diff options
Diffstat (limited to 'README.md')
| -rw-r--r-- | README.md | 68 |
1 files changed, 16 insertions, 52 deletions
@@ -1,64 +1,28 @@ # Text Recognizer -Implementing the text recognizer project from the course ["Full Stack Deep Learning Course"](https://fullstackdeeplearning.com/march2019) in PyTorch in order to learn best practices when building a deep learning project. I have expanded on this project by adding additional feature and ideas given by Claudio Jolowicz in ["Hypermodern Python"](https://cjolowicz.github.io/posts/hypermodern-python-01-setup/). +Implementing the text recognizer project from the course ["Full Stack Deep Learning Course"](https://fullstackdeeplearning.com/march2019) (FSDL) in PyTorch in order to learn best practices when building a deep learning project. I have expanded on this project by adding additional feature and ideas given by Claudio Jolowicz in ["Hypermodern Python"](https://cjolowicz.github.io/posts/hypermodern-python-01-setup/). ## Setup TBC -## Todo -- [x] subsampling -- [x] Be able to run experiments -- [x] Train models -- [x] Fix input size in base model -- [x] Fix s.t. the best weights are saved -- [x] Implement total training time -- [x] Fix tqdm and logging output -- [x] Fix basic test to load model -- [x] Fix loading previous experiments -- [x] Able to set verbosity level on the logger to terminal output -- [x] Implement Callbacks for training - - [x] Implement early stopping - - [x] Implement wandb - - [x] Implement lr scheduler as a callback - - [x] Implement save checkpoint callback - - [x] Implement TQDM progress bar (Low priority) -- [ ] Check that dataset exists, otherwise download it form the web. Do this in run_experiment.py. -- [x] Create repr func for data loaders -- [ ] Be able to restart with lr scheduler (May skip this) -- [ ] Implement population based training -- [x] Implement Bayesian hyperparameter search (with W&B maybe) -- [x] Try to fix shell cmd security issues S404, S602 -- [x] Change prepare_experiment.py to print statements st it can be run with tasks/prepare_sample_experiments.sh | parallel -j1 -- [x] Fix caption in WandbImageLogger -- [x] Rename val_accuracy in metric -- [x] Start implementing callback list stuff in train.py -- [x] Fix s.t. callbacks can be loaded in run_experiment.py -- [x] Lift out Emnist dataset out of Emnist dataloaders -- [x] Finish Emnist line dataset -- [x] SentenceGenerator -- [x] Write a Emnist line data loader -- [x] Implement ctc line model - - [x] Implement CNN encoder (ResNet style) - - [x] Implement the RNN + output layer - - [x] Construct/implement the CTC loss -- [x] Sweep base config yaml file -- [x] sweep.py -- [x] sweep.yaml -- [x] Fix dataset splits. -- [x] Implement predict on image -- [x] CTC decoder -- [x] IAM dataset -- [x] IAM Lines dataset -- [x] IAM paragraphs dataset -- [ ] CNN + Transformer (!!) -- [ ] CNN + GPT -- [ ] fix nosec problem -- [x] common Dataset class -- [x] Fix CTC blank stuff and varying length -- [x] Metric Learning for backbone training + + +## Todo +- [ ] create wordpieces + - [x] make_wordpieces.py + - [x] build_transitions.py + - [ ] transform that encodes iam targets to wordpieces + - [ ] transducer loss function +- [ ] 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 + - [ ] + ## Run Sweeps Run the following commands to execute hyperparameter search with W&B: |