本文参考内容：
有官方文档内容可知，我们对网络参数的保存不仅可以使用model来自动保存，也可以使用
from mindspore.train.serialization import save_checkpoint
来进行手动保存。
===========================================================
自动保存参数：
给出在模型训练过程中自动保存参数的代码demo：
#!/usr/bin python # encoding:UTF-8 """" 对输入的超参数进行处理 """ import os import argparse """ 设置运行的背景context """ from mindspore import context """ 对数据集进行预处理 """ import mindspore.dataset as ds import mindspore.dataset.transforms.c_transforms as C import mindspore.dataset.vision.c_transforms as CV from mindspore.dataset.vision import Inter from mindspore import dtype as mstype """ 构建神经网络 """ import mindspore.nn as nn from mindspore.common.initializer import Normal """ 训练时对模型参数的保存 """ from mindspore.train.callback import ModelCheckpoint, CheckpointConfig from mindspore.train.serialization import save_checkpoint """ 导入模型训练需要的库 """ from mindspore.nn import Accuracy from mindspore.train.callback import LossMonitor from mindspore import Model import os os.system('rm -f *.ckpt *.meta') parser = argparse.ArgumentParser(description='MindSpore LeNet Example') parser.add_argument('--device_target', type=str, default="CPU", choices=['Ascend', 'GPU', 'CPU']) args = parser.parse_known_args()[0] # 为mindspore设置运行背景context context.set_context(mode=context.GRAPH_MODE, device_target=args.device_target) def create_dataset(data_path, batch_size=32, repeat_size=1, num_parallel_workers=1): # 定义数据集 mnist_ds = ds.MnistDataset(data_path) resize_height, resize_width = 32, 32 rescale = 1.0 / 255.0 shift = 0.0 rescale_nml = 1 / 0.3081 shift_nml = -1 * 0.1307 / 0.3081 # 定义所需要操作的map映射 resize_op = CV.Resize((resize_height, resize_width), interpolation=Inter.LINEAR) rescale_nml_op = CV.Rescale(rescale_nml, shift_nml) rescale_op = CV.Rescale(rescale, shift) hwc2chw_op = CV.HWC2CHW() type_cast_op = C.TypeCast(mstype.int32) # 使用map映射函数，将数据操作应用到数据集 mnist_ds = mnist_ds.map(operations=type_cast_op, input_columns="label", num_parallel_workers=num_parallel_workers) mnist_ds = mnist_ds.map(operations=resize_op, input_columns="image", num_parallel_workers=num_parallel_workers) mnist_ds = mnist_ds.map(operations=rescale_op, input_columns="image", num_parallel_workers=num_parallel_workers) mnist_ds = mnist_ds.map(operations=rescale_nml_op, input_columns="image", num_parallel_workers=num_parallel_workers) mnist_ds = mnist_ds.map(operations=hwc2chw_op, input_columns="image", num_parallel_workers=num_parallel_workers) # 进行shuffle、batch、repeat操作 buffer_size = 10000 mnist_ds = mnist_ds.shuffle(buffer_size=buffer_size) mnist_ds = mnist_ds.batch(batch_size, drop_remainder=True) mnist_ds = mnist_ds.repeat(repeat_size) return mnist_ds class LeNet5(nn.Cell): """ Lenet网络结构 """ def __init__(self, num_class=10, num_channel=1): super(LeNet5, self).__init__() # 定义所需要的运算 self.conv1 = nn.Conv2d(num_channel, 6, 5, pad_mode='valid') self.conv2 = nn.Conv2d(6, 16, 5, pad_mode='valid') self.fc1 = nn.Dense(16 * 5 * 5, 120, weight_init=Normal(0.02)) self.fc2 = nn.Dense(120, 84, weight_init=Normal(0.02)) self.fc3 = nn.Dense(84, num_class, weight_init=Normal(0.02)) self.relu = nn.ReLU() self.max_pool2d = nn.MaxPool2d(kernel_size=2, stride=2) self.flatten = nn.Flatten() def construct(self, x): # 使用定义好的运算构建前向网络 x = self.conv1(x) x = self.relu(x) x = self.max_pool2d(x) x = self.conv2(x) x = self.relu(x) x = self.max_pool2d(x) x = self.flatten(x) x = self.fc1(x) x = self.relu(x) x = self.fc2(x) x = self.relu(x) x = self.fc3(x) return x # 实例化网络 net = LeNet5() # 定义损失函数 net_loss = nn.SoftmaxCrossEntropyWithLogits(sparse=True, reduction='mean') # 定义优化器 net_opt = nn.Momentum(net.trainable_params(), learning_rate=0.01, momentum=0.9) # 设置模型保存参数 # 每125steps保存一次模型参数，最多保留15个文件 config_ck = CheckpointConfig(save_checkpoint_steps=125, keep_checkpoint_max=15) # 应用模型保存参数 ckpoint = ModelCheckpoint(prefix="checkpoint_lenet", config=config_ck) def train_net(args, model, epoch_size, data_path, repeat_size, ckpoint_cb, sink_mode): """定义训练的方法""" # 加载训练数据集 ds_train = create_dataset(os.path.join(data_path, "train"), 32, repeat_size) model.train(epoch_size, ds_train, callbacks=[ckpoint_cb, LossMonitor(125)], dataset_sink_mode=sink_mode) def test_net(network, model, data_path): """定义验证的方法""" ds_eval = create_dataset(os.path.join(data_path, "test")) acc = model.eval(ds_eval, dataset_sink_mode=False) print("{}".format(acc)) mnist_path = "./datasets/MNIST_Data" train_epoch = 1 dataset_size = 1 model = Model(net, net_loss, net_opt, metrics={"Accuracy": Accuracy()}) train_net(args, model, train_epoch, mnist_path, dataset_size, ckpoint, False) #test_net(net, model, mnist_path)
