Мой проект представляет собой мультимодальное распознавание именованных объектов. Я пытаюсь создать мультимодальный BERT. Сначала я создаю векторы изображений с помощью ResNET и векторы текста с помощью BERT. Затем он индивидуально передается в мультимодальный BERT. (Внутри BERT имеется механизм многоголового внимания, перекрестного внимания и прямой связи). Затем эта модель должна генерировать последовательность векторов, комбинацию векторов текста и изображения. Я также хочу, чтобы эта модель генерировала метки. ), а затем эта последовательность векторов затем передается в Bi-LSTM только для контекстного улучшения, а затем в CRF для генерации меток.
Я написал код для мультимодального BERT и получаю ошибку
import os
import re
import random
import torch
import torch.nn as nn
from torch.utils.data import Dataset, DataLoader
from torchvision import transforms
from PIL import Image
from transformers import BertModel, BertTokenizer, AdamW
from TorchCRF import CRF
# Prepare the BERT tokenizer and image preprocessing
tokenizer = BertTokenizer.from_pretrained('bert-base-uncased')
# Preprocess images for ResNet
preprocess_image = transforms.Compose([
transforms.Resize(256),
transforms.CenterCrop(224),
transforms.ToTensor(),
transforms.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]),
])
# Function to load and preprocess an image
def load_image(image_path):
image = Image.open(image_path).convert('RGB')
image = preprocess_image(image).unsqueeze(0) # Add batch dimension
return image
# Function to parse the text file
def parse_text_file(file_path):
data = {}
current_imgid = None
accumulated_sentence = []
accumulated_labels = []
with open(file_path, 'r', encoding='utf-8') as f:
for line in f:
line = line.strip()
if line.startswith("IMGID:"):
if current_imgid and accumulated_sentence:
# Save previous IMGID's sentence and labels
data[current_imgid] = {
'sentence': " ".join(accumulated_sentence),
'labels': accumulated_labels
}
# Start processing new IMGID
current_imgid = line.split(':')[1].strip()
accumulated_sentence = []
accumulated_labels = []
elif line: # Non-empty line (word-label pairs)
word, label = line.split()
if not re.match(r'http?://\S+', word): # Ignore URLs
accumulated_sentence.append(word)
accumulated_labels.append(label)
# Process last IMGID
if current_imgid and accumulated_sentence:
data[current_imgid] = {
'sentence': " ".join(accumulated_sentence),
'labels': accumulated_labels
}
return data
# Dataset class combining images and text
class MultimodalDataset(Dataset):
def __init__(self, image_dir, text_file):
self.image_dir = image_dir
self.data = parse_text_file(text_file)
self.tokenizer = BertTokenizer.from_pretrained('bert-base-uncased')
def __len__(self):
return len(self.data)
def __getitem__(self, idx):
# Get image ID
imgid = list(self.data.keys())[idx]
img_path = os.path.join(self.image_dir, f"{imgid}.jpg")
# Load and preprocess the image
image = load_image(img_path)
# Get text and labels
text_data = self.data[imgid]
sentence = text_data['sentence']
labels = text_data['labels']
# Tokenize text and convert to BERT format
inputs = self.tokenizer(sentence, return_tensors='pt', padding='max_length', truncation=True, max_length=128)
# Pad labels to the same length as inputs
padded_labels = [labels if i < len(labels) else "O" for i in range(128)] # Replace with "O" for padding
padded_labels = padded_labels[:128] # Truncate if longer
padded_labels += ["O"] * (128 - len(padded_labels)) # Pad with "O"
return {
'input_ids': inputs['input_ids'].squeeze(0),
'attention_mask': inputs['attention_mask'].squeeze(0),
'image': image.squeeze(0),
'labels': padded_labels # Use padded labels
}
# Function to randomly mask tokens in the input text
def mask_tokens(inputs, tokenizer, mask_prob=0.15):
labels = inputs.clone()
rand = torch.rand(inputs.shape)
mask_arr = (rand < mask_prob) & (inputs != tokenizer.cls_token_id) & \
(inputs != tokenizer.sep_token_id) & (inputs != tokenizer.pad_token_id)
inputs[mask_arr] = tokenizer.mask_token_id
labels[~mask_arr] = -100 # Only compute loss on masked tokens
return inputs, labels
# Function to randomly mask regions of the image
def mask_image(image, mask_prob=0.15):
masked_image = image.clone()
rand = torch.rand(image.shape)
mask_arr = rand < mask_prob
masked_image[mask_arr] = 0 # Masked regions set to 0
return masked_image
# Multimodal BERT Model
class MultimodalBERT(nn.Module):
def __init__(self, hidden_dim=768, num_labels=9): # num_labels = number of NER tags
super(MultimodalBERT, self).__init__()
# Image and text embedding dimensions (768 for BERT)
self.image_linear = nn.Linear(768, hidden_dim) # Map image embeddings to the same size as BERT embeddings
self.text_bert = BertModel.from_pretrained('bert-base-uncased') # BERT model for text embeddings
# Attention mechanism for cross-modal interaction
self.multihead_attention = nn.MultiheadAttention(hidden_dim, num_heads=8, dropout=0.1)
# Feed-forward neural network after attention
self.feed_forward = nn.Sequential(
nn.Linear(hidden_dim, hidden_dim),
nn.ReLU(),
nn.Linear(hidden_dim, hidden_dim)
)
# Linear layer to map to the number of NER labels
self.classifier = nn.Linear(hidden_dim, num_labels) # Map hidden_dim to num_labels (e.g., 9 NER tags)
