Add custom model definitions (model_definitions.py)
Browse files- model_definitions.py +137 -0
model_definitions.py
ADDED
@@ -0,0 +1,137 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import torch
|
2 |
+
import torch.nn as nn
|
3 |
+
from torchvision.models import resnet50, ResNet50_Weights
|
4 |
+
import torch.nn.functional as F
|
5 |
+
from huggingface_hub import PyTorchModelHubMixin # Import the mixin
|
6 |
+
|
7 |
+
# --- Custom Model Definitions ---
|
8 |
+
|
9 |
+
class Identity(nn.Module):
|
10 |
+
def forward(self, x: torch.Tensor) -> torch.Tensor:
|
11 |
+
return x
|
12 |
+
|
13 |
+
class AdditiveAttention(nn.Module):
|
14 |
+
def __init__(self, d_model: int, hidden_dim: int = 128):
|
15 |
+
super(AdditiveAttention, self).__init__()
|
16 |
+
self.query_projection = nn.Linear(d_model, hidden_dim)
|
17 |
+
self.key_projection = nn.Linear(d_model, hidden_dim)
|
18 |
+
self.value_projection = nn.Linear(d_model, hidden_dim)
|
19 |
+
self.attention_mechanism = nn.Linear(hidden_dim, hidden_dim) # Output hidden_dim
|
20 |
+
|
21 |
+
def forward(self, query: torch.Tensor) -> torch.Tensor:
|
22 |
+
keys = self.key_projection(query)
|
23 |
+
values = self.value_projection(query)
|
24 |
+
queries = self.query_projection(query)
|
25 |
+
|
26 |
+
attention_scores = torch.tanh(queries + keys)
|
27 |
+
attention_weights = F.softmax(self.attention_mechanism(attention_scores), dim=1)
|
28 |
+
|
29 |
+
attended_values = values * attention_weights # Element-wise product
|
30 |
+
return attended_values
|
31 |
+
|
32 |
+
class ResNet50Custom(nn.Module, PyTorchModelHubMixin): # Inherit from PyTorchModelHubMixin
|
33 |
+
def __init__(self, input_channels: int, num_classes: int, **kwargs):
|
34 |
+
super(ResNet50Custom, self).__init__()
|
35 |
+
|
36 |
+
# Store config for PyTorchModelHubMixin to serialize to config.json
|
37 |
+
self.config = {
|
38 |
+
"input_channels": input_channels,
|
39 |
+
"num_classes": num_classes,
|
40 |
+
**kwargs
|
41 |
+
}
|
42 |
+
|
43 |
+
self.input_channels = input_channels
|
44 |
+
|
45 |
+
self.model = resnet50(weights=ResNet50_Weights.IMAGENET1K_V1)
|
46 |
+
|
47 |
+
self.model.conv1 = nn.Conv2d(input_channels, 64, kernel_size=7, stride=2, padding=3, bias=False)