其中，保存参数代码主要为：
# 设置模型保存参数
# 每125steps保存一次模型参数，最多保留15个文件
config_ck = CheckpointConfig(save_checkpoint_steps=125, keep_checkpoint_max=15)
# 应用模型保存参数
ckpoint = ModelCheckpoint(prefix="checkpoint_lenet", config=config_ck)
model.train(epoch_size, ds_train, callbacks=[ckpoint_cb, LossMonitor(125)], dataset_sink_mode=sink_mode)
在模型训练过程中，自动保存参数，不仅会保存网络参数，同时也会保存优化器参数。
========================================================================
手动保存参数：
demo：
#!/usr/bin python # encoding:UTF-8 """" 对输入的超参数进行处理 """ import os import argparse """ 设置运行的背景context """ from mindspore import context """ 对数据集进行预处理 """ import mindspore.dataset as ds import mindspore.dataset.transforms.c_transforms as C import mindspore.dataset.vision.c_transforms as CV from mindspore.dataset.vision import Inter from mindspore import dtype as mstype """ 构建神经网络 """ import mindspore.nn as nn from mindspore.common.initializer import Normal """ 训练时对模型参数的保存 """ from mindspore.train.callback import ModelCheckpoint, CheckpointConfig from mindspore.train.serialization import save_checkpoint """ 导入模型训练需要的库 """ from mindspore.nn import Accuracy from mindspore.train.callback import LossMonitor from mindspore import Model import os os.system('rm -f *.ckpt *.meta') parser = argparse.ArgumentParser(description='MindSpore LeNet Example') parser.add_argument('--device_target', type=str, default="CPU", choices=['Ascend', 'GPU', 'CPU']) args = parser.parse_known_args()[0] # 为mindspore设置运行背景context context.set_context(mode=context.GRAPH_MODE, device_target=args.device_target) def create_dataset(data_path, batch_size=32, repeat_size=1, num_parallel_workers=1): # 定义数据集 mnist_ds = ds.MnistDataset(data_path) resize_height, resize_width = 32, 32 rescale = 1.0 / 255.0 shift = 0.0 rescale_nml = 1 / 0.3081 shift_nml = -1 * 0.1307 / 0.3081 # 定义所需要操作的map映射 resize_op = CV.Resize((resize_height, resize_width), interpolation=Inter.LINEAR) rescale_nml_op = CV.Rescale(rescale_nml, shift_nml) rescale_op = CV.Rescale(rescale, shift) hwc2chw_op = CV.HWC2CHW() type_cast_op = C.TypeCast(mstype.int32) # 使用map映射函数，将数据操作应用到数据集 mnist_ds = mnist_ds.map(operations=type_cast_op, input_columns="label", num_parallel_workers=num_parallel_workers) mnist_ds = mnist_ds.map(operations=resize_op, input_columns="image", num_parallel_workers=num_parallel_workers) mnist_ds = mnist_ds.map(operations=rescale_op, input_columns="image", num_parallel_workers=num_parallel_workers) mnist_ds = mnist_ds.map(operations=rescale_nml_op, input_columns="image", num_parallel_workers=num_parallel_workers) mnist_ds = mnist_ds.map(operations=hwc2chw_op, input_columns="image", num_parallel_workers=num_parallel_workers) # 进行shuffle、batch、repeat操作 buffer_size = 10000 mnist_ds = mnist_ds.shuffle(buffer_size=buffer_size) mnist_ds = mnist_ds.batch(batch_size, drop_remainder=True) mnist_ds = mnist_ds.repeat(repeat_size) return mnist_ds class LeNet5(nn.Cell): """ Lenet网络结构 """ def __init__(self, num_class=10, num_channel=1): super(LeNet5, self).