def diff_model(self):
with tf.device('cpu'):
ds = get_tr_ds(
original_pattern='/opt/dataset/tr2_cropped/data/*.png',
mask_pattern='/opt/dataset/tr2_cropped/label/*.png',
batch_size=2)
def reshape_fn(d, l):
d = tf.cast(tf.reshape(d, (-1, 1024, 1024, 3)),
tf.float32) / 255.0
l = tf.reshape(l, (-1, 1024, 1024, 3))
return d, l
ds = ds.map(reshape_fn)
model = Deeplab(num_classes=3)
# model.trainable = True
model.compile(
optimizer=tf.keras.optimizers.Adam(),
loss=tf.keras.losses.BinaryCrossentropy(),
# loss=FocalLoss()
)
# metrics=[tf.keras.metrics.MeanIoU(num_classes=3)])
routine = TrainRoutine(ds=ds, model=model)
routine.run(exp_dir='exp/01', epochs=1)
self.assertTrue(os.path.exists('exp/01/config/exp_config.json'))
def train(loss_list,
freeze_expdir_list,
unfreeze_expdir_list,
batch_size=32,
resize=True,
freeze_epochs=10,
unfreeze_epochs=5):
with tf.device('cpu'):
ds = get_tr_ds(original_pattern='/opt/dataset/tr3_cropped/data/*.png',
mask_pattern='/opt/dataset/tr3_cropped/label/*.png',
batch_size=batch_size)
def reshape_fn(d, l):
d = tf.cast(tf.reshape(d, (-1, 1024, 1024, 3)), tf.float32)
l = l[:, :, :, 1]
l = tf.one_hot(l, 4)
l = tf.reshape(l, (-1, 1024, 1024, 4))
if resize:
l = tf.image.resize(l, [224, 224])
return d, l
ds = ds.map(reshape_fn)
for i in range(len(loss_list)):
model = Deeplab(num_classes=4)
model.layers[0].trainable = False
model.compile(optimizer=tf.keras.optimizers.Adam(),
loss=loss_list[i],
metrics=[MeanIou(num_classes=4)])
routine = TrainRoutine(ds=ds, model=model)
routine.run(exp_dir=freeze_expdir_list[i], epochs=freeze_epochs)
model.layers[0].trainable = True
model.compile(optimizer=tf.keras.optimizers.Adam(),
loss=loss_list[i],
metrics=[MeanIou(num_classes=4)])
routine = TrainRoutine(ds=ds, model=model)
routine.run(exp_dir=unfreeze_expdir_list[i], epochs=unfreeze_epochs)
def apply_clustering_to_conv2D(layer):
if isinstance(layer, tf.keras.layers.Conv2D):
return cluster_weights(layer, **clustering_params)
return layer
clustered_model = tf.keras.models.clone_model(
base_model,
clone_function=apply_clustering_to_conv2D,
)
clustered_model.summary()
with tf.device('/cpu:0'):
ds = get_tr_ds(original_pattern='/opt/dataset/tr2_cropped/data/*.png',
mask_pattern='/opt/dataset/tr2_cropped/label/*.png',
batch_size=1)
def reshape_fn(d, l):
d = tf.cast(tf.reshape(d, (-1, 1024, 1024, 3)), tf.float32) / 255.0
l = l[:, :, :, 1]
l = tf.one_hot(l, 4)
l = tf.reshape(l, (-1, 1024, 1024, 4))
return d, l
ds = ds.map(reshape_fn)
clustered_model.compile(optimizer=tf.keras.optimizers.Adam(),
loss=LovaszLoss(),
metrics=[MeanIou(num_classes=4)])
clustered_model.fit(ds, epochs=1)
def get_tr_ds_demo(self):
ds = get_tr_ds(original_pattern='/opt/dataset/tr2_cropped/data/*.png',
mask_pattern='/opt/dataset/tr2_cropped/label/*.png')
for d in ds.take(1):
self.assertEqual(np.array(d).shape[1:], (2, 1024, 1024, 3))