Preface:

In recent run data vgg16, but the parameters of the amount is too great, running very slow, although the final results are fairly impressive. Followed by another study a little GoogLeNet, write their own code inceptionV1, because the computer ran vgg16 been occupied, so we did not run.

During this period, little brother on the computer ran MobileNet V1, can not find the right pytorch pre-training version of the heavy, so to start from scratch to run, over-fitting is very serious, look after and so what is the reason, speculation is too unbalanced datasets serious and no pre-training.

Today we called directly run the code data of pytorch inception V3, and for this network has not been studied in depth, so when it came to run some of the more low-level problem, but also not search on Baidu, finally found a solution on Google .

problem:

'InceptionOuputs' object has no attribute 'log_softmax'

In the error Torch 0.4.0 should be reported:

'tuple' object has no attribute 'log_softmax'

Solution:

inception series has an auxiliary classifier in training when you can choose whether to use the auxiliary classifier, I chose to use the time in training, so the output of the network are:

       if self.training and self.aux_logits:
           return _InceptionOuputs(x, aux)

And_InceptionOuputsIt is defined as the top of the code:

_InceptionOuputs = namedtuple('InceptionOuputs', ['logits', 'aux_logits'])

The return value is a tuple type, but in the train.py

outputs = model(inputs)

The return value is assigned to only one value, so in this caseoutputsIs atupleType value, and thenoutputsInput to function calculates the loss function to go and they will report the above error.

But why is version 0.4.0, when reported errors are not the same?

This is because the modified version of the code to return the output of the network is so written:

        if self.training and self.aux_logits:
            return x, aux

Simply not defined_InceptionOuputs！

But the search in Google are also'tuple' object has no attribute 'log_softmax'The solution to this problem, about the new version of the problem has not yet seen.