-
Notifications
You must be signed in to change notification settings - Fork 0
/
vid_cap.py
204 lines (171 loc) · 7.84 KB
/
vid_cap.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
import tensorflow as tf
import re, h5py
import numpy as np
from tensorflow.keras.layers import Layer, Dropout, Dense
from tensorflow.keras import Model
from transformers import T5Tokenizer
from tx_helper import EncoderDecoder
class VidCap_Base(Model):
def __init__(self,
**kwargs):
super().__init__(**kwargs)
def set_strategy(self,
strategy):
self.strategy = strategy
@tf.function
def _init_0(self,
vid_inputs,
txt_inputs,
vid_inputs_attn_mask,
txt_inputs_attn_mask):
self.call(vid_inputs=vid_inputs,
txt_inputs=txt_inputs,
vid_inputs_attn_mask=vid_inputs_attn_mask,
txt_inputs_attn_mask=txt_inputs_attn_mask,
training=False)
@tf.function
def _init_d(self,
vid_inputs,
txt_inputs,
vid_inputs_attn_mask,
txt_inputs_attn_mask):
self.strategy.run(self._init_0,args=(vid_inputs,
txt_inputs,
vid_inputs_attn_mask,
txt_inputs_attn_mask,))
def _dummy_inputs(self,
batch_size,
vid_inputs=None,
from_features=True,
d_vid=0):
txt_max_length = 1
vid_max_length = 1
if from_features:
vid_inputs = tf.random.uniform([batch_size,1,d_vid], minval=0, maxval=255, dtype=tf.float32)
else:
vid_inputs = tf.cast(vid_inputs, dtype=tf.uint8)
txt_inputs = tf.random.uniform([batch_size,1], minval=0, maxval=10, dtype=tf.int32)
vid_inputs_attn_mask = tf.random.uniform([batch_size,1], minval=0, maxval=1, dtype=tf.int32)
txt_inputs_attn_mask = tf.random.uniform([batch_size,1], minval=0, maxval=1, dtype=tf.int32)
return vid_inputs, txt_inputs, vid_inputs_attn_mask, txt_inputs_attn_mask
def init(self,
batch_size,
vid_inputs=None,
from_features=True,
d_vid=0):
(vid_inputs,
txt_inputs,
vid_inputs_attn_mask,
txt_inputs_attn_mask,
) = self._dummy_inputs(batch_size, vid_inputs, from_features, d_vid)
self._init_0(vid_inputs, txt_inputs, vid_inputs_attn_mask, txt_inputs_attn_mask)
def distributed_init(self,
batch_size,
vid_inputs=None,
from_features=True,
d_vid=0):
(vid_inputs,
txt_inputs,
vid_inputs_attn_mask,
txt_inputs_attn_mask,
) = self._dummy_inputs(batch_size, vid_inputs, from_features, d_vid)
self._init_d(vid_inputs, txt_inputs, vid_inputs_attn_mask, txt_inputs_attn_mask)
def freeze_decoder(self,
skip_embeddings=True):
self.tx.decoder.trainable = False
self.tx.decoder_embeddings.trainable = skip_embeddings
class VidCapModel(VidCap_Base):
def __init__(self,
config,
name='VidCapModel',
**kwargs):
super().__init__(name=name, **kwargs)
config_tx = config['TX_CONFIG']
config_tx['GENERATION']['MAX_LENGTH'] = config['DATA']['MAX_TXT_LENGTH']
config_tx['ENCODER']['MAX_SPT_POSITIONS'] = config['DATA']['MAX_REGIONS']
config_tx['ENCODER']['MAX_TMP_POSITIONS'] = config['DATA']['MAX_FRAMES']
config_tx['ENCODER']['ROLE_TYPES'] = eval(config_tx['ENCODER']['ROLE_TYPES'])
config_tx['DECODER']['MAX_TMP_POSITIONS'] = config['DATA']['MAX_TXT_LENGTH']
