def get_segmentation_model(freeze):
segmentation_model = pspnet.pspnet(8, 192, 192)
x = segmentation_model.get_layer("activation_10").output
# Explicitly define new model input and output by slicing out old model layers
model_new = Model(inputs=segmentation_model.layers[0].input, outputs=x)
if freeze:
for layer in model_new.layers:
layer.trainable = False
return model_new
filenames.sort()
images = []
for img in filenames:
n = cv2.imread(img)
binary = n.astype("float") / 255.0
img = img.replace("mask", "mask_binary")
cv2.imwrite(img, binary)
from keras_segmentation.models.pspnet import pspnet
from keras_segmentation.pretrained import model_from_checkpoint_path
from keras_segmentation.train import find_latest_checkpoint
#train model
model = pspnet(n_classes=2)
model.train(train_images="dataset/image",
train_annotations="dataset/mask",
checkpoints_path="/usr/code/tmp/checkpoints",
epochs=5)
#load model
model_config = {
'model_class': 'pspnet',
'n_classes': 2,
"input_height": 384,
"input_width": 576
}
latest_weight = find_latest_checkpoint("/usr/code/tmp/checkpoints")
model = model_from_checkpoint_path(model_config, latest_weight)
data_path = "/work/LAS/jannesar-lab/mburke/image-segmentation-keras/cityscape/prepped/"
# data_path = "/work/LAS/jannesar-lab/mburke/image-segmentation-keras/dataset1/"
# data_path = "/Users/MatthewBurke/PycharmProjects/image-segmentation-keras/cityscape/prepped/"
print("data path is ", data_path)
# pret_model = pspnet_101_cityscapes() # load the pretrained model trained on Cityscapes dataset
# print("pret_model")
# # evaluating the pretrained model
# print(pret_model.evaluate_segmentation(inp_images_dir=data_path + "images_prepped_test/",
# annotations_dir=data_path + "annotations_prepped_test/"))
print("loading pspnet")
# psp_101 produces OOM error when training
# input_height=1024, input_width=2048 actual image dims, use defaults of input_height=384, input_width=576 instead
pspnet = pspnet(20) # n_classes changed from 19 to 20
print("model beginning training is ", pspnet.model_name)
pspnet.train(
train_images=data_path + "images_prepped_train/",
train_annotations=data_path + "annotations_prepped_train/",
input_height=None,
input_width=None,
n_classes=None,
verify_dataset=True,
checkpoints_path="./checkpoints/pspnet",
epochs=5, # doesn't do anything now
batch_size=4, # default 2
validate=True,
val_images=data_path + "images_prepped_val",
val_annotations=data_path + "annotations_prepped_val",
val_steps_per_epoch=512,
gen_use_multiprocessing=True, # default False
optimizer_name='adadelta',
do_augment=False,
history_csv="./checkpoints/fcn8/model_history_log.csv")
print("Evaluating ", fcn8.name)
# evaluating the model
print(
fcn8.evaluate_segmentation(
inp_images_dir=data_path + "images_prepped_test/",
annotations_dir=data_path + "annotations_prepped_test/"))
# psp_101 produces OOM error when training
psp_gtfine = pspnet(
20, input_height=1024,
input_width=2048) # change to vgg_unet? # n_classes changed from 19 to 20
print("model beginning training is ", psp_gtfine.name)
psp_gtfine.train(
train_images=data_path + "images_prepped_train/",
train_annotations=data_path + "annotations_prepped_train/",
input_height=None,
input_width=None,
n_classes=None,
verify_dataset=True,
checkpoints_path="./checkpoints/psp_gtfine",
epochs=5, # doesn't do anything now
batch_size=4, # default 2
validate=True,
val_images=data_path + "images_prepped_val",
optimizer='adam',
metrics=['accuracy',
tf.keras.metrics.AUC()])
gan_segnet_reg = gan_disc.make_gan_reg(gen_segnet, disc_segnet_reg)
gan_segnet_reg.compile(loss='binary_crossentropy',
optimizer='adam',
metrics=['accuracy', tf.keras.metrics.AUC()])
train_alternately(gen_model=gen_segnet,
d_model=disc_segnet_reg,
gan_model=gan_segnet_reg,
gen_model_name="segnet",
reg_or_stacked="reg",
train_gen_first=False)
# --------------------- pspnet ---------------------------------------
gen_pspnet = pspnet(20, input_height=128,
input_width=256) # n_classes changed from 19 to 20
gen_pspnet.compile(loss='categorical_crossentropy',
optimizer='adam',
metrics=['accuracy'])
gen_checkpoints_path = get_path("gen_pspnet")
train_gen(gen_pspnet, gen_checkpoints_path, data_path=data_path)
# gen_pspnet.load_weights("")
# Train my stacked input gan
disc_pspnet_stacked = gan_disc.discriminator(gen_pspnet)
disc_pspnet_stacked.compile(loss='binary_crossentropy',
optimizer='adam',
metrics=['accuracy'])
gan_pspnet_stacked = gan_disc.make_gan(gen_pspnet, disc_pspnet_stacked)
gan_pspnet_stacked.compile(loss='binary_crossentropy',
optimizer='adam',