How many tokens will the image occupy?
#9
by
wiccanmind
- opened
To economize on GPU memory while fine-tuning, I reduced the --model_max_length from the default 2048 to 512. This resulted in the loss becoming 0.0. I realized that the model might not be learning anything because all 512 tokens were masked as -100 (input sequence). So my question is: How many tokens will the image occupy? Additionally, how can I find or calculate that in the model configuration?