-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathmodel.py
278 lines (214 loc) · 10.1 KB
/
model.py
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
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
import torch
import torch.nn as nn
import torch.nn.functional as F
from torch.nn.utils.rnn import pad_sequence
import math
import numpy as np
from torch_geometric.nn import MessagePassing
from torch_geometric.utils import add_self_loops, degree
from torch_geometric.nn import GCNConv, SAGEConv,TransformerConv,GATConv
def get_emb(sin_inp):
"""
Gets a base embedding for one dimension with sin and cos intertwined
"""
emb = torch.stack((sin_inp.sin(), sin_inp.cos()), dim=-1)
return torch.flatten(emb, -2, -1)
class PositionalEncoding1D(nn.Module):
def __init__(self, channels):
"""
:param channels: The last dimension of the tensor you want to apply pos emb to.
"""
super(PositionalEncoding1D, self).__init__()
self.org_channels = channels
channels = int(np.ceil(channels / 2) * 2)
self.channels = channels
inv_freq = 1.0 / (10000 ** (torch.arange(0, channels, 2).float() / channels))
self.register_buffer("inv_freq", inv_freq)
self.cached_penc = None
def forward(self, tensor):
"""
:param tensor: A 3d tensor of size (batch_size, x, ch)
:return: Positional Encoding Matrix of size (batch_size, x, ch)
"""
if len(tensor.shape) != 3:
raise RuntimeError("The input tensor has to be 3d!")
if self.cached_penc is not None and self.cached_penc.shape == tensor.shape:
return self.cached_penc
self.cached_penc = None
batch_size, x, orig_ch = tensor.shape
pos_x = torch.arange(x, device=tensor.device).type(self.inv_freq.type())
sin_inp_x = torch.einsum("i,j->ij", pos_x, self.inv_freq)
emb_x = get_emb(sin_inp_x)
emb = torch.zeros((x, self.channels), device=tensor.device).type(tensor.type())
emb[:, : self.channels] = emb_x
self.cached_penc = emb[None, :, :orig_ch].repeat(batch_size, 1, 1)
return self.cached_penc
class TransformerModel(nn.Module):
def __init__(self, output_dim, d_input , d_model, nhead, num_layers, dropout=0.1):
super(TransformerModel, self).__init__()
self.d_model = d_model
self.pose_embed = nn.Linear(d_input, d_model)
self.tgt_query = nn.Embedding(output_dim+1, d_model)
self.d_input = d_input
self.class_token = torch.nn.Parameter(
torch.randn(1, 1, self.d_input)
)
# self.pos_encoder = PositionalEncoding(d_model, dropout)
# LAYERS
self.positional_encoder = PositionalEncoding1D(d_model)
self.transformer_encoder = nn.TransformerEncoder(
nn.TransformerEncoderLayer(d_model=d_model, nhead=nhead, dropout=dropout,batch_first=True),
num_layers=num_layers)
self.transformer_decoder = nn.TransformerDecoder(
nn.TransformerDecoderLayer(d_model=d_model, nhead=nhead, dropout=dropout,batch_first=True),
num_layers=num_layers)
self.fc = nn.Sequential(
nn.Linear(d_model, 128),
nn.Dropout(0.1),
nn.Linear(128, output_dim)
)
self.fc_enc = nn.Sequential(
nn.Linear(d_model, 128),
nn.Dropout(0.1),
nn.Linear(128, output_dim)
)
# self.fc_cls = nn.Sequential(
# nn.Linear(d_model, 128),
# nn.Dropout(0.1),
# nn.Linear(128, 2)
# )
self.fc_cls = nn.Sequential(nn.LayerNorm(d_model), nn.Linear(d_model, 2))
self.transformer = nn.Transformer(
d_model=d_model,
nhead=nhead,
num_encoder_layers=num_layers,
num_decoder_layers=num_layers,
dropout=dropout,
batch_first=True,
)
def forward(self, poses,tgt=None):
# Embedding layer for poses
# poses = poses.view(1,-1,63)
# pose_embedded = self.pose_embed(poses)
# print(poses.shape)
bs = poses.shape[0]
poses = poses.view(bs,-1,self.d_input)
poses = torch.cat([self.class_token, poses], dim=1)
pose_embedded = self.pose_embed(poses)
pos_embd = self.positional_encoder(pose_embedded)
# print(self.tgt_query.shape)
# Positional encoding
batch_size , pose_len, pose_dim = pose_embedded.size()
# pos = torch.arange(0, pose_len).unsqueeze(1).repeat(1, batch_size).to(pose_embedded.device)
# # print(pos.shape)
# pos_embedded = self.pos_embed(pos).permute(1, 0, 2)
src = pose_embedded + pos_embd.to(pose_embedded.device)
# transformer_out = self.transformer(src, tgt)
# Transformer encoder
encoder = self.transformer_encoder(src) # B x L x D
encoder_out = self.fc_enc(encoder[:,1:,:]) # L x B x V
cls_out = self.fc_cls(encoder[:,0,:]) # L x B x V
if tgt!=None:
tgt = tgt.reshape(bs,-1,1)
pos_tgt = self.positional_encoder(tgt)
tgt = self.tgt_query(tgt).reshape(bs,-1,self.d_model)
