重磅！Pytorch寶典來(lái)了！
來(lái)源：Jack Cui 作者：科技猛獸
一、基本配置
導(dǎo)入包和版本查詢：
import torch
import torch.nn as nn
import torchvision
print(torch.__version__)
print(torch.version.cuda)#cuda版本查詢
print(torch.backends.cudnn.version())#cudnn版本查詢
print(torch.cuda.get_device_name(0))#設(shè)備名
可復(fù)現(xiàn)性
在硬件設(shè)備（CPU、GPU）不同時(shí)，完全的可復(fù)現(xiàn)性無(wú)法保證，即使隨機(jī)種子相同。但是，在同一個(gè)設(shè)備上，應(yīng)該保證可復(fù)現(xiàn)性。具體做法是，在程序開始的時(shí)候固定torch的隨機(jī)種子，同時(shí)也把numpy的隨機(jī)種子固定。
np.random.seed(0)
torch.manual_seed(0)#為CPU設(shè)置種子用于生成隨機(jī)數(shù)，以使得結(jié)果是確定的
torch.cuda.manual_seed_all(0)#為所有的GPU設(shè)置種子，以使得結(jié)果是確定的
torch.backends.cudnn.deterministic = True
torch.backends.cudnn.benchmark = False
解釋一下：torch.backends.cudnn.benchmark = true
總的來(lái)說，大部分情況下，設(shè)置這個(gè) flag 可以讓內(nèi)置的 cuDNN 的 auto-tuner 自動(dòng)尋找最適合當(dāng)前配置的高效算法，來(lái)達(dá)到優(yōu)化運(yùn)行效率的問題。
一般來(lái)講，應(yīng)該遵循以下準(zhǔn)則：
如果網(wǎng)絡(luò)的輸入數(shù)據(jù)維度或類型上變化不大，設(shè)置 torch.backends.cudnn.benchmark = true 可以增加運(yùn)行效率； 如果網(wǎng)絡(luò)的輸入數(shù)據(jù)在每次 iteration 都變化的話，會(huì)導(dǎo)致 cnDNN 每次都會(huì)去尋找一遍最優(yōu)配置，這樣反而會(huì)降低運(yùn)行效率。
torch.backends.cudnn.deterministic是啥？顧名思義，將這個(gè) flag 置為True的話，每次返回的卷積算法將是確定的，即默認(rèn)算法。如果配合上設(shè)置 Torch 的隨機(jī)種子為固定值的話，應(yīng)該可以保證每次運(yùn)行網(wǎng)絡(luò)的時(shí)候相同輸入的輸出是固定的。
顯卡設(shè)置
如果只需要一張顯卡：
# Device configuration
device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
如果需要指定多張顯卡，比如0，1號(hào)顯卡：
import os
os.environ['CUDA_VISIBLE_DEVICES'] = '0,1'
也可以在命令行運(yùn)行代碼時(shí)設(shè)置顯卡：
CUDA_VISIBLE_DEVICES=0,1 python train.py
清除顯存
torch.cuda.empty_cache()
也可以使用在命令行重置GPU的指令
nvidia-smi --gpu-reset -i [gpu_id]
二、張量(Tensor)處理
張量的數(shù)據(jù)類型
PyTorch有9種CPU張量類型和9種GPU張量類型。
張量基本信息
tensor = torch.randn(3,4,5)
print(tensor.type()) # 數(shù)據(jù)類型
print(tensor.size()) # 張量的shape，是個(gè)元組
print(tensor.dim()) # 維度的數(shù)量
命名張量
張量命名是一個(gè)非常有用的方法，這樣可以方便地使用維度的名字來(lái)做索引或其他操作，大大提高了可讀性、易用性，防止出錯(cuò)。
# 在PyTorch 1.3之前，需要使用注釋
# Tensor[N, C, H, W]
images = torch.randn(32, 3, 56, 56)
images.sum(dim=1)
images.select(dim=1, index=0)
# PyTorch 1.3之后
NCHW = [‘N’, ‘C’, ‘H’, ‘W’]
images = torch.randn(32, 3, 56, 56, names=NCHW)
images.sum('C')
images.select('C', index=0)
# 也可以這么設(shè)置
tensor = torch.rand(3,4,1,2,names=('C', 'N', 'H', 'W'))
# 使用align_to可以對(duì)維度方便地排序
tensor = tensor.align_to('N', 'C', 'H', 'W')
數(shù)據(jù)類型轉(zhuǎn)換
# 設(shè)置默認(rèn)類型，pytorch中的FloatTensor遠(yuǎn)遠(yuǎn)快于DoubleTensor
torch.set_default_tensor_type(torch.FloatTensor)
# 類型轉(zhuǎn)換
tensor = tensor.cuda()
tensor = tensor.cpu()
tensor = tensor.float()
tensor = tensor.long()
torch.Tensor與np.ndarray轉(zhuǎn)換
除了CharTensor，其他所有CPU上的張量都支持轉(zhuǎn)換為numpy格式然后再轉(zhuǎn)換回來(lái)。
ndarray = tensor.cpu().numpy()
tensor = torch.from_numpy(ndarray).float()
tensor = torch.from_numpy(ndarray.copy()).float() # If ndarray has negative stride.