__init__() # 定义所需要的运算 self.conv1 = nn.Conv2d(num_channel, 6, 5, pad_mode='valid') self.conv2 = nn.Conv2d(6, 16, 5, pad_mode='valid') self.fc1 = nn.Dense(16 * 5 * 5, 120, weight_init=Normal(0.02)) self.fc2 = nn.Dense(120, 84, weight_init=Normal(0.02)) self.fc3 = nn.Dense(84, num_class, weight_init=Normal(0.02)) self.relu = nn.ReLU() self.max_pool2d = nn.MaxPool2d(kernel_size=2, stride=2) self.flatten = nn.Flatten() def construct(self, x): # 使用定义好的运算构建前向网络 x = self.conv1(x) x = self.relu(x) x = self.max_pool2d(x) x = self.conv2(x) x = self.relu(x) x = self.max_pool2d(x) x = self.flatten(x) x = self.fc1(x) x = self.relu(x) x = self.fc2(x) x = self.relu(x) x = self.fc3(x) return x # 实例化网络 net = LeNet5() # 定义损失函数 net_loss = nn.SoftmaxCrossEntropyWithLogits(sparse=True, reduction='mean') # 定义优化器 net_opt = nn.Momentum(net.trainable_params(), learning_rate=0.01, momentum=0.9) # 设置模型保存参数 # 每125steps保存一次模型参数，最多保留15个文件 config_ck = CheckpointConfig(save_checkpoint_steps=125, keep_checkpoint_max=15) # 应用模型保存参数 ckpoint = ModelCheckpoint(prefix="checkpoint_lenet", config=config_ck) def train_net(args, model, epoch_size, data_path, repeat_size, ckpoint_cb, sink_mode): """定义训练的方法""" # 加载训练数据集 ds_train = create_dataset(os.path.join(data_path, "train"), 32, repeat_size) model.train(epoch_size, ds_train, callbacks=[ckpoint_cb, LossMonitor(125)], dataset_sink_mode=sink_mode) def test_net(network, model, data_path): """定义验证的方法""" ds_eval = create_dataset(os.path.join(data_path, "test")) acc = model.eval(ds_eval, dataset_sink_mode=False) print("{}".format(acc)) mnist_path = "./datasets/MNIST_Data" train_epoch = 1 dataset_size = 1 model = Model(net, net_loss, net_opt, metrics={"Accuracy": Accuracy()}) #train_net(args, model, train_epoch, mnist_path, dataset_size, ckpoint, False) #test_net(net, model, mnist_path) save_checkpoint(net, './net_parameters.ckpt') save_checkpoint(net_opt, './net_opt_parameters.ckpt')
主要代码：
from mindspore.train.serialization import save_checkpoint
save_checkpoint(net, './net_parameters.ckpt')
save_checkpoint(net_opt, './net_opt_parameters.ckpt')
其中，
save_checkpoint(net, './net_parameters.ckpt') # 是网络模型net中的参数保存为 net_parameters.ckpt 文件
save_checkpoint(net_opt, './net_opt_parameters.ckpt') # 是将优化器 net_opt 中的参数保存为 net_opt_parameters.ckpt 文件
可以看到，上面的操作是把网络参数和优化器参数分别保存为了两个文件。
==============================================================================
当然，我们也可以把网络参数和优化器参数保存到一个文件里面， 如下：
#!/usr/bin python # encoding:UTF-8 """" 对输入的超参数进行处理 """ import os import argparse """ 设置运行的背景context """ from mindspore import context """ 对数据集进行预处理 """ import mindspore.dataset as ds import mindspore.dataset.transforms.c_transforms as C import mindspore.dataset.vision.c_transforms as CV from mindspore.dataset.vision import Inter from mindspore import dtype as mstype """ 构建神经网络 """ import mindspore.nn as nn from mindspore.common.initializer import Normal """ 训练时对模型参数的保存 """ from mindspore.train.callback import ModelCheckpoint, CheckpointConfig from mindspore.train.serialization import save_checkpoint """ 导入模型训练需要的库 """ from mindspore.nn import Accuracy from mindspore.train.callback import LossMonitor from mindspore import Model import os os.system('rm -f *.ckpt *.meta') parser = argparse.ArgumentParser(description='MindSpore LeNet Example') parser.add_argument('--device_target', type=str, default="CPU", choices=['Ascend', 'GPU', 'CPU']) args = parser.parse_known_args()[0] # 为mindspore设置运行背景context context.set_context(mode=context.GRAPH_MODE, device_target=args.device_target) def create_dataset(data_path, batch_size=32, repeat_size=1, num_parallel_workers=1): # 定义数据集 mnist_ds = ds.MnistDataset(data_path) resize_height, resize_width = 32, 32 rescale = 1.0 / 255.0 shift = 0.0 rescale_nml = 1 / 0.3081 shift_nml = -1 * 0.1307 / 0.3081 # 定义所需要操作的map映射 resize_op = CV.Resize((resize_height, resize_width), interpolation=Inter.LINEAR) rescale_nml_op = CV.Rescale(rescale_nml, shift_nml) rescale_op = CV.Rescale(rescale, shift) hwc2chw_op = CV.HWC2CHW() type_cast_op = C.TypeCast(mstype.int32) # 使用map映射函数，将数据操作应用到数据集 mnist_ds = mnist_ds.map(operations=type_cast_op, input_columns="label", num_parallel_workers=num_parallel_workers) mnist_ds = mnist_ds.map(operations=resize_op, input_columns="image", num_parallel_workers=num_parallel_workers) mnist_ds = mnist_ds.map(operations=rescale_op, input_columns="image", num_parallel_workers=num_parallel_workers) mnist_ds = mnist_ds.map(operations=rescale_nml_op, input_columns="image", num_parallel_workers=num_parallel_workers) mnist_ds = mnist_ds.map(operations=hwc2chw_op, input_columns="image", num_parallel_workers=num_parallel_workers) # 进行shuffle、batch、repeat操作 buffer_size = 10000 mnist_ds = mnist_ds.shuffle(buffer_size=buffer_size) mnist_ds = mnist_ds.batch(batch_size, drop_remainder=True) mnist_ds = mnist_ds.repeat(repeat_size) return mnist_ds class LeNet5(nn.Cell): """ Lenet网络结构 """ def __init__(self, num_class=10, num_channel=1): super(LeNet5, self).