File size: 8,565 Bytes
f291f4a
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
from torch import nn


class SingleVisualizationModel(nn.Module):
    def __init__(self, input_dims, output_dims, units, hidden_layer=3):
        super(SingleVisualizationModel, self).__init__()

        self.input_dims = input_dims
        self.output_dims = output_dims
        self.units = units
        self.hidden_layer = hidden_layer
        self._init_autoencoder()
    
    # TODO find the best model architecture
    def _init_autoencoder(self):
        self.encoder = nn.Sequential(
            nn.Linear(self.input_dims, self.units),
            nn.ReLU(True))
        for h in range(self.hidden_layer):
            self.encoder.add_module("{}".format(2*h+2), nn.Linear(self.units, self.units))
            self.encoder.add_module("{}".format(2*h+3), nn.ReLU(True))
        self.encoder.add_module("{}".format(2*(self.hidden_layer+1)), nn.Linear(self.units, self.output_dims))

        self.decoder = nn.Sequential(
            nn.Linear(self.output_dims, self.units),
            nn.ReLU(True))
        for h in range(self.hidden_layer):
            self.decoder.add_module("{}".format(2*h+2), nn.Linear(self.units, self.units))
            self.decoder.add_module("{}".format(2*h+3), nn.ReLU(True))
        self.decoder.add_module("{}".format(2*(self.hidden_layer+1)), nn.Linear(self.units, self.input_dims))

    def forward(self, edge_to, edge_from):
        outputs = dict()
        embedding_to = self.encoder(edge_to)
        embedding_from = self.encoder(edge_from)
        recon_to = self.decoder(embedding_to)
        recon_from = self.decoder(embedding_from)
        
        outputs["umap"] = (embedding_to, embedding_from)
        outputs["recon"] = (recon_to, recon_from)

        return outputs

class VisModel(nn.Module):
    """define you own visualizatio model by specifying the structure

    """
    def __init__(self, encoder_dims, decoder_dims):
        """define you own visualizatio model by specifying the structure

        Parameters
        ----------
        encoder_dims : list of int
            the neuron number of your encoder
            for example, [100,50,2], denote two fully connect layers, with shape (100,50) and (50,2)
        decoder_dims : list of int
            same as encoder_dims
        """
        super(VisModel, self).__init__()
        assert len(encoder_dims) > 1
        assert len(decoder_dims) > 1
        self.encoder_dims = encoder_dims
        self.decoder_dims = decoder_dims
        self._init_autoencoder()
    
    def _init_autoencoder(self):
        self.encoder = nn.Sequential()
        for i in range(0, len(self.encoder_dims)-2):
            self.encoder.add_module("{}".format(len(self.encoder)), nn.Linear(self.encoder_dims[i], self.encoder_dims[i+1]))
            self.encoder.add_module("{}".format(len(self.encoder)), nn.ReLU(True))
        self.encoder.add_module("{}".format(len(self.encoder)), nn.Linear(self.encoder_dims[-2], self.encoder_dims[-1]))
        
        self.decoder = nn.Sequential()
        for i in range(0, len(self.decoder_dims)-2):
            self.decoder.add_module("{}".format(len(self.decoder)), nn.Linear(self.decoder_dims[i], self.decoder_dims[i+1]))
            self.decoder.add_module("{}".format(len(self.decoder)), nn.ReLU(True))
        self.decoder.add_module("{}".format(len(self.decoder)), nn.Linear(self.decoder_dims[-2], self.decoder_dims[-1]))


    def forward(self, edge_to, edge_from):
        outputs = dict()
        embedding_to = self.encoder(edge_to)
        embedding_from = self.encoder(edge_from)
        recon_to = self.decoder(embedding_to)
        recon_from = self.decoder(embedding_from)
        
        outputs["umap"] = (embedding_to, embedding_from)
        outputs["recon"] = (recon_to, recon_from)

        return outputs


'''
The visualization model definition class
'''
import tensorflow as tf
from tensorflow import keras
class tfModel(keras.Model):
    def __init__(self, optimizer, loss, loss_weights, encoder_dims, decoder_dims, batch_size, withoutB=True, attention=True, prev_trainable_variables=None):

        super(tfModel, self).__init__()
        self._init_autoencoder(encoder_dims, decoder_dims)
        self.optimizer = optimizer  # optimizer
        self.withoutB = withoutB
        self.attention = attention

