def train_net():
# Create a new train directory, where to save all.log and config.json
#H['output_dir'] = 'output_dir/train_dir/%s' % datetime.now().strftime('%Y_%m_%d_%H.%M')
if not tf.gfile.Exists(H['output_dir']):
tf.gfile.MakeDirs(H['output_dir'])
with open(H['output_dir'] + '/config.json', 'w') as conf_file:
json.dump(H, conf_file, indent=4)
shutil.copy('../tf_scripts/architecture/model_def/resnet2D.py',
H['output_dir'] + '/resnet2D.py')
shutil.copytree('../tf_scripts/architecture/',
H['output_dir'] + '/architecture')
shutil.copytree('../tf_scripts/io_modules/',
H['output_dir'] + '/io_modules')
initialize_logger(folder=H['output_dir'])
with tf.Graph().as_default(), tf.device('/cpu:0'):
train_data_iter = List_Iterator(
H,
img_lst=H['train_lst'],
img_shape=H['image_shape'],
label_shape=H['label_shape'],
batch_size=H['batch_size'],
num_preprocess_threads=4,
shuffle=True,
is_training=True,
)
valid_data_iter = List_Iterator(
H,
img_lst=H['val_lst'],
img_shape=H['image_shape'],
label_shape=H['label_shape'],
batch_size=H['batch_size'],
num_preprocess_threads=4,
shuffle=True,
is_training=False,
)
model = resnet2D
update_scopes = []
#update_scopes.append('logits')
# Loss operations
loss_op = metrics.logloss
# Additional Evaluation metrics
metric_ops = None #metrics.logloss_rank
H['train_image_summary'] = image_summaries.classification_image_summary
H['validation_image_summary'] = image_summaries.classification_image_summary
H['model_graph'] = model
H['loss'] = loss_op
H['metrics'] = metric_ops
H['train_scopes'] = update_scopes
H['train_iter'] = train_data_iter
H['valid_iter'] = valid_data_iter
H['VARIABLES_TO_RESTORE'] = tf.contrib.slim.get_variables_to_restore()
H['UPDATE_OPS_COLLECTION'] = tf.GraphKeys.UPDATE_OPS
args = []
train(*args, **H)
def train_net():
# Create a new train directory, where to save all.log and config.json
#H['output_dir'] = 'output_dir/train_dir/%s' % datetime.now().strftime('%Y_%m_%d_%H.%M')
if not tf.gfile.Exists(H['output_dir']):
tf.gfile.MakeDirs(H['output_dir'])
with open(H['output_dir'] + '/config.json', 'w') as conf_file:
json.dump(H, conf_file, indent=4)
initialize_logger(folder=H['output_dir'])
shutil.copy('../tf_scripts/architecture/model_def/unet.py',
H['output_dir'] + '/unet.py')
shutil.copytree('../tf_scripts/architecture/',
H['output_dir'] + '/architecture')
shutil.copytree('../tf_scripts/io_modules/',
H['output_dir'] + '/io_modules')
with tf.Graph().as_default(), tf.device('/cpu:0'):
train_data_iter = NPY_construct_Iter(
H,
img_lst=H['train_lst'],
npy_lst=H['train_npy_path'],
img_shape=H['image_shape'],
label_shape=H['label_shape'],
batch_size=H['batch_size'],
num_preprocess_threads=4,
shuffle=True,
is_training=True,
)
valid_data_iter = NPY_construct_Iter(
H,
img_lst=H['val_lst'],
npy_lst=H['val_npy_path'],
img_shape=H['image_shape'],
label_shape=H['label_shape'],
batch_size=H['batch_size'],
num_preprocess_threads=4,
shuffle=True,
is_training=False,
)
model = unet_model.unet
update_scopes = []
#update_scopes.append('logits')
# Loss operations
# here correct LOSS is to choose
# - jaccard_separated_channels_with_center
# - logloss_with_center
# - MSE_segmentation_with_center
loss_op = metrics.MSE_segmentation_with_center
# Additional Evaluation metrics
metric_ops = metrics.MSE_and_sensitivity_on_center
H['train_image_summary'] = image_summaries.segmentation_image_summary
H['validation_image_summary'] = image_summaries.segmentation_image_summary
H['model_graph'] = model
H['loss'] = loss_op
H['metrics'] = metric_ops
H['train_scopes'] = update_scopes
H['train_iter'] = train_data_iter
H['valid_iter'] = valid_data_iter
H['VARIABLES_TO_RESTORE'] = tf.contrib.slim.get_variables_to_restore()
H['UPDATE_OPS_COLLECTION'] = tf.GraphKeys.UPDATE_OPS
args = []
train(*args, **H)