train and config.auxiliary_classification else -1
74
-
75
- # loading image and label
76
- if self.is_train:
77
- image, label = preproc(image, label, preproc_methods=config.preproc_methods)
78
- # else:
79
- # if _label.shape[0] > 2048 or _label.shape[1] > 2048:
80
- # _image = cv2.resize(_image, (2048, 2048), interpolation=cv2.INTER_LINEAR)
81
- # _label = cv2.resize(_label, (2048, 2048), interpolation=cv2.INTER_LINEAR)
82
-
83
- image, label = self.transform_image(image), self.transform_label(label)
84
-
85
- if self.is_train:
86
- return image, label, class_label
87
- else:
88
- return image, label, self.label_paths[index]
89
-
90
- def __len__(self):
91
- return len(self.image_paths)
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
models/backbones/build_backbone.py DELETED
@@ -1,44 +0,0 @@
1
- import torch
2
- import torch.nn as nn
3
- from collections import OrderedDict
4
- from torchvision.models import vgg16, vgg16_bn, VGG16_Weights, VGG16_BN_Weights, resnet50, ResNet50_Weights
5
- from models.backbones.pvt_v2 import pvt_v2_b0, pvt_v2_b1, pvt_v2_b2, pvt_v2_b5
6
- from models.backbones.swin_v1 import swin_v1_t, swin_v1_s, swin_v1_b, swin_v1_l
7
- from config import Config
8
-
9
-
10
- config = Config()
11
-
12
- def build_backbone(bb_name, pretrained=True, params_settings=''):
13
- if bb_name == 'vgg16':
14
- bb_net = list(vgg16(pretrained=VGG16_Weights.DEFAULT if pretrained else None).children())[0]
15
- bb = nn.Sequential(OrderedDict({'conv1': bb_net[:4], 'conv2': bb_net[4:9], 'conv3': bb_net[9:16], 'conv4': bb_net[16:23]}))
16
- elif bb_name == 'vgg16bn':
17
- bb_net = list(vgg16_bn(pretrained=VGG16_BN_Weights.DEFAULT if pretrained else None).children())[0]
18
- bb = nn.Sequential(OrderedDict({'conv1': bb_net[:6], 'conv2': bb_net[6:13], 'conv3': bb_net[13:23], 'conv4': bb_net[23:33]}))
19
- elif bb_name == 'resnet50':
20
- bb_net = list(resnet50(pretrained=ResNet50_Weights.DEFAULT if pretrained else None).children())
21
- bb = nn.Sequential(OrderedDict({'conv1': nn.Sequential(*bb_net[0:3]), 'conv2': bb_net[4], 'conv3': bb_net[5], 'conv4': bb_net[6]}))
22
- else:
23
- bb = eval('{}({})'.format(bb_name, params_settings))
24
- if pretrained:
25
- bb = load_weights(bb, bb_name)
26
- return bb
27
-
28
- def load_weights(model, model_name):
29
- save_model = torch.load(config.weights[model_name])
30
- model_dict = model.state_dict()
31
- state_dict = {k: v if v.size() == model_dict[k].size() else model_dict[k] for k, v in save_model.items() if k in model_dict.keys()}
32
- # to ignore the weights with mismatched size when I modify the backbone itself.
33
- if not state_dict:
34
- save_model_keys = list(save_model.keys())
35
- sub_item = save_model_keys[0] if len(save_model_keys) == 1 else None
36
- state_dict = {k: v if v.size() == model_dict[k].size() else model_dict[k] for k, v in save_model[sub_item].items() if k in model_dict.keys()}
37
- if not state_dict or not sub_item:
38
- print('Weights are not successully loaded. Check the state dict of weights file.')
39
- return None
40
- else:
41
- print('Found correct weights in the "{}" item of loaded state_dict.'.format(sub_item))
42
- model_dict.update(state_dict)
43
- model.load_state_dict(model_dict)
44
- return model
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
models/backbones/pvt_v2.py DELETED
@@ -1,435 +0,0 @@
1
- import torch
2
- import torch.nn as nn
3
- from functools import partial
4
-
5
- from timm.models.layers import DropPath, to_2tuple, trunc_normal_
6
- from timm.models.registry import register_model
7
-
8
- import math
9
-
10
- from config import Config
11
-
12
- config = Config()
13
-
14
- class Mlp(nn.Module):
15
- def __init__(self, in_features, hidden_features=None, out_features=None, act_layer=nn.GELU, drop=0.):
16
- super().__init__()
17
- out_features = out_features or in_features
18
- hidden_features = hidden_features or in_features
19
- self.fc1 = nn.Linear(in_features, hidden_features)
20
- self.dwconv = DWConv(hidden_features)
21
- self.act = act_layer()
22
- self.fc2 = nn.Linear(hidden_features, out_features)
23
- self.drop = nn.Dropout(drop)
24
-
25
- self.apply(self._init_weights)
26
-
27
- def _init_weights(self, m):
28
- if isinstance(m, nn.Linear):
29
- trunc_normal_(m.weight, std=.02)
30
- if isinstance(m, nn.Linear) and m.bias is not None:
31
- nn.init.constant_(m.bias, 0)
32
- elif isinstance(m, nn.LayerNorm):
33
- nn.init.constant_(m.bias, 0)
34
- nn.init.constant_(m.weight, 1.0)
35
- elif isinstance(m, nn.Conv2d):
36
- fan_out = m.kernel_size[0] * m.kernel_size[1] * m.out_channels
37
- fan_out //= m.groups
38
- m.weight.data.normal_(0, math.sqrt(2.0 / fan_out))
39
- if m.bias is not None:
40
- m.bias.data.zero_()
41
-
42
- def forward(self, x, H, W):
43
- x = self.fc1(x)
44
- x = self.dwconv(x, H, W)
45
- x = self.act(x)
46
- x = self.drop(x)
47
- x = self.fc2(x)
48
- x = self.drop(x)
49
- return x
50
-
51
-
52
- class Attention(nn.Module):
53
- def __init__(self, dim, num_heads=8, qkv_bias=False, qk_scale=None, attn_drop=0., proj_drop=0., sr_ratio=1):
54
- super().__init__()
55
- assert dim % num_heads == 0, f"dim {dim} should be divided by num_heads {num_heads}."
56
-
57
- self.dim = dim
58
- self.num_heads = num_heads
59
- head_dim = dim // num_heads
60
- self.scale = qk_scale or head_dim ** -0.5
61
-
62
- self.q = nn.Linear(dim, dim, bias=qkv_bias)
63
- self.kv = nn.Linear(dim, dim * 2, bias=qkv_bias)
64
- self.attn_drop_prob = attn_drop
65
- self.attn_drop = nn.Dropout(attn_drop)
66
- self.proj = nn.Linear(dim, dim)
67
- self.proj_drop = nn.Dropout(proj_drop)
68
-
69
- self.sr_ratio = sr_ratio
70
- if sr_ratio > 1:
71
- self.sr = nn.Conv2d(dim, dim, kernel_size=sr_ratio, stride=sr_ratio)
72
- self.norm = nn.LayerNorm(dim)
73
-
74
- self.apply(self._init_weights)
75
-
76
- def _init_weights(self, m):
77
- if isinstance(m, nn.Linear):
78
- trunc_normal_(m.weight, std=.02)
79
- if isinstance(m, nn.Linear) and m.bias is not None:
80
- nn.init.constant_(m.bias, 0)
81
- elif isinstance(m, nn.LayerNorm):
82
- nn.init.constant_(m.bias, 0)
83
- nn.init.constant_(m.weight, 1.0)
84
- elif isinstance(m, nn.Conv2d):
85
- fan_out = m.kernel_size[0] * m.kernel_size[1] * m.out_channels
86
- fan_out //= m.groups
87
- m.weight.data.normal_(0, math.sqrt(2.0 / fan_out))
88
- if m.bias is not None:
89
- m.bias.data.zero_()
90
-
91
- def forward(self, x, H, W):
92
- B, N, C = x.shape
93
- q = self.q(x).reshape(B, N, self.num_heads, C // self.num_heads).permute(0, 2, 1, 3)
94
-
95
- if self.sr_ratio > 1:
96
- x_ = x.permute(0, 2, 1).reshape(B, C, H, W)
97
- x_ = self.sr(x_).reshape(B, C, -1).permute(0, 2, 1)
98
- x_ = self.norm(x_)
99
- kv = self.kv(x_).reshape(B, -1, 2, self.num_heads, C // self.num_heads).permute(2, 0, 3, 1, 4)
100
- else:
101
- kv = self.kv(x).reshape(B, -1, 2, self.num_heads, C // self.num_heads).permute(2, 0, 3, 1, 4)
102
- k, v = kv[0], kv[1]
103
-
104
- if config.SDPA_enabled:
105
- x = torch.nn.functional.scaled_dot_product_attention(
106
- q, k, v,
107
- attn_mask=None, dropout_p=self.attn_drop_prob, is_causal=False
108
- ).transpose(1, 2).reshape(B, N, C)
109
- else:
110
- attn = (q @ k.transpose(-2, -1)) * self.scale
111
- attn = attn.softmax(dim=-1)
112
- attn = self.attn_drop(attn)
113
-
114
- x = (attn @ v).transpose(1, 2).reshape(B, N, C)
115
- x = self.proj(x)
116
- x = self.proj_drop(x)
117
-
118
- return x
119
-
120
-
121
- class Block(nn.Module):
122
-
123
- def __init__(self, dim, num_heads, mlp_ratio=4., qkv_bias=False, qk_scale=None, drop=0., attn_drop=0.,
124
- drop_path=0., act_layer=nn.GELU, norm_layer=nn.LayerNorm, sr_ratio=1):
125
- super().__init__()
126
- self.norm1 = norm_layer(dim)
127
- self.attn = Attention(
128
- dim,
129
- num_heads=num_heads, qkv_bias=qkv_bias, qk_scale=qk_scale,
130
- attn_drop=attn_drop, proj_drop=drop, sr_ratio=sr_ratio)
131
- # NOTE: drop path for stochastic depth, we shall see if this is better than dropout here
132
- self.drop_path = DropPath(drop_path) if drop_path > 0. else nn.Identity()
133
- self.norm2 = norm_layer(dim)
134
- mlp_hidden_dim = int(dim * mlp_ratio)
135
- self.mlp = Mlp(in_features=dim, hidden_features=mlp_hidden_dim, act_layer=act_layer, drop=drop)
136
-
137
- self.apply(self._init_weights)
138
-
139
- def _init_weights(self, m):
140
- if isinstance(m, nn.Linear):
141
- trunc_normal_(m.weight, std=.02)
142
- if isinstance(m, nn.Linear) and m.bias is not None:
143
- nn.init.constant_(m.bias, 0)
144
- elif isinstance(m, nn.LayerNorm):
145
- nn.init.constant_(m.bias, 0)
146
- nn.init.constant_(m.weight, 1.0)
147
- elif isinstance(m, nn.Conv2d):
148
- fan_out = m.kernel_size[0] * m.kernel_size[1] * m.out_channels
149
- fan_out //= m.groups
150
- m.weight.data.normal_(0, math.sqrt(2.0 / fan_out))
151
- if m.bias is not None:
152
- m.bias.data.zero_()
153
-
154
- def forward(self, x, H, W):
155
- x = x + self.drop_path(self.attn(self.norm1(x), H, W))
156
- x = x + self.drop_path(self.mlp(self.norm2(x), H, W))
157
-
158
- return x
159
-
160
-
161
- class OverlapPatchEmbed(nn.Module):
162
- """ Image to Patch Embedding
163
- """
164
-
165
- def __init__(self, img_size=224, patch_size=7, stride=4, in_channels=3, embed_dim=768):
166
- super().__init__()
167
- img_size = to_2tuple(img_size)
168
- patch_size = to_2tuple(patch_size)
169
-
170
- self.img_size = img_size
171
- self.patch_size = patch_size
172
- self.H, self.W = img_size[0] // patch_size[0], img_size[1] // patch_size[1]
173
- self.num_patches = self.H * self.W
174
- self.proj = nn.Conv2d(in_channels, embed_dim, kernel_size=patch_size, stride=stride,
175
- padding=(patch_size[0] // 2, patch_size[1] // 2))
176
- self.norm = nn.LayerNorm(embed_dim)
177
-
178
- self.apply(self._init_weights)
179
-
180
- def _init_weights(self, m):
181
- if isinstance(m, nn.Linear):
182
- trunc_normal_(m.weight, std=.02)
183
- if isinstance(m, nn.Linear) and m.bias is not None:
184
- nn.init.constant_(m.bias, 0)
185
- elif isinstance(m, nn.LayerNorm):
186
- nn.init.constant_(m.bias, 0)
187
- nn.init.constant_(m.weight, 1.0)
188
- elif isinstance(m, nn.Conv2d):
189
- fan_out = m.kernel_size[0] * m.kernel_size[1] * m.out_channels
190
- fan_out //= m.groups
191
- m.weight.data.normal_(0, math.sqrt(2.0 / fan_out))
192
- if m.bias is not None:
193
- m.bias.data.zero_()
194
-
195
- def forward(self, x):
196
- x = self.proj(x)
197
- _, _, H, W = x.shape
198
- x = x.flatten(2).transpose(1, 2)
199
- x = self.norm(x)
200
-
201
- return x, H, W
202
-
203
-
204
- class PyramidVisionTransformerImpr(nn.Module):
205
- def __init__(self, img_size=224, patch_size=16, in_channels=3, num_classes=1000, embed_dims=[64, 128, 256, 512],
206
- num_heads=[1, 2, 4, 8], mlp_ratios=[4, 4, 4, 4], qkv_bias=False, qk_scale=None, drop_rate=0.,
207
- attn_drop_rate=0., drop_path_rate=0., norm_layer=nn.LayerNorm,
208
- depths=[3, 4, 6, 3], sr_ratios=[8, 4, 2, 1]):
209
- super().__init__()
210
- self.num_classes = num_classes
211
- self.depths = depths
212
-
213
- # patch_embed
214
- self.patch_embed1 = OverlapPatchEmbed(img_size=img_size, patch_size=7, stride=4, in_channels=in_channels,
215
- embed_dim=embed_dims[0])
216
- self.patch_embed2 = OverlapPatchEmbed(img_size=img_size // 4, patch_size=3, stride=2, in_channels=embed_dims[0],
217
- embed_dim=embed_dims[1])
218
- self.patch_embed3 = OverlapPatchEmbed(img_size=img_size // 8, patch_size=3, stride=2, in_channels=embed_dims[1],
219
- embed_dim=embed_dims[2])
220
- self.patch_embed4 = OverlapPatchEmbed(img_size=img_size // 16, patch_size=3, stride=2, in_channels=embed_dims[2],
221
- embed_dim=embed_dims[3])
222
-
223
- # transformer encoder
224
- dpr = [x.item() for x in torch.linspace(0, drop_path_rate, sum(depths))] # stochastic depth decay rule
225
- cur = 0
226
- self.block1 = nn.ModuleList([Block(
227
- dim=embed_dims[0], num_heads=num_heads[0], mlp_ratio=mlp_ratios[0], qkv_bias=qkv_bias, qk_scale=qk_scale,
228
- drop=drop_rate, attn_drop=attn_drop_rate, drop_path=dpr[cur + i], norm_layer=norm_layer,
229
- sr_ratio=sr_ratios[0])
230
- for i in range(depths[0])])
231
- self.norm1 = norm_layer(embed_dims[0])
232
-
233
- cur += depths[0]
234
- self.block2 = nn.ModuleList([Block(
235
- dim=embed_dims[1], num_heads=num_heads[1], mlp_ratio=mlp_ratios[1], qkv_bias=qkv_bias, qk_scale=qk_scale,
236
- drop=drop_rate, attn_drop=attn_drop_rate, drop_path=dpr[cur + i], norm_layer=norm_layer,
237
- sr_ratio=sr_ratios[1])
238
- for i in range(depths[1])])
239
- self.norm2 = norm_layer(embed_dims[1])
240
-
241
- cur += depths[1]
242
- self.block3 = nn.ModuleList([Block(
243
- dim=embed_dims[2], num_heads=num_heads[2], mlp_ratio=mlp_ratios[2], qkv_bias=qkv_bias, qk_scale=qk_scale,
244
- drop=drop_rate, attn_drop=attn_drop_rate, drop_path=dpr[cur + i], norm_layer=norm_layer,
245
- sr_ratio=sr_ratios[2])
246
- for i in range(depths[2])])
247
- self.norm3 = norm_layer(embed_dims[2])
248
-
249
- cur += depths[2]
250
- self.block4 = nn.ModuleList([Block(
251
- dim=embed_dims[3], num_heads=num_heads[3], mlp_ratio=mlp_ratios[3], qkv_bias=qkv_bias, qk_scale=qk_scale,
252
- drop=drop_rate, attn_drop=attn_drop_rate, drop_path=dpr[cur + i], norm_layer=norm_layer,
253
- sr_ratio=sr_ratios[3])
254
- for i in range(depths[3])])
255
- self.norm4 = norm_layer(embed_dims[3])
256
-
257
- # classification head
258
- # self.head = nn.Linear(embed_dims[3], num_classes) if num_classes > 0 else nn.Identity()
259
-
260
- self.apply(self._init_weights)
261
-
262
- def _init_weights(self, m):
263
- if isinstance(m, nn.Linear):
264
- trunc_normal_(m.weight, std=.02)
265
- if isinstance(m, nn.Linear) and m.bias is not None:
266
- nn.init.constant_(m.bias, 0)
267
- elif isinstance(m, nn.LayerNorm):
268
- nn.init.constant_(m.bias, 0)
269
- nn.init.constant_(m.weight, 1.0)
270
- elif isinstance(m, nn.Conv2d):
271
- fan_out = m.kernel_size[0] * m.kernel_size[1] * m.out_channels
272
- fan_out //= m.groups
273
- m.weight.data.normal_(0, math.sqrt(2.0 / fan_out))
274
- if m.bias is not None:
275
- m.bias.data.zero_()
276
-
277
- def init_weights(self, pretrained=None):
278
- if isinstance(pretrained, str):
279
- logger = 1
280
- #load_checkpoint(self, pretrained, map_location='cpu', strict=False, logger=logger)
281
-
282
- def reset_drop_path(self, drop_path_rate):
283
- dpr = [x.item() for x in torch.linspace(0, drop_path_rate, sum(self.depths))]
284
- cur = 0
285
- for i in range(self.depths[0]):
286
- self.block1[i].drop_path.drop_prob = dpr[cur + i]
287
-
288
- cur += self.depths[0]
289
- for i in range(self.depths[1]):
290
- self.block2[i].drop_path.drop_prob = dpr[cur + i]
291
-
292
- cur += self.depths[1]
293
- for i in range(self.depths[2]):
294
- self.block3[i].drop_path.drop_prob = dpr[cur + i]
295
-
296
- cur += self.depths[2]
297
- for i in range(self.depths[3]):
298
- self.block4[i].drop_path.drop_prob = dpr[cur + i]
299
-
300
- def freeze_patch_emb(self):
301
- self.patch_embed1.requires_grad = False
302
-
303
- @torch.jit.ignore
304
- def no_weight_decay(self):
305
- return {'pos_embed1', 'pos_embed2', 'pos_embed3', 'pos_embed4', 'cls_token'} # has pos_embed may be better
306
-
307
- def get_classifier(self):
308
- return self.head
309
-
310
- def reset_classifier(self, num_classes, global_pool=''):
311
- self.num_classes = num_classes
312
- self.head = nn.Linear(self.embed_dim, num_classes) if num_classes > 0 else nn.Identity()
313
-
314
- def forward_features(self, x):
315
- B = x.shape[0]
316
- outs = []
317
-
318
- # stage 1
319
- x, H, W = self.patch_embed1(x)
320
- for i, blk in enumerate(self.block1):
321
- x = blk(x, H, W)
322
- x = self.norm1(x)
323
- x = x.reshape(B, H, W, -1).permute(0, 3, 1, 2).contiguous()
324
- outs.append(x)
325
-
326
- # stage 2
327
- x, H, W = self.patch_embed2(x)
328
- for i, blk in enumerate(self.block2):
329
- x = blk(x, H, W)
330
- x = self.norm2(x)
331
- x = x.reshape(B, H, W, -1).permute(0, 3, 1, 2).contiguous()
332
- outs.append(x)
333
-
334
- # stage 3
335
- x, H, W = self.patch_embed3(x)
336
- for i, blk in enumerate(self.block3):
337
- x = blk(x, H, W)
338
- x = self.norm3(x)
339
- x = x.reshape(B, H, W, -1).permute(0, 3, 1, 2).contiguous()
340
- outs.append(x)
341
-
342
- # stage 4
343
- x, H, W = self.patch_embed4(x)
344
- for i, blk in enumerate(self.block4):
345
- x = blk(x, H, W)
346
- x = self.norm4(x)
347
- x = x.reshape(B, H, W, -1).permute(0, 3, 1, 2).contiguous()
348
- outs.append(x)
349
-
350
- return outs
351
-
352
- # return x.mean(dim=1)
353
-
354
- def forward(self, x):
355
- x = self.forward_features(x)
356
- # x = self.head(x)
357
-
358
- return x
359
-
360
-
361
- class DWConv(nn.Module):
362
- def __init__(self, dim=768):
363
- super(DWConv, self).__init__()
364
- self.dwconv = nn.Conv2d(dim, dim, 3, 1, 1, bias=True, groups=dim)
365
-
366
- def forward(self, x, H, W):
367
- B, N, C = x.shape
368
- x = x.transpose(1, 2).view(B, C, H, W).contiguous()
369
- x = self.dwconv(x)
370
- x = x.flatten(2).transpose(1, 2)
371
-
372
- return x
373
-
374
-
375
- def _conv_filter(state_dict, patch_size=16):
376
- """ convert patch embedding weight from manual patchify + linear proj to conv"""
377
- out_dict = {}
378
- for k, v in state_dict.items():
379
- if 'patch_embed.proj.weight' in k:
380
- v = v.reshape((v.shape[0], 3, patch_size, patch_size))
381
- out_dict[k] = v
382
-
383
- return out_dict
384
-
385
-
386
- ## @register_model
387
- class pvt_v2_b0(PyramidVisionTransformerImpr):
388
- def __init__(self, **kwargs):
389
- super(pvt_v2_b0, self).__init__(
390
- patch_size=4, embed_dims=[32, 64, 160, 256], num_heads=[1, 2, 5, 8], mlp_ratios=[8, 8, 4, 4],
391
- qkv_bias=True, norm_layer=partial(nn.LayerNorm, eps=1e-6), depths=[2, 2, 2, 2], sr_ratios=[8, 4, 2, 1],
392
- drop_rate=0.0, drop_path_rate=0.1)
393
-
394
-
395
-
396
- ## @register_model
397
- class pvt_v2_b1(PyramidVisionTransformerImpr):
398
- def __init__(self, **kwargs):
399
- super(pvt_v2_b1, self).__init__(
400
- patch_size=4, embed_dims=[64, 128, 320, 512], num_heads=[1, 2, 5, 8], mlp_ratios=[8, 8, 4, 4],
401
- qkv_bias=True, norm_layer=partial(nn.LayerNorm, eps=1e-6), depths=[2, 2, 2, 2], sr_ratios=[8, 4, 2, 1],
402
- drop_rate=0.0, drop_path_rate=0.1)
403
-
404
- ## @register_model
405
- class pvt_v2_b2(PyramidVisionTransformerImpr):
406
- def __init__(self, in_channels=3, **kwargs):
407
- super(pvt_v2_b2, self).__init__(
408
- patch_size=4, embed_dims=[64, 128, 320, 512], num_heads=[1, 2, 5, 8], mlp_ratios=[8, 8, 4, 4],
409
- qkv_bias=True, norm_layer=partial(nn.LayerNorm, eps=1e-6), depths=[3, 4, 6, 3], sr_ratios=[8, 4, 2, 1],
410
- drop_rate=0.0, drop_path_rate=0.1, in_channels=in_channels)
411
-
412
- ## @register_model
413
- class pvt_v2_b3(PyramidVisionTransformerImpr):
414
- def __init__(self, **kwargs):
415
- super(pvt_v2_b3, self).__init__(
416
- patch_size=4, embed_dims=[64, 128, 320, 512], num_heads=[1, 2, 5, 8], mlp_ratios=[8, 8, 4, 4],
417
- qkv_bias=True, norm_layer=partial(nn.LayerNorm, eps=1e-6), depths=[3, 4, 18, 3], sr_ratios=[8, 4, 2, 1],
418
- drop_rate=0.0, drop_path_rate=0.1)
419
-
420
- ## @register_model
421
- class pvt_v2_b4(PyramidVisionTransformerImpr):
422
- def __init__(self, **kwargs):
423
- super(pvt_v2_b4, self).__init__(
424
- patch_size=4, embed_dims=[64, 128, 320, 512], num_heads=[1, 2, 5, 8], mlp_ratios=[8, 8, 4, 4],
425
- qkv_bias=True, norm_layer=partial(nn.LayerNorm, eps=1e-6), depths=[3, 8, 27, 3], sr_ratios=[8, 4, 2, 1],
426
- drop_rate=0.0, drop_path_rate=0.1)
427
-
428
-
429
- ## @register_model
430
- class pvt_v2_b5(PyramidVisionTransformerImpr):
431
- def __init__(self, **kwargs):
432
- super(pvt_v2_b5, self).__init__(
433
- patch_size=4, embed_dims=[64, 128, 320, 512], num_heads=[1, 2, 5, 8], mlp_ratios=[4, 4, 4, 4],
434
- qkv_bias=True, norm_layer=partial(nn.LayerNorm, eps=1e-6), depths=[3, 6, 40, 3], sr_ratios=[8, 4, 2, 1],
435
- drop_rate=0.0, drop_path_rate=0.1)
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
models/backbones/swin_v1.py DELETED
@@ -1,627 +0,0 @@
1
- # --------------------------------------------------------
2
- # Swin Transformer
3
- # Copyright (c) 2021 Microsoft
4
- # Licensed under The MIT License [see LICENSE for details]
5
- # Written by Ze Liu, Yutong Lin, Yixuan Wei
6
- # --------------------------------------------------------
7
-
8
- import torch
9
- import torch.nn as nn
10
- import torch.nn.functional as F
11
- import torch.utils.checkpoint as checkpoint
12
- import numpy as np
13
- from timm.models.layers import DropPath, to_2tuple, trunc_normal_
14
-
15
- from config import Config
16
-
17
-
18
- config = Config()
19
-
20
- class Mlp(nn.Module):
21
- """ Multilayer perceptron."""
22
-
23
- def __init__(self, in_features, hidden_features=None, out_features=None, act_layer=nn.GELU, drop=0.):
24
- super().__init__()
25
- out_features = out_features or in_features
26
- hidden_features = hidden_features or in_features
27
- self.fc1 = nn.Linear(in_features, hidden_features)
28
- self.act = act_layer()
29
- self.fc2 = nn.Linear(hidden_features, out_features)
30
- self.drop = nn.Dropout(drop)
31
-
32
- def forward(self, x):
33
- x = self.fc1(x)
34
- x = self.act(x)
35
- x = self.drop(x)
36
- x = self.fc2(x)
37
- x = self.drop(x)
38
- return x
39
-
40
-
41
- def window_partition(x, window_size):
42
- """
43
- Args:
44
- x: (B, H, W, C)
45
- window_size (int): window size
46
-
47
- Returns:
48
- windows: (num_windows*B, window_size, window_size, C)
49
- """
50
- B, H, W, C = x.shape
51
- x = x.view(B, H // window_size, window_size, W // window_size, window_size, C)
52
- windows = x.permute(0, 1, 3, 2, 4, 5).contiguous().view(-1, window_size, window_size, C)
53
- return windows
54
-
55
-
56
- def window_reverse(windows, window_size, H, W):
57
- """
58
- Args:
59
- windows: (num_windows*B, window_size, window_size, C)
60
- window_size (int): Window size
61
- H (int): Height of image
62
- W (int): Width of image
63
-
64
- Returns:
65
- x: (B, H, W, C)
66
- """
67
- B = int(windows.shape[0] / (H * W / window_size / window_size))
68
- x = windows.view(B, H // window_size, W // window_size, window_size, window_size, -1)
69
- x = x.permute(0, 1, 3, 2, 4, 5).contiguous().view(B, H, W, -1)
70
- return x
71
-
72
-
73
- class WindowAttention(nn.Module):
74
- """ Window based multi-head self attention (W-MSA) module with relative position bias.
75
- It supports both of shifted and non-shifted window.
76
-
77
- Args:
78
- dim (int): Number of input channels.
79
- window_size (tuple[int]): The height and width of the window.
80
- num_heads (int): Number of attention heads.
