def main():
print(tf.__version__)
cp_callback_coarse = tf.keras.callbacks.ModelCheckpoint(
filepath=COARSE_CHECKPOINT_PATH, save_weights_only=True, verbose=1)
cp_callback_refine = tf.keras.callbacks.ModelCheckpoint(
filepath=REFINED_CHECKPOINT_PATH, save_weights_only=True, verbose=1)
csv_logger = CSVLogger('log.csv', append=False, separator=',')
nyu_data_generator = NyuDepthGenerator(batch_size=10,
csv_path='data/train.csv')
# eval_data_generator = NyuDepthGenerator(batch_size=1, csv_path='data/dev.csv')
latest_checkpoint_refine = tf.train.latest_checkpoint(
REFINED_CHECKPOINT_DIR)
latest_checkpoint_coarse = tf.train.latest_checkpoint(
COARSE_CHECKPOINT_DIR)
if RUN_REFINE:
refine_model, coarse_model = models.refined_network_model()
model = refine_model
if latest_checkpoint_refine:
print("\nRestored refine model from checkpoint")
refine_model.load_weights(latest_checkpoint_refine)
elif latest_checkpoint_coarse:
print("\nRestored coarse model from checkpoint")
coarse_model.load_weights(latest_checkpoint_coarse)
else:
print(
"\nCoarse model not restored. Please exit and run coarse model first"
)
print("\nStarting one pass training")
else:
coarse_model, _, _ = models.coarse_network_model()
model = coarse_model
if latest_checkpoint_coarse:
print("\nRestored coarse model from checkpoint")
coarse_model.load_weights(latest_checkpoint_coarse)
else:
print("\nNo coarse checkpoint saved")
model.compile(
optimizer=keras.optimizers.Adam(), # Optimizer
# Loss function to minimize
loss=models.depth_loss,
# List of metrics to monitor
metrics=[
metrics.rmse, metrics.abs_relative_diff,
metrics.squared_relative_diff
])
predict_while_train = PredictWhileTrain(nyu_data_generator)
if not os.path.isdir(PREDICT_FILE_PATH):
os.mkdir(TRAIN_PREDICT_FILE_PATH)
print('Fit model on training data')
if RUN_REFINE:
history = model.fit(
x=nyu_data_generator,
epochs=30,
callbacks=[cp_callback_refine, csv_logger, predict_while_train])
else:
history = model.fit(
x=nyu_data_generator,
epochs=30,
callbacks=[cp_callback_coarse, csv_logger, predict_while_train])
print('\nHistory dict:', history.history)
result = model.evaluate(x=nyu_data_generator, steps=1)
print("Final eval loss: ", result)
if not os.path.isdir(PREDICT_FILE_PATH):
os.mkdir(PREDICT_FILE_PATH)
predictions = model.predict(x=nyu_data_generator, steps=1)
print("Prediction dim: " + str(predictions.shape))
for i in range(predictions.shape[0]):
predictions[i] = (predictions[i] / np.max(predictions[i])) * 255.0
image_name = os.path.join(PREDICT_FILE_PATH, '%05d_d.png' % i)
image_im = Image.fromarray(np.uint8(predictions[i].reshape(
TARGET_HEIGHT, TARGET_WIDTH)),
mode="L")
image_im.save(image_name)
def main():
print(tf.__version__)
x_train = []
y_train = []
x_eval = []
y_eval = []
h5file = h5py.File(NYU_FILE_PATH, 'r')
file_count = h5file['images'].shape[0]
dev_split = 0.9
train_count = file_count * dev_split
for i in range(file_count):
if i % 10 == 0:
print("processing file " + str(i))
image = np.transpose(h5file['images'][i], (2, 1, 0))
depth = np.transpose(h5file['depths'][i], (1, 0))
image_im = Image.fromarray(np.uint8(image))
image_im = image_im.resize((IMAGE_WIDTH, IMAGE_HEIGHT))
image_np_arr = np.array(image_im)
# print ("image_np_arr shape: " + str(image_np_arr.shape))
depth_scaled = (depth / 10.0) * 255.0
depth_im = Image.fromarray(np.uint8(depth_scaled))
depth_im = depth_im.resize((TARGET_WIDTH, TARGET_HEIGHT))
depth_np_arr = np.array(depth_im)
