-
Notifications
You must be signed in to change notification settings - Fork 80
/
model.py
355 lines (275 loc) · 13.7 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
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
# Copyright (c) 2022 Graphcore Ltd. All rights reserved.
# Copyright (c) 2021 OpenAI
# This file has been modified by Graphcore
from collections import OrderedDict
import numpy as np
import poptorch
import torch
from torch import nn
class QuickGELU(nn.Module):
def forward(self, x: torch.Tensor):
return x * torch.sigmoid(1.702 * x)
def recomputation_checkpoint(module):
"""
Annotates the output of a module to be checkpointed instead of recomputed
"""
def recompute_outputs(module, inputs, outputs):
if type(outputs) is tuple:
return tuple(poptorch.recomputationCheckpoint(y) for y in outputs)
else:
return poptorch.recomputationCheckpoint(outputs)
module.register_forward_hook(recompute_outputs)
class ResidualAttentionBlock(nn.Module):
def __init__(self, d_model: int, n_head: int, attn_mask: torch.Tensor = None):
super().__init__()
self.attn = nn.MultiheadAttention(d_model, n_head)
self.ln_1 = nn.LayerNorm(d_model)
self.mlp = nn.Sequential(
OrderedDict(
[
("c_fc", nn.Linear(d_model, d_model * 4)),
("gelu", QuickGELU()),
("c_proj", nn.Linear(d_model * 4, d_model)),
]
)
)
self.ln_2 = nn.LayerNorm(d_model)
self.attn_mask = attn_mask
def attention(self, x: torch.Tensor):
if self.attn_mask is not None:
self.attn_mask = self.attn_mask.to(x.device)
self.attn_mask = self.attn_mask.type(x.dtype)
return self.attn(x, x, x, need_weights=False, attn_mask=self.attn_mask)[0]
def forward(self, x: torch.Tensor):
x = x + self.attention(self.ln_1(x))
x = x + self.mlp(self.ln_2(x))
return x
class Transformer(nn.Module):
def __init__(self, width: int, layers: int, heads: int, attn_mask: torch.Tensor = None):
super().__init__()
self.width = width
self.layers = layers
self.resblocks = nn.Sequential(*[ResidualAttentionBlock(width, heads, attn_mask) for _ in range(layers)])
def forward(self, x: torch.Tensor):
return self.resblocks(x)
class VisionTransformer(nn.Module):
def __init__(self, input_resolution: int, patch_size: int, width: int, layers: int, heads: int, output_dim: int):
super().__init__()
self.input_resolution = input_resolution
self.output_dim = output_dim
self.conv1 = nn.Conv2d(in_channels=3, out_channels=width, kernel_size=patch_size, stride=patch_size, bias=False)
scale = width**-0.5
self.class_embedding = nn.Parameter(scale * torch.randn(width))
self.positional_embedding = nn.Parameter(scale * torch.randn((input_resolution // patch_size) ** 2 + 1, width))
self.ln_pre = nn.LayerNorm(width)
self.transformer = Transformer(width, layers, heads, attn_mask=None)
self.ln_post = nn.LayerNorm(width)
# Scale the weight of self.proj
self.proj = nn.Parameter(scale * torch.randn(width, output_dim))
def forward(self, x: torch.Tensor):
x = self.conv1(x) # shape = [*, width, grid, grid]
x = x.reshape(x.shape[0], x.shape[1], -1) # shape = [*, width, grid ** 2]
x = x.permute(0, 2, 1) # shape = [*, grid ** 2, width]
x = torch.cat(
[self.class_embedding + torch.zeros(x.shape[0], 1, x.shape[-1], device=x.device, dtype=x.dtype), x], dim=1
) # shape = [*, grid ** 2 + 1, width]
x = x + self.positional_embedding
x = self.ln_pre(x)
x = x.permute(1, 0, 2) # NLD -> LND
x = self.transformer(x)
x = x.permute(1, 0, 2) # LND -> NLD
x = self.ln_post(x[:, 0, :])
if self.proj is not None:
x = x @ self.proj
return x
class CLIP(nn.Module):
def __init__(self, config):
super().__init__()
self.context_length = config.context_length
self.batch_size = config.batch_size
self.memory_size = config.memory_size
self.embed_dim = config.embed_dim
vision_heads = config.vision_width // 64
self.visual = VisionTransformer(
input_resolution=config.image_resolution,
patch_size=config.vision_patch_size,
width=config.vision_width,
layers=config.vision_layers,
heads=vision_heads,
output_dim=config.embed_dim,
)
self.transformer = Transformer(
width=config.transformer_width,
layers=config.transformer_layers,
heads=config.transformer_heads,
attn_mask=self.build_attention_mask(),
)
self.vocab_size = config.vocab_size
self.token_embedding = torch.nn.Embedding(config.vocab_size, config.transformer_width)
self.positional_embedding = nn.Parameter(torch.empty(self.context_length, config.transformer_width))
self.ln_final = nn.LayerNorm(config.transformer_width)
self.text_projection = nn.Parameter(torch.empty(config.transformer_width, config.embed_dim))
self.logit_scale = nn.Parameter(torch.ones([]) * np.log(1 / 0.07))
self.initialize_parameters()
# Allocate the register buffers to store the features of the passed steps
self.register_buffer(
"image_fea_queue",
torch.normal(
mean=0.0,
std=(self.transformer.width**-0.5) * ((2 * self.transformer.layers) ** -0.5),
size=(self.memory_size * self.batch_size, self.embed_dim),
),
)
self.register_buffer(
"text_fea_queue",
torch.normal(
mean=0.0,
std=(self.transformer.width**-0.5) * ((2 * self.transformer.layers) ** -0.5),
size=(self.memory_size * self.batch_size, self.embed_dim),
),
)
# Loss
self.loss = nn.CrossEntropyLoss()
def initialize_parameters(self):
nn.init.normal_(self.token_embedding.weight, std=0.02)
nn.init.normal_(self.positional_embedding, std=0.01)
proj_std = (self.transformer.width**-0.5) * ((2 * self.transformer.layers) ** -0.5)
attn_std = self.transformer.width**-0.5
fc_std = (2 * self.transformer.width) ** -0.5
for block in self.transformer.resblocks:
nn.init.normal_(block.attn.in_proj_weight, std=attn_std)
nn.init.normal_(block.attn.out_proj.weight, std=proj_std)
nn.init.normal_(block.mlp.c_fc.weight, std=fc_std)
nn.init.normal_(block.mlp.c_proj.weight, std=proj_std)
if self.text_projection is not None:
nn.init.normal_(self.text_projection, std=self.transformer.width**-0.5)
@torch.no_grad()
def dequeue_enqueue(self, image_fea, text_fea):
# Update the image features encoded in the current step to the register buffer
last_image = self.image_fea_queue[: (self.memory_size - 1) * self.batch_size, :]
update_image = torch.cat([image_fea, last_image], dim=0)
self.image_fea_queue.copy_(update_image)
# Update the text features encoded in the current step to the register buffer
last_text = self.text_fea_queue[: (self.memory_size - 1) * self.batch_size, :]
update_text = torch.cat([text_fea, last_text], dim=0)
self.text_fea_queue.copy_(update_text)
def build_attention_mask(self):
# Lazily create causal attention mask, with full attention between the vision tokens
# pytorch uses additive attention mask; fill with -64512 which is an invalid value
mask = torch.empty(self.context_length, self.context_length)
mask.fill_(-64512)
mask.triu_(1) # Zero out the lower diagonal
return mask
def encode_image(self, image):
return self.visual(image)
def encode_text(self, text):
x = self.token_embedding(text) # [batch_size, n_ctx, d_model]
x = x + self.positional_embedding
# Because the batch_first = False
x = x.permute(1, 0, 2) # NLD -> LND
x = self.transformer(x)
x = x.permute(1, 0, 2) # LND -> NLD
x = self.ln_final(x)
# x.shape = [batch_size, n_ctx, transformer.width]
# Take features from the eot embedding (eot_token is the highest number in each sequence)
x = x[torch.arange(x.shape[0], device=x.device).long(), text.argmax(dim=-1)] @ self.text_projection
return x
def parallelize(self, log):
"""
Transform the model to run in an IPU pipeline.
