PyTorch custom loss function

How should a custom loss function be implemented ? Using below code is causing error :

import torch
import torch.nn as nn
import torchvision
import torchvision.transforms as transforms
import numpy as np
import matplotlib.pyplot as plt
import torch.utils.data as data_utils
import torch.nn as nn
import torch.nn.functional as F

num_epochs = 20

x1 = np.array([0,0])
x2 = np.array([0,1])
x3 = np.array([1,0])
x4 = np.array([1,1])

num_epochs = 200

class cus2(torch.nn.Module):

    def __init__(self):
        super(cus2,self).__init__()

    def forward(self, outputs, labels):
        # reshape labels to give a flat vector of length batch_size*seq_len
        labels = labels.view(-1)  

        # mask out 'PAD' tokens
        mask = (labels >= 0).float()

        # the number of tokens is the sum of elements in mask
        num_tokens = int(torch.sum(mask).data[0])

        # pick the values corresponding to labels and multiply by mask
        outputs = outputs[range(outputs.shape[0]), labels]*mask

        # cross entropy loss for all non 'PAD' tokens
        return -torch.sum(outputs)/num_tokens


x = torch.tensor([x1,x2,x3,x4]).float()

y = torch.tensor([0,1,1,0]).long()

train = data_utils.TensorDataset(x,y)
train_loader = data_utils.DataLoader(train , batch_size=2 , shuffle=True)

device = 'cpu'

input_size = 2
hidden_size = 100 
num_classes = 2

learning_rate = .0001

class NeuralNet(nn.Module) : 
    def __init__(self, input_size, hidden_size, num_classes) : 
        super(NeuralNet, self).__init__()
        self.fc1 = nn.Linear(input_size , hidden_size)
        self.relu = nn.ReLU()
        self.fc2 = nn.Linear(hidden_size , num_classes)

    def forward(self, x) : 
        out = self.fc1(x)
        out = self.relu(out)
        out = self.fc2(out)
        return out

for i in range(0 , 1) :

        model = NeuralNet(input_size, hidden_size, num_classes).to(device)

        criterion = nn.CrossEntropyLoss()
#         criterion = Regress_Loss()
#         criterion = cus2()
        optimizer = torch.optim.Adam(model.parameters(), lr=learning_rate)

        total_step = len(train_loader)
        for epoch in range(num_epochs) : 
            for i,(images , labels) in enumerate(train_loader) : 
                images = images.reshape(-1 , 2).to(device)
                labels = labels.to(device)

                outputs = model(images)
                loss = criterion(outputs , labels)

                optimizer.zero_grad()
                loss.backward()
                optimizer.step()
#                 print(loss)

        outputs = model(x)

        print(outputs.data.max(1)[1])

makes perfect predictions on training data :

tensor([0, 1, 1, 0])

Using a custom loss function from https://cs230-stanford.github.io/pytorch-nlp.html#writing-a-custom-loss-function :

is implemented in above code as cus2

Un-commenting code # criterion = cus2() to use this loss function returns :

tensor([0, 0, 0, 0])

A warning is also returned :

UserWarning: invalid index of a 0-dim tensor. This will be an error in PyTorch 0.5. Use tensor.item() to convert a 0-dim tensor to a Python number

I've not implemented the custom loss function correctly ?

回答1:

Your loss function is programmatically correct except for below:

    # the number of tokens is the sum of elements in mask
    num_tokens = int(torch.sum(mask).data[0])

When you do torch.sum it returns a 0-dimensional tensor and hence the warning that it can't be indexed. To fix this do int(torch.sum(mask).item()) as suggested or int(torch.sum(mask)) will work too.

Now, are you trying to emulate the CE loss using the custom loss? If yes, then you are missing the log_softmax

To fix that add outputs = torch.nn.functional.log_softmax(outputs, dim=1) before statement 4. Note that in case of tutorial that you have attached, log_softmax is already done in the forward call. You can do that too.

Also, I noticed that the learning rate is slow and even with CE loss, results are not consistent. Increasing the learning rate to 1e-3 works well for me in case of custom as well as CE loss.

回答2:

If you use torch functions you should be fine

import torch 

def my_custom_loss(output, target):
    loss = torch.mean((output-target*2)**3)
    return loss

# Forward pass to the Network
# then, 
loss.backward()