JIT compatible way to convert base64 encoded image to a Tensor

[anjali-chadha]

Hello -
I have an image which is encoded as a base64 string. I want to take this string as an input and convert it to a Tensor.

How can I achieve this in a way which is torchscriptable?

So far, I have tried the two step process: 1) Convert base64 encoded image to Pillow Image using pillow library, 2) convert PIL Image to Tensor using transformations provided in torchvision
However, this approach is not torchscript compatible.

Any recommendations on how to do this in torchscript-compatible manner?

Thank you!

[mthrok]

The best I can think of is to use the underlying implementation of torchaudio's StreamReader.
It is not public API, and, base64 data has to be decoded separately.

import base64

import torch
from torchaudio.io import StreamReader


@torch.jit.script
def decode_png(data):
    s = torch.classes.torchaudio.ffmpeg_StreamReaderTensor(data, "png_pipe", None, 8046)
    s.add_video_stream(
        s.find_best_video_stream(),  # stream_index
        -1,  # frames_per_chunk
        3,  # buffer_chunk_size
        "format=pix_fmts=rgb24",  # filter_desc
        None,  # decoder
        None,  # decoder_option
        None,  # hw_accel
    )
    s.process_all_packets()
    img, = s.pop_chunks()
    return img


if __name__ == '__main__':

    # From https://codepen.io/jamiekane/pen/YayWOa
    data = b"iVBORw0KGgoAAAANSUhEUgAAABgAAAAYCAYAAADgdz34AAAABHNCSVQICAgIfAhkiAAAAAlwSFlzAAAApgAAAKYB3X3/OAAAABl0RVh0U29mdHdhcmUAd3d3Lmlua3NjYXBlLm9yZ5vuPBoAAANCSURBVEiJtZZPbBtFFMZ/M7ubXdtdb1xSFyeilBapySVU8h8OoFaooFSqiihIVIpQBKci6KEg9Q6H9kovIHoCIVQJJCKE1ENFjnAgcaSGC6rEnxBwA04Tx43t2FnvDAfjkNibxgHxnWb2e/u992bee7tCa00YFsffekFY+nUzFtjW0LrvjRXrCDIAaPLlW0nHL0SsZtVoaF98mLrx3pdhOqLtYPHChahZcYYO7KvPFxvRl5XPp1sN3adWiD1ZAqD6XYK1b/dvE5IWryTt2udLFedwc1+9kLp+vbbpoDh+6TklxBeAi9TL0taeWpdmZzQDry0AcO+jQ12RyohqqoYoo8RDwJrU+qXkjWtfi8Xxt58BdQuwQs9qC/afLwCw8tnQbqYAPsgxE1S6F3EAIXux2oQFKm0ihMsOF71dHYx+f3NND68ghCu1YIoePPQN1pGRABkJ6Bus96CutRZMydTl+TvuiRW1m3n0eDl0vRPcEysqdXn+jsQPsrHMquGeXEaY4Yk4wxWcY5V/9scqOMOVUFthatyTy8QyqwZ+kDURKoMWxNKr2EeqVKcTNOajqKoBgOE28U4tdQl5p5bwCw7BWquaZSzAPlwjlithJtp3pTImSqQRrb2Z8PHGigD4RZuNX6JYj6wj7O4TFLbCO/Mn/m8R+h6rYSUb3ekokRY6f/YukArN979jcW+V/S8g0eT/N3VN3kTqWbQ428m9/8k0P/1aIhF36PccEl6EhOcAUCrXKZXXWS3XKd2vc/TRBG9O5ELC17MmWubD2nKhUKZa26Ba2+D3P+4/MNCFwg59oWVeYhkzgN/JDR8deKBoD7Y+ljEjGZ0sosXVTvbc6RHirr2reNy1OXd6pJsQ+gqjk8VWFYmHrwBzW/n+uMPFiRwHB2I7ih8ciHFxIkd/3Omk5tCDV1t+2nNu5sxxpDFNx+huNhVT3/zMDz8usXC3ddaHBj1GHj/As08fwTS7Kt1HBTmyN29vdwAw+/wbwLVOJ3uAD1wi/dUH7Qei66PfyuRj4Ik9is+hglfbkbfR3cnZm7chlUWLdwmprtCohX4HUtlOcQjLYCu+fzGJH2QRKvP3UNz8bWk1qMxjGTOMThZ3kvgLI5AzFfo379UAAAAASUVORK5CYII="

    data = base64.b64decode(data)

    data = torch.frombuffer(data, dtype=torch.uint8)
    img = decode_png(data)

    import matplotlib.pyplot as plt

    plt.imshow(img[0].permute(1, 2, 0))
    plt.show()

[YosuaMichael]

Similar to @mthrok , we can do this with torchvision.io.image API for decoding image. However need to decode base64 separately (I can't find a torchscript-compatible way to decode base64).

import base64
import torch
from torchvision.io import image, ImageReadMode

raw_image_b64 = b"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"
raw_image_bytes = base64.b64decode(raw_image_b64)
raw_image_tensor = torch.frombuffer(raw_image_bytes, dtype=torch.uint8)

# This is jit scripted decode_image function
decode_image_script = torch.jit.script(image.decode_image)

decoded_image = decode_image_script(raw_image_tensor, mode=ImageReadMode.RGB)

import matplotlib.pyplot as plt

plt.imshow(decoded_image.permute(1, 2, 0))
plt.show()

[vadimkantorov]

I wonder if there's a simple way of transforming SSE vectorized base64 decoding to tensor ops: http://www.alfredklomp.com/programming/sse-base64/, https://r-libre.teluq.ca/1362/1/base64.pdf -> this probably could be a feature request to core pytorch as well...