Spaces:
Sleeping
Sleeping
Refactor: Production-grade modular layout, NLTK permissions fix, and editorial audit dashboard overhaul
405a3f0 Download src/model.py from amadmateen/SentinelReview-AI: direct link, hf CLI and curl.
- Browser
- Download file 3.12 kB
-
https://huggingface.co/spaces/amadmateen/SentinelReview-AI/resolve/main/src/model.py
- Command line
-
hf download hf://spaces/amadmateen/SentinelReview-AI/src/model.py
-
curl -L -o model.py https://huggingface.co/spaces/amadmateen/SentinelReview-AI/resolve/main/src/model.py
3.12 kB
| import torch | |
| import torch.nn as nn | |
| from torchvision import models | |
| from transformers import DistilBertModel | |
| class ExactDistilBertResNetModel(nn.Module): | |
| """ | |
| Multimodal neural network architecture combining DistilBERT (text representation) | |
| and ResNet18 (image representation) with a metadata feed-forward layer. | |
| NOTE ON ARCHITECTURE CONSTRAINTS: | |
| -------------------------------- | |
| Although a ResNet18 module is initialized and its weights exist inside the | |
| pretrained checkpoint, the forward pass hardcodes a zero tensor placeholder for | |
| the ResNet features: | |
| `other_features = torch.zeros(batch_size, 512, ...)` | |
| This is preserved to guarantee identical prediction behavior without retraining, | |
| since the classifier weights corresponding to these dimensions are random initializations | |
| that never received gradients during training. | |
| """ | |
| def __init__(self, num_classes=2): | |
| super(ExactDistilBertResNetModel, self).__init__() | |
| # DistilBERT component | |
| self.distilbert = DistilBertModel.from_pretrained('distilbert-base-uncased') | |
| distilbert_output_dim = 768 # DistilBERT hidden output size | |
| # ResNet component - without the final fully connected classification layer | |
| self.resnet = models.resnet18(weights=None) | |
| self.resnet.fc = nn.Identity() | |
| # Metadata linear projection (6 input features -> 32 projection dimensions) | |
| self.metadata_fc = nn.Linear(6, 32) | |
| # Combined multimodal classifier | |
| # Inputs: 768 (DistilBERT CLS) + 32 (metadata) + 512 (ResNet image placeholder) = 1312 dimensions | |
| self.classifier = nn.Sequential( | |
| nn.Linear(1312, 128), | |
| nn.ReLU(), | |
| nn.Dropout(0.1), | |
| nn.Linear(128, num_classes) | |
| ) | |
| def forward(self, input_ids=None, attention_mask=None, metadata=None): | |
| # 1. Process text with DistilBERT and grab [CLS] token representation | |
| distilbert_outputs = self.distilbert( | |
| input_ids=input_ids, | |
| attention_mask=attention_mask | |
| ) | |
| text_features = distilbert_outputs.last_hidden_state[:, 0, :] # Shape: [batch_size, 768] | |
| # 2. Process review-level metadata features | |
| if metadata is not None: | |
| metadata_features = self.metadata_fc(metadata) # Shape: [batch_size, 32] | |
| else: | |
| batch_size = text_features.shape[0] | |
| metadata_features = torch.zeros(batch_size, 32, device=text_features.device) | |
| # 3. Apply the ResNet features placeholder (strictly 0s as in original model checkpoint configuration) | |
| batch_size = text_features.shape[0] | |
| other_features = torch.zeros(batch_size, 512, device=text_features.device) | |
| # 4. Concatenate vectors along features dimension | |
| combined_features = torch.cat([text_features, metadata_features, other_features], dim=1) | |
| # 5. Execute final classification head layers | |
| logits = self.classifier(combined_features) | |
| return logits | |