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
|
import caffe2.python.hypothesis_test_util as hu
import hypothesis.strategies as st
import numpy as np
import numpy.testing as npt
from caffe2.python import core, layer_model_instantiator, regularizer, schema, workspace
from caffe2.python.layer_test_util import LayersTestCase
from caffe2.python.optimizer import SgdOptimizer
from caffe2.python.regularizer import L1Norm, RegularizationBy
from caffe2.python.regularizer_context import RegularizerContext, UseRegularizer
from hypothesis import given
class TestRegularizerContext(LayersTestCase):
@given(X=hu.arrays(dims=[2, 5]))
def test_regularizer_context(self, X):
weight_reg_out = L1Norm(0.2)
bias_reg_out = L1Norm(0)
regularizers = {"WEIGHT": weight_reg_out, "BIAS": bias_reg_out}
output_dims = 2
input_record = self.new_record(schema.Scalar((np.float32, (5,))))
schema.FeedRecord(input_record, [X])
with UseRegularizer(regularizers):
weight_reg = RegularizerContext.current().get_regularizer("WEIGHT")
bias_reg = RegularizerContext.current().get_regularizer("BIAS")
optim = SgdOptimizer(0.15)
assert (
weight_reg == weight_reg_out
), "fail to get correct weight reg from context"
assert bias_reg == bias_reg_out, "fail to get correct bias reg from context"
fc_output = self.model.FC(
input_record,
output_dims,
weight_optim=optim,
bias_optim=optim,
weight_reg=weight_reg,
bias_reg=bias_reg,
)
# model.output_schema has to a struct
self.model.output_schema = schema.Struct(("fc_output", fc_output))
self.assertEqual(schema.Scalar((np.float32, (output_dims,))), fc_output)
_, train_net = layer_model_instantiator.generate_training_nets(self.model)
ops = train_net.Proto().op
ops_type_list = [ops[i].type for i in range(len(ops))]
assert ops_type_list.count("LpNorm") == 2
assert ops_type_list.count("Scale") == 4
assert ops_type_list.count("LpNormGradient") == 2
class TestRegularizer(LayersTestCase):
@given(X=hu.arrays(dims=[2, 5], elements=hu.floats(min_value=-1.0, max_value=1.0)))
def test_log_barrier(self, X):
param = core.BlobReference("X")
workspace.FeedBlob(param, X)
train_init_net, train_net = self.get_training_nets()
reg = regularizer.LogBarrier(1.0)
output = reg(train_net, train_init_net, param, by=RegularizationBy.ON_LOSS)
reg(
train_net,
train_init_net,
param,
grad=None,
by=RegularizationBy.AFTER_OPTIMIZER,
)
workspace.RunNetOnce(train_init_net)
workspace.RunNetOnce(train_net)
def ref(X):
return (
np.array(np.sum(-np.log(np.clip(X, 1e-9, None))) * 0.5).astype(
np.float32
),
np.clip(X, 1e-9, None),
)
for x, y in zip(workspace.FetchBlobs([output, param]), ref(X)):
npt.assert_allclose(x, y, rtol=1e-3)
@given(
X=hu.arrays(dims=[2, 5], elements=hu.floats(min_value=-1.0, max_value=1.0)),
left_open=st.booleans(),
right_open=st.booleans(),
eps=hu.floats(min_value=1e-6, max_value=1e-4),
ub=hu.floats(min_value=-1.0, max_value=1.0),
lb=hu.floats(min_value=-1.0, max_value=1.0),
**hu.gcs_cpu_only
)
def test_bounded_grad_proj(self, X, left_open, right_open, eps, ub, lb, gc, dc):
if ub - (eps if right_open else 0.) < lb + (eps if left_open else 0.):
return
param = core.BlobReference("X")
workspace.FeedBlob(param, X)
train_init_net, train_net = self.get_training_nets()
reg = regularizer.BoundedGradientProjection(
lb=lb, ub=ub, left_open=left_open, right_open=right_open, epsilon=eps
)
output = reg(train_net, train_init_net, param, by=RegularizationBy.ON_LOSS)
reg(
train_net,
train_init_net,
param,
grad=None,
by=RegularizationBy.AFTER_OPTIMIZER,
)
workspace.RunNetOnce(train_init_net)
workspace.RunNetOnce(train_net)
def ref(X):
return np.clip(
X, lb + (eps if left_open else 0.), ub - (eps if right_open else 0.)
