模型保存与载入

一、保存载入体系简介

1.1 基础API保存载入体系

飞桨框架2.1对模型与参数的保存与载入相关接口进行了梳理:对于训练调优场景,我们推荐使用paddle.save/load保存和载入模型;对于推理部署场景,我们推荐使用paddle.jit.save/load(动态图)和paddle.static.save/load_inference_model(静态图)保存载入模型。

飞桨保存载入相关接口包括:

paddle.save

paddle.load

paddle.jit.save

paddle.jit.load

paddle.static.save_inference_model

paddle.static.load_inference_model

各接口关系如下图所示:

../../_images/paddle_save_load_2.1.png ../../_images/paddle_jit_save_load_2.1.png

1.2 高阶API保存载入体系

  • paddle.Model.fit (训练接口,同时带有参数保存的功能)

  • paddle.Model.save

  • paddle.Model.load

飞桨框架2.0高阶API仅有一套Save/Load接口,表意直观,体系清晰,若有需要,建议直接阅读相关API文档,此处不再赘述。

注解

本教程着重介绍飞桨框架2.1的各个保存载入接口的关系及各种使用场景,不对接口参数进行详细介绍,如果需要了解具体接口参数的含义,请直接阅读对应API文档。

模型保存常见问题

二、训练调优场景的模型&参数保存载入

2.1 动态图参数保存载入

若仅需要保存/载入模型的参数,可以使用 paddle.save/load 结合Layer和Optimizer的state_dict达成目的,此处state_dict是对象的持久参数的载体,dict的key为参数名,value为参数真实的numpy array值。

结合以下简单示例,介绍参数保存和载入的方法,以下示例完成了一个简单网络的训练过程:

import numpy as np
import paddle
import paddle.nn as nn
import paddle.optimizer as opt

BATCH_SIZE = 16
BATCH_NUM = 4
EPOCH_NUM = 4

IMAGE_SIZE = 784
CLASS_NUM = 10

# define a random dataset
class RandomDataset(paddle.io.Dataset):
    def __init__(self, num_samples):
        self.num_samples = num_samples

    def __getitem__(self, idx):
        image = np.random.random([IMAGE_SIZE]).astype('float32')
        label = np.random.randint(0, CLASS_NUM - 1, (1, )).astype('int64')
        return image, label

    def __len__(self):
        return self.num_samples

class LinearNet(nn.Layer):
    def __init__(self):
        super(LinearNet, self).__init__()
        self._linear = nn.Linear(IMAGE_SIZE, CLASS_NUM)

    def forward(self, x):
        return self._linear(x)

def train(layer, loader, loss_fn, opt):
    for epoch_id in range(EPOCH_NUM):
        for batch_id, (image, label) in enumerate(loader()):
            out = layer(image)
            loss = loss_fn(out, label)
            loss.backward()
            opt.step()
            opt.clear_grad()
            print("Epoch {} batch {}: loss = {}".format(
                epoch_id, batch_id, np.mean(loss.numpy())))

# create network
layer = LinearNet()
loss_fn = nn.CrossEntropyLoss()
adam = opt.Adam(learning_rate=0.001, parameters=layer.parameters())

# create data loader
dataset = RandomDataset(BATCH_NUM * BATCH_SIZE)
loader = paddle.io.DataLoader(dataset,
    batch_size=BATCH_SIZE,
    shuffle=True,
    drop_last=True,
    num_workers=2)

# train
train(layer, loader, loss_fn, adam)

2.1.1 参数保存

参数保存时,先获取目标对象(Layer或者Optimzier)的state_dict,然后将state_dict保存至磁盘,示例如下(接前述示例):

# save
paddle.save(layer.state_dict(), "linear_net.pdparams")
paddle.save(adam.state_dict(), "adam.pdopt")

2.1.2 参数载入

参数载入时,先从磁盘载入保存的state_dict,然后通过set_state_dict方法配置到目标对象中,示例如下(接前述示例):

# load
layer_state_dict = paddle.load("linear_net.pdparams")
opt_state_dict = paddle.load("adam.pdopt")

layer.set_state_dict(layer_state_dict)
adam.set_state_dict(opt_state_dict)

2.2 静态图模型&参数保存载入

若仅需要保存/载入模型的参数,可以使用 paddle.save/load 结合Program的state_dict达成目的,此处state_dict与动态图state_dict概念类似,dict的key为参数名,value为参数真实的值。若想保存整个模型,需要使用``paddle.save``将Program和state_dict都保存下来。

