-
Notifications
You must be signed in to change notification settings - Fork 2
/
multi_model_inference.py
217 lines (167 loc) · 8.06 KB
/
multi_model_inference.py
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
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
from torch.utils.data import DataLoader, Dataset, random_split
from torchvision.transforms import transforms
import torch.nn as nn
import torch
from transformers import DistilBertModel
from PIL import Image
from transformers import DistilBertTokenizer
from torchvision import models
import torch.nn.functional as F
from io import BytesIO
import logging
import ssl
ssl._create_default_https_context = ssl._create_unverified_context
# We will first prepare the dataset for inference
class InferenceDataset(Dataset):
def __init__(self, dataframe, tokenizer, max_len, transform, client, bucket_name):
self.dataframe = dataframe
self.tokenizer = tokenizer
self.max_len = max_len
self.transform = transform
self.minio_client = client
self.bucket_name = bucket_name
def __len__(self):
return len(self.dataframe)
def __getitem__(self, idx):
row = self.dataframe.iloc[idx]
text = row['title']
img_path = row['image_path']
# Processing text
encoding = self.tokenizer.encode_plus(
text,
add_special_tokens=True,
max_length=self.max_len,
return_token_type_ids=False,
padding='max_length',
return_attention_mask=True,
return_tensors='pt',
truncation=True
)
bucket, _, object_name = img_path.partition('/')
if self.minio_client:
image_data = self.minio_client.get_obj(self.bucket_name, object_name)
image_bytes = image_data.read()
image = Image.open(BytesIO(image_bytes)).convert("RGB")
else:
image = Image.open(img_path).convert("RGB")
if self.transform:
image = self.transform(image)
return {
'input_ids': encoding['input_ids'].flatten(),
'attention_mask': encoding['attention_mask'].flatten(),
'image': image
}
text = DistilBertModel.from_pretrained('distilbert-base-uncased')
img = models.efficientnet_v2_m(pretrained=True)
class MultiHeadCrossAttention(nn.Module):
def __init__(self, text_dim, image_dim, num_heads, hidden_dim, output_dim):
super(MultiHeadCrossAttention, self).__init__()
self.num_heads = num_heads
self.head_dim = hidden_dim // num_heads
# These linear layers project the inputs to multiple heads
self.text_query = nn.Linear(text_dim, hidden_dim, bias=False)
self.text_key = nn.Linear(text_dim, hidden_dim, bias=False)
self.text_value = nn.Linear(text_dim, hidden_dim, bias=False)
self.image_query = nn.Linear(image_dim, hidden_dim, bias=False)
self.image_key = nn.Linear(image_dim, hidden_dim, bias=False)
self.image_value = nn.Linear(image_dim, hidden_dim, bias=False)
# Final projection layer
self.out_proj = nn.Linear(hidden_dim, output_dim, bias=False)
def forward(self, text_features, image_features):
Q_text = self.text_query(text_features)
K_text = self.text_key(text_features)
V_text = self.text_value(text_features)
Q_image = self.image_query(image_features)
K_image = self.image_key(image_features)
V_image = self.image_value(image_features)
# Split the hidden dimension into num_heads
Q_text = Q_text.view(Q_text.size(0), -1, self.num_heads, self.head_dim).transpose(1, 2)
K_text = K_text.view(K_text.size(0), -1, self.num_heads, self.head_dim).transpose(1, 2)
V_text = V_text.view(V_text.size(0), -1, self.num_heads, self.head_dim).transpose(1, 2)
Q_image = Q_image.view(Q_image.size(0), -1, self.num_heads, self.head_dim).transpose(1, 2)
K_image = K_image.view(K_image.size(0), -1, self.num_heads, self.head_dim).transpose(1, 2)
V_image = V_image.view(V_image.size(0), -1, self.num_heads, self.head_dim).transpose(1, 2)
