Feature: Support for Seq2Seq (LSTM) model for next word prediction

[chekoduadarsh]

trafficstars

Issue

I would like to take up the task to implement Seq2Seq models on deep_AutoViML. This will allow this library to perform operations like next word prediction, Text summarization etc.

Proposed approach

keras_model_type = "next_word_prediction" 
deepauto.fit(train_datafile, target, keras_model_type=keras_model_type,
		project_name=project_name, keras_options=keras_options, model_options=model_options, 
		save_model_flag=False, verbose=1)

We can use keras_model_type in deep_autoviml.py to check for the string next word prediction, here the data will be preprocessed and an appropriate model will be chosen. After this chosen model will be trained for the given data. Users can either enter or use the default early stopping, epochs and other features.

if keras_model_type.lower() in ['image', 'images', "image_classification"]:
   # Train Image classification

elif keras_model_type.lower() in ['text classification', "text_classification"]:
   # Train for Text classification

elif keras_model_type.lower() in ['next word prediction', "next_word_prediction"]:
   # Train for next word prediction

Similarly, We can create a model for time series prediction.

@AutoViML: If you have a better approach to solving this problem let me know

[AutoViML]

Hi @chekoduadarsh 👍 I have added you as a collaborator to this repository - once you accept the invite I will be able to assign a project I have created in this repo. You can see it here. https://github.com/users/AutoViML/projects/2/views/1

[chekoduadarsh]

@AutoViML Thank you, I have accepted your offer Soon I will make PR request on the above feature. Thank you for this opportunity.

[AutoViML]

Hi @chekoduadarsh 👍 Any updates? AutoViML

[chekoduadarsh]

Hello @AutoViML ,

Sorry for the delayed response. (I was busy with medical and personal things)

please have a look into the forked repository https://github.com/chekoduadarsh/deep_autoviml

Timeseries prediction -> Status update

I have implemented LSTM, GRU and RNN Models for time series prediction and I have tested it with energydata_complete from https://archive.ics.uci.edu

• [x] Make LSTM, GRU and RNN Models
• [x] Use TimeseriesGenerator and parse test/train data
• [x] Successfully train a model and make a example notebook
• [x] Comment the code and mention the source

ref: https://github.com/srivatsan88/End-to-End-Time-Series/blob/master/Multivariate_Time_Series_Modeling_using_LSTM.ipynb ref: https://www.youtube.com/watch?v=i4vGKgbtf1U&list=PL3N9eeOlCrP5cK0QRQxeJd6GrQvhAtpBK&index=12

Please review the changes and let me know if you want me to do something before i make the pull request. :pray:

Thank you Adarsh C

[AutoViML]

Hi @chekoduadarsh 👍 I will test the code with a few classic time series datasets and then approve. thanks for the quick turnaround! AutoViML

[AutoViML]

Hi @chekoduadarsh : can you create a Pull Request for my github?? don't have much time and would appreciate the help, Thanks AutoVimal

[chekoduadarsh]

Hello @AutoViML

I will create a PR! But I don't think the solution is complete! I am sorry but I couldn't finalize it I think

[AutoViML]

Ok let me see if I can work on it. Closing this for now. Thank you for all your help. Auto Vimal