Default image classfier model handler fails on black & white JPG

[fbbradheintz]

Repro: In the quick start, replace kitten.jpg with the attached file. There's a crash in image_classifier.py when it attempts to convert the image to a tensor, because it assumes 3 color channels.

[fbbradheintz]

Sample input to induce failure

[harshbafna]

@fbbradheintz @mycpuorg : The pre-processing and post-processing for the default image classification handler is implemented for torchvision's imagenet based models like densenet, resnet etc. Thus we use the standard normalization for imagenet models which is based on RGB images.

We expect the user to implement a custom handler for other image-classification models, for e.g. MNIST digit classifier or other models which are not trained using imagenet dataset, which uses different normalization parameters. We will add a documentation describing the expected input and output of all default handlers.

E.g :

• default image_classifer expects a RGB image file as input and returns the probability of top 5 results
• default text_classifier expects a text file as input and returns the classification of the text.

[mycpuorg]

@harshbafna
We can explicitly say that image_classifier handles only RGB images and so on for other default handlers. That is reasonable, however, it is not OK for the classifier handler to crash. We must handle it as an exception and return an error code at a minimum.

[fbbradheintz]

If you can put up the doc change, I'll close this.

[harshbafna]

Added documentation in stage_release branch.

commit : 4ab6f7c

[fbbradheintz]

Looks good. Please close after merge.

[fbbradheintz]

Regression in the 4/10 build.

[harshbafna]

@fbbradheintz : We agreed to document this and the documentation is available in master branch :
https://github.com/pytorch/serve/blob/master/docs/default_handlers.md

[fbbradheintz]

Sounds good. Thanks!