81
- qkv_bias (bool, optional): If True, add a learnable bias to query, key, value. Default: True
82
- qk_scale (float | None, optional): Override default qk scale of head_dim ** -0.5 if set
83
- attn_drop (float, optional): Dropout ratio of attention weight. Default: 0.0
84
- proj_drop (float, optional): Dropout ratio of output. Default: 0.0
85
- """
86
-
87
- def __init__(self, dim, window_size, num_heads, qkv_bias=True, qk_scale=None, attn_drop=0., proj_drop=0.):
88
-
89
- super().__init__()
90
- self.dim = dim
91
- self.window_size = window_size # Wh, Ww
92
- self.num_heads = num_heads
93
- head_dim = dim // num_heads
94
- self.scale = qk_scale or head_dim ** -0.5
95
-
96
- # define a parameter table of relative position bias
97
- self.relative_position_bias_table = nn.Parameter(
98
- torch.zeros((2 * window_size[0] - 1) * (2 * window_size[1] - 1), num_heads)) # 2*Wh-1 * 2*Ww-1, nH
99
-
100
- # get pair-wise relative position index for each token inside the window
101
- coords_h = torch.arange(self.window_size[0])
102
- coords_w = torch.arange(self.window_size[1])
103
- coords = torch.stack(torch.meshgrid([coords_h, coords_w], indexing='ij')) # 2, Wh, Ww
104
- coords_flatten = torch.flatten(coords, 1) # 2, Wh*Ww
105
- relative_coords = coords_flatten[:, :, None] - coords_flatten[:, None, :] # 2, Wh*Ww, Wh*Ww
106
- relative_coords = relative_coords.permute(1, 2, 0).contiguous() # Wh*Ww, Wh*Ww, 2
107
- relative_coords[:, :, 0] += self.window_size[0] - 1 # shift to start from 0
108
- relative_coords[:, :, 1] += self.window_size[1] - 1
109
- relative_coords[:, :, 0] *= 2 * self.window_size[1] - 1
110
- relative_position_index = relative_coords.sum(-1) # Wh*Ww, Wh*Ww
111
- self.register_buffer("relative_position_index", relative_position_index)
112
-
113
- self.qkv = nn.Linear(dim, dim * 3, bias=qkv_bias)
114
- self.attn_drop_prob = attn_drop
115
- self.attn_drop = nn.Dropout(attn_drop)
116
- self.proj = nn.Linear(dim, dim)
117
- self.proj_drop = nn.Dropout(proj_drop)
118
-
119
- trunc_normal_(self.relative_position_bias_table, std=.02)
120
- self.softmax = nn.Softmax(dim=-1)
121
-
122
- def forward(self, x, mask=None):
123
- """ Forward function.
124
-
125
- Args:
126
- x: input features with shape of (num_windows*B, N, C)
127
- mask: (0/-inf) mask with shape of (num_windows, Wh*Ww, Wh*Ww) or None
128
- """
129
- B_, N, C = x.shape
130
- qkv = self.qkv(x).reshape(B_, N, 3, self.num_heads, C // self.num_heads).permute(2, 0, 3, 1, 4)
131
- q, k, v = qkv[0], qkv[1], qkv[2] # make torchscript happy (cannot use tensor as tuple)
132
-
133
- q = q * self.scale
134
-
135
- if config.SDPA_enabled:
136
- x = torch.nn.functional.scaled_dot_product_attention(
137
- q, k, v,
138
- attn_mask=None, dropout_p=self.attn_drop_prob, is_causal=False
139
- ).transpose(1, 2).reshape(B_, N, C)
140
- else:
141
- attn = (q @ k.transpose(-2, -1))
142
-
143
- relative_position_bias = self.relative_position_bias_table[self.relative_position_index.view(-1)].view(
144
- self.window_size[0] * self.window_size[1], self.window_size[0] * self.window_size[1], -1) # Wh*Ww,Wh*Ww,nH
145
- relative_position_bias = relative_position_bias.permute(2, 0, 1).contiguous() # nH, Wh*Ww, Wh*Ww
146
- attn = attn + relative_position_bias.unsqueeze(0)
147
-
148
- if mask is not None:
149
- nW = mask.shape[0]
150
- attn = attn.view(B_ // nW, nW, self.num_heads, N, N) + mask.unsqueeze(1).unsqueeze(0)
151
- attn = attn.view(-1, self.num_heads, N, N)
152
- attn = self.softmax(attn)
153
- else:
154
- attn = self.softmax(attn)
155
-
156
- attn = self.attn_drop(attn)
157
-
158
- x = (attn @ v).transpose(1, 2).reshape(B_, N, C)
159
- x = self.proj(x)
160
- x = self.proj_drop(x)
161
- return x
162
-
163
-
164
- class SwinTransformerBlock(nn.Module):
165
- """ Swin Transformer Block.
166
-
167
- Args:
168
- dim (int): Number of input channels.
169
- num_heads (int): Number of attention heads.
170
- window_size (int): Window size.
171
- shift_size (int): Shift size for SW-MSA.
172
- mlp_ratio (float): Ratio of mlp hidden dim to embedding dim.
173
- qkv_bias (bool, optional): If True, add a learnable bias to query, key, value. Default: True
174
- qk_scale (float | None, optional): Override default qk scale of head_dim ** -0.5 if set.
175
- drop (float, optional): Dropout rate. Default: 0.0
176
- attn_drop (float, optional): Attention dropout rate. Default: 0.0
177
- drop_path (float, optional): Stochastic depth rate. Default: 0.0
178
- act_layer (nn.Module, optional): Activation layer. Default: nn.GELU
179
- norm_layer (nn.Module, optional): Normalization layer. Default: nn.LayerNorm
180
- """
181
-
182
- def __init__(self, dim, num_heads, window_size=7, shift_size=0,
183
- mlp_ratio=4., qkv_bias=True, qk_scale=None, drop=0., attn_drop=0., drop_path=0.,
184
- act_layer=nn.GELU, norm_layer=nn.LayerNorm):
185
- super().__init__()
186
- self.dim = dim
187
- self.num_heads = num_heads
188
- self.window_size = window_size
189
- self.shift_size = shift_size
190
- self.mlp_ratio = mlp_ratio
191
- assert 0 <= self.shift_size < self.window_size, "shift_size must in 0-window_size"
192
-
193
- self.norm1 = norm_layer(dim)
194
- self.attn = WindowAttention(
195
- dim, window_size=to_2tuple(self.window_size), num_heads=num_heads,
196
- qkv_bias=qkv_bias, qk_scale=qk_scale, attn_drop=attn_drop, proj_drop=drop)
197
-
198
- self.drop_path = DropPath(drop_path) if drop_path > 0. else nn.Identity()
199
- self.norm2 = norm_layer(dim)
200
- mlp_hidden_dim = int(dim * mlp_ratio)
201
- self.mlp = Mlp(in_features=dim, hidden_features=mlp_hidden_dim, act_layer=act_layer, drop=drop)
202
-
203
- self.H = None
204
- self.W = None
205
-
206
- def forward(self, x, mask_matrix):
207
- """ Forward function.
208
-
209
- Args:
210
- x: Input feature, tensor size (B, H*W, C).
211
- H, W: Spatial resolution of the input feature.
212
- mask_matrix: Attention mask for cyclic shift.
213
- """
214
- B, L, C = x.shape
215
- H, W = self.H, self.W
216
- assert L == H * W, "input feature has wrong size"
217
-
218
- shortcut = x
219
- x = self.norm1(x)
220
- x = x.view(B, H, W, C)
221
-
222
- # pad feature maps to multiples of window size
223
- pad_l = pad_t = 0
224
- pad_r = (self.window_size - W % self.window_size) % self.window_size
225
- pad_b = (self.window_size - H % self.window_size) % self.window_size
226
- x = F.pad(x, (0, 0, pad_l, pad_r, pad_t, pad_b))
227
- _, Hp, Wp, _ = x.shape
228
-
229
- # cyclic shift
230
- if self.shift_size > 0:
231
- shifted_x = torch.roll(x, shifts=(-self.shift_size, -self.shift_size), dims=(1, 2))
232
- attn_mask = mask_matrix
233
- else:
234
- shifted_x = x
235
- attn_mask = None
236
-
237
- # partition windows
238
- x_windows = window_partition(shifted_x, self.window_size) # nW*B, window_size, window_size, C
239
- x_windows = x_windows.view(-1, self.window_size * self.window_size, C) # nW*B, window_size*window_size, C
240
-
241
- # W-MSA/SW-MSA
242
- attn_windows = self.attn(x_windows, mask=attn_mask) # nW*B, window_size*window_size, C
243
-
244
- # merge windows
245
- attn_windows = attn_windows.view(-1, self.window_size, self.window_size, C)
246
- shifted_x = window_reverse(attn_windows, self.window_size, Hp, Wp) # B H' W' C
247
-
248
- # reverse cyclic shift
249
- if self.shift_size > 0:
250
- x = torch.roll(shifted_x, shifts=(self.shift_size, self.shift_size), dims=(1, 2))
251
- else:
252
- x = shifted_x
253
-
254
- if pad_r > 0 or pad_b > 0:
255
- x = x[:, :H, :W, :].contiguous()
256
-
257
- x = x.view(B, H * W, C)
258
-
259
- # FFN
260
- x = shortcut + self.drop_path(x)
261
- x = x + self.drop_path(self.mlp(self.norm2(x)))
262
-
263
- return x
264
-
265
-
266
- class PatchMerging(nn.Module):
267
- """ Patch Merging Layer
268
-
269
- Args:
270
- dim (int): Number of input channels.
271
- norm_layer (nn.Module, optional): Normalization layer. Default: nn.LayerNorm
272
- """
273
- def __init__(self, dim, norm_layer=nn.LayerNorm):
274
- super().__init__()
275
- self.dim = dim
276
- self.reduction = nn.Linear(4 * dim, 2 * dim, bias=False)
277
- self.norm = norm_layer(4 * dim)
278
-
279
- def forward(self, x, H, W):
280
- """ Forward function.
281
-
282
- Args:
283
- x: Input feature, tensor size (B, H*W, C).
284
- H, W: Spatial resolution of the input feature.
285
- """
286
- B, L, C = x.shape
287
- assert L == H * W, "input feature has wrong size"
288
-
289
- x = x.view(B, H, W, C)
290
-
291
- # padding
292
- pad_input = (H % 2 == 1) or (W % 2 == 1)
293
- if pad_input:
294
- x = F.pad(x, (0, 0, 0, W % 2, 0, H % 2))
295
-
296
- x0 = x[:, 0::2, 0::2, :] # B H/2 W/2 C
297
- x1 = x[:, 1::2, 0::2, :] # B H/2 W/2 C
298
- x2 = x[:, 0::2, 1::2, :] # B H/2 W/2 C
299
- x3 = x[:, 1::2, 1::2, :] # B H/2 W/2 C
300
- x = torch.cat([x0, x1, x2, x3], -1) # B H/2 W/2 4*C
301
- x = x.view(B, -1, 4 * C) # B H/2*W/2 4*C
302
-
303
- x = self.norm(x)
304
- x = self.reduction(x)
305
-
306
- return x
307
-
308
-
309
- class BasicLayer(nn.Module):
310
- """ A basic Swin Transformer layer for one stage.
311
-
312
- Args:
313
- dim (int): Number of feature channels
314
- depth (int): Depths of this stage.
315
- num_heads (int): Number of attention head.
316
- window_size (int): Local window size. Default: 7.
317
- mlp_ratio (float): Ratio of mlp hidden dim to embedding dim. Default: 4.
318
- qkv_bias (bool, optional): If True, add a learnable bias to query, key, value. Default: True
319
- qk_scale (float | None, optional): Override default qk scale of head_dim ** -0.5 if set.
320
- drop (float, optional): Dropout rate. Default: 0.0
321
- attn_drop (float, optional): Attention dropout rate. Default: 0.0
322
- drop_path (float | tuple[float], optional): Stochastic depth rate. Default: 0.0
323
- norm_layer (nn.Module, optional): Normalization layer. Default: nn.LayerNorm
324
- downsample (nn.Module | None, optional): Downsample layer at the end of the layer. Default: None
325
- use_checkpoint (bool): Whether to use checkpointing to save memory. Default: False.
326
- """
327
-
328
- def __init__(self,
329
- dim,
330
- depth,
331
- num_heads,
332
- window_size=7,
333
- mlp_ratio=4.,
334
- qkv_bias=True,
335
- qk_scale=None,
336
- drop=0.,
337
- attn_drop=0.,
338
- drop_path=0.,
339
- norm_layer=nn.LayerNorm,
340
- downsample=None,
341
- use_checkpoint=False):
342
- super().__init__()
343
- self.window_size = window_size
344
- self.shift_size = window_size // 2
345
- self.depth = depth
346
- self.use_checkpoint = use_checkpoint
347
-
348
- # build blocks
349
- self.blocks = nn.ModuleList([
350
- SwinTransformerBlock(
351
- dim=dim,
352
- num_heads=num_heads,
353
- window_size=window_size,
354
- shift_size=0 if (i % 2 == 0) else window_size // 2,
355
- mlp_ratio=mlp_ratio,
356
- qkv_bias=qkv_bias,
357
- qk_scale=qk_scale,
358
- drop=drop,
359
- attn_drop=attn_drop,
360
- drop_path=drop_path[i] if isinstance(drop_path, list) else drop_path,
361
- norm_layer=norm_layer)
362
- for i in range(depth)])
363
-
364
- # patch merging layer
365
- if downsample is not None:
366
- self.downsample = downsample(dim=dim, norm_layer=norm_layer)
367
- else:
368
- self.downsample = None
369
-
370
- def forward(self, x, H, W):
371
- """ Forward function.
372
-
373
- Args:
374
- x: Input feature, tensor size (B, H*W, C).
375
- H, W: Spatial resolution of the input feature.
376
- """
377
-
378
- # calculate attention mask for SW-MSA
379
- Hp = int(np.ceil(H / self.window_size)) * self.window_size
380
- Wp = int(np.ceil(W / self.window_size)) * self.window_size
381
- img_mask = torch.zeros((1, Hp, Wp, 1), device=x.device) # 1 Hp Wp 1
382
- h_slices = (slice(0, -self.window_size),
383
- slice(-self.window_size, -self.shift_size),
384
- slice(-self.shift_size, None))
385
- w_slices = (slice(0, -self.window_size),
386
- slice(-self.window_size, -self.shift_size),
387
- slice(-self.shift_size, None))
388
- cnt = 0
389
- for h in h_slices:
390
- for w in w_slices:
391
- img_mask[:, h, w, :] = cnt
392
- cnt += 1
393
-
394
- mask_windows = window_partition(img_mask, self.window_size) # nW, window_size, window_size, 1
395
- mask_windows = mask_windows.view(-1, self.window_size * self.window_size)
396
- attn_mask = mask_windows.unsqueeze(1) - mask_windows.unsqueeze(2)
397
- attn_mask = attn_mask.masked_fill(attn_mask != 0, float(-100.0)).masked_fill(attn_mask == 0, float(0.0))
398
-
399
- for blk in self.blocks:
400
- blk.H, blk.W = H, W
401
- if self.use_checkpoint:
402
- x = checkpoint.checkpoint(blk, x, attn_mask)
403
- else:
404
- x = blk(x, attn_mask)
405
- if self.downsample is not None:
406
- x_down = self.downsample(x, H, W)
407
- Wh, Ww = (H + 1) // 2, (W + 1) // 2
408
- return x, H, W, x_down, Wh, Ww
409
- else:
410
- return x, H, W, x, H, W
411
-
412
-
413
- class PatchEmbed(nn.Module):
414
- """ Image to Patch Embedding
415
-
416
- Args:
417
- patch_size (int): Patch token size. Default: 4.
