[Pytorch & Tensorflow2.x 기초] CNN 모델로 Pytorch와 Tensorflow2.x의 기본적인 모델 구현 확인하기
2020. 9. 1. 21:44ㆍDL in Python
TensorFlow 2.0¶
In [1]:
import tensorflow as tf
from tensorflow.keras import layers
from tensorflow.keras import datasets
Hyperparameter Tunning¶
In [2]:
num_epochs = 1
batch_size = 64
learning_rate = 0.001
dropout_rate = 0.7
input_shape = (28, 28, 1)
num_classes = 10
Preprocess¶
In [3]:
# data load
(train_x, train_y), (test_x, test_y) = datasets.mnist.load_data()
In [5]:
# 차원 늘리기
train_x = train_x[..., tf.newaxis]
test_x = test_x[..., tf.newaxis]
# normalization
train_x = train_x / 255.
test_x = test_x / 255.
Build Model¶
In [6]:
inputs = layers.Input(input_shape)
net = layers.Conv2D(32, (3, 3), padding='SAME')(inputs)
net = layers.Activation('relu')(net)
net = layers.Conv2D(32, (3, 3), padding='SAME')(net)
net = layers.Activation('relu')(net)
net = layers.MaxPooling2D(pool_size=(2, 2))(net)
net = layers.Dropout(dropout_rate)(net)
net = layers.Conv2D(64, (3, 3), padding='SAME')(net)
net = layers.Activation('relu')(net)
net = layers.Conv2D(64, (3, 3), padding='SAME')(net)
net = layers.Activation('relu')(net)
net = layers.MaxPooling2D(pool_size=(2, 2))(net)
net = layers.Dropout(dropout_rate)(net)
net = layers.Flatten()(net)
net = layers.Dense(512)(net)
net = layers.Activation('relu')(net)
net = layers.Dropout(dropout_rate)(net)
net = layers.Dense(num_classes)(net)
net = layers.Activation('softmax')(net)
model = tf.keras.Model(inputs=inputs, outputs=net, name='Basic_CNN')
In [7]:
# Model is the full model w/o custom layers
model.compile(optimizer=tf.keras.optimizers.Adam(learning_rate), # Optimization
loss='sparse_categorical_crossentropy', # Loss Function
metrics=['accuracy']) # Metrics / Accuracy
Training¶
In [8]:
model.fit(train_x, train_y,
batch_size=batch_size,
shuffle=True)
model.evaluate(test_x, test_y, batch_size=batch_size)
Out[8]:
PyTorch¶
In [10]:
import torch
# 모델 설계를 위한 라이브러리
import torch.nn as nn
import torch.nn.functional as F
# optimization을 위한 라이브러리
import torch.optim as optim
# data load와 rescale을 위한 라이브러리
from torchvision import datasets, transforms
Hyperparameter Tunning¶
In [15]:
seed = 1
lr = 0.001
momentum = 0.5
batch_size = 64
test_batch_size = 64
epochs = 1
# gpu 사용?
no_cuda = False
# pytorch는 progress bar가 돌아가지 않아서 직접 출력을 통해 확인하기 위한 학습 로그 interval을 생성
log_interval = 100
Model¶
In [12]:
class Net(nn.Module):
def __init__(self):
super(Net, self).__init__()
self.conv1 = nn.Conv2d(1, 20, 5, 1) # tensorflow는 in-channel없이 out-channel만 사용, 반면 pytorch는 둘다 필요
self.conv2 = nn.Conv2d(20, 50, 5, 1)
self.fc1 = nn.Linear(4*4*50, 500) # tensorflow의 dense 역할
self.fc2 = nn.Linear(500, 10)
def forward(self, x):
x = F.relu(self.conv1(x))
x = F.max_pool2d(x, 2, 2)
x = F.relu(self.conv2(x))
x = F.max_pool2d(x, 2, 2)
x = x.view(-1, 4*4*50)
x = F.relu(self.fc1(x))
x = self.fc2(x)
return F.log_softmax(x, dim=1)
Preprocess¶
In [13]:
torch.manual_seed(seed)
# gpu 있으면 사용
use_cuda = not no_cuda and torch.cuda.is_available()
device = torch.device("cuda" if use_cuda else "cpu")
kwargs = {'num_workers': 1, 'pin_memory': True} if use_cuda else {}
# data loader
train_loader = torch.utils.data.DataLoader(
datasets.MNIST('../data', train=True, download=True,
transform=transforms.Compose([
transforms.ToTensor(),
transforms.Normalize((0.1307,), (0.3081,))
])),
batch_size=batch_size, shuffle=True, **kwargs)
test_loader = torch.utils.data.DataLoader(
datasets.MNIST('../data', train=False, transform=transforms.Compose([
transforms.ToTensor(),
transforms.Normalize((0.1307,), (0.3081,))
])),
batch_size=test_batch_size, shuffle=True, **kwargs)
Optimization¶
In [14]:
model = Net().to(device)
optimizer = optim.SGD(model.parameters(), lr=lr, momentum=momentum)
Training¶
In [16]:
for epoch in range(1, epochs + 1):
# Train Mode
model.train()
for batch_idx, (data, target) in enumerate(train_loader):
data, target = data.to(device), target.to(device)
optimizer.zero_grad() # backpropagation 계산하기 전에 0으로 기울기 계산
output = model(data)
loss = F.nll_loss(output, target) # https://pytorch.org/docs/stable/nn.html#nll-loss
loss.backward() # 계산한 기울기를
optimizer.step()
if batch_idx % log_interval == 0:
print('Train Epoch: {} [{}/{} ({:.0f}%)]\tLoss: {:.6f}'.format(
epoch, batch_idx * len(data), len(train_loader.dataset),
100. * batch_idx / len(train_loader), loss.item()))
# Test mode
model.eval() # batch norm이나 dropout 등을 train mode 변환
test_loss = 0
correct = 0
with torch.no_grad(): # autograd engine, 즉 backpropagatin이나 gradient 계산 등을 꺼서 memory usage를 줄이고 속도를 높임
for data, target in test_loader:
data, target = data.to(device), target.to(device)
output = model(data)
test_loss += F.nll_loss(output, target, reduction='sum').item() # sum up batch loss
pred = output.argmax(dim=1, keepdim=True) # get the index of the max log-probability
correct += pred.eq(target.view_as(pred)).sum().item() # pred와 target과 같은지 확인
test_loss /= len(test_loader.dataset)
print('\nTest set: Average loss: {:.4f}, Accuracy: {}/{} ({:.0f}%)\n'.format(
test_loss, correct, len(test_loader.dataset),
100. * correct / len(test_loader.dataset)))