• 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# ==============================================================================
15from tqdm import tqdm
16import numpy as np
17import mindspore as ms
18import mindspore.nn as nn
19from mindspore.dataset import NumpySlicesDataset
20from mindspore import context, Tensor
21
22context.set_context(mode=context.GRAPH_MODE, device_target="Ascend")
23
24class AutoEncoderTrainNetwork(nn.Cell):
25    def __init__(self):
26        super(AutoEncoderTrainNetwork, self).__init__()
27        self.loss_fun = nn.MSELoss()
28        self.net = nn.CellList([nn.Dense(2, 32), nn.Dense(32, 2)])
29        self.relu = nn.ReLU()
30
31    def reconstruct_sample(self, x: Tensor):
32        for _, layer in enumerate(self.net):
33            x = layer(x)
34            x = self.relu(x)
35        return x
36
37    def construct(self, x: Tensor):
38        recon_x = self.reconstruct_sample(x)
39        return self.loss_fun(recon_x, x)
40
41    def sample_2d_data(self, n_normals=2000, n_outliers=400):
42        z = np.random.randn(n_normals, 2)
43        outliers = np.random.uniform(low=-6, high=6, size=(n_outliers, 2))
44        centers = np.array([(2., 0), (-2., 0)])
45        sigma = 0.3
46        normal_points = sigma * z + centers[np.random.randint(len(centers), size=(n_normals,))]
47        return np.vstack((normal_points, outliers))
48
49    def create_synthetic_dataset(self):
50        transformed_dataset = self.sample_2d_data()
51        for dim in range(transformed_dataset.shape[1]):
52            min_val = transformed_dataset[:, dim].min()
53            max_val = transformed_dataset[:, dim].max()
54            if min_val != max_val:
55                transformed_dataset[:, dim] = (transformed_dataset[:, dim] - min_val) / (max_val - min_val)
56            elif min_val != 1:
57                transformed_dataset[:, dim] = transformed_dataset[:, dim] / min_val
58        transformed_dataset = transformed_dataset.astype(np.float32)
59        return transformed_dataset
60
61
62def test_auto_monad_layer():
63    ae_with_loss = AutoEncoderTrainNetwork()
64    transformed_dataset = ae_with_loss.create_synthetic_dataset()
65    dataloader = NumpySlicesDataset(data=(transformed_dataset,), shuffle=True)
66    dataloader = dataloader.batch(batch_size=16)
67    optim = nn.RMSProp(params=ae_with_loss.trainable_params(), learning_rate=0.002,)
68    train_net = nn.TrainOneStepCell(ae_with_loss, optim)
69    train_net.set_train()
70    gen_samples = dict()
71    num_epoch = 21
72    for epoch in tqdm(range(num_epoch)):
73        loss = []
74        for _, (batch,) in enumerate(dataloader):
75            batch = Tensor(batch, dtype=ms.float32)
76            loss_ = train_net(batch)
77            loss.append(loss_.asnumpy())
78        avg_loss = np.array(loss).mean()
79        if epoch % 10 == 0:
80            gen_samples[epoch] = ae_with_loss.reconstruct_sample(Tensor(transformed_dataset)).asnumpy()
81        print(f"epoch: {epoch}/{num_epoch}, avg loss: {avg_loss}")
82