Upload 2 files

loss_fn.py

@@ -0,0 +1,201 @@
+# https://github.com/Alibaba-MIIL/ASL/blob/main/src/loss_functions/losses.py
+
+# MIT License
+
+# Copyright (c) 2020 Alibaba-MIIL
+
+# Permission is hereby granted, free of charge, to any person obtaining a copy
+# of this software and associated documentation files (the "Software"), to deal
+# in the Software without restriction, including without limitation the rights
+# to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+# copies of the Software, and to permit persons to whom the Software is
+# furnished to do so, subject to the following conditions:
+
+# The above copyright notice and this permission notice shall be included in all
+# copies or substantial portions of the Software.
+
+# THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+# IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+# FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+# AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+# LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+# OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+# SOFTWARE.
+
+
+import torch
+import torch.nn as nn
+
+
+class AsymmetricLoss(nn.Module):
+    def __init__(
+        self,
+        gamma_neg=4,
+        gamma_pos=1,
+        clip=0.05,
+        eps=1e-8,
+        disable_torch_grad_focal_loss=True,
+    ):
+        super(AsymmetricLoss, self).__init__()
+
+        self.gamma_neg = gamma_neg
+        self.gamma_pos = gamma_pos
+        self.clip = clip
+        self.disable_torch_grad_focal_loss = disable_torch_grad_focal_loss
+        self.eps = eps
+
+    def forward(self, x, y):
+        """ "
+        Parameters
+        ----------
+        x: input logits
+        y: targets (multi-label binarized vector)
+        """
+
+        # Calculating Probabilities
+        x_sigmoid = torch.sigmoid(x)
+        xs_pos = x_sigmoid
+        xs_neg = 1 - x_sigmoid
+
+        # Asymmetric Clipping
+        if self.clip is not None and self.clip > 0:
+            xs_neg = (xs_neg + self.clip).clamp(max=1)
+
+        # Basic CE calculation
+        los_pos = y * torch.log(xs_pos.clamp(min=self.eps))
+        los_neg = (1 - y) * torch.log(xs_neg.clamp(min=self.eps))
+        loss = los_pos + los_neg
+
+        # Asymmetric Focusing
+        if self.gamma_neg > 0 or self.gamma_pos > 0:
+            if self.disable_torch_grad_focal_loss:
+                torch.set_grad_enabled(False)
+            pt0 = xs_pos * y
+            pt1 = xs_neg * (1 - y)  # pt = p if t > 0 else 1-p
+            pt = pt0 + pt1
+            one_sided_gamma = self.gamma_pos * y + self.gamma_neg * (1 - y)
+            one_sided_w = torch.pow(1 - pt, one_sided_gamma)
+            if self.disable_torch_grad_focal_loss:
+                torch.set_grad_enabled(True)
+            loss *= one_sided_w
+
+        return -loss.sum()
+
+
+class AsymmetricLossOptimized(nn.Module):
+    """Notice - optimized version, minimizes memory allocation and gpu uploading,
+    favors inplace operations"""
+
+    def __init__(
+        self,
+        gamma_neg=4,
+        gamma_pos=1,
+        clip=0.05,
+        eps=1e-8,
+        disable_torch_grad_focal_loss=False,
+    ):
+        super(AsymmetricLossOptimized, self).__init__()
+
+        self.gamma_neg = gamma_neg
+        self.gamma_pos = gamma_pos
+        self.clip = clip
+        self.disable_torch_grad_focal_loss = disable_torch_grad_focal_loss
+        self.eps = eps
+
+        # prevent memory allocation and gpu uploading every iteration, and encourages inplace operations
+        self.targets = self.anti_targets = self.xs_pos = self.xs_neg = (
+            self.asymmetric_w
+        ) = self.loss = None
+
+    def forward(self, x, y):
+        """ "
+        Parameters
+        ----------
+        x: input logits
+        y: targets (multi-label binarized vector)
+        """
+
+        self.targets = y
+        self.anti_targets = 1 - y
+
+        # Calculating Probabilities
+        self.xs_pos = torch.sigmoid(x)
+        self.xs_neg = 1.0 - self.xs_pos
+
+        # Asymmetric Clipping
+        if self.clip is not None and self.clip > 0:
+            self.xs_neg.add_(self.clip).clamp_(max=1)
+
+        # Basic CE calculation
+        self.loss = self.targets * torch.log(self.xs_pos.clamp(min=self.eps))
+        self.loss.add_(self.anti_targets * torch.log(self.xs_neg.clamp(min=self.eps)))
+
+        # Asymmetric Focusing
+        if self.gamma_neg > 0 or self.gamma_pos > 0:
+            if self.disable_torch_grad_focal_loss:
+                torch.set_grad_enabled(False)
+            self.xs_pos = self.xs_pos * self.targets
+            self.xs_neg = self.xs_neg * self.anti_targets
+            self.asymmetric_w = torch.pow(
+                1 - self.xs_pos - self.xs_neg,
+                self.gamma_pos * self.targets + self.gamma_neg * self.anti_targets,
+            )
+            if self.disable_torch_grad_focal_loss:
+                torch.set_grad_enabled(True)
+            self.loss *= self.asymmetric_w
+
+        return -self.loss.sum()
+
+
+class ASLSingleLabel(nn.Module):
+    """
+    This loss is intended for single-label classification problems
+    """
+
+    def __init__(self, gamma_pos=0, gamma_neg=4, eps: float = 0.1, reduction="mean"):
+        super(ASLSingleLabel, self).__init__()
+
+        self.eps = eps
+        self.logsoftmax = nn.LogSoftmax(dim=-1)
+        self.targets_classes = []
+        self.gamma_pos = gamma_pos
+        self.gamma_neg = gamma_neg
+        self.reduction = reduction
+
+    def forward(self, inputs, target):
+        """
+        "input" dimensions: - (batch_size,number_classes)
+        "target" dimensions: - (batch_size)
+        """
+        num_classes = inputs.size()[-1]
+        log_preds = self.logsoftmax(inputs)
+        self.targets_classes = torch.zeros_like(inputs).scatter_(
+            1, target.long().unsqueeze(1), 1
+        )
+
+        # ASL weights
+        targets = self.targets_classes
+        anti_targets = 1 - targets
+        xs_pos = torch.exp(log_preds)
+        xs_neg = 1 - xs_pos
+        xs_pos = xs_pos * targets
+        xs_neg = xs_neg * anti_targets
+        asymmetric_w = torch.pow(
+            1 - xs_pos - xs_neg,
+            self.gamma_pos * targets + self.gamma_neg * anti_targets,
+        )
+        log_preds = log_preds * asymmetric_w
+
+        if self.eps > 0:  # label smoothing
+            self.targets_classes = self.targets_classes.mul(1 - self.eps).add(
+                self.eps / num_classes
+            )
+
+        # loss calculation
+        loss = -self.targets_classes.mul(log_preds)
+
+        loss = loss.sum(dim=-1)
+        if self.reduction == "mean":
+            loss = loss.mean()
+
+        return loss

