Kevin

Can you try changing the line 292 in /libs/datasets/download_and_convert_coco.py `img = np.array(Image.open(img_name)) ` to `img = np.asarray(Image.open(img_name)) ` ? This might not be the solution, it is just a lucky...

The error comes from the fact that if you open a file with `Image.open(...)` It creates an Image object which is from the PIL-library, which is called 'JpegImageFile' . The...

@Designbook1 I would love to tell you but I am not able to reproduce the problem. You can give the command `np.array(Image.open(img_name)) `a dataype so it should convert it automatically...

A nice feature request would be a checksum to check if the downloaded data is ok.

An idea for further investigation: 1. Try to replace `Image.open `with `cv2.imread`, to do so you must install opencv (pip install opencv-python) If this fixes the error there is an...

Maybe this would be a better description to be put into the README ? @CharlesShang

Did one of you manage to actually train the network completly with a decent loss ?

Can you "reproduce" the random occurance with a certain seed for the random initialization?

This is adressed in the original paper of the the Mask RCNN. They trained for 2 days on a 8 GPU machine. Since parallelism does not speed up linear this...

Since the original paper uses the Faster RCNN hyperparameters mini-batching was used, which means that a batch-size greater than 1 must be used. However the batchsize will affect the learning...