ya_train[tbr_train] = np.take(ya_train, good_candidates_train)
xa_train[tbr_train, ] = np.take(xa_train, good_candidates_train, 0)
# Accelerometer and silhouette training vectors are the same, they're both
# calories. They originally contain different NaN, but once these are replaced
# only one training vector is needed
y_train = ya_train
# Remove Nan data from test
nan_test = np.logical_or(np.isnan(ya_test), np.isnan(yv_test))
xv_test = np.delete(xv_test, np.where(nan_test), 0)
xa_test = np.delete(xa_test, np.where(nan_test), 0)
y_test = np.delete(yv_test, np.where(nan_test))
# Network
model = NetworkCombined(img_rows, img_cols, Nv_chan, acc_buffersiz, Na_chan)
model.summary()
opt = keras.optimizers.RMSprop() #(lr=3e-3)
model.compile(optimizer=opt, loss='mean_squared_error', metrics=['accuracy'])
best_model_name = '%s_best.h5' % network_name
# Callbacks
filepath = os.path.join(save_dir, best_model_name)
save_history = SaveHistory()
save_history.SetFilename(network_name, save_dir)
if Set['bluecrystal']:
callbacks = [save_history]
nan_test = np.logical_or(np.isnan(ya_test), np.isnan(yv_test))
xv_test = np.delete(xv_test, np.where(nan_test), 0)
xa_test = np.delete(xa_test, np.where(nan_test), 0)
y_test = np.delete(yv_test, np.where(nan_test))
lab_test = np.delete(lab_test, np.where(nan_test))
# Initialize the network structure
if si == 1:
# Initialise error per label
Nlabels = 11
R_per_lab = np.zeros(Nlabels)
err_per_lab = np.zeros(Nlabels)
N_per_lab = np.zeros(Nlabels)
# Network
model = NetworkCombined(img_rows, img_cols, Nv_chan, acc_buffersiz,
Na_chan)
# Read the model weights
file_name = os.path.join(model_path,
model_name % (buffer_tested[bi], si))
if os.path.isfile(file_name):
print('Loading model for leave %s out' % file_name)
model.load_weights(file_name)
else:
print(
'FILE NOT FOUND! PREDICTING WITH RANDOM WEIGHTS! **********************************************'
)
# Predict the results
print('Predicting the result...')