# CRF layer for named entity recognition
self.crf = CRF(num_labels, batch_first=True)
def forward(self, text_inputs, text_attention_mask, image_features, labels=None, mlm_labels=None):
# Process text using BERT
text_outputs = self.text_bert(input_ids=text_inputs, attention_mask=text_attention_mask)
text_embeddings = text_outputs.last_hidden_state # [batch_size, seq_len, hidden_dim]
# Process image features
image_embeddings = self.image_linear(image_features) # [batch_size, num_regions (49), hidden_dim]
# Multimodal cross-attention between image and text
multimodal_output, _ = self.multihead_attention(text_embeddings.permute(1, 0, 2), # Query: text embeddings
image_embeddings.permute(1, 0, 2), # Key: image embeddings
image_embeddings.permute(1, 0, 2)) # Value: image embeddings
multimodal_output = multimodal_output.permute(1, 0, 2) # Reshape back to [batch_size, seq_len, hidden_dim]
# Feed forward layer
output = self.feed_forward(multimodal_output)
# Classify each token for NER
emissions = self.classifier(output) # [batch_size, seq_len, num_labels]
# Predict NER labels using CRF
if labels is not None:
log_likelihood = self.crf(emissions, labels, mask=text_attention_mask.bool())
return -log_likelihood
else:
predictions = self.crf.decode(emissions, mask=text_attention_mask.bool())
return predictions
# Training Function
def train_model(model, dataloader, optimizer, num_epochs=10, save_path="multimodal_bert.pth"):
model.train()
for epoch in range(num_epochs):
total_loss = 0
for batch in dataloader:
text_inputs, attention_masks, images, labels = batch['input_ids'], batch['attention_mask'], batch['image'], batch['labels']
# Apply MLM masking
mlm_inputs, mlm_labels = mask_tokens(text_inputs, tokenizer)
masked_images = mask_image(images)
# Forward pass
loss = model(mlm_inputs, attention_masks, masked_images, labels=labels, mlm_labels=mlm_labels)
# Backpropagation
optimizer.zero_grad()
loss.backward()
optimizer.step()
total_loss += loss.item()
print(f'Epoch [{epoch+1}/{num_epochs}], Loss: {total_loss / len(dataloader)}')
# Save the model
torch.save(model.state_dict(), save_path)
print(f'Model saved to {save_path}')
# Function to load model
def load_model(model, load_path="multimodal_bert.pth"):
model.load_state_dict(torch.load(load_path))
model.eval()
print(f'Model loaded from {load_path}')
return model
# Main function to run training
if __name__ == "__main__":
# Define paths
image_dir = "C:/Python312/MNER_Proj/images"# Update with your image directory
text_file = "C:/Python312/MNER_Proj/images_sentences.txt" # Update with your text file path
# Create dataset and dataloader
dataset = MultimodalDataset(image_dir, text_file)
dataloader = DataLoader(dataset, batch_size=16, shuffle=True)
# Instantiate model and optimizer
model = MultimodalBERT(hidden_dim=768, num_labels=9)
optimizer = AdamW(model.parameters(), lr=5e-5)
# Train the model
train_model(model, dataloader, optimizer, num_epochs=10, save_path="multimodal_bert.pth")```
Я получаю эту ошибку:
PS C:\Python312\MNER_Proj> python main_2.py
C:\Python312\MNER_Proj\venv\Lib\site-packages\transformers\optimization.py:591: FutureWarning: This implementation of AdamW is deprecated and will be removed in a future version. Use the PyTorch implementation torch.optim.AdamW instead, or set `no_deprecation_warning=True` to disable this warning
warnings.warn(
Traceback (most recent call last):
File "C:\Python312\MNER_Proj\main_2.py", line 228, in
train_model(model, dataloader, optimizer, num_epochs=10, save_path="multimodal_bert.pth")
File "C:\Python312\MNER_Proj\main_2.py", line 191, in train_model
loss = model(mlm_inputs, attention_masks, masked_images, labels=labels, mlm_labels=mlm_labels)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "C:\Python312\MNER_Proj\venv\Lib\site-packages\torch\nn\modules\module.py", line 1553, in _wrapped_call_impl
return self._call_impl(*args, **kwargs)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "C:\Python312\MNER_Proj\venv\Lib\site-packages\torch\nn\modules\module.py", line 1562, in _call_impl
return forward_call(*args, **kwargs)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "C:\Python312\MNER_Proj\main_2.py", line 155, in forward
image_embeddings = self.image_linear(image_features) # [batch_size, num_regions (49), hidden_dim]
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "C:\Python312\MNER_Proj\venv\Lib\site-packages\torch\nn\modules\module.py", line 1553, in _wrapped_call_impl
return self._call_impl(*args, **kwargs)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "C:\Python312\MNER_Proj\venv\Lib\site-packages\torch\nn\modules\module.py", line 1562, in _call_impl
return forward_call(*args, **kwargs)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "C:\Python312\MNER_Proj\venv\Lib\site-packages\torch\nn\modules\linear.py", line 117, in forward
return F.linear(input, self.weight, self.bias)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
RuntimeError: mat1 and mat2 shapes cannot be multiplied (2688x224 and 768x768)
Подробнее здесь: https://stackoverflow.com/questions/790 ... ecognition
