diff --git a/feedforward/README.md b/feedforward/README.md index d2113dbcdd55cbb5e8669ebfb102549c271837c5..5bba8254eedf3956990ff1f300b16017423477ce 100644 --- a/feedforward/README.md +++ b/feedforward/README.md @@ -209,8 +209,8 @@ def get_data(): ```python train_x, train_y, test_x, test_y = get_data() -train_x = train_x.reshape(-1, 1, 28, 28) -test_x = test_x.reshape(-1, 1, 28, 28) +train_x = train_x.reshape(-1, 1, cfg.image_height, cfg.image_width) +test_x = test_x.reshape(-1, 1, cfg.image_height, cfg.image_width) train_x = train_x / 255.0 test_x = test_x / 255.0 train_x = train_x.astype('Float32') @@ -269,14 +269,12 @@ class Forward_fashion(nn.Cell): self.fc1 = nn.Dense(cfg.channel * cfg.image_height * cfg.image_width, 128) self.relu = nn.ReLU() self.fc2 = nn.Dense(128, self.num_class) - self.softmax = nn.Softmax() def construct(self, x): x = self.flatten(x) x = self.fc1(x) x = self.relu(x) x = self.fc2(x) - x = self.softmax(x) return x ``` @@ -305,65 +303,65 @@ model.train(cfg.epoch_size, ds_train, callbacks=[ckpoint_cb, loss_cb], dataset_s ``` ============== Starting Training ============== - epoch: 1 step 1000, loss is 1.6292920112609863 - Epoch time: 8827.199, per step time: 8.827, avg loss: 1.629 + epoch: 1 step 1000, loss is 0.567767322063446 + Epoch time: 11428.370, per step time: 11.428, avg loss: 0.568 ************************************************************ - epoch: 2 step 1000, loss is 1.6026073694229126 - Epoch time: 1601.618, per step time: 1.602, avg loss: 1.603 + epoch: 2 step 1000, loss is 0.3970850706100464 + Epoch time: 2018.074, per step time: 2.018, avg loss: 0.397 ************************************************************ - epoch: 3 step 1000, loss is 1.6622530221939087 - Epoch time: 1629.635, per step time: 1.630, avg loss: 1.662 + epoch: 3 step 1000, loss is 0.31815576553344727 + Epoch time: 1971.219, per step time: 1.971, avg loss: 0.318 ************************************************************ - epoch: 4 step 1000, loss is 1.6305657625198364 - Epoch time: 1471.701, per step time: 1.472, avg loss: 1.631 + epoch: 4 step 1000, loss is 0.3128049373626709 + Epoch time: 1974.937, per step time: 1.975, avg loss: 0.313 ************************************************************ - epoch: 5 step 1000, loss is 1.5535054206848145 - Epoch time: 1770.755, per step time: 1.771, avg loss: 1.554 + epoch: 5 step 1000, loss is 0.3095005750656128 + Epoch time: 2029.930, per step time: 2.030, avg loss: 0.310 ************************************************************ - epoch: 6 step 1000, loss is 1.5950586795806885 - Epoch time: 1985.995, per step time: 1.986, avg loss: 1.595 + epoch: 6 step 1000, loss is 0.25628671050071716 + Epoch time: 1934.886, per step time: 1.935, avg loss: 0.256 ************************************************************ - epoch: 7 step 1000, loss is 1.6165529489517212 - Epoch time: 1928.856, per step time: 1.929, avg loss: 1.617 + epoch: 7 step 1000, loss is 0.24347715079784393 + Epoch time: 1897.307, per step time: 1.897, avg loss: 0.243 ************************************************************ - epoch: 8 step 1000, loss is 1.5757038593292236 - Epoch time: 1814.812, per step time: 1.815, avg loss: 1.576 + epoch: 8 step 1000, loss is 0.28936269879341125 + Epoch time: 1921.264, per step time: 1.921, avg loss: 0.289 ************************************************************ - epoch: 9 step 1000, loss is 1.5905802249908447 - Epoch time: 1558.252, per step time: 1.558, avg loss: 1.591 + epoch: 9 step 1000, loss is 0.4469510316848755 + Epoch time: 1875.093, per step time: 1.875, avg loss: 0.447 ************************************************************ - epoch: 10 step 1000, loss is 1.5414245128631592 - Epoch time: 1456.539, per step time: 1.457, avg loss: 1.541 + epoch: 10 step 1000, loss is 0.2915213108062744 + Epoch time: 1876.605, per step time: 1.877, avg loss: 0.292 ************************************************************ - epoch: 11 step 1000, loss is 1.5789177417755127 - Epoch time: 1591.151, per step time: 1.591, avg loss: 1.579 + epoch: 11 step 1000, loss is 0.24928903579711914 + Epoch time: 1910.094, per step time: 1.910, avg loss: 0.249 ************************************************************ - epoch: 12 step 1000, loss is 1.5879883766174316 - Epoch time: 2011.590, per step time: 2.012, avg loss: 1.588 + epoch: 12 step 1000, loss is 0.12853321433067322 + Epoch time: 1974.167, per step time: 1.974, avg loss: 0.129 ************************************************************ - epoch: 13 step 1000, loss is 1.5823071002960205 - Epoch time: 1663.497, per step time: 1.663, avg loss: 1.582 + epoch: 13 step 1000, loss is 0.14836660027503967 + Epoch time: 1841.105, per step time: 1.841, avg loss: 0.148 ************************************************************ - epoch: 14 step 1000, loss is 1.6213573217391968 - Epoch time: 1914.475, per step time: 1.914, avg loss: 1.621 + epoch: 14 step 1000, loss is 0.26581835746765137 + Epoch time: 1694.728, per step time: 1.695, avg loss: 0.266 ************************************************************ - epoch: 15 step 1000, loss is 1.6282684803009033 - Epoch time: 1921.290, per step time: 1.921, avg loss: 1.628 + epoch: 15 step 1000, loss is 0.2012856900691986 + Epoch time: 1937.829, per step time: 1.938, avg loss: 0.201 ************************************************************ - epoch: 16 step 1000, loss is 1.5011317729949951 - Epoch time: 1899.634, per step time: 1.900, avg loss: 1.501 + epoch: 16 step 1000, loss is 0.14978612959384918 + Epoch time: 1793.748, per step time: 1.794, avg loss: 0.150 ************************************************************ - epoch: 17 step 1000, loss is 1.566664457321167 - Epoch time: 1495.834, per step time: 1.496, avg loss: 1.567 + epoch: 17 step 1000, loss is 0.3085048198699951 + Epoch time: 1667.389, per step time: 1.667, avg loss: 0.309 ************************************************************ - epoch: 18 step 1000, loss is 1.552886724472046 - Epoch time: 1449.513, per step time: 1.450, avg loss: 1.553 + epoch: 18 step 1000, loss is 0.17254383862018585 + Epoch time: 1558.955, per step time: 1.559, avg loss: 0.173 ************************************************************ - epoch: 19 step 1000, loss is 1.6042685508728027 - Epoch time: 1721.564, per step time: 1.722, avg loss: 1.604 + epoch: 19 step 1000, loss is 0.10585948824882507 + Epoch time: 1567.354, per step time: 1.567, avg loss: 0.106 ************************************************************ - epoch: 20 step 1000, loss is 1.5174891948699951 - Epoch time: 1633.871, per step time: 1.634, avg loss: 1.517 + epoch: 20 step 1000, loss is 0.27113234996795654 + Epoch time: 1589.239, per step time: 1.589, avg