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
|
from caffe2.proto import caffe2_pb2
from caffe2.python import model_helper, workspace, core, rnn_cell, test_util
from caffe2.python.attention import AttentionType
import numpy as np
import unittest
import caffe2.python.hypothesis_test_util as hu
import hypothesis.strategies as st
from hypothesis import given, settings
class TestRNNExecutor(test_util.TestCase):
def setUp(self):
super(TestRNNExecutor, self).setUp()
self.batch_size = 8
self.input_dim = 20
self.hidden_dim = 30
self.encoder_dim = 40
@given(
T=st.integers(10, 100),
forward_only=st.booleans(),
**hu.gcs)
@settings(deadline=10000)
def test_lstm_with_attention_equal_simplenet(self, T, forward_only, gc, dc):
self.Tseq = [T, T // 2, T // 2 + T // 4, T, T // 2 + 1]
workspace.ResetWorkspace()
with core.DeviceScope(gc):
print("Run with device: {}, forward only: {}".format(
gc, forward_only))
workspace.FeedBlob(
"seq_lengths",
np.array([T] * self.batch_size, dtype=np.int32)
)
workspace.FeedBlob("target", np.random.rand(
T, self.batch_size, self.hidden_dim).astype(np.float32))
workspace.FeedBlob("hidden_init", np.zeros(
[1, self.batch_size, self.hidden_dim], dtype=np.float32
))
workspace.FeedBlob("cell_init", np.zeros(
[1, self.batch_size, self.hidden_dim], dtype=np.float32
))
model = model_helper.ModelHelper(name="lstm")
model.net.AddExternalInputs(["input"])
init_blobs = []
hidden_init, cell_init, encoder_outputs = model.net.AddExternalInputs(
"hidden_init",
"cell_init",
"encoder_outputs"
)
awec_init = model.net.AddExternalInputs([
'initial_attention_weighted_encoder_context',
])
init_blobs.extend([hidden_init, cell_init])
workspace.FeedBlob(
awec_init,
np.random.rand(1, self.batch_size, self.encoder_dim).astype(
np.float32),
)
workspace.FeedBlob(
encoder_outputs,
np.random.rand(1, self.batch_size, self.encoder_dim).astype(
np.float32),
)
outputs = rnn_cell.LSTMWithAttention(
model=model,
decoder_inputs="input",
decoder_input_lengths="seq_lengths",
initial_decoder_hidden_state=hidden_init,
initial_decoder_cell_state=cell_init,
initial_attention_weighted_encoder_context=awec_init,
encoder_output_dim=self.encoder_dim,
encoder_outputs=encoder_outputs,
encoder_lengths=None,
decoder_input_dim=self.input_dim,
decoder_state_dim=self.hidden_dim,
scope="",
attention_type=AttentionType.Recurrent,
forward_only=forward_only,
outputs_with_grads=[0],
)
output = outputs[0]
print(outputs)
loss = model.AveragedLoss(
model.SquaredL2Distance([output, "target"], "dist"),
"loss"
)
# Add gradient ops
if not forward_only:
model.AddGradientOperators([loss])
# init
for init_blob in init_blobs:
workspace.FeedBlob(init_blob, np.zeros(
[1, self.batch_size, self.hidden_dim], dtype=np.float32
))
self._compare(model, forward_only)
def init_lstm_model(self, T, num_layers, forward_only, use_loss=True):
workspace.FeedBlob(
"seq_lengths",
np.array([T] * self.batch_size, dtype=np.int32)
)
workspace.FeedBlob("target", np.random.rand(
T, self.batch_size, self.hidden_dim).astype(np.float32))
workspace.FeedBlob("hidden_init", np.zeros(
[1, self.batch_size, self.hidden_dim], dtype=np.float32
))
workspace.FeedBlob("cell_init", np.zeros(
[1, self.batch_size, self.hidden_dim], dtype=np.float32
))
model = model_helper.ModelHelper(name="lstm")
model.net.AddExternalInputs(["input"])
init_blobs = []
for i in range(num_layers):
hidden_init, cell_init = model.net.AddExternalInputs(
"hidden_init_{}".format(i),
"cell_init_{}".format(i)
)
init_blobs.extend([hidden_init, cell_init])
output, last_hidden, _, last_state = rnn_cell.LSTM(
model=model,
input_blob="input",
seq_lengths="seq_lengths",
initial_states=init_blobs,
dim_in=self.input_dim,
dim_out=[self.hidden_dim] * num_layers,
scope="",
drop_states=True,
forward_only=forward_only,
return_last_layer_only=True,
)
if use_loss:
loss = model.AveragedLoss(
model.SquaredL2Distance([output, "target"], "dist"),
"loss"
