一、pytorch finetuning 自己的圖片進行訓(xùn)練
這種讀取圖片的方式用的是torch自帶的 ImageFolder，讀取的文件夾必須在一個大的子文件下，按類別歸好類。
就像我現(xiàn)在要區(qū)分三個類別。
#perpare data set
#train data
train_data=torchvision.datasets.ImageFolder("F:/eyeDataSet/trainData",transform=transforms.Compose(
[
transforms.Scale(256),
transforms.CenterCrop(224),
transforms.ToTensor()
]))
print(len(train_data))
train_loader=DataLoader(train_data,batch_size=20,shuffle=True)
然后就是fine tuning自己的網(wǎng)絡(luò)，在torch中可以對整個網(wǎng)絡(luò)修改后，訓(xùn)練全部的參數(shù)也可以只訓(xùn)練其中的一部分，我這里就只訓(xùn)練最后一個全連接層。
torchvision中提供了很多常用的模型，比如resnet ，Vgg，Alexnet等等
# prepare model
mode1_ft_res18=torchvision.models.resnet18(pretrained=True)
for param in mode1_ft_res18.parameters():
param.requires_grad=False
num_fc=mode1_ft_res18.fc.in_features
mode1_ft_res18.fc=torch.nn.Linear(num_fc,3)
定義自己的優(yōu)化器，注意這里的參數(shù)只傳入最后一層的
#loss function and optimizer criterion=torch.nn.CrossEntropyLoss() #parameters only train the last fc layer optimizer=torch.optim.Adam(mode1_ft_res18.fc.parameters(),lr=0.001)
然后就可以開始訓(xùn)練了，定義好各種參數(shù)。
#start train
#label not one-hot encoder
EPOCH=1
for epoch in range(EPOCH):
train_loss=0.
train_acc=0.
for step,data in enumerate(train_loader):
batch_x,batch_y=data
batch_x,batch_y=Variable(batch_x),Variable(batch_y)
#batch_y not one hot
#out is the probability of eatch class
# such as one sample[-1.1009 0.1411 0.0320],need to calculate the max index
# out shape is batch_size * class
out=mode1_ft_res18(batch_x)
loss=criterion(out,batch_y)
train_loss+=loss.data[0]
# pred is the expect class
#batch_y is the true label
pred=torch.max(out,1)[1]
train_correct=(pred==batch_y).sum()
train_acc+=train_correct.data[0]
optimizer.zero_grad()
loss.backward()
optimizer.step()
if step%14==0:
print("Epoch: ",epoch,"Step",step,
"Train_loss: ",train_loss/((step+1)*20),"Train acc: ",train_acc/((step+1)*20))
測試部分和訓(xùn)練部分類似這里就不一一說明。
這樣就完整了對自己網(wǎng)絡(luò)的訓(xùn)練測試，完整代碼如下：
import torch
import numpy as np
import torchvision
from torchvision import transforms,utils
from torch.utils.data import DataLoader
from torch.autograd import Variable
#perpare data set
#train data
train_data=torchvision.datasets.ImageFolder("F:/eyeDataSet/trainData",transform=transforms.Compose(
[
transforms.Scale(256),
transforms.CenterCrop(224),
transforms.ToTensor()
]))
print(len(train_data))
train_loader=DataLoader(train_data,batch_size=20,shuffle=True)
#test data
test_data=torchvision.datasets.ImageFolder("F:/eyeDataSet/testData",transform=transforms.Compose(
[
transforms.Scale(256),
transforms.CenterCrop(224),
transforms.ToTensor()
]))
test_loader=DataLoader(test_data,batch_size=20,shuffle=True)
# prepare model
mode1_ft_res18=torchvision.models.resnet18(pretrained=True)
for param in mode1_ft_res18.parameters():
param.requires_grad=False
num_fc=mode1_ft_res18.fc.in_features
mode1_ft_res18.fc=torch.nn.Linear(num_fc,3)
#loss function and optimizer
criterion=torch.nn.CrossEntropyLoss()
#parameters only train the last fc layer
optimizer=torch.optim.Adam(mode1_ft_res18.fc.parameters(),lr=0.001)
#start train
#label not one-hot encoder
EPOCH=1
for epoch in range(EPOCH):
train_loss=0.
train_acc=0.
