if name in ['new_output', 'fc6', 'fc7']:
layer_lr = learning_rate
else:
layer_lr = learning_rate / 10
if name != 'fc8':
layer_updates = lasagne.updates.nesterov_momentum(
loss, layer_params, learning_rate=layer_lr, momentum=0.9)
updates.update(layer_updates)
# Compile functions for training, validation and prediction
train_fn = theano.function([X_sym, y_sym], [loss, acc], updates=updates)
val_fn = theano.function([X_sym, y_sym], [loss, acc])
pred_fn = theano.function([X_sym], prediction)
X_train, y_train, ids_train = load_train_data()
X_test, ids_test = load_test_data(img_start_ix=0, max_img=1000)
y_pseudo = load_pseudo_labels()
pos = np.array(range(4777))
X, y, ids = np.concatenate((X_train, X_test)), np.concatenate(
(y_train, y_pseudo)), np.concatenate((ids_train, ids_test))
X, y, ids, pos = sklearn.utils.shuffle(X, y, ids, pos, random_state=0)
new_pseudo_label_pos = []
for i in range(3777, 4777):
k = np.where(pos == i)
new_pseudo_label_pos.append(int(k[0]))
# generator splitting an iterable into chunks of maximum length N
def batches(iterable, N):
chunk = []
tr_loss_tot /= len(t_ix[:2000])
tr_acc_tot /= len(t_ix[:2000])
val_loss_tot /= len(v_ix)
val_acc_tot /= len(v_ix)
res = [epoch, round(tr_loss_tot,3), round(tr_acc_tot*100,3), round(val_loss_tot,3), round(val_acc_tot * 100,3)]
print('\t'.join(map(str,res)))
#print(epoch, val_loss_tot, val_acc_tot * 100)
# if loss is too high, convergence failure so don't bother to make predictions
if val_loss_tot > 1.0:
continue
print("Predicting with test images...")
imgs_per_batch = 99
for j in range(0, 1001, imgs_per_batch):
X_sub_part, sub_ids_part = load_test_data(j, imgs_per_batch)
y_proba_part = pred_fn(X_sub_part)
if j==0:
y_proba = y_proba_part
ids = sub_ids_part
else:
y_proba = np.append(y_proba, y_proba_part, axis=0)
ids = np.append(ids, sub_ids_part, axis=0)
if j%99==0:
print(j)
make_submission('../output/submission_vgs_' + str(i) + '.csv', y_proba, ids)
