modeling_trendyol_dinov2.py

"""
Hugging Face compatible model implementation for Trendyol DinoV2
"""
import torch
import torch.nn as nn
from transformers import PreTrainedModel, PretrainedConfig
from transformers.modeling_outputs import BaseModelOutput
from typing import Optional, Tuple, Union
import torch.nn.functional as F


class TrendyolDinoV2Config(PretrainedConfig):
    """
    Configuration class for TrendyolDinoV2 model.
    """
    model_type = "trendyol_dinov2"
    
    def __init__(
        self,
        embedding_dim=256,
        input_size=224,
        hidden_size=256,
        backbone_name="dinov2_vitb14",
        in_features=768,
        downscale_size=332,
        pad_color=255,
        jpeg_quality=90,
        **kwargs
    ):
        super().__init__(**kwargs)
        self.embedding_dim = embedding_dim
        self.input_size = input_size
        self.hidden_size = hidden_size
        self.backbone_name = backbone_name
        self.in_features = in_features
        self.downscale_size = downscale_size
        self.pad_color = pad_color
        self.jpeg_quality = jpeg_quality


class TYArcFaceDinoV2(nn.Module):
    """Core model architecture"""
    def __init__(self, config):
        super(TYArcFaceDinoV2, self).__init__()
        self.config = config
        
        # Load DinoV2 backbone
        try:
            self.backbone = torch.hub.load('facebookresearch/dinov2', config.backbone_name)
        except Exception as e:
            raise RuntimeError(f"Failed to load DinoV2 backbone: {e}")
            
        self.hidden_size = config.hidden_size
        self.in_features = config.in_features
        self.embedding_dim = config.embedding_dim
        
        self.bn1 = nn.BatchNorm2d(self.in_features)
        # Freeze backbone
        self.backbone.requires_grad_(False)
        
        # Projection layers
        self.fc11 = nn.Linear(self.in_features * self.hidden_size, self.embedding_dim)
        self.bn11 = nn.BatchNorm1d(self.embedding_dim)

    def forward(self, pixel_values):
        try:
            features = self.backbone.get_intermediate_layers(
                pixel_values, return_class_token=True, reshape=True
            )
            features = features[0][0]  # Get the features
            features = self.bn1(features)
            features = features.flatten(start_dim=1)
            features = self.fc11(features)
            features = self.bn11(features)
            features = F.normalize(features)
            return features
        except Exception as e:
            raise RuntimeError(f"Forward pass failed: {e}")


class TrendyolDinoV2Model(PreTrainedModel):
    """
    Hugging Face compatible wrapper for TrendyolDinoV2
    """
    config_class = TrendyolDinoV2Config
    base_model_prefix = "model"
    
    def __init__(self, config):
        super().__init__(config)
        self.model = TYArcFaceDinoV2(config)
        
        # Initialize weights
        self.init_weights()
    
    def _init_weights(self, module):
        """Initialize weights (required by PreTrainedModel)"""
        if isinstance(module, nn.Linear):
            module.weight.data.normal_(mean=0.0, std=0.02)
            if module.bias is not None:
                module.bias.data.zero_()
        elif isinstance(module, nn.BatchNorm1d) or isinstance(module, nn.BatchNorm2d):
            module.bias.data.zero_()
            module.weight.data.fill_(1.0)
    
    def init_weights(self):
        """Initialize all weights in the model"""
        self.apply(self._init_weights)
    
    def forward(
        self,
        pixel_values: Optional[torch.Tensor] = None,
        output_hidden_states: Optional[bool] = None,
        return_dict: Optional[bool] = None,
        **kwargs
    ):
        return_dict = return_dict if return_dict is not None else getattr(self.config, 'use_return_dict', True)
        
        if pixel_values is None:
            raise ValueError("pixel_values cannot be None")
        
        # Get embeddings from the model
        embeddings = self.model(pixel_values)
        
        if not return_dict:
            return (embeddings,)
        
        return BaseModelOutput(
            last_hidden_state=embeddings,
            hidden_states=None,
            attentions=None
        )
    
    def get_embeddings(self, pixel_values):
        """Convenience method to get embeddings directly"""
        with torch.no_grad():
            outputs = self.forward(pixel_values, return_dict=True)
            return outputs.last_hidden_state


# Register the configuration
TrendyolDinoV2Config.register_for_auto_class()
TrendyolDinoV2Model.register_for_auto_class("AutoModel")