Image for training and prediction have different size？

[chptx]

in predict.py, image stream is cropped before prediction.

def _convert_stream_to_array(stream):
    ...
    image = image.crop((0, PICAMERA_RESOLUTION_HEIGHT / 2, PICAMERA_RESOLUTION_WIDTH, PICAMERA_RESOLUTION_HEIGHT))
    ...

but in train.py, it not do the same step in load_images_to_array, is this a bug?

[Vignesh1115]

Hi @chptx, We are cropping the image in predict.py because we expect the images supplied while training to be cropped. Since the most useful information for predicting the direction to move in would be in the lower half of the image, we expect the lower half of the image to be supplied as the training data. We chose not to do the cropping of the image programatically in case anyone wanted to use the complete image data. I do see that this crucial aspect is not very clear in the README as it is in the blog post and will be updating it to reflect the same.

[chptx]

I think it should crop the image in train.py to work correctly out-of-the-box for developer who use this project.

Pull requests #15