You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
I am implementing a single node training script for the decoder and it seems @lucidrains has implemented a wrapper script for this purpose that is already feature-full. Currently, the forward pass is implemented as follows:
This lacks the ability to substitute our own image embeddings in the case where we have precomputed embedding-image pairs. The functionality is already mostly supported by the Decoder network where image_embed can be passed to the forward method so this could be implemented by simply adding the image_embed parameter as a pass though to decoder.forward. However, it would also be convenient to make the clip model optional in the Decoder constructor. I already started on this a week ago in this branch by adding the ability to set clip_image_size and channels separately from a clip model.
There are only a few small changes that would be necessary to implement this feature so I could put together a pull request to do this.
The text was updated successfully, but these errors were encountered:
I am implementing a single node training script for the decoder and it seems @lucidrains has implemented a wrapper script for this purpose that is already feature-full. Currently, the forward pass is implemented as follows:
DALLE2-pytorch/dalle2_pytorch/train.py
Lines 189 to 199 in 1d5dc08
This lacks the ability to substitute our own image embeddings in the case where we have precomputed embedding-image pairs. The functionality is already mostly supported by the Decoder network where
image_embed
can be passed to the forward method so this could be implemented by simply adding theimage_embed
parameter as a pass though todecoder.forward
. However, it would also be convenient to make the clip model optional in the Decoder constructor. I already started on this a week ago in this branch by adding the ability to setclip_image_size
andchannels
separately from a clip model.There are only a few small changes that would be necessary to implement this feature so I could put together a pull request to do this.
The text was updated successfully, but these errors were encountered: