def get_train_valid(state):
datapath = os.path.join(state['data_dir'], state['resolution'])
train_path = datapath + '_train'
valid_path = datapath + '_valid'
x_train, y_train, label_to_idx, _ = load_featurized(train_path)
x_valid, y_valid, label_to_idx, _ = load_featurized(valid_path)
return x_train, y_train, x_valid, y_valid, label_to_idx
def med_vs_lowres(params, n_art):
log = get_medres_vs_lowres_outfile(params['logpath'])
low_xs, low_ys, y_to_idx, ids = load_featurized(params['lowres_path'])
med_xs, med_ys, y_to_idx, ids = load_featurized(params['medres_path'])
n_classes = len(y_to_idx)
assert np.array_equal(low_ys, med_ys)
ys = low_ys
low_model = load_model(params['lowres_id'], params['model_dir'])
med_model = load_model(params['medres_id'], params['model_dir'])
low_preds = get_pred(low_model, low_xs, n_classes)
med_preds = get_pred(med_model, med_xs, n_classes)
low_acc = accuracy_score(ys, low_preds)
med_acc = accuracy_score(ys, med_preds)
p_diff = art_test(ys,
low_preds,
med_preds,
absolute=True,
n=n_art,
scoring=precision_recall_fscore_micro,
return_distribution=False)[2]
r = [
n_art, med_acc, low_acc, med_acc - low_acc, p_diff,
p_diff * params['bonferroni']
]
csv.writer(log).writerow(r)
log.close()
def per_vs_cnn(state, n_art):
log = get_per_vs_cnn_outfile(state['logpath'])
low_xs, low_ys, y_to_idx, ids = load_featurized(state['lowres_path'])
med_xs, med_ys, y_to_idx, ids = load_featurized(state['medres_path'])
n_classes = len(y_to_idx)
assert np.array_equal(low_ys, med_ys)
ys = med_ys
per_medres = load_model(state['medres_id'], state['per_dir'])
per_lowres = load_model(state['lowres_id'], state['per_dir'])
cnn_medres = load_model(state['medres_id'], state['cnn_dir'])
cnn_lowres = load_model(state['lowres_id'], state['cnn_dir'])
preds_per_medres = get_pred(per_medres, med_xs, n_classes)
preds_per_lowres = get_pred(per_lowres, low_xs, n_classes)
preds_cnn_medres = get_pred(cnn_medres, med_xs, n_classes)
preds_cnn_lowres = get_pred(cnn_lowres, low_xs, n_classes)
acc_per_medres = accuracy_score(ys, preds_per_medres)
acc_per_lowres = accuracy_score(ys, preds_per_lowres)
acc_cnn_medres = accuracy_score(ys, preds_cnn_medres)
acc_cnn_lowres = accuracy_score(ys, preds_cnn_lowres, n_classes)
p_diff_ch = art_test(ys,
preds_per_lowres,
preds_cnn_lowres,
absolute=True,
n=n_art,
scoring=precision_recall_fscore_micro,
return_distribution=False)[2]
p_diff_plain = art_test(ys,
preds_per_medres,
preds_cnn_medres,
absolute=True,
n=n_art,
scoring=precision_recall_fscore_micro,
return_distribution=False)[2]
r = [
n_art, 'low_res', acc_cnn_lowres, acc_per_lowres,
acc_cnn_lowres - acc_per_lowres, p_diff_ch,
p_diff_ch * state['bonferroni']
]
csv.writer(log).writerow(r)
r = [
n_art, 'med_res', acc_cnn_medres, acc_per_medres,
acc_cnn_medres - acc_per_medres, p_diff_ch,
p_diff_plain * state['bonferroni']
]
csv.writer(log).writerow(r)
log.close()
def plot_spectrograms(highres_path, medres_path, lowres_path, plot_dir):
high_xs, ys, label_to_idx, x_ids = load_featurized(highres_path)
med_xs, ys, label_to_idx, x_ids = load_featurized(medres_path)
low_xs, ys, label_to_idx, x_ids = load_featurized(lowres_path)
for res, xs in [('highres', high_xs), ('medres', med_xs),
('lowres', low_xs)]:
for i, x, in enumerate(xs[:100]):
plot_path = os.path.join(plot_dir, '%s_%s' % (i, res))
plot_spectrogram(x, plot_path)