modeling_siglip.py

@@ -0,0 +1,83 @@
+from dataclasses import dataclass
+
+import torch
+import torch.nn as nn
+from torch.nn import BCEWithLogitsLoss, CrossEntropyLoss, MSELoss
+
+from transformers import SiglipVisionModel, SiglipPreTrainedModel, SiglipVisionConfig
+from transformers.utils import ModelOutput
+
+from loss_fn import AsymmetricLossOptimized
+
+
+@dataclass
+class SiglipForImageClassifierOutput(ModelOutput):
+    loss: torch.FloatTensor | None = None
+    logits: torch.FloatTensor | None = None
+    pooler_output: torch.FloatTensor | None = None
+    hidden_states: tuple[torch.FloatTensor, ...] | None = None
+    attentions: tuple[torch.FloatTensor, ...] | None = None
+
+
+class SiglipForImageClassification(SiglipPreTrainedModel):
+    config_class = SiglipVisionConfig
+    main_input_name = "pixel_values"
+
+    def __init__(
+        self,
+        config,
+    ):
+        super().__init__(config)
+
+        self.num_labels = config.num_labels
+        self.siglip = SiglipVisionModel(config)
+
+        # Classifier head
+        self.classifier = (
+            nn.Linear(config.hidden_size, config.num_labels)
+            if config.num_labels > 0
+            else nn.Identity()
+        )
+
+        # Initialize weights and apply final processing
+        self.post_init()
+
+    def forward(
+        self, pixel_values: torch.FloatTensor, labels: torch.LongTensor | None = None
+    ):
+        outputs = self.siglip(pixel_values)
+        pooler_output = outputs.pooler_output
+        logits = self.classifier(pooler_output)
+
+        loss = None
+        if labels is not None:
+            if self.config.problem_type is None:
+                if self.num_labels == 1:
+                    self.config.problem_type = "regression"
+                elif self.num_labels > 1 and (
+                    labels.dtype == torch.long or labels.dtype == torch.int
+                ):
+                    self.config.problem_type = "single_label_classification"
+                else:
+                    self.config.problem_type = "multi_label_classification"
+
+            if self.config.problem_type == "regression":
+                loss_fct = MSELoss()
+                if self.num_labels == 1:
+                    loss = loss_fct(logits.squeeze(), labels.squeeze())
+                else:
+                    loss = loss_fct(logits, labels)
+            elif self.config.problem_type == "single_label_classification":
+                loss_fct = CrossEntropyLoss()
+                loss = loss_fct(logits.view(-1, self.num_labels), labels.view(-1))
+            elif self.config.problem_type == "multi_label_classification":
+                loss_fct = AsymmetricLossOptimized()
+                loss = loss_fct(logits, labels)
+
+        return SiglipForImageClassifierOutput(
+            loss=loss,
+            logits=logits,
+            pooler_output=outputs.pooler_output,
+            hidden_states=outputs.hidden_states,
+            attentions=outputs.attentions,
+        )