)
assert output is None
npt.assert_allclose(workspace.blobs[param], ref(X), atol=1e-7)
@given(
output_dim=st.integers(1, 10),
input_num=st.integers(3, 30),
reg_weight=st.integers(0, 10)
)
def test_group_l1_norm(self, output_dim, input_num, reg_weight):
"""
1. create a weight blob
2. create random group splits
3. run group_l1_nrom with the weight blob
4. run equivalent np operations to calculate group l1 norm
5. compare if the results from 3 and 4 are equal
"""
def compare_reference(weight, group_boundaries, reg_lambda, output):
group_splits = np.hsplit(weight, group_boundaries[1:-1])
l2_reg = np.sqrt([np.sum(np.square(g)) for g in group_splits])
l2_normalized = np.multiply(l2_reg,
np.array([np.sqrt(g.shape[1]) for g in group_splits]))
result = np.multiply(np.sum(l2_normalized), reg_lambda)
npt.assert_almost_equal(result, workspace.blobs[output], decimal=2)
weight = np.random.rand(output_dim, input_num).astype(np.float32)
feature_num = np.random.randint(low=1, high=input_num - 1)
group_boundaries = [0]
group_boundaries = np.append(
group_boundaries,
np.sort(
np.random.choice(range(1, input_num - 1), feature_num, replace=False)
),
)
group_boundaries = np.append(group_boundaries, [input_num])
split_info = np.diff(group_boundaries)
weight_blob = core.BlobReference("weight_blob")
workspace.FeedBlob(weight_blob, weight)
train_init_net, train_net = self.get_training_nets()
reg = regularizer.GroupL1Norm(reg_weight * 0.1, split_info.tolist())
output = reg(
train_net, train_init_net, weight_blob, by=RegularizationBy.ON_LOSS
)
workspace.RunNetOnce(train_init_net)
workspace.RunNetOnce(train_net)
compare_reference(weight, group_boundaries, reg_weight * 0.1, output)
@given(
param_dim=st.integers(10, 30),
k=st.integers(5, 9),
reg_weight=st.integers(0, 10)
)
def test_l1_norm_trimmed(self, param_dim, k, reg_weight):
weight = np.random.rand(param_dim).astype(np.float32)
weight_blob = core.BlobReference("weight_blob")
workspace.FeedBlob(weight_blob, weight)
train_init_net, train_net = self.get_training_nets()
reg = regularizer.L1NormTrimmed(reg_weight * 0.1, k)
output = reg(
train_net, train_init_net, weight_blob, by=RegularizationBy.ON_LOSS
)
workspace.RunNetOnce(train_init_net)
workspace.RunNetOnce(train_net)
result = np.sum(np.sort(np.absolute(weight))[:(param_dim - k)]) * reg_weight * 0.1
npt.assert_almost_equal(result, workspace.blobs[output], decimal=2)
@given(
param_dim=st.integers(10, 30),
k=st.integers(5, 9),
l1=st.integers(0, 10),
l2=st.integers(0, 10)
)
def test_elastic_l1_norm_trimmed(self, param_dim, k, l1, l2):
weight = np.random.rand(param_dim).astype(np.float32)
weight_blob = core.BlobReference("weight_blob")
workspace.FeedBlob(weight_blob, weight)
train_init_net, train_net = self.get_training_nets()
reg = regularizer.ElasticNetL1NormTrimmed(l1 * 0.1, l2 * 0.1, k)
output = reg(
train_net, train_init_net, weight_blob, by=RegularizationBy.ON_LOSS
)
workspace.RunNetOnce(train_init_net)
workspace.RunNetOnce(train_net)
l1_norm = np.sum(np.sort(np.absolute(weight))[:(param_dim - k)])
l2_norm = np.sum(np.square(weight))
result = l1_norm * l1 * 0.1 + l2_norm * l2 * 0.1
npt.assert_almost_equal(result, workspace.blobs[output], decimal=2)
@given(
row_dim=st.integers(5, 10),
norm=st.floats(min_value=1.0, max_value=4.0),
data_strategy=st.data(),
)
def test_fp16_max_norm(self, row_dim, norm, data_strategy):
weight = np.random.rand(row_dim, 5).astype(np.float16)
grad = np.random.rand(row_dim, 5).astype(np.float16)
# generate indices that will be updated
indices = data_strategy.draw(
hu.tensor(
dtype=np.int64,
min_dim=1,
max_dim=1,
elements=st.sampled_from(np.arange(weight.shape[0])),
)
)
indices = np.unique(indices)
# compute expected result
result = weight.copy()
# prevent dived by zero
eps = 1e-12
norms = np.sqrt(np.sum(result[indices, ] ** 2, axis=1, keepdims=True))
# if the norms are smaller than max_norm, then it doesn't need update
desired = np.clip(norms, 0, norm)
# apply max norm
result[indices, ] *= desired / (eps + norms)
weight_blob = core.BlobReference("weight_blob")
workspace.FeedBlob(weight_blob, weight)
grad_blob = core.BlobReference("grad_blob")
workspace.FeedBlob(grad_blob, grad)
indices_blob = core.BlobReference("indices")
workspace.FeedBlob(indices_blob, indices)
grad_blob_slice = core.GradientSlice(indices=indices_blob, values=grad_blob)
train_init_net, train_net = self.get_training_nets()
reg = regularizer.MaxNorm(norm, dtype='fp16')
reg(
train_net, train_init_net, weight_blob, grad_blob_slice, by=RegularizationBy.AFTER_OPTIMIZER
)
workspace.RunNetOnce(train_init_net)
workspace.RunNetOnce(train_net)
npt.assert_almost_equal(result, workspace.FetchBlob('weight_blob'), decimal=2)
|