depth_np_arr = depth_np_arr / 255.0 * 10.0
# print ("depth_np_arr shape: " + str(depth_np_arr.shape))
# print ("depth_np_arr: " + str(depth_np_arr))
if i < train_count:
x_train.append(image_np_arr)
y_train.append(depth_np_arr)
else:
x_eval.append(image_np_arr)
y_eval.append(depth_np_arr)
x_train = np.array(x_train) / 255.0
y_train = np.array(y_train)
x_eval = np.array(x_eval) / 255.0
y_eval = np.array(y_eval)
print(x_train.shape)
print(y_train.shape)
print(x_eval.shape)
print(y_eval.shape)
cp_callback_coarse = tf.keras.callbacks.ModelCheckpoint(
filepath=COARSE_CHECKPOINT_PATH,
save_weights_only=True,
verbose=1,
period=10)
cp_callback_refine = tf.keras.callbacks.ModelCheckpoint(
filepath=REFINED_CHECKPOINT_PATH,
save_weights_only=True,
verbose=1,
period=10)
csv_logger = CSVLogger('log.csv', append=False, separator=',')
latest_checkpoint_refine = tf.train.latest_checkpoint(
REFINED_CHECKPOINT_DIR)
latest_checkpoint_coarse = tf.train.latest_checkpoint(
COARSE_CHECKPOINT_DIR)
if RUN_REFINE:
refine_model, coarse_model = models.refined_network_model()
model = refine_model
if latest_checkpoint_refine:
print("\nRestored refine model from checkpoint")
refine_model.load_weights(latest_checkpoint_refine)
elif latest_checkpoint_coarse:
print("\nRestored coarse model from checkpoint")
coarse_model.load_weights(latest_checkpoint_coarse)
else:
print(
"\nCoarse model not restored. Please exit and run coarse model first"
)
print("\nStarting one pass training")
else:
coarse_model, _, _ = models.coarse_network_model()
model = coarse_model
if latest_checkpoint_coarse:
print("\nRestored coarse model from checkpoint")
coarse_model.load_weights(latest_checkpoint_coarse)
else:
print("\nNo coarse checkpoint saved")
model.compile(
optimizer=keras.optimizers.Adam(), # Optimizer
# Loss function to minimize
loss=models.depth_loss_2,
metrics=[
metrics.abs_relative_diff, metrics.squared_relative_diff,
metrics.rmse, metrics.rmse_log, metrics.rmse_scale_invariance_log
])
predict_while_train = PredictWhileTrain(x_train)
early_stop = tf.keras.callbacks.EarlyStopping(monitor='val_loss',
patience=5)
if not os.path.isdir(TRAIN_PREDICT_FILE_PATH):
os.mkdir(TRAIN_PREDICT_FILE_PATH)
print('Fit model on training data')
if RUN_REFINE:
history = model.fit(x=x_train,
y=y_train,
validation_data=(x_eval, y_eval),
epochs=300,
callbacks=[cp_callback_refine, csv_logger])
else:
history = model.fit(x=x_train,
y=y_train,
validation_data=(x_eval, y_eval),
epochs=300,
callbacks=[cp_callback_coarse, csv_logger])
print('\nHistory dict:', history.history)
result = model.evaluate(x=x_eval, y=y_eval)
print("Final eval loss on validation: ", result)
def main():
print(tf.__version__)
cp_callback_coarse = tf.keras.callbacks.ModelCheckpoint(
filepath=COARSE_CHECKPOINT_PATH,
save_weights_only=True,
verbose=1,
period=10)
cp_callback_refine = tf.keras.callbacks.ModelCheckpoint(
filepath=REFINED_CHECKPOINT_PATH,
save_weights_only=True,
verbose=1,
period=10)
csv_logger = CSVLogger('log.csv', append=False, separator=',')
x_train, y_train = csv_inputs(csv_file_path='data/train.csv')
x_eval, y_eval = csv_inputs(csv_file_path='data/dev.csv')
latest_checkpoint_refine = tf.train.latest_checkpoint(
REFINED_CHECKPOINT_DIR)
latest_checkpoint_coarse = tf.train.latest_checkpoint(
COARSE_CHECKPOINT_DIR)
if RUN_REFINE:
refine_model, coarse_model = models.refined_network_model()
model = refine_model
if latest_checkpoint_refine:
