RL training: Base64 image decode errors (Incorrect padding / Invalid base64-encoded string)

[Ryann-Ran]

Hi, thanks for your great work!

I’m trying to reproduce the RL training and I’m consistently seeing errors like:

• [ERROR code] Failed to decode image 0: Incorrect padding
• [ERROR code] Failed to decode image 0: Invalid base64-encoded string: number of data characters (336877) cannot be 1 more than a multiple of 4

My main question is whether these errors correspond to the issue described in the README (i.e., caused by network pressure):

During RL training, transmitting a large number of high-resolution images to a single node in a short period may saturate the bandwidth, leading to retransmissions and timeouts.

I tried launching 3 code servers, each with 4 processes, on a single machine with 8× A800 (80GB), but it didn’t work out. I used commands like:

IMAGE=chenshawn6915/multimodal-ipython-sandbox:oss-v2
docker run -p 18901:18901 -p 18902:18902 -p 18903:18903 -p 18904:18904 $IMAGE
docker run -p 28901:18901 -p 28902:18902 -p 28903:18903 -p 28904:18904 $IMAGE
docker run -p 38901:18901 -p 38902:18902 -p 38903:18903 -p 38904:18904 $IMAGE

[Ryann-Ran]

I solved this problem by revising the function in reinforcement_learning/verl/workers/agent/envs/deepeyesv2/deepeyesv2.py:

def base64_to_pil_image(base64_string: str) -> Image.Image:
    payload = str(base64_string).strip()
    if payload.startswith("data:image"):
        payload = payload.split(",", 1)[1]
    payload = payload.replace("\n", "").replace(" ", "")
    payload += "=" * (-len(payload) % 4)
    image_data = base64.b64decode(payload)
    image = Image.open(BytesIO(image_data))
    image.load()
    return image