| 115 | |
| 116 | |
| 117 | class CLIPLoss(nn.Module): |
| 118 | def get_logits( |
| 119 | self, |
| 120 | image_features: torch.FloatTensor, |
| 121 | spectrum_features: torch.FloatTensor, |
| 122 | logit_scale: float, |
| 123 | ) -> Tuple[torch.FloatTensor, torch.FloatTensor]: |
| 124 | # Normalize image features |
| 125 | image_features = F.normalize(image_features, dim=-1, eps=1e-3) |
| 126 | |
| 127 | # Normalize spectrum features |
| 128 | spectrum_features = F.normalize(spectrum_features, dim=-1, eps=1e-3) |
| 129 | |
| 130 | # Calculate the logits for the image and spectrum features |
| 131 | logits_per_image = logit_scale * image_features @ spectrum_features.T |
| 132 | return logits_per_image, logits_per_image.T |
| 133 | |
| 134 | def forward( |
| 135 | self, |
| 136 | image_features: torch.FloatTensor, |
| 137 | spectrum_features: torch.FloatTensor, |
| 138 | logit_scale: float, |
| 139 | output_dict: bool = False, |
| 140 | ) -> torch.FloatTensor: |
| 141 | # Get the logits for the image and spectrum features |
| 142 | logits_per_image, logits_per_spectrum = self.get_logits( |
| 143 | image_features, spectrum_features, logit_scale |
| 144 | ) |
| 145 | |
| 146 | # Calculate the contrastive loss |
| 147 | labels = torch.arange( |
| 148 | logits_per_image.shape[0], device=image_features.device, dtype=torch.long |
| 149 | ) |
| 150 | total_loss = ( |
| 151 | F.cross_entropy(logits_per_image, labels) |
| 152 | + F.cross_entropy(logits_per_spectrum, labels) |
| 153 | ) / 2 |
| 154 | return {"contrastive_loss": total_loss} if output_dict else total_loss |
| 155 | |
| 156 | |
| 157 | class ImageHead(nn.Module): |