Encode array with alpha channel

[phatngy]

Describe the question.

I have a numpy array with alpha channel and want to use nvimagecodec to encode it. But it doesn't seem to support it and returns None.

encoder = nvimgcodec.Encoder()
arr = np.random.randint(0, 255, (1024, 1024, 4), dtype=np.uint8)
enc_bytes = encoder.encode(arr, "png")

I see nvimagecodec can transcode a PNG image with alpha channel.

decoder = nvimgcodec.Decoder()
png = nvimgcodec.CodeStream("image.png") // image with alpha channel
img_d = decoder.decode(png, nvimgcodec.DecodeParams(color_spec=nvimgcodec.ColorSpec.UNCHANGED))
enc_bytes = encoder.encode(img_d, "png")

So I tried to convert numpy array to nvimagecodec.Image and got this error:
RuntimeError: Unexpected number of channels. Only 3 channels for RGB or 1 channel for gray scale are accepted.

Check for duplicates

• I have searched the open bugs/issues and have found no duplicates for this bug report

[smatysik-nv]

Thank you @phatngy for raising this - it's indeed a missing case that we should support.

As the log indicates, currently only 3 channels for RGB or 1 channel for grayscale are accepted when creating an nvimgcodec.Image. When you pass a NumPy array to encode, it first converts it to an nvimgcodec.Image, which is why you encounter this error.

In the case of transcoding, you first decode using nvimgcodec.ColorSpec.UNCHANGED, which allows an image with 4 channels to be created. Since encode doesn’t perform any conversions in this scenario, it accepts the image as-is.

We initially limited support to 3 channels because not all encoders are equipped to handle more, and they should at least fail gracefully if unsupported. The PNG encoder itself is capable of handling 4 channels, but due to this general limitation, it's currently blocked.

Apologies for the inconvenience - we plan to address this in the next release.