tgt = tgt + pos_tgt.to(pose_embedded.device)
tgt_mask = self.get_tgt_mask(tgt.size(1)).to(pose_embedded.device)
decoder = self.transformer_decoder(tgt, encoder, tgt_mask=tgt_mask)
logits = self.fc(decoder) # L x B x V
return cls_out ,logits, encoder_out # B x L x V
return cls_out, encoder_out # B x L x V
@torch.no_grad()
def return_scores(self, poses, strings, vocab_map):
scores = []
bs = poses.shape[0]
poses = poses.view(bs,-1,self.d_input)
poses = torch.cat([self.class_token, poses], dim=1)
pose_embedded = self.pose_embed(poses)
pos_embd = self.positional_encoder(pose_embedded)
# print(self.tgt_query.shape)
# Positional encoding
batch_size , pose_len, pose_dim = pose_embedded.size()
# pos = torch.arange(0, pose_len).unsqueeze(1).repeat(1, batch_size).to(pose_embedded.device)
# # print(pos.shape)
# pos_embedded = self.pos_embed(pos).permute(1, 0, 2)
src = pose_embedded + pos_embd.to(pose_embedded.device)
# transformer_out = self.transformer(src, tgt)
# Transformer encoder
encoder = self.transformer_encoder(src) # B x L x D
encoder_out = self.fc_enc(encoder[:,1:,:]) # L x B x V
cls_out = self.fc_cls(encoder[:,0,:]) # L x B x V
for string in strings:
if len(string)<1 :
scores.append(0)
continue
criterion = nn.CrossEntropyLoss()
tgt = torch.tensor([[32]], dtype=torch.long, device=poses.device)
input = [32] + list(map(lambda x: vocab_map[x], string))
grt = torch.tensor(input[1:], dtype=torch.long, device=poses.device)
tgt = torch.tensor(input[:-1], dtype=torch.long, device=poses.device)
tgt = tgt.reshape(bs,-1,1)
pos_tgt = self.positional_encoder(tgt)
tgt = self.tgt_query(tgt).reshape(bs,-1,self.d_model)
tgt = tgt + pos_tgt.to(pose_embedded.device)
tgt_mask = self.get_tgt_mask(tgt.size(1)).to(pose_embedded.device)
decoder = self.transformer_decoder(tgt, encoder, tgt_mask=tgt_mask)
logits = self.fc(decoder) # L x B x V
loss = criterion(logits[0], grt)
scores.append(-(len(input)-1)*loss.item())
return scores
def create_pad_mask(self, matrix: torch.tensor, pad_token: int):
# If matrix = [1,2,3,0,0,0] where pad_token=0, the result mask is
# [False, False, False, True, True, True]
return (matrix == pad_token)
def get_tgt_mask(self, size):
# Generates a squeare matrix where the each row allows one word more to be seen
mask = torch.tril(torch.ones(size, size) == 1) # Lower triangular matrix
mask = mask.float()
mask = mask.masked_fill(mask == 0, float('-inf')) # Convert zeros to -inf
mask = mask.masked_fill(mask == 1, float(0.0)) # Convert ones to 0
# EX for size=5:
# [[0., -inf, -inf, -inf, -inf],
# [0., 0., -inf, -inf, -inf],
# [0., 0., 0., -inf, -inf],
# [0., 0., 0., 0., -inf],
# [0., 0., 0., 0., 0.]]
return mask
def collate_fn(batch):
# Pad sequences to the same length
# Not used right now
poses, labels = zip(*batch)
poses = pad_sequence(poses, batch_first=True)
labels = pad_sequence(labels, batch_first=True)
return poses, labels
class GCN(torch.nn.Module):
def __init__(self, input_chanels, hidden_channels1, hidden_channels2, hidden_channels3, output_channels):
super().__init__( )
# torch.manual_seed(1234567)
self.pose_embed = nn.Linear(input_chanels, hidden_channels1)
self.positional_encoder = PositionalEncoding1D(hidden_channels1)
self.transformer1 = TransformerConv(hidden_channels1, hidden_channels2, heads=8, dropout=0.1)
self.head_transformer1 =nn.Linear(hidden_channels1*8, hidden_channels1)
self.transformer2 = TransformerConv(hidden_channels1, hidden_channels2, heads=8, dropout=0.1)
self.head_transformer2 =nn.Linear(hidden_channels1*8, hidden_channels1)
self.conv3 = GCNConv(hidden_channels2, hidden_channels3)
self.conv4 = GCNConv(hidden_channels3, output_channels)
self.non_linearity = nn.ELU()
self.activation = torch.nn.Sigmoid()
def forward(self, data):
# print(f'Inside model - num graphs: {data.num_graphs},', f'device: {data.batch.device}')
x, edge_index = data.x, data.edge_index
# import pdb; pdb.set_trace()
x = x.view(-1,63)
x = self.pose_embed(x)
pos_embd = self.positional_encoder(x.unsqueeze(dim=0))
x = x + pos_embd.to(x.device)[0]
x = self.transformer1(x,edge_index)
x = self.head_transformer1(x)
x = self.transformer2(x,edge_index)
x = self.head_transformer2(x)
x = self.conv3(x,edge_index)
x = F.dropout(x, p=0.2, training=self.training)
x = self.conv4(x,edge_index)
# import pdb; pdb.set_trace()
return x.unsqueeze(0)