File size: 10,237 Bytes
e775f6d
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
import torch; torch.manual_seed(0)
import torch.nn as nn
import torch.nn.functional as F
import torch.utils
import torch.distributions
import matplotlib.pyplot as plt; plt.rcParams['figure.dpi'] = 200

device = 'cuda' if torch.cuda.is_available() else 'cpu'

def get_activation(activation):
    if activation == 'tanh':
        activ = F.tanh
    elif activation == 'relu':
        activ = F.relu
    elif activation == 'mish':
        activ = F.mish
    elif activation == 'sigmoid':
        activ = F.sigmoid
    elif activation == 'leakyrelu':
        activ = F.leaky_relu
    elif activation == 'exp':
        activ = torch.exp
    else:
        raise ValueError
    return activ

class IngredientEncoder(nn.Module):
    def __init__(self, input_dim, deepset_latent_dim, hidden_dims, activation, dropout):
        super(IngredientEncoder, self).__init__()
        self.linears = nn.ModuleList()
        self.dropouts = nn.ModuleList()
        dims = [input_dim] + hidden_dims + [deepset_latent_dim]
        for d_in, d_out in zip(dims[:-1], dims[1:]):
            self.linears.append(nn.Linear(d_in, d_out))
            self.dropouts.append(nn.Dropout(dropout))
        self.activation = get_activation(activation)
        self.n_layers = len(self.linears)
        self.layer_range = range(self.n_layers)

    def forward(self, x):
        for i_layer, layer, dropout in zip(self.layer_range, self.linears, self.dropouts):
            x = layer(x)
            if i_layer != self.n_layers - 1:
                x = self.activation(dropout(x))
        return x  # do not use dropout on last layer?

class DeepsetCocktailEncoder(nn.Module):
    def __init__(self, input_dim, deepset_latent_dim, hidden_dims_ing, activation,
                 hidden_dims_cocktail, latent_dim, aggregation, dropout):
        super(DeepsetCocktailEncoder, self).__init__()
        self.input_dim = input_dim  # dimension of ingredient representation + quantity
        self.ingredient_encoder = IngredientEncoder(input_dim, deepset_latent_dim, hidden_dims_ing, activation, dropout)  # encode each ingredient separately
        self.deepset_latent_dim = deepset_latent_dim  # dimension of the deepset aggregation
        self.aggregation = aggregation
        self.latent_dim = latent_dim
        # post aggregation network
        self.linears = nn.ModuleList()
        self.dropouts = nn.ModuleList()
        dims = [deepset_latent_dim] + hidden_dims_cocktail
        for d_in, d_out in zip(dims[:-1], dims[1:]):
            self.linears.append(nn.Linear(d_in, d_out))
            self.dropouts.append(nn.Dropout(dropout))
        self.FC_mean  = nn.Linear(hidden_dims_cocktail[-1], latent_dim)
        self.FC_logvar   = nn.Linear(hidden_dims_cocktail[-1], latent_dim)
        self.softplus = nn.Softplus()

        self.activation = get_activation(activation)
        self.n_layers = len(self.linears)
        self.layer_range = range(self.n_layers)

    def forward(self, nb_ingredients, x):

        # reshape x in (batch size * nb ingredients, dim_ing_rep)
        batch_size = x.shape[0]
        all_ingredients = []
        for i in range(batch_size):
            for j in range(nb_ingredients[i]):
                all_ingredients.append(x[i, self.input_dim * j: self.input_dim * (j + 1)].reshape(1, -1))
        x = torch.cat(all_ingredients, dim=0)
        # encode ingredients in parallel
        ingredients_encodings = self.ingredient_encoder(x)
        assert ingredients_encodings.shape == (torch.sum(nb_ingredients), self.deepset_latent_dim)

        # aggregate
        x = []
        index_first = 0
        for i in range(batch_size):
            index_last = index_first + nb_ingredients[i]
            # aggregate
            if self.aggregation == 'sum':
                x.append(torch.sum(ingredients_encodings[index_first:index_last], dim=0).reshape(1, -1))
            elif self.aggregation == 'mean':
                x.append(torch.mean(ingredients_encodings[index_first:index_last], dim=0).reshape(1, -1))
            else:
                raise ValueError
            index_first = index_last
        x = torch.cat(x, dim=0)
        assert x.shape[0] == batch_size

        for i_layer, layer, dropout in zip(self.layer_range, self.linears, self.dropouts):
            x = self.activation(dropout(layer(x)))
        mean = self.FC_mean(x)
        logvar = self.FC_logvar(x)
        return mean, logvar

class Decoder(nn.Module):
    def __init__(self, latent_dim, hidden_dims, num_ingredients, activation, dropout, filter_output=None):
        super(Decoder, self).__init__()
        self.linears = nn.ModuleList()
        self.dropouts = nn.ModuleList()
        dims = [latent_dim] + hidden_dims + [num_ingredients]
        for d_in, d_out in zip(dims[:-1], dims[1:]):
            self.linears.append(nn.Linear(d_in, d_out))
            self.dropouts.append(nn.Dropout(dropout))
        self.activation = get_activation(activation)
        self.n_layers = len(self.linears)
        self.layer_range = range(self.n_layers)
        self.filter = filter_output

