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
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
|
# Owner(s): ["oncall: quantization"]
import copy
import torch
import torch.nn as nn
import torch.ao.nn.quantized as nnq
import torch.nn.intrinsic as nni
import torch.nn.intrinsic.quantized as nniq
import torch.nn.intrinsic.qat as nniqat
from torch.ao.quantization import (
quantize,
prepare,
convert,
prepare_qat,
quantize_qat,
fuse_modules,
fuse_modules_qat,
QConfig,
default_qconfig,
default_qat_qconfig,
)
from torch.testing._internal.common_quantization import (
QuantizationTestCase,
ModelForFusion,
ModelWithSequentialFusion,
ModelForLinearBNFusion,
ModelForFusionWithBias,
ModelForConvTransposeBNFusion,
test_only_eval_fn,
test_only_train_fn,
skipIfNoFBGEMM,
)
from torch.testing._internal.common_quantized import (
override_quantized_engine,
supported_qengines,
)
@skipIfNoFBGEMM
class TestFuseEager(QuantizationTestCase):
def test_fuse_module_train(self):
model = ModelForFusion(default_qat_qconfig).train()
# Test step by step fusion
model = fuse_modules_qat(model, ['conv1', 'bn1', 'relu1'])
model = fuse_modules_qat(model, ['sub1.conv', 'sub1.bn'])
self.assertEqual(type(model.conv1), nni.ConvBnReLU2d,
msg="Fused Conv + BN + Relu first layer")
self.assertEqual(type(model.bn1), torch.nn.Identity,
msg="Fused Conv + BN + Relu (skipped BN)")
self.assertEqual(type(model.relu1), torch.nn.Identity,
msg="Fused Conv + BN + Relu (skipped Relu)")
self.assertEqual(type(model.sub1.conv), nni.ConvBn2d,
msg="Fused submodule Conv + BN")
self.assertEqual(type(model.sub1.bn), torch.nn.Identity,
msg="Fused submodule Conv + BN (skipped BN)")
self.assertEqual(type(model.sub2.conv), torch.nn.Conv2d,
msg="Non-fused submodule Conv")
self.assertEqual(type(model.sub2.relu), torch.nn.ReLU,
msg="Non-fused submodule ReLU")
model = prepare_qat(model)
self.checkObservers(model)
def checkQAT(model):
self.assertEqual(type(model.conv1), nniqat.ConvBnReLU2d)
self.assertEqual(type(model.bn1), nn.Identity)
self.assertEqual(type(model.relu1), nn.Identity)
self.assertEqual(type(model.sub1.conv), nniqat.ConvBn2d)
self.assertEqual(type(model.sub1.bn), nn.Identity)
self.assertEqual(type(model.sub2.conv), nn.Conv2d)
self.assertEqual(type(model.sub2.relu), nn.ReLU)
checkQAT(model)
test_only_train_fn(model, self.img_data_1d_train)
model = convert(model)
def checkQuantized(model):
self.assertEqual(type(model.conv1), nniq.ConvReLU2d)
self.assertEqual(type(model.bn1), nn.Identity)
self.assertEqual(type(model.relu1), nn.Identity)
self.assertEqual(type(model.sub1.conv), nnq.Conv2d)
self.assertEqual(type(model.sub1.bn), nn.Identity)
self.assertEqual(type(model.sub2.conv), nn.Conv2d)
self.assertEqual(type(model.sub2.relu), nn.ReLU)
test_only_eval_fn(model, self.img_data_1d)
self.checkNoQconfig(model)
with self.assertRaisesRegex(RuntimeError, "Could not run 'aten::native_batch_norm' with arguments from the 'QuantizedCPU'"):
checkQuantized(model)
model = ModelForFusion(default_qat_qconfig).train()
model = fuse_modules_qat(
model,
[['conv1', 'bn1', 'relu1'],
['sub1.conv', 'sub1.bn']])
model = quantize_qat(model, test_only_train_fn, [self.img_data_1d_train])
with self.assertRaisesRegex(RuntimeError, "Could not run 'aten::native_batch_norm' with arguments from the 'QuantizedCPU'"):
checkQuantized(model)
def test_fuse_module_eval(self):
model = ModelForFusion(default_qconfig)
model.eval()
model = fuse_modules(
model,
[['conv3', 'bn3', 'relu4'],
['conv1', 'bn1', 'relu1'],
['conv2', 'relu2'],
['bn2', 'relu3'],
['sub1.conv', 'sub1.bn']])
self.assertEqual(type(model.conv1), nni.ConvReLU2d,
