Conversion of .pt to .ptl for classification models giving very poor results

1

I am trying to build a object detection model which detects flies on cow. I have done a good job detecting flies using yolo frameworks but this only happens when the quality of the image is top notch as well as the position of the cow is closer to the screen. i.e If the image is taken from a short distance. So, I planned to reject the bad images ( quality as well as position ) using a classification model. I build a 5 class classification model -

bad class - having blurry images and cow which are far away.
difficult - some what decent images as compared to bad
good - good quality and short distance cows.
no_cow - images in which cows are absent. ( Basically the dataset consists of green pasture images without any cow in them )
not_cow - Animals other than cow.

I trained the model in pytorch and the .pt is doing a good job in predicting the classes with a validation accuracy of 77%. But when I convert the .pt file to .ptl and run inference on mobile using same set of images ( I am using react-native-pytorch-core library ) I get a validation accuracy of 62%. But the .ptl model 99% of the time is only predicting among the good,no_cow and not_cow images whereas the .pt model is somewhat equally distributed among all the classes. What may me the reason?

2

Okay, so after trying various methods we figured out above issues to some extent. Writing this as it may help somebody someday.

As we were using a pre-trained ImageNet model , normalizing the tensor with values MEAN = [0.485, 0.456, 0.406] and STD = [0.229, 0.224, 0.225] instead of simply passing [0.5,0.5,0.5] solved few issues. Next we compared the image tensors which were getting passed as input to the model and we found out that they were quiet different. We were using pytorch in the web code and react-native-pytorch-core library for our mobile code. Visualizing the image tensors helped us a lot. As per the documentation of react-native-pytorch-core and the example provided by them for image classification, we had used centerCrop function due to which the image was getting cropped and hence the results varied significantly between web and mobile. Removing the centerCrop step and resizing the image to the desired size reduced the difference between web and mobile results to a great extent. Some difference is also attributable to the differences in underlying resizing functions in opencv and PIL.