evaluate_speaker_split.sh

#!/bin/bash
# this is the script for final evaluation

# the architecture was chosen in architecture exploration
# the hyperparam are chosen in hyperparameter tuning

set -e

export OUTPUT_FOLDER=speakers_split/results/
# chosen architecture: three stages with attention
export THREE_STAGES=true_highway
export ATTENTION=both


# this time test on folds 1-4 and test on 5
export MODE=eval

export BATCH_SIZE=2

# the chosen hyperparams
export LABEL_EMB_SIZE=64
export LSTM_SIZE=128

DATASET="huric/speakers_split/en_nation/true" make train_joint
DATASET="huric/speakers_split/en_nation/false" make train_joint
DATASET="huric/speakers_split/native/solid english speaker" make train_joint
DATASET="huric/speakers_split/native/weak english speaker" make train_joint
DATASET="huric/speakers_split/native/yes" make train_joint
DATASET="huric/speakers_split/proficiency/yes" make train_joint
DATASET="huric/speakers_split/proficiency/no" make train_joint

# evaluation steps
python nlunetwork/results_aggregator.py nlunetwork/results/speakers_split/results/eval_loss_both_slottype_full_we_large_recurrent_cell_lstm_attention_both_three_stages_true_highway___hyper:LABEL_EMB_SIZE=64,LSTM_SIZE=128,BATCH_SIZE=2,MAX_EPOCHS=100/huric/speakers_split/en_nation ''
python nlunetwork/results_aggregator.py nlunetwork/results/speakers_split/results/eval_loss_both_slottype_full_we_large_recurrent_cell_lstm_attention_both_three_stages_true_highway___hyper:LABEL_EMB_SIZE=64,LSTM_SIZE=128,BATCH_SIZE=2,MAX_EPOCHS=100/huric/speakers_split/native ''
python nlunetwork/results_aggregator.py nlunetwork/results/speakers_split/results/eval_loss_both_slottype_full_we_large_recurrent_cell_lstm_attention_both_three_stages_true_highway___hyper:LABEL_EMB_SIZE=64,LSTM_SIZE=128,BATCH_SIZE=2,MAX_EPOCHS=100/huric/speakers_split/proficiency ''