File size: 17,992 Bytes
6a4f823 |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 |
import os, sys
currentdir = os.path.dirname(os.path.realpath(__file__))
parentdir = os.path.dirname(currentdir)
sys.path.append(parentdir) # PYTHON > 3.3 does not allow relative referencing
import numpy as np
import tensorflow as tf
from tensorflow.keras.callbacks import ModelCheckpoint, TensorBoard, EarlyStopping
from tensorflow.keras import Input
from tensorflow.keras.models import Model
from tensorflow.python.keras.utils import Progbar
from tensorflow.python.framework.errors import InvalidArgumentError
import voxelmorph as vxm
import neurite as ne
import h5py
from datetime import datetime
import pickle
import DeepDeformationMapRegistration.utils.constants as C
from DeepDeformationMapRegistration.utils.misc import try_mkdir, function_decorator
from DeepDeformationMapRegistration.utils.nifti_utils import save_nifti
from DeepDeformationMapRegistration.losses import NCC, HausdorffDistanceErosion, GeneralizedDICEScore, StructuralSimilarity_simplified
from DeepDeformationMapRegistration.layers import AugmentationLayer
from DeepDeformationMapRegistration.ms_ssim_tf import MultiScaleStructuralSimilarity, _MSSSIM_WEIGHTS
from DeepDeformationMapRegistration.utils.acummulated_optimizer import AdamAccumulated
from Brain_study.data_generator import BatchGenerator
from Brain_study.utils import SummaryDictionary, named_logs
import time
def launch_train(dataset_folder, validation_folder, output_folder, gpu_num=0, lr=1e-4, rw=5e-3, simil='ssim', segm='hd'):
assert dataset_folder is not None and output_folder is not None
os.environ['CUDA_DEVICE_ORDER'] = C.DEV_ORDER
os.environ['CUDA_VISIBLE_DEVICES'] = str(gpu_num) # Check availability before running using 'nvidia-smi'
C.GPU_NUM = str(gpu_num)
output_folder = os.path.join(output_folder + '_' + datetime.now().strftime("%H%M%S-%d%m%Y"))
os.makedirs(output_folder, exist_ok=True)
log_file = open(os.path.join(output_folder, 'log.txt'), 'w')
C.TRAINING_DATASET = dataset_folder
C.VALIDATION_DATASET = validation_folder
C.ACCUM_GRADIENT_STEP = 16
C.BATCH_SIZE = 2 if C.ACCUM_GRADIENT_STEP == 1 else C.ACCUM_GRADIENT_STEP
C.EARLY_STOP_PATIENCE = 10 * (C.ACCUM_GRADIENT_STEP / 2 if C.ACCUM_GRADIENT_STEP != 1 else 1)
C.LEARNING_RATE = lr
C.LIMIT_NUM_SAMPLES = None
C.EPOCHS = 10000
aux = "[{}]\tINFO:\nTRAIN DATASET: {}\nVALIDATION DATASET: {}\n" \
"GPU: {}\n" \
"BATCH SIZE: {}\n" \
"LR: {}\n" \
"SIMILARITY: {}\n" \
"REG. WEIGHT: {}\n" \
"EPOCHS: {:d}\n" \
"ACCUM. GRAD: {}\n" \
"EARLY STOP PATIENCE: {}".format(datetime.now().strftime('%H:%M:%S\t%d/%m/%Y'),
C.TRAINING_DATASET,
C.VALIDATION_DATASET,
C.GPU_NUM,
C.BATCH_SIZE,
C.LEARNING_RATE,
simil,
rw,
C.EPOCHS,
C.ACCUM_GRADIENT_STEP,
C.EARLY_STOP_PATIENCE)
log_file.write(aux)
print(aux)
# Load data
# Build data generator
data_generator = BatchGenerator(C.TRAINING_DATASET, C.BATCH_SIZE if C.ACCUM_GRADIENT_STEP == 1 else 1, True,
C.TRAINING_PERC, labels=['all'], combine_segmentations=False,
directory_val=C.VALIDATION_DATASET)
train_generator = data_generator.get_train_generator()
validation_generator = data_generator.get_validation_generator()
image_input_shape = train_generator.get_data_shape()[1][:-1]
image_output_shape = [64] * 3
nb_labels = len(train_generator.get_segmentation_labels())
# Config the training sessions
config = tf.compat.v1.ConfigProto() # device_count={'GPU':0})
config.gpu_options.allow_growth = True
config.log_device_placement = False ## to log device placement (on which device the operation ran)