msg="Fused Conv + BN + Relu first layer (BN is folded)")
self.assertEqual(type(model.conv1[0]), nn.Conv2d,
msg="Fused Conv + BN + Relu (Conv + folded BN only)")
self.assertEqual(type(model.conv1[1]), nn.ReLU,
msg="Fused Conv + BN + Relu second layer (Relu only)")
self.assertEqual(type(model.bn1), nn.Identity,
msg="Fused Conv + BN + Relu second layer (Skipped BN)")
self.assertEqual(type(model.relu1), nn.Identity,
msg="Fused Conv + BN + Relu second layer (Skipped Relu)")
self.assertEqual(type(model.conv2), nni.ConvReLU3d,
msg="Fused Conv + BN + Relu first layer (BN is folded)")
self.assertEqual(type(model.bn2), nni.BNReLU3d,
msg="Fused BN + Relu first layer (Relu is folded))")
self.assertEqual(type(model.relu3), nn.Identity,
msg="Fused BN + Relu second layer (Skipped Relu)")
self.assertEqual(type(model.conv2[0]), nn.Conv3d,
msg="Fused Conv + BN + Relu (Conv + folded BN only)")
self.assertEqual(type(model.conv2[1]), nn.ReLU,
msg="Fused Conv + BN + Relu second layer (Relu only)")
self.assertEqual(type(model.relu2), nn.Identity,
msg="Fused Conv + BN + Relu second layer (Skipped Relu)")
self.assertEqual(type(model.conv3), nni.ConvReLU1d,
msg="Fused Conv + Relu for Conv1d (folded BN)")
self.assertEqual(type(model.conv3[0]), nn.Conv1d,
msg="Fused Conv + Relu for Conv1d ")
self.assertEqual(type(model.conv3[1]), nn.ReLU,
msg="Fused Conv + Relu for Conv1d")
self.assertEqual(type(model.bn3), nn.Identity,
msg="Fused Conv + BN + Relu for Conv1d (Skipped BN)")
self.assertEqual(type(model.sub1.conv), nn.Conv2d,
msg="Fused submodule Conv + folded BN")
self.assertEqual(type(model.sub1.bn), nn.Identity,
msg="Fused submodule (skipped BN)")
self.assertEqual(type(model.sub2.conv), nn.Conv2d,
msg="Non-fused submodule Conv")
self.assertEqual(type(model.sub2.relu), torch.nn.ReLU,
msg="Non-fused submodule ReLU")
model = prepare(model)
self.checkObservers(model)
test_only_eval_fn(model, self.img_data_1d)
model = convert(model)
def checkQuantized(model):
self.assertEqual(type(model.conv3), nniq.ConvReLU1d)
self.assertEqual(type(model.conv1), nniq.ConvReLU2d)
self.assertEqual(type(model.bn1), nn.Identity)
self.assertEqual(type(model.relu1), nn.Identity)
self.assertEqual(type(model.sub1.conv), nnq.Conv2d)
self.assertEqual(type(model.sub1.bn), nn.Identity)
self.assertEqual(type(model.sub2.conv), nn.Conv2d)
self.assertEqual(type(model.sub2.relu), nn.ReLU)
self.assertEqual(type(model.bn2), nniq.BNReLU3d)
test_only_eval_fn(model, self.img_data_1d)
self.checkNoQconfig(model)
checkQuantized(model)
model = ModelForFusion(default_qconfig).eval()
model = fuse_modules(
model,
[['conv1', 'bn1', 'relu1'],
['conv2', 'relu2'],
['bn2', 'relu3'],
['sub1.conv', 'sub1.bn'],
['conv3', 'bn3', 'relu4']])
model = quantize(model, test_only_eval_fn, [self.img_data_1d])
checkQuantized(model)
def test_fusion_sequential_model_train(self):
for qengine in supported_qengines:
with override_quantized_engine(qengine):
model = ModelWithSequentialFusion().train()
model.to(torch.float)
fuse_modules_qat(
model, [['conv1', 'relu1'] ,
['features.0.0', 'features.0.1', 'features.0.2'],
['features.1.0', 'features.1.1', 'features.1.2'],
['features.2.0', 'features.2.1', 'features.2.2'],
['classifier.0', 'classifier.1']],
inplace=True)
self.assertEqual(type(model.conv1), nni.ConvReLU2d,
msg="Fused Conv + Relu: nni.ConvReLU2d")
self.assertEqual(type(model.conv1[0]), nn.Conv2d,
msg="Fused Conv + Relu: Conv2d")
self.assertEqual(type(model.conv1[1]), nn.ReLU,
msg="Fused Conv + Relu: Relu")
self.assertEqual(type(model.relu1), nn.Identity,
msg="Fused Conv + Relu: Identity")
for i in range(3):
self.assertEqual(type(model.features[i][0]), nni.ConvBnReLU2d,
