-
Notifications
You must be signed in to change notification settings - Fork 180
/
Copy pathrec_srn_head.py
196 lines (169 loc) · 7.06 KB
/
rec_srn_head.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
import os, sys
import numpy as np
import paddle
# paddle.enable_static()
import paddle.fluid as fluid
from paddle import ParamAttr
import paddle.nn as nn
import paddle.nn.functional as F
from collections import OrderedDict
import torch
from pp_rec_srn_head import SRNHead as PPHead
from pt_rec_srn_head import SRNHead as PTHead
SEED = 666
input_shape = (1, 512, 8, 32)
in_channels = 512
out_channels = 38
max_text_length = 25
num_heads = 8
num_encoder_TUs = 2
num_decoder_TUs = 4
hidden_dims = 512
def paddle_func():
np.random.seed(SEED)
# x = np.load('input.npy', allow_pickle=True)
x = np.random.rand(input_shape[0], input_shape[1], input_shape[2], input_shape[3]).astype(np.float32)
print('pp input size: {}'.format(x.shape))
# x1 = np.random.rand(1, 256, 1).astype(np.float32)
x1 = np.load('encoder_word_pos_list.npy', allow_pickle=True)
# x2 = np.random.rand(1, 25, 1).astype(np.int)
x2 = np.load('gsrm_word_pos_list.npy', allow_pickle=True)
# x3 = np.random.rand(1, 8, 25, 25).astype(np.float32)
x3 = np.load('gsrm_slf_attn_bias1_list.npy', allow_pickle=True)
# x4 = np.random.rand(1, 8, 25, 25).astype(np.float32)
x4 = np.load('gsrm_slf_attn_bias2_list.npy', allow_pickle=True)
# np.save('org.npy', x)
with fluid.dygraph.guard():
layer = PPHead(in_channels=in_channels,
out_channels=out_channels,
max_text_length=max_text_length,
num_heads=num_heads,
num_encoder_TUs=num_encoder_TUs,
num_decoder_TUs=num_decoder_TUs,
hidden_dims=hidden_dims)
# sd = np.load('lstm.npy', allow_pickle=True).tolist()
# lstm.set_state_dict(sd)
state_dict = layer.state_dict()
sd = OrderedDict()
for key, value in state_dict.items():
v = value.numpy()
# print(key, value.shape, np.sum(v), np.mean(v), np.max(v), np.min(v))
sd[key] = v
np.save('srnhead.npy', sd)
inputs = fluid.dygraph.to_variable(x)
x1 = fluid.dygraph.to_variable(x1)
x2 = fluid.dygraph.to_variable(x2)
x3 = fluid.dygraph.to_variable(x3)
x4 = fluid.dygraph.to_variable(x4)
out = layer(inputs, [x1,x2,x3,x4])
predict = out['predict']
pvam_feature = out['pvam_feature']
decoded_out = out['decoded_out']
word_out = out['word_out']
gsrm_out = out['gsrm_out']
return predict.numpy(), pvam_feature.numpy(), decoded_out.numpy(), word_out.numpy(), gsrm_out.numpy()
# return alpha.numpy()
def torch_func():
np.random.seed(SEED)
# x = np.load('input.npy', allow_pickle=True)
x = np.random.rand(input_shape[0], input_shape[1], input_shape[2], input_shape[3]).astype(np.float32)
print('pt input size: {}'.format(x.shape))
# x1 = np.random.rand(1, 256, 1).astype(np.float32)
x1 = np.load('encoder_word_pos_list.npy', allow_pickle=True)
# x2 = np.random.rand(1, 25, 1).astype(np.int)
x2 = np.load('gsrm_word_pos_list.npy', allow_pickle=True)
# x3 = np.random.rand(1, 8, 25, 25).astype(np.float32)
x3 = np.load('gsrm_slf_attn_bias1_list.npy', allow_pickle=True)
# x4 = np.random.rand(1, 8, 25, 25).astype(np.float32)
x4 = np.load('gsrm_slf_attn_bias2_list.npy', allow_pickle=True)
inputs = torch.Tensor(x)
x1 = torch.Tensor(x1)
x2 = torch.Tensor(x2)
x3 = torch.Tensor(x3)
x4 = torch.Tensor(x4)
layer = PTHead(in_channels=in_channels,
out_channels=out_channels,
max_text_length=max_text_length,
num_heads=num_heads,
num_encoder_TUs=num_encoder_TUs,
num_decoder_TUs=num_decoder_TUs,
hidden_dims=hidden_dims)
sd = np.load('srnhead.npy', allow_pickle=True)
sd = sd.tolist()
for key, value in layer.state_dict().items():
name = key
keyword = 'block_list.'
if keyword in name:
# replace: block_list.
name = name.replace(keyword, '')
else:
name = name
# for srn
keyword = 'base_block.'
if keyword in name:
# replace: base_block.
name = name.replace(keyword, '')
keyword = 'base_block_2.0.'
if keyword in name:
# replace: base_block_2.0. -> base_block_2.
name = name.replace(keyword, 'base_block_2.')
# for srn head
keyword = 'encoder_layers.'
if keyword in name:
# replace: encoder_layers.
name = name.replace(keyword, '')
keyword = 'functors.'
if keyword in name:
# replace: functors.
name = name.replace(keyword, '')
if name.endswith('num_batches_tracked'):
continue
if name.endswith('running_mean'):
ppname = name.replace('running_mean', '_mean')
elif name.endswith('running_var'):
ppname = name.replace('running_var', '_variance')
elif name.endswith('bias') or name.endswith('weight'):
ppname = name
elif 'lstm' in name:
ppname = name
elif 'attention_cell' in name:
ppname = name
else:
print('Redundance:')
print(name)
raise ValueError
try:
if key.endswith('.weight'):
if len(sd[ppname].shape) == len(layer.state_dict()[key].shape) == 2 \
and sd[ppname].shape[0] == layer.state_dict()[key].shape[1] \
and sd[ppname].shape[1] == layer.state_dict()[key].shape[0]:
layer.state_dict()[key].copy_(torch.Tensor(sd[ppname].T))
else:
layer.state_dict()[key].copy_(torch.Tensor(sd[ppname]))
else:
layer.state_dict()[key].copy_(torch.Tensor(sd[ppname]))
except Exception as e:
print('except: pt: ', key)
print('except: pp: ', ppname)
print('except: pt: ', layer.state_dict()[key].shape)
print('except: pp: ', sd[ppname].shape)
raise e
out = layer(inputs, [x1,x2,x3,x4])
predict = out['predict']
pvam_feature = out['pvam_feature']
decoded_out = out['decoded_out']
word_out = out['word_out']
gsrm_out = out['gsrm_out']
return predict.data.numpy(), pvam_feature.data.numpy(), decoded_out.data.numpy(), word_out.data.numpy(), gsrm_out.data.numpy()
# return alpha.data.numpy()
def print_cmp(inp, name=None):
print('{}: shape-{}, sum: {}, mean: {}, max: {}, min: {}'.format(name, inp.shape,
np.sum(inp), np.mean(inp),
np.max(inp), np.min(inp)))
if __name__ == '__main__':
print('==========paddle=================')
predict, pvam_feature, decoded_out, word_out, gsrm_out = paddle_func()
print_cmp(predict, name='predict')
print('===========pytorch================')
predict, pvam_feature, decoded_out, word_out, gsrm_out = torch_func()
print_cmp(predict, name='predict')