• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1# Copyright 2021 Huawei Technologies Co., Ltd
2#
3# Licensed under the Apache License, Version 2.0 (the "License");
4# you may not use this file except in compliance with the License.
5# You may obtain a copy of the License at
6#
7# http://www.apache.org/licenses/LICENSE-2.0
8#
9# Unless required by applicable law or agreed to in writing, software
10# distributed under the License is distributed on an "AS IS" BASIS,
11# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
12# See the License for the specific language governing permissions and
13# limitations under the License.
14# ============================================================================
15"""Generate the mindir for bprop"""
16import os
17import numpy as np
18
19import mindspore.nn as nn
20from mindspore import Tensor, Parameter
21from mindspore.ops import operations as P
22import mindspore.ops as ops
23from mindspore.ops.operations import _inner_ops as inner
24import mindspore.common.dtype as mstype
25from mindspore.common.initializer import initializer
26import mindspore.ops._grad as g
27from mindspore.ops.bprop_mindir import serializable_bprop_ops
28from mindspore._c_expression import load_mindir
29
30
31class Net(nn.Cell):
32    def __init__(self, op):
33        super(Net, self).__init__()
34        self.op = op
35
36    def construct(self, *inputs, a=0, b=1):
37        c = a + b
38        return c, self.op(*inputs)
39
40
41class TupleInputNet(nn.Cell):
42    def __init__(self, op):
43        super(TupleInputNet, self).__init__()
44        self.op = op
45
46    def construct(self, x):
47        return self.op((x,))
48
49
50class GradNet(nn.Cell):
51    def __init__(self, network):
52        super(GradNet, self).__init__()
53        self.grad = ops.GradOperation(get_all=True)
54        self.network = network
55
56    def construct(self, *inputs):
57        gout = self.grad(self.network)(*inputs)
58        return gout
59
60
61def test_remove_mindir_dir():
62    bprop_path = g.__file__
63    bprop_installed_dir = bprop_path[: bprop_path.rindex('/')]
64    bprop_mindir_export_dir = bprop_installed_dir + "/../bprop_mindir"
65    os.rename(bprop_mindir_export_dir, bprop_mindir_export_dir + "_bak")
66    x = Tensor(np.array([[[[-1, 1, 10],
67                           [1, -1, 1],
68                           [10, 1, -1]]]]).astype(np.float32))
69    relu = Net(P.ReLU())
70    grad = GradNet(relu)
71    grad.compile(x)
72    os.rename(bprop_mindir_export_dir + "_bak", bprop_mindir_export_dir)
73
74
75def test_load_mindir_dir():
76    bprop_path = g.__file__
77    bprop_installed_dir = bprop_path[: bprop_path.rindex('/')]
78    bprop_mindir_export_dir = bprop_installed_dir + "/../bprop_mindir"
79    for op in serializable_bprop_ops:
80        file_name = bprop_mindir_export_dir + "/" + op.name + "_bprop.mindir"
81        graph = load_mindir(file_name)
82        assert not graph is None
83
84
85def test_relu():
86    x = Tensor(np.array([[[[-1, 1, 10],
87                           [1, -1, 1],
88                           [10, 1, -1]]]]).astype(np.float32))
89    relu = Net(P.ReLU())
90    grad = GradNet(relu)
91    grad.compile(x)
92
93
94def test_identity():
95    x = Tensor(np.array([1, 2, 3, 4]).astype(np.int64))
96    identity = Net(P.Identity())
97    grad = GradNet(identity)
98    grad.compile(x)
99
100
101def test_range():
102    x = Tensor(np.array([1, 2, 3, 2]).astype(np.int64))
103    range_net = Net(inner.Range(1.0, 8.0, 2.0))
104    grad = GradNet(range_net)
105    grad.compile(x)
106
107
108def test_ones_like():
109    x = Tensor(np.array([[0, 1], [2, 1]]).astype(np.int32))
110    ones_like = Net(P.OnesLike())
111    grad = GradNet(ones_like)
112    grad.compile(x)
113
114
115def test_zeros_like():
116    x = Tensor(np.array([[0, 1], [2, 1]]).astype(np.int32))
117    zeros_like = Net(P.ZerosLike())
118    grad = GradNet(zeros_like)
119    grad.compile(x)
120
121
122def test_argmax():
123    x = Tensor(np.array([[1, 20, 5], [67, 8, 9], [130, 24, 15]]).astype(np.float32))
124    argmax = Net(P.Argmax())
125    grad = GradNet(argmax)
126    grad.compile(x)
127
128
129def test_argmin():
130    x = Tensor(np.array([[1, 20, 5], [67, 8, 9], [130, 24, 15]]).astype(np.float32))
131    argmin = Net(P.Argmin())
132    grad = GradNet(argmin)
133    grad.compile(x)
134
135
136def test_broadcast():
137    x = Tensor(np.array([1, 2, 5, 2]).astype(np.float32))
138    broadcast = TupleInputNet(P.Broadcast(1))
139    grad = GradNet(broadcast)
140    grad.compile(x)
141
142
143def test_is_finite():
144    x = Tensor(np.ones([2, 4]).astype(np.int32))
145    is_finite = Net(P.IsFinite())
146    grad = GradNet(is_finite)
147    grad.compile(x)
148
149
150def test_approximate_equal():
151    x = Tensor(np.array([1, 2, 3]).astype(np.float32))
152    y = Tensor(np.array([2, 4, 6]).astype(np.float32))
153    approximate_equal = Net(P.ApproximateEqual(2.))
