def main(argv):
data = get_data(FLAGS.train_data, FLAGS.num_classes)
train_data = data.batch(16, drop_remainder=True)
model = UNet(num_classes=FLAGS.num_classes)
model.compile(optimizer=tf.keras.optimizers.Adam(),
loss='binary_crossentropy')
model.fit(train_data, epochs=25)
for index, (image, label) in enumerate(data.batch(1).take(5)):
prediction = model.predict(image)
plot_result(f'results/{index}.png', image, label, prediction)
lr_base = 0.01 * (float(batch_size) / 16)
input_shape = (320, 320, 3)
with tf.device("/cpu:0"):
model = UNet(input_shape, weight_decay=3e-3, classes=21)
if GPU_COUNT > 1:
model.summary()
from keras.utils.training_utils import multi_gpu_model
model = multi_gpu_model(model, gpus=GPU_COUNT)
model.compile(
# loss=crossentropy_without_ambiguous,
loss=crossentropy_without_ambiguous,
optimizer=Adam(lr=0.001),
# optimizer = SGD(lr=lr_base, momentum=0.9),
metrics=[
'accuracy', categorical_accuracy_without_ambiguous,
categorical_accuracy_only_valid_classes
])
model.summary()
current_dir = os.getcwd()
save_path = os.path.join(current_dir, 'output/unet')
if not os.path.exists(save_path):
os.mkdir(save_path)
checkpoint_path = os.path.join(save_path, 'checkpoint_weights.hdf5')
if RESUME:
print('checkPoint file:{}'.format(checkpoint_path))
def train_model(name, experiment, image_size, training_data_list, training_mask_list,
model_spec=[16, 32, 64, 128, 256], preprocess_list=None,
preprocess_stretch=False, preprocess_mask=None,
preprocess_fisher=False, keep_image=True,
load_model=False, epochs=15):
# make copies of the input array before shuffling
training_data_list = list(training_data_list)
training_mask_list = list(training_mask_list)
random.Random(experiment*42).shuffle(training_data_list)
random.Random(experiment*42).shuffle(training_mask_list)
# we're augmenting data -- expand the list of training data
train_input_img_paths = training_data_list[:-(test_samples
+ val_samples)] * random_factor
train_target_img_paths = training_mask_list[:-(test_samples
+ val_samples)] * random_factor
val_input_img_paths = training_data_list[-(
test_samples + val_samples):-val_samples]
val_target_img_paths = training_mask_list[-(
test_samples + val_samples):-val_samples]
test_input_img_paths = training_data_list[-test_samples:]
test_target_img_paths = training_mask_list[-test_samples:]
pp = None
# Chain of preprocessing functions, first one added is performed first
if preprocess_list is not None:
# Instantiate data Sequences for each split
if not preprocess_stretch:
pp = ImagePreprocessGradient(preprocess_list, keep_image, pp)
else:
pp = ImagePreprocessStretchedGradient(preprocess_list, pp)
if preprocess_mask is not None:
# Apply mask after gradients - masking first only gets overwritten
pp = ImagePreprocessMask(preprocess_mask, pp)
if preprocess_fisher is True:
pp = ImagePreprocessFisherize(pp)
if pp is not None:
# Instantiate pre-processed data sequences for each split
train_gen = RoadSeq(batch_size, image_size,
train_input_img_paths, train_target_img_paths, augment_data=True,
preprocess_fn=pp.preprocess())
val_gen = RoadSeq(batch_size, image_size,
val_input_img_paths, val_target_img_paths, augment_data=False,
preprocess_fn=pp.preprocess())
test_gen = RoadSeq(len(test_input_img_paths), image_size,
test_input_img_paths, test_target_img_paths, augment_data=False,
preprocess_fn=pp.preprocess())
else:
# use the images as they are
train_gen = RoadSeq(batch_size, image_size,
train_input_img_paths, train_target_img_paths, augment_data=True)
val_gen = RoadSeq(batch_size, image_size,
val_input_img_paths, val_target_img_paths, augment_data=False)
test_gen = RoadSeq(len(test_input_img_paths), image_size,
