all_B_over_U = data['all_B_over_U']
print 'Done'
N = 500
random_hyperparams = {
'count__ngram_range': [(1, n) for n in npr.randint(1, 4, N)],
'count__max_df': npr.uniform(.75, 1.0, N),
'tfidf__norm': npr.choice(['l1', 'l2', None], N),
'tfidf__use_idf': npr.choice([True, False], N),
'svc__base_estimator__C': npr.choice(np.logspace(-3, 4, 10000), N)
}
score_kwds = {
'xs': train_data['text'],
'ys': train_data['is_multiple'],
'bs': train_data['is_biased'],
'all_B_over_U': all_B_over_U,
'fit_weight_kwd': 'svc__sample_weight',
'n_cv_splits': 5,
'random_seed': random_seed
}
print 'Starting Experiments...'
t0 = time()
experiments = random_search(model, random_hyperparams,
'best_svm_mult_silver.pkl', **score_kwds)
print 'Done {}:{}:{} seconds. Writing out experiments'.format(*hms(time() -
t0))
joblib.dump(experiments, 'svm_mult_silver_dev.pkl')
print 'All done'
test_summaries_writer.add_summary(
u.explicit_summaries({"xent": test_loss}), step)
test_summaries_writer.add_summary(u.explicit_summaries(tag_values), step)
for idx, img in enumerate(stats['debug_imgs']):
debug_img_summary = u.pil_image_to_tf_summary(img,
tag="debug_img_%d" % idx)
test_summaries_writer.add_summary(debug_img_summary, step)
test_summaries_writer.flush()
# report one liner
log = []
log.append("step %d/%d" % (step, opts.steps))
log.append("time %d" % int(time.time() - start_time))
log.append("train_loss %f" % train_loss)
log.append("test_loss %s" % test_loss)
log.append("test stats { p:%0.2f, r:%0.2f, f1:%0.2f }" %
tuple([stats[k] for k in ['precision', 'recall', 'f1']]))
print("\t".join(log))
# check if done by steps or time
step += 1 # TODO: fetch global_step from keras model (?)
if step >= opts.steps:
done = True
if opts.secs is not None:
run_time = time.time() - start_time
remaining_time = opts.secs - run_time
print("run_time %s remaining_time %s" %
(u.hms(run_time), u.hms(remaining_time)))
if remaining_time < 0:
done = True