Torch.tensor與PIL.Image轉(zhuǎn)換
# pytorch中的張量默認(rèn)采用[N, C, H, W]的順序，并且數(shù)據(jù)范圍在[0,1]，需要進(jìn)行轉(zhuǎn)置和規(guī)范化
# torch.Tensor -> PIL.Image
image = PIL.Image.fromarray(torch.clamp(tensor*255, min=0, max=255).byte().permute(1,2,0).cpu().numpy())
image = torchvision.transforms.functional.to_pil_image(tensor) # Equivalently way
# PIL.Image -> torch.Tensor
path = r'./figure.jpg'
tensor = torch.from_numpy(np.asarray(PIL.Image.open(path))).permute(2,0,1).float() / 255
tensor = torchvision.transforms.functional.to_tensor(PIL.Image.open(path)) # Equivalently way
對(duì)應(yīng)的2種方法的示意圖分別為：
torch.clamp把每個(gè)元素夾緊到區(qū)間[min, max]中。第1種方法用到的2個(gè)函數(shù)是：PIL.Image.fromarray() torch.from_numpy(np.asarray())
np.ndarray與PIL.Image的轉(zhuǎn)換
image = PIL.Image.fromarray(ndarray.astype(np.uint8))
ndarray = np.asarray(PIL.Image.open(path))
從只包含一個(gè)元素的張量中提取值
value = torch.rand(1).item()
張量形變
# 在將卷積層輸入全連接層的情況下通常需要對(duì)張量做形變處理，
# 相比torch.view，torch.reshape可以自動(dòng)處理輸入張量不連續(xù)的情況。
tensor = torch.rand(2,3,4)
shape = (6, 4)
tensor = torch.reshape(tensor, shape)
打亂順序
tensor = tensor[torch.randperm(tensor.size(0))] # 打亂第一個(gè)維度
水平翻轉(zhuǎn)
# pytorch不支持tensor[::-1]這樣的負(fù)步長(zhǎng)操作，水平翻轉(zhuǎn)可以通過張量索引實(shí)現(xiàn)
# 假設(shè)張量的維度為[N, D, H, W].
tensor = tensor[:,:,:,torch.arange(tensor.size(3) - 1, -1, -1).long()]
復(fù)制張量
# Operation | New/Shared memory | Still in computation graph |
tensor.clone() # | New | Yes |
tensor.detach() # | Shared | No |
tensor.detach.clone()() # | New | No |
張量拼接
'''
注意torch.cat和torch.stack的區(qū)別在于torch.cat沿著給定的維度拼接，
而torch.stack會(huì)新增一維。例如當(dāng)參數(shù)是3個(gè)10x5的張量，torch.cat的結(jié)果是30x5的張量，
而torch.stack的結(jié)果是3x10x5的張量。
'''
tensor = torch.cat(list_of_tensors, dim=0)
tensor = torch.stack(list_of_tensors, dim=0)
將整數(shù)標(biāo)簽轉(zhuǎn)為one-hot編碼
# pytorch的標(biāo)記默認(rèn)從0開始
tensor = torch.tensor([0, 2, 1, 3])
N = tensor.size(0)
num_classes = 4
one_hot = torch.zeros(N, num_classes).long()
one_hot.scatter_(dim=1, index=torch.unsqueeze(tensor, dim=1), src=torch.ones(N, num_classes).long())
得到非零元素
torch.nonzero(tensor) # index of non-zero elements
torch.nonzero(tensor==0) # index of zero elements
torch.nonzero(tensor).size(0) # number of non-zero elements
torch.nonzero(tensor == 0).size(0) # number of zero elements
判斷兩個(gè)張量相等
torch.allclose(tensor1, tensor2) # float tensor
torch.equal(tensor1, tensor2) # int tensor
張量擴(kuò)展
# Expand tensor of shape 64*512 to shape 64*512*7*7.
tensor = torch.rand(64,512)
torch.reshape(tensor, (64, 512, 1, 1)).expand(64, 512, 7, 7)
