zeeraktalat / mlearn Goto Github PK

Could the encoded data be stored in the object, provided that it is only the indices of the tokens that are stored?

This could potentially save time as indices would not need to be recomputed, only situating it in the tensor would need to be done.

Set up Hyper parameter fees for (hyper-)parameter optimisation for neural nets and sklearn models.

• Optuna
• WandB

GeneralDataset.load_labels is not implemented.

In 2 weeks

Tomorrow

• Write tests for resuable iter

We want to update the prediction writer so that it it writes a file with format

docID, original_Text, model-1 prediction, model-2 prediction, ..., model-n prediction

The head for the models should be the name of the model, not the hyper-parameters (hyper-parameters should be available from file detailing which model is the best).

Ensure a log file is written for each module individually as well as a single log for the entire API

Methods to wrap around sklearn API so that they become a single line call.

When using the tensorisation I created, the ML model tends to not learn anything. Should shift to Cython/C implementation of the tensorisation and have python call the C module.

Add wrapper around the GP library so that single line call is all it requires to run a model

Two different logs:

1. Log that can record performance for each batch
2. Log that can record performance for each epoch

Wednesday 11 AM

In a week

Today

Tomorrow

Add missing tests for modules:

• data_processing/fileio.py
• utils/early_stopping.py
• data_processing/batching.py
• data_processing/loaders.py
• modeling/embedding.py
• modeling/evaluate.py
• modeling/linear.py
• modeling/multitask.py
• modeling/onehot.py
• modeling/train.py
• utils/metrics.py
• utils/pipeline.py

Add tests for missing lines in:

• data_processing/data.py
• data_processing/clean.py

Next week

Error is thrown in loading preprocessor if liwc_dir = None when preprocessor is initialized.

E.g. https://github.com/ZeerakW/mlearn/blob/800472c4c25c8ab6959406b0651c4095143d8adc/mlearn/data/dataset.py#L528

Megan: "can't have lists as default args has to be e.g. None and then at the top of the function: if blah is None: blah = that list"

Add layer over GPy

The LIWC computation takes a subword into account and not just the full word when considering kleene-starred tokens, e.g. abuse*.

#OnlineAbuse triggers abuse*,AFFECT, abuse*,NEGEMOTION etc. It should trigger UNK.

20th of June

Add layer over GPy

Today

Add logging for when exceptions are caught.

In a week

Early stopping will overwrite the best performing model if multiple models are used. E.g. cnn, lstm, mlp are all used then it'll first store the best cnn, then the best lstm, and finally the best mlp in the same file name.

Consider moving away from SpaCy to Ekphrasis

June 13, 2018

17th of June, 2018

14th of June