结合以下简单示例,介绍参数保存和载入的方法:

import paddle
import paddle.static as static

paddle.enable_static()

# create network
x = paddle.static.data(name="x", shape=[None, 224], dtype='float32')
z = paddle.static.nn.fc(x, 10)

place = paddle.CPUPlace()
exe = paddle.static.Executor(place)
exe.run(paddle.static.default_startup_program())
prog = paddle.static.default_main_program()

2.2.1 静态图模型&参数保存

参数保存时,先获取Program的state_dict,然后将state_dict保存至磁盘,示例如下(接前述示例):

paddle.save(prog.state_dict(), "temp/model.pdparams")

如果想要保存整个静态图模型,除了state_dict还需要保存Program

paddle.save(prog, "temp/model.pdmodel")

2.2.2 静态图模型&参数载入

如果只保存了state_dict,可以跳过此段代码,直接载入state_dict。如果模型文件中包含Program和state_dict,请先载入Program,示例如下(接前述示例):

prog = paddle.load("temp/model.pdmodel")

参数载入时,先从磁盘载入保存的state_dict,然后通过set_state_dict方法配置到Program中,示例如下(接前述示例):

state_dict = paddle.load("temp/model.pdparams")
prog.set_state_dict(state_dict)

三、训练部署场景的模型&参数保存载入

3.1 动态图模型&参数保存载入(训练推理)

若要同时保存/载入动态图模型结构和参数,可以使用 paddle.jit.save/load 实现。

3.1.1 动态图模型&参数保存

模型&参数存储根据训练模式不同,有两种使用情况:

  1. 动转静训练 + 模型&参数保存

  2. 动态图训练 + 模型&参数保存

3.1.1.1 动转静训练 + 模型&参数保存

动转静训练相比直接使用动态图训练具有更好的执行性能,训练完成后,直接将目标Layer传入 paddle.jit.save 保存即可。:

一个简单的网络训练示例如下:

import numpy as np
import paddle
import paddle.nn as nn
import paddle.optimizer as opt

BATCH_SIZE = 16
BATCH_NUM = 4
EPOCH_NUM = 4

IMAGE_SIZE = 784
CLASS_NUM = 10

# define a random dataset
class RandomDataset(paddle.io.Dataset):
    def __init__(self, num_samples):
        self.num_samples = num_samples

    def __getitem__(self, idx):
        image = np.random.random([IMAGE_SIZE]).astype('float32')
        label = np.random.randint(0, CLASS_NUM - 1, (1, )).astype('int64')
        return image, label

    def __len__(self):
        return self.num_samples

class LinearNet(nn.Layer):
    def __init__(self):
        super(LinearNet, self).__init__()
        self._linear = nn.Linear(IMAGE_SIZE, CLASS_NUM)

    @paddle.jit.to_static
    def forward(self, x):
        return self._linear(x)

def train(layer, loader, loss_fn, opt):
    for epoch_id in range(EPOCH_NUM):
        for batch_id, (image, label) in enumerate(loader()):
            out = layer(image)
            loss = loss_fn(out, label)
            loss.backward()
            opt.step()
            opt.clear_grad()
            print("Epoch {} batch {}: loss = {}".format(
                epoch_id, batch_id, np.mean(loss.numpy())))

# create network
layer = LinearNet()
loss_fn = nn.CrossEntropyLoss()
adam = opt.Adam(learning_rate=0.001, parameters=layer.parameters())

# create data loader
dataset = RandomDataset(BATCH_NUM * BATCH_SIZE)
loader = paddle.io.DataLoader(dataset,
    batch_size=BATCH_SIZE,
    shuffle=True,
    drop_last=True,
    num_workers=2)

# train
train(layer, loader, loss_fn, adam)

随后使用 paddle.jit.save 对模型和参数进行存储(接前述示例):

# save
path = "example.model/linear"
paddle.jit.save(layer, path)

通过动转静训练后保存模型&参数,有以下三项注意点:

  1. Layer对象的forward方法需要经由 paddle.jit.to_static 装饰

经过 paddle.jit.to_static 装饰forward方法后,相应Layer在执行时,会先生成描述模型的Program,然后通过执行Program获取计算结果,示例如下:

import paddle
import paddle.nn as nn

IMAGE_SIZE = 784
CLASS_NUM = 10

class LinearNet(nn.Layer):
    def __init__(self):
        super(LinearNet, self).__init__()
        self._linear = nn.Linear(IMAGE_SIZE, CLASS_NUM)

