Spaces:
Running
Running
File size: 11,783 Bytes
c9baa67 |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 |
import functools
import math
import numpy as np
import torch
import torch.nn as nn
import torch.nn.functional as F
from .layers import GaussianFilterNd
def encode_scanpath_features(x_hist, y_hist, size, device=None, include_x=True, include_y=True, include_duration=False):
assert include_x
assert include_y
assert not include_duration
height = size[0]
width = size[1]
xs = torch.arange(width, dtype=torch.float32).to(device)
ys = torch.arange(height, dtype=torch.float32).to(device)
YS, XS = torch.meshgrid(ys, xs, indexing='ij')
XS = torch.repeat_interleave(
torch.repeat_interleave(
XS[np.newaxis, np.newaxis, :, :],
repeats=x_hist.shape[0],
dim=0,
),
repeats=x_hist.shape[1],
dim=1,
)
YS = torch.repeat_interleave(
torch.repeat_interleave(
YS[np.newaxis, np.newaxis, :, :],
repeats=y_hist.shape[0],
dim=0,
),
repeats=y_hist.shape[1],
dim=1,
)
XS -= x_hist.unsqueeze(2).unsqueeze(3)
YS -= y_hist.unsqueeze(2).unsqueeze(3)
distances = torch.sqrt(XS**2 + YS**2)
return torch.cat((XS, YS, distances), axis=1)
class FeatureExtractor(torch.nn.Module):
def __init__(self, features, targets):
super().__init__()
self.features = features
self.targets = targets
#print("Targets are {}".format(targets))
self.outputs = {}
for target in targets:
layer = dict([*self.features.named_modules()])[target]
layer.register_forward_hook(self.save_outputs_hook(target))
def save_outputs_hook(self, layer_id: str):
def fn(_, __, output):
self.outputs[layer_id] = output.clone()
return fn
def forward(self, x):
self.outputs.clear()
self.features(x)
return [self.outputs[target] for target in self.targets]
def upscale(tensor, size):
tensor_size = torch.tensor(tensor.shape[2:]).type(torch.float32)
target_size = torch.tensor(size).type(torch.float32)
factors = torch.ceil(target_size / tensor_size)
factor = torch.max(factors).type(torch.int64).to(tensor.device)
assert factor >= 1
tensor = torch.repeat_interleave(tensor, factor, dim=2)
tensor = torch.repeat_interleave(tensor, factor, dim=3)
tensor = tensor[:, :, :size[0], :size[1]]
return tensor
class Finalizer(nn.Module):
"""Transforms a readout into a gaze prediction
A readout network returns a single, spatial map of probable gaze locations.
This module bundles the common processing steps necessary to transform this into
the predicted gaze distribution:
- resizing to the stimulus size
- smoothing of the prediction using a gaussian filter
- removing of channel and time dimension
- weighted addition of the center bias
- normalization
"""
def __init__(
self,
sigma,
kernel_size=None,
learn_sigma=False,
center_bias_weight=1.0,
learn_center_bias_weight=True,
saliency_map_factor=4,
):
"""Creates a new finalizer
Args:
size (tuple): target size for the predictions
sigma (float): standard deviation of the gaussian kernel used for smoothing
kernel_size (int, optional): size of the gaussian kernel
learn_sigma (bool, optional): If True, the standard deviation of the gaussian kernel will
be learned (default: False)
center_bias (string or tensor): the center bias
center_bias_weight (float, optional): initial weight of the center bias
learn_center_bias_weight (bool, optional): If True, the center bias weight will be
learned (default: True)
"""
super(Finalizer, self).__init__()
self.saliency_map_factor = saliency_map_factor
self.gauss = GaussianFilterNd([2, 3], sigma, truncate=3, trainable=learn_sigma)
self.center_bias_weight = nn.Parameter(torch.Tensor([center_bias_weight]), requires_grad=learn_center_bias_weight)
def forward(self, readout, centerbias):
"""Applies the finalization steps to the given readout"""
downscaled_centerbias = F.interpolate(
centerbias.view(centerbias.shape[0], 1, centerbias.shape[1], centerbias.shape[2]),
scale_factor=1 / self.saliency_map_factor,
recompute_scale_factor=False,
)[:, 0, :, :]
out = F.interpolate(
readout,
size=[downscaled_centerbias.shape[1], downscaled_centerbias.shape[2]]
)
# apply gaussian filter
out = self.gauss(out)
# remove channel dimension
out = out[:, 0, :, :]
# add to center bias
out = out + self.center_bias_weight * downscaled_centerbias
out = F.interpolate(out[:, np.newaxis, :, :], size=[centerbias.shape[1], centerbias.shape[2]])[:, 0, :, :]
# normalize
out = out - out.logsumexp(dim=(1, 2), keepdim=True)
return out
class DeepGazeII(torch.nn.Module):
def __init__(self, features, readout_network, downsample=2, readout_factor=16, saliency_map_factor=2, initial_sigma=8.0):
super().__init__()
self.readout_factor = readout_factor
self.saliency_map_factor = saliency_map_factor
self.features = features
for param in self.features.parameters():
param.requires_grad = False
self.features.eval()
self.readout_network = readout_network
self.finalizer = Finalizer(
sigma=initial_sigma,
learn_sigma=True,
saliency_map_factor=self.saliency_map_factor,
)
self.downsample = downsample
def forward(self, x, centerbias):