config.allow_soft_placement = True # https://github.com/tensorflow/tensorflow/issues/30782
sess = tf.Session(config=config)
tf.keras.backend.set_session(sess)
# Build model
input_layer_augm = Input(shape=train_generator.get_data_shape()[0], name='input_augmentation')
augm_layer = AugmentationLayer(max_displacement=C.MAX_AUG_DISP, # Max 30 mm in isotropic space
max_deformation=C.MAX_AUG_DEF, # Max 6 mm in isotropic space
max_rotation=C.MAX_AUG_ANGLE, # Max 10 deg in isotropic space
num_control_points=C.NUM_CONTROL_PTS_AUG,
num_augmentations=C.NUM_AUGMENTATIONS,
gamma_augmentation=C.GAMMA_AUGMENTATION,
brightness_augmentation=C.BRIGHTNESS_AUGMENTATION,
in_img_shape=image_input_shape,
out_img_shape=image_output_shape,
only_image=False,
only_resize=False,
trainable=False)
augm_model = Model(inputs=input_layer_augm, outputs=augm_layer(input_layer_augm))
enc_features = [16, 32, 32, 32]# const.ENCODER_FILTERS
dec_features = [32, 32, 32, 32, 32, 16, 16]# const.ENCODER_FILTERS[::-1]
nb_features = [enc_features, dec_features]
network = vxm.networks.VxmDenseSemiSupervisedSeg(inshape=image_output_shape,
nb_labels=nb_labels,
nb_unet_features=nb_features,
int_steps=0,
int_downsize=1,
seg_downsize=1)
# Compile the model
SSIM_KER_SIZE = 5
MS_SSIM_WEIGHTS = _MSSSIM_WEIGHTS[:3]
MS_SSIM_WEIGHTS /= np.sum(MS_SSIM_WEIGHTS)
if simil=='ssim':
loss_simil = StructuralSimilarity_simplified(patch_size=SSIM_KER_SIZE, dim=3, dynamic_range=1.).loss
elif simil=='ms_ssim':
loss_simil = MultiScaleStructuralSimilarity(max_val=1., filter_size=SSIM_KER_SIZE, power_factors=MS_SSIM_WEIGHTS).loss
elif simil=='ncc':
loss_simil = NCC(image_input_shape).loss
elif simil=='ms_ssim__ncc' or simil=='ncc__ms_ssim':
@function_decorator('MS_SSIM_NCC__loss')
def loss_simil(y_true, y_pred):
return MultiScaleStructuralSimilarity(max_val=1., filter_size=SSIM_KER_SIZE, power_factors=MS_SSIM_WEIGHTS).loss(y_true, y_pred) + NCC(image_input_shape).loss(y_true, y_pred)
elif simil=='ms_ssim__mse' or simil=='mse__ms_ssim':
@function_decorator('MS_SSIM_MSE__loss')
def loss_simil(y_true, y_pred):
return MultiScaleStructuralSimilarity(max_val=1., filter_size=SSIM_KER_SIZE, power_factors=MS_SSIM_WEIGHTS).loss(y_true, y_pred) + vxm.losses.MSE().loss(y_true, y_pred)
elif simil=='ssim__ncc' or simil=='ncc__ssim' :
@function_decorator('SSIM_NCC__loss')
def loss_simil(y_true, y_pred):
return StructuralSimilarity_simplified(patch_size=SSIM_KER_SIZE, dim=3, dynamic_range=1.).loss(y_true, y_pred) + NCC(image_input_shape).loss(y_true, y_pred)
elif simil=='ssim__mse' or simil=='mse__ssim':
@function_decorator('SSIM_MSE__loss')
def loss_simil(y_true, y_pred):
return StructuralSimilarity_simplified(patch_size=SSIM_KER_SIZE, dim=3, dynamic_range=1.).loss(y_true, y_pred) + vxm.losses.MSE().loss(y_true, y_pred)
else:
loss_simil = vxm.losses.MSE().loss
if segm == 'hd':
loss_segm = HausdorffDistanceErosion(3, 10, im_shape=image_output_shape + [nb_labels]).loss
elif segm == 'dice':
loss_segm = GeneralizedDICEScore(image_output_shape + [nb_labels], num_labels=nb_labels).loss
else:
raise ValueError('No valid value for segm')
losses = {'transformer': loss_simil,
'seg_transformer': loss_segm,
'flow': vxm.losses.Grad('l2').loss}
loss_weights = {'transformer': 1,
'seg_transformer': 1.,
'flow': 5e-3}
metrics = {'transformer': [vxm.losses.MSE().loss, NCC(image_input_shape).metric, MultiScaleStructuralSimilarity(max_val=1., filter_size=SSIM_KER_SIZE, power_factors=MS_SSIM_WEIGHTS).metric],
'seg_transformer': [GeneralizedDICEScore(image_output_shape + [train_generator.get_data_shape()[2][-1]], num_labels=nb_labels).metric,