        self.loss = loss  # dict of 3 losses {"total", "umap", "reconstrunction", "regularization"}
        self.loss_weights = loss_weights  # weights for each loss (in total 3 losses)

        self.prev_trainable_variables = prev_trainable_variables  # weights for previous iteration
        self.batch_size = batch_size
    
    def _init_autoencoder(self, encoder_dims, decoder_dims):
        self.encoder = tf.keras.Sequential([
            tf.keras.layers.InputLayer(input_shape=(encoder_dims[0],)),
            tf.keras.layers.Flatten(),
        ])
        for i in range(1, len(encoder_dims)-1, 1):
            self.encoder.add(tf.keras.layers.Dense(units=encoder_dims[i], activation="relu"))
        self.encoder.add(tf.keras.layers.Dense(units=encoder_dims[-1]),)

        self.decoder = tf.keras.Sequential([
            tf.keras.layers.InputLayer(input_shape=(decoder_dims[0],)),
        ])
        for i in range(1, len(decoder_dims)-1, 1):
            self.decoder.add(tf.keras.layers.Dense(units=decoder_dims[i], activation="relu"))
        self.decoder.add(tf.keras.layers.Dense(units=decoder_dims[-1]))
        print(self.encoder.summary())
        print(self.decoder.summary())

    def train_step(self, x):

        to_x, from_x, to_alpha, from_alpha, n_rate, weight = x[0]
        to_x = tf.cast(to_x, dtype=tf.float32)
        from_x = tf.cast(from_x, dtype=tf.float32)
        to_alpha = tf.cast(to_alpha, dtype=tf.float32)
        from_alpha = tf.cast(from_alpha, dtype=tf.float32)
        n_rate = tf.cast(n_rate, dtype=tf.float32)
        weight = tf.cast(weight, dtype=tf.float32)

        # Forward pass
        with tf.GradientTape(persistent=True) as tape:

            # parametric embedding
            embedding_to = self.encoder(to_x)  # embedding for instance 1
            embedding_from = self.encoder(from_x)  # embedding for instance 1
            embedding_to_recon = self.decoder(embedding_to)  # reconstruct instance 1
            embedding_from_recon = self.decoder(embedding_from)  # reconstruct instance 1

            # concatenate embedding1 and embedding2 to prepare for umap loss
            embedding_to_from = tf.concat((embedding_to, embedding_from, weight),
                                          axis=1)
            # reconstruction loss
            if self.attention:
                reconstruct_loss = self.loss["reconstruction"](to_x, from_x, embedding_to_recon, embedding_from_recon,to_alpha, from_alpha)
            else:
                self.loss["reconstruction"] = tf.keras.losses.MeanSquaredError()
                reconstruct_loss = self.loss["reconstruction"](y_true=to_x, y_pred=embedding_to_recon)/2 + self.loss["reconstruction"](y_true=from_x, y_pred=embedding_from_recon)/2

            # umap loss
            umap_loss = self.loss["umap"](None, embed_to_from=embedding_to_from)  # w_(t-1), no gradient

            # compute alpha bar
            alpha_mean = tf.cast(tf.reduce_mean(tf.stop_gradient(n_rate)), dtype=tf.float32)
            # L2 norm of w current - w for last epoch (subject model's epoch)
            # dummy zeros-loss if no previous epoch
            if self.prev_trainable_variables is None:
                prev_trainable_variables = [tf.stop_gradient(x) for x in self.trainable_variables]
            else:
                prev_trainable_variables = self.prev_trainable_variables
            regularization_loss = self.loss["regularization"](w_prev=prev_trainable_variables,w_current=self.trainable_variables, to_alpha=alpha_mean)

                # aggregate loss, weighted average
            loss = tf.add(tf.add(tf.math.multiply(tf.constant(self.loss_weights["reconstruction"]), reconstruct_loss),
                                    tf.math.multiply(tf.constant(self.loss_weights["umap"]), umap_loss)),
                            tf.math.multiply(tf.constant(self.loss_weights["regularization"]), regularization_loss))

        # Compute gradients
        trainable_vars = self.trainable_variables
        grads = tape.gradient(loss, trainable_vars)

        # Update weights
        self.optimizer.apply_gradients(zip(grads, trainable_vars))

        return {"loss": loss, "umap": umap_loss, "reconstruction": reconstruct_loss,
                "regularization": regularization_loss}