在 Pytorch 中使用自定义损失训练模型如何设置优化器并运行训练？

Question

I am new to pytorch and I am trying to run a github model I found and test it. So the author's provided the model and the loss function.

like this:

#1. Inference the model
model = PhysNet_padding_Encoder_Decoder_MAX(frames=128)
rPPG, x_visual, x_visual3232, x_visual1616 = model(inputs)

#2. Normalized the Predicted rPPG signal and GroundTruth BVP signal
rPPG = (rPPG-torch.mean(rPPG)) /torch.std(rPPG)     # normalize
BVP_label = (BVP_label-torch.mean(BVP_label)) /torch.std(BVP_label)     # normalize

#3. Calculate the loss
loss_ecg = Neg_Pearson(rPPG, BVP_label)

Dataloading

    train_loader = torch.utils.data.DataLoader(train_set, batch_size = 20, shuffle = True)

    batch = next(iter(train_loader))

    data, label1, label2 = batch

    inputs= data

Let's say I want to train this model for 15 epochs. So this is what I have so far: I am trying to set the optimizer and training, but I am not sure how to tie the custom loss and data loading to the model and set the 15 epoch training correctly.

optimizer = optim.SGD(model.parameters(), lr=0.001, momentum=0.9)

for epoch in range(15):
  ....

Any suggestions?

Answer 1

I assumed BVP_label is label1 of train_loader

train_loader = torch.utils.data.DataLoader(train_set, batch_size = 20, shuffle = True)

# Using GPU
device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")

model = PhysNet_padding_Encoder_Decoder_MAX(frames=128)
model.to(device)

optimizer = optim.SGD(model.parameters(), lr=0.001, momentum=0.9)

for epoch in range(15):
    model.train()
    for inputs, label1, label2 in train_loader:
        rPPG, x_visual, x_visual3232, x_visual1616 = model(inputs)
        BVP_label = label1 # assumed BVP_label is label1

        rPPG = (rPPG-torch.mean(rPPG)) /torch.std(rPPG)
        BVP_label = (BVP_label-torch.mean(BVP_label)) /torch.std(BVP_label)
        
        loss_ecg = Neg_Pearson(rPPG, BVP_label)
        
        optimizer.zero_grad()
        loss_ecg.backward()
        optimizer.step()

PyTorch training steps are as belows.

• Create DataLoader
• Initialize model and optimizer
• Create a device object and move model to the device

in the train loop

• select a mini-batch of data
• use the model to make predictions
• calculate the loss
• loss.backward() updates the gradients of the model
• update the parameters using optimizer

As you may know you can also check PyTorch Tutorials.

Learning PyTorch with Examples

What is torch.nn really?