-
Notifications
You must be signed in to change notification settings - Fork 631
/
TabTransformer.py
290 lines (229 loc) · 12.6 KB
/
TabTransformer.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
279
280
281
282
283
284
285
286
287
288
289
290
# AUTOGENERATED! DO NOT EDIT! File to edit: ../../nbs/064_models.TabTransformer.ipynb.
# %% auto 0
__all__ = ['ifnone', 'SharedEmbedding', 'FullEmbeddingDropout', 'TabTransformer']
# %% ../../nbs/064_models.TabTransformer.ipynb 4
# This is an unofficial TabTransformer implementation in Pytorch developed by Ignacio Oguiza - oguiza@timeseriesAI.co based on:
# Huang, X., Khetan, A., Cvitkovic, M., & Karnin, Z. (2020).
# TabTransformer: Tabular Data Modeling Using Contextual Embeddings.
# arXiv preprint https://arxiv.org/pdf/2012.06678
# Official repo: https://github.com/awslabs/autogluon/tree/master/tabular/src/autogluon/tabular/models/tab_transformer
import torch
import torch.nn as nn
import torch.nn.functional as F
def ifnone(a, b):
# From fastai.fastcore
"`b` if `a` is None else `a`"
return b if a is None else a
def _trunc_normal_(x, mean=0., std=1.):
"Truncated normal initialization (approximation)"
# From fastai.layers
# From https://discuss.pytorch.org/t/implementing-truncated-normal-initializer/4778/12
return x.normal_().fmod_(2).mul_(std).add_(mean)
class _Embedding(nn.Embedding):
"Embedding layer with truncated normal initialization"
# From fastai.layers
def __init__(self, ni, nf, std=0.01):
super(_Embedding, self).__init__(ni, nf)
_trunc_normal_(self.weight.data, std=std)
class SharedEmbedding(nn.Module):
def __init__(self, num_embeddings, embedding_dim, shared_embed=True, add_shared_embed=False, shared_embed_div=8):
super().__init__()
if shared_embed:
if add_shared_embed:
shared_embed_dim = embedding_dim
self.embed = _Embedding(num_embeddings, embedding_dim)
else:
shared_embed_dim = embedding_dim // shared_embed_div
self.embed = _Embedding(num_embeddings, embedding_dim - shared_embed_dim)
self.shared_embed = nn.Parameter(torch.empty(1, 1, shared_embed_dim))
_trunc_normal_(self.shared_embed.data, std=0.01)
self.add_shared_embed = add_shared_embed
else:
self.embed = _Embedding(num_embeddings, embedding_dim)
self.shared_embed = None
def forward(self, x):
out = self.embed(x).unsqueeze(1)
if self.shared_embed is None: return out
if self.add_shared_embed:
out += self.shared_embed
else:
shared_embed = self.shared_embed.expand(out.shape[0], -1, -1)
out = torch.cat((out, shared_embed), dim=-1)
return out
class FullEmbeddingDropout(nn.Module):
'''From https://github.com/jrzaurin/pytorch-widedeep/blob/be96b57f115e4a10fde9bb82c35380a3ac523f52/pytorch_widedeep/models/tab_transformer.py#L153'''
def __init__(self, dropout: float):
super().__init__()
self.dropout = dropout
def forward(self, x):
mask = x.new().resize_((x.size(1), 1)).bernoulli_(1 - self.dropout).expand_as(x) / (1 - self.dropout)
return mask * x
class _MLP(nn.Module):
def __init__(self, dims, bn=False, act=None, skip=False, dropout=0., bn_final=False):
super().__init__()
dims_pairs = list(zip(dims[:-1], dims[1:]))
layers = []
for i, (dim_in, dim_out) in enumerate(dims_pairs):
is_last = i >= (len(dims) - 2)
if bn and (not is_last or bn_final): layers.append(nn.BatchNorm1d(dim_in))
if dropout and not is_last:
layers.append(nn.Dropout(dropout))
layers.append(nn.Linear(dim_in, dim_out))
if is_last: break
layers.append(ifnone(act, nn.ReLU()))
self.mlp = nn.Sequential(*layers)
self.shortcut = nn.Linear(dims[0], dims[-1]) if skip else None
def forward(self, x):
if self.shortcut is not None:
return self.mlp(x) + self.shortcut(x)
else:
