【深度學(xué)習(xí)】PyTorch 常用 Tricks 總結(jié)

from torchsummary import summarysummary(your_model, input_size=(channels, H, W))

import torch.nn as nn
outputs = model(data)loss= loss_fn(outputs, target)optimizer.zero_grad()loss.backward()nn.utils.clip_grad_norm_(model.parameters(), max_norm=20, norm_type=2)optimizer.step()

import cv2import torch
image = cv2.imread(img_path)image = torch.tensor(image)print(image.size())
img = image.view(1, *image.size())print(img.size())
# output:# torch.Size([h, w, c])# torch.Size([1, h, w, c])

import cv2import numpy as np
image = cv2.imread(img_path)print(image.shape)img = image[np.newaxis, :, :, :]print(img.shape)
# output:# (h, w, c)# (1, h, w, c)

import cv2import torch
image = cv2.imread(img_path)image = torch.tensor(image)print(image.size())
img = image.unsqueeze(dim=0)  print(img.size())
img = img.squeeze(dim=0)print(img.size())
# output:# torch.Size([(h, w, c)])# torch.Size([1, h, w, c])# torch.Size([h, w, c])

import torchclass_num = 8batch_size = 4
def one_hot(label):    """    將一維列表轉(zhuǎn)換為獨(dú)熱編碼    """    label = label.resize_(batch_size, 1)    m_zeros = torch.zeros(batch_size, class_num)    # 從 value 中取值，然后根據(jù) dim 和 index 給相應(yīng)位置賦值    onehot = m_zeros.scatter_(1, label, 1)  # (dim,index,value)
    return onehot.numpy()  # Tensor -> Numpy
label = torch.LongTensor(batch_size).random_() % class_num  # 對(duì)隨機(jī)數(shù)取余print(one_hot(label))
# output:[[0. 0. 0. 1. 0. 0. 0. 0.] [0. 0. 0. 0. 1. 0. 0. 0.] [0. 0. 1. 0. 0. 0. 0. 0.] [0. 1. 0. 0. 0. 0. 0. 0.]]

with torch.no_grad():    # 使用model進(jìn)行預(yù)測的代碼????pass

import torch.optim as optimfrom torch.optim import lr_scheduler
# 訓(xùn)練前的初始化optimizer = optim.Adam(net.parameters(), lr=0.001)scheduler = lr_scheduler.StepLR(optimizer, 10, 0.1)  # # 每過10個(gè)epoch，學(xué)習(xí)率乘以0.1
# 訓(xùn)練過程中for n in n_epoch:    scheduler.step()    ...

net = Network()  # 獲取自定義網(wǎng)絡(luò)結(jié)構(gòu)for name, value in net.named_parameters():    print('name: {0},\t grad: {1}'.format(name, value.requires_grad))

name: cnn.VGG_16.convolution1_1.weight,   grad: Truename: cnn.VGG_16.convolution1_1.bias,   grad: Truename: cnn.VGG_16.convolution1_2.weight,   grad: Truename: cnn.VGG_16.convolution1_2.bias,   grad: Truename: cnn.VGG_16.convolution2_1.weight,   grad: Truename: cnn.VGG_16.convolution2_1.bias,   grad: Truename: cnn.VGG_16.convolution2_2.weight,   grad: Truename: cnn.VGG_16.convolution2_2.bias,   grad: True

no_grad = [    'cnn.VGG_16.convolution1_1.weight',    'cnn.VGG_16.convolution1_1.bias',    'cnn.VGG_16.convolution1_2.weight',    'cnn.VGG_16.convolution1_2.bias']

net = Net.CTPN()  # 獲取網(wǎng)絡(luò)結(jié)構(gòu)for name, value in net.named_parameters():    if name in no_grad:        value.requires_grad = False    else:        value.requires_grad = True

name: cnn.VGG_16.convolution1_1.weight,   grad: Falsename: cnn.VGG_16.convolution1_1.bias,   grad: Falsename: cnn.VGG_16.convolution1_2.weight,   grad: Falsename: cnn.VGG_16.convolution1_2.bias,   grad: Falsename: cnn.VGG_16.convolution2_1.weight,   grad: Truename: cnn.VGG_16.convolution2_1.bias,   grad: Truename: cnn.VGG_16.convolution2_2.weight,   grad: Truename: cnn.VGG_16.convolution2_2.bias,   grad: True

optimizer = optim.Adam(filter(lambda p: p.requires_grad, net.parameters()), lr=0.01)

net = Network()  # 獲取自定義網(wǎng)絡(luò)結(jié)構(gòu)for name, value in net.named_parameters():    print('name: {}'.format(name))
# 輸出：# name: cnn.VGG_16.convolution1_1.weight# name: cnn.VGG_16.convolution1_1.bias# name: cnn.VGG_16.convolution1_2.weight# name: cnn.VGG_16.convolution1_2.bias# name: cnn.VGG_16.convolution2_1.weight# name: cnn.VGG_16.convolution2_1.bias# name: cnn.VGG_16.convolution2_2.weight# name: cnn.VGG_16.convolution2_2.bias

conv1_params = []conv2_params = []
for name, parms in net.named_parameters():    if "convolution1" in name:        conv1_params += [parms]    else:        conv2_params += [parms]
# 然后在優(yōu)化器中進(jìn)行如下操作：optimizer = optim.Adam(    [        {"params": conv1_params, 'lr': 0.01},        {"params": conv2_params, 'lr': 0.001},    ],    weight_decay=1e-3,)