for step,data in enumerate(train_loader):
batch_x,batch_y=data
batch_x,batch_y=Variable(batch_x),Variable(batch_y)
#batch_y not one hot
#out is the probability of eatch class
# such as one sample[-1.1009 0.1411 0.0320],need to calculate the max index
# out shape is batch_size * class
out=mode1_ft_res18(batch_x)
loss=criterion(out,batch_y)
train_loss+=loss.data[0]
# pred is the expect class
#batch_y is the true label
pred=torch.max(out,1)[1]
train_correct=(pred==batch_y).sum()
train_acc+=train_correct.data[0]
optimizer.zero_grad()
loss.backward()
optimizer.step()
if step%14==0:
print("Epoch: ",epoch,"Step",step,
"Train_loss: ",train_loss/((step+1)*20),"Train acc: ",train_acc/((step+1)*20))
#print("Epoch: ", epoch, "Train_loss: ", train_loss / len(train_data), "Train acc: ", train_acc / len(train_data))
# test model
mode1_ft_res18.eval()
eval_loss=0
eval_acc=0
for step ,data in enumerate(test_loader):
batch_x,batch_y=data
batch_x,batch_y=Variable(batch_x),Variable(batch_y)
out=mode1_ft_res18(batch_x)
loss = criterion(out, batch_y)
eval_loss += loss.data[0]
# pred is the expect class
# batch_y is the true label
pred = torch.max(out, 1)[1]
test_correct = (pred == batch_y).sum()
eval_acc += test_correct.data[0]
optimizer.zero_grad()
loss.backward()
optimizer.step()
print( "Test_loss: ", eval_loss / len(test_data), "Test acc: ", eval_acc / len(test_data))
二、PyTorch 利用預(yù)訓(xùn)練模型進行Fine-tuning
在Deep Learning領(lǐng)域，很多子領(lǐng)域的應(yīng)用，比如一些動物識別，食物的識別等，公開的可用的數(shù)據(jù)庫相對于ImageNet等數(shù)據(jù)庫而言，其規(guī)模太小了，無法利用深度網(wǎng)絡(luò)模型直接train from scratch，容易引起過擬合，這時就需要把一些在大規(guī)模數(shù)據(jù)庫上已經(jīng)訓(xùn)練完成的模型拿過來，在目標數(shù)據(jù)庫上直接進行Fine-tuning（微調(diào)），這個已經(jīng)經(jīng)過訓(xùn)練的模型對于目標數(shù)據(jù)集而言，只是一種相對較好的參數(shù)初始化方法而已，尤其是大數(shù)據(jù)集與目標數(shù)據(jù)集結(jié)構(gòu)比較相似的話，經(jīng)過在目標數(shù)據(jù)集上微調(diào)能夠得到不錯的效果。
Fine-tune預(yù)訓(xùn)練網(wǎng)絡(luò)的步驟：
1. 首先更改預(yù)訓(xùn)練模型分類層全連接層的數(shù)目，因為一般目標數(shù)據(jù)集的類別數(shù)與大規(guī)模數(shù)據(jù)庫的類別數(shù)不一致，更改為目標數(shù)據(jù)集上訓(xùn)練集的類別數(shù)目即可，一致的話則無需更改；
2. 把分類器前的網(wǎng)絡(luò)的所有層的參數(shù)固定，即不讓它們參與學習，不進行反向傳播，只訓(xùn)練分類層的網(wǎng)絡(luò)，這時學習率可以設(shè)置的大一點，如是原來初始學習率的10倍或幾倍或0.01等，這時候網(wǎng)絡(luò)訓(xùn)練的比較快，因為除了分類層，其它層不需要進行反向傳播，可以多嘗試不同的學習率設(shè)置。
3.接下來是設(shè)置相對較小的學習率，對整個網(wǎng)絡(luò)進行訓(xùn)練，這時網(wǎng)絡(luò)訓(xùn)練變慢啦。
下面對利用PyTorch深度學習框架Fine-tune預(yù)訓(xùn)練網(wǎng)絡(luò)的過程中涉及到的固定可學習參數(shù)，對不同的層設(shè)置不同的學習率等進行詳細講解。
1. PyTorch對某些層固定網(wǎng)絡(luò)的可學習參數(shù)的方法：
class Net(nn.Module):
def __init__(self, num_classes=546):
super(Net, self).__init__()
self.features = nn.Sequential(
nn.Conv2d(1, 64, kernel_size=3, stride=2, padding=1),
nn.BatchNorm2d(64),
nn.ReLU(inplace=True),
nn.Conv2d(64, 64, kernel_size=3, stride=1, padding=1),
nn.BatchNorm2d(64),
nn.ReLU(inplace=True),
)
