1# Copyright 2019 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# ============================================================================ 15import numpy as np 16 17import mindspore.context as context 18import mindspore.nn as nn 19from mindspore import Tensor 20from mindspore.common.api import ms_function 21from mindspore.ops import operations as P 22 23context.set_context(device_target="Ascend") 24 25 26class Net(nn.Cell): 27 def __init__(self, is_grad=False): 28 super(Net, self).__init__() 29 self.SparseSoftmaxCrossEntropyWithLogits = P.SparseSoftmaxCrossEntropyWithLogits(is_grad=is_grad) 30 31 @ms_function 32 def construct(self, features, labels): 33 return self.SparseSoftmaxCrossEntropyWithLogits(features, labels) 34 35 36def np_sparse_softmax_cross_entropy_with_logits(labels_shape, logits_shape, logits_dtype): 37 num_class = logits_shape[1] 38 labels = np.random.randint(low=0, high=num_class - 1, size=labels_shape).astype(np.int32) 39 logits = np.random.rand(*logits_shape).astype(logits_dtype) 40 features = logits 41 features_reshape = np.reshape(features, [-1, num_class]) 42 labels_reshape = np.reshape(labels, [-1]) 43 batch_dim = 0 44 class_dim = 1 45 batch_size = features_reshape.shape[batch_dim] 46 e = np.exp(features_reshape - np.reshape(np.amax(features_reshape, axis=class_dim), [batch_size, 1])) 47 probs = e / np.reshape(np.sum(e, axis=class_dim), [batch_size, 1]) 48 labels_mat = np.zeros_like(probs).astype(probs.dtype) 49 labels_mat[np.arange(batch_size), labels_reshape] = 1.0 50 bp = (probs - labels_mat) 51 loss = -np.sum(labels_mat * np.log(probs + 1.0e-20), axis=1) 52 bp_res = np.reshape(bp, features.shape) 53 loss_res = np.reshape(loss, labels.shape) 54 loss_res = np.sum(loss_res, axis=0) / loss_res.shape[0] 55 return labels, logits, loss_res, bp_res 56 57 58def test_net(): 59 '''Compare Numpy with MS type is float32''' 60 labels_shape = (32,) 61 logits_shape = [32, 1001] 62 labels, logits, loss_np, _ = np_sparse_softmax_cross_entropy_with_logits(labels_shape, logits_shape, np.float32) 63 expect = loss_np 64 SparseSoftmaxCrossEntropyWithLogits = Net() 65 loss_me = SparseSoftmaxCrossEntropyWithLogits(Tensor(logits), Tensor(labels)) 66# assert 67 assert np.allclose(expect.flatten(), loss_me.asnumpy().flatten(), 0.01, 0.01) 68 print(loss_me.asnumpy().flatten()) 69 print("-------------------------") 70 print(expect) 71 72 73test_net() 74