AttributeError: 'tuple' object has no attribute 'log_softmax'

Question

while trying to finetune inception_V3 for my own dataset by changing the last fc layer like

last_layer =nn.Linear(n_inputs, len(classes))
inception_v3.fc = last_layer

after that when I train it got this error on this position

    # on training loop
    output = inception_v3(data)
    # calculate the batch loss
    loss = criterion(output, target)

Error is

 AttributeError: 'tuple' object has no attribute 'log_softmax'

Harshit Kumar · Accepted Answer · 2018-11-26 07:48:37Z

14

This is a well known problem.

Try one the following solutions:

disable aux_logits when the model is created here by also passing aux_logits=False to the inception_v3 function.
edit your train function to accept and unpack the returned tuple to be something like: output, aux = model(input_var)

Check the following link for more info.

answered Nov 26, 2018 at 7:48

Harshit Kumar

13.1k10 gold badges62 silver badges78 bronze badges

Sign up to request clarification or add additional context in comments.

1 Comment

solved by doing inception_v3 = models.inception_v3(pretrained=True) inception_v3.aux_logits=False

prosti · Accepted Answer · 2019-07-08 23:09:02Z

4

This problem seams to me like instead you define F:

import torch.nn.functional as F

You in accident have set F to some tuple

F=(1,2)

And then when you call F.log_softmax you get exactly this error.

answered Jul 8, 2019 at 23:09

prosti

46.9k19 gold badges199 silver badges161 bronze badges