418
- in_channels (int): Number of input image channels. Default: 3.
419
- embed_dim (int): Number of linear projection output channels. Default: 96.
420
- norm_layer (nn.Module, optional): Normalization layer. Default: None
421
- """
422
-
423
- def __init__(self, patch_size=4, in_channels=3, embed_dim=96, norm_layer=None):
424
- super().__init__()
425
- patch_size = to_2tuple(patch_size)
426
- self.patch_size = patch_size
427
-
428
- self.in_channels = in_channels
429
- self.embed_dim = embed_dim
430
-
431
- self.proj = nn.Conv2d(in_channels, embed_dim, kernel_size=patch_size, stride=patch_size)
432
- if norm_layer is not None:
433
- self.norm = norm_layer(embed_dim)
434
- else:
435
- self.norm = None
436
-
437
- def forward(self, x):
438
- """Forward function."""
439
- # padding
440
- _, _, H, W = x.size()
441
- if W % self.patch_size[1] != 0:
442
- x = F.pad(x, (0, self.patch_size[1] - W % self.patch_size[1]))
443
- if H % self.patch_size[0] != 0:
444
- x = F.pad(x, (0, 0, 0, self.patch_size[0] - H % self.patch_size[0]))
445
-
446
- x = self.proj(x) # B C Wh Ww
447
- if self.norm is not None:
448
- Wh, Ww = x.size(2), x.size(3)
449
- x = x.flatten(2).transpose(1, 2)
450
- x = self.norm(x)
451
- x = x.transpose(1, 2).view(-1, self.embed_dim, Wh, Ww)
452
-
453
- return x
454
-
455
-
456
- class SwinTransformer(nn.Module):
457
- """ Swin Transformer backbone.
458
- A PyTorch impl of : `Swin Transformer: Hierarchical Vision Transformer using Shifted Windows` -
459
- https://arxiv.org/pdf/2103.14030
460
-
461
- Args:
462
- pretrain_img_size (int): Input image size for training the pretrained model,
463
- used in absolute postion embedding. Default 224.
464
- patch_size (int | tuple(int)): Patch size. Default: 4.
465
- in_channels (int): Number of input image channels. Default: 3.
466
- embed_dim (int): Number of linear projection output channels. Default: 96.
467
- depths (tuple[int]): Depths of each Swin Transformer stage.
468
- num_heads (tuple[int]): Number of attention head of each stage.
469
- window_size (int): Window size. Default: 7.
470
- mlp_ratio (float): Ratio of mlp hidden dim to embedding dim. Default: 4.
471
- qkv_bias (bool): If True, add a learnable bias to query, key, value. Default: True
472
- qk_scale (float): Override default qk scale of head_dim ** -0.5 if set.
473
- drop_rate (float): Dropout rate.
474
- attn_drop_rate (float): Attention dropout rate. Default: 0.
475
- drop_path_rate (float): Stochastic depth rate. Default: 0.2.
476
- norm_layer (nn.Module): Normalization layer. Default: nn.LayerNorm.
477
- ape (bool): If True, add absolute position embedding to the patch embedding. Default: False.
478
- patch_norm (bool): If True, add normalization after patch embedding. Default: True.
479
- out_indices (Sequence[int]): Output from which stages.
480
- frozen_stages (int): Stages to be frozen (stop grad and set eval mode).
481
- -1 means not freezing any parameters.
482
- use_checkpoint (bool): Whether to use checkpointing to save memory. Default: False.
483
- """
484
-
485
- def __init__(self,
486
- pretrain_img_size=224,
487
- patch_size=4,
488
- in_channels=3,
489
- embed_dim=96,
490
- depths=[2, 2, 6, 2],
491
- num_heads=[3, 6, 12, 24],
492
- window_size=7,
493
- mlp_ratio=4.,
494
- qkv_bias=True,
495
- qk_scale=None,
496
- drop_rate=0.,
497
- attn_drop_rate=0.,
498
- drop_path_rate=0.2,
499
- norm_layer=nn.LayerNorm,
500
- ape=False,
501
- patch_norm=True,
502
- out_indices=(0, 1, 2, 3),
503
- frozen_stages=-1,
504
- use_checkpoint=False):
505
- super().__init__()
506
-
507
- self.pretrain_img_size = pretrain_img_size
508
- self.num_layers = len(depths)
509
- self.embed_dim = embed_dim
510
- self.ape = ape
511
- self.patch_norm = patch_norm
512
- self.out_indices = out_indices
513
- self.frozen_stages = frozen_stages
514
-
515
- # split image into non-overlapping patches
516
- self.patch_embed = PatchEmbed(
517
- patch_size=patch_size, in_channels=in_channels, embed_dim=embed_dim,
518
- norm_layer=norm_layer if self.patch_norm else None)
519
-
520
- # absolute position embedding
521
- if self.ape:
522
- pretrain_img_size = to_2tuple(pretrain_img_size)
523
- patch_size = to_2tuple(patch_size)
524
- patches_resolution = [pretrain_img_size[0] // patch_size[0], pretrain_img_size[1] // patch_size[1]]
525
-
526
- self.absolute_pos_embed = nn.Parameter(torch.zeros(1, embed_dim, patches_resolution[0], patches_resolution[1]))
527
- trunc_normal_(self.absolute_pos_embed, std=.02)
528
-
529
- self.pos_drop = nn.Dropout(p=drop_rate)
530
-
531
- # stochastic depth
532
- dpr = [x.item() for x in torch.linspace(0, drop_path_rate, sum(depths))] # stochastic depth decay rule
533
-
534
- # build layers
535
- self.layers = nn.ModuleList()
536
- for i_layer in range(self.num_layers):
537
- layer = BasicLayer(
538
- dim=int(embed_dim * 2 ** i_layer),
539
- depth=depths[i_layer],
540
- num_heads=num_heads[i_layer],
541
- window_size=window_size,
542
- mlp_ratio=mlp_ratio,
543
- qkv_bias=qkv_bias,
544
- qk_scale=qk_scale,
545
- drop=drop_rate,
546
- attn_drop=attn_drop_rate,
547
- drop_path=dpr[sum(depths[:i_layer]):sum(depths[:i_layer + 1])],
548
- norm_layer=norm_layer,
549
- downsample=PatchMerging if (i_layer < self.num_layers - 1) else None,
550
- use_checkpoint=use_checkpoint)
551
- self.layers.append(layer)
552
-
553
- num_features = [int(embed_dim * 2 ** i) for i in range(self.num_layers)]
554
- self.num_features = num_features
555
-
556
- # add a norm layer for each output
557
- for i_layer in out_indices:
558
- layer = norm_layer(num_features[i_layer])
559
- layer_name = f'norm{i_layer}'
560
- self.add_module(layer_name, layer)
561
-
562
- self._freeze_stages()
563
-
564
- def _freeze_stages(self):
565
- if self.frozen_stages >= 0:
566
- self.patch_embed.eval()
567
- for param in self.patch_embed.parameters():
568
- param.requires_grad = False
569
-
570
- if self.frozen_stages >= 1 and self.ape:
571
- self.absolute_pos_embed.requires_grad = False
572
-
573
- if self.frozen_stages >= 2:
574
- self.pos_drop.eval()
575
- for i in range(0, self.frozen_stages - 1):
576
- m = self.layers[i]
577
- m.eval()
578
- for param in m.parameters():
579
- param.requires_grad = False
580
-
581
-
582
- def forward(self, x):
583
- """Forward function."""
584
- x = self.patch_embed(x)
585
-
586
- Wh, Ww = x.size(2), x.size(3)
587
- if self.ape:
588
- # interpolate the position embedding to the corresponding size
589
- absolute_pos_embed = F.interpolate(self.absolute_pos_embed, size=(Wh, Ww), mode='bicubic')
590
- x = (x + absolute_pos_embed) # B Wh*Ww C
591
-
592
- outs = []#x.contiguous()]
593
- x = x.flatten(2).transpose(1, 2)
594
- x = self.pos_drop(x)
595
- for i in range(self.num_layers):
596
- layer = self.layers[i]
597
- x_out, H, W, x, Wh, Ww = layer(x, Wh, Ww)
598
-
599
- if i in self.out_indices:
600
- norm_layer = getattr(self, f'norm{i}')
601
- x_out = norm_layer(x_out)
602
-
603
- out = x_out.view(-1, H, W, self.num_features[i]).permute(0, 3, 1, 2).contiguous()
604
- outs.append(out)
605
-
606
- return tuple(outs)
607
-
608
- def train(self, mode=True):
609
- """Convert the model into training mode while keep layers freezed."""
610
- super(SwinTransformer, self).train(mode)
611
- self._freeze_stages()
612
-
613
- def swin_v1_t():
614
- model = SwinTransformer(embed_dim=96, depths=[2, 2, 6, 2], num_heads=[3, 6, 12, 24], window_size=7)
615
- return model
616
-
617
- def swin_v1_s():
618
- model = SwinTransformer(embed_dim=96, depths=[2, 2, 18, 2], num_heads=[3, 6, 12, 24], window_size=7)
619
- return model
620
-
621
- def swin_v1_b():
622
- model = SwinTransformer(embed_dim=128, depths=[2, 2, 18, 2], num_heads=[4, 8, 16, 32], window_size=12)
623
- return model
624
-
625
- def swin_v1_l():
626
- model = SwinTransformer(embed_dim=192, depths=[2, 2, 18, 2], num_heads=[6, 12, 24, 48], window_size=12)
627
- return model
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
models/birefnet.py DELETED
@@ -1,317 +0,0 @@
1
- import torch
2
- import torch.nn as nn
3
- from collections import OrderedDict
4
- import torch
5
- import torch.nn as nn
6
- import torch.nn.functional as F
7
- from torchvision.models import vgg16, vgg16_bn
8
- from torchvision.models import resnet50
9
- from kornia.filters import laplacian
10
-
11
- from config import Config
12
- from dataset import class_labels_TR_sorted
13
- from models.backbones.build_backbone import build_backbone
14
- from models.modules.decoder_blocks import BasicDecBlk, ResBlk, HierarAttDecBlk
15
- from models.modules.lateral_blocks import BasicLatBlk
16
- from models.modules.aspp import ASPP, ASPPDeformable
17
- from models.modules.ing import *
18
- from models.refinement.refiner import Refiner, RefinerPVTInChannels4, RefUNet
19
- from models.refinement.stem_layer import StemLayer
20
-
21
-
22
- class BiRefNet(nn.Module):
23
- def __init__(self, bb_pretrained=True):
24
- super(BiRefNet, self).__init__()
25
- self.config = Config()
26
- self.epoch = 1
27
- self.bb = build_backbone(self.config.bb, pretrained=bb_pretrained)
28
-
29
- channels = self.config.lateral_channels_in_collection
30
-
31
- if self.config.auxiliary_classification:
32
- self.avgpool = nn.AdaptiveAvgPool2d((1, 1))
33
- self.cls_head = nn.Sequential(
34
- nn.Linear(channels[0], len(class_labels_TR_sorted))
35
- )
36
-
37
- if self.config.squeeze_block:
38
- self.squeeze_module = nn.Sequential(*[
39
- eval(self.config.squeeze_block.split('_x')[0])(channels[0]+sum(self.config.cxt), channels[0])
40
- for _ in range(eval(self.config.squeeze_block.split('_x')[1]))
41
- ])
42
-
43
- self.decoder = Decoder(channels)
44
-
45
- if self.config.ender:
46
- self.dec_end = nn.Sequential(
47
- nn.Conv2d(1, 16, 3, 1, 1),
48
- nn.Conv2d(16, 1, 3, 1, 1),
49
- nn.ReLU(inplace=True),
50
- )
51
-
52
- # refine patch-level segmentation
53
- if self.config.refine:
54
- if self.config.refine == 'itself':
55
- self.stem_layer = StemLayer(in_channels=3+1, inter_channels=48, out_channels=3, norm_layer='BN' if self.config.batch_size > 1 else 'LN')
56
- else:
57
- self.refiner = eval('{}({})'.format(self.config.refine, 'in_channels=3+1'))
58
-
59
- if self.config.freeze_bb:
60
- # Freeze the backbone...