矩陣乘法
# Matrix multiplcation: (m*n) * (n*p) * -> (m*p).
result = torch.mm(tensor1, tensor2)
# Batch matrix multiplication: (b*m*n) * (b*n*p) -> (b*m*p)
result = torch.bmm(tensor1, tensor2)
# Element-wise multiplication.
result = tensor1 * tensor2
計(jì)算兩組數(shù)據(jù)之間的兩兩歐式距離
利用broadcast機(jī)制
dist = torch.sqrt(torch.sum((X1[:,None,:] - X2) ** 2, dim=2))
三、模型定義和操作
一個(gè)簡(jiǎn)單兩層卷積網(wǎng)絡(luò)的示例
# convolutional neural network (2 convolutional layers)
class ConvNet(nn.Module):
def __init__(self, num_classes=10):
super(ConvNet, self).__init__()
self.layer1 = nn.Sequential(
nn.Conv2d(1, 16, kernel_size=5, stride=1, padding=2),
nn.BatchNorm2d(16),
nn.ReLU(),
nn.MaxPool2d(kernel_size=2, stride=2))
self.layer2 = nn.Sequential(
nn.Conv2d(16, 32, kernel_size=5, stride=1, padding=2),
nn.BatchNorm2d(32),
nn.ReLU(),
nn.MaxPool2d(kernel_size=2, stride=2))
self.fc = nn.Linear(7*7*32, num_classes)
def forward(self, x):
out = self.layer1(x)
out = self.layer2(out)
out = out.reshape(out.size(0), -1)
out = self.fc(out)
return out
model = ConvNet(num_classes).to(device)
雙線性匯合（bilinear pooling）
X = torch.reshape(N, D, H * W) # Assume X has shape N*D*H*W
X = torch.bmm(X, torch.transpose(X, 1, 2)) / (H * W) # Bilinear pooling
assert X.size() == (N, D, D)
X = torch.reshape(X, (N, D * D))
X = torch.sign(X) * torch.sqrt(torch.abs(X) + 1e-5) # Signed-sqrt normalization
X = torch.nn.functional.normalize(X) # L2 normalization
多卡同步 BN（Batch normalization）
當(dāng)使用 torch.nn.DataParallel 將代碼運(yùn)行在多張 GPU 卡上時(shí)，PyTorch 的 BN 層默認(rèn)操作是各卡上數(shù)據(jù)獨(dú)立地計(jì)算均值和標(biāo)準(zhǔn)差，同步 BN 使用所有卡上的數(shù)據(jù)一起計(jì)算 BN 層的均值和標(biāo)準(zhǔn)差，緩解了當(dāng)批量大?。╞atch size）比較小時(shí)對(duì)均值和標(biāo)準(zhǔn)差估計(jì)不準(zhǔn)的情況，是在目標(biāo)檢測(cè)等任務(wù)中一個(gè)有效的提升性能的技巧。
sync_bn = torch.nn.SyncBatchNorm(num_features, eps=1e-05, momentum=0.1, affine=True,
track_running_stats=True)
將已有網(wǎng)絡(luò)的所有BN層改為同步BN層
def convertBNtoSyncBN(module, process_group=None):
'''Recursively replace all BN layers to SyncBN layer.
Args:
module[torch.nn.Module]. Network
'''
if isinstance(module, torch.nn.modules.batchnorm._BatchNorm):
sync_bn = torch.nn.SyncBatchNorm(module.num_features, module.eps, module.momentum,
module.affine, module.track_running_stats, process_group)
sync_bn.running_mean = module.running_mean
sync_bn.running_var = module.running_var
if module.affine:
sync_bn.weight = module.weight.clone().detach()
sync_bn.bias = module.bias.clone().detach()
return sync_bn
else:
for name, child_module in module.named_children():
setattr(module, name) = convert_syncbn_model(child_module, process_group=process_group))
return module
affine定義了BN層的參數(shù)γ和β是否是可學(xué)習(xí)的(不可學(xué)習(xí)默認(rèn)是常數(shù)1和0)。
類似 BN 滑動(dòng)平均
如果要實(shí)現(xiàn)類似 BN 滑動(dòng)平均的操作，在 forward 函數(shù)中要使用原地（inplace）操作給滑動(dòng)平均賦值。
class BN(torch.nn.Module)
def __init__(self):
...
self.register_buffer('running_mean', torch.zeros(num_features))
def forward(self, X):
...