    @paddle.jit.to_static
    def forward(self, x):
        return self._linear(x)

若最终需要生成的描述模型的Program支持动态输入,可以同时指明模型的 InputSepc ,示例如下:

import paddle
import paddle.nn as nn
from paddle.static import InputSpec

IMAGE_SIZE = 784
CLASS_NUM = 10

class LinearNet(nn.Layer):
    def __init__(self):
        super(LinearNet, self).__init__()
        self._linear = nn.Linear(IMAGE_SIZE, CLASS_NUM)

    @paddle.jit.to_static(input_spec=[InputSpec(shape=[None, 784], dtype='float32')])
    def forward(self, x):
        return self._linear(x)
  1. 请确保Layer.forward方法中仅实现预测功能,避免将训练所需的loss计算逻辑写入forward方法

Layer更准确的语义是描述一个具有预测功能的模型对象,接收输入的样本数据,输出预测的结果,而loss计算是仅属于模型训练中的概念。将loss计算的实现放到Layer.forward方法中,会使Layer在不同场景下概念有所差别,并且增大Layer使用的复杂性,这不是良好的编码行为,同时也会在最终保存预测模型时引入剪枝的复杂性,因此建议保持Layer实现的简洁性,下面通过两个示例对比说明:

错误示例如下:

import paddle
import paddle.nn as nn

IMAGE_SIZE = 784
CLASS_NUM = 10

class LinearNet(nn.Layer):
    def __init__(self):
        super(LinearNet, self).__init__()
        self._linear = nn.Linear(IMAGE_SIZE, CLASS_NUM)

    @paddle.jit.to_static
    def forward(self, x, label=None):
        out = self._linear(x)
        if label:
            loss = nn.functional.cross_entropy(out, label)
            avg_loss = nn.functional.mean(loss)
            return out, avg_loss
        else:
            return out

正确示例如下:

import paddle
import paddle.nn as nn

IMAGE_SIZE = 784
CLASS_NUM = 10

class LinearNet(nn.Layer):
    def __init__(self):
        super(LinearNet, self).__init__()
        self._linear = nn.Linear(IMAGE_SIZE, CLASS_NUM)

    @paddle.jit.to_static
    def forward(self, x):
        return self._linear(x)
  1. 如果你需要保存多个方法,需要用 paddle.jit.to_static 装饰每一个需要被保存的方法。

注解

只有在forward之外还需要保存其他方法时才用这个特性,如果仅装饰非forward的方法,而forward没有被装饰,是不符合规范的。此时 paddle.jit.saveinput_spec 参数必须为None。

示例代码如下:

import paddle
import paddle.nn as nn
from paddle.static import InputSpec

IMAGE_SIZE = 784
CLASS_NUM = 10

class LinearNet(nn.Layer):
    def __init__(self):
        super(LinearNet, self).__init__()
        self._linear = nn.Linear(IMAGE_SIZE, CLASS_NUM)
        self._linear_2 = nn.Linear(IMAGE_SIZE, CLASS_NUM)

    @paddle.jit.to_static(input_spec=[InputSpec(shape=[None, IMAGE_SIZE], dtype='float32')])
    def forward(self, x):
        return self._linear(x)

    @paddle.jit.to_static(input_spec=[InputSpec(shape=[None, IMAGE_SIZE], dtype='float32')])
    def another_forward(self, x):
        return self._linear_2(x)

inps = paddle.randn([1, IMAGE_SIZE])
layer = LinearNet()
before_0 = layer.another_forward(inps)
before_1 = layer(inps)
# save and load
path = "example.model/linear"
paddle.jit.save(layer, path)

保存的模型命名规则:forward的模型名字为:模型名+后缀,其他函数的模型名字为:模型名+函数名+后缀。每个函数有各自的pdmodel和pdiparams的文件,所有函数共用pdiparams.info。上述代码将在 example.model 文件夹下产生5个文件: linear.another_forward.pdiparams、 linear.pdiparams、 linear.pdmodel、 linear.another_forward.pdmodel、 linear.pdiparams.info

  1. 当使用 jit.save 保存函数时,jit.save 只保存这个函数对应的静态图 Program ,不会保存和这个函数相关的参数。如果你必须保存参数,请使用Layer封装这个函数。

示例代码如下:

def fun(inputs):
    return paddle.tanh(inputs)

path = 'func/model'
inps = paddle.rand([3, 6])
origin = fun(inps)

paddle.jit.save(
    fun,
    path,
    input_spec=[
        InputSpec(
            shape=[None, 6], dtype='float32', name='x'),
    ])
load_func = paddle.jit.load(path)
load_result = load_func(inps)

3.1.1.2 动态图训练 + 模型&参数保存

动态图模式相比动转静模式更加便于调试,如果你仍需要使用动态图直接训练,也可以在动态图训练完成后调用 paddle.jit.save 直接保存模型和参数。

同样是一个简单的网络训练示例:

import numpy as np
import paddle
import paddle.nn as nn
import paddle.optimizer as opt
from paddle.static import InputSpec

BATCH_SIZE = 16
BATCH_NUM = 4
EPOCH_NUM = 4

IMAGE_SIZE = 784
CLASS_NUM = 10

# define a random dataset
class RandomDataset(paddle.io.Dataset):
    def __init__(self, num_samples):
        self.num_samples = num_samples

    def __getitem__(self, idx):
        image = np.random.random([IMAGE_SIZE]).astype('float32')
        label = np.random.randint(0, CLASS_NUM - 1, (1, )).astype('int64')
        return image, label

    def __len__(self):
        return self.num_samples

class LinearNet(nn.Layer):
    def __init__(self):
        super(LinearNet, self).__init__()
        self._linear = nn.Linear(IMAGE_SIZE, CLASS_NUM)

    def forward(self, x):
        return self._linear(x)

def train(layer, loader, loss_fn, opt):
    for epoch_id in range(EPOCH_NUM):
        for batch_id, (image, label) in enumerate(loader()):
            out = layer(image)
            loss = loss_fn(out, label)
            loss.backward()
            opt.step()
            opt.clear_grad()
            print("Epoch {} batch {}: loss = {}".format(
                epoch_id, batch_id, np.mean(loss.numpy())))

# create network
layer = LinearNet()
loss_fn = nn.CrossEntropyLoss()
adam = opt.Adam(learning_rate=0.001, parameters=layer.parameters())

# create data loader
dataset = RandomDataset(BATCH_NUM * BATCH_SIZE)
loader = paddle.io.DataLoader(dataset,
    batch_size=BATCH_SIZE,
    shuffle=True,
    drop_last=True,
    num_workers=2)

# train
train(layer, loader, loss_fn, adam)

训练完成后使用 paddle.jit.save 对模型和参数进行存储:

# save
path = "example.dy_model/linear"
paddle.jit.save(
    layer=layer,
    path=path,
    input_spec=[InputSpec(shape=[None, 784], dtype='float32')])

动态图训练后使用 paddle.jit.save 保存模型和参数注意点如下:

  1. 相比动转静训练,Layer对象的forward方法不需要额外装饰,保持原实现即可

  2. 与动转静训练相同,请确保Layer.forward方法中仅实现预测功能,避免将训练所需的loss计算逻辑写入forward方法

  3. 在最后使用 paddle.jit.save 时,需要指定Layer的 InputSpec ,Layer对象forward方法的每一个参数均需要对应的 InputSpec 进行描述,不能省略。这里的 input_spec 参数支持两种类型的输入:

  • InputSpec 列表

使用InputSpec描述forward输入参数的shape,dtype和name,如前述示例(此处示例中name省略,name省略的情况下会使用forward的对应参数名作为name,所以这里的name为 x ):

paddle.jit.save(
    layer=layer,
    path=path,
    input_spec=[InputSpec(shape=[None, 784], dtype='float32')])
  • Example Tensor 列表

除使用InputSpec之外,也可以直接使用forward训练时的示例输入,此处可以使用前述示例中迭代DataLoader得到的 image ,示例如下:

paddle.jit.save(
    layer=layer,
    path=path,
    input_spec=[image])

3.1.2 动态图模型&参数载入

载入模型参数,使用 paddle.jit.load 载入即可,载入后得到的是一个Layer的派生类对象 TranslatedLayerTranslatedLayer 具有Layer具有的通用特征,支持切换 train 或者 eval 模式,可以进行模型调优或者预测。

注解

为了规避变量名字冲突,载入之后会重命名变量。

载入模型及参数,示例如下:

import numpy as np
import paddle
import paddle.nn as nn
import paddle.optimizer as opt

BATCH_SIZE = 16
BATCH_NUM = 4
EPOCH_NUM = 4

IMAGE_SIZE = 784
CLASS_NUM = 10

# load
path = "example.model/linear"
loaded_layer = paddle.jit.load(path)

载入模型及参数后进行预测,示例如下(接前述示例):

