RoBERTa-base hidden_size Mismatch Error

[toututuya]

I am encountering an issue while working with the roberta-base model configuration. I noticed that the following configuration is being used:
RobertaConf.hidden_size = 256
RobertaConf.intermediate_size = 1024
RobertaConf.num_attention_heads = 8
However, when running the model, I encounter the following error:
RuntimeError: Given normalized_shape=[768], expected input with shape [*, 768], but got input of size[1, 198, 256]
The error suggests a shape mismatch where the model expects an input with a hidden_size of 768, but the input has a size of [1, 198, 256]. Given that the hidden_size has been changed to 256, I am unsure how to handle this discrepancy.

[Vision-Kek]

I ran into the same error, for me this issue was that I used a piece of code that called the resizer (768->256) after text encoding, whereas apparently it shoulnd't for ParSeDA_v2 with GLIP_attn as I found in the below snippet which solved it

if args.fusion_type == "GLIP_attn":
    text_memory_resized = text_memory
else:
    text_memory_resized = model.module.transformer.resizer(text_memory)

[toututuya]

I ran into the same error, for me this issue was that I used a piece of code that called the resizer (768->256) after text encoding, whereas apparently it shoulnd't for ParSeDA_v2 with GLIP_attn as I found in the below snippet which solved it

if args.fusion_type == "GLIP_attn":
text_memory_resized = text_memory
else:
text_memory_resized = model.module.transformer.resizer(text_memory)

Thank you so much for your detailed explanation!