Do I need to normalize data for inception-v4?

[protossw512]

Hi, I was trying to remove fully connected layer and only use pre-trained convolution layers to do fine tuning. However, the result was pretty bad. I saw in evaluate_image.py you applied: def preprocess_input(x): x = np.divide(x, 255.0) x = np.subtract(x, 1.0) x = np.multiply(x, 2.0) return x

It seems like we need to normalize input image to range [-2, 0]. Then I followed this procedure, it do improved a little bit, but still very bad.

I used pretrained ResNet 50 from Kreas to fine tune the model, and got 92% accuracy on testing data, but when I applied the same way, but with Inception-v4 net, I only got 70% accuracy. Something must be wrong, but I couldn't figure out.

3/18 edit:

I found that someone said (http://stackoverflow.com/a/39597537/7555390) for inception the input has to be normalized to [-1, 1], but your code seems like to make it [-2, 0], is that a typo?

[kentsommer]

The preprocessing applied as

def preprocess_input(x): 
   x = np.divide(x, 255.0) 
   x = np.subtract(x, 1.0) 
   x = np.multiply(x, 2.0) 
   return x

This is not a typo.

Without knowing what your training data is, it is hard to pinpoint what might be going wrong. Within the next couple days I will be releasing a Keras-2 version of this model with a few extra training related hyperparameters that I did not include in this original release. Perhaps that is what is causing issues for you.

[protossw512]

It is a private dataset. It is interesting since the processing method actually normalizes each pixel to [-2, 0]. I switched to tensorflow backend and it worked like a magic. I also tried normalize data to [-2, 0] and [-1, 1], there is little difference on accuracy.

Thank you for your reply and your effort on providing inception implementation on Keras.

[pengpaiSH]

+1. I came into the same problem. With TensorFlow backend, I could get 80% classification accuracy with ResNet50 while I only got 70% with InceptionV4. Something goes wrong... By the way, I have modified the codes in Keras 2 style.

[igrekun]

I ran resnet-50 and inception-v4 on the same dataset for 100 epoch and here's what i got. In case this might be useful. Accuracy Loss

Tensorflow backend, keras-v2. Both models have last layer replaced by 2 dense. Can confirm inception strangely stops training at ~70% accuracy

[zyxue]

The slow convergence doesn't seem to be concordant with the comparison in the InceptionV4 paper, the Figure 25

[kentsommer]

@zyxue @igrekun @pengpaiSH @protossw512

Hi all! Sorry for taking so long to address these issues. I've added a bunch of training parameters and tweaked a few others so that the Keras version matches the tf.slim training parameters exactly. Hopefully this will fix the issues mentioned above. Thanks again for being so patient!

[kentsommer]

I'll close this out for now, if anyone runs into similar issues feel free to either open a new issue or comment on this one and I'll re-open it.