def get_seg_model(seg_model_name, input_dims):
#return segmentation model based on name
if seg_model_name.lower() == 'block2d':
full_model = BlockModel2D(input_dims, filt_num=16, numBlocks=4)
elif seg_model_name.lower() == 'resunet':
full_model = res_unet(input_dims)
return full_model
def GetBlockModelMasks(weights_path, test_imgs, batch_size):
# Create model
tqdm.write('Loading segmentation model...')
model = BlockModel2D(input_shape=im_dims+(n_channels,),
filt_num=16, numBlocks=4)
# Load weights
model.load_weights(weights_path)
# convert to linear output layer- for better ensembling
model = ConvertModelOutputToLinear(model)
# Get predicted masks
tqdm.write('Getting predicted masks...')
masks = model.predict(test_imgs, batch_size=batch_size, verbose=0)
del model
return masks
trainX, valX, trainY, valY = train_test_split(img_files,
mask_files,
test_size=val_split,
random_state=rng,
shuffle=True)
train_dict = dict([(f, mf) for f, mf in zip(trainX, trainY)])
val_dict = dict([(f, mf) for f, mf in zip(valX, valY)])
# Setup datagens
train_gen = PngDataGenerator(trainX, train_dict, **train_params)
val_gen = PngDataGenerator(valX, val_dict, **val_params)
# Create model
model = BlockModel2D(input_shape=im_dims + (n_channels, ),
filt_num=16,
numBlocks=4)
# Load pretrain weights, if provided
if pretrain_weights_filepath is not None:
model.load_weights(pretrain_weights_filepath)
# Compile model
model.compile(Adam(), loss=dice_coef_loss)
# Create callbacks
cb_check = ModelCheckpoint(weight_filepath,
monitor='val_loss',
verbose=1,
save_best_only=True,
save_weights_only=True,
mode='auto',
def _get_model(self, filt_num=16, numBlocks=4):
return BlockModel2D(input_shape=self.dims+(self.n_channels,),
filt_num=filt_num, numBlocks=numBlocks)
history2 = model.fit_generator(generator=train_gen,
epochs=epochs[1],
verbose=1,
callbacks=[cb_check, cb_plateau],
validation_data=val_gen)
# %% ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
# ~~~~~~ Full Size Training ~~~~~~~
# ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
print('Setting up 1024 training')
# make full-size model
full_model = BlockModel2D((1024, 1024, n_channels), filt_num=16, numBlocks=4)
full_model.load_weights(best_weight_path)
# Compile model
full_model.compile(Adam(lr=learnRate), loss=dice_coef_loss)
# Set weight paths
best_weight_path = best_weight_filepath.format('1024train')
# Setup full size datagens with only large masks
train_gen, val_gen = get_seg_datagen(
pos_img_filt_path, pos_mask_filt_path, full_train_params, full_val_params, val_split)
# Create callbacks
cb_check = ModelCheckpoint(best_weight_path, monitor='val_loss',
verbose=1, save_best_only=True,