return self.mlp(x)
class _ScaledDotProductAttention(nn.Module):
def __init__(self, d_k:int, res_attention:bool=False):
super().__init__()
self.d_k,self.res_attention = d_k,res_attention
def forward(self, q, k, v, prev=None, attn_mask=None):
# MatMul (q, k) - similarity scores for all pairs of positions in an input sequence
scores = torch.matmul(q, k) # scores : [bs x n_heads x q_len x q_len]
# Scale
scores = scores / (self.d_k ** 0.5)
# Attention mask (optional)
if attn_mask is not None: # mask with shape [q_len x q_len]
if attn_mask.dtype == torch.bool:
scores.masked_fill_(attn_mask, float('-inf'))
else:
scores += attn_mask
# SoftMax
if prev is not None: scores = scores + prev
attn = F.softmax(scores, dim=-1) # attn : [bs x n_heads x q_len x q_len]
# MatMul (attn, v)
context = torch.matmul(attn, v) # context: [bs x n_heads x q_len x d_v]
if self.res_attention: return context, attn, scores
else: return context, attn
class _MultiheadAttention(nn.Module):
def __init__(self, d_model:int, n_heads:int, d_k:int, d_v:int, res_attention:bool=False):
"""Input shape: Q, K, V:[batch_size (bs) x q_len x d_model], mask:[q_len x q_len]"""
super().__init__()
self.n_heads, self.d_k, self.d_v = n_heads, d_k, d_v
self.W_Q = nn.Linear(d_model, d_k * n_heads, bias=False)
self.W_K = nn.Linear(d_model, d_k * n_heads, bias=False)
self.W_V = nn.Linear(d_model, d_v * n_heads, bias=False)
self.W_O = nn.Linear(n_heads * d_v, d_model, bias=False)
self.res_attention = res_attention
# Scaled Dot-Product Attention (multiple heads)
if self.res_attention:
self.sdp_attn = _ScaledDotProductAttention(self.d_k, self.res_attention)
else:
self.sdp_attn = _ScaledDotProductAttention(self.d_k)
def forward(self, Q, K, V, prev=None, attn_mask=None):
bs = Q.size(0)
# Linear (+ split in multiple heads)
q_s = self.W_Q(Q).view(bs, -1, self.n_heads, self.d_k).transpose(1,2) # q_s : [bs x n_heads x q_len x d_k]
k_s = self.W_K(K).view(bs, -1, self.n_heads, self.d_k).permute(0,2,3,1) # k_s : [bs x n_heads x d_k x q_len] - transpose(1,2) + transpose(2,3)
v_s = self.W_V(V).view(bs, -1, self.n_heads, self.d_v).transpose(1,2) # v_s : [bs x n_heads x q_len x d_v]
# Scaled Dot-Product Attention (multiple heads)
if self.res_attention:
context, attn, scores = self.sdp_attn(q_s, k_s, v_s, prev=prev, attn_mask=attn_mask)
else:
context, attn = self.sdp_attn(q_s, k_s, v_s, attn_mask=attn_mask)
# context: [bs x n_heads x q_len x d_v], attn: [bs x n_heads x q_len x q_len]
# Concat
context = context.transpose(1, 2).contiguous().view(bs, -1, self.n_heads * self.d_v) # context: [bs x q_len x n_heads * d_v]
# Linear
output = self.W_O(context) # context: [bs x q_len x d_model]
if self.res_attention: return output, attn, scores
else: return output, attn # output: [bs x q_len x d_model]
class _TabEncoderLayer(nn.Module):
def __init__(self, q_len, d_model, n_heads, d_k=None, d_v=None, d_ff=None,
res_dropout=0.1, activation="gelu", res_attention=False):
super().__init__()
assert not d_model%n_heads, f"d_model ({d_model}) must be divisible by n_heads ({n_heads})"
d_k = ifnone(d_k, d_model // n_heads)
d_v = ifnone(d_v, d_model // n_heads)
d_ff = ifnone(d_ff, d_model * 4)
# Multi-Head attention
self.res_attention = res_attention
self.self_attn = _MultiheadAttention(d_model, n_heads, d_k, d_v, res_attention=res_attention)
# Add & Norm
self.dropout_attn = nn.Dropout(res_dropout)
self.layernorm_attn = nn.LayerNorm(d_model)
# Position-wise Feed-Forward
self.ff = nn.Sequential(nn.Linear(d_model, d_ff), self._get_activation_fn(activation), nn.Linear(d_ff, d_model))
# Add & Norm
self.dropout_ffn = nn.Dropout(res_dropout)
