def create_model():
fcn_vgg16 = FCN(input_shape=(PATCH_SIZE, PATCH_SIZE, 3), classes=3,
weights='imagenet', trainable_encoder=True)
sgd = optimizers.SGD(lr=0.001, decay=1e-6, momentum=0.9, nesterov=True)
fcn_vgg16.compile(optimizer=sgd, loss='categorical_crossentropy',
metrics=['accuracy'])
return fcn_vgg16
def test_fcn_vgg16_shape():
"""Test output shape."""
if K.image_data_format() == 'channels_first':
input_shape = (3, 500, 500)
else:
input_shape = (500, 500, 3)
fcn_vgg16 = FCN(input_shape=input_shape, classes=21)
layers = [l.name for l in fcn_vgg16.layers]
assert 'upscore_feat1' in layers
assert 'upscore_feat2' in layers
assert 'upscore_feat3' in layers
for l in fcn_vgg16.layers:
if l.name == 'block1_pool':
test_shape = (None, 250, 250, 64)
assert is_same_shape(l.output_shape, test_shape)
elif l.name == 'block2_pool':
test_shape = (None, 125, 125, 128)
assert is_same_shape(l.output_shape, test_shape)
elif l.name == 'block3_pool':
test_shape = (None, 63, 63, 256)
assert is_same_shape(l.output_shape, test_shape)
elif l.name == 'block4_pool':
test_shape = (None, 32, 32, 512)
assert is_same_shape(l.output_shape, test_shape)
elif l.name == 'block5_pool':
test_shape = (None, 16, 16, 512)
assert is_same_shape(l.output_shape, test_shape)
elif l.name == 'dropout_2':
test_shape = (None, 16, 16, 4096)
assert is_same_shape(l.output_shape, test_shape)
elif l.name == 'upscore_feat1':
test_shape = (None, 32, 32, 21)
assert is_same_shape(l.output_shape, test_shape)
elif l.name == 'upscore_feat2':
test_shape = (None, 63, 63, 21)
assert is_same_shape(l.output_shape, test_shape)
elif l.name == 'upscore_feat3':
test_shape = (None, 500, 500, 21)
assert is_same_shape(l.output_shape, test_shape)
elif l.name == 'score':
test_shape = (None, 500, 500, 21)
assert is_same_shape(l.output_shape, test_shape)
assert is_same_shape(fcn_vgg16.output_shape, (None, 500, 500, 21))
input_shape = (1366, 768, 3)
fcn_vgg16 = FCN(input_shape=input_shape, classes=21)
assert is_same_shape(fcn_vgg16.output_shape, (None, 1366, 768, 21))
def main(args):
# set the necessary list
# train_list = pd.read_csv(args.train_list,header=None)
# val_list = pd.read_csv(args.val_list,header=None)
#
# # set the necessary directories
# trainimg_dir = args.trainimg_dir
# trainmsk_dir = args.trainmsk_dir
# valimg_dir = args.valimg_dir
# valmsk_dir = args.valmsk_dir
#
# train_gen = data_gen_small(trainimg_dir, trainmsk_dir, train_list, args.batch_size, [args.input_shape[0], args.input_shape[1]], args.n_labels)
# val_gen = data_gen_small(valimg_dir, valmsk_dir, val_list, args.batch_size, [args.input_shape[0], args.input_shape[1]], args.n_labels)
train_gen, val_gen = train_val_generator(args.batch_size)
# segnet = CreateSegNet(args.input_shape, args.n_labels, args.kernel, args.pool_size, args.output_mode)
from keras_fcn import FCN
segnet = FCN(input_shape=args.input_shape,
classes=args.n_labels,
weights='imagenet',
trainable_encoder=True)
print(segnet.summary())
checkpointer = ModelCheckpoint(
filepath="model/weights.{epoch:02d}-{val_acc:.4f}.hdf5",
verbose=1,
save_best_only=False)
mycallback = MyCallBack()
segnet.compile(loss=args.loss,
optimizer=args.optimizer,
metrics=["accuracy"])
segnet.fit_generator(train_gen,
steps_per_epoch=args.epoch_steps,
epochs=args.n_epochs,
validation_data=val_gen,
validation_steps=args.val_steps,
callbacks=[checkpointer, mycallback])
segnet.save_weights("model/finalSegNet" + str(args.n_epochs) + ".hdf5")
print("sava weight done..")
json_string = segnet.to_json()
open("model/LIP_SegNet.json", "w").write(json_string)
def test_fcn_vgg16_correctness():
"""Test output not NaN."""
if K.image_data_format() == 'channels_first':
input_shape = (3, 500, 500)
x = np.random.rand(1, 3, 500, 500)
y = np.random.randint(21, size=(1, 500, 500))
y = np.eye(21)[y]
y = np.transpose(y, (0, 3, 1, 2))
else:
input_shape = (500, 500, 3)
x = np.random.rand(1, 500, 500, 3)
y = np.random.randint(21, size=(1, 500, 500))
y = np.eye(21)[y]
fcn_vgg16 = FCN(classes=21, input_shape=input_shape)
fcn_vgg16.compile(optimizer='rmsprop',
loss='categorical_crossentropy',
metrics=['accuracy'])
fcn_vgg16.fit(x, y, batch_size=1, epochs=1)
loss = fcn_vgg16.evaluate(x, y, batch_size=1)
assert not np.any(np.isinf(loss))
assert not np.any(np.isnan(loss))
y_pred = fcn_vgg16.predict(x, batch_size=1)
assert not np.any(np.isinf(y_pred))
assert not np.any(np.isnan(y_pred))
def test_fcn_vgg16_correctness():
"""Test output not NaN."""
if K.image_data_format() == 'channels_first':
input_shape = (3, 500, 500)
x = np.random.rand(1, 3, 500, 500)
y = np.random.randint(21, size=(1, 500, 500))
y = np.eye(21)[y]
y = np.transpose(y, (0, 3, 1, 2))
else:
input_shape = (500, 500, 3)
x = np.random.rand(1, 500, 500, 3)
y = np.random.randint(21, size=(1, 500, 500))
y = np.eye(21)[y]
fcn_vgg16 = FCN(classes=21, input_shape=input_shape)
fcn_vgg16.compile(optimizer='rmsprop',
loss='categorical_crossentropy',
metrics=['accuracy'])
fcn_vgg16.fit(x, y, batch_size=1, epochs=1)
loss = fcn_vgg16.evaluate(x, y, batch_size=1)
assert not np.any(np.isinf(loss))
assert not np.any(np.isnan(loss))
y_pred = fcn_vgg16.predict(x, batch_size=1)
assert not np.any(np.isinf(y_pred))
assert not np.any(np.isnan(y_pred))
#check_num = CheckNumericsOps(validation_data=[np.random.random((1, 224, 224, 3)), 1],
# histogram_freq=100)
datagen = PascalVocGenerator(image_shape=[224, 224, 3],
image_resample=True,
pixelwise_center=True,
pixel_mean=[115.85100, 110.50989, 102.16182],
pixelwise_std_normalization=True,
pixel_std=[70.30930, 69.41244, 72.60676])
train_loader = ImageSetLoader(**init_args['image_set_loader']['train'])
val_loader = ImageSetLoader(**init_args['image_set_loader']['val'])
fcn_vgg16 = FCN(input_shape=(224, 224, 3),
classes=21,
weight_decay=3e-3,
weights='imagenet',
trainable_encoder=True)
optimizer = keras.optimizers.Adam(1e-4)
fcn_vgg16.compile(optimizer=optimizer,
loss='categorical_crossentropy',
metrics=['accuracy'])
fcn_vgg16.fit_generator(
datagen.flow_from_imageset(class_mode='categorical',
classes=21,
batch_size=1,
shuffle=True,
image_set_loader=train_loader),
steps_per_epoch=1112,
#%%
print(len(X_train_all))
#%%
generator = ImageDataGenerator(
horizontal_flip=True,
vertical_flip=True,
)
#%%
# u-net
model = get_unet(X_train_all[0].shape[0], 3)
history = model.fit(X_train_all, y_train_all, batch_size=32, epochs=10)
#%%
fcn_vgg16 = FCN(input_shape=X_train[0].shape, classes=3,
weights='None', trainable_encoder=True)
sgd = optimizers.SGD(lr=0.001, decay=1e-6, momentum=0.9, nesterov=True)
fcn_vgg16.compile(optimizer=sgd,
loss='categorical_crossentropy',
metrics=['accuracy'])
# history = fcn_vgg16.fit(X_train_all, y_train_all, batch_size=32, epochs=20)
batch_size = 64
epochs = 50
history = fcn_vgg16.fit_generator(generator.flow(X_train_all, y_train_all, batch_size=batch_size),
steps_per_epoch=len(X_train_all) / batch_size, epochs=epochs)
#%%
plot_history(history)