目录
- 模型的保存与复用
- 模型定义和参数打印
- 模型保存
- 模型推理
- 模型再训练
- 模型迁移
- 参考文献
本文整理了Pytorch框架下模型的保存、复用、推理、再训练和迁移等实现。
模型的保存与复用
模型定义和参数打印
# 定义模型结构
class LenNet(nn.Module):
def __init__(self):
super(LenNet, self).__init__()
self.conv = nn.Sequential( # [batch, 1, 28, 28]
nn.Conv2d(1, 8, 5, 2), # [batch, 1, 28, 28]
nn.ReLU(inplace=True),
nn.MaxPool2d(2, 2), # [batch, 8, 14, 14]
nn.Conv2d(8, 16, 5), # [batch, 16, 10, 10]
nn.ReLU(inplace=True),
nn.MaxPool2d(2, 2), # [batch, 16, 5, 5]
)
self.fc = nn.Sequential(
nn.Flatten(),
nn.Linear(16*5*5, 128),
nn.ReLU(inplace=True),
nn.Linear(128, 64),
nn.ReLU(inplace=True),
nn.Linear(64, 10)
)
def forward(self, X):
return self.fc(self.conv(X))
# 查看模型参数
# 网络模型中的参数model.state_dict()是以字典形式保存（实质上是collections模块中的OrderedDict）
model = LenNet()
print("Model's state_dict:")
for param_tensor in model.state_dict():
print(param_tensor, "\t", model.state_dict()[param_tensor].size())
# 参数名中的fc和conv前缀是根据定义nn.Sequential()时的名字所确定。
# 参数名中的数字表示每个Sequential()中网络层所在的位置。
print(model.state_dict().keys()) # 打印键
print(model.state_dict().values()) # 打印值
# 优化器optimizer的参数打印类似
optimizer = torch.optim.SGD(model.parameters(), lr=0.001, momentum=0.9)
print("Optimizer's state_dict:")
for var_name in optimizer.state_dict():
print(var_name, "\t", optimizer.state_dict()[var_name])
模型保存
import os
# 指定保存的模型名称时Pytorch官方建议的后缀为.pt或者.pth
model_save_dir = './model_logs/'
model_save_path = os.path.join(model_save_dir, 'LeNet.pt')
torch.save(model.state_dict(), model_save_path)
# 在训练过程中保存某个条件下的最优模型，可以如下操作
best_model_state = deepcopy(model.state_dict())
torch.save(best_model_state, model_save_path)
# 下面这种方法是错误的，因为best_model_state只是model.state_dict()的引用，会随着训练的改变而改变
best_model_state = model.state_dict()
torch.save(best_model_state, model_save_path)
模型推理
def inference(data_iter, device, model_save_dir):
model = LeNet() # 初始化现有模型的权重参数
model.to(device)
model_save_path = os.path.join(model_save_dir, 'LeNet.pt')
# 如果本地存在模型，则加载本地模型参数覆盖原有模型
if os.path.exists(model_save_path):
loaded_paras = torch.load(model_save_path)
model.load_state_dict(loaded_paras)
model.eval()
with torch.no_grad(): # 开始推理
acc_sum, n = 0., 0
for x, y in data_iter:
x, y = x.to(device), y.to(device)
logits = model(x)
acc_sum += (logits.argmax(1) == y).float().sum().item()
n += len(y)
print("Accuracy in test data is : ", acc_sum / n)
模型再训练
class MyModel:
def __init__(self,
batch_size=64,
epochs=5,
learning_rate=0.001,
model_save_dir='./MODEL'):
self.batch_size = batch_size
self.epochs = epochs
self.learning_rate = learning_rate
self.model_save_dir = model_save_dir
self.model = LeNet()
def train(self):
train_iter, test_iter = load_dataset(self.batch_size)
# 在训练过程中只保存网络权重，在再训练时只载入网络权重参数初始化网络训练。这里是核心部分，开始。
if not os.path.exists(self.model_save_dir):
os.makedirs(self.model_save_dir)
model_save_path = os.path.join(self.model_save_dir, 'model.pt')
if os.path.exists(model_save_path):
loaded_paras = torch.load(model_save_path)
self.model.load_state_dict(loaded_paras)
print("#### 成功载入已有模型，进行再训练...")