msg="Fused submodule Conv + folded BN")
self.assertEqual(type(model.features[i][1]), nn.Identity,
msg="Fused submodule (skipped BN)")
self.assertEqual(type(model.features[i][2]), nn.Identity,
msg="Non-fused submodule Conv")
self.assertEqual(type(model.classifier[0]), nni.LinearReLU)
self.assertEqual(type(model.classifier[1]), nn.Identity)
model.qconfig = torch.ao.quantization.get_default_qat_qconfig(qengine)
prepare_qat(model, inplace=True)
self.checkObservers(model)
model(self.img_data_2d[0][0])
def checkQAT(model):
self.assertEqual(type(model.conv1), nniqat.ConvReLU2d)
self.assertEqual(type(model.relu1), nn.Identity)
for i in range(3):
self.assertEqual(type(model.features[i][0]), nniqat.ConvBnReLU2d,
msg="Fused submodule Conv + folded BN")
self.assertEqual(type(model.features[i][1]), nn.Identity,
msg="Fused submodule (skipped BN)")
self.assertEqual(type(model.features[i][2]), nn.Identity,
msg="Non-fused submodule Conv")
self.assertEqual(type(model.classifier[0]), nniqat.LinearReLU)
self.assertEqual(type(model.classifier[1]), nn.Identity)
checkQAT(model)
model(self.img_data_2d[1][0])
convert(model, inplace=True)
model(self.img_data_2d[1][0])
self.checkModelWithSequentialQuantized(model)
def test_fusion_sequential_model_eval(self):
for qengine in supported_qengines:
with override_quantized_engine(qengine):
model = ModelWithSequentialFusion().eval()
model.to(torch.float)
fuse_modules(
model,
[['conv1', 'relu1'],
['features.0.0', 'features.0.1', 'features.0.2'],
['features.1.0', 'features.1.1', 'features.1.2'],
['features.2.0', 'features.2.1', 'features.2.2'],
['classifier.0', 'classifier.1']],
inplace=True)
self.assertEqual(type(model.conv1), nni.ConvReLU2d,
msg="Fused Conv + Relu: nni.ConvReLU2d")
self.assertEqual(type(model.conv1[0]), nn.Conv2d,
msg="Fused Conv + Relu: Conv2d")
self.assertEqual(type(model.conv1[1]), nn.ReLU,
msg="Fused Conv + Relu: Relu")
self.assertEqual(type(model.relu1), nn.Identity,
msg="Fused Conv + Relu: Identity")
for i in range(3):
self.assertEqual(type(model.features[i][0]), nni.ConvReLU2d,
msg="Fused submodule Conv + folded BN")
self.assertEqual(type(model.features[i][1]), nn.Identity,
msg="Fused submodule (skipped BN)")
self.assertEqual(type(model.features[i][2]), nn.Identity,
msg="Non-fused submodule Conv")
self.assertEqual(type(model.classifier[0]), nni.LinearReLU)
self.assertEqual(type(model.classifier[1]), nn.Identity)
model.qconfig = torch.ao.quantization.get_default_qconfig(qengine)
prepare(model, inplace=True)
self.checkObservers(model)
model(self.img_data_2d[0][0])
convert(model, inplace=True)
model(self.img_data_2d[1][0])
self.checkModelWithSequentialQuantized(model)
def checkModelWithSequentialQuantized(self, model):
self.assertEqual(type(model.conv1), nniq.ConvReLU2d)
self.assertEqual(type(model.relu1), nn.Identity)
for i in range(3):
self.assertEqual(type(model.features[i][0]), nniq.ConvReLU2d)
self.assertEqual(type(model.features[i][1]), nn.Identity)
self.assertEqual(type(model.features[i][2]), nn.Identity)
self.assertEqual(type(model.classifier[0]), nniq.LinearReLU)
self.assertEqual(type(model.classifier[1]), nn.Identity)
def test_fusion_conv_with_bias(self):
for qengine in supported_qengines:
with override_quantized_engine(qengine):
model_orig = ModelForFusionWithBias().train()
# reference model
model_ref = copy.deepcopy(model_orig)
# output with no fusion.
out_ref = model_ref(self.img_data_2d[0][0])
# fused model
model_orig.qconfig = QConfig(activation=torch.nn.Identity,
weight=torch.nn.Identity)
model = fuse_modules_qat(
model_orig,
[["conv1", "bn1", "relu1"],
["conv2", "bn2"]])
prep_model = prepare_qat(model, inplace=False)