154    grad = GradNet(approximate_equal)
155    grad.compile(x, y)
156
157
158def test_logical_not():
159    x = Tensor(np.array([True, False, True]).astype(np.bool))
160    logical_not = Net(P.LogicalNot())
161    grad = GradNet(logical_not)
162    grad.compile(x)
163
164
165def test_sign():
166    x = Tensor(np.array([[2.0, 0.0, -1.0]]).astype(np.float32))
167    sign = Net(P.Sign())
168    grad = GradNet(sign)
169    grad.compile(x)
170
171
172def test_round():
173    x = Tensor(np.array([0.8, 1.5, 2.3, 2.5, -4.5]).astype(np.float32))
174    round_net = Net(P.Round())
175    grad = GradNet(round_net)
176    grad.compile(x)
177
178
179def test_lin_space():
180    start = Tensor(1, mstype.float32)
181    stop = Tensor(10, mstype.float32)
182    num = 5
183    lin_space = Net(P.LinSpace())
184    grad = GradNet(lin_space)
185    grad.compile(start, stop, num)
186
187
188def test_dropout_gen_mask():
189    x = (2, 4, 2, 2)
190    keep_prob = Tensor(1.0, mstype.float32)
191    dropout_gen_mask = Net(P.DropoutGenMask(10, 28))
192    grad = GradNet(dropout_gen_mask)
193    grad.compile(x, keep_prob)
194
195
196def test_onehot():
197    indices = Tensor(np.array([0, 1, 2]).astype(np.int32))
198    depth, on_value, off_value = 3, Tensor(1.0, mstype.float32), Tensor(0.0, mstype.float32)
199    one_hot = Net(P.OneHot())
200    grad = GradNet(one_hot)
201    grad.compile(indices, depth, on_value, off_value)
202
203
204def test_assign():
205    class AssignNet(nn.Cell):
206        def __init__(self):
207            super(AssignNet, self).__init__()
208            self.assign = P.Assign()
209            self.variable = Parameter(Tensor([1.0], mstype.float32), name="variable")
210
211        def construct(self, x):
212            return self.assign(self.variable, x)
213
214    value = Tensor([2.0], mstype.float32)
215    assign = AssignNet()
216    grad = GradNet(assign)
217    grad.compile(value)
218
219
220def test_assign_add():
221    class AssignAddNet(nn.Cell):
222        def __init__(self):
223            super(AssignAddNet, self).__init__()
224            self.assign_add = P.AssignAdd()
225            self.variable = Parameter(initializer(1, [1], mstype.int64), name="global_step")
226
227        def construct(self, x):
228            return self.assign_add(self.variable, x)
229
230    value = Tensor(np.ones([1]).astype(np.int64) * 100)
231    assign_add = AssignAddNet()
232    grad = GradNet(assign_add)
233    grad.compile(value)
234
235
236def test_assign_sub():
237    class AssignSubNet(nn.Cell):
238        def __init__(self):
239            super(AssignSubNet, self).__init__()
240            self.assign = P.AssignSub()
241            self.variable = Parameter(initializer(1, [1], mstype.int32), name="global_step")
242
243        def construct(self, x):
244            return self.assign(self.variable, x)
245
246    value = Tensor(np.ones([1]).astype(np.int32) * 100)
247    assign_sub = AssignSubNet()
248    grad = GradNet(assign_sub)
249    grad.compile(value)
250
251
252def test_iou():