    def forward(self, x, to_filter=False):
        for i_layer, layer, dropout in zip(self.layer_range, self.linears, self.dropouts):
            x = layer(x)
            if i_layer != self.n_layers - 1:
                x = self.activation(dropout(x))
        if to_filter:
            x = self.filter(x)
        return x

class PredictorHead(nn.Module):
    def __init__(self, latent_dim, dim_output, final_activ):
        super(PredictorHead, self).__init__()
        self.linear = nn.Linear(latent_dim, dim_output)
        if final_activ != None:
            self.final_activ = get_activation(final_activ)
            self.use_final_activ = True
        else:
            self.use_final_activ = False

    def forward(self, x):
        x = self.linear(x)
        if self.use_final_activ: x = self.final_activ(x)
        return x


class VAEModel(nn.Module):
    def __init__(self, encoder, decoder, auxiliaries_dict):
        super(VAEModel, self).__init__()
        self.encoder = encoder
        self.decoder = decoder
        self.latent_dim = self.encoder.latent_dim
        self.auxiliaries_str = []
        self.auxiliaries = nn.ModuleList()
        for aux_str in sorted(auxiliaries_dict.keys()):
            if aux_str == 'taste_reps':
                self.taste_reps_decoder = PredictorHead(self.latent_dim, auxiliaries_dict[aux_str]['dim_output'], auxiliaries_dict[aux_str]['final_activ'])
            else:
                self.auxiliaries_str.append(aux_str)
                self.auxiliaries.append(PredictorHead(self.latent_dim, auxiliaries_dict[aux_str]['dim_output'], auxiliaries_dict[aux_str]['final_activ']))

    def reparameterization(self, mean, logvar):
        std = torch.exp(0.5 * logvar)
        epsilon = torch.randn_like(std).to(device)        # sampling epsilon
        z = mean + std * epsilon                          # reparameterization trick
        return z


    def sample(self, n=1):
        z = torch.randn(size=(n, self.latent_dim))
        return self.decoder(z)

    def get_all_auxiliaries(self, x):
        return [aux(x) for aux in self.auxiliaries]

    def get_auxiliary(self, z, aux_str):
        if aux_str == 'taste_reps':
            return self.taste_reps_decoder(z)
        else:
            index = self.auxiliaries_str.index(aux_str)
            return self.auxiliaries[index](z)

    def forward_direct(self, x, aux_str=None, to_filter=False):
        mean, logvar = self.encoder(x)
        z = self.reparameterization(mean, logvar)  # takes exponential function (log var -> std)
        x_hat = self.decoder(mean, to_filter=to_filter)
        if aux_str is not None:
            return x_hat, z, mean, logvar, self.get_auxiliary(z, aux_str), [aux_str]
        else:
            return x_hat, z, mean, logvar, self.get_all_auxiliaries(z), self.auxiliaries_str

    def forward(self, nb_ingredients, x, aux_str=None, to_filter=False):
        assert False
        mean, std = self.encoder(nb_ingredients, x)
        z = self.reparameterization(mean, std)  # takes exponential function (log var -> std)
        x_hat = self.decoder(mean, to_filter=to_filter)
        if aux_str is not None:
            return x_hat, z, mean, std, self.get_auxiliary(z, aux_str), [aux_str]
        else:
            return x_hat, z, mean, std, self.get_all_auxiliaries(z), self.auxiliaries_str




class SimpleEncoder(nn.Module):

    def __init__(self, input_dim, hidden_dims, latent_dim, activation, dropout):
        super(SimpleEncoder, self).__init__()
        self.latent_dim = latent_dim
        # post aggregation network
        self.linears = nn.ModuleList()
        self.dropouts = nn.ModuleList()
        dims = [input_dim] + hidden_dims
        for d_in, d_out in zip(dims[:-1], dims[1:]):
            self.linears.append(nn.Linear(d_in, d_out))
            self.dropouts.append(nn.Dropout(dropout))
        self.FC_mean = nn.Linear(hidden_dims[-1], latent_dim)
        self.FC_logvar = nn.Linear(hidden_dims[-1], latent_dim)
        # self.softplus = nn.Softplus()

        self.activation = get_activation(activation)
        self.n_layers = len(self.linears)
        self.layer_range = range(self.n_layers)

    def forward(self, x):
        for i_layer, layer, dropout in zip(self.layer_range, self.linears, self.dropouts):
            x = self.activation(dropout(layer(x)))
        mean = self.FC_mean(x)
        logvar = self.FC_logvar(x)
        return mean, logvar

def get_vae_model(input_dim, deepset_latent_dim, hidden_dims_ing, activation,
                  hidden_dims_cocktail, hidden_dims_decoder, num_ingredients, latent_dim, aggregation, dropout, auxiliaries_dict,
                  filter_decoder_output):
    # encoder = DeepsetCocktailEncoder(input_dim, deepset_latent_dim, hidden_dims_ing, activation,
    #                                  hidden_dims_cocktail, latent_dim, aggregation, dropout)
    encoder = SimpleEncoder(num_ingredients, hidden_dims_cocktail, latent_dim, activation, dropout)
    decoder = Decoder(latent_dim, hidden_dims_decoder, num_ingredients, activation, dropout, filter_output=filter_decoder_output)
    vae = VAEModel(encoder, decoder, auxiliaries_dict)
    return vae