self.running_mean += momentum * (current - self.running_mean)
計(jì)算模型整體參數(shù)量
num_parameters = sum(torch.numel(parameter) for parameter in model.parameters())
查看網(wǎng)絡(luò)中的參數(shù)
可以通過model.state_dict()或者model.named_parameters()函數(shù)查看現(xiàn)在的全部可訓(xùn)練參數(shù)（包括通過繼承得到的父類中的參數(shù)）
params = list(model.named_parameters())
(name, param) = params[28]
print(name)
print(param.grad)
print('-------------------------------------------------')
(name2, param2) = params[29]
print(name2)
print(param2.grad)
print('----------------------------------------------------')
(name1, param1) = params[30]
print(name1)
print(param1.grad)
模型權(quán)重初始化
注意 model.modules() 和 model.children() 的區(qū)別：model.modules() 會(huì)迭代地遍歷模型的所有子層，而 model.children() 只會(huì)遍歷模型下的一層。
# Common practise for initialization.
for layer in model.modules():
if isinstance(layer, torch.nn.Conv2d):
torch.nn.init.kaiming_normal_(layer.weight, mode='fan_out',
nonlinearity='relu')
if layer.bias is not None:
torch.nn.init.constant_(layer.bias, val=0.0)
elif isinstance(layer, torch.nn.BatchNorm2d):
torch.nn.init.constant_(layer.weight, val=1.0)
torch.nn.init.constant_(layer.bias, val=0.0)
elif isinstance(layer, torch.nn.Linear):
torch.nn.init.xavier_normal_(layer.weight)
if layer.bias is not None:
torch.nn.init.constant_(layer.bias, val=0.0)
# Initialization with given tensor.
layer.weight = torch.nn.Parameter(tensor)
提取模型中的某一層
modules()會(huì)返回模型中所有模塊的迭代器，它能夠訪問到最內(nèi)層，比如self.layer1.conv1這個(gè)模塊，還有一個(gè)與它們相對(duì)應(yīng)的是name_children()屬性以及named_modules(),這兩個(gè)不僅會(huì)返回模塊的迭代器，還會(huì)返回網(wǎng)絡(luò)層的名字。
# 取模型中的前兩層
new_model = nn.Sequential(*list(model.children())[:2]
# 如果希望提取出模型中的所有卷積層，可以像下面這樣操作：
for layer in model.named_modules():
if isinstance(layer[1],nn.Conv2d):
conv_model.add_module(layer[0],layer[1])
部分層使用預(yù)訓(xùn)練模型
注意如果保存的模型是 torch.nn.DataParallel，則當(dāng)前的模型也需要是
model.load_state_dict(torch.load('model.pth'), strict=False)
將在 GPU 保存的模型加載到 CPU
model.load_state_dict(torch.load('model.pth', map_location='cpu'))
導(dǎo)入另一個(gè)模型的相同部分到新的模型
模型導(dǎo)入?yún)?shù)時(shí)，如果兩個(gè)模型結(jié)構(gòu)不一致，則直接導(dǎo)入?yún)?shù)會(huì)報(bào)錯(cuò)。用下面方法可以把另一個(gè)模型的相同的部分導(dǎo)入到新的模型中。
# model_new代表新的模型
# model_saved代表其他模型，比如用torch.load導(dǎo)入的已保存的模型
model_new_dict = model_new.state_dict()
model_common_dict = {k:v for k, v in model_saved.items() if k in model_new_dict.keys()}
model_new_dict.update(model_common_dict)
model_new.load_state_dict(model_new_dict)
四、數(shù)據(jù)處理
計(jì)算數(shù)據(jù)集的均值和標(biāo)準(zhǔn)差
import os
import cv2
import numpy as np
from torch.utils.data import Dataset
from PIL import Image
def compute_mean_and_std(dataset):
# 輸入PyTorch的dataset，輸出均值和標(biāo)準(zhǔn)差
mean_r = 0
mean_g = 0
mean_b = 0
for img, _ in dataset:
img = np.asarray(img) # change PIL Image to numpy array
mean_b += np.mean(img[:, :, 0])
mean_g += np.mean(img[:, :, 1])
mean_r += np.mean(img[:, :, 2])
mean_b /= len(dataset)
mean_g /= len(dataset)
mean_r /= len(dataset)
diff_r = 0
diff_g = 0
diff_b = 0
N = 0
for img, _ in dataset:
img = np.asarray(img)
diff_b += np.sum(np.power(img[:, :, 0] - mean_b, 2))
diff_g += np.sum(np.power(img[:, :, 1] - mean_g, 2))
diff_r += np.sum(np.power(img[:, :, 2] - mean_r, 2))
N += np.prod(img[:, :, 0].shape)
std_b = np.sqrt(diff_b / N)
std_g = np.sqrt(diff_g / N)
std_r = np.sqrt(diff_r / N)
mean = (mean_b.item() / 255.0, mean_g.item() / 255.0, mean_r.item() / 255.0)
std = (std_b.item() / 255.0, std_g.item() / 255.0, std_r.item() / 255.0)
return mean, std
常用訓(xùn)練和驗(yàn)證數(shù)據(jù)預(yù)處理
其中 ToTensor 操作會(huì)將 PIL.Image 或形狀為 H×W×D，數(shù)值范圍為 [0, 255] 的 np.ndarray 轉(zhuǎn)換為形狀為 D×H×W，數(shù)值范圍為 [0.0, 1.0] 的 torch.Tensor。
train_transform = torchvision.transforms.Compose([
torchvision.transforms.RandomResizedCrop(size=224,
scale=(0.08, 1.0)),
torchvision.transforms.RandomHorizontalFlip(),
torchvision.transforms.ToTensor(),
torchvision.transforms.Normalize(mean=(0.485, 0.456, 0.406),
