def main(num_epochs=NUM_EPOCHS):
print("Loading data...")
load_path = '/idiap/user/dwu/spyder/Codalab_Age_estimation/'
fname = load_path+"data_with_label_DPM_Color_NoCROP.hdf5"
X, y, y_mean, y_std, y_train_variance, X_mean, X_std = load(fname)
dataloader = DataLoader(X, y, validation_num=VALIDATIONNUM, batch_size=BATCH_SIZE)
if True:
print("Building model and compiling functions...")
output_layer = build_model()
x_ = _shared(np.empty((BATCH_SIZE,3,96,96)))
y_ = _shared(np.empty((BATCH_SIZE,1)))
iter_funcs = create_iter_functions(
x_, y_,
output_layer,
X_tensor_type=T.tensor4,
learning_rate=LEARNING_RATE
)
if True:
print("Loading the trained parameters")
params_all = pickle.load(open('Lasagne_age_regressor_E200.pkl', 'rb'))
lasagne.layers.set_all_param_values(output_layer, params_all)
print("Starting training...")
now = time.time()
valid_loss = np.inf
try:
for epoch in train(iter_funcs, dataloader, x_, y_):
print("Epoch {} of {} took {:.3f}s".format(
epoch['number'], num_epochs, time.time() - now))
now = time.time()
print(" training loss:\t\t{:.6f}".format(epoch['train_loss']))
print(" validation loss:\t\t{:.6f}".format(epoch['valid_loss']))
if epoch['number'] >= num_epochs:
break
# we also save the best validation parameters:
if epoch['valid_loss'] < valid_loss:
valid_loss = epoch['valid_loss']
print "saving best validation parameters with best parameter epoch: "+str(epoch['number'])
all_params = lasagne.layers.get_all_param_values(output_layer)
saved_name = 'Lasagne_age_regressor_best.pkl'
with(open(saved_name, 'wb')) as file:
pickle.dump(all_params, file)
except KeyboardInterrupt:
all_params = lasagne.layers.get_all_param_values(output_layer)
with(open('Lasagne_age_regressor.pkl', 'wb')) as file:
pickle.dump(all_params, file)
all_params = lasagne.layers.get_all_param_values(output_layer)
with(open('Lasagne_age_regressor.pkl', 'wb')) as file:
pickle.dump(all_params, file)
print "best parameter epoch: "+str(epoch['number'])
return output_layer
def main():
print("Loading data...")
load_path = '/idiap/user/dwu/spyder/Codalab_Age_estimation/'
fname = load_path+"data_with_label_DPM_Color_NoCROP.hdf5"
X, y, y_mean, y_std, y_train_variance, X_mean, X_std = load(fname, test=True)
dataloader = DataLoaderTest(X)
print("Building model and compiling functions...")
