Huawei_Technology
/
mindspore
Not watched
1
0
Fork 0
Code
Releases 13 Wiki evaluate Activity
Issues 0 Pull Requests 0 Datasets Model Cloudbrain HPC
You can not select more than 25 topics Topics must start with a chinese character,a letter or number, can include dashes ('-') and can be up to 35 characters long.
Tree: 1a54785fe2
Branches Tags
${ item.name }
Create branch ${ searchTerm }
from '1a54785fe2'
${ noResults }
mindspore/tests/st/ops/gpu/test_ctcloss_op.py

120 lines
6.6 kB
Raw Blame History 

  1. # Copyright 2020 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. 

  16. import numpy as np

  17. import pytest

  18. 

  19. import mindspore.context as context

  20. import mindspore.nn as nn

  21. from mindspore import Tensor

  22. from mindspore.ops import operations as P

  23. from mindspore.common import dtype as mstype

  24. from mindspore.ops.composite import GradOperation

  25. 

  26. class Net(nn.Cell):

  27.     def __init__(self):

  28.         super(Net, self).__init__()

  29.         self.loss = P.CTCLossV2()

  30.         self.div = P.RealDiv()

  31.         self.cast = P.Cast()

  32.         self.mean = P.ReduceMean()

  33. 

  34.     def construct(self, probs, label, input_length, label_length):

  35.         x, _ = self.loss(probs, label, input_length, label_length)

  36.         x = self.div(x, self.cast(label_length, mstype.float32))

  37.         x = self.mean(x)

  38.         return x

  39. 

  40. class GradData(nn.Cell):

  41.     def __init__(self, network):

  42.         super(GradData, self).__init__()

  43.         self.grad = GradOperation(get_all=True, sens_param=False)

  44.         self.network = network

  45. 

  46.     def construct(self, probs, labels, input_lengths, label_lengths):

  47.         return self.grad(self.network)(probs, labels, input_lengths, label_lengths)

  48. 

  49. 

  50. @pytest.mark.level0

  51. @pytest.mark.platform_x86_gpu_training

  52. @pytest.mark.env_onecard

  53. def test_ctcloss():

  54.     probs = Tensor([[[-4.4131, -4.6093, -3.4333, -3.9268, -2.8917, -3.4093, -4.2243, -1.1379, -7.1046, -0.6902],

  55.                      [-2.5109, -3.3397, -4.9384, -1.2723, -1.1443, -2.4683, -2.6768, -4.1282, -2.7062, -3.1906],

  56.                      [-2.5092, -1.6392, -2.0864, -4.0059, -1.5610, -2.3223, -2.4816, -2.9922, -3.1412, -2.3311]],

  57. 

  58.                     [[-2.1243, -3.5773, -3.1108, -4.4253, -2.7080, -1.9653, -2.0499, -2.4418, -1.8620, -1.5229],

  59.                      [-2.2479, -3.5128, -1.4189, -2.8701, -1.8562, -2.2752, -2.7019, -2.1865, -2.5634, -2.9869],

  60.                      [-3.2144, -1.3986, -3.1083, -3.9634, -3.5131, -3.2317, -2.6200, -1.7938, -1.8159, -1.7255]],

  61. 

  62.                     [[-3.1301, -2.1649, -0.9286, -2.9452, -2.5992, -2.0263, -2.9201, -3.2155, -2.8302, -3.3636],

  63.                      [-1.4661, -3.6311, -2.4781, -4.6180, -2.7308, -1.7019, -1.5570, -2.6012, -4.0788, -2.3073],

  64.                      [-2.6833, -1.5033, -3.6922, -2.6360, -2.6974, -2.6847, -2.7579, -2.1396, -1.4093, -2.9630]],

  65. 

  66.                     [[-2.0094, -2.3024, -3.3673, -1.0220, -2.8326, -2.2613, -3.0535, -2.9879, -3.7015, -2.4510],

  67.                      [-1.9071, -3.2603, -2.3229, -2.0572, -4.3450, -2.1284, -2.6306, -1.3824, -2.9815, -2.5061],

  68.                      [-2.7931, -3.7631, -3.2440, -4.3887, -1.0271, -3.8851, -1.2418, -4.5123, -2.2993, -2.4607]],

  69. 

  70.                     [[-1.5763, -2.7539, -3.6941, -3.8166, -1.2599, -2.6903, -2.5826, -4.8208, -2.9562, -1.6321],

  71.                      [-3.3031, -3.0087, -1.9982, -1.9081, -3.8731, -2.8764, -2.2485, -2.3808, -1.4283, -2.1625],

  72.                      [-2.4516, -3.2394, -4.2053, -4.3541, -2.5229, -4.0717, -1.4894, -2.3151, -1.1098, -2.3465]]],

  73.                    dtype=mstype.float32)

  74.     labels = Tensor([9, 4, 6, 4, 7, 1, 4, 6, 6, 8], dtype=mstype.int32)

  75.     input_lengths = Tensor([5, 5, 5], dtype=mstype.int32)

  76.     label_lengths = Tensor([3, 3, 4], dtype=mstype.int32)

  77. 