HausdorffDistanceErosion(3, 10, im_shape=image_output_shape + [train_generator.get_data_shape()[2][-1]]).metric
]}
metrics_weights = {'transformer': 1,
'seg_transformer': 1,
'flow': rw}
# Train
os.makedirs(output_folder, exist_ok=True)
os.makedirs(os.path.join(output_folder, 'checkpoints'), exist_ok=True)
os.makedirs(os.path.join(output_folder, 'tensorboard'), exist_ok=True)
os.makedirs(os.path.join(output_folder, 'history'), exist_ok=True)
callback_best_model = ModelCheckpoint(os.path.join(output_folder, 'checkpoints', 'best_model.h5'),
save_best_only=True, monitor='val_loss', verbose=1, mode='min')
# callback_save_model = ModelCheckpoint(os.path.join(output_folder, 'checkpoints', 'weights.{epoch:05d}-{val_loss:.2f}.h5'),
# save_best_only=True, save_weights_only=True, monitor='val_loss', verbose=0, mode='min')
# CSVLogger(train_log_name, ';'),
# UpdateLossweights([haus_weight, dice_weight], [const.MODEL+'_resampler_seg', const.MODEL+'_resampler_seg'])
callback_tensorboard = TensorBoard(log_dir=os.path.join(output_folder, 'tensorboard'),
batch_size=C.BATCH_SIZE, write_images=False, histogram_freq=0,
update_freq='epoch', # or 'batch' or integer
write_graph=True, write_grads=True
)
callback_early_stop = EarlyStopping(monitor='val_loss', verbose=1, patience=C.EARLY_STOP_PATIENCE, min_delta=0.00001)
# Compile the model
optimizer = AdamAccumulated(C.ACCUM_GRADIENT_STEP, lr=C.LEARNING_RATE)
network.compile(optimizer=optimizer,
loss=losses,
loss_weights=loss_weights,
metrics=metrics)
callback_tensorboard.set_model(network)
callback_best_model.set_model(network)
# callback_save_model.set_model(network)
callback_early_stop.set_model(network)
summary = SummaryDictionary(network, C.BATCH_SIZE, C.ACCUM_GRADIENT_STEP)
names = network.metrics_names # It give both the loss and metric names
log_file.write('\n\n[{}]\tINFO:\tStart training\n\n'.format(datetime.now().strftime('%H:%M:%S\t%d/%m/%Y')))
with sess.as_default():
#sess.run(tf.global_variables_initializer())
callback_tensorboard.on_train_begin()
callback_early_stop.on_train_begin()
callback_best_model.on_train_begin()
# callback_save_model.on_train_begin()
for epoch in range(C.EPOCHS):
callback_tensorboard.on_epoch_begin(epoch)
callback_early_stop.on_epoch_begin(epoch)
callback_best_model.on_epoch_begin(epoch)
# callback_save_model.on_epoch_begin(epoch)
print("\nEpoch {}/{}".format(epoch, C.EPOCHS))
print('TRAINING')
log_file.write('\n\n[{}]\tINFO:\tTraining epoch {}\n\n'.format(datetime.now().strftime('%H:%M:%S\t%d/%m/%Y'), epoch))
progress_bar = Progbar(len(train_generator), width=30, verbose=1)
t0 = time.time()
for step, (in_batch, _) in enumerate(train_generator, 1):
#print('Loaded in {} s'.format(time.time() - t0))
# callback_tensorboard.on_train_batch_begin(step)
callback_best_model.on_train_batch_begin(step)
# callback_save_model.on_train_batch_begin(step)
callback_early_stop.on_train_batch_begin(step)
try:
t0 = time.time()
fix_img, mov_img, fix_seg, mov_seg = augm_model.predict(in_batch)
#print('Augmented in {} s'.format(time.time() - t0))
np.nan_to_num(fix_img, copy=False)
np.nan_to_num(mov_img, copy=False)
if np.isnan(np.sum(mov_img)) or np.isnan(np.sum(fix_img)) or np.isinf(np.sum(mov_img)) or np.isinf(np.sum(fix_img)):
msg = 'CORRUPTED DATA!! Unique: Fix: {}\tMoving: {}'.format(np.unique(fix_img),
np.unique(mov_img))
print(msg)
log_file.write('\n\n[{}]\tWAR: {}'.format(datetime.now().strftime('%H:%M:%S\t%d/%m/%Y'), msg))
except InvalidArgumentError as err:
print('TF Error : {}'.format(str(err)))
