def test(model, batch_size, save):
# evaluate
model.eval()
torch_device = torch.device('cuda')
# Loss and optimizer 定义损失函数， 使用的是最小平方误差函数
criterion = nn.MSELoss()
criterion = criterion.to(torch_device)
eval_loss_dict = []
test_loader = get_data.get_test_loader(batch_size)
batch_tqdm = tqdm(enumerate(test_loader, 0), total=test_loader.__len__())
for i, data in batch_tqdm:
inputs, labels = data
labels = get_data.get_size_labels(1, labels)
inputs = inputs.cuda()
labels = labels.cuda()
outputs = model(inputs)
loss = criterion(outputs, labels)
loss = loss.cuda()
eval_loss_dict.append(loss.item())
if save:
for j in range(10):
save_imgname = 'G:/evocnn/image_set/output/' + str(i * 10 + j) + '.tif'
img = TIFF.open(save_imgname, 'w')
# 要保存之前先要改成CPU
output_img = outputs[j].cpu().detach().numpy()
img.write_image(output_img, write_rgb=True)
mean_test_loss = np.mean(eval_loss_dict)
std_test_loss = np.std(eval_loss_dict)
print("valid mean:{},std:{}".format(mean_test_loss, std_test_loss))
def train(model, batch_size, optimizer, model_type):
torch_device = torch.device('cuda')
model.cuda()
if model_type == 1:
train_loader = get_data.get_final_train_loader(batch_size)
elif model_type == 2:
train_loader = get_data.get_gauss50_final_train_loader(batch_size)
else:
train_loader = get_data.get_mixed_final_train_loader(batch_size)
# Loss and optimizer 3.定义损失函数， 使用的是最小平方误差函数
criterion = nn.MSELoss()
criterion = criterion.to(torch_device)
loss_dict = []
num_epochs = train_loader.__len__()
# Train the model 5. 迭代训练
model.train()
batch_tqdm = tqdm(enumerate(train_loader, 0), total=num_epochs, ncols=100)
for i, data in batch_tqdm:
# Convert numpy arrays to torch tensors 5.1 准备tensor的训练数据和标签
inputs, labels = data
labels = get_data.get_size_labels(1, labels)
inputs = inputs.cuda()
labels = labels.cuda()
# labels = get_data.get_size_labels(indi.get_layer_size(),labels)
# Forward pass 5.2 前向传播计算网络结构的输出结果
optimizer.zero_grad()
outputs = model(inputs)
# 5.3 计算损失函数
loss = criterion(outputs, labels)
loss = loss.cuda()
# Backward and optimize 5.4 反向传播更新参数
loss.backward()
optimizer.step()
# 可选 5.5 打印训练信息和保存loss
loss_dict.append(loss.item())
print('Loss: {:.4f}'.format(np.mean(loss_dict)))
file_path = os.getcwd() + '/loss.txt'
with open(file_path, 'a') as myfile:
myfile.write(str(np.mean(loss_dict)))
myfile.write("\n")
def parse_individual(self, indi):
torch_device = torch.device('cuda')
cnn = CNN(indi)
cnn.cuda()
print(cnn)
complexity = get_total_params(cnn.cuda(), (220, 30, 30))
train_loader = get_data.get_mixed_train_loader(self.batch_size)
# Loss and optimizer 3.定义损失函数， 使用的是最小平方误差函数
criterion = nn.MSELoss()
criterion = criterion.to(torch_device)
# 4.定义迭代优化算法， 使用的是Adam，SGD不行
learning_rate = 0.004
optimizer = torch.optim.Adam(cnn.parameters(), lr=learning_rate)
loss_dict = []
num_epochs = train_loader.__len__()
# Train the model 5. 迭代训练
cnn.train()
for i, data in enumerate(train_loader, 0):
# Convert numpy arrays to torch tensors 5.1 准备tensor的训练数据和标签
inputs, labels = data
labels = get_data.get_size_labels(1, labels)
inputs = inputs.cuda()
labels = labels.cuda()
# labels = get_data.get_size_labels(indi.get_layer_size(),labels)
# Forward pass 5.2 前向传播计算网络结构的输出结果
optimizer.zero_grad()
outputs = cnn(inputs)
# 5.3 计算损失函数
loss = criterion(outputs, labels)
loss = loss.cuda()
# Backward and optimize 5.4 反向传播更新参数
loss.backward()
optimizer.step()
# 可选 5.5 打印训练信息和保存loss
loss_dict.append(loss.item())
if (i + 1) % 100 == 0:
print('Epoch [{}/{}], Loss: {:.4f}'.format(i + 1, num_epochs, loss.item()))
# evaluate
cnn.eval()
eval_loss_dict = []
valid_loader = get_data.get_mixed_validate_loader(self.batch_size)
for i, data in enumerate(valid_loader, 0):
inputs, labels = data
labels = get_data.get_size_labels(1, labels)
inputs = inputs.cuda()
labels = labels.cuda()
outputs = cnn(inputs)
loss = criterion(outputs, labels)
loss = loss.cuda()
eval_loss_dict.append(loss.item())
mean_test_loss = np.mean(eval_loss_dict)
std_test_loss = np.std(eval_loss_dict)
print("valid mean:{},std:{}".format(mean_test_loss, std_test_loss))
return mean_test_loss, std_test_loss, complexity