# inference
loaded_layer.eval()
x = paddle.randn([1, IMAGE_SIZE], 'float32')
pred = loaded_layer(x)

载入模型及参数后进行调优,示例如下(接前述示例):

# define a random dataset
class RandomDataset(paddle.io.Dataset):
    def __init__(self, num_samples):
        self.num_samples = num_samples

    def __getitem__(self, idx):
        image = np.random.random([IMAGE_SIZE]).astype('float32')
        label = np.random.randint(0, CLASS_NUM - 1, (1, )).astype('int64')
        return image, label

    def __len__(self):
        return self.num_samples

def train(layer, loader, loss_fn, opt):
    for epoch_id in range(EPOCH_NUM):
        for batch_id, (image, label) in enumerate(loader()):
            out = layer(image)
            loss = loss_fn(out, label)
            loss.backward()
            opt.step()
            opt.clear_grad()
            print("Epoch {} batch {}: loss = {}".format(
                epoch_id, batch_id, np.mean(loss.numpy())))

# fine-tune
loaded_layer.train()
dataset = RandomDataset(BATCH_NUM * BATCH_SIZE)
loader = paddle.io.DataLoader(dataset,
    batch_size=BATCH_SIZE,
    shuffle=True,
    drop_last=True,
    num_workers=2)
loss_fn = nn.CrossEntropyLoss()
adam = opt.Adam(learning_rate=0.001, parameters=loaded_layer.parameters())
train(loaded_layer, loader, loss_fn, adam)
# save after fine-tuning
paddle.jit.save(loaded_layer, "fine-tune.model/linear", input_spec=[x])

此外, paddle.jit.save 同时保存了模型和参数,如果你只需要从存储结果中载入模型的参数,可以使用 paddle.load 接口载入,返回所存储模型的state_dict,示例如下:

import paddle
import paddle.nn as nn

IMAGE_SIZE = 784
CLASS_NUM = 10

class LinearNet(nn.Layer):
    def __init__(self):
        super(LinearNet, self).__init__()
        self._linear = nn.Linear(IMAGE_SIZE, CLASS_NUM)

    @paddle.jit.to_static
    def forward(self, x):
        return self._linear(x)

# create network
layer = LinearNet()

# load
path = "example.model/linear"
state_dict = paddle.load(path)

# inference
layer.set_state_dict(state_dict, use_structured_name=False)
layer.eval()
x = paddle.randn([1, IMAGE_SIZE], 'float32')
pred = layer(x)

3.2 静态图模型&参数保存载入(推理部署)

保存/载入静态图推理模型,可以通过 paddle.static.save/load_inference_model 实现。示例如下:

import paddle
import numpy as np

paddle.enable_static()

# Build the model
startup_prog = paddle.static.default_startup_program()
main_prog = paddle.static.default_main_program()
with paddle.static.program_guard(main_prog, startup_prog):
    image = paddle.static.data(name="img", shape=[64, 784])
    w = paddle.create_parameter(shape=[784, 200], dtype='float32')
    b = paddle.create_parameter(shape=[200], dtype='float32')
    hidden_w = paddle.matmul(x=image, y=w)
    hidden_b = paddle.add(hidden_w, b)
exe = paddle.static.Executor(paddle.CPUPlace())
exe.run(startup_prog)

3.2.1 静态图推理模型&参数保存

静态图导出推理模型需要指定导出路径、输入、输出变量以及执行器。 save_inference_model 会裁剪Program的冗余部分,并导出两个文件: path_prefix.pdmodelpath_prefix.pdiparams 。示例如下(接前述示例):

# Save the inference model
path_prefix = "./infer_model"
paddle.static.save_inference_model(path_prefix, [image], [hidden_b], exe)

3.2.2 静态图推理模型&参数载入

载入静态图推理模型时,输入给 load_inference_model 的路径必须与 save_inference_model 的一致。示例如下(接前述示例):

[inference_program, feed_target_names, fetch_targets] = (
    paddle.static.load_inference_model(path_prefix, exe))
tensor_img = np.array(np.random.random((64, 784)), dtype=np.float32)
results = exe.run(inference_program,
                feed={feed_target_names[0]: tensor_img},
                fetch_list=fetch_targets)

四、旧保存格式兼容载入

如果你是从飞桨框架1.x切换到2.1,曾经使用飞桨框架1.x的fluid相关接口保存模型或者参数,飞桨框架2.1也对这种情况进行了兼容性支持,请参考 兼容载入旧格式模型