61
- print(self.named_parameters())
62
- for key, value in self.named_parameters():
63
- if 'bb.' in key and 'refiner.' not in key:
64
- value.requires_grad = False
65
-
66
- def forward_enc(self, x):
67
- if self.config.bb in ['vgg16', 'vgg16bn', 'resnet50']:
68
- x1 = self.bb.conv1(x); x2 = self.bb.conv2(x1); x3 = self.bb.conv3(x2); x4 = self.bb.conv4(x3)
69
- else:
70
- x1, x2, x3, x4 = self.bb(x)
71
- if self.config.mul_scl_ipt == 'cat':
72
- B, C, H, W = x.shape
73
- x1_, x2_, x3_, x4_ = self.bb(F.interpolate(x, size=(H//2, W//2), mode='bilinear', align_corners=True))
74
- x1 = torch.cat([x1, F.interpolate(x1_, size=x1.shape[2:], mode='bilinear', align_corners=True)], dim=1)
75
- x2 = torch.cat([x2, F.interpolate(x2_, size=x2.shape[2:], mode='bilinear', align_corners=True)], dim=1)
76
- x3 = torch.cat([x3, F.interpolate(x3_, size=x3.shape[2:], mode='bilinear', align_corners=True)], dim=1)
77
- x4 = torch.cat([x4, F.interpolate(x4_, size=x4.shape[2:], mode='bilinear', align_corners=True)], dim=1)
78
- elif self.config.mul_scl_ipt == 'add':
79
- B, C, H, W = x.shape
80
- x1_, x2_, x3_, x4_ = self.bb(F.interpolate(x, size=(H//2, W//2), mode='bilinear', align_corners=True))
81
- x1 = x1 + F.interpolate(x1_, size=x1.shape[2:], mode='bilinear', align_corners=True)
82
- x2 = x2 + F.interpolate(x2_, size=x2.shape[2:], mode='bilinear', align_corners=True)
83
- x3 = x3 + F.interpolate(x3_, size=x3.shape[2:], mode='bilinear', align_corners=True)
84
- x4 = x4 + F.interpolate(x4_, size=x4.shape[2:], mode='bilinear', align_corners=True)
85
- class_preds = self.cls_head(self.avgpool(x4).view(x4.shape[0], -1)) if self.training and self.config.auxiliary_classification else None
86
- if self.config.cxt:
87
- x4 = torch.cat(
88
- (
89
- *[
90
- F.interpolate(x1, size=x4.shape[2:], mode='bilinear', align_corners=True),
91
- F.interpolate(x2, size=x4.shape[2:], mode='bilinear', align_corners=True),
92
- F.interpolate(x3, size=x4.shape[2:], mode='bilinear', align_corners=True),
93
- ][-len(self.config.cxt):],
94
- x4
95
- ),
96
- dim=1
97
- )
98
- return (x1, x2, x3, x4), class_preds
99
-
100
- def forward_ori(self, x):
101
- ########## Encoder ##########
102
- (x1, x2, x3, x4), class_preds = self.forward_enc(x)
103
- if self.config.squeeze_block:
104
- x4 = self.squeeze_module(x4)
105
- ########## Decoder ##########
106
- features = [x, x1, x2, x3, x4]
107
- if self.training and self.config.out_ref:
108
- features.append(laplacian(torch.mean(x, dim=1).unsqueeze(1), kernel_size=5))
109
- scaled_preds = self.decoder(features)
110
- return scaled_preds, class_preds
111
-
112
- # def forward_ref(self, x, pred):
113
- # # refine patch-level segmentation
114
- # if pred.shape[2:] != x.shape[2:]:
115
- # pred = F.interpolate(pred, size=x.shape[2:], mode='bilinear', align_corners=True)
116
- # # pred = pred.sigmoid()
117
- # if self.config.refine == 'itself':
118
- # x = self.stem_layer(torch.cat([x, pred], dim=1))
119
- # scaled_preds, class_preds = self.forward_ori(x)
120
- # else:
121
- # scaled_preds = self.refiner([x, pred])
122
- # class_preds = None
123
- # return scaled_preds, class_preds
124
-
125
- # def forward_ref_end(self, x):
126
- # # remove the grids of concatenated preds
127
- # return self.dec_end(x) if self.config.ender else x
128
-
129
-
130
- # def forward(self, x):
131
- # if self.config.refine:
132
- # scaled_preds, class_preds_ori = self.forward_ori(F.interpolate(x, size=(x.shape[2]//4, x.shape[3]//4), mode='bilinear', align_corners=True))
133
- # class_preds_lst = [class_preds_ori]
134
- # for _ in range(self.config.refine_iteration):
135
- # scaled_preds_ref, class_preds_ref = self.forward_ref(x, scaled_preds[-1])
136
- # scaled_preds += scaled_preds_ref
137
- # class_preds_lst.append(class_preds_ref)
138
- # else:
139
- # scaled_preds, class_preds = self.forward_ori(x)
140
- # class_preds_lst = [class_preds]
141
- # return [scaled_preds, class_preds_lst] if self.training else scaled_preds
142
-
143
- def forward(self, x):
144
- scaled_preds, class_preds = self.forward_ori(x)
145
- class_preds_lst = [class_preds]
146
- return [scaled_preds, class_preds_lst] if self.training else scaled_preds
147
-
148
-
149
- class Decoder(nn.Module):
150
- def __init__(self, channels):
151
- super(Decoder, self).__init__()
152
- self.config = Config()
153
- DecoderBlock = eval(self.config.dec_blk)
154
- LateralBlock = eval(self.config.lat_blk)
155
-
156
- if self.config.dec_ipt:
157
- self.split = self.config.dec_ipt_split
158
- N_dec_ipt = 64
159
- DBlock = SimpleConvs
160
- ic = 64
161
- ipt_cha_opt = 1
162
- self.ipt_blk5 = DBlock(2**10*3 if self.split else 3, [N_dec_ipt, channels[0]//8][ipt_cha_opt], inter_channels=ic)
163
- self.ipt_blk4 = DBlock(2**8*3 if self.split else 3, [N_dec_ipt, channels[0]//8][ipt_cha_opt], inter_channels=ic)
164
- self.ipt_blk3 = DBlock(2**6*3 if self.split else 3, [N_dec_ipt, channels[1]//8][ipt_cha_opt], inter_channels=ic)
165
- self.ipt_blk2 = DBlock(2**4*3 if self.split else 3, [N_dec_ipt, channels[2]//8][ipt_cha_opt], inter_channels=ic)
166
- self.ipt_blk1 = DBlock(2**0*3 if self.split else 3, [N_dec_ipt, channels[3]//8][ipt_cha_opt], inter_channels=ic)
167
- else:
168
- self.split = None
169
-
170
- self.decoder_block4 = DecoderBlock(channels[0]+([N_dec_ipt, channels[0]//8][ipt_cha_opt] if self.config.dec_ipt else 0), channels[1])
171
- self.decoder_block3 = DecoderBlock(channels[1]+([N_dec_ipt, channels[0]//8][ipt_cha_opt] if self.config.dec_ipt else 0), channels[2])
172
- self.decoder_block2 = DecoderBlock(channels[2]+([N_dec_ipt, channels[1]//8][ipt_cha_opt] if self.config.dec_ipt else 0), channels[3])
173
- self.decoder_block1 = DecoderBlock(channels[3]+([N_dec_ipt, channels[2]//8][ipt_cha_opt] if self.config.dec_ipt else 0), channels[3]//2)
174
- self.conv_out1 = nn.Sequential(nn.Conv2d(channels[3]//2+([N_dec_ipt, channels[3]//8][ipt_cha_opt] if self.config.dec_ipt else 0), 1, 1, 1, 0))
175
-
176
- self.lateral_block4 = LateralBlock(channels[1], channels[1])
177
- self.lateral_block3 = LateralBlock(channels[2], channels[2])
178
- self.lateral_block2 = LateralBlock(channels[3], channels[3])
179
-
180
- if self.config.ms_supervision:
181
- self.conv_ms_spvn_4 = nn.Conv2d(channels[1], 1, 1, 1, 0)
182
- self.conv_ms_spvn_3 = nn.Conv2d(channels[2], 1, 1, 1, 0)
183
- self.conv_ms_spvn_2 = nn.Conv2d(channels[3], 1, 1, 1, 0)
184
-
185
- if self.config.out_ref:
186
- _N = 16
187
- self.gdt_convs_4 = nn.Sequential(nn.Conv2d(channels[1], _N, 3, 1, 1), nn.BatchNorm2d(_N) if self.config.batch_size > 1 else nn.Identity(), nn.ReLU(inplace=True))
188
- self.gdt_convs_3 = nn.Sequential(nn.Conv2d(channels[2], _N, 3, 1, 1), nn.BatchNorm2d(_N) if self.config.batch_size > 1 else nn.Identity(), nn.ReLU(inplace=True))
189
- self.gdt_convs_2 = nn.Sequential(nn.Conv2d(channels[3], _N, 3, 1, 1), nn.BatchNorm2d(_N) if self.config.batch_size > 1 else nn.Identity(), nn.ReLU(inplace=True))
190
-
191
- self.gdt_convs_pred_4 = nn.Sequential(nn.Conv2d(_N, 1, 1, 1, 0))
192
- self.gdt_convs_pred_3 = nn.Sequential(nn.Conv2d(_N, 1, 1, 1, 0))
193
- self.gdt_convs_pred_2 = nn.Sequential(nn.Conv2d(_N, 1, 1, 1, 0))
194
-
195
- self.gdt_convs_attn_4 = nn.Sequential(nn.Conv2d(_N, 1, 1, 1, 0))
196
- self.gdt_convs_attn_3 = nn.Sequential(nn.Conv2d(_N, 1, 1, 1, 0))
197
- self.gdt_convs_attn_2 = nn.Sequential(nn.Conv2d(_N, 1, 1, 1, 0))
198
-
199
-
200
- def get_patches_batch(self, x, p):
201
- _size_h, _size_w = p.shape[2:]
202
- patches_batch = []
203
- for idx in range(x.shape[0]):
204
- columns_x = torch.split(x[idx], split_size_or_sections=_size_w, dim=-1)
205
- patches_x = []
206
- for column_x in columns_x:
207
- patches_x += [p.unsqueeze(0) for p in torch.split(column_x, split_size_or_sections=_size_h, dim=-2)]
208
- patch_sample = torch.cat(patches_x, dim=1)
209
- patches_batch.append(patch_sample)
210
- return torch.cat(patches_batch, dim=0)
211
-
212
- def forward(self, features):
213
- if self.training and self.config.out_ref:
214
- outs_gdt_pred = []
215
- outs_gdt_label = []
216
- x, x1, x2, x3, x4, gdt_gt = features
217
- else:
218
- x, x1, x2, x3, x4 = features
219
- outs = []
220
-
221
- if self.config.dec_ipt:
222
- patches_batch = self.get_patches_batch(x, x4) if self.split else x
223
- x4 = torch.cat((x4, self.ipt_blk5(F.interpolate(patches_batch, size=x4.shape[2:], mode='bilinear', align_corners=True))), 1)
224
- p4 = self.decoder_block4(x4)
225
- m4 = self.conv_ms_spvn_4(p4) if self.config.ms_supervision else None
226
- if self.config.out_ref:
227
- p4_gdt = self.gdt_convs_4(p4)
228
- if self.training:
229
- # >> GT:
230
- m4_dia = m4
231
- gdt_label_main_4 = gdt_gt * F.interpolate(m4_dia, size=gdt_gt.shape[2:], mode='bilinear', align_corners=True)
232
- outs_gdt_label.append(gdt_label_main_4)
233
- # >> Pred:
234
- gdt_pred_4 = self.gdt_convs_pred_4(p4_gdt)
235
- outs_gdt_pred.append(gdt_pred_4)
236
- gdt_attn_4 = self.gdt_convs_attn_4(p4_gdt).sigmoid()
237
- # >> Finally:
238
- p4 = p4 * gdt_attn_4
239
- _p4 = F.interpolate(p4, size=x3.shape[2:], mode='bilinear', align_corners=True)
240
- _p3 = _p4 + self.lateral_block4(x3)
241
-
242
- if self.config.dec_ipt:
243
- patches_batch = self.get_patches_batch(x, _p3) if self.split else x
244
- _p3 = torch.cat((_p3, self.ipt_blk4(F.interpolate(patches_batch, size=x3.shape[2:], mode='bilinear', align_corners=True))), 1)
245
- p3 = self.decoder_block3(_p3)
246
- m3 = self.conv_ms_spvn_3(p3) if self.config.ms_supervision else None
247
- if self.config.out_ref:
248
- p3_gdt = self.gdt_convs_3(p3)
249
- if self.training:
250
- # >> GT:
251
- # m3 --dilation--> m3_dia
252
- # G_3^gt * m3_dia --> G_3^m, which is the label of gradient
253
- m3_dia = m3
254
- gdt_label_main_3 = gdt_gt * F.interpolate(m3_dia, size=gdt_gt.shape[2:], mode='bilinear', align_corners=True)
255
- outs_gdt_label.append(gdt_label_main_3)
256
- # >> Pred:
257
- # p3 --conv--BN--> F_3^G, where F_3^G predicts the \hat{G_3} with xx
258
- # F_3^G --sigmoid--> A_3^G
259
- gdt_pred_3 = self.gdt_convs_pred_3(p3_gdt)
260
- outs_gdt_pred.append(gdt_pred_3)
261
- gdt_attn_3 = self.gdt_convs_attn_3(p3_gdt).sigmoid()
262
- # >> Finally:
263
- # p3 = p3 * A_3^G
264
- p3 = p3 * gdt_attn_3
265
- _p3 = F.interpolate(p3, size=x2.shape[2:], mode='bilinear', align_corners=True)
266
- _p2 = _p3 + self.lateral_block3(x2)
267
-
268
- if self.config.dec_ipt:
269
- patches_batch = self.get_patches_batch(x, _p2) if self.split else x
270
- _p2 = torch.cat((_p2, self.ipt_blk3(F.interpolate(patches_batch, size=x2.shape[2:], mode='bilinear', align_corners=True))), 1)
271
- p2 = self.decoder_block2(_p2)
272
- m2 = self.conv_ms_spvn_2(p2) if self.config.ms_supervision else None
273
- if self.config.out_ref:
274
- p2_gdt = self.gdt_convs_2(p2)
275
- if self.training:
276
- # >> GT:
277
- m2_dia = m2
278
- gdt_label_main_2 = gdt_gt * F.interpolate(m2_dia, size=gdt_gt.shape[2:], mode='bilinear', align_corners=True)
279
- outs_gdt_label.append(gdt_label_main_2)
280
- # >> Pred:
281
- gdt_pred_2 = self.gdt_convs_pred_2(p2_gdt)
282
- outs_gdt_pred.append(gdt_pred_2)
283
- gdt_attn_2 = self.gdt_convs_attn_2(p2_gdt).sigmoid()
284
- # >> Finally:
285
- p2 = p2 * gdt_attn_2
286
- _p2 = F.interpolate(p2, size=x1.shape[2:], mode='bilinear', align_corners=True)
287
- _p1 = _p2 + self.lateral_block2(x1)
288
-
289
- if self.config.dec_ipt:
290
- patches_batch = self.get_patches_batch(x, _p1) if self.split else x
291
- _p1 = torch.cat((_p1, self.ipt_blk2(F.interpolate(patches_batch, size=x1.shape[2:], mode='bilinear', align_corners=True))), 1)
292
- _p1 = self.decoder_block1(_p1)
293
- _p1 = F.interpolate(_p1, size=x.shape[2:], mode='bilinear', align_corners=True)
294
-
295
- if self.config.dec_ipt:
296
- patches_batch = self.get_patches_batch(x, _p1) if self.split else x
297
- _p1 = torch.cat((_p1, self.ipt_blk1(F.interpolate(patches_batch, size=x.shape[2:], mode='bilinear', align_corners=True))), 1)
298
- p1_out = self.conv_out1(_p1)
299
-
300
- if self.config.ms_supervision:
301
- outs.append(m4)
302
- outs.append(m3)
303
- outs.append(m2)
304
- outs.append(p1_out)
305
- return outs if not (self.config.out_ref and self.training) else ([outs_gdt_pred, outs_gdt_label], outs)
306
-
307
-
308
- class SimpleConvs(nn.Module):
309
- def __init__(
310
- self, in_channels: int, out_channels: int, inter_channels=64
311
- ) -> None:
312
- super().__init__()
313
- self.conv1 = nn.Conv2d(in_channels, inter_channels, 3, 1, 1)
314
- self.conv_out = nn.Conv2d(inter_channels, out_channels, 3, 1, 1)
315
-
316
- def forward(self, x):
317
- return self.conv_out(self.conv1(x))