__init__() # 定义所需要的运算 self.conv1 = nn.Conv2d(num_channel, 6, 5, pad_mode='valid') self.conv2 = nn.Conv2d(6, 16, 5, pad_mode='valid') self.fc1 = nn.Dense(16 * 5 * 5, 120, weight_init=Normal(0.02)) self.fc2 = nn.Dense(120, 84, weight_init=Normal(0.02)) self.fc3 = nn.Dense(84, num_class, weight_init=Normal(0.02)) self.relu = nn.ReLU() self.max_pool2d = nn.MaxPool2d(kernel_size=2, stride=2) self.flatten = nn.Flatten() def construct(self, x): # 使用定义好的运算构建前向网络 x = self.conv1(x) x = self.relu(x) x = self.max_pool2d(x) x = self.conv2(x) x = self.relu(x) x = self.max_pool2d(x) x = self.flatten(x) x = self.fc1(x) x = self.relu(x) x = self.fc2(x) x = self.relu(x) x = self.fc3(x) return x # 实例化网络 net = LeNet5() # 定义损失函数 net_loss = nn.SoftmaxCrossEntropyWithLogits(sparse=True, reduction='mean') # 定义优化器 net_opt = nn.Momentum(net.trainable_params(), learning_rate=0.01, momentum=0.9) # 设置模型保存参数 # 每125steps保存一次模型参数，最多保留15个文件 config_ck = CheckpointConfig(save_checkpoint_steps=125, keep_checkpoint_max=15) # 应用模型保存参数 ckpoint = ModelCheckpoint(prefix="checkpoint_lenet", config=config_ck) def train_net(args, model, epoch_size, data_path, repeat_size, ckpoint_cb, sink_mode): """定义训练的方法""" # 加载训练数据集 ds_train = create_dataset(os.path.join(data_path, "train"), 32, repeat_size) model.train(epoch_size, ds_train, callbacks=[ckpoint_cb, LossMonitor(125)], dataset_sink_mode=sink_mode) def test_net(network, model, data_path): """定义验证的方法""" ds_eval = create_dataset(os.path.join(data_path, "test")) acc = model.eval(ds_eval, dataset_sink_mode=False) print("{}".format(acc)) mnist_path = "./datasets/MNIST_Data" train_epoch = 1 dataset_size = 1 model = Model(net, net_loss, net_opt, metrics={"Accuracy": Accuracy()}) #train_net(args, model, train_epoch, mnist_path, dataset_size, ckpoint, False) #test_net(net, model, mnist_path) a = net.trainable_params() b = net_opt.trainable_params() x = [] for param in a: c = dict() c['name'] = param.name c['data'] = param x.append(c) for param in b: c = dict() c['name'] = param.name c['data'] = param x.append(c) save_checkpoint(x, './parameters.ckpt')
其主要思想就是传给 save_checkpoint 中的不一定是 nn.cell, 也可以是一个 list 。
list 里面存的是每一个参数的参数字典， 参数字典的key为参数的name , value则为参数。
按照这个思想，手动保存参数还可以写成下面形式：
#!/usr/bin python # encoding:UTF-8 """" 对输入的超参数进行处理 """ import os import argparse """ 设置运行的背景context """ from mindspore import context """ 对数据集进行预处理 """ import mindspore.dataset as ds import mindspore.dataset.transforms.c_transforms as C import mindspore.dataset.vision.c_transforms as CV from mindspore.dataset.vision import Inter from mindspore import dtype as mstype """ 构建神经网络 """ import mindspore.nn as nn from mindspore.common.initializer import Normal """ 训练时对模型参数的保存 """ from mindspore.train.callback import ModelCheckpoint, CheckpointConfig from mindspore.train.serialization import save_checkpoint """ 导入模型训练需要的库 """ from mindspore.nn import Accuracy from mindspore.train.callback import LossMonitor from mindspore import Model import os os.system('rm -f *.ckpt *.meta') parser = argparse.ArgumentParser(description='MindSpore LeNet Example') parser.add_argument('--device_target', type=str, default="CPU", choices=['Ascend', 'GPU', 'CPU']) args = parser.parse_known_args()[0] # 为mindspore设置运行背景context context.set_context(mode=context.GRAPH_MODE, device_target=args.device_target) def create_dataset(data_path, batch_size=32, repeat_size=1, num_parallel_workers=1): # 定义数据集 mnist_ds = ds.MnistDataset(data_path) resize_height, resize_width = 32, 32 rescale = 1.0 / 255.0 shift = 0.0 rescale_nml = 1 / 0.3081 shift_nml = -1 * 0.1307 / 0.3081 # 定义所需要操作的map映射 resize_op = CV.Resize((resize_height, resize_width), interpolation=Inter.LINEAR) rescale_nml_op = CV.Rescale(rescale_nml, shift_nml) rescale_op = CV.Rescale(rescale, shift) hwc2chw_op = CV.HWC2CHW() type_cast_op = C.TypeCast(mstype.int32) # 使用map映射函数，将数据操作应用到数据集 mnist_ds = mnist_ds.map(operations=type_cast_op, input_columns="label", num_parallel_workers=num_parallel_workers) mnist_ds = mnist_ds.map(operations=resize_op, input_columns="image", num_parallel_workers=num_parallel_workers) mnist_ds = mnist_ds.map(operations=rescale_op, input_columns="image", num_parallel_workers=num_parallel_workers) mnist_ds = mnist_ds.map(operations=rescale_nml_op, input_columns="image", num_parallel_workers=num_parallel_workers) mnist_ds = mnist_ds.map(operations=hwc2chw_op, input_columns="image", num_parallel_workers=num_parallel_workers) # 进行shuffle、batch、repeat操作 buffer_size = 10000 mnist_ds = mnist_ds.shuffle(buffer_size=buffer_size) mnist_ds = mnist_ds.batch(batch_size, drop_remainder=True) mnist_ds = mnist_ds.repeat(repeat_size) return mnist_ds class LeNet5(nn.Cell): """ Lenet网络结构 """ def __init__(self, num_class=10, num_channel=1): super(LeNet5, self).__init__() # 定义所需要的运算 self.conv1 = nn.Conv2d(num_channel, 6, 5, pad_mode='valid') self.conv2 = nn.Conv2d(6, 16, 5, pad_mode='valid') self.fc1 = nn.Dense(16 * 5 * 5, 120, weight_init=Normal(0.02)) self.fc2 = nn.Dense(120, 84, weight_init=Normal(0.02)) self.fc3 = nn.Dense(84, num_class, weight_init=Normal(0.02)) self.relu = nn.ReLU() self.max_pool2d = nn.MaxPool2d(kernel_size=2, stride=2) self.flatten = nn.Flatten() def construct(self, x): # 使用定义好的运算构建前向网络 x = self.conv1(x) x = self.relu(x) x = self.max_pool2d(x) x = self.conv2(x) x = self.relu(x) x = self.max_pool2d(x) x = self.flatten(x) x = self.fc1(x) x = self.relu(x) x = self.fc2(x) x = self.relu(x) x = self.fc3(x) return x # 实例化网络 net = LeNet5() # 定义损失函数 net_loss = nn.SoftmaxCrossEntropyWithLogits(sparse=True, reduction='mean') # 定义优化器 net_opt = nn.Momentum(net.trainable_params(), learning_rate=0.01, momentum=0.9) # 设置模型保存参数 # 每125steps保存一次模型参数，最多保留15个文件 config_ck = CheckpointConfig(save_checkpoint_steps=125, keep_checkpoint_max=15) # 应用模型保存参数 ckpoint = ModelCheckpoint(prefix="checkpoint_lenet", config=config_ck) def train_net(args, model, epoch_size, data_path, repeat_size, ckpoint_cb, sink_mode): """定义训练的方法""" # 加载训练数据集 ds_train = create_dataset(os.path.join(data_path, "train"), 32, repeat_size) model.train(epoch_size, ds_train, callbacks=[ckpoint_cb, LossMonitor(125)], dataset_sink_mode=sink_mode) def test_net(network, model, data_path): """定义验证的方法""" ds_eval = create_dataset(os.path.join(data_path, "test")) acc = model.eval(ds_eval, dataset_sink_mode=False) print("{}".format(acc)) mnist_path = "./datasets/MNIST_Data" train_epoch = 1 dataset_size = 1 model = Model(net, net_loss, net_opt, metrics={"Accuracy": Accuracy()}) #train_net(args, model, train_epoch, mnist_path, dataset_size, ckpoint, False) #test_net(net, model, mnist_path) a = net.parameters_and_names() b = net_opt.parameters_and_names() x = [] for name, param in a: c = dict() c['name'] = name c['data'] = param x.append(c) for name, param in b: c = dict() c['name'] = name c['data'] = param x.append(c) save_checkpoint(x, './parameters.ckpt')
===============================================================================================
纠正一个问题：
前面我们讨论的时候都是认为优化器中参数是不包含网络参数的，但是实际中优化器的参数是包括网络参数的，给出代码：