self.Conv1_1 = nn.Sequential(
nn.Conv2d(64, 64, kernel_size=3, stride=1, padding=1),
nn.BatchNorm2d(64),
nn.ReLU(inplace=True),
nn.Conv2d(64, 64, kernel_size=3, stride=1, padding=1),
nn.BatchNorm2d(64),
)
for p in self.parameters():
p.requires_grad=False
self.Conv1_2 = nn.Sequential(
nn.Conv2d(64, 64, kernel_size=3, stride=1, padding=1),
nn.BatchNorm2d(64),
nn.ReLU(inplace=True),
nn.Conv2d(64, 64, kernel_size=3, stride=1, padding=1),
nn.BatchNorm2d(64),
)
如上述代碼，則模型Net網(wǎng)絡(luò)中self.features與self.Conv1_1層中的參數(shù)便是固定，不可學習的。這主要看代碼：
for p in self.parameters():
p.requires_grad=False
插入的位置，這段代碼前的所有層的參數(shù)是不可學習的，也就沒有反向傳播過程。也可以指定某一層的參數(shù)不可學習，如下：
for p in self.features.parameters():
p.requires_grad=False
則 self.features層所有參數(shù)均是不可學習的。
注意，上述代碼設(shè)置若要真正生效，在訓(xùn)練網(wǎng)絡(luò)時需要在設(shè)置優(yōu)化器如下：
optimizer = torch.optim.SGD(filter(lambda p: p.requires_grad, model.parameters()), args.lr,
momentum=args.momentum,
weight_decay=args.weight_decay)
2. PyTorch之為不同的層設(shè)置不同的學習率
model = Net()
conv1_2_params = list(map(id, model.Conv1_2.parameters()))
base_params = filter(lambda p: id(p) not in conv1_2_params,
model.parameters())
optimizer = torch.optim.SGD([
{"params": base_params},
{"params": model.Conv1_2.parameters(), "lr": 10 * args.lr}], args.lr,
momentum=args.momentum, weight_decay=args.weight_decay)
上述代碼表示將模型Net網(wǎng)絡(luò)的 self.Conv1_2層的學習率設(shè)置為傳入學習率的10倍，base_params的學習沒有明確設(shè)置，則默認為傳入的學習率args.lr。
注意：
[{"params": base_params}, {"params": model.Conv1_2.parameters(), "lr": 10 * args.lr}]
表示為列表中的字典結(jié)構(gòu)。
這種方法設(shè)置不同的學習率顯得不夠靈活，可以為不同的層設(shè)置靈活的學習率，可以采用如下方法在adjust_learning_rate函數(shù)中設(shè)置：
def adjust_learning_rate(optimizer, epoch, args):
lre = []
lre.extend([0.01] * 10)
lre.extend([0.005] * 10)
lre.extend([0.0025] * 10)
lr = lre[epoch]
optimizer.param_groups[0]["lr"] = 0.9 * lr
optimizer.param_groups[1]["lr"] = 10 * lr
print(param_group[0]["lr"])
print(param_group[1]["lr"])
上述代碼中的optimizer.param_groups[0]就代表[{"params": base_params}, {"params": model.Conv1_2.parameters(), "lr": 10 * args.lr}]中的"params": base_params}，optimizer.param_groups[1]代表{"params": model.Conv1_2.parameters(), "lr": 10 * args.lr}，這里設(shè)置的學習率會把args.lr給覆蓋掉，個人認為上述代碼在設(shè)置學習率方面更靈活一些。上述代碼也可如下變成實現(xiàn)（注意學習率隨便設(shè)置的，未與上述代碼保持一致）：
def adjust_learning_rate(optimizer, epoch, args):
lre = np.logspace(-2, -4, 40)
lr = lre[epoch]
for i in range(len(optimizer.param_groups)):
param_group = optimizer.param_groups[i]
if i == 0:
param_group["lr"] = 0.9 * lr
else:
param_group["lr"] = 10 * lr
print(param_group["lr"])
下面貼出SGD優(yōu)化器的PyTorch實現(xiàn)，及其每個參數(shù)的設(shè)置和表示意義，具體如下：
import torch
from .optimizer import Optimizer, required
class SGD(Optimizer):
r"""Implements stochastic gradient descent (optionally with momentum).
