STEP_SIZE_TRAIN = train_generator.__len__()
STEP_SIZE_VALID = validation_generator.__len__()
# Set new model name.
model_name = 'headpose' + str(int(time.time()))
model_path = model_dir + model_name + '.h5'
loss_csv = model_dir + model_name + '_loss.csv'
# Configure a callback for logging train progress in a .csv file.
csv_logger = CSVLogger(loss_csv)
# Configure estimator model from architecture parameters set before.
model = mpatacchiola_generic(in_size, num_conv_blocks, num_filters_start,
num_dense_layers, dense_layer_size, dropout_rate)
model.compile(loss='mean_squared_error', optimizer='adam', metrics=['mae'])
# Train the configured model on the train generator.
history = model.fit_generator(generator=train_generator,
steps_per_epoch=STEP_SIZE_TRAIN,
validation_data=validation_generator,
validation_steps=STEP_SIZE_VALID,
epochs=epochs,
callbacks=[reduce_lr, stop, csv_logger],
verbose=verbose)
# Get score for the dataset (tilt, pan and global error).
pred = model.predict((test_array / 255.0 - mean) / std)
mean = 0.408808
std = 0.237583
t_mean = -0.041212
t_std = 0.323931
p_mean = -0.000276
p_std = 0.540958
# Models.
head_detector = ssd_512(image_size=(in_size_detector, in_size_detector, 3), n_classes=1, min_scale=0.1, max_scale=1, mode='inference')
head_detector.load_weights(detector_path)
pose_estimator = mpatacchiola_generic(in_size_estimator, num_conv_blocks, num_filters_start, num_dense_layers, dense_layer_size)
pose_estimator.load_weights(estimator_path)
# Get video source.
video_source = input("Input stream: ")
# If video source is a device convert the string identifying it to integer.
try:
video_source = int(video_source)
except:
pass
# Initialize cam.