253    anchor_boxes = Tensor(np.random.randint(1.0, 5.0, [3, 4]).astype(np.float16))
254    gt_boxes = Tensor(np.random.randint(1.0, 5.0, [3, 4]).astype(np.float16))
255    iou = Net(P.IOU())
256    grad = GradNet(iou)
257    grad.compile(anchor_boxes, gt_boxes)
258
259
260def test_bn_training_reduce():
261    x = Tensor(np.ones([128, 3, 32, 3]).astype(np.float32))
262    bn_training_reduce = Net(P.BNTrainingReduce())
263    grad = GradNet(bn_training_reduce)
264    grad.compile(x)
265
266
267def test_equal():
268    x = Tensor([2.0], mstype.float32)
269    y = Tensor([2.0], mstype.float32)
270    equal = Net(P.Equal())
271    grad = GradNet(equal)
272    grad.compile(x, y)
273
274
275def test_not_equal():
276    x = Tensor([2.0], mstype.float32)
277    y = Tensor([2.0], mstype.float32)
278    not_equal = Net(P.NotEqual())
279    grad = GradNet(not_equal)
280    grad.compile(x, y)
281
282
283def test_greater():
284    x = Tensor(np.array([1, 2, 3]), mstype.int32)
285    y = Tensor(np.array([1, 1, 4]), mstype.int32)
286    greater = Net(P.Greater())
287    grad = GradNet(greater)
288    grad.compile(x, y)
289
290
291def test_greater_equal():
292    x = Tensor(np.array([1, 2, 3]), mstype.int32)
293    y = Tensor(np.array([1, 1, 4]), mstype.int32)
294    greater_equal = Net(P.GreaterEqual())
295    grad = GradNet(greater_equal)
296    grad.compile(x, y)
297
298
299def test_less():
300    x = Tensor(np.array([1, 2, 3]), mstype.int32)
301    y = Tensor(np.array([1, 1, 4]), mstype.int32)
302    less = Net(P.Less())
303    grad = GradNet(less)
304    grad.compile(x, y)
305
306
307def test_less_equal():
308    x = Tensor(np.array([1, 2, 3]), mstype.int32)
309    y = Tensor(np.array([1, 1, 4]), mstype.int32)
310    less_equal = Net(P.LessEqual())
311    grad = GradNet(less_equal)
312    grad.compile(x, y)
313
314
315def test_logical_and():
316    x = Tensor(np.array([True, False, True]), mstype.bool_)
317    y = Tensor(np.array([True, True, False]), mstype.bool_)
318    logical_and = Net(P.LogicalAnd())
319    grad = GradNet(logical_and)
320    grad.compile(x, y)
321
322
323def test_logical_or():
324    x = Tensor(np.array([True, False, True]), mstype.bool_)
325    y = Tensor(np.array([True, True, False]), mstype.bool_)
326    logical_or = Net(P.LogicalOr())
327    grad = GradNet(logical_or)
328    grad.compile(x, y)
329
330
331def test_reduce_all():
332    x = Tensor(np.array([[True, False], [True, True]]))
333    reduce_all = Net(P.ReduceAll(keep_dims=True))
334    grad = GradNet(reduce_all)
335    grad.compile(x)
336
337
338def test_reduce_any():
339    x = Tensor(np.array([[True, False], [True, True]]))
340    reduce_all = Net(P.ReduceAny(keep_dims=True))
341    grad = GradNet(reduce_all)
342    grad.compile(x)
343
344
345def test_dropout_do_mask():
346    input_x = Tensor(np.ones([2, 2, 3]), mstype.float32)
347    keep_prob = Tensor(0.5, mstype.float32)
348    mask = Tensor(np.ones([2]), mstype.uint8)
349    dropout_do_mask = Net(P.DropoutDoMask())
350    grad = GradNet(dropout_do_mask)
351    grad.compile(input_x, mask, keep_prob)
352