esalesky / visrep Goto Github PK

Hello,

Currently, the encode method in hub_interface.py only returns tensors from image slices and decode is only called in translate and is also not the decoder of the transformer model.

Is there an easier way to access encoded tensors than _build_batches, which are actually generated from VisualTextDataset? Including decoded tensors generated in SequenceGenerator and called in inference_step.

Thank you

As discussed in person, I am interested in using this to translate from SignWriting to spoke language text.

SignWriting, can be linearly represented like so:

However the way it is intended to be read by humans is like so:

While the specification allows for up to 500x500 representation for each sign, in practice, signs are a lot smaller.
For example, the white space in this sign is huge:

I will need to find what size covers 99% of the data.

Looking forward to multi README so I know how to prep the image data, increase the patch size/use multiple vertical and horizontal patches, and train the model

Hi, I have been trying to implement the code on MTTT dataset as given in the paper. But while loading the data during fairseq train, I am getting the following error 'No such dataset implementation None' probably while loading the data.

Any kind of help would be greatly appreciated

Command line for traing in case i am doing something wrong
CUDA_VISIBLE_DEVICES=0 fairseq-train data-bin/MMMT.tokenized.en-tr --task visual_text --source-lang en --target-lang tr --target-dict dict.tr.txt --arch visual_text_transformer --image-window 15 --image-stride 10 --image-font-path fairseq/data/visual/fonts/NotoSans-Regular.ttf --image-embed-normalize --image-embed-type 1layer --share-decoder-input-output-embed --optimizer adam --adam-betas '(0.9, 0.98)' --clip-norm 0.0 --lr 5e-4 --lr-scheduler inverse_sqrt --warmup-updates 4000 --dropout 0.3 --weight-decay 0.0001 --criterion label_smoothed_cross_entropy --label-smoothing 0.1 --max-tokens 4096 --max-epoch 50

Hi,

Currently, there is no method to load the checkpoint together with their dictionaries.
It would be great if you have a PyTorch Hub interface for the model. Inference would be easier, and it could be reused for further applications.

Thanks