Nesterov momentum is based on the formula from
`On the importance of initialization and momentum in deep learning`__.
Args:
params (iterable): iterable of parameters to optimize or dicts defining
parameter groups
lr (float): learning rate
momentum (float, optional): momentum factor (default: 0)
weight_decay (float, optional): weight decay (L2 penalty) (default: 0)
dampening (float, optional): dampening for momentum (default: 0)
nesterov (bool, optional): enables Nesterov momentum (default: False)
Example:
>>> optimizer = torch.optim.SGD(model.parameters(), lr=0.1, momentum=0.9)
>>> optimizer.zero_grad()
>>> loss_fn(model(input), target).backward()
>>> optimizer.step()
__ http://www.cs.toronto.edu/%7Ehinton/absps/momentum.pdf
.. note::
The implementation of SGD with Momentum/Nesterov subtly differs from
Sutskever et. al. and implementations in some other frameworks.
Considering the specific case of Momentum, the update can be written as
.. math::
v =
ho * v + g
p = p - lr * v
where p, g, v and :math:`
ho` denote the parameters, gradient,
velocity, and momentum respectively.
This is in contrast to Sutskever et. al. and
other frameworks which employ an update of the form
.. math::
v =
ho * v + lr * g
p = p - v
The Nesterov version is analogously modified.
"""
def __init__(self, params, lr=required, momentum=0, dampening=0,
weight_decay=0, nesterov=False):
if lr is not required and lr < 0.0:
raise ValueError("Invalid learning rate: {}".format(lr))
if momentum < 0.0:
raise ValueError("Invalid momentum value: {}".format(momentum))
if weight_decay < 0.0:
raise ValueError("Invalid weight_decay value: {}".format(weight_decay))
defaults = dict(lr=lr, momentum=momentum, dampening=dampening,
weight_decay=weight_decay, nesterov=nesterov)
if nesterov and (momentum <= 0 or dampening != 0):
raise ValueError("Nesterov momentum requires a momentum and zero dampening")
super(SGD, self).__init__(params, defaults)
def __setstate__(self, state):
super(SGD, self).__setstate__(state)
for group in self.param_groups:
group.setdefault("nesterov", False)
def step(self, closure=None):
"""Performs a single optimization step.
Arguments:
closure (callable, optional): A closure that reevaluates the model
and returns the loss.
"""
loss = None
if closure is not None:
loss = closure()
for group in self.param_groups:
weight_decay = group["weight_decay"]
momentum = group["momentum"]
dampening = group["dampening"]
nesterov = group["nesterov"]
for p in group["params"]:
if p.grad is None:
continue
d_p = p.grad.data
if weight_decay != 0:
d_p.add_(weight_decay, p.data)
if momentum != 0:
param_state = self.state[p]
if "momentum_buffer" not in param_state:
buf = param_state["momentum_buffer"] = torch.zeros_like(p.data)
buf.mul_(momentum).add_(d_p)
else:
buf = param_state["momentum_buffer"]
buf.mul_(momentum).add_(1 - dampening, d_p)
if nesterov:
d_p = d_p.add(momentum, buf)
else:
d_p = buf
p.data.add_(-group["lr"], d_p)
return loss
經(jīng)驗總結(jié)：
在Fine-tuning時最好不要隔層設(shè)置層的參數(shù)的可學習與否，這樣做一般效果餅不理想，一般準則即可，即先Fine-tuning分類層，學習率設(shè)置的大一些，然后在將整個網(wǎng)絡(luò)設(shè)置一個較小的學習率，所有層一起訓(xùn)練。
至于不先經(jīng)過Fine-tune分類層，而是將整個網(wǎng)絡(luò)所有層一起訓(xùn)練，只是分類層的學習率相對設(shè)置大一些，這樣做也可以，至于哪個效果更好，沒評估過。當用三元組損失（triplet loss）微調(diào)用softmax loss訓(xùn)練的網(wǎng)絡(luò)時，可以設(shè)置階梯型的較小學習率，整個網(wǎng)絡(luò)所有層一起訓(xùn)練，效果比較好，而不用先Fine-tune分類層前一層的輸出。
以上為個人經(jīng)驗，希望能給大家一個參考，也希望大家多多支持服務(wù)器之家。
原文鏈接：https://blog.csdn.net/xiexu911/article/details/81227126
