The purpose of using poetry is to "Freeze" the modules here. If you need to update the dependencies, you're going to have a bad time.

Sorry for the mess that is the training of this model. Anyways, here is the newsetup for the final design change. Architecture: We use conv1d w/ 3 filters and 2 stride to map time as past window, current window, future window stride converts it all into spatiotemporal space. Feed into LSTM

OntoMapAnalysis P = participants p - participant in P

1. Gen train data on participant selected for training
   • Leave out three questions [validation] train/px/validation/gaze_qid.arff
   • remainign questions are [training] train/px/train/gaze_qid.arff
2. Gen test data on participant for testing
   • Preselect (|P| * 0.2) participants for training
   • Leave 3 questions for [retraining] test/px/retrain/retrain/gaze_qid.arff
   • Leave 3 questions for [testing] real world test/test/gaze_qid.arff Python Model Training for each p in P:
   1. Get dir of px
   2. load validation
   3. load train
   4. Train model
3. Save weights
4. Plot over each participant? define evaluation x horizon, what do we want to see change as the model trains?

Python Evaluation 2) Load model with pretrained weights 3) for each p in testP:

• (Reset hidden state), set back to preloaded state
• Train using [retraining] (3 questions)
• Predict the remaining questions
• Output accuracy
• Store in testPids

plot histogram using testPids