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
models/modules/aspp.py DELETED
@@ -1,119 +0,0 @@
1
- import torch
2
- import torch.nn as nn
3
- import torch.nn.functional as F
4
- from models.modules.deform_conv import DeformableConv2d
5
- from config import Config
6
-
7
-
8
- config = Config()
9
-
10
-
11
- class _ASPPModule(nn.Module):
12
- def __init__(self, in_channels, planes, kernel_size, padding, dilation):
13
- super(_ASPPModule, self).__init__()
14
- self.atrous_conv = nn.Conv2d(in_channels, planes, kernel_size=kernel_size,
15
- stride=1, padding=padding, dilation=dilation, bias=False)
16
- self.bn = nn.BatchNorm2d(planes) if config.batch_size > 1 else nn.Identity()
17
- self.relu = nn.ReLU(inplace=True)
18
-
19
- def forward(self, x):
20
- x = self.atrous_conv(x)
21
- x = self.bn(x)
22
-
23
- return self.relu(x)
24
-
25
-
26
- class ASPP(nn.Module):
27
- def __init__(self, in_channels=64, out_channels=None, output_stride=16):
28
- super(ASPP, self).__init__()
29
- self.down_scale = 1
30
- if out_channels is None:
31
- out_channels = in_channels
32
- self.in_channelster = 256 // self.down_scale
33
- if output_stride == 16:
34
- dilations = [1, 6, 12, 18]
35
- elif output_stride == 8:
36
- dilations = [1, 12, 24, 36]
37
- else:
38
- raise NotImplementedError
39
-
40
- self.aspp1 = _ASPPModule(in_channels, self.in_channelster, 1, padding=0, dilation=dilations[0])
41
- self.aspp2 = _ASPPModule(in_channels, self.in_channelster, 3, padding=dilations[1], dilation=dilations[1])
42
- self.aspp3 = _ASPPModule(in_channels, self.in_channelster, 3, padding=dilations[2], dilation=dilations[2])
43
- self.aspp4 = _ASPPModule(in_channels, self.in_channelster, 3, padding=dilations[3], dilation=dilations[3])
44
-
45
- self.global_avg_pool = nn.Sequential(nn.AdaptiveAvgPool2d((1, 1)),
46
- nn.Conv2d(in_channels, self.in_channelster, 1, stride=1, bias=False),
47
- nn.BatchNorm2d(self.in_channelster) if config.batch_size > 1 else nn.Identity(),
48
- nn.ReLU(inplace=True))
49
- self.conv1 = nn.Conv2d(self.in_channelster * 5, out_channels, 1, bias=False)
50
- self.bn1 = nn.BatchNorm2d(out_channels) if config.batch_size > 1 else nn.Identity()
51
- self.relu = nn.ReLU(inplace=True)
52
- self.dropout = nn.Dropout(0.5)
53
-
54
- def forward(self, x):
55
- x1 = self.aspp1(x)
56
- x2 = self.aspp2(x)
57
- x3 = self.aspp3(x)
58
- x4 = self.aspp4(x)
59
- x5 = self.global_avg_pool(x)
60
- x5 = F.interpolate(x5, size=x1.size()[2:], mode='bilinear', align_corners=True)
61
- x = torch.cat((x1, x2, x3, x4, x5), dim=1)
62
-
63
- x = self.conv1(x)
64
- x = self.bn1(x)
65
- x = self.relu(x)
66
-
67
- return self.dropout(x)
68
-
69
-
70
- ##################### Deformable
71
- class _ASPPModuleDeformable(nn.Module):
72
- def __init__(self, in_channels, planes, kernel_size, padding):
73
- super(_ASPPModuleDeformable, self).__init__()
74
- self.atrous_conv = DeformableConv2d(in_channels, planes, kernel_size=kernel_size,
75
- stride=1, padding=padding, bias=False)
76
- self.bn = nn.BatchNorm2d(planes) if config.batch_size > 1 else nn.Identity()
77
- self.relu = nn.ReLU(inplace=True)
78
-
79
- def forward(self, x):
80
- x = self.atrous_conv(x)
81
- x = self.bn(x)
82
-
83
- return self.relu(x)
84
-
85
-
86
- class ASPPDeformable(nn.Module):
87
- def __init__(self, in_channels, out_channels=None, parallel_block_sizes=[1, 3, 7]):
88
- super(ASPPDeformable, self).__init__()
89
- self.down_scale = 1
90
- if out_channels is None:
91
- out_channels = in_channels
92
- self.in_channelster = 256 // self.down_scale
93
-
94
- self.aspp1 = _ASPPModuleDeformable(in_channels, self.in_channelster, 1, padding=0)
95
- self.aspp_deforms = nn.ModuleList([
96
- _ASPPModuleDeformable(in_channels, self.in_channelster, conv_size, padding=int(conv_size//2)) for conv_size in parallel_block_sizes
97
- ])
98
-
99
- self.global_avg_pool = nn.Sequential(nn.AdaptiveAvgPool2d((1, 1)),
100
- nn.Conv2d(in_channels, self.in_channelster, 1, stride=1, bias=False),
101
- nn.BatchNorm2d(self.in_channelster) if config.batch_size > 1 else nn.Identity(),
102
- nn.ReLU(inplace=True))
103
- self.conv1 = nn.Conv2d(self.in_channelster * (2 + len(self.aspp_deforms)), out_channels, 1, bias=False)
104
- self.bn1 = nn.BatchNorm2d(out_channels) if config.batch_size > 1 else nn.Identity()
105
- self.relu = nn.ReLU(inplace=True)
106
- self.dropout = nn.Dropout(0.5)
107
-
108
- def forward(self, x):
109
- x1 = self.aspp1(x)
110
- x_aspp_deforms = [aspp_deform(x) for aspp_deform in self.aspp_deforms]
111
- x5 = self.global_avg_pool(x)
112
- x5 = F.interpolate(x5, size=x1.size()[2:], mode='bilinear', align_corners=True)
113
- x = torch.cat((x1, *x_aspp_deforms, x5), dim=1)
114
-
115
- x = self.conv1(x)
116
- x = self.bn1(x)
117
- x = self.relu(x)
118
-
119
- return self.dropout(x)
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
models/modules/attentions.py DELETED
@@ -1,93 +0,0 @@
1
- import numpy as np
2
- import torch
3
- from torch import nn
4
- from torch.nn import init
5
-
6
-
7
- class SEWeightModule(nn.Module):
8
- def __init__(self, channels, reduction=16):
9
- super(SEWeightModule, self).__init__()
10
- self.avg_pool = nn.AdaptiveAvgPool2d(1)
11
- self.fc1 = nn.Conv2d(channels, channels//reduction, kernel_size=1, padding=0)
12
- self.relu = nn.ReLU(inplace=True)
13
- self.fc2 = nn.Conv2d(channels//reduction, channels, kernel_size=1, padding=0)
14
- self.sigmoid = nn.Sigmoid()
15
-
16
- def forward(self, x):
17
- out = self.avg_pool(x)
18
- out = self.fc1(out)
19
- out = self.relu(out)
20
- out = self.fc2(out)
21
- weight = self.sigmoid(out)
22
- return weight
23
-
24
-
25
- class PSA(nn.Module):
26
-
27
- def __init__(self, in_channels, S=4, reduction=4):
28
- super().__init__()
29
- self.S = S
30
-
31
- _convs = []
32
- for i in range(S):
33
- _convs.append(nn.Conv2d(in_channels//S, in_channels//S, kernel_size=2*(i+1)+1, padding=i+1))
34
- self.convs = nn.ModuleList(_convs)
35
-
36
- self.se_block = SEWeightModule(in_channels//S, reduction=S*reduction)
37
-
38
- self.softmax = nn.Softmax(dim=1)
39
-
40
- def forward(self, x):
41
- b, c, h, w = x.size()
42
-
43
- # Step1: SPC module
44
- SPC_out = x.view(b, self.S, c//self.S, h, w) #bs,s,ci,h,w
45
- for idx, conv in enumerate(self.convs):
46
- SPC_out[:,idx,:,:,:] = conv(SPC_out[:,idx,:,:,:].clone())
47
-
48
- # Step2: SE weight
49
- se_out=[]
50
- for idx in range(self.S):
51
- se_out.append(self.se_block(SPC_out[:, idx, :, :, :]))
52
- SE_out = torch.stack(se_out, dim=1)
53
- SE_out = SE_out.expand_as(SPC_out)
54
-
55
- # Step3: Softmax
56
- softmax_out = self.softmax(SE_out)
57
-
58
- # Step4: SPA
59
- PSA_out = SPC_out * softmax_out
60
- PSA_out = PSA_out.view(b, -1, h, w)
61
-
62
- return PSA_out
63
-
64
-
65
- class SGE(nn.Module):
66
-
67
- def __init__(self, groups):
68
- super().__init__()
69
- self.groups=groups
70
- self.avg_pool = nn.AdaptiveAvgPool2d(1)
71
- self.weight=nn.Parameter(torch.zeros(1,groups,1,1))
72
- self.bias=nn.Parameter(torch.zeros(1,groups,1,1))
73
- self.sig=nn.Sigmoid()
74
-
75
- def forward(self, x):
76
- b, c, h,w=x.shape
77
- x=x.view(b*self.groups,-1,h,w) #bs*g,dim//g,h,w
78
- xn=x*self.avg_pool(x) #bs*g,dim//g,h,w
79
- xn=xn.sum(dim=1,keepdim=True) #bs*g,1,h,w
80
- t=xn.view(b*self.groups,-1) #bs*g,h*w
81
-
82
- t=t-t.mean(dim=1,keepdim=True) #bs*g,h*w
83
- std=t.std(dim=1,keepdim=True)+1e-5
84
- t=t/std #bs*g,h*w
85
- t=t.view(b,self.groups,h,w) #bs,g,h*w
86
-
87
- t=t*self.weight+self.bias #bs,g,h*w
88
- t=t.view(b*self.groups,1,h,w) #bs*g,1,h*w
89
- x=x*self.sig(t)
90
- x=x.view(b,c,h,w)
91
-
92
- return x
93
-
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
models/modules/decoder_blocks.py DELETED
@@ -1,101 +0,0 @@
1
- import torch
2
- import torch.nn as nn
3
- from models.modules.aspp import ASPP, ASPPDeformable
4
- from models.modules.attentions import PSA, SGE
5
- from config import Config
6
-
7
-
8
- config = Config()
9
-
10
-
11
- class BasicDecBlk(nn.Module):
12
- def __init__(self, in_channels=64, out_channels=64, inter_channels=64):
13
- super(BasicDecBlk, self).__init__()
14
- inter_channels = in_channels // 4 if config.dec_channels_inter == 'adap' else 64
15
- self.conv_in = nn.Conv2d(in_channels, inter_channels, 3, 1, padding=1)
16
- self.relu_in = nn.ReLU(inplace=True)
17
- if config.dec_att == 'ASPP':
18
- self.dec_att = ASPP(in_channels=inter_channels)
19
- elif config.dec_att == 'ASPPDeformable':
20
- self.dec_att = ASPPDeformable(in_channels=inter_channels)
21
- self.conv_out = nn.Conv2d(inter_channels, out_channels, 3, 1, padding=1)
22
- self.bn_in = nn.BatchNorm2d(inter_channels) if config.batch_size > 1 else nn.Identity()
23
- self.bn_out = nn.BatchNorm2d(out_channels) if config.batch_size > 1 else nn.Identity()
24
-
25
- def forward(self, x):
26
- x = self.conv_in(x)
27
- x = self.bn_in(x)
28
- x = self.relu_in(x)
29
- if hasattr(self, 'dec_att'):
30
- x = self.dec_att(x)
31
- x = self.conv_out(x)
32
- x = self.bn_out(x)
33
- return x
34
-
35
-
36
- class ResBlk(nn.Module):
37
- def __init__(self, in_channels=64, out_channels=None, inter_channels=64):
38
- super(ResBlk, self).__init__()
39
- if out_channels is None:
40
- out_channels = in_channels
41
- inter_channels = in_channels // 4 if config.dec_channels_inter == 'adap' else 64
42
-
43
- self.conv_in = nn.Conv2d(in_channels, inter_channels, 3, 1, padding=1)
44
- self.bn_in = nn.BatchNorm2d(inter_channels) if config.batch_size > 1 else nn.Identity()
45
- self.relu_in = nn.ReLU(inplace=True)
46
-
47
- if config.dec_att == 'ASPP':
48
- self.dec_att = ASPP(in_channels=inter_channels)
49
- elif config.dec_att == 'ASPPDeformable':
50
- self.dec_att = ASPPDeformable(in_channels=inter_channels)
51
-
52
- self.conv_out = nn.Conv2d(inter_channels, out_channels, 3, 1, padding=1)
53
- self.bn_out = nn.BatchNorm2d(out_channels) if config.batch_size > 1 else nn.Identity()
54
-
55
- self.conv_resi = nn.Conv2d(in_channels, out_channels, 1, 1, 0)
56
-
57
- def forward(self, x):
58
- _x = self.conv_resi(x)
59
- x = self.conv_in(x)
60
- x = self.bn_in(x)
61
- x = self.relu_in(x)
62
- if hasattr(self, 'dec_att'):
63
- x = self.dec_att(x)
64
- x = self.conv_out(x)
65
- x = self.bn_out(x)
66
- return x + _x
67
-
68
-
69
- class HierarAttDecBlk(nn.Module):
70
- def __init__(self, in_channels=64, out_channels=None, inter_channels=64):
71
- super(HierarAttDecBlk, self).__init__()
72
- if out_channels is None:
73
- out_channels = in_channels
74
- inter_channels = in_channels // 4 if config.dec_channels_inter == 'adap' else 64
75
- self.split_y = 8 # must be divided by channels of all intermediate features
76
- self.split_x = 8
77
-
78
- self.conv_in = nn.Conv2d(in_channels, inter_channels, 3, 1, 1)
79
-
80
- self.psa = PSA(inter_channels*self.split_y*self.split_x, S=config.batch_size)
81
- self.sge = SGE(groups=config.batch_size)
82
-
83
- if config.dec_att == 'ASPP':
84
- self.dec_att = ASPP(in_channels=inter_channels)
85
- elif config.dec_att == 'ASPPDeformable':
86
- self.dec_att = ASPPDeformable(in_channels=inter_channels)
87
- self.conv_out = nn.Conv2d(inter_channels, out_channels, 3, 1, 1)
88
-
89
- def forward(self, x):
90
- x = self.conv_in(x)
91
- N, C, H, W = x.shape
92
- x_patchs = x.reshape(N, -1, H//self.split_y, W//self.split_x)
93
-
94
- # Hierarchical attention: group attention X patch spatial attention
95
- x_patchs = self.psa(x_patchs) # Group Channel Attention -- each group is a single image
96
- x_patchs = self.sge(x_patchs) # Patch Spatial Attention
97
- x = x.reshape(N, C, H, W)
98
- if hasattr(self, 'dec_att'):
99
- x = self.dec_att(x)
100
- x = self.conv_out(x)
101
- return x
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
models/modules/deform_conv.py DELETED
@@ -1,66 +0,0 @@
1
- import torch
2
- import torch.nn as nn
3
- from torchvision.ops import deform_conv2d
4
-
5
-
6
- class DeformableConv2d(nn.Module):
7
- def __init__(self,
8
- in_channels,
9
- out_channels,
10
- kernel_size=3,
11
- stride=1,
12
- padding=1,
13
- bias=False):
14
-
15
- super(DeformableConv2d, self).__init__()
16
-
17
- assert type(kernel_size) == tuple or type(kernel_size) == int
18
-
19
- kernel_size = kernel_size if type(kernel_size) == tuple else (kernel_size, kernel_size)
20
- self.stride = stride if type(stride) == tuple else (stride, stride)
21
- self.padding = padding
22
-
23
- self.offset_conv = nn.Conv2d(in_channels,
24
- 2 * kernel_size[0] * kernel_size[1],
25
- kernel_size=kernel_size,
26
- stride=stride,
27
- padding=self.padding,
28
- bias=True)
29
-
30
- nn.init.constant_(self.offset_conv.weight, 0.)