|
48 |
+
|
49 |
+
# The final FC layer of ResNet50Custom will be used *only* when ResNet50Custom is a standalone classifier.
|
50 |
+
# When used as a feature extractor within MultiModalModel, this layer will be temporarily replaced by Identity().
|
51 |
+
self.model.fc = nn.Linear(self.model.fc.in_features, num_classes)
|
52 |
+
|
53 |
+
def forward(self, x: torch.Tensor) -> torch.Tensor:
|
54 |
+
return self.model(x)
|
55 |
+
|
56 |
+
def get_feature_size(self) -> int:
|
57 |
+
return self.model.fc.in_features
|
58 |
+
|
59 |
+
|
60 |
+
class MultiModalModel(nn.Module, PyTorchModelHubMixin): # Inherit from PyTorchModelHubMixin
|
61 |
+
def __init__(self,
|
62 |
+
image_input_channels: int,
|
63 |
+
bathy_input_channels: int,
|
64 |
+
sss_input_channels: int,
|
65 |
+
num_classes: int,
|
66 |
+
attention_type: str = "scaled_dot_product",
|
67 |
+
**kwargs): # Added **kwargs for mixin compatibility
|
68 |
+
super(MultiModalModel, self).__init__()
|
69 |
+
|
70 |
+
# Store config for PyTorchModelHubMixin to serialize to config.json
|
71 |
+
self.config = {
|
72 |
+
"image_input_channels": image_input_channels,
|
73 |
+
"bathy_input_channels": bathy_input_channels,
|
74 |
+
"sss_input_channels": sss_input_channels,
|
75 |
+
"num_classes": num_classes,
|
76 |
+
"attention_type": attention_type,
|
77 |
+
**kwargs # Pass along any extra kwargs for mixin
|
78 |
+
}
|
79 |
+
|
80 |
+
# Instantiate feature extraction models *inside* MultiModalModel
|
81 |
+
# Their final FC layers will be treated as Identity for feature extraction
|
82 |
+
self.image_model_feat = ResNet50Custom(input_channels=image_input_channels, num_classes=num_classes)
|
83 |
+
self.bathy_model_feat = ResNet50Custom(input_channels=bathy_input_channels, num_classes=num_classes)
|
84 |
+
self.sss_model_feat = ResNet50Custom(input_channels=sss_input_channels, num_classes=num_classes)
|
85 |
+
|
86 |
+
# The ResNet50's feature output size is 2048 before its final FC layer
|
87 |
+
feature_dim = self.image_model_feat.get_feature_size() # Should be 2048
|
88 |
+
|
89 |
+
# Attention layers (AdditiveAttention uses d_model and outputs hidden_dim)
|
90 |
+
attention_hidden_dim = 128 # This matches your fc layer input calculation (3*128=384)
|
91 |
+
self.attention_image = AdditiveAttention(feature_dim, hidden_dim=attention_hidden_dim)
|
92 |
+
self.attention_bathy = AdditiveAttention(feature_dim, hidden_dim=attention_hidden_dim)
|
93 |
+
self.attention_sss = AdditiveAttention(feature_dim, hidden_dim=attention_hidden_dim)
|
94 |
+
|
95 |
+
# Final classification layers
|
96 |
+
self.fc = nn.Linear(3 * attention_hidden_dim, 1284)
|
97 |
+
self.fc1 = nn.Linear(1284, 32)
|
98 |
+
# Ensure num_classes is int for the linear layer
|
99 |
+
num_classes_int = int(num_classes)
|
100 |
+
if not isinstance(num_classes_int, int):
|
101 |
+
raise TypeError("num_classes must be an integer after casting")
|
102 |
+
self.fc2 = nn.Linear(32, num_classes_int)
|
103 |
+
self.attention_type = attention_type
|
104 |
+
|
105 |
+
def forward(self, inputs: torch.Tensor, bathy_tensor: torch.Tensor, sss_image: torch.Tensor) -> torch.Tensor:
|
106 |
+
# Temporarily replace the final FC layer of the feature extractors with Identity
|
107 |
+
# to get the 2048 features, then restore them.
|
108 |
+
original_image_fc = self.image_model_feat.model.fc
|
109 |
+
original_bathy_fc = self.bathy_model_feat.model.fc
|
110 |
+
original_sss_fc = self.sss_model_feat.model.fc
|
111 |
+
|
112 |
+
self.image_model_feat.model.fc = Identity()
|
113 |
+
self.bathy_model_feat.model.fc = Identity()
|
114 |
+
self.sss_model_feat.model.fc = Identity()
|
115 |
+
|
116 |
+
image_features = self.image_model_feat(inputs)
|
117 |
+
bathy_features = self.bathy_model_feat(bathy_tensor)
|
118 |
+
sss_features = self.sss_model_feat(sss_image)
|
119 |
+
|
120 |
+
# Restore original FC layers on the feature extractors
|
121 |
+
self.image_model_feat.model.fc = original_image_fc
|
122 |
+
self.bathy_model_feat.model.fc = original_bathy_fc
|
123 |
+
self.sss_model_feat.model.fc = original_sss_fc
|
124 |
+
|
125 |
+
# Apply attention
|
126 |
+
image_features_attended = self.attention_image(image_features)
|
127 |
+
bathy_features_attended = self.attention_bathy(bathy_features)
|
128 |
+
sss_features_attended = self.attention_sss(sss_features)
|
129 |
+
|
130 |
+
# Concatenate attended features
|
131 |
+
combined_features = torch.cat([image_features_attended, bathy_features_attended, sss_features_attended], dim=1)
|
132 |
+
|
133 |
+
# Pass through final classification layers
|
134 |
+
outputs_1 = self.fc(combined_features)
|
135 |
+
output_2 = self.fc1(outputs_1)
|
136 |
+
outputs = self.fc2(output_2)
|
137 |
+
return outputs
|