model.py

import torch.nn as nn
import torchvision.models as models
import torch
from transformers import AutoTokenizer, AutoModel, AutoConfig

class ImageModel(nn.Module):
    def __init__(self, num_genre) -> None:
        super(ImageModel, self).__init__()
        # Feature extraction layer.
        # Input 200x200
        self.features = models.mobilenet_v3_large(weights="IMAGENET1K_V2")
        in_features = self.features.classifier[0].in_features
        self.features.classifier = nn.Identity()
        self.head_score = nn.Sequential(
            nn.Dropout(p=0.5),
            nn.Linear(in_features=in_features, out_features=512, bias=True),
            nn.SiLU(),
            nn.Dropout(p=0.2),
            nn.Linear(in_features=512, out_features=1, bias=True)
        )
        self.head_award = nn.Sequential(
            nn.Dropout(p=0.5),
            nn.Linear(in_features=in_features, out_features=512, bias=True),
            nn.SiLU(),
            nn.Dropout(p=0.2),
            nn.Linear(in_features=512, out_features=1, bias=True)
        )
        self.head_genre = nn.Sequential(
            nn.Dropout(p=0.5),
            nn.Linear(in_features=in_features, out_features=1024, bias=True),
            nn.Hardswish(),
            nn.Dropout(p=0.2),
            nn.Linear(in_features=1024, out_features=num_genre, bias=True)
        )
        
        # Initialize model weights.
        self._initialize_weights()

    def forward(self, x: torch.Tensor) -> torch.Tensor:
        x = self.features(x)
        return self.head_score(x), self.head_award(x), self.head_genre(x)


    # The filter weight of each layer is a Gaussian distribution with zero mean and standard deviation initialized by random extraction 0.001 (deviation is 0).
    def _initialize_weights(model):
        """
        Initializes weights of all layers in a PyTorch model.

        Args:
            model (nn.Module): The model to initialize weights for.
        """
        for m in model.modules():
            if isinstance(m, nn.Conv2d):
                nn.init.xavier_normal_(m.weight)
            elif isinstance(m, nn.Linear):
                nn.init.xavier_normal_(m.weight)
            elif isinstance(m, nn.BatchNorm2d):
                nn.init.constant_(m.weight, 1)
                nn.init.constant_(m.bias, 0)
                
                
class TextModel(nn.Module):
    def __init__(self, model_name, num_genre):
        super(TextModel, self).__init__()
        config = AutoConfig.from_pretrained(model_name)
        in_features = config.hidden_size
        self.transformer = AutoModel.from_pretrained(model_name)
        
        self.head_score = nn.Sequential(
            # nn.Dropout(p=0.5),
            # nn.Linear(in_features=in_features, out_features=512, bias=True),
            # nn.SiLU(),
            nn.Dropout(p=0.2),
            nn.Linear(in_features=in_features, out_features=1, bias=True)
        )
        self.head_award = nn.Sequential(
            # nn.Dropout(p=0.5),
            # nn.Linear(in_features=in_features, out_features=512, bias=True),
            # nn.SiLU(),
            nn.Dropout(p=0.2),
            nn.Linear(in_features=in_features, out_features=1, bias=True)
        )
        self.head_genre = nn.Sequential(
            # nn.Linear(in_features=in_features, out_features=1024, bias=True),
            # nn.Hardswish(),
            nn.Dropout(p=0.2),
            nn.Linear(in_features=in_features, out_features=num_genre, bias=True)
        )

    def forward(self, x):
        x = self.transformer(input_ids=x[0], attention_mask=x[1])['pooler_output']
        return self.head_score(x), self.head_award(x), self.head_genre(x)