LABELS_DICT = {
1: 'SOL',
2: 'GM',
3: 'GL',
4: 'TA',
5: 'ELD',
6: 'PE',
}
'''
MODEL_RESOLUTION = np.array([1.037037, 1.037037])
MODEL_SIZE = (432, 432)
MODEL_SIZE_SPLIT = (250, 250)
image_list, mask_list = pretrain.common_input_process_split(
LABELS_DICT, MODEL_RESOLUTION, MODEL_SIZE, MODEL_SIZE_SPLIT, {
'image_list': image_list,
'resolution': MODEL_RESOLUTION
}, seg_list)
ch = mask_list[0].shape[2]
aggregated_masks = []
mask_list_no_overlap = []
for masks in mask_list:
agg, new_masks = pretrain.calc_aggregated_masks_and_remove_overlap(masks)
aggregated_masks.append(agg)
mask_list_no_overlap.append(new_masks)
for slice_number in range(len(image_list)):
img = image_list[slice_number]
segmentation = model.predict(
np.expand_dims(np.stack([img, np.zeros(MODEL_SIZE_SPLIT)], axis=-1),
def thigh_incremental_mem(modelObj: DynamicDLModel,
trainingData: dict,
trainingOutputs,
bs=5,
minTrainImages=5):
import dl.common.preprocess_train as pretrain
from dl.common.DataGenerators import DataGeneratorMem
import os
from keras.callbacks import ModelCheckpoint
from keras import optimizers
import time
try:
np
except:
import numpy as np
from dl.labels.thigh import inverse_labels
MODEL_RESOLUTION = np.array([1.037037, 1.037037])
MODEL_SIZE = (432, 432)
MODEL_SIZE_SPLIT = (250, 250)
BAND = 49
BATCH_SIZE = bs
CHECKPOINT_PATH = os.path.join(".", "Weights_incremental_split", "thigh")
MIN_TRAINING_IMAGES = minTrainImages
os.makedirs(CHECKPOINT_PATH, exist_ok=True)
t = time.time()
print('Image preprocess')
image_list, mask_list = pretrain.common_input_process_split(
inverse_labels, MODEL_RESOLUTION, MODEL_SIZE, MODEL_SIZE_SPLIT,
trainingData, trainingOutputs)
print('Done. Elapsed', time.time() - t)
nImages = len(image_list)
if nImages < MIN_TRAINING_IMAGES:
print("Not enough images for training")
return
print("image shape", image_list[0].shape)
print("mask shape", mask_list[0].shape)
print('Weight calculation')
t = time.time()
output_data_structure = pretrain.input_creation_mem(
image_list, mask_list, BAND)
print('Done. Elapsed', time.time() - t)
card = len(image_list)
steps = int(float(card) / BATCH_SIZE)
print(f'Incremental learning for thigh with {nImages} images')
t = time.time()
netc = modelObj.model
checkpoint_files = os.path.join(
CHECKPOINT_PATH, "weights - {epoch: 02d} - {loss: .2f}.hdf5")
training_generator = DataGeneratorMem(output_data_structure,
list_X=list(range(steps *
BATCH_SIZE)),
batch_size=BATCH_SIZE,
dim=MODEL_SIZE_SPLIT)
#check = ModelCheckpoint(filepath=checkpoint_files, monitor='loss', verbose=0, save_best_only=False,save_weights_only=True, mode='auto', period=10)
check = ModelCheckpoint(
filepath=checkpoint_files,
monitor='loss',
verbose=0,
save_best_only=True, # save_freq='epoch',
save_weights_only=True,
mode='auto')
adamlr = optimizers.Adam(learning_rate=0.001,
beta_1=0.9,
beta_2=0.999,
epsilon=1e-08,
amsgrad=True)
netc.compile(loss=pretrain.weighted_loss, optimizer=adamlr)
#history = netc.fit_generator(generator=training_generator, steps_per_epoch=steps, epochs=5, callbacks=[check], verbose=1)
history = netc.fit(x=training_generator,
steps_per_epoch=steps,
epochs=5,
callbacks=[check],
verbose=1)
print('Done. Elapsed', time.time() - t)