  78.     context.set_context(mode=context.GRAPH_MODE, device_target="GPU")

  79.     net = Net()

  80.     ctc_loss = net(probs, labels, input_lengths, label_lengths)

  81.     expect_loss = [2.4099]

  82.     assert np.allclose(ctc_loss.asnumpy(), expect_loss)

  83. 

  84.     grad = GradData(net)(probs, labels, input_lengths, label_lengths)

  85.     expect_grad = [[[8.8442e-05, 1.1065e-03, 3.5867e-03, 2.1896e-03, 6.1646e-03,

  86.                      3.6738e-03, 1.6262e-03, 3.5610e-02, 9.1258e-05, -5.4134e-02],

  87.                     [-3.7523e-03, 3.9386e-03, 7.9623e-04, 3.1132e-02, -6.2954e-02,

  88.                      9.4143e-03, 7.6425e-03, 1.7902e-03, 7.4211e-03, 4.5719e-03],

  89.                     [6.7778e-03, 1.6178e-02, 1.0344e-02, 1.5173e-03, -6.5840e-02,

  90.                      8.1707e-03, 6.9674e-03, 4.1814e-03, 3.6026e-03, 8.0991e-03]],

  91. 

  92.                    [[-1.2581e-02, 3.1057e-03, 4.9517e-03, 1.3301e-03, -2.6320e-02,

  93.                      1.5568e-02, 1.4305e-02, 9.6671e-03, 1.7262e-02, -2.7292e-02],

  94.                     [-1.5566e-02, 3.3126e-03, 2.6887e-02, 6.2993e-03, -3.9716e-02,

  95.                      1.1420e-02, 7.4531e-03, -1.4252e-02, 8.5603e-03, 5.6048e-03],

  96.                     [3.3483e-03, 2.0579e-02, 3.7231e-03, 1.5832e-03, 2.4837e-03,

  97.                      3.2909e-03, -7.7267e-02, 1.3861e-02, 1.3558e-02, 1.4840e-02]],

  98. 

  99.                    [[-8.0007e-03, 1.2751e-02, 4.3901e-02, 5.8435e-03, -7.2627e-02,

  100.                      1.4647e-02, -8.0584e-03, 4.4595e-03, 6.5557e-03, 5.2891e-04],

  101.                     [-3.6006e-02, 1.5308e-03, 9.3225e-03, 1.0969e-03, -2.5098e-03,

  102.                      2.0260e-02, 2.3419e-02, -3.0053e-02, 1.8809e-03, 1.1059e-02],

  103.                     [-7.7639e-02, 1.8533e-02, 2.0764e-03, 5.9706e-03, 5.6150e-03,

  104.                      5.6868e-03, 5.2854e-03, 9.8085e-03, 2.0360e-02, 4.3053e-03]],

  105. 

  106.                    [[-2.6776e-02, 1.1113e-02, 3.8314e-03, 3.9986e-02, -1.6020e-02,

  107.                      1.1579e-02, -4.1635e-02, 5.5992e-03, 2.7429e-03, 9.5786e-03],

  108.                     [-6.8619e-03, -6.4066e-03, 1.0888e-02, 1.4201e-02, 1.4413e-03,

  109.                      1.3225e-02, 8.0039e-03, -4.9191e-02, 5.6352e-03, 9.0651e-03],

  110.                     [5.1026e-03, 1.9343e-03, 3.2506e-03, 1.0347e-03, 2.9837e-02,

  111.                      1.7121e-03, -5.9261e-02, 9.1443e-04, 8.3608e-03, 7.1146e-03]],

  112. 

  113.                    [[-2.0848e-02, 7.0754e-03, 2.7633e-03, 2.4447e-03, 3.1520e-02,

  114.                      7.5401e-03, -5.8895e-02, 8.9559e-04, 5.7796e-03, 2.1724e-02],

  115.                     [-1.3499e-03, -1.0019e-01, 1.5064e-02, 1.6485e-02, 2.3104e-03,

  116.                      6.2597e-03, 1.1729e-02, 1.0275e-02, 2.6635e-02, 1.2782e-02],

  117.                     [7.1796e-03, 3.2656e-03, 1.2430e-03, 1.0712e-03, 6.6856e-03,

  118.                      1.4207e-03, 1.8792e-02, 8.2297e-03, -5.5865e-02, 7.9753e-03]]]

  119.     assert np.allclose(grad[0].asnumpy(), expect_grad, atol=1e-5)