print("\nRestored refine model from checkpoint")
refine_model.load_weights(latest_checkpoint_refine)
elif latest_checkpoint_coarse:
print("\nRestored coarse model from checkpoint")
coarse_model.load_weights(latest_checkpoint_coarse)
else:
print(
"\nCoarse model not restored. Please exit and run coarse model first"
)
print("\nStarting one pass training")
else:
coarse_model, _, _ = models.coarse_network_model()
model = coarse_model
if latest_checkpoint_coarse:
print("\nRestored coarse model from checkpoint")
coarse_model.load_weights(latest_checkpoint_coarse)
else:
print("\nNo coarse checkpoint saved")
model.compile(
optimizer=keras.optimizers.Adam(), # Optimizer
# Loss function to minimize
loss=models.rmse_scale_invariance_log_loss,
metrics=[
metrics.abs_relative_diff, metrics.squared_relative_diff,
metrics.rmse, metrics.rmse_log, metrics.rmse_scale_invariance_log
])
predict_while_train = PredictWhileTrain(x_train)
early_stop = tf.keras.callbacks.EarlyStopping(monitor='val_loss',
patience=5)
if not os.path.isdir(TRAIN_PREDICT_FILE_PATH):
os.mkdir(TRAIN_PREDICT_FILE_PATH)
print('Fit model on training data')
if RUN_REFINE:
history = model.fit(
x=x_train,
y=y_train,
validation_data=(x_eval, y_eval),
epochs=2005,
callbacks=[cp_callback_refine, csv_logger, predict_while_train])
else:
history = model.fit(
x=x_train,
y=y_train,
validation_data=(x_eval, y_eval),
epochs=2005,
callbacks=[cp_callback_coarse, csv_logger, predict_while_train])
print('\nHistory dict:', history.history)
result = model.evaluate(x=x_eval, y=y_eval)
print("Final eval loss on validation: ", result)
def main():
print(tf.__version__)
latest_checkpoint_refine = tf.train.latest_checkpoint(
REFINED_CHECKPOINT_DIR)
latest_checkpoint_coarse = tf.train.latest_checkpoint(
COARSE_CHECKPOINT_DIR)
if RUN_REFINE:
refine_model, coarse_model = models.refined_network_model()
model = refine_model
if latest_checkpoint_refine:
print("\nRestored refine model from checkpoint")
refine_model.load_weights(latest_checkpoint_refine)
elif latest_checkpoint_coarse:
print("\nRestored coarse model from checkpoint")
coarse_model.load_weights(latest_checkpoint_coarse)
else:
print(
"\nCoarse model not restored. Please exit and run coarse model first"
)
print("\nStarting one pass training")
else:
coarse_model, _, _ = models.coarse_network_model()
model = coarse_model
if latest_checkpoint_coarse:
print("\nRestored coarse model from checkpoint")
coarse_model.load_weights(latest_checkpoint_coarse)
else:
print("\nNo coarse checkpoint saved")
model.compile(
optimizer=keras.optimizers.Adam(), # Optimizer
# Loss function to minimize
loss=models.depth_loss,
# List of metrics to monitor
metrics=None)
x_eval, y_eval = csv_inputs(csv_file_path='data/dev.csv')
result = model.evaluate(x=x_eval, y=y_eval)
print("Final eval loss on validation: ", result)
if not os.path.isdir(PREDICT_FILE_PATH):
os.mkdir(PREDICT_FILE_PATH)
predictions = model.predict(x=x_eval)
print("Prediction dim: " + str(predictions.shape))
for i in range(predictions.shape[0]):
predictions[i] = (predictions[i] / np.max(predictions[i])) * 255.0
prediction_name = os.path.join(PREDICT_FILE_PATH,
'%05d_predict.png' % i)
prediction_im = Image.fromarray(
np.uint8(predictions[i].reshape(TARGET_HEIGHT, TARGET_WIDTH)))
prediction_im.save(prediction_name)
color_name = os.path.join(PREDICT_FILE_PATH, '%05d_c.png' % i)
color_im = Image.fromarray(np.uint8(x_eval[i] * 255.0))
