autocaption_prefix and autocaption_suffix not always added to captions

[erikbrodnick]

I observed in the training output logs that the captions did not always include the autocaption_prefix or autocaption_suffix. This is pretty important as it contains the unique trigger token.

After looking at caption.py I noticed this block:

if autocaption_prefix:
    inp += f"\n\nYou must start the caption with '{autocaption_prefix}'. "

if autocaption_suffix:
    inp += f"\n\nYou must end the caption with '{autocaption_suffix}'."

Instead of relying on the llm to add these which it is clearly failing to do, I suggest manually adding to the resulting output as follows:

output = self.tokenizer.batch_decode(output_ids, skip_special_tokens=True)[
    0
].strip()

if autocaption_prefix:
    output = f"{autocaption_prefix} {output}"
if autocaption_suffix:
    output = f"{output} {autocaption_suffix}"

print(f"Caption for {image_path}: {output}")

[tomhillable]

Also have this issue and am seeing only about 50% of the captions having the prefix or suffix enforced.

[dmund95]

Yeah I agree.

[dmd]

I am seeing less than 20% of captions having the prefix or suffix. I have worked around it for the time being by allowing autocaptioning to proceed, then cancelling the run as soon as it finishes. Then I grab the logs and manually insert my prefix on each one, and write each line to a matching .txt file, turn off autocaptioning, and start over using a new zipfile with the included .txts.

This is annoying.