[FIXED] How to convert raw image bytes, read from stream, to a tensor of a shape valid to perform Conv2d?

Issue

Community,

I have a file of bytes organized into chunks of 16384 bytes. Each of the chunks contains an uncompressed image of size 64x64 pixels. The format of the image pixel is 8-bit ABGR.

Let's say I have successfully read the chunk into numpy.array:

buf = numpy.fromfile( dataFile, dtype=np.uint8, count=16384, offset=offs)

The question is how could I convert this array of bytes into a Pytorch tensor so that to be able to perform convolution (Conv2d)?

If I understand it properly, the aforementioned convolution (Conv2d) expects the tensor to have separate channel planes instead of a single plane of multichannel pixels.

And the extra question is how to get rid of alpha-channel in the meanwhile?

Solution

In the following code, I'm assuming that your image is stored in a row-major fashion (so your bytes are abgrabgrabgr.... and not aaaa....bbbb....gggg....rrrr)

# First I reshape the buf into an image with 4 channels
buf = buf.reshape((4, 64, 64))
# Then I remove the alpha channel by taking only the 3 last :
bgr_buf = buf[1:, :, :]
# Then I make it into a pytorch tensor : 
tensor = torch.from_numpy(bgr_buf)
# And finally, pytorch convolutions expects batchs of images. So let's make it a batch of 1
batch = tensor.view(1, 3, 64, 64)

Answered By - trialNerror