color_im.save(color_name)
depth_name = os.path.join(PREDICT_FILE_PATH, '%05d_d.png' % i)
depth_im = Image.fromarray(np.uint8(y_eval[i] * 255.0))
depth_im.save(depth_name)
def main():
print(tf.__version__)
x_eval = []
y_eval = []
h5file = h5py.File(NYU_FILE_PATH, 'r')
file_count = h5file['images'].shape[0]
# file_count = 100
dev_split = 0.9
train_count = file_count * dev_split
for i in range(file_count):
if i < train_count:
continue
else:
print("processing image " + str(i))
image = np.transpose(h5file['images'][i], (2, 1, 0))
depth = np.transpose(h5file['depths'][i], (1, 0))
image_im = Image.fromarray(np.uint8(image))
image_im = image_im.resize((IMAGE_WIDTH, IMAGE_HEIGHT))
image_np_arr = np.array(image_im)
# print ("image_np_arr shape: " + str(image_np_arr.shape))
depth_scaled = (depth / 10.0) * 255.0
depth_im = Image.fromarray(np.uint8(depth_scaled))
depth_im = depth_im.resize((TARGET_WIDTH, TARGET_HEIGHT))
depth_np_arr = np.array(depth_im)
depth_np_arr = depth_np_arr / 255.0 * 10.0
# print ("depth_np_arr shape: " + str(depth_np_arr.shape))
# print ("depth_np_arr: " + str(depth_np_arr))
x_eval.append(image_np_arr)
y_eval.append(depth_np_arr)
x_eval = np.array(x_eval) / 255.0
y_eval = np.array(y_eval)
latest_checkpoint_refine = tf.train.latest_checkpoint(
REFINED_CHECKPOINT_DIR)
latest_checkpoint_coarse = tf.train.latest_checkpoint(
COARSE_CHECKPOINT_DIR)
if RUN_REFINE:
refine_model, coarse_model = models.refined_network_model()
model = refine_model
if latest_checkpoint_refine:
print("\nRestored refine model from checkpoint")
refine_model.load_weights(latest_checkpoint_refine)
elif latest_checkpoint_coarse:
print("\nRestored coarse model from checkpoint")
coarse_model.load_weights(latest_checkpoint_coarse)
else:
print(
"\nCoarse model not restored. Please exit and run coarse model first"
)
print("\nStarting one pass training")
else:
coarse_model, _, _ = models.coarse_network_model()
model = coarse_model
if latest_checkpoint_coarse:
print("\nRestored coarse model from checkpoint")
coarse_model.load_weights(latest_checkpoint_coarse)
else:
print("\nNo coarse checkpoint saved")
model.compile(
optimizer=keras.optimizers.Adam(), # Optimizer
# Loss function to minimize
loss=models.depth_loss_2,
# List of metrics to monitor
# metrics= [metrics.rmse])
metrics=[
metrics.abs_relative_diff, metrics.squared_relative_diff,
metrics.rmse, metrics.rmse_log, metrics.rmse_scale_invariance_log
])
result = model.evaluate(x=x_eval, y=y_eval, batch_size=32)
print("Final eval loss on validation: ", result)
# print("truth: " + str(y_eval))
# return
if not os.path.isdir(PREDICT_FILE_PATH):
os.mkdir(PREDICT_FILE_PATH)
predictions = model.predict(x=x_eval)
print("Prediction dim: " + str(predictions.shape))
for i in range(predictions.shape[0]):
# print("prediction before: " + str(predictions[i]))
predictions[i] = (predictions[i] / 10.0) * 255.0
# print("prediction after: " + str(predictions[i]))
prediction_name = os.path.join(PREDICT_FILE_PATH,
'%05d_predict.png' % i)
prediction_im = Image.fromarray(
np.uint8(predictions[i].reshape(TARGET_HEIGHT, TARGET_WIDTH)))
prediction_im.save(prediction_name)
color_name = os.path.join(PREDICT_FILE_PATH, '%05d_c.png' % i)
color_im = Image.fromarray(np.uint8(x_eval[i] * 255.0))
color_im.save(color_name)
depth_name = os.path.join(PREDICT_FILE_PATH, '%05d_d.png' % i)
depth_im = Image.fromarray(np.uint8((y_eval[i] / 10.0) * 255.0))
depth_im.save(depth_name)