Upload 3 files
Browse files- image_encoder.py +22 -0
- projection_head.py +25 -0
- text_encoder.py +22 -0
image_encoder.py
ADDED
@@ -0,0 +1,22 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
from torch import nn
|
2 |
+
import timm
|
3 |
+
|
4 |
+
from configuration import CFG
|
5 |
+
|
6 |
+
class ImageEncoder(nn.Module):
|
7 |
+
"""
|
8 |
+
Encode images to a fixed size vector
|
9 |
+
"""
|
10 |
+
|
11 |
+
def __init__(
|
12 |
+
self, model_name=CFG.model_name, pretrained=CFG.pretrained, trainable=CFG.trainable
|
13 |
+
):
|
14 |
+
super().__init__()
|
15 |
+
self.model = timm.create_model(
|
16 |
+
model_name, pretrained, num_classes=0, global_pool="avg"
|
17 |
+
)
|
18 |
+
for p in self.model.parameters():
|
19 |
+
p.requires_grad = trainable
|
20 |
+
|
21 |
+
def forward(self, x):
|
22 |
+
return self.model(x)
|
projection_head.py
ADDED
@@ -0,0 +1,25 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
from torch import nn
|
2 |
+
from configuration import CFG
|
3 |
+
|
4 |
+
class ProjectionHead(nn.Module):
|
5 |
+
def __init__(
|
6 |
+
self,
|
7 |
+
embedding_dim,
|
8 |
+
projection_dim=CFG.projection_dim,
|
9 |
+
dropout=CFG.dropout
|
10 |
+
):
|
11 |
+
super().__init__()
|
12 |
+
self.projection = nn.Linear(embedding_dim, projection_dim)
|
13 |
+
self.gelu = nn.GELU()
|
14 |
+
self.fc = nn.Linear(projection_dim, projection_dim)
|
15 |
+
self.dropout = nn.Dropout(dropout)
|
16 |
+
self.layer_norm = nn.LayerNorm(projection_dim)
|
17 |
+
|
18 |
+
def forward(self, x):
|
19 |
+
projected = self.projection(x)
|
20 |
+
x = self.gelu(projected)
|
21 |
+
x = self.fc(x)
|
22 |
+
x = self.dropout(x)
|
23 |
+
x = x + projected
|
24 |
+
x = self.layer_norm(x)
|
25 |
+
return x
|
text_encoder.py
ADDED
@@ -0,0 +1,22 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
from torch import nn
|
2 |
+
from configuration import CFG
|
3 |
+
from transformers import DistilBertModel, DistilBertConfig, DistilBertTokenizer
|
4 |
+
|
5 |
+
class TextEncoder(nn.Module):
|
6 |
+
def __init__(self, model_name=CFG.text_encoder_model, pretrained=CFG.pretrained, trainable=CFG.trainable):
|
7 |
+
super().__init__()
|
8 |
+
if pretrained:
|
9 |
+
self.model = DistilBertModel.from_pretrained(model_name)
|
10 |
+
else:
|
11 |
+
self.model = DistilBertModel(config=DistilBertConfig())
|
12 |
+
|
13 |
+
for p in self.model.parameters():
|
14 |
+
p.requires_grad = trainable
|
15 |
+
|
16 |
+
# we are using the CLS token hidden representation as the sentence's embedding
|
17 |
+
self.target_token_idx = 0
|
18 |
+
|
19 |
+
def forward(self, input_ids, attention_mask):
|
20 |
+
output = self.model(input_ids=input_ids, attention_mask=attention_mask)
|
21 |
+
last_hidden_state = output.last_hidden_state
|
22 |
+
return last_hidden_state[:, self.target_token_idx, :]
|