output_layer = build_model(input_width=96, input_height=96, output_dim=1,
batch_size=BATCH_SIZE, dimshuffle=True)
#output_layer = build_model()
print("Loading the trained parameters")
params_all = pickle.load(open('Lasagne_age_regressor_best.pkl', 'rb'))
lasagne.layers.set_all_param_values(output_layer, params_all)
x_ = _shared(np.empty((BATCH_SIZE,3,96,96)))
iter_funcs = create_test_functions(x_, output_layer, X_tensor_type=T.tensor4)
now = time.time()
predAges = np.empty(shape=(dataloader.X.shape[0],1))
valid_image_list = '/idiap/user/dwu/spyder/Codalab_Age_estimation/valid_matlab_squence.csv'
Valid_image_name = []
with open(valid_image_list, 'r') as valid_list_file:
for image_name in valid_list_file:
#print image_name[:-1]
Valid_image_name.append(image_name[:-1])
image_count = 0
print "start predicting"
for b in range(dataloader.n_iter_test):
dataloader.next_test_batch(x_)
predAge_batch = iter_funcs()
#print predAge_batch.shape
predAges[b*BATCH_SIZE:b*BATCH_SIZE+predAge_batch.shape[0],:] = predAge_batch * y_std + y_mean
if True:
#let's save all the faces
valid_image_dir = '/idiap/user/dwu/spyder/Codalab_Age_estimation/Validation'
valid_save_dir = '/idiap/user/dwu/spyder/Codalab_Age_estimation/Valid_crop'
font = cv2.FONT_HERSHEY_SIMPLEX
for index in range(len(predAge_batch)):
color_image = cv2.imread(os.path.join(valid_image_dir, Valid_image_name[image_count+index]))
cv2.putText(color_image, str(predAge_batch[index] * y_std + y_mean), (0,50), font, 1, (255,0,255))
cv2.imwrite(os.path.join(valid_save_dir, Valid_image_name[image_count+index]), color_image)
image_count += BATCH_SIZE
y_dict = {}
y_dict['y_pred'] = predAges
with open(load_path+'valid_y_pred.pkl','wb') as handle:
pickle.dump(y_dict, handle)
print("Finish predicting {} images, using time {:.3f}s".format(dataloader.X.shape[0], time.time()-now))
def main():
print("Loading data...")
load_path = '/idiap/user/dwu/spyder/Codalab_Age_estimation/'
fname = load_path + "data_with_label_DPM_Color_NoCROP.hdf5"
X, y, y_mean, y_std, y_train_variance, X_mean, X_std = load(fname,
test=True)
dataloader = DataLoaderTest(X)
print("Building model and compiling functions...")
output_layer = build_model(input_width=96,
input_height=96,
output_dim=1,
batch_size=BATCH_SIZE,
dimshuffle=True)
#output_layer = build_model()
print("Loading the trained parameters")
params_all = pickle.load(open('Lasagne_age_regressor_best.pkl', 'rb'))
lasagne.layers.set_all_param_values(output_layer, params_all)
x_ = _shared(np.empty((BATCH_SIZE, 3, 96, 96)))
iter_funcs = create_test_functions(x_,
output_layer,
X_tensor_type=T.tensor4)
now = time.time()
predAges = np.empty(shape=(dataloader.X.shape[0], 1))
valid_image_list = '/idiap/user/dwu/spyder/Codalab_Age_estimation/valid_matlab_squence.csv'
Valid_image_name = []
with open(valid_image_list, 'r') as valid_list_file:
for image_name in valid_list_file:
#print image_name[:-1]
Valid_image_name.append(image_name[:-1])
image_count = 0
print "start predicting"
for b in range(dataloader.n_iter_test):
dataloader.next_test_batch(x_)
predAge_batch = iter_funcs()
#print predAge_batch.shape
predAges[b * BATCH_SIZE:b * BATCH_SIZE +
predAge_batch.shape[0], :] = predAge_batch * y_std + y_mean
if True:
#let's save all the faces
valid_image_dir = '/idiap/user/dwu/spyder/Codalab_Age_estimation/Validation'
valid_save_dir = '/idiap/user/dwu/spyder/Codalab_Age_estimation/Valid_crop'
font = cv2.FONT_HERSHEY_SIMPLEX
for index in range(len(predAge_batch)):
color_image = cv2.imread(
os.path.join(valid_image_dir,
Valid_image_name[image_count + index]))
cv2.putText(color_image,
str(predAge_batch[index] * y_std + y_mean),
(0, 50), font, 1, (255, 0, 255))
cv2.imwrite(
os.path.join(valid_save_dir,
Valid_image_name[image_count + index]),
color_image)
image_count += BATCH_SIZE
y_dict = {}
y_dict['y_pred'] = predAges
with open(load_path + 'valid_y_pred.pkl', 'wb') as handle:
pickle.dump(y_dict, handle)
print("Finish predicting {} images, using time {:.3f}s".format(
dataloader.X.shape[0],
time.time() - now))
