How to get the probability of prediction in the transfer learning

1

Hi I am using the transfer learning tutorial for my classification (two labels) with the following link: https://pytorch.org/tutorials/beginner/transfer_learning_tutorial.html.

except data loader part, The code about mode is only changed with following code which change the class numbers:

model_ft = models.resnet18(pretrained=True)
num_ftrs = model_ft.fc.in_features
model_ft.fc = nn.Linear(num_ftrs, 2)

I found the prediction of outputs in the following code:

 with torch.set_grad_enabled(phase == 'train'):
                    outputs = model(inputs)
                    print(outputs.shape, outputs)
                    _, preds = torch.max(outputs, 1)

The output like the following result:

torch.Size([32, 2]) tensor([[ 3.1194, -1.4058],
        [ 1.4559, -1.3360],
        [ 1.9011, -1.0287],
        [ 0.8083, -1.0037],
        [ 0.7831, -1.1790],
        [ 0.7990, -1.3758],

I found that the number is not probability of prediction like softmax function in the keras output .

Now I am using 5-folds to predict the test data, and I want to get the probability of each folds
,and average of the sum all folds probabilities.

I don’t know how to change the code for my goal of getting probability of prediction.

2

This tutorial uses torch.nn.CrossEntropyLoss() as the loss function, which means that you should just run torch.nn.functional.softmax on your outputs, since it learned log-softmax values.

1 Like
3

thanks for your answer.