diff options
Diffstat (limited to 'README.md')
| -rw-r--r-- | README.md | 34 |
1 files changed, 8 insertions, 26 deletions
@@ -11,44 +11,26 @@ TBC Extract text from the iam dataset: ``` -poetry run extract-iam-text --use_words --save_text train.txt --save_tokens letters.txt +poetry run python extract-iam-text --use_words --save_text train.txt --save_tokens letters.txt ``` Create word pieces from the extracted training text: ``` -poetry run make-wordpieces --output_prefix iamdb_1kwp --text_file train.txt --num_pieces 100 +poetry run python make-wordpieces --output_prefix iamdb_1kwp --text_file train.txt --num_pieces 100 ``` Optionally, build a transition graph for word pieces: ``` -poetry run build-transitions --tokens iamdb_1kwp_tokens_1000.txt --lexicon iamdb_1kwp_lex_1000.txt --blank optional --self_loops --save_path 1kwp_prune_0_10_optblank.bin --prune 0 10 +poetry run python build-transitions --tokens iamdb_1kwp_tokens_1000.txt --lexicon iamdb_1kwp_lex_1000.txt --blank optional --self_loops --save_path 1kwp_prune_0_10_optblank.bin --prune 0 10 ``` (TODO: Not working atm, needed for GTN loss function) ## Todo -- [x] create wordpieces - - [x] make_wordpieces.py - - [x] build_transitions.py - - [x] transform that encodes iam targets to wordpieces - - [x] transducer loss function -- [ ] Train with word pieces - - [ ] Pad word pieces index to same length -- [ ] Local attention in first layer of transformer -- [ ] Halonet encoder -- [ ] Implement CPC - - [ ] https://arxiv.org/pdf/1905.09272.pdf - - [ ] https://pytorch-lightning-bolts.readthedocs.io/en/latest/self_supervised_models.html?highlight=byol - - -- [ ] Predictive coding - - https://arxiv.org/pdf/1807.03748.pdf - - https://arxiv.org/pdf/1904.05862.pdf - - https://arxiv.org/pdf/1910.05453.pdf - - https://blog.evjang.com/2016/11/tutorial-categorical-variational.html - - - - +- [ ] Reimplement transformer from scratch +- [ ] Implement Nyström attention (for efficient attention) +- [ ] Dino +- [ ] Efficient-net b0 + transformer decoder +- [ ] Test encoder pre-training ViT (CvT?) with Dino, then train decoder in a separate step ## Run Sweeps |