#!/usr/bin python # encoding:UTF-8 """" 对输入的超参数进行处理 """ import os import argparse """ 设置运行的背景context """ from mindspore import context, Tensor """ 对数据集进行预处理 """ import mindspore.dataset as ds import mindspore.dataset.transforms.c_transforms as C import mindspore.dataset.vision.c_transforms as CV from mindspore.dataset.vision import Inter from mindspore import dtype as mstype """ 构建神经网络 """ import mindspore.nn as nn from mindspore.common.initializer import Normal """ 训练时对模型参数的保存 """ from mindspore.train.callback import ModelCheckpoint, CheckpointConfig """ 导入模型训练需要的库 """ from mindspore.nn import Accuracy from mindspore.train.callback import LossMonitor from mindspore import Model import os os.system('rm -f *.ckpt *.meta') parser = argparse.ArgumentParser(description='MindSpore LeNet Example') parser.add_argument('--device_target', type=str, default="CPU", choices=['Ascend', 'GPU', 'CPU']) args = parser.parse_known_args()[0] # 为mindspore设置运行背景context context.set_context(mode=context.GRAPH_MODE, device_target=args.device_target) def create_dataset(data_path, batch_size=32, repeat_size=1, num_parallel_workers=1): # 定义数据集 mnist_ds = ds.MnistDataset(data_path) resize_height, resize_width = 32, 32 rescale = 1.0 / 255.0 shift = 0.0 rescale_nml = 1 / 0.3081 shift_nml = -1 * 0.1307 / 0.3081 # 定义所需要操作的map映射 resize_op = CV.Resize((resize_height, resize_width), interpolation=Inter.LINEAR) rescale_nml_op = CV.Rescale(rescale_nml, shift_nml) rescale_op = CV.Rescale(rescale, shift) hwc2chw_op = CV.HWC2CHW() type_cast_op = C.TypeCast(mstype.int32) # 使用map映射函数，将数据操作应用到数据集 mnist_ds = mnist_ds.map(operations=type_cast_op, input_columns="label", num_parallel_workers=num_parallel_workers) mnist_ds = mnist_ds.map(operations=resize_op, input_columns="image", num_parallel_workers=num_parallel_workers) mnist_ds = mnist_ds.map(operations=rescale_op, input_columns="image", num_parallel_workers=num_parallel_workers) mnist_ds = mnist_ds.map(operations=rescale_nml_op, input_columns="image", num_parallel_workers=num_parallel_workers) mnist_ds = mnist_ds.map(operations=hwc2chw_op, input_columns="image", num_parallel_workers=num_parallel_workers) # 进行shuffle、batch、repeat操作 buffer_size = 10000 mnist_ds = mnist_ds.shuffle(buffer_size=buffer_size) mnist_ds = mnist_ds.batch(batch_size, drop_remainder=True) mnist_ds = mnist_ds.repeat(repeat_size) return mnist_ds class LeNet5(nn.Cell): """ Lenet网络结构 """ def __init__(self, num_class=10, num_channel=1): super(LeNet5, self).__init__() # 定义所需要的运算 self.conv1 = nn.Conv2d(num_channel, 6, 5, pad_mode='valid') self.conv2 = nn.Conv2d(6, 16, 5, pad_mode='valid') self.fc1 = nn.Dense(16 * 5 * 5, 120, weight_init=Normal(0.02)) self.fc2 = nn.Dense(120, 84, weight_init=Normal(0.02)) self.fc3 = nn.Dense(84, num_class, weight_init=Normal(0.02)) self.relu = nn.ReLU() self.max_pool2d = nn.MaxPool2d(kernel_size=2, stride=2) self.flatten = nn.Flatten() def construct(self, x): # 使用定义好的运算构建前向网络 x = self.conv1(x) x = self.relu(x) x = self.max_pool2d(x) x = self.conv2(x) x = self.relu(x) x = self.max_pool2d(x) x = self.flatten(x) x = self.fc1(x) x = self.relu(x) x = self.fc2(x) x = self.relu(x) x = self.fc3(x) return x # 实例化网络 net = LeNet5() # 定义损失函数 net_loss = nn.SoftmaxCrossEntropyWithLogits(sparse=True, reduction='mean') # 定义优化器 net_opt = nn.Momentum(net.trainable_params(), learning_rate=0.01, momentum=0.9) for name, para in net.parameters_and_names(): print(name) print(para.requires_grad) if name == 'conv1.weight': print(Tensor(para)) print("*"*40) for name, para in net_opt.parameters_and_names(): print(name) print(para.requires_grad) if name == 'conv1.weight': print(Tensor(para))
运行结果：
conv1.weight
True
[[[[-0.00123636 -0.00018152 0.00089368 -0.00056078 -0.02275192]
[ 0.00409119 0.0010135 -0.00466038 -0.00555031 -0.00918423]
[-0.00442939 0.00621323 0.00214683 -0.00155054 -0.00546987]
[ 0.01697116 -0.00559946 0.00357803 0.03168541 0.00407573]
[-0.00518463 -0.01200203 0.01070325 -0.02007808 0.00738484]]]
[[[ 0.00341292 0.0079666 -0.00499046 -0.00656943 -0.00331597]
[-0.01387733 0.00665028 -0.01610895 -0.00282408 0.0092861 ]
[-0.01939811 -0.01994145 0.01014557 0.00459681 0.00120816]
[-0.00354739 0.00169169 0.00359304 0.00019773 0.00124371]
[-0.0075929 -0.02099637 0.01632461 -0.02093766 0.00231244]]]
[[[ 0.0088163 -0.01221289 -0.01604474 -0.00574877 -0.00278494]
[ 0.0068464 -0.01448571 -0.00408135 -0.00037711 0.01360335]
[ 0.00826573 0.0063943 -0.00635501 -0.01091845 -0.01706182]
[-0.01376995 0.00267098 -0.01873252 -0.00560728 -0.0133691 ]
[ 0.00562847 0.0048407 0.01391821 0.00568764 0.01011486]]]
[[[ 0.00413718 0.00476703 0.00920789 -0.01249459 0.01619304]
[ 0.01443657 -0.02348764 0.0085768 0.00959142 -0.00631981]
[-0.00826734 0.00130019 0.00431718 -0.01096678 0.00586409]
[-0.01054094 0.01216885 0.00910433 -0.00326026 0.00994863]
[ 0.00993542 0.00768977 -0.00420083 0.00905468 -0.0049615 ]]]
[[[-0.00463446 -0.00677943 -0.00506198 -0.00308914 0.01606419]
[ 0.00844193 -0.00854285 0.0003332 -0.01010361 0.01140079]
[ 0.00595709 0.00572435 -0.00393711 0.00326021 -0.00986465]
[-0.0090545 -0.00300089 0.00010969 -0.03852516 0.00215564]
[-0.01172458 -0.01011858 0.01508922 0.00723284 0.00269153]]]
[[[-0.00602197 -0.00078419 -0.00048669 0.00453082 0.00515535]
[ 0.00237266 -0.00097092 -0.00680392 -0.00715334 0.01152472]
[-0.00824045 -0.0188182 0.00147573 -0.00263265 -0.00235698]
[ 0.00553491 -0.01060611 0.01170796 0.00063573 0.00259822]
[-0.00482674 -0.01767036 0.01275289 0.00904524 0.00328132]]]]
conv2.weight
True
fc1.weight
True
fc1.bias
True
fc2.weight
True
fc2.bias
True
fc3.weight
True
fc3.bias
True
****************************************
learning_rate
True
conv1.weight
True
[[[[-0.00123636 -0.00018152 0.00089368 -0.00056078 -0.02275192]
[ 0.00409119 0.0010135 -0.00466038 -0.00555031 -0.00918423]
[-0.00442939 0.00621323 0.00214683 -0.00155054 -0.00546987]
[ 0.01697116 -0.00559946 0.00357803 0.03168541 0.00407573]
[-0.00518463 -0.01200203 0.01070325 -0.02007808 0.00738484]]]
[[[ 0.00341292 0.0079666 -0.00499046 -0.00656943 -0.00331597]
[-0.01387733 0.00665028 -0.01610895 -0.00282408 0.0092861 ]
[-0.01939811 -0.01994145 0.01014557 0.00459681 0.00120816]
[-0.00354739 0.00169169 0.00359304 0.00019773 0.00124371]
[-0.0075929 -0.02099637 0.01632461 -0.02093766 0.00231244]]]
[[[ 0.0088163 -0.01221289 -0.01604474 -0.00574877 -0.00278494]
[ 0.0068464 -0.01448571 -0.00408135 -0.00037711 0.01360335]
[ 0.00826573 0.0063943 -0.00635501 -0.01091845 -0.01706182]
[-0.01376995 0.00267098 -0.01873252 -0.00560728 -0.0133691 ]
[ 0.00562847 0.0048407 0.01391821 0.00568764 0.01011486]]]
[[[ 0.00413718 0.00476703 0.00920789 -0.01249459 0.01619304]
[ 0.01443657 -0.02348764 0.0085768 0.00959142 -0.00631981]
[-0.00826734 0.00130019 0.00431718 -0.01096678 0.00586409]
[-0.01054094 0.01216885 0.00910433 -0.00326026 0.00994863]
[ 0.00993542 0.00768977 -0.00420083 0.00905468 -0.0049615 ]]]
[[[-0.00463446 -0.00677943 -0.00506198 -0.00308914 0.01606419]
[ 0.00844193 -0.00854285 0.0003332 -0.01010361 0.01140079]
[ 0.00595709 0.00572435 -0.00393711 0.00326021 -0.00986465]
[-0.0090545 -0.00300089 0.00010969 -0.03852516 0.00215564]
[-0.01172458 -0.01011858 0.01508922 0.00723284 0.00269153]]]
[[[-0.00602197 -0.00078419 -0.00048669 0.00453082 0.00515535]
[ 0.00237266 -0.00097092 -0.00680392 -0.00715334 0.01152472]
[-0.00824045 -0.0188182 0.00147573 -0.00263265 -0.00235698]
[ 0.00553491 -0.01060611 0.01170796 0.00063573 0.00259822]
[-0.00482674 -0.01767036 0.01275289 0.00904524 0.00328132]]]]
conv2.weight
True
fc1.weight
True
fc1.bias
True
fc2.weight
True
fc2.bias
True
fc3.weight
True
fc3.bias
True
momentum
True
moments.conv1.weight
True
moments.conv2.weight
True
moments.fc1.weight
True
moments.fc1.bias
True
moments.fc2.weight
True
moments.fc2.bias
True
moments.fc3.weight
True
moments.fc3.bias
True
从运行结果中可以看到， 优化器的参数中本身就包含了网络的训练参数，那么上面的 手动保存参数的文件就只写优化器那部分就可以了，具体如下：