continue
t0 = time.time()
ret = network.train_on_batch(x=(mov_img, fix_img, mov_seg),
y=(fix_img, fix_img, fix_seg))
# print("Trained on batch in {} s".format(time.time() - t0))
if np.isnan(ret).any():
os.makedirs(os.path.join(output_folder, 'corrupted'), exist_ok=True)
save_nifti(mov_img, os.path.join(output_folder, 'corrupted', 'mov_img_nan.nii.gz'))
save_nifti(fix_img, os.path.join(output_folder, 'corrupted', 'fix_img_nan.nii.gz'))
pred_img, dm = network((mov_img, fix_img))
save_nifti(pred_img, os.path.join(output_folder, 'corrupted', 'pred_img_nan.nii.gz'))
save_nifti(dm, os.path.join(output_folder, 'corrupted', 'dm_nan.nii.gz'))
log_file.write('\n\n[{}]\tERR: Corruption error'.format(datetime.now().strftime('%H:%M:%S\t%d/%m/%Y')))
raise ValueError('CORRUPTION ERROR: Halting training')
summary.on_train_batch_end(ret)
# callback_tensorboard.on_train_batch_end(step, named_logs(network, ret))
callback_best_model.on_train_batch_end(step, named_logs(network, ret))
# callback_save_model.on_train_batch_end(step, named_logs(network, ret))
callback_early_stop.on_train_batch_end(step, named_logs(network, ret))
progress_bar.update(step, zip(names, ret))
log_file.write('\t\tStep {:03d}: {}'.format(step, ret))
t0 = time.time()
print('End of epoch{}: '.format(step), ret, '\n')
val_values = progress_bar._values.copy()
ret = [val_values[x][0]/val_values[x][1] for x in names]
print('\nVALIDATION')
log_file.write('\n\n[{}]\tINFO:\tValidation epoch {}\n\n'.format(datetime.now().strftime('%H:%M:%S\t%d/%m/%Y'), epoch))
progress_bar = Progbar(len(validation_generator), width=30, verbose=1)
for step, (in_batch, _) in enumerate(validation_generator, 1):
# callback_tensorboard.on_test_batch_begin(step) # This is cursed, don't do it again
# callback_early_stop.on_test_batch_begin(step)
try:
fix_img, mov_img, fix_seg, mov_seg = augm_model.predict(in_batch)
except InvalidArgumentError as err:
print('TF Error : {}'.format(str(err)))
continue
ret = network.test_on_batch(x=(mov_img, fix_img, mov_seg),
y=(fix_img, fix_img, fix_seg))
# pred_segm = network.register(mov_segm, fix_segm)
summary.on_validation_batch_end(ret)
# callback_early_stop.on_test_batch_end(step, named_logs(network, ret))
# callback_tensorboard.on_test_batch_end(step, named_logs(network, ret)) # This is cursed, don't do it again
progress_bar.update(step, zip(names, ret))
log_file.write('\t\tStep {:03d}: {}'.format(step, ret))
val_values = progress_bar._values.copy()
ret = [val_values[x][0]/val_values[x][1] for x in names]
train_generator.on_epoch_end()
validation_generator.on_epoch_end()
epoch_summary = summary.on_epoch_end()
callback_tensorboard.on_epoch_end(epoch, epoch_summary)
callback_early_stop.on_epoch_end(epoch, epoch_summary)
callback_best_model.on_epoch_end(epoch, epoch_summary)
# callback_save_model.on_epoch_end(epoch, named_logs(network, ret, True))
callback_tensorboard.on_train_end()
# callback_save_model.on_train_end()
callback_best_model.on_train_end()
callback_early_stop.on_train_end()
if __name__ == '__main__':
os.environ['CUDA_DEVICE_ORDER'] = C.DEV_ORDER
os.environ['CUDA_VISIBLE_DEVICES'] = '0' # Check availability before running using 'nvidia-smi'
config = tf.compat.v1.ConfigProto() # device_count={'GPU':0})
config.gpu_options.allow_growth = True
config.log_device_placement = False ## to log device placement (on which device the operation ran)
tf.keras.backend.set_session(tf.Session(config=config))
launch_train('/mnt/EncryptedData1/Users/javier/Brain_study/ERASE',
'TrainOutput/THESIS/UW_None_mse_ssim_haus',
0)
|