- Adds pipeline stages to the model
- Replaces self-attention layers with fused-qkv self-attention layers
- (If enabled) Replaces the word embedding with a SerializedEmbedding
- Adds recomputation checkpoints
Recommended usage:
```
model = CLIP(config).parallelize().half()
```
"""
log.logger.info("---------- Device Allocation -----------")
log.logger.info("image_encoder 0 --> IPU 0")
for index in range(1):
layer = self.visual.transformer.resblocks[index]
recomputation_checkpoint(layer)
self.visual.transformer.resblocks[index] = poptorch.BeginBlock(
layer, f"image_encoder_layer{index}", ipu_id=0
)
log.logger.info("image_encoder 1 ~ 3 --> IPU 1")
for index in range(1, 4):
layer = self.visual.transformer.resblocks[index]
recomputation_checkpoint(layer)
self.visual.transformer.resblocks[index] = poptorch.BeginBlock(
layer, f"image_encoder_layer{index}", ipu_id=1
)
log.logger.info("image_encoder 4 ~ 6 --> IPU 2")
for index in range(4, 7):
layer = self.visual.transformer.resblocks[index]
recomputation_checkpoint(layer)
self.visual.transformer.resblocks[index] = poptorch.BeginBlock(
layer, f"image_encoder_layer{index}", ipu_id=2
)
log.logger.info("image_encoder 7 ~ 9 --> IPU 3")
for index in range(7, 10):
layer = self.visual.transformer.resblocks[index]
recomputation_checkpoint(layer)
self.visual.transformer.resblocks[index] = poptorch.BeginBlock(
layer, f"image_encoder_layer{index}", ipu_id=3
)
log.logger.info("image_encoder 10 ~ 11 --> IPU 4")
for index in range(10, 12):
layer = self.visual.transformer.resblocks[index]
recomputation_checkpoint(layer)
self.visual.transformer.resblocks[index] = poptorch.BeginBlock(
layer, f"image_encoder_layer{index}", ipu_id=4
)
log.logger.info("token_embedding --> IPU 5")
self.token_embedding = poptorch.BeginBlock(self.token_embedding, "embedding", ipu_id=5)
log.logger.info("text_enocder 0 --> IPU 5")
layer = self.transformer.resblocks[0]
recomputation_checkpoint(layer)
self.transformer.resblocks[0] = poptorch.BeginBlock(layer, "text_encoder_layer0", ipu_id=5)
log.logger.info("text_enocder 1 ~ 5 --> IPU 6")
for index in range(1, 6):
layer = self.transformer.resblocks[index]
recomputation_checkpoint(layer)
self.transformer.resblocks[index] = poptorch.BeginBlock(layer, f"text_encoder_layer{index}", ipu_id=6)
log.logger.info("text_enocder 6 ~ 11 --> IPU 7")
for index in range(6, 12):
layer = self.transformer.resblocks[index]
recomputation_checkpoint(layer)
self.transformer.resblocks[index] = poptorch.BeginBlock(layer, f"text_encoder_layer{index}", ipu_id=7)
log.logger.info("loss --> IPU 7")
self.loss = poptorch.BeginBlock(self.loss, f"loss", ipu_id=7)
log.logger.info("---------------------------------------")
return self
def forward(self, images=None, texts=None):
if self.training:
image_features = self.encode_image(images)
text_features = self.encode_text(texts)
image_features = image_features / image_features.norm(dim=1, keepdim=True)
text_features = text_features / text_features.norm(dim=1, keepdim=True)
# Concat the features encoded in the current step and from the register buffer
n_text_features = torch.cat([text_features, self.text_fea_queue.clone().detach()], dim=0)
n_image_features = torch.cat([image_features, self.image_fea_queue.clone().detach()], dim=0)
logits_per_image = self.logit_scale.exp() * torch.mm(n_image_features, n_text_features.t())
logits_per_text = logits_per_image.t()
labels = torch.arange(logits_per_image.size()[0], device=logits_per_text.device).long()
i_loss = self.loss(logits_per_image, labels)
t_loss = self.loss(logits_per_text, labels)
loss = (i_loss + t_loss) / 2.0
self.dequeue_enqueue(image_features, text_features)
return poptorch.identity_loss(loss, reduction="mean")
else:
# Only encode all the texts for zeroshot test
if images.mean() == 0:
class_embeddings = self.encode_text(texts)
class_embeddings = class_embeddings / class_embeddings.norm(dim=1, keepdim=True)
class_embedding = class_embeddings.mean(dim=0)
class_embedding /= class_embedding.norm()
return class_embedding
# Only encode the images for zeroshot test
else:
image_features = self.encode_image(images)
image_features = image_features / image_features.norm(dim=1, keepdim=True)
logits_per_image = 100.0 * self.logit_scale.exp() * image_features @ texts
return logits_per_image