31
- nn.init.constant_(self.offset_conv.bias, 0.)
32
-
33
- self.modulator_conv = nn.Conv2d(in_channels,
34
- 1 * kernel_size[0] * kernel_size[1],
35
- kernel_size=kernel_size,
36
- stride=stride,
37
- padding=self.padding,
38
- bias=True)
39
-
40
- nn.init.constant_(self.modulator_conv.weight, 0.)
41
- nn.init.constant_(self.modulator_conv.bias, 0.)
42
-
43
- self.regular_conv = nn.Conv2d(in_channels,
44
- out_channels=out_channels,
45
- kernel_size=kernel_size,
46
- stride=stride,
47
- padding=self.padding,
48
- bias=bias)
49
-
50
- def forward(self, x):
51
- #h, w = x.shape[2:]
52
- #max_offset = max(h, w)/4.
53
-
54
- offset = self.offset_conv(x)#.clamp(-max_offset, max_offset)
55
- modulator = 2. * torch.sigmoid(self.modulator_conv(x))
56
-
57
- x = deform_conv2d(
58
- input=x,
59
- offset=offset,
60
- weight=self.regular_conv.weight,
61
- bias=self.regular_conv.bias,
62
- padding=self.padding,
63
- mask=modulator,
64
- stride=self.stride,
65
- )
66
- return x
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
models/modules/ing.py DELETED
@@ -1,29 +0,0 @@
1
- import torch.nn as nn
2
- from models.modules.mlp import MLPLayer
3
-
4
-
5
- class BlockA(nn.Module):
6
- def __init__(self, in_channels=64, out_channels=64, inter_channels=64, mlp_ratio=4.):
7
- super(BlockA, self).__init__()
8
- inter_channels = in_channels
9
- self.conv = nn.Conv2d(in_channels, inter_channels, 3, 1, 1)
10
- self.norm1 = nn.LayerNorm(inter_channels)
11
- self.ffn = MLPLayer(in_features=inter_channels,
12
- hidden_features=int(inter_channels * mlp_ratio),
13
- act_layer=nn.GELU,
14
- drop=0.)
15
- self.norm2 = nn.LayerNorm(inter_channels)
16
-
17
- def forward(self, x):
18
- B, C, H, W = x.shape
19
- _x = self.conv(x)
20
- _x = _x.flatten(2).transpose(1, 2)
21
- _x = self.norm1(_x)
22
- x = x.reshape(B, H, W, -1).permute(0, 3, 1, 2).contiguous()
23
-
24
- x = x + _x
25
- _x1 = self.ffn(x)
26
- _x1 = self.norm2(_x1)
27
- _x1 = _x1.reshape(B, H, W, -1).permute(0, 3, 1, 2).contiguous()
28
- x = x + _x1
29
- return x
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
models/modules/lateral_blocks.py DELETED
@@ -1,21 +0,0 @@
1
- import numpy as np
2
- import torch
3
- import torch.nn as nn
4
- import torch.nn.functional as F
5
- from functools import partial
6
-
7
- from config import Config
8
-
9
-
10
- config = Config()
11
-
12
-
13
- class BasicLatBlk(nn.Module):
14
- def __init__(self, in_channels=64, out_channels=64, inter_channels=64):
15
- super(BasicLatBlk, self).__init__()
16
- inter_channels = in_channels // 4 if config.dec_channels_inter == 'adap' else 64
17
- self.conv = nn.Conv2d(in_channels, out_channels, 1, 1, 0)
18
-
19
- def forward(self, x):
20
- x = self.conv(x)
21
- return x
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
models/modules/mlp.py DELETED
@@ -1,118 +0,0 @@
1
- import torch
2
- import torch.nn as nn
3
- from functools import partial
4
-
5
- from timm.models.layers import DropPath, to_2tuple, trunc_normal_
6
- from timm.models.registry import register_model
7
-
8
- import math
9
-
10
-
11
- class MLPLayer(nn.Module):
12
- def __init__(self, in_features, hidden_features=None, out_features=None, act_layer=nn.GELU, drop=0.):
13
- super().__init__()
14
- out_features = out_features or in_features
15
- hidden_features = hidden_features or in_features
16
- self.fc1 = nn.Linear(in_features, hidden_features)
17
- self.act = act_layer()
18
- self.fc2 = nn.Linear(hidden_features, out_features)
19
- self.drop = nn.Dropout(drop)
20
-
21
- def forward(self, x):
22
- x = self.fc1(x)
23
- x = self.act(x)
24
- x = self.drop(x)
25
- x = self.fc2(x)
26
- x = self.drop(x)
27
- return x
28
-
29
-
30
- class Attention(nn.Module):
31
- def __init__(self, dim, num_heads=8, qkv_bias=False, qk_scale=None, attn_drop=0., proj_drop=0., sr_ratio=1):
32
- super().__init__()
33
- assert dim % num_heads == 0, f"dim {dim} should be divided by num_heads {num_heads}."
34
-
35
- self.dim = dim
36
- self.num_heads = num_heads
37
- head_dim = dim // num_heads
38
- self.scale = qk_scale or head_dim ** -0.5
39
-
40
- self.q = nn.Linear(dim, dim, bias=qkv_bias)
41
- self.kv = nn.Linear(dim, dim * 2, bias=qkv_bias)
42
- self.attn_drop = nn.Dropout(attn_drop)
43
- self.proj = nn.Linear(dim, dim)
44
- self.proj_drop = nn.Dropout(proj_drop)
45
-
46
- self.sr_ratio = sr_ratio
47
- if sr_ratio > 1:
48
- self.sr = nn.Conv2d(dim, dim, kernel_size=sr_ratio, stride=sr_ratio)
49
- self.norm = nn.LayerNorm(dim)
50
-
51
- def forward(self, x, H, W):
52
- B, N, C = x.shape
53
- q = self.q(x).reshape(B, N, self.num_heads, C // self.num_heads).permute(0, 2, 1, 3)
54
-
55
- if self.sr_ratio > 1:
56
- x_ = x.permute(0, 2, 1).reshape(B, C, H, W)
57
- x_ = self.sr(x_).reshape(B, C, -1).permute(0, 2, 1)
58
- x_ = self.norm(x_)
59
- kv = self.kv(x_).reshape(B, -1, 2, self.num_heads, C // self.num_heads).permute(2, 0, 3, 1, 4)
60
- else:
61
- kv = self.kv(x).reshape(B, -1, 2, self.num_heads, C // self.num_heads).permute(2, 0, 3, 1, 4)
62
- k, v = kv[0], kv[1]
63
-
64
- attn = (q @ k.transpose(-2, -1)) * self.scale
65
- attn = attn.softmax(dim=-1)
66
- attn = self.attn_drop(attn)
67
-
68
- x = (attn @ v).transpose(1, 2).reshape(B, N, C)
69
- x = self.proj(x)
70
- x = self.proj_drop(x)
71
- return x
72
-
73
-
74
- class Block(nn.Module):
75
- def __init__(self, dim, num_heads, mlp_ratio=4., qkv_bias=False, qk_scale=None, drop=0., attn_drop=0.,
76
- drop_path=0., act_layer=nn.GELU, norm_layer=nn.LayerNorm, sr_ratio=1):
77
- super().__init__()
78
- self.norm1 = norm_layer(dim)
79
- self.attn = Attention(
80
- dim,
81
- num_heads=num_heads, qkv_bias=qkv_bias, qk_scale=qk_scale,
82
- attn_drop=attn_drop, proj_drop=drop, sr_ratio=sr_ratio)
83
- # NOTE: drop path for stochastic depth, we shall see if this is better than dropout here
84
- self.drop_path = DropPath(drop_path) if drop_path > 0. else nn.Identity()
85
- self.norm2 = norm_layer(dim)
86
- mlp_hidden_dim = int(dim * mlp_ratio)
87
- self.mlp = MLPLayer(in_features=dim, hidden_features=mlp_hidden_dim, act_layer=act_layer, drop=drop)
88
-
89
- def forward(self, x, H, W):
90
- x = x + self.drop_path(self.attn(self.norm1(x), H, W))
91
- x = x + self.drop_path(self.mlp(self.norm2(x), H, W))
92
- return x
93
-
94
-
95
- class OverlapPatchEmbed(nn.Module):
96
- """ Image to Patch Embedding
97
- """
98
-
99
- def __init__(self, img_size=224, patch_size=7, stride=4, in_channels=3, embed_dim=768):
100
- super().__init__()
101
- img_size = to_2tuple(img_size)
102
- patch_size = to_2tuple(patch_size)
103
-
104
- self.img_size = img_size
105
- self.patch_size = patch_size
106
- self.H, self.W = img_size[0] // patch_size[0], img_size[1] // patch_size[1]
107
- self.num_patches = self.H * self.W
108
- self.proj = nn.Conv2d(in_channels, embed_dim, kernel_size=patch_size, stride=stride,
109
- padding=(patch_size[0] // 2, patch_size[1] // 2))
110
- self.norm = nn.LayerNorm(embed_dim)
111
-
112
- def forward(self, x):
113
- x = self.proj(x)
114
- _, _, H, W = x.shape
115
- x = x.flatten(2).transpose(1, 2)
116
- x = self.norm(x)
117
- return x, H, W
118
-
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
models/modules/utils.py DELETED
@@ -1,54 +0,0 @@
1
- import torch.nn as nn
2
-
3
-
4
- def build_act_layer(act_layer):
5
- if act_layer == 'ReLU':
6
- return nn.ReLU(inplace=True)
7
- elif act_layer == 'SiLU':
8
- return nn.SiLU(inplace=True)
9
- elif act_layer == 'GELU':
10
- return nn.GELU()
11
-
12
- raise NotImplementedError(f'build_act_layer does not support {act_layer}')
13
-
14
-
15
- def build_norm_layer(dim,
16
- norm_layer,
17
- in_format='channels_last',
18
- out_format='channels_last',
19
- eps=1e-6):
20
- layers = []
21
- if norm_layer == 'BN':
22
- if in_format == 'channels_last':
23
- layers.append(to_channels_first())
24
- layers.append(nn.BatchNorm2d(dim))
25
- if out_format == 'channels_last':
26
- layers.append(to_channels_last())
27
- elif norm_layer == 'LN':
28
- if in_format == 'channels_first':
29
- layers.append(to_channels_last())
30
- layers.append(nn.LayerNorm(dim, eps=eps))
31
- if out_format == 'channels_first':
32
- layers.append(to_channels_first())
33
- else:
34
- raise NotImplementedError(
35
- f'build_norm_layer does not support {norm_layer}')
36
- return nn.Sequential(*layers)
37
-
38
-
39
- class to_channels_first(nn.Module):
40
-
41
- def __init__(self):
42
- super().__init__()
43
-
44
- def forward(self, x):
45
- return x.permute(0, 3, 1, 2)
46
-
47
-
48
- class to_channels_last(nn.Module):
49
-
50
- def __init__(self):
51
- super().__init__()
52
-
53
- def forward(self, x):
54
- return x.permute(0, 2, 3, 1)
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
models/refinement/refiner.py DELETED
@@ -1,253 +0,0 @@
1
- import torch
2
- import torch.nn as nn
3
- from collections import OrderedDict
4
- import torch
5
- import torch.nn as nn
6
- import torch.nn.functional as F
7
- from torchvision.models import vgg16, vgg16_bn
8
- from torchvision.models import resnet50
9
-
10
- from config import Config
11
- from dataset import class_labels_TR_sorted
12
- from models.backbones.build_backbone import build_backbone
13
- from models.modules.decoder_blocks import BasicDecBlk
14
- from models.modules.lateral_blocks import BasicLatBlk
15
- from models.modules.ing import *
16
- from models.refinement.stem_layer import StemLayer
17
-
18
-
19
- class RefinerPVTInChannels4(nn.Module):
20
- def __init__(self, in_channels=3+1):
21
- super(RefinerPVTInChannels4, self).__init__()
22
- self.config = Config()
23
- self.epoch = 1
24
- self.bb = build_backbone(self.config.bb, params_settings='in_channels=4')
25
-
26
- lateral_channels_in_collection = {
27
- 'vgg16': [512, 256, 128, 64], 'vgg16bn': [512, 256, 128, 64], 'resnet50': [1024, 512, 256, 64],
28
- 'pvt_v2_b2': [512, 320, 128, 64], 'pvt_v2_b5': [512, 320, 128, 64],
29
- 'swin_v1_b': [1024, 512, 256, 128], 'swin_v1_l': [1536, 768, 384, 192],
30
- }
31
- channels = lateral_channels_in_collection[self.config.bb]
32
- self.squeeze_module = BasicDecBlk(channels[0], channels[0])
33
-
34
- self.decoder = Decoder(channels)
35
-
36
- if 0:
37
- for key, value in self.named_parameters():
38
- if 'bb.' in key:
39
- value.requires_grad = False
40
-
41
- def forward(self, x):
42
- if isinstance(x, list):
43
- x = torch.cat(x, dim=1)
44
- ########## Encoder ##########
45
- if self.config.bb in ['vgg16', 'vgg16bn', 'resnet50']:
46
- x1 = self.bb.conv1(x)
47
- x2 = self.bb.conv2(x1)
48
- x3 = self.bb.conv3(x2)
49
- x4 = self.bb.conv4(x3)
50
- else:
51
- x1, x2, x3, x4 = self.bb(x)
52
-
53
- x4 = self.squeeze_module(x4)
54
-
55
- ########## Decoder ##########
56
-
57
- features = [x, x1, x2, x3, x4]
58
- scaled_preds = self.decoder(features)
59
-
60
- return scaled_preds
61
-