self.layernorm_ffn = nn.LayerNorm(d_model)
def forward(self, src, prev=None, attn_mask=None):
# Multi-Head attention sublayer
## Multi-Head attention
if self.res_attention:
src2, attn, scores = self.self_attn(src, src, src, prev, attn_mask=attn_mask)
else:
src2, attn = self.self_attn(src, src, src, attn_mask=attn_mask)
self.attn = attn
## Add & Norm
src = src + self.dropout_attn(src2) # Add: residual connection with residual dropout
src = self.layernorm_attn(src) # Norm: layernorm
# Feed-forward sublayer
## Position-wise Feed-Forward
src2 = self.ff(src)
## Add & Norm
src = src + self.dropout_ffn(src2) # Add: residual connection with residual dropout
src = self.layernorm_ffn(src) # Norm: layernorm
if self.res_attention:
return src, scores
else:
return src
def _get_activation_fn(self, activation):
if callable(activation): return activation()
elif activation.lower() == "relu": return nn.ReLU()
elif activation.lower() == "gelu": return nn.GELU()
raise ValueError(f'{activation} is not available. You can use "relu", "gelu", or a callable')
class _TabEncoder(nn.Module):
def __init__(self, q_len, d_model, n_heads, d_k=None, d_v=None, d_ff=None, res_dropout=0.1, activation='gelu', res_attention=False, n_layers=1):
super().__init__()
self.layers = nn.ModuleList([_TabEncoderLayer(q_len, d_model, n_heads=n_heads, d_k=d_k, d_v=d_v, d_ff=d_ff, res_dropout=res_dropout,
activation=activation, res_attention=res_attention) for i in range(n_layers)])
self.res_attention = res_attention
def forward(self, src, attn_mask=None):
output = src
scores = None
if self.res_attention:
for mod in self.layers: output, scores = mod(output, prev=scores, attn_mask=attn_mask)
return output
else:
for mod in self.layers: output = mod(output, attn_mask=attn_mask)
return output
class TabTransformer(nn.Module):
def __init__(self, classes, cont_names, c_out, column_embed=True, add_shared_embed=False, shared_embed_div=8, embed_dropout=0.1, drop_whole_embed=False,
d_model=32, n_layers=6, n_heads=8, d_k=None, d_v=None, d_ff=None, res_attention=True, attention_act='gelu', res_dropout=0.1, norm_cont=True,
mlp_mults=(4, 2), mlp_dropout=0., mlp_act=None, mlp_skip=False, mlp_bn=False, bn_final=False):
super().__init__()
n_cat = len(classes)
n_classes = [len(v) for v in classes.values()]
n_cont = len(cont_names)
self.embeds = nn.ModuleList([SharedEmbedding(ni, d_model, shared_embed=column_embed, add_shared_embed=add_shared_embed,
shared_embed_div=shared_embed_div) for ni in n_classes])
n_emb = sum(n_classes)
self.n_emb,self.n_cont = n_emb,n_cont
self.emb_drop = None
if embed_dropout:
self.emb_drop = FullEmbeddingDropout(embed_dropout) if drop_whole_embed else nn.Dropout(embed_dropout)
self.transformer = _TabEncoder(n_cat, d_model, n_heads=n_heads, d_k=d_k, d_v=d_v, d_ff=d_ff, res_dropout=res_dropout,
activation=attention_act, res_attention=res_attention, n_layers=n_layers)
self.norm = nn.LayerNorm(n_cont) if norm_cont else None
mlp_input_size = (d_model * n_cat) + n_cont
hidden_dimensions = list(map(lambda t: int(mlp_input_size * t), mlp_mults))
all_dimensions = [mlp_input_size, *hidden_dimensions, c_out]
self.mlp = _MLP(all_dimensions, act=mlp_act, skip=mlp_skip, bn=mlp_bn, dropout=mlp_dropout, bn_final=bn_final)
def forward(self, x_cat, x_cont=None):
if self.n_emb != 0:
x = [e(x_cat[:,i]) for i,e in enumerate(self.embeds)]
x = torch.cat(x, 1)
if self.emb_drop is not None: x = self.emb_drop(x)
x = self.transformer(x)
x = x.flatten(1)
if self.n_cont != 0:
if self.norm is not None: x_cont = self.norm(x_cont)
x = torch.cat([x, x_cont], 1) if self.n_emb != 0 else x_cont
x = self.mlp(x)
return x