)
# Add gradient ops
if not forward_only:
model.AddGradientOperators([loss])
# init
for init_blob in init_blobs:
workspace.FeedBlob(init_blob, np.zeros(
[1, self.batch_size, self.hidden_dim], dtype=np.float32
))
return model, output
def test_empty_sequence(self):
'''
Test the RNN executor's handling of empty input sequences
'''
Tseq = [0, 1, 2, 3, 0, 1]
workspace.ResetWorkspace()
with core.DeviceScope(caffe2_pb2.DeviceOption()):
model, output = self.init_lstm_model(
T=4, num_layers=1, forward_only=True, use_loss=False)
workspace.RunNetOnce(model.param_init_net)
self.enable_rnn_executor(model.net, 1, True)
np.random.seed(10022015)
first_call = True
for seq_len in Tseq:
input_shape = [seq_len, self.batch_size, self.input_dim]
workspace.FeedBlob(
"input", np.random.rand(*input_shape).astype(np.float32))
workspace.FeedBlob(
"target",
np.random.rand(
seq_len, self.batch_size, self.hidden_dim
).astype(np.float32))
if first_call:
workspace.CreateNet(model.net, overwrite=True)
first_call = False
workspace.RunNet(model.net.Proto().name)
val = workspace.FetchBlob(output)
self.assertEqual(val.shape[0], seq_len)
@given(
num_layers=st.integers(1, 8),
T=st.integers(4, 100),
forward_only=st.booleans(),
**hu.gcs)
@settings(deadline=10000)
def test_lstm_equal_simplenet(self, num_layers, T, forward_only, gc, dc):
'''
Test that the RNN executor produces same results as
the non-executor (i.e running step nets as sequence of simple nets).
'''
self.Tseq = [T, T // 2, T // 2 + T // 4, T, T // 2 + 1]
workspace.ResetWorkspace()
with core.DeviceScope(gc):
print("Run with device: {}, forward only: {}".format(
gc, forward_only))
model, _ = self.init_lstm_model(T, num_layers, forward_only)
self._compare(model, forward_only)
def _compare(self, model, forward_only):
# Store list of blobs that exist in the beginning
workspace.RunNetOnce(model.param_init_net)
init_ws = {k: workspace.FetchBlob(k) for k in workspace.Blobs()}
# Run with executor
for enable_executor in [0, 1]:
self.enable_rnn_executor(model.net, enable_executor, forward_only)
workspace.ResetWorkspace()
# Reset original state
for k, v in init_ws.items():
workspace.FeedBlob(k, v)
np.random.seed(10022015)
ws = {}
for j in range(len(self.Tseq)):
input_shape = [self.Tseq[j], self.batch_size, self.input_dim]
workspace.FeedBlob(
"input", np.random.rand(*input_shape).astype(np.float32))
workspace.FeedBlob(
"target",
np.random.rand(
self.Tseq[j], self.batch_size, self.hidden_dim
).astype(np.float32))
if j == 0:
workspace.CreateNet(model.net, overwrite=True)
workspace.RunNet(model.net.Proto().name)
# Store results for each iteration
for k in workspace.Blobs():
ws[k + "." + str(j)] = workspace.FetchBlob(k)
if enable_executor:
rnn_exec_ws = ws
else:
non_exec_ws = ws
# Test that all blobs are equal after running with executor
# or without.
self.assertEqual(list(non_exec_ws.keys()), list(rnn_exec_ws.keys()))
mismatch = False
for k in rnn_exec_ws.keys():
non_exec_v = non_exec_ws[k]
rnn_exec_v = rnn_exec_ws[k]
if type(non_exec_v) is np.ndarray:
if not np.allclose(non_exec_v, rnn_exec_v):
print("Mismatch: {}".format(k))
nv = non_exec_v.flatten()
rv = rnn_exec_v.flatten()
c = 0
for j in range(len(nv)):
if rv[j] != nv[j]:
print(j, rv[j], nv[j])
c += 1
if c == 10:
break
mismatch = True
self.assertFalse(mismatch)
def enable_rnn_executor(self, net, value, forward_only):
num_found = 0
for op in net.Proto().op:
if op.type.startswith("RecurrentNetwork"):
for arg in op.arg:
if arg.name == 'enable_rnn_executor':
arg.i = value
num_found += 1
# This sanity check is so that if someone changes the
# enable_rnn_executor parameter name, the test will
# start failing as this function will become defective.
self.assertEqual(1 if forward_only else 2, num_found)
if __name__ == "__main__":
import random
random.seed(2603)
workspace.GlobalInit([
'caffe2',
'--caffe2_log_level=0',
'--caffe2_rnn_executor=1'])
unittest.main()
|