# output with fusion but no observers.
out_fused = prep_model(self.img_data_2d[0][0])
self.assertEqual(out_ref, out_fused)
def checkBN(bn_ref, bn):
self.assertEqual(bn_ref.weight, bn.weight)
self.assertEqual(bn_ref.bias, bn.bias)
self.assertEqual(bn_ref.running_mean, bn.running_mean)
self.assertEqual(bn_ref.running_var, bn.running_var)
checkBN(model_ref.bn1, prep_model.conv1.bn)
checkBN(model_ref.bn2, prep_model.conv2.bn)
model.qconfig = torch.ao.quantization.get_default_qconfig(qengine)
prepare_qat(model, inplace=True)
model(self.img_data_2d[0][0])
def checkQAT(model):
self.assertEqual(type(model.conv1), nniqat.ConvBnReLU2d)
self.assertEqual(type(model.bn1), nn.Identity)
self.assertEqual(type(model.relu1), nn.Identity)
self.assertEqual(type(model.conv2), nniqat.ConvBn2d)
self.assertEqual(type(model.bn2), nn.Identity)
checkQAT(model)
def test_fusion_linear_bn_eval(self):
model = ModelForLinearBNFusion().train()
inp1 = torch.randn(8, 20)
inp2 = torch.randn(8, 20)
# Get some interesting values into the running mean and variance.
model(inp1)
model.eval()
golden = model(inp2)
model = fuse_modules(model, [["fc", "bn"]])
self.assertEqual(type(model.bn), nn.Identity)
self.assertEqual(golden, model(inp2))
def test_fusion_convtranspose_bn_eval(self):
model = ModelForConvTransposeBNFusion().train()
inp1 = torch.randn(8, 3, 16)
inp2 = torch.randn(8, 3, 16)
# Get some interesting values into the running mean and variance.
model(inp1)
model.eval()
golden = model(inp2)
model = fuse_modules(model, [["conv1", "bn1"], ["conv2", "bn2"], ["conv3", "bn3"]])
self.assertEqual(type(model.bn1), nn.Identity)
self.assertEqual(type(model.bn2), nn.Identity)
self.assertEqual(type(model.bn3), nn.Identity)
self.assertEqual(golden, model(inp2))
def test_forward_hooks_preserved(self):
r"""Test case that checks whether forward pre hooks of the first module and
post forward hooks of the last module in modules list passed to fusion function preserved.
(e.g. before fusion: [nn.Conv2d (with pre forward hooks), nn.BatchNorm2d, nn.ReLU (with post forward hooks)]
after fusion: [nni.ConvBnReLU2d (with pre and post hooks), nn.Identity, nn.Identity])
"""
model = ModelForFusion(default_qat_qconfig).train()
counter = {
'pre_forwards': 0,
'forwards': 0,
}
fused = False
def fw_pre_hook(fused_module_class, h_module, input):
if fused:
self.assertEqual(type(h_module), fused_module_class,
"After fusion owner of the first module's forward pre hook is not a fused module")
counter['pre_forwards'] += 1
def fw_hook(fused_module_class, h_module, input, output):
if fused:
self.assertEqual(type(h_module), fused_module_class,
"After fusion owner of the last module's forward hook is not a fused module")
counter['forwards'] += 1
# Registering two pre and two post forward hooks, thus expecting counter increment by two each inference
model.conv1.register_forward_pre_hook(lambda *args: fw_pre_hook(nni.ConvBnReLU2d, *args))
model.sub1.conv.register_forward_pre_hook(lambda *args: fw_pre_hook(nni.ConvBn2d, *args))
model.relu1.register_forward_hook(lambda *args: fw_hook(nni.ConvBnReLU2d, *args))
model.sub1.bn.register_forward_hook(lambda *args: fw_hook(nni.ConvBn2d, *args))
test_only_eval_fn(model, self.img_data_1d)
self.assertEqual(counter['pre_forwards'], 2 * len(self.img_data_1d))
self.assertEqual(counter['forwards'], 2 * len(self.img_data_1d))
model = fuse_modules_qat(model, ['conv1', 'bn1', 'relu1'])
model = fuse_modules_qat(model, ['sub1.conv', 'sub1.bn'])
fused = True
before_fusion_pre_count = counter['pre_forwards']
before_fusion_post_count = counter['forwards']
test_only_eval_fn(model, self.img_data_1d)
self.assertEqual(counter['pre_forwards'] - before_fusion_pre_count, 2 * len(self.img_data_1d))
self.assertEqual(counter['forwards'] - before_fusion_post_count, 2 * len(self.img_data_1d))
if __name__ == '__main__':
raise RuntimeError("This test file is not meant to be run directly, use:\n\n"
"\tpython test/test_quantization.py TESTNAME\n\n"
"instead.")
|
