import os
import logging
import tensorflow as tf
from medseg_dl import parameters
from medseg_dl.utils import utils_data, utils_misc
from medseg_dl.model import input_fn
from medseg_dl.model import model_fn
from medseg_dl.model import evaluation
def main(dir_model, non_local = 'disable', attgate = 'disable', device = None):
# Since this is the evaluation script all ops should switch to prediction
b_training = False
# Load parameters from model file
file_params = os.path.join(dir_model, 'params.yaml')
assert os.path.isfile(file_params) # file has to exist!
params = parameters.Params(path_yaml=file_params)
# Set logger
utils_misc.set_logger(os.path.join(params.dict['dir_model'], 'eval.log'), params.dict['log_level'])
# Set random seed
if params.dict['b_use_seed']:
tf.set_random_seed(params.dict['random_seed'])
# Set device for graph calc
if device:
os.environ['CUDA_VISIBLE_DEVICES'] = device
else:
os.environ['CUDA_VISIBLE_DEVICES'] = params.dict['device']
""" Fetch data, generate pipeline and model """
tf.reset_default_graph()
# Fetch datasets, atm. saved as json
logging.info('Fetching the datasets...')
filenames_train, filenames_eval = utils_data.load_sets(params.dict['dir_data'],
params.dict['dir_model'],
path_parser_cfg=params.dict['path_parser_cfg'],
set_split=params.dict['set_split'],
b_recreate=True)
# Create a tf.data pipeline
# Note: one can either do evaluation on small little patches or the whole image -> has to be reflected in the model
logging.info('Creating the pipeline...')
logging.info(f'Evaluating {filenames_eval}')
spec_pipeline = input_fn.gen_pipeline_eval_patch(filenames=filenames_eval,
shape_image=params.dict['shape_image_eval'],
shape_input=params.dict['shape_input'],
shape_output=params.dict['shape_output'],
size_batch=params.dict['size_batch_eval'],
channels_out=params.dict['channels_out'],
size_buffer=params.dict['size_buffer'],
num_parallel_calls=params.dict['num_parallel_calls'],
b_with_labels=params.dict['b_eval_labels_patch'],
b_verbose=True)
spec_pipeline_metrics = input_fn.gen_pipeline_eval_image(filenames=filenames_eval,
shape_image=params.dict['shape_image_eval'],
channels_out=params.dict['channels_out'],
size_batch=1,
size_buffer=params.dict['size_buffer'],
num_parallel_calls=params.dict['num_parallel_calls'],
b_with_labels=params.dict['b_eval_labels_image'],
b_verbose=True)
# Create the model
logging.info('Creating the model...')
spec_model = model_fn.model_fn(spec_pipeline,
spec_pipeline_metrics,
b_training=b_training,
channels=params.dict['channels'],
channels_out=params.dict['channels_out'],
batch_size=params.dict['size_batch_eval'],
b_dynamic_pos_mid=params.dict['b_dynamic_pos_mid'],
b_dynamic_pos_end=params.dict['b_dynamic_pos_end'],
non_local=non_local,
non_local_num=1,
attgate=attgate,
filters=params.dict['filters'],
dense_layers=params.dict['dense_layers'],
alpha=params.dict['alpha'],
dropout_rate=params.dict['rate_dropout'],
b_verbose=False)
# Evaluate a saved model
logging.info('Starting evaluation...')
evaluation.sess_eval(spec_pipeline, spec_pipeline_metrics, spec_model, params, filenames_eval)
if __name__ == '__main__':
model_dir_base = '/home/stage13_realshuffle_attgate_batch24_l2_repeat10_shuffle_eval_reminder_timout6000_K'
run = 'catalog_in_model_dir_base'
model_dir = os.path.join(model_dir_base, run)
device = '2'
non_local = 'disable'
attgate = 'active'
main(model_dir, non_local=non_local, attgate=attgate, device=device)