config_tx['DECODER']['SELF_ROLE_TYPES'] = eval(config_tx['DECODER']['SELF_ROLE_TYPES'])
config_tx['DECODER']['CROSS_ROLE_TYPES'] = eval(config_tx['DECODER']['CROSS_ROLE_TYPES'])
self.gen_cfg = config_tx['GENERATION']
self.gen_fn = None
self.tokenizer = T5Tokenizer.from_pretrained('t5-base')
self.tx = EncoderDecoder(config_tx,
name='transformer')
self.proj = Dense(self.tx.encoder.d_model,
use_bias=False,
name='vid_proj')
@staticmethod
def _check_generation_graph_capability(gen_cfg):
use_cache = gen_cfg['USE_CACHE']
num_beams = gen_cfg['NUM_BEAMS']
length_penalty = gen_cfg['LENGTH_PENALTY']
repetition_penalty = gen_cfg['REPETITION_PENALTY']
no_repeat_ngram_size = gen_cfg['NO_REPEAT_NGRAM_SIZE']
autograph = True
autograph = autograph and (not use_cache)
autograph = autograph and (num_beams == 1)
autograph = autograph and (length_penalty == 1.0)
autograph = autograph and (repetition_penalty == 1.0)
autograph = autograph and (no_repeat_ngram_size == 0)
return autograph
@tf.function
def get_vid_features(self,
vid_inputs,
training):
vid_pos_ids = vid_inputs[:,:,-5:]
vid_features = vid_inputs[:,:,:-5]
vid_features = self.proj(vid_features)
vid_features = tf.nn.relu(vid_features)
return vid_features, vid_pos_ids
@tf.function
def call(self,
vid_inputs=None,
txt_inputs=None,
vid_inputs_attn_mask=None,
txt_inputs_attn_mask=None,
training=False):
vid_features, vid_pos_ids = self.get_vid_features(vid_inputs, training)
outputs = self.tx(inputs=None,
inputs_embeds=vid_features,
attention_mask=vid_inputs_attn_mask,
encoder_pos_ids=vid_pos_ids,
decoder_input_ids=txt_inputs,
decoder_attention_mask=txt_inputs_attn_mask,
use_cache=False,
training=training)
logits = outputs['decoder_head_logits']
classes = tf.nn.softmax(logits)
loss = outputs['loss']
outputs.update({'predictions': classes, 'loss': loss})
return outputs
@tf.function
def generate(self,
vid_inputs,
vid_mask,
**kwargs):
vid_features, vid_pos_ids = self.get_vid_features(vid_inputs, False)
if self.gen_fn is None:
#initialize self.gen_fn
self.autograph = VidCapModel._check_generation_graph_capability(self.gen_cfg)
if self.autograph:
self.gen_fn = tf.function(func=self.tx.generate, experimental_autograph_options=tf.autograph.experimental.Feature.EQUALITY_OPERATORS)
else:
self.gen_fn = self.tx.generate
decoded = self.gen_fn(input_embeds=vid_features,
attention_mask=vid_mask,
encoder_pos_ids=vid_pos_ids,
max_length=self.gen_cfg['MAX_LENGTH'],
early_stopping=self.gen_cfg['EARLY_STOPPING'],
num_beams=self.gen_cfg['NUM_BEAMS'],
no_repeat_ngram_size=self.gen_cfg['NO_REPEAT_NGRAM_SIZE'],
repetition_penalty=self.gen_cfg['REPETITION_PENALTY'],
length_penalty=self.gen_cfg['LENGTH_PENALTY'],
pad_token_id=self.gen_cfg['PAD_TOKEN_ID'],
eos_token_id=self.gen_cfg['EOS_TOKEN_ID'],
decoder_start_token_id=self.gen_cfg['DECODER_START_TOKEN_ID'],
use_cache=self.gen_cfg['USE_CACHE'],
autograph=self.autograph,
**kwargs)
return decoded