File size: 3,751 Bytes
df511be 259423f df511be |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 |
import torch
import torch.nn as nn
from transformers import CLIPTextModel, RobertaModel, CLIPVisionModel
from timm import create_model
EMBEDDING_DIM = 512
class ImageEncoder(nn.Module):
def __init__(self):
super(ImageEncoder, self).__init__()
# Load the Swin Transformer with features_only=True
self.swin = create_model("swin-tiny-patch4-window7-224", pretrained=True, features_only=True)
for param in self.swin.parameters():
param.requires_grad = True
# Get the feature size of the final stage
self.swin_output_dim = self.swin.feature_info.channels()[-1] # Last stage: 1024 channels
# Define FC layer
self.fc1 = nn.Linear(self.swin_output_dim * 7 * 7, EMBEDDING_DIM) # Flattened input size
nn.init.xavier_uniform_(self.fc1.weight)
nn.init.zeros_(self.fc1.bias)
for param in self.fc1.parameters():
param.requires_grad = True
def forward(self, x):
# Extract features from Swin
swin_features = self.swin(x)[-1] # Use the last stage feature map (e.g., [B, 1024, 7, 7])
# Flatten feature map
swin_features = swin_features.view(swin_features.size(0), -1) # Shape: (B, 1024*7*7)
# Pass through FC layer
output = self.fc1(swin_features) # Shape: (B, embedding_dim)
return output
from transformers import RobertaModel
class RobertaEncoder(nn.Module):
def __init__(self, roberta_model_path="roberta-base"):
super(RobertaEncoder, self).__init__()
# Load pre-trained RoBERTa model
self.roberta = RobertaModel.from_pretrained(roberta_model_path)
# Add a linear projection layer to reduce dimensionality
self.projection = nn.Linear(self.roberta.config.hidden_size, EMBEDDING_DIM)
# Initialize the projection layer weights
nn.init.xavier_uniform_(self.projection.weight)
nn.init.zeros_(self.projection.bias)
# Allow fine-tuning of the RoBERTa model
for param in self.roberta.parameters():
param.requires_grad = True
def forward(self, input_ids, attention_mask):
"""
Forward pass through RoBERTa.
Args:
input_ids: Tensor of shape (batch_size, seq_length)
attention_mask: Tensor of shape (batch_size, seq_length)
Returns:
Embedding: Tensor of shape (batch_size, EMBEDDING_DIM)
"""
roberta_output = self.roberta(input_ids=input_ids, attention_mask=attention_mask)
cls_token = roberta_output.last_hidden_state[:, 0, :] # Use CLS token
pooled_output = torch.mean(roberta_output.last_hidden_state, dim=1) # Mean pooling
return self.projection(cls_token+pooled_output)
from transformers import AutoTokenizer, Siglip2TextModel,AutoModel
class SigLIP2TextEncoder(nn.Module):
def __init__(self, embedding_dim=512):
super(SigLIP2TextEncoder, self).__init__()
model = AutoModel.from_pretrained("google/siglip2-base-patch16-224")
self.text_encoder = model.text_model
hidden_size = self.text_encoder.config.hidden_size
self.projection = nn.Linear(hidden_size, embedding_dim)
nn.init.xavier_uniform_(self.projection.weight)
nn.init.zeros_(self.projection.bias)
for param in self.text_encoder.parameters():
param.requires_grad = True
for param in self.projection.parameters():
param.requires_grad = True
def forward(self, tokens):
"""
Args:
tokens:
Returns:
Tensor of shape (batch_size, embedding_dim)
"""
outputs = self.text_encoder(**tokens)
return self.projection(outputs.pooler_output)
|