orig_shape = x.shape
x = F.interpolate(
x,
scale_factor=1 / self.downsample,
recompute_scale_factor=False,
)
x = self.features(x)
readout_shape = [math.ceil(orig_shape[2] / self.downsample / self.readout_factor), math.ceil(orig_shape[3] / self.downsample / self.readout_factor)]
x = [F.interpolate(item, readout_shape) for item in x]
x = torch.cat(x, dim=1)
x = self.readout_network(x)
x = self.finalizer(x, centerbias)
return x
def train(self, mode=True):
self.features.eval()
self.readout_network.train(mode=mode)
self.finalizer.train(mode=mode)
class DeepGazeIII(torch.nn.Module):
def __init__(self, features, saliency_network, scanpath_network, fixation_selection_network, downsample=2, readout_factor=2, saliency_map_factor=2, included_fixations=-2, initial_sigma=8.0):
super().__init__()
self.downsample = downsample
self.readout_factor = readout_factor
self.saliency_map_factor = saliency_map_factor
self.included_fixations = included_fixations
self.features = features
for param in self.features.parameters():
param.requires_grad = False
self.features.eval()
self.saliency_network = saliency_network
self.scanpath_network = scanpath_network
self.fixation_selection_network = fixation_selection_network
self.finalizer = Finalizer(
sigma=initial_sigma,
learn_sigma=True,
saliency_map_factor=self.saliency_map_factor,
)
def forward(self, x, centerbias, x_hist=None, y_hist=None, durations=None):
orig_shape = x.shape
x = F.interpolate(x, scale_factor=1 / self.downsample)
x = self.features(x)
readout_shape = [math.ceil(orig_shape[2] / self.downsample / self.readout_factor), math.ceil(orig_shape[3] / self.downsample / self.readout_factor)]
x = [F.interpolate(item, readout_shape) for item in x]
x = torch.cat(x, dim=1)
x = self.saliency_network(x)
if self.scanpath_network is not None:
scanpath_features = encode_scanpath_features(x_hist, y_hist, size=(orig_shape[2], orig_shape[3]), device=x.device)
#scanpath_features = F.interpolate(scanpath_features, scale_factor=1 / self.downsample / self.readout_factor)
scanpath_features = F.interpolate(scanpath_features, readout_shape)
y = self.scanpath_network(scanpath_features)
else:
y = None
x = self.fixation_selection_network((x, y))
x = self.finalizer(x, centerbias)
return x
def train(self, mode=True):
self.features.eval()
self.saliency_network.train(mode=mode)
if self.scanpath_network is not None:
self.scanpath_network.train(mode=mode)
self.fixation_selection_network.train(mode=mode)
self.finalizer.train(mode=mode)
class DeepGazeIIIMixture(torch.nn.Module):
def __init__(self, features, saliency_networks, scanpath_networks, fixation_selection_networks, finalizers, downsample=2, readout_factor=2, saliency_map_factor=2, included_fixations=-2, initial_sigma=8.0):
super().__init__()
self.downsample = downsample
self.readout_factor = readout_factor
self.saliency_map_factor = saliency_map_factor
self.included_fixations = included_fixations
self.features = features
for param in self.features.parameters():
param.requires_grad = False
self.features.eval()
self.saliency_networks = torch.nn.ModuleList(saliency_networks)
self.scanpath_networks = torch.nn.ModuleList(scanpath_networks)
self.fixation_selection_networks = torch.nn.ModuleList(fixation_selection_networks)
self.finalizers = torch.nn.ModuleList(finalizers)
def forward(self, x, centerbias, x_hist=None, y_hist=None, durations=None):
orig_shape = x.shape
x = F.interpolate(
x,
scale_factor=1 / self.downsample,
recompute_scale_factor=False,
)
x = self.features(x)
readout_shape = [math.ceil(orig_shape[2] / self.downsample / self.readout_factor), math.ceil(orig_shape[3] / self.downsample / self.readout_factor)]
x = [F.interpolate(item, readout_shape) for item in x]
x = torch.cat(x, dim=1)
predictions = []
readout_input = x
for saliency_network, scanpath_network, fixation_selection_network, finalizer in zip(
self.saliency_networks, self.scanpath_networks, self.fixation_selection_networks, self.finalizers
):
x = saliency_network(readout_input)
if scanpath_network is not None:
scanpath_features = encode_scanpath_features(x_hist, y_hist, size=(orig_shape[2], orig_shape[3]), device=x.device)
scanpath_features = F.interpolate(scanpath_features, readout_shape)
y = scanpath_network(scanpath_features)
else:
y = None
x = fixation_selection_network((x, y))
x = finalizer(x, centerbias)
predictions.append(x[:, np.newaxis, :, :])
predictions = torch.cat(predictions, dim=1) - np.log(len(self.saliency_networks))
prediction = predictions.logsumexp(dim=(1), keepdim=True)
return prediction
class MixtureModel(torch.nn.Module):
def __init__(self, models):
super().__init__()
self.models = torch.nn.ModuleList(models)
def forward(self, *args, **kwargs):
predictions = [model.forward(*args, **kwargs) for model in self.models]
predictions = torch.cat(predictions, dim=1)
predictions -= np.log(len(self.models))
prediction = predictions.logsumexp(dim=(1), keepdim=True)
return prediction
|