std=(0.229, 0.224, 0.225)),
])
val_transform = torchvision.transforms.Compose([
torchvision.transforms.Resize(256),
torchvision.transforms.CenterCrop(224),
torchvision.transforms.ToTensor(),
torchvision.transforms.Normalize(mean=(0.485, 0.456, 0.406),
std=(0.229, 0.224, 0.225)),
])
五、模型訓(xùn)練和測(cè)試
分類模型訓(xùn)練代碼
# Loss and optimizer
criterion = nn.CrossEntropyLoss()
optimizer = torch.optim.Adam(model.parameters(), lr=learning_rate)
# Train the model
total_step = len(train_loader)
for epoch in range(num_epochs):
for i ,(images, labels) in enumerate(train_loader):
images = images.to(device)
labels = labels.to(device)
# Forward pass
outputs = model(images)
loss = criterion(outputs, labels)
# Backward and optimizer
optimizer.zero_grad()
loss.backward()
optimizer.step()
if (i+1) % 100 == 0:
print('Epoch: [{}/{}], Step: [{}/{}], Loss: {}'
.format(epoch+1, num_epochs, i+1, total_step, loss.item()))
分類模型測(cè)試代碼
# Test the model
model.eval() # eval mode(batch norm uses moving mean/variance
#instead of mini-batch mean/variance)
with torch.no_grad():
correct = 0
total = 0
for images, labels in test_loader:
images = images.to(device)
labels = labels.to(device)
outputs = model(images)
_, predicted = torch.max(outputs.data, 1)
total += labels.size(0)
correct += (predicted == labels).sum().item()
print('Test accuracy of the model on the 10000 test images: {} %'
.format(100 * correct / total))
自定義loss
繼承torch.nn.Module類寫自己的loss。
class MyLoss(torch.nn.Module):
def __init__(self):
super(MyLoss, self).__init__()
def forward(self, x, y):
loss = torch.mean((x - y) ** 2)
return loss
標(biāo)簽平滑（label smoothing）
寫一個(gè)label_smoothing.py的文件，然后在訓(xùn)練代碼里引用，用LSR代替交叉熵?fù)p失即可。label_smoothing.py內(nèi)容如下：
import torch
import torch.nn as nn
class LSR(nn.Module):
def __init__(self, e=0.1, reduction='mean'):
super().__init__()
self.log_softmax = nn.LogSoftmax(dim=1)
self.e = e
self.reduction = reduction
def _one_hot(self, labels, classes, value=1):
"""
Convert labels to one hot vectors
Args:
labels: torch tensor in format [label1, label2, label3, ...]
classes: int, number of classes
value: label value in one hot vector, default to 1
Returns:
return one hot format labels in shape [batchsize, classes]
"""
one_hot = torch.zeros(labels.size(0), classes)
#labels and value_added size must match
labels = labels.view(labels.size(0), -1)
value_added = torch.Tensor(labels.size(0), 1).fill_(value)
value_added = value_added.to(labels.device)
one_hot = one_hot.to(labels.device)
one_hot.scatter_add_(1, labels, value_added)
return one_hot
def _smooth_label(self, target, length, smooth_factor):
"""convert targets to one-hot format, and smooth
them.
Args:
target: target in form with [label1, label2, label_batchsize]
length: length of one-hot format(number of classes)
smooth_factor: smooth factor for label smooth
Returns:
smoothed labels in one hot format
"""
one_hot = self._one_hot(target, length, value=1 - smooth_factor)
one_hot += smooth_factor / (length - 1)
return one_hot.to(target.device)
def forward(self, x, target):
if x.size(0) != target.size(0):
raise ValueError('Expected input batchsize ({}) to match target batch_size({})'
.format(x.size(0), target.size(0)))
if x.dim() < 2:
raise ValueError('Expected input tensor to have least 2 dimensions(got {})'
.format(x.size(0)))
if x.dim() != 2:
raise ValueError('Only 2 dimension tensor are implemented, (got {})'
.format(x.size()))
smoothed_target = self._smooth_label(target, x.size(1), self.e)
x = self.log_softmax(x)
loss = torch.sum(- x * smoothed_target, dim=1)
if self.reduction == 'none':
return loss
elif self.reduction == 'sum':
return torch.sum(loss)
elif self.reduction == 'mean':
return torch.mean(loss)
else:
raise ValueError('unrecognized option, expect reduction to be one of none, mean, sum')
或者直接在訓(xùn)練文件里做label smoothing
for images, labels in train_loader:
images, labels = images.cuda(), labels.cuda()
N = labels.size(0)
