Why is it called Enc_only instead of Decoder in the model?

[redEddie]

Hi, thanks for releasing this codebase!

I’m running the LIBERO BC experiments with the following command:

bash scripts/bc/libero_so_bc_mamba_dec.sh

When I look at the YAML file for the model being executed, it appears that the Enc_only class is being used (line 15). I noticed that Enc_only returns pred_actions in its forward() method. This made me wonder why it is named Enc_only instead of something like Decoder, since it seems to be part of the decoding process.

Here are the specific lines I’m referring to:

YAML file: bc_mamba_dec.yaml#L15

Code: transformer.py#L296

Could someone clarify the reason behind this naming choice? Thanks in advance!

[awesomeNabi]

The class is named Enc_only because it uses only an encoder-style Transformer block internally.
It does not use:

• a decoder block,

• decoder layers,

• causal self-attention,

Even though the final output is named pred_actions, this output does not make the module a decoder.In diffusion-policy,this is called a BERT-like architecture
It simply means:“The encoder processed the entire noisy action sequence and output its denoised prediction.”
This is typical for diffusion models, where the denoiser is usually:a UNet (encoder-like),or an encoder-only Transformer,not an autoregressive decoder.
So the name Enc_only describes the architecture, not the output’s meaning.