#!/usr/bin python # encoding:UTF-8 """" 对输入的超参数进行处理 """ import os import argparse """ 设置运行的背景context """ from mindspore import context """ 对数据集进行预处理 """ import mindspore.dataset as ds import mindspore.dataset.transforms.c_transforms as C import mindspore.dataset.vision.c_transforms as CV from mindspore.dataset.vision import Inter from mindspore import dtype as mstype """ 构建神经网络 """ import mindspore.nn as nn from mindspore.common.initializer import Normal """ 训练时对模型参数的保存 """ from mindspore.train.callback import ModelCheckpoint, CheckpointConfig from mindspore.train.serialization import save_checkpoint """ 导入模型训练需要的库 """ from mindspore.nn import Accuracy from mindspore.train.callback import LossMonitor from mindspore import Model import os os.system('rm -f *.ckpt *.meta') parser = argparse.ArgumentParser(description='MindSpore LeNet Example') parser.add_argument('--device_target', type=str, default="CPU", choices=['Ascend', 'GPU', 'CPU']) args = parser.parse_known_args()[0] # 为mindspore设置运行背景context context.set_context(mode=context.GRAPH_MODE, device_target=args.device_target) def create_dataset(data_path, batch_size=32, repeat_size=1, num_parallel_workers=1): # 定义数据集 mnist_ds = ds.MnistDataset(data_path) resize_height, resize_width = 32, 32 rescale = 1.0 / 255.0 shift = 0.0 rescale_nml = 1 / 0.3081 shift_nml = -1 * 0.1307 / 0.3081 # 定义所需要操作的map映射 resize_op = CV.Resize((resize_height, resize_width), interpolation=Inter.LINEAR) rescale_nml_op = CV.Rescale(rescale_nml, shift_nml) rescale_op = CV.Rescale(rescale, shift) hwc2chw_op = CV.HWC2CHW() type_cast_op = C.TypeCast(mstype.int32) # 使用map映射函数，将数据操作应用到数据集 mnist_ds = mnist_ds.map(operations=type_cast_op, input_columns="label", num_parallel_workers=num_parallel_workers) mnist_ds = mnist_ds.map(operations=resize_op, input_columns="image", num_parallel_workers=num_parallel_workers) mnist_ds = mnist_ds.map(operations=rescale_op, input_columns="image", num_parallel_workers=num_parallel_workers) mnist_ds = mnist_ds.map(operations=rescale_nml_op, input_columns="image", num_parallel_workers=num_parallel_workers) mnist_ds = mnist_ds.map(operations=hwc2chw_op, input_columns="image", num_parallel_workers=num_parallel_workers) # 进行shuffle、batch、repeat操作 buffer_size = 10000 mnist_ds = mnist_ds.shuffle(buffer_size=buffer_size) mnist_ds = mnist_ds.batch(batch_size, drop_remainder=True) mnist_ds = mnist_ds.repeat(repeat_size) return mnist_ds class LeNet5(nn.Cell): """ Lenet网络结构 """ def __init__(self, num_class=10, num_channel=1): super(LeNet5, self).__init__() # 定义所需要的运算 self.conv1 = nn.Conv2d(num_channel, 6, 5, pad_mode='valid') self.conv2 = nn.Conv2d(6, 16, 5, pad_mode='valid') self.fc1 = nn.Dense(16 * 5 * 5, 120, weight_init=Normal(0.02)) self.fc2 = nn.Dense(120, 84, weight_init=Normal(0.02)) self.fc3 = nn.Dense(84, num_class, weight_init=Normal(0.02)) self.relu = nn.ReLU() self.max_pool2d = nn.MaxPool2d(kernel_size=2, stride=2) self.flatten = nn.Flatten() def construct(self, x): # 使用定义好的运算构建前向网络 x = self.conv1(x) x = self.relu(x) x = self.max_pool2d(x) x = self.conv2(x) x = self.relu(x) x = self.max_pool2d(x) x = self.flatten(x) x = self.fc1(x) x = self.relu(x) x = self.fc2(x) x = self.relu(x) x = self.fc3(x) return x # 实例化网络 net = LeNet5() # 定义损失函数 net_loss = nn.SoftmaxCrossEntropyWithLogits(sparse=True, reduction='mean') # 定义优化器 net_opt = nn.Momentum(net.trainable_params(), learning_rate=0.01, momentum=0.9) # 设置模型保存参数 # 每125steps保存一次模型参数，最多保留15个文件 config_ck = CheckpointConfig(save_checkpoint_steps=125, keep_checkpoint_max=15) # 应用模型保存参数 ckpoint = ModelCheckpoint(prefix="checkpoint_lenet", config=config_ck) def train_net(args, model, epoch_size, data_path, repeat_size, ckpoint_cb, sink_mode): """定义训练的方法""" # 加载训练数据集 ds_train = create_dataset(os.path.join(data_path, "train"), 32, repeat_size) model.train(epoch_size, ds_train, callbacks=[ckpoint_cb, LossMonitor(125)], dataset_sink_mode=sink_mode) def test_net(network, model, data_path): """定义验证的方法""" ds_eval = create_dataset(os.path.join(data_path, "test")) acc = model.eval(ds_eval, dataset_sink_mode=False) print("{}".format(acc)) mnist_path = "./datasets/MNIST_Data" train_epoch = 1 dataset_size = 1 model = Model(net, net_loss, net_opt, metrics={"Accuracy": Accuracy()}) #train_net(args, model, train_epoch, mnist_path, dataset_size, ckpoint, False) #test_net(net, model, mnist_path) save_checkpoint(net_opt, './net_opt_parameters.ckpt')
而，./net_opt_parameters.ckpt 文件中就已经包含了优化器的参数以及网络中的可训练参数。
所以，手动保存参数的最终代码形式就是：
save_checkpoint(net_opt, './net_opt_parameters.ckpt')