# Calculate the attention scores
attn_scores_text_image = torch.matmul(Q_text, K_image.transpose(-1, -2)) / (self.head_dim ** 0.5)
attn_scores_image_text = torch.matmul(Q_image, K_text.transpose(-1, -2)) / (self.head_dim ** 0.5)
# Normalize scores
attn_probs_text_image = F.softmax(attn_scores_text_image, dim=-1)
attn_probs_image_text = F.softmax(attn_scores_image_text, dim=-1)
# Apply attention
attn_output_text_image = torch.matmul(attn_probs_text_image, V_image)
attn_output_image_text = torch.matmul(attn_probs_image_text, V_text)
# Concatenate the results across the heads
attn_output_text_image = attn_output_text_image.transpose(1, 2).contiguous().view(text_features.size(0), -1)
attn_output_image_text = attn_output_image_text.transpose(1, 2).contiguous().view(image_features.size(0), -1)
# Project to output dimension
output_text_image = self.out_proj(attn_output_text_image)
output_image_text = self.out_proj(attn_output_image_text)
return output_text_image, output_image_text
class MultiModalModel(nn.Module):
def __init__(self, num_labels):
super(MultiModalModel, self).__init__()
# Load pre-trained models
self.bert = text
self.resnet = img
# Remove the final classification layer of ResNet
self.resnet = nn.Sequential(*list(self.resnet.children())[:-1])
self.mhca = MultiHeadCrossAttention(text_dim=768, image_dim=1280, num_heads=4, hidden_dim=512, output_dim=2048)
self.classifier = nn.Sequential(
nn.Linear(2816, 512),
nn.ReLU(),
nn.Dropout(0.2),
nn.Linear(512, num_labels)
)
def forward(self, input_ids, attention_mask, image, device):
# Forward pass through BERT
outputs = self.bert(input_ids=input_ids, attention_mask=attention_mask)
text_features = outputs['last_hidden_state'][:, 0, :] # CLS token output as text feature
# Forward pass through ResNet
image_features = self.resnet(image)
image_features = image_features.view(image_features.size(0), -1) # Flatten the output
if text_features.dim() == 2:
text_features = text_features.unsqueeze(1)
if image_features.dim() == 2:
image_features = image_features.unsqueeze(1)
attended_text, attended_image = self.mhca(text_features, image_features)
attended_text = attended_text.squeeze(1) # shape: [16, 768]
attended_image = attended_image.squeeze(1) # shape: [16, 2048]
self.image_projection = torch.nn.Linear(2048, 768).to(device)
attended_image = self.image_projection(attended_image)
combined_features = torch.cat((attended_text, attended_image), dim=-1)
logits = self.classifier(combined_features)
return logits
# Function to run inference
def run_inference(model, dataloader, device):
model.eval()
predictions = []
with torch.no_grad():
for batch in dataloader:
input_ids = batch['input_ids'].to(device)
attention_mask = batch['attention_mask'].to(device)
images = batch['image'].to(device)
outputs = model(input_ids=input_ids, attention_mask=attention_mask, image=images, device=device)
preds = torch.argmax(outputs, dim=1)
predictions.extend(preds.cpu().numpy())
return predictions
def get_inference_data(dataset, minio_client, bucket_name):
MAX_LEN = 128
BATCH_SIZE = 16
tokenizer = DistilBertTokenizer.from_pretrained('distilbert-base-uncased')
transform = transforms.Compose([
transforms.RandomHorizontalFlip(),
transforms.RandomRotation(10),
transforms.RandomResizedCrop(224, scale=(0.8, 1.0)),
transforms.ColorJitter(brightness=0.2, contrast=0.2, saturation=0.2, hue=0.1),
transforms.ToTensor(),
])
# Prepare dataset and dataloader for inference
inference_dataset = InferenceDataset(
dataframe=dataset,
tokenizer=tokenizer,
max_len=MAX_LEN,
transform=transform,
client=minio_client,
bucket_name = bucket_name
)
return DataLoader(inference_dataset, batch_size=BATCH_SIZE, shuffle=False)