loss: 0.271 ************************************************************ #### 评估测试 @@ -374,7 +372,7 @@ metric = model.eval(ds_test) print(metric) ``` - {'acc': 0.8715863453815261} + {'acc': 0.8862449799196788} #### 预测 @@ -384,22 +382,27 @@ test_ = ds_test.create_dict_iterator().get_next() test = Tensor(test_['x'], mindspore.float32) predictions = model.predict(test) predictions = predictions.asnumpy() -for i in range(10): +for i in range(15): p_np = predictions[i, :] p_list = p_np.tolist() print('第' + str(i) + '个sample预测结果：', p_list.index(max(p_list)), ' 真实结果：', test_['y'][i]) ``` - 第0个sample预测结果： 3 真实结果： 3 - 第1个sample预测结果： 2 真实结果： 2 - 第2个sample预测结果： 3 真实结果： 3 - 第3个sample预测结果： 9 真实结果： 9 + 第0个sample预测结果： 1 真实结果： 1 + 第1个sample预测结果： 0 真实结果： 0 + 第2个sample预测结果： 2 真实结果： 2 + 第3个sample预测结果： 2 真实结果： 2 第4个sample预测结果： 8 真实结果： 8 第5个sample预测结果： 4 真实结果： 4 第6个sample预测结果： 4 真实结果： 4 - 第7个sample预测结果： 8 真实结果： 8 - 第8个sample预测结果： 3 真实结果： 3 - 第9个sample预测结果： 1 真实结果： 1 + 第7个sample预测结果： 1 真实结果： 1 + 第8个sample预测结果： 6 真实结果： 2 + 第9个sample预测结果： 8 真实结果： 8 + 第10个sample预测结果： 5 真实结果： 5 + 第11个sample预测结果： 8 真实结果： 0 + 第12个sample预测结果： 5 真实结果： 5 + 第13个sample预测结果： 6 真实结果： 6 + 第14个sample预测结果： 9 真实结果： 9 #### 对预测结果可视化 @@ -471,29 +474,18 @@ args, unknown = parser.parse_known_args() MindSpore暂时没有提供直接访问OBS数据的接口，需要通过MoXing提供的API与OBS交互。将OBS中存储的数据拷贝至执行容器： -- 方式一，拷贝自己账户下OBS桶内的数据集。 - - ```python - import moxing - moxing.file.copy_parallel(src_url=args.data_url, dst_url='Fashion-MNIST/') - ``` - -- 方式二，拷贝他人账户下OBS桶内的数据集，前提是他人账户下的OBS桶已设为公共读/公共读写，且需要他人账户的访问密钥、私有访问密钥、OBS桶-概览-基本信息-Endpoint。 - - ```python - import moxing - # set moxing/obs auth info, ak:Access Key Id, sk:Secret Access Key, server:endpoint of obs bucket - moxing.file.set_auth(ak='VCT2GKI3GJOZBQYJG5WM', sk='t1y8M4Z6bHLSAEGK2bCeRYMjo2S2u0QBqToYbxzB', - server="obs.cn-north-4.myhuaweicloud.com") - # copy dataset from obs bucket to container/cache - moxing.file.copy_parallel(src_url="s3://share-course/dataset/fashion-mnist/", dst_url='Fashion-MNIST/') - ``` +拷贝自己账户下OBS桶内的数据集。 + +```python +import moxing +moxing.file.copy_parallel(src_url=args.data_url, dst_url='Fashion-MNIST/') +``` 如需将训练输出（如模型Checkpoint）从执行容器拷贝至OBS，请参考： ```python -import moxing as mox -mox.file.copy_parallel(src_url='model_fashion', dst_url=args.train_url) +import moxing +moxing.file.copy_parallel(src_url='model_fashion', dst_url=args.train_url) ``` ### 创建训练作业 diff --git a/feedforward/images/output_2.png b/feedforward/images/output_2.png index ed7de9222d607d5e35b9a9405605954431f63b1e..e89f0284873ac0a429604a26ec95f39073d0cf4c 100644 Binary files a/feedforward/images/output_2.png and b/feedforward/images/output_2.png differ diff --git a/feedforward/main.py b/feedforward/main.py index 7bd301a642bc2b6e0a7ac05c7f3fcecf1dbf11c4..71cf4af8e9bad486d1bd86a11223db8a1c8fecd0 100644 --- a/feedforward/main.py +++ b/feedforward/main.py @@ -18,6 +18,37 @@ from mindspore import Tensor context.set_context(mode=context.GRAPH_MODE, device_target='Ascend') +import argparse +parser = argparse.ArgumentParser() +parser.add_argument('--data_url', required=True, default=None, help='Location of data.') +parser.add_argument('--train_url', required=True, default=None, help='Location of training outputs.') +args, unknown = parser.parse_known_args() + +import moxing + +# copy dataset from your own OBS bucket. +moxing.file.copy_parallel(src_url=args.data_url, dst_url='Fashion-MNIST') + + +cfg = edict({ + 'train_size': 60000, # 训练集大小 + 'test_size': 10000, # 测试集大小 + 'channel': 1, # 图片通道数 + 'image_height': 28, # 图片高度 + 'image_width': 28, # 图片宽度 + 'batch_size': 60, + 'num_classes': 10, # 分类类别 + 'lr': 0.001, # 学习率 + 'epoch_size': 20, # 训练次数 + 'data_dir_train': os.path.join('Fashion-MNIST', 'train'), + 'data_dir_test': os.path.join('Fashion-MNIST', 'test'), + 'save_checkpoint_steps': 1, # 多少步保存一次模型 + 'keep_checkpoint_max': 3, # 最多保存多少个模型 + 'output_directory': './model_fashion', # 保存模型路径 + 'output_prefix': "checkpoint_fashion_forward" # 保存模型文件名字 +}) + + def read_image(file_name): ''' :param file_name: 文件路径 @@ -94,57 +125,18 @@ class Forward_fashion(nn.Cell): self.fc1 = nn.Dense(cfg.channel * cfg.image_height * cfg.image_width, 128) self.relu = nn.ReLU() self.fc2 = nn.Dense(128, self.num_class) - self.softmax = nn.Softmax() def construct(self, x): x = self.flatten(x) x = self.fc1(x) x = self.relu(x) x = self.fc2(x) - x = self.softmax(x) return x -import argparse -parser = argparse.ArgumentParser() -parser.add_argument('--data_url', required=True, default=None, help='Location of data.') -parser.add_argument('--train_url', required=True, default=None, help='Location of training outputs.') -args, unknown = parser.parse_known_args() - -import moxing - -# WAY1: copy dataset from your own OBS bucket. -# moxing.file.copy_parallel(src_url=args.data_url, dst_url='Fashion-MNIST') - -# WAY2: copy dataset from other's OBS bucket, which has been set public read or public read&write. -# set moxing/obs auth info, ak:Access Key Id, sk:Secret Access Key, server:endpoint of obs bucket -moxing.file.set_auth(ak='VCT2GKI3GJOZBQYJG5WM', sk='t1y8M4Z6bHLSAEGK2bCeRYMjo2S2u0QBqToYbxzB', - server="obs.cn-north-4.myhuaweicloud.com") -# copy dataset from obs bucket to container/cache -moxing.file.copy_parallel(src_url="s3://share-course/dataset/fashion-mnist/", dst_url='Fashion-MNIST/') - - -cfg = edict({ - 'train_size': 60000, # 训练集大小 - 'test_size': 10000, # 测试集大小 - 'channel': 1, # 图片通道数 - 'image_height': 28, # 图片高度 - 'image_width': 28, # 图片宽度 - 'batch_size': 60, - 'num_classes': 10, # 分类类别 - 'lr': 0.001, # 学习率 - 'epoch_size': 20, # 训练次数 - 'data_dir_train': os.path.join('Fashion-MNIST', 'train'), - 'data_dir_test': os.path.join('Fashion-MNIST', 'test'), - 'save_checkpoint_steps': 1, # 多少步保存一次模型 - 'keep_checkpoint_max': 3, # 最多保存多少个模型 - 'output_directory': './model_fashion', # 保存模型路径 - 'output_prefix': "checkpoint_fashion_forward" # 保存模型文件名字 -}) - train_x, train_y, test_x, test_y = get_data() -train_x = train_x.reshape(-1, 1, 28, 28) -test_x = test_x.reshape(-1, 1, 28, 28) +train_x = train_x.reshape(-1, 1, cfg.image_height, cfg.image_width) +test_x = test_x.reshape(-1, 1, cfg.image_height, cfg.image_width) train_x = train_x / 255.0 test_x = test_x / 255.0 train_x = train_x.astype('Float32') @@ -190,9 +182,9 @@ test_ = ds_test.create_dict_iterator().get_next() test = Tensor(test_['x'], mindspore.float32) predictions = model.predict(test) predictions = predictions.asnumpy() -for i in range(10): +for i in range(15): p_np = predictions[i, :] p_list = p_np.tolist() print('第' + str(i) + '个sample预测结果：', p_list.index(max(p_list)), ' 真实结果：', test_['y'][i]) -mox.file.copy_parallel(src_url='model_fashion', dst_url=args.train_url) +moxing.file.copy_parallel(src_url='model_fashion', dst_url=args.train_url)