# C is the number of classes.
smoothed_labels = torch.full(size=(N, C), fill_value=0.1 / (C - 1)).cuda()
smoothed_labels.scatter_(dim=1, index=torch.unsqueeze(labels, dim=1), value=0.9)
score = model(images)
log_prob = torch.nn.functional.log_softmax(score, dim=1)
loss = -torch.sum(log_prob * smoothed_labels) / N
optimizer.zero_grad()
loss.backward()
optimizer.step()
Mixup訓(xùn)練
beta_distribution = torch.distributions.beta.Beta(alpha, alpha)
for images, labels in train_loader:
images, labels = images.cuda(), labels.cuda()
# Mixup images and labels.
lambda_ = beta_distribution.sample([]).item()
index = torch.randperm(images.size(0)).cuda()
mixed_images = lambda_ * images + (1 - lambda_) * images[index, :]
label_a, label_b = labels, labels[index]
# Mixup loss.
scores = model(mixed_images)
loss = (lambda_ * loss_function(scores, label_a)
+ (1 - lambda_) * loss_function(scores, label_b))
optimizer.zero_grad()
loss.backward()
optimizer.step()
L1 正則化
l1_regularization = torch.nn.L1Loss(reduction='sum')
loss = ... # Standard cross-entropy loss
for param in model.parameters():
loss += torch.sum(torch.abs(param))
loss.backward()
不對(duì)偏置項(xiàng)進(jìn)行權(quán)重衰減（weight decay）
pytorch里的weight decay相當(dāng)于l2正則
bias_list = (param for name, param in model.named_parameters() if name[-4:] == 'bias')
others_list = (param for name, param in model.named_parameters() if name[-4:] != 'bias')
parameters = [{'parameters': bias_list, 'weight_decay': 0},
{'parameters': others_list}]
optimizer = torch.optim.SGD(parameters, lr=1e-2, momentum=0.9, weight_decay=1e-4)
得到當(dāng)前學(xué)習(xí)率
# If there is one global learning rate (which is the common case).
lr = next(iter(optimizer.param_groups))['lr']
# If there are multiple learning rates for different layers.
all_lr = []
for param_group in optimizer.param_groups:
all_lr.append(param_group['lr'])
另一種方法，在一個(gè)batch訓(xùn)練代碼里，當(dāng)前的lr是optimizer.param_groups[0]['lr']
學(xué)習(xí)率衰減
# Reduce learning rate when validation accuarcy plateau.
scheduler = torch.optim.lr_scheduler.ReduceLROnPlateau(optimizer, mode='max', patience=5, verbose=True)
for t in range(0, 80):
train(...)
val(...)
scheduler.step(val_acc)
# Cosine annealing learning rate.
scheduler = torch.optim.lr_scheduler.CosineAnnealingLR(optimizer, T_max=80)
# Reduce learning rate by 10 at given epochs.
scheduler = torch.optim.lr_scheduler.MultiStepLR(optimizer, milestones=[50, 70], gamma=0.1)
for t in range(0, 80):
scheduler.step()
train(...)
val(...)
# Learning rate warmup by 10 epochs.
scheduler = torch.optim.lr_scheduler.LambdaLR(optimizer, lr_lambda=lambda t: t / 10)
for t in range(0, 10):
scheduler.step()
train(...)
val(...)
