Using pretrained weights of ResNet50

VMS · February 16, 2022, 7:48am

Hi .

I am using transfer learning for image classification of yoga poses . This is a part of learning journey in ML for computer vision.
I have tried to building CNN models . As well as tried transfer learning using pre-trained models from keras application.
What is puzzling to me is that while using pre-trained weights of ResNet50 , with out preprocessing the images I get an accuracy of 0.99 . How ever , when I subject the images to preprocess_input function , the accuracy drops to 0.8 .
Any one can point out why ?
Thx!

Bhack · February 16, 2022, 11:11am

Generally the default preprocess is in Caffe mode:

github.com

keras-team/keras/blob/master/keras/applications/imagenet_utils.py#L69-L80

      
        
            PREPROCESS_INPUT_MODE_DOC = """
                mode: One of "caffe", "tf" or "torch". Defaults to "caffe".
                  - caffe: will convert the images from RGB to BGR,
                      then will zero-center each color channel with
                      respect to the ImageNet dataset,
                      without scaling.
                  - tf: will scale pixels between -1 and 1,
                      sample-wise.
                  - torch: will scale pixels between 0 and 1 and then
                      will normalize each channel with respect to the
                      ImageNet dataset.
              """

Have you tried to finetune a little bit this model on your dataset:

VMS · February 16, 2022, 11:35am

Hi . No haven’t . Since with out preprocessing , accuracy of 0.98 - 0.99 was achieved on the test data . But unable to understand as to why should accuracy reduce after preprocessing . Will try though.

Bhack · February 16, 2022, 3:52pm

I cannot see if there is something specific with your images statistics as if you see the images are standardized over ImageNet stats:

zero-center each color channel with respect to the ImageNet dataset

VMS · February 17, 2022, 4:06am

Thanks , will check that .