62
-
63
- class Refiner(nn.Module):
64
- def __init__(self, in_channels=3+1):
65
- super(Refiner, self).__init__()
66
- self.config = Config()
67
- self.epoch = 1
68
- self.stem_layer = StemLayer(in_channels=in_channels, inter_channels=48, out_channels=3, norm_layer='BN' if self.config.batch_size > 1 else 'LN')
69
- self.bb = build_backbone(self.config.bb)
70
-
71
- lateral_channels_in_collection = {
72
- 'vgg16': [512, 256, 128, 64], 'vgg16bn': [512, 256, 128, 64], 'resnet50': [1024, 512, 256, 64],
73
- 'pvt_v2_b2': [512, 320, 128, 64], 'pvt_v2_b5': [512, 320, 128, 64],
74
- 'swin_v1_b': [1024, 512, 256, 128], 'swin_v1_l': [1536, 768, 384, 192],
75
- }
76
- channels = lateral_channels_in_collection[self.config.bb]
77
- self.squeeze_module = BasicDecBlk(channels[0], channels[0])
78
-
79
- self.decoder = Decoder(channels)
80
-
81
- if 0:
82
- for key, value in self.named_parameters():
83
- if 'bb.' in key:
84
- value.requires_grad = False
85
-
86
- def forward(self, x):
87
- if isinstance(x, list):
88
- x = torch.cat(x, dim=1)
89
- x = self.stem_layer(x)
90
- ########## Encoder ##########
91
- if self.config.bb in ['vgg16', 'vgg16bn', 'resnet50']:
92
- x1 = self.bb.conv1(x)
93
- x2 = self.bb.conv2(x1)
94
- x3 = self.bb.conv3(x2)
95
- x4 = self.bb.conv4(x3)
96
- else:
97
- x1, x2, x3, x4 = self.bb(x)
98
-
99
- x4 = self.squeeze_module(x4)
100
-
101
- ########## Decoder ##########
102
-
103
- features = [x, x1, x2, x3, x4]
104
- scaled_preds = self.decoder(features)
105
-
106
- return scaled_preds
107
-
108
-
109
- class Decoder(nn.Module):
110
- def __init__(self, channels):
111
- super(Decoder, self).__init__()
112
- self.config = Config()
113
- DecoderBlock = eval('BasicDecBlk')
114
- LateralBlock = eval('BasicLatBlk')
115
-
116
- self.decoder_block4 = DecoderBlock(channels[0], channels[1])
117
- self.decoder_block3 = DecoderBlock(channels[1], channels[2])
118
- self.decoder_block2 = DecoderBlock(channels[2], channels[3])
119
- self.decoder_block1 = DecoderBlock(channels[3], channels[3]//2)
120
-
121
- self.lateral_block4 = LateralBlock(channels[1], channels[1])
122
- self.lateral_block3 = LateralBlock(channels[2], channels[2])
123
- self.lateral_block2 = LateralBlock(channels[3], channels[3])
124
-
125
- if self.config.ms_supervision:
126
- self.conv_ms_spvn_4 = nn.Conv2d(channels[1], 1, 1, 1, 0)
127
- self.conv_ms_spvn_3 = nn.Conv2d(channels[2], 1, 1, 1, 0)
128
- self.conv_ms_spvn_2 = nn.Conv2d(channels[3], 1, 1, 1, 0)
129
- self.conv_out1 = nn.Sequential(nn.Conv2d(channels[3]//2, 1, 1, 1, 0))
130
-
131
- def forward(self, features):
132
- x, x1, x2, x3, x4 = features
133
- outs = []
134
- p4 = self.decoder_block4(x4)
135
- _p4 = F.interpolate(p4, size=x3.shape[2:], mode='bilinear', align_corners=True)
136
- _p3 = _p4 + self.lateral_block4(x3)
137
-
138
- p3 = self.decoder_block3(_p3)
139
- _p3 = F.interpolate(p3, size=x2.shape[2:], mode='bilinear', align_corners=True)
140
- _p2 = _p3 + self.lateral_block3(x2)
141
-
142
- p2 = self.decoder_block2(_p2)
143
- _p2 = F.interpolate(p2, size=x1.shape[2:], mode='bilinear', align_corners=True)
144
- _p1 = _p2 + self.lateral_block2(x1)
145
-
146
- _p1 = self.decoder_block1(_p1)
147
- _p1 = F.interpolate(_p1, size=x.shape[2:], mode='bilinear', align_corners=True)
148
- p1_out = self.conv_out1(_p1)
149
-
150
- if self.config.ms_supervision:
151
- outs.append(self.conv_ms_spvn_4(p4))
152
- outs.append(self.conv_ms_spvn_3(p3))
153
- outs.append(self.conv_ms_spvn_2(p2))
154
- outs.append(p1_out)
155
- return outs
156
-
157
-
158
- class RefUNet(nn.Module):
159
- # Refinement
160
- def __init__(self, in_channels=3+1):
161
- super(RefUNet, self).__init__()
162
- self.encoder_1 = nn.Sequential(
163
- nn.Conv2d(in_channels, 64, 3, 1, 1),
164
- nn.Conv2d(64, 64, 3, 1, 1),
165
- nn.BatchNorm2d(64),
166
- nn.ReLU(inplace=True)
167
- )
168
-
169
- self.encoder_2 = nn.Sequential(
170
- nn.MaxPool2d(2, 2, ceil_mode=True),
171
- nn.Conv2d(64, 64, 3, 1, 1),
172
- nn.BatchNorm2d(64),
173
- nn.ReLU(inplace=True)
174
- )
175
-
176
- self.encoder_3 = nn.Sequential(
177
- nn.MaxPool2d(2, 2, ceil_mode=True),
178
- nn.Conv2d(64, 64, 3, 1, 1),
179
- nn.BatchNorm2d(64),
180
- nn.ReLU(inplace=True)
181
- )
182
-
183
- self.encoder_4 = nn.Sequential(
184
- nn.MaxPool2d(2, 2, ceil_mode=True),
185
- nn.Conv2d(64, 64, 3, 1, 1),
186
- nn.BatchNorm2d(64),
187
- nn.ReLU(inplace=True)
188
- )
189
-
190
- self.pool4 = nn.MaxPool2d(2, 2, ceil_mode=True)
191
- #####
192
- self.decoder_5 = nn.Sequential(
193
- nn.Conv2d(64, 64, 3, 1, 1),
194
- nn.BatchNorm2d(64),
195
- nn.ReLU(inplace=True)
196
- )
197
- #####
198
- self.decoder_4 = nn.Sequential(
199
- nn.Conv2d(128, 64, 3, 1, 1),
200
- nn.BatchNorm2d(64),
201
- nn.ReLU(inplace=True)
202
- )
203
-
204
- self.decoder_3 = nn.Sequential(
205
- nn.Conv2d(128, 64, 3, 1, 1),
206
- nn.BatchNorm2d(64),
207
- nn.ReLU(inplace=True)
208
- )
209
-
210
- self.decoder_2 = nn.Sequential(
211
- nn.Conv2d(128, 64, 3, 1, 1),
212
- nn.BatchNorm2d(64),
213
- nn.ReLU(inplace=True)
214
- )
215
-
216
- self.decoder_1 = nn.Sequential(
217
- nn.Conv2d(128, 64, 3, 1, 1),
218
- nn.BatchNorm2d(64),
219
- nn.ReLU(inplace=True)
220
- )
221
-
222
- self.conv_d0 = nn.Conv2d(64, 1, 3, 1, 1)
223
-
224
- self.upscore2 = nn.Upsample(scale_factor=2, mode='bilinear', align_corners=True)
225
-
226
- def forward(self, x):
227
- outs = []
228
- if isinstance(x, list):
229
- x = torch.cat(x, dim=1)
230
- hx = x
231
-
232
- hx1 = self.encoder_1(hx)
233
- hx2 = self.encoder_2(hx1)
234
- hx3 = self.encoder_3(hx2)
235
- hx4 = self.encoder_4(hx3)
236
-
237
- hx = self.decoder_5(self.pool4(hx4))
238
- hx = torch.cat((self.upscore2(hx), hx4), 1)
239
-
240
- d4 = self.decoder_4(hx)
241
- hx = torch.cat((self.upscore2(d4), hx3), 1)
242
-
243
- d3 = self.decoder_3(hx)
244
- hx = torch.cat((self.upscore2(d3), hx2), 1)
245
-
246
- d2 = self.decoder_2(hx)
247
- hx = torch.cat((self.upscore2(d2), hx1), 1)
248
-
249
- d1 = self.decoder_1(hx)
250
-
251
- x = self.conv_d0(d1)
252
- outs.append(x)
253
- return outs
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
models/refinement/stem_layer.py DELETED
@@ -1,45 +0,0 @@
1
- import torch.nn as nn
2
- from models.modules.utils import build_act_layer, build_norm_layer
3
-
4
-
5
- class StemLayer(nn.Module):
6
- r""" Stem layer of InternImage
7
- Args:
8
- in_channels (int): number of input channels
9
- out_channels (int): number of output channels
10
- act_layer (str): activation layer
11
- norm_layer (str): normalization layer
12
- """
13
-
14
- def __init__(self,
15
- in_channels=3+1,
16
- inter_channels=48,
17
- out_channels=96,
18
- act_layer='GELU',
19
- norm_layer='BN'):
20
- super().__init__()
21
- self.conv1 = nn.Conv2d(in_channels,
22
- inter_channels,
23
- kernel_size=3,
24
- stride=1,
25
- padding=1)
26
- self.norm1 = build_norm_layer(
27
- inter_channels, norm_layer, 'channels_first', 'channels_first'
28
- )
29
- self.act = build_act_layer(act_layer)
30
- self.conv2 = nn.Conv2d(inter_channels,
31
- out_channels,
32
- kernel_size=3,
33
- stride=1,
34
- padding=1)
35
- self.norm2 = build_norm_layer(
36
- out_channels, norm_layer, 'channels_first', 'channels_first'
37
- )
38
-
39
- def forward(self, x):
40
- x = self.conv1(x)
41
- x = self.norm1(x)
42
- x = self.act(x)
43
- x = self.conv2(x)
44
- x = self.norm2(x)
45
- return x
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
preproc.py DELETED
@@ -1,85 +0,0 @@
1
- from PIL import Image, ImageEnhance
2
- import random
3
- import numpy as np
4
- import random
5
-
6
-
7
- def preproc(image, label, preproc_methods=['flip']):
8
- if 'flip' in preproc_methods:
9
- image, label = cv_random_flip(image, label)
10
- if 'crop' in preproc_methods:
11
- image, label = random_crop(image, label)
12
- if 'rotate' in preproc_methods:
13
- image, label = random_rotate(image, label)
14
- if 'enhance' in preproc_methods:
15
- image = color_enhance(image)
16
- if 'pepper' in preproc_methods:
17
- label = random_pepper(label)
18
- return image, label
19
-
20
-
21
- def cv_random_flip(img, label):
22
- if random.random() > 0.5:
23
- img = img.transpose(Image.FLIP_LEFT_RIGHT)
24
- label = label.transpose(Image.FLIP_LEFT_RIGHT)
25
- return img, label
26
-
27
-
28
- def random_crop(image, label):
29
- border = 30
30
- image_width = image.size[0]
31
- image_height = image.size[1]
32
- border = int(min(image_width, image_height) * 0.1)
33
- crop_win_width = np.random.randint(image_width - border, image_width)
34
- crop_win_height = np.random.randint(image_height - border, image_height)
35
- random_region = (
36
- (image_width - crop_win_width) >> 1, (image_height - crop_win_height) >> 1, (image_width + crop_win_width) >> 1,
37
- (image_height + crop_win_height) >> 1)
38
- return image.crop(random_region), label.crop(random_region)
39
-
40
-
41
- def random_rotate(image, label, angle=15):
42
- mode = Image.BICUBIC
43
- if random.random() > 0.8:
44
- random_angle = np.random.randint(-angle, angle)
45
- image = image.rotate(random_angle, mode)
46
- label = label.rotate(random_angle, mode)
47
- return image, label
48
-
49
-
50
- def color_enhance(image):
51
- bright_intensity = random.randint(5, 15) / 10.0
52
- image = ImageEnhance.Brightness(image).enhance(bright_intensity)
53
- contrast_intensity = random.randint(5, 15) / 10.0
54
- image = ImageEnhance.Contrast(image).enhance(contrast_intensity)
55
- color_intensity = random.randint(0, 20) / 10.0
56
- image = ImageEnhance.Color(image).enhance(color_intensity)
57
- sharp_intensity = random.randint(0, 30) / 10.0
58
- image = ImageEnhance.Sharpness(image).enhance(sharp_intensity)
59
- return image
60
-
61
-
62
- def random_gaussian(image, mean=0.1, sigma=0.35):
63
- def gaussianNoisy(im, mean=mean, sigma=sigma):
64
- for _i in range(len(im)):
65
- im[_i] += random.gauss(mean, sigma)
66
- return im
67
-
68
- img = np.asarray(image)
69
- width, height = img.shape
70
- img = gaussianNoisy(img[:].flatten(), mean, sigma)
71
- img = img.reshape([width, height])
72
- return Image.fromarray(np.uint8(img))
73
-
74
-
75
- def random_pepper(img, N=0.0015):
76
- img = np.array(img)
77
- noiseNum = int(N * img.shape[0] * img.shape[1])
78
- for i in range(noiseNum):
79
- randX = random.randint(0, img.shape[0] - 1)
80
- randY = random.randint(0, img.shape[1] - 1)
81
- if random.randint(0, 1) == 0:
82
- img[randX, randY] = 0
83
- else:
84
- img[randX, randY] = 255
85
- return Image.fromarray(img)
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
requirements.txt CHANGED
@@ -8,3 +8,5 @@ scipy==1.12.0
8
  scikit-image==0.22.0
9
  kornia==0.7.1
10
  gradio_imageslider==0.0.18
 
 
 
8
  scikit-image==0.22.0
9
  kornia==0.7.1
10
  gradio_imageslider==0.0.18
11
+ transformers==4.42.4
12
+ huggingface_hub==0.23.4