Tracking progress

[patrickphat]

Date: 19/03 Task:

• [x] Prepare data: Crop image into batches of text-lines given ground truth coordinates @patrickphatnguyen
• [x] Train CRNN baseline model: Train a CRNN model with prepared data @patrickphatnguyen
• [x] Write documentation on loading prepared dataset and labels for OCR task @patrickphatnguyen here

[patrickphat]

Date: 20/03 Task:

• [x] Train CTPN for layout detection module @0902338471
• [x] Write document on loading prepared file and training for this module @0902338471
• [x] Train CRNN baseline model: Train a CRNN model with prepared data (need to learn more about preparing data for CTC loss) urgent @patrickphatnguyen

[patrickphat]

22/03

• [x] Train CRNN for OCR @0902338471

[patrickphat]

23/03

• [x] Prepare data for Graph CNN model @patrickphatnguyen
• [x] Read paper on Graph CNN @patrickphatnguyen
• [x] Read paper on Graph CNN @0902338471

[patrickphat]

24/03

• [x] Write report here @0902338471
• [x] Implement Graph CNN @patrickphatnguyen

[patrickphat]

25/03

• [x] Fix GCN Model @0902338471
• [ ] Read explainer paper @patrickphatnguyen @0902338471
• [ ] Implementing explainer model @patrickphatnguyen
• [ ] Write report for KV part @patrickphatnguyen

[patrickphat]

28/03

• [x] Re-process data @0902338471

• Access notebook here

• Access data here here

• Remove the "blocking" sight part in building graph, if direct sight then connect, no matter what

• Get rid of the BERT model

• Add special tokens: all numbers map to [number] (e.g: RM98 -> RM 98 -> RM [number] ), all date map to [date]

• Use BoW to represent feature vector (limit to 300-700 most frequent words if the dictionary is too large)

• There's a lot of total price bounding boxes, only consider the last one to the bottom to be the total price :D

• Access training notebook here to train the model again

• [x] Write report for KV (I have school deadline on 28/03) @patrickphatnguyen

[patrickphat]

New GCN notebook here

• [x] implement accuracy and f1-score (prioritized) (consider all classes other than "other" positive")@0902338471
• [ ] Try weights for labels @0902338471
• [x] date token to "ddate" instead of "ddate" @0902338471
• [x] time token to "ttime" instead of "time" @0902338471
• [x] money token to "mmoney" instead of "money" @0902338471
• [ ] try node_labels 1st version (Without taking last one to be the price) @0902338471
• [x] implement GNN Explainer @patrickphatnguyen
• [x] implement interactive explainer @patrickphatnguyen

[patrickphat]

31/3

• [x] Make slides @0902338471 @patrickphatnguyen

• Access slides here

• Read slides layout here

• [x] Implement interactive explainer @patrickphatnguyen

[patrickphat]

1/4

• [x] Write KV module documentation here on how to load data and train @0902338471
• [x] Write Explainer module documentation here @patrickphatnguyen
• [x] Write KV + Explainer report here @patrickphatnguyen
• [x] Write Full-flow flax evaluation on report here @0902338471

[patrickphat]

Xong nhiệm vụ, issue closed