優(yōu)化器鏈?zhǔn)礁?/span>
從1.4版本開始，torch.optim.lr_scheduler 支持鏈?zhǔn)礁拢╟haining），即用戶可以定義兩個(gè) schedulers，并交替在訓(xùn)練中使用。
import torch
from torch.optim import SGD
from torch.optim.lr_scheduler import ExponentialLR, StepLR
model = [torch.nn.Parameter(torch.randn(2, 2, requires_grad=True))]
optimizer = SGD(model, 0.1)
scheduler1 = ExponentialLR(optimizer, gamma=0.9)
scheduler2 = StepLR(optimizer, step_size=3, gamma=0.1)
for epoch in range(4):
print(epoch, scheduler2.get_lr()[0])
optimizer.step()
scheduler1.step()
scheduler2.step()
模型訓(xùn)練可視化
PyTorch可以使用tensorboard來(lái)可視化訓(xùn)練過程。
安裝和運(yùn)行TensorBoard。
pip install tensorboard
tensorboard --logdir=runs
使用SummaryWriter類來(lái)收集和可視化相應(yīng)的數(shù)據(jù)，放了方便查看，可以使用不同的文件夾，比如'Loss/train'和'Loss/test'。
from torch.utils.tensorboard import SummaryWriter
import numpy as np
writer = SummaryWriter()
for n_iter in range(100):
writer.add_scalar('Loss/train', np.random.random(), n_iter)
writer.add_scalar('Loss/test', np.random.random(), n_iter)
writer.add_scalar('Accuracy/train', np.random.random(), n_iter)
writer.add_scalar('Accuracy/test', np.random.random(), n_iter)
保存與加載斷點(diǎn)
注意為了能夠恢復(fù)訓(xùn)練，我們需要同時(shí)保存模型和優(yōu)化器的狀態(tài)，以及當(dāng)前的訓(xùn)練輪數(shù)。
start_epoch = 0
# Load checkpoint.
if resume: # resume為參數(shù)，第一次訓(xùn)練時(shí)設(shè)為0，中斷再訓(xùn)練時(shí)設(shè)為1
model_path = os.path.join('model', 'best_checkpoint.pth.tar')
assert os.path.isfile(model_path)
checkpoint = torch.load(model_path)
best_acc = checkpoint['best_acc']
start_epoch = checkpoint['epoch']
model.load_state_dict(checkpoint['model'])
optimizer.load_state_dict(checkpoint['optimizer'])
print('Load checkpoint at epoch {}.'.format(start_epoch))
print('Best accuracy so far {}.'.format(best_acc))
# Train the model
for epoch in range(start_epoch, num_epochs):
...
# Test the model
...
# save checkpoint
is_best = current_acc > best_acc
best_acc = max(current_acc, best_acc)
checkpoint = {
'best_acc': best_acc,
'epoch': epoch + 1,
'model': model.state_dict(),
'optimizer': optimizer.state_dict(),
}
model_path = os.path.join('model', 'checkpoint.pth.tar')
best_model_path = os.path.join('model', 'best_checkpoint.pth.tar')
torch.save(checkpoint, model_path)
if is_best:
shutil.copy(model_path, best_model_path)
提取 ImageNet 預(yù)訓(xùn)練模型某層的卷積特征
# VGG-16 relu5-3 feature.
model = torchvision.models.vgg16(pretrained=True).features[:-1]
# VGG-16 pool5 feature.
model = torchvision.models.vgg16(pretrained=True).features
# VGG-16 fc7 feature.
model = torchvision.models.vgg16(pretrained=True)
model.classifier = torch.nn.Sequential(*list(model.classifier.children())[:-3])
# ResNet GAP feature.
model = torchvision.models.resnet18(pretrained=True)
model = torch.nn.Sequential(collections.OrderedDict(
list(model.named_children())[:-1]))
with torch.no_grad():
model.eval()
conv_representation = model(image)
提取 ImageNet 預(yù)訓(xùn)練模型多層的卷積特征
class FeatureExtractor(torch.nn.Module):
"""Helper class to extract several convolution features from the given
pre-trained model.
Attributes:
_model, torch.nn.Module.
