Encoder-Decoder Analysis

Model Architecture



In [ ]:

    
report_file = 'reports/encdec_200_512_2.json'
log_file = 'logs/encdec_200_512_logs.json'

import json
import matplotlib.pyplot as plt
with open(report_file) as f:
    report = json.loads(f.read())
with open(log_file) as f:
    logs = json.loads(f.read())
print'Encoder: \n\n', report['architecture']['encoder']
print'Decoder: \n\n', report['architecture']['decoder']

Perplexity on Each Dataset



In [ ]:

    
print('Train Perplexity: ', report['train_perplexity'])
print('Valid Perplexity: ', report['valid_perplexity'])
print('Test Perplexity: ', report['test_perplexity'])

Loss vs. Epoch



In [ ]:

    
%matplotlib inline
for k in logs.keys():
    plt.plot(logs[k][0], logs[k][1], label=str(k) + ' (train)')
    plt.plot(logs[k][0], logs[k][2], label=str(k) + ' (valid)')
plt.title('Loss v. Epoch')
plt.xlabel('Epoch')
plt.ylabel('Loss')
plt.legend()
plt.show()

Perplexity vs. Epoch



In [ ]:

    
%matplotlib inline
for k in logs.keys():
    plt.plot(logs[k][0], logs[k][3], label=str(k) + ' (train)')
    plt.plot(logs[k][0], logs[k][4], label=str(k) + ' (valid)')
plt.title('Perplexity v. Epoch')
plt.xlabel('Epoch')
plt.ylabel('Perplexity')
plt.legend()
plt.show()

Generations



In [ ]:

    
def print_sample(sample, best_bleu=None):
    enc_input = ' '.join([w for w in sample['encoder_input'].split(' ') if w != '<pad>'])
    gold = ' '.join([w for w in sample['gold'].split(' ') if w != '<mask>'])
    print('Input: '+ enc_input + '\n')
    print('Gend: ' + sample['generated'] + '\n')
    print('True: ' + gold + '\n')
    if best_bleu is not None:
        cbm = ' '.join([w for w in best_bleu['best_match'].split(' ') if w != '<mask>'])
        print('Closest BLEU Match: ' + cbm + '\n')
        print('Closest BLEU Score: ' + str(best_bleu['best_score']) + '\n')
    print('\n')



In [ ]:

    
for i, sample in enumerate(report['train_samples']):
    print_sample(sample, report['best_bleu_matches_train'][i] if 'best_bleu_matches_train' in report else None)



In [ ]:

    
for i, sample in enumerate(report['valid_samples']):
    print_sample(sample, report['best_bleu_matches_valid'][i] if 'best_bleu_matches_valid' in report else None)



In [ ]:

    
for i, sample in enumerate(report['test_samples']):
    print_sample(sample, report['best_bleu_matches_test'][i] if 'best_bleu_matches_test' in report else None)

BLEU Analysis



In [1]:

    
def print_bleu(blue_struct):
    print 'Overall Score: ', blue_struct['score'], '\n'
    print '1-gram Score: ', blue_struct['components']['1']
    print '2-gram Score: ', blue_struct['components']['2']
    print '3-gram Score: ', blue_struct['components']['3']
    print '4-gram Score: ', blue_struct['components']['4']



In [ ]:

    
# Training Set BLEU Scores
print_bleu(report['train_bleu'])



In [ ]:

    
# Validation Set BLEU Scores
print_bleu(report['valid_bleu'])



In [ ]:

    
# Test Set BLEU Scores
print_bleu(report['test_bleu'])



In [ ]:

    
# All Data BLEU Scores
print_bleu(report['combined_bleu'])

N-pairs BLEU Analysis

This analysis randomly samples 1000 pairs of generations/ground truths and treats them as translations, giving their BLEU score. We can expect very low scores in the ground truth and high scores can expose hyper-common generations



In [ ]:

    
# Training Set BLEU n-pairs Scores
print_bleu(report['n_pairs_bleu_train'])



In [ ]:

    
# Validation Set n-pairs BLEU Scores
print_bleu(report['n_pairs_bleu_valid'])



In [ ]:

    
# Test Set n-pairs BLEU Scores
print_bleu(report['n_pairs_bleu_test'])



In [ ]:

    
# Combined n-pairs BLEU Scores
print_bleu(report['n_pairs_bleu_all'])



In [ ]:

    
# Ground Truth n-pairs BLEU Scores
print_bleu(report['n_pairs_bleu_gold'])

Alignment Analysis

This analysis computs the average Smith-Waterman alignment score for generations, with the same intuition as N-pairs BLEU, in that we expect low scores in the ground truth and hyper-common generations to raise the scores



In [ ]:

    
print 'Average (Train) Generated Score: ', report['average_alignment_train']
print 'Average (Valid) Generated Score: ', report['average_alignment_valid']
print 'Average (Test) Generated Score: ', report['average_alignment_test']
print 'Average (All) Generated Score: ', report['average_alignment_all']
print 'Average Gold Score: ', report['average_alignment_gold']