# 结束
optimizer = torch.optim.Adam(self.model.parameters(), lr=self.learning_rate)
device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
self.model.to(device)
for epoch in range(self.epochs):
for i, (x, y) in enumerate(train_iter):
x, y = x.to(device), y.to(device)
loss, logits = self.model(x)
optimizer.zero_grad()
loss.backward()
optimizer.step()
if i % 100 == 0:
acc = (logits.argmax(1) == y).float().mean()
print("Epochs[{}/{}]---batch[{}/{}]---acc {:.4}---loss {:.4}".format(
epoch, self.epochs, len(train_iter), i, acc, loss.item()))
print("Epochs[{}/{}]--acc on test {:.4}".format(epoch, self.epochs,
self.evaluate(test_iter, self.model, device)))
torch.save(self.model.state_dict(), model_save_path)
@staticmethod
def evaluate(data_iter, model, device):
with torch.no_grad():
acc_sum, n = 0.0, 0
for x, y in data_iter:
x, y = x.to(device), y.to(device)
logits = model(x)
acc_sum += (logits.argmax(1) == y).float().sum().item()
n += len(y)
return acc_sum / n
# 在保存参数的时候，将优化器参数、损失值等可一同保存，然后在恢复模型时连同其它参数一起恢复
model_save_path = os.path.join(model_save_dir, 'LeNet.pt')
torch.save({
'epoch': epoch,
'model_state_dict': model.state_dict(),
'optimizer_state_dict': optimizer.state_dict(),
'loss': loss,
...
}, model_save_path)
# 加载方式如下
checkpoint = torch.load(model_save_path)
model.load_state_dict(checkpoint['model_state_dict'])
optimizer.load_state_dict(checkpoint['optimizer_state_dict'])
epoch = checkpoint['epoch']
loss = checkpoint['loss']
模型迁移
# 定义新模型NewLeNet 和LeNet区别在于新增了一个全连接层
class NewLenNet(nn.Module):
def __init__(self):
super(NewLenNet, self).__init__()
self.conv = nn.Sequential( # [batch, 1, 28, 28]
nn.Conv2d(1, 8, 5, 2), # [batch, 1, 28, 28]
nn.ReLU(inplace=True),
nn.MaxPool2d(2, 2), # [batch, 8, 14, 14]
nn.Conv2d(8, 16, 5), # [batch, 16, 10, 10]
nn.ReLU(inplace=True),
nn.MaxPool2d(2, 2), # [batch, 16, 5, 5]
)
self.fc = nn.Sequential(
nn.Flatten(),
nn.Linear(16*5*5, 128),
nn.ReLU(inplace=True),
nn.Linear(128, 64), # 这层以前和LeNet结构一致 可以用LeNet的参数来进行替换
nn.ReLU(inplace=True),
nn.Linear(64, 32),
nn.ReLU(inplace=True),
nn.Linear(32, 10)
)
def forward(self, X):
return self.fc(self.conv(X))
# 定义替换函数 匹配两个网络 size相同处地方进行参数替换
def para_state_dict(model, model_save_dir):
state_dict = deepcopy(model.state_dict())
model_save_path = os.path.join(model_save_dir, 'model.pt')
if os.path.exists(model_save_path):
loaded_paras = torch.load(model_save_path)
for key in state_dict: # 在新的网络模型中遍历对应参数
if key in loaded_paras and state_dict[key].size() == loaded_paras[key].size():
print("成功初始化参数:", key)
state_dict[key] = loaded_paras[key]
return state_dict
# 更新一下模型迁移后的训练代码
def train(self):
train_iter, test_iter = load_dataset(self.batch_size)
if not os.path.exists(self.model_save_dir):
os.makedirs(self.model_save_dir)
model_save_path = os.path.join(self.model_save_dir, 'model_new.pt')
old_model = os.path.join(self.model_save_dir, 'LeNet.pt')
if os.path.exists(old_model):
state_dict = para_state_dict(self.model, self.model_save_dir) # 调用迁移代码 将LeNet的前几层参数迁移到NewLeNet
self.model.load_state_dict(state_dict)
print("#### 成功载入已有模型，进行再训练...")
optimizer = torch.optim.Adam(self.model.parameters(), lr=self.learning_rate)
device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
self.model.to(device)
for epoch in range(self.epochs):
for i, (x, y) in enumerate(train_iter):
x, y = x.to(device), y.to(device)
loss, logits = self.model(x)
optimizer.zero_grad()
loss.backward()
optimizer.step()
if i % 100 == 0:
acc = (logits.argmax(1) == y).float().mean()
print("Epochs[{}/{}]---batch[{}/{}]---acc {:.4}---loss {:.4}".format(
epoch, self.epochs, len(train_iter), i, acc, loss.item()))
print("Epochs[{}/{}]--acc on test {:.4}".format(epoch, self.epochs,
self.evaluate(test_iter, self.model, device)))
torch.save(self.model.state_dict(), model_save_path)
# 这里更新未进行训练的推理
def inference(data_iter, device, model_save_dir='./MODEL'):
model = NewLeNet() # 初始化现有模型的权重参数
print("初始化参数 conv.0.bias 为：", model.state_dict()['conv.0.bias'])
model.to(device)
state_dict = para_state_dict(model, model_save_dir) # 迁移模型参数
model.load_state_dict(state_dict)
model.eval()
print("载入本地模型重新初始化 conv.0.bias 为：", model.state_dict()['conv.0.bias'])
with torch.no_grad():
acc_sum, n = 0.0, 0
for x, y in data_iter:
x, y = x.to(device), y.to(device)
logits = model(x)
acc_sum += (logits.argmax(1) == y).float().sum().item()
n += len(y)
print("Accuracy in test data is :", acc_sum / n)