_layers_to_extract, list or set
Example:
>>> model = torchvision.models.resnet152(pretrained=True)
>>> model = torch.nn.Sequential(collections.OrderedDict(
list(model.named_children())[:-1]))
>>> conv_representation = FeatureExtractor(
pretrained_model=model,
layers_to_extract={'layer1', 'layer2', 'layer3', 'layer4'})(image)
"""
def __init__(self, pretrained_model, layers_to_extract):
torch.nn.Module.__init__(self)
self._model = pretrained_model
self._model.eval()
self._layers_to_extract = set(layers_to_extract)
def forward(self, x):
with torch.no_grad():
conv_representation = []
for name, layer in self._model.named_children():
x = layer(x)
if name in self._layers_to_extract:
conv_representation.append(x)
return conv_representation
微調(diào)全連接層
model = torchvision.models.resnet18(pretrained=True)
for param in model.parameters():
param.requires_grad = False
model.fc = nn.Linear(512, 100) # Replace the last fc layer
optimizer = torch.optim.SGD(model.fc.parameters(), lr=1e-2, momentum=0.9, weight_decay=1e-4)
以較大學(xué)習(xí)率微調(diào)全連接層，較小學(xué)習(xí)率微調(diào)卷積層
model = torchvision.models.resnet18(pretrained=True)
finetuned_parameters = list(map(id, model.fc.parameters()))
conv_parameters = (p for p in model.parameters() if id(p) not in finetuned_parameters)
parameters = [{'params': conv_parameters, 'lr': 1e-3},
{'params': model.fc.parameters()}]
optimizer = torch.optim.SGD(parameters, lr=1e-2, momentum=0.9, weight_decay=1e-4)
六、其他注意事項(xiàng)
不要使用太大的線性層。因?yàn)閚n.Linear(m,n)使用的是O(m,n)的內(nèi)存，線性層太大很容易超出現(xiàn)有顯存。 不要在太長(zhǎng)的序列上使用RNN。因?yàn)镽NN反向傳播使用的是BPTT算法，其需要的內(nèi)存和輸入序列的長(zhǎng)度呈線性關(guān)系。 model(x) 前用 model.train() 和 model.eval() 切換網(wǎng)絡(luò)狀態(tài)。 不需要計(jì)算梯度的代碼塊用 with torch.no_grad() 包含起來(lái)。 model.eval() 和 torch.no_grad() 的區(qū)別在于，model.eval() 是將網(wǎng)絡(luò)切換為測(cè)試狀態(tài)，例如 BN 和dropout在訓(xùn)練和測(cè)試階段使用不同的計(jì)算方法。torch.no_grad() 是關(guān)閉 PyTorch 張量的自動(dòng)求導(dǎo)機(jī)制，以減少存儲(chǔ)使用和加速計(jì)算，得到的結(jié)果無(wú)法進(jìn)行 loss.backward()。 model.zero_grad()會(huì)把整個(gè)模型的參數(shù)的梯度都?xì)w零, 而optimizer.zero_grad()只會(huì)把傳入其中的參數(shù)的梯度歸零。 torch.nn.CrossEntropyLoss 的輸入不需要經(jīng)過 Softmax。torch.nn.CrossEntropyLoss 等價(jià)于 torch.nn.functional.log_softmax + torch.nn.NLLLoss。 loss.backward() 前用 optimizer.zero_grad() 清除累積梯度。 torch.utils.data.DataLoader 中盡量設(shè)置 pin_memory=True，對(duì)特別小的數(shù)據(jù)集如 MNIST 設(shè)置 pin_memory=False 反而更快一些。num_workers 的設(shè)置需要在實(shí)驗(yàn)中找到最快的取值。 用 del 及時(shí)刪除不用的中間變量，節(jié)約 GPU 存儲(chǔ)。 使用 inplace 操作可節(jié)約 GPU 存儲(chǔ)，如
x = torch.nn.functional.relu(x, inplace=True)
減少 CPU 和 GPU 之間的數(shù)據(jù)傳輸。 例如如果你想知道一個(gè) epoch 中每個(gè) mini-batch 的 loss 和準(zhǔn)確率，先將它們累積在 GPU 中等一個(gè) epoch 結(jié)束之后一起傳輸回 CPU 會(huì)比每個(gè) mini-batch 都進(jìn)行一次 GPU 到 CPU 的傳輸更快。 使用半精度浮點(diǎn)數(shù) half() 會(huì)有一定的速度提升，具體效率依賴于 GPU 型號(hào)。需要小心數(shù)值精度過低帶來(lái)的穩(wěn)定性問題。 時(shí)常使用 assert tensor.size() == (N, D, H, W) 作為調(diào)試手段，確保張量維度和你設(shè)想中一致。 除了標(biāo)記 y 外，盡量少使用一維張量，使用 n*1 的二維張量代替，可以避免一些意想不到的一維張量計(jì)算結(jié)果。 統(tǒng)計(jì)代碼各部分耗時(shí)
with torch.autograd.profiler.profile(enabled=True, use_cuda=False) as profile:
...
print(profile)
# 或者在命令行運(yùn)行
python -m torch.utils.bottleneck main.py
使用TorchSnooper來(lái)調(diào)試PyTorch代碼，程序在執(zhí)行的時(shí)候，就會(huì)自動(dòng) print 出來(lái)每一行的執(zhí)行結(jié)果的 tensor 的形狀、數(shù)據(jù)類型、設(shè)備、是否需要梯度的信息。
# pip install torchsnooper
import torchsnooper
# 對(duì)于函數(shù)，使用修飾器
@torchsnooper.snoop()
# 如果不是函數(shù)，使用 with 語(yǔ)句來(lái)激活 TorchSnooper，把訓(xùn)練的那個(gè)循環(huán)裝進(jìn) with 語(yǔ)句中去。
with torchsnooper.snoop():
原本的代碼
評(píng)論
圖片
表情