hyperMM/HyperMM.py at main · robustml-eurecom/hyperMM · GitHub

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
import numpy as np

import torch
import torch.nn as nn
from torchvision import models

class HyperMMPretrain(nn.Module):
    def __init__(self):
        super().__init__()
        self.emb_dim = 5

        self.pretrained_model = models.vgg11(pretrained=True)
        for param in self.pretrained_model.parameters():
            param.requires_grad = False

        self.pretrained_model.eval()

        self.encoder = nn.Sequential(
            nn.Conv2d(512, 512, kernel_size=1)
        )
        self.decoder = nn.ConvTranspose2d(512, 512, kernel_size=1)
        self.classifier = nn.Sequential(nn.Linear(512 * 7 * 7, 512), nn.ReLU(inplace=True), nn.Linear(512, 1))

        self.total_params = self.encoder[-1].weight.nelement() + self.encoder[-1].bias.nelement()
        self.embedding = nn.Embedding(20, self.emb_dim)

        self.hypernet = nn.Sequential(
            nn.Linear(self.emb_dim, 20),
            nn.ReLU(),
            nn.Linear(20, self.total_params)
        )

    def sample_weights(self, c):
        weights = self.hypernet(self.embedding(c))
        weights = weights.reshape([-1])

        next_idx = 0
        layer = self.encoder[-1]
        if isinstance(layer, nn.Conv2d):
            # Get the current and next layer's neuron counts for the splice
            cur_idx = next_idx
            weight_size = np.prod(layer.weight.shape)
            bias_size = np.prod(layer.bias.shape)
            next_idx += weight_size + bias_size

            weight_splice = weights[cur_idx : cur_idx + weight_size].reshape(layer.weight.shape)
            bias_splice = weights[cur_idx + weight_size : next_idx].reshape(layer.bias.shape)

            # Copy over the generated weights into the parameters of the dynamics network
            # Note that this delete is important to properly establish the computation graph link
            del layer.weight
            layer.weight = weight_splice
            del layer.bias
            layer.bias = bias_splice


    def forward(self, x, c):
        self.sample_weights(c)
        x = torch.squeeze(x, dim=0)

        input_features = self.pretrained_model.features(x)
        pooled_features = torch.max(input_features, 0, keepdim=True)[0]

        features = self.encoder(pooled_features)
        output_features = self.decoder(features)
        output_classif = self.classifier(torch.flatten(features, 1))

        return features, pooled_features, output_features, output_classif


class HyperMMNet(nn.Module):
    def __init__(self, feature_extractor):
        super().__init__()

        self.phi = feature_extractor
        for param in self.phi.parameters():
            param.requires_grad = False

        self.rho = nn.Sequential(
            nn.Linear(512 * 7 * 7, 512),
            nn.ReLU(),
            nn.Dropout(0.5),
            nn.Linear(512, 1),
        )

    def forward(self, X):
        feats = []
        for x in X:
            img = torch.squeeze(x[0], dim=0)
            c = x[1]
            features,_,_,_ = self.phi(img, c)
            feats.append(torch.flatten(features, 1))

        latent = torch.stack(tuple(feats), 0)

        sum_latent = latent.mean(dim=0)
        output = self.rho(sum_latent)
        return output

"""
class HyperMMNet(nn.Module):
    def __init__(self, feature_extractor):
        super().__init__()

        self.feature_extractor = feature_extractor
        for param in self.feature_extractor.parameters():
            param.requires_grad = False

        self.phi = nn.Sequential(
            nn.Linear(256 * 6 * 6, 256 * 6 * 6),
            #nn.ReLU(inplace=True),
            #nn.Linear(256 * 6 * 6, 256 * 6 * 6),
            nn.ReLU(inplace=True)
        )

        self.rho = nn.Sequential(
            nn.Linear(256 * 6 * 6, 512),
            nn.ReLU(),
            nn.Dropout(0.2),
            nn.Linear(512, 256),
            nn.ReLU(),
            nn.Linear(256, 128),
            nn.ReLU(),
            nn.Linear(128, 1),
        )

    def forward(self, X):
        feats = []
        for x in X:
            img = torch.squeeze(x[0], dim=0)
            c = x[1]
            features,_,_ = self.feature_extractor(img, c)
            feats.append(torch.flatten(features, 1))

        feats = torch.cat(tuple(feats), 0)

        latent = self.phi(feats)
        sum_latent = latent.mean(dim=0)
        output = self.rho(sum_latent)
        return output
"""