test_input_img_paths, test_target_img_paths, augment_data=False)
model_name = name+'.'+str(experiment)+'.h5'
model = UNet(image_size, model_spec)
model.compile(optimizer="adam",
loss="binary_crossentropy", metrics=["acc"])
if load_model:
model.load_weights(model_name)
model.summary()
callbacks = [
keras.callbacks.ModelCheckpoint(
model_name, save_best_only=True)
]
model.fit(train_gen, epochs=epochs, verbose=1,
validation_data=val_gen, callbacks=callbacks)
x, y = test_gen.__getitem__(0)
start = timer()
results = model.predict(x)
end = timer()
prediction_time = (end - start) / len(results)
results = np.array(results > 0.5).astype(np.uint8)
return calculate_error(results, test_target_img_paths) + (prediction_time,)
optimizer = tf.keras.optimizers.Adam(lr=config.lr)
# compute class weights for the loss: inverse-frequency balanced
# note: we set to 0 the weights for the classes "no_data"(0) and "clouds"(1) to ignore these
class_weight = (1 / LCD.TRAIN_CLASS_COUNTS
) * LCD.TRAIN_CLASS_COUNTS.sum() / (LCD.N_CLASSES)
class_weight[LCD.IGNORED_CLASSES_IDX] = 0.
print(f"Will use class weights: {class_weight}")
#loss = tf.keras.losses.SparseCategoricalCrossentropy(from_logits=False)
loss = WeightedSparseCategoricalCrossEntropy()
#loss = dice_loss()
#loss = jaccard_loss()
print("Compile model")
model.compile(optimizer=optimizer, loss=loss, metrics=[])
# metrics = [tf.keras.metrics.Precision(),
# tf.keras.metrics.Recall(),
# tf.keras.metrics.MeanIoU(num_classes=LCD.N_CLASSES)]) # TODO segmentation metrics
# Launch training
model.fit(
train_dataset,
epochs=config.epochs,
callbacks=callbacks,
steps_per_epoch=trainset_size // config.batch_size,
validation_data=val_dataset,
validation_steps=valset_size // config.batch_size,
)
model.save('/content/experiments/saved')
def run_demo(name,
experiment,
image_size,
training_data_list,
training_mask_list,
model_spec=[16, 32, 64, 128, 256],
preprocess_list=None,
preprocess_stretch=False,
preprocess_mask=None,
keep_image=True):
# make copies of the input array before shuffling
training_data_list = list(training_data_list)
random.Random(experiment * 42).shuffle(training_data_list)
test_input_img_paths = training_data_list[-test_samples:]
if training_mask_list is not None:
training_mask_list = list(training_mask_list)
random.Random(experiment * 42).shuffle(training_mask_list)
test_target_img_paths = training_mask_list[-test_samples:]
else:
test_target_img_paths = None
pp = None
# Chain of preprocessing functions, first one added is performed first
if preprocess_list is not None:
# Instantiate data Sequences for each split
if not preprocess_stretch:
pp = ImagePreprocessGradient(preprocess_list, keep_image, pp)
else:
pp = ImagePreprocessStretchedGradient(preprocess_list, pp)
if preprocess_mask is not None:
# Apply mask after gradients - masking first only gets overwritten
pp = ImagePreprocessMask(preprocess_mask, pp)
if pp is not None:
# Instantiate pre-processed data sequences for each split
test_gen = RoadSeq(len(test_input_img_paths),
image_size,
test_input_img_paths,
test_target_img_paths,
augment_data=False,
preprocess_fn=pp.preprocess())
else:
# use the images as they are
test_gen = RoadSeq(len(test_input_img_paths),
image_size,
test_input_img_paths,
test_target_img_paths,
augment_data=False)
model_name = name + '.' + str(experiment) + '.h5'
model = UNet(image_size, model_spec)
model.compile(optimizer="adam",
loss="binary_crossentropy",
metrics=["acc"])
model.load_weights(model_name)
x, y = test_gen.__getitem__(0)
results = model.predict(x)
results = np.array(results > 0.5).astype(np.uint8)
display_results(test_input_img_paths, test_target_img_paths, [results])
return results
