I want to submit a pytorch question about the model structure of GoogLeNet provided by pytorch. When using GoogLeNet provided by pytorch, I found that there is a problem with the structure of the model. The specific situation is as follows，The following figure shows the structure of the model mentioned in the GoogLeNet paper
In the third branch, a 5*5 convolution kernel is used, but in the model provided by pytorch, a 3*3 convolution kernel is used. The following figure is the network structure provided by pytorch. We can see that in the branch The number of large and small convolution kernels in the three is 3*3
I discovered this problem when migrating the pre-training weights: the model provided by pytorc is different from the model in the paper. I don’t know if I am wrong or if pytorch was negligent when providing the model, so I am here to ask, I hope it can help Got an answer, thank you