aug_dict=aug_dict_preprocessing,
transform_ratio=0.,
batch_size=1,
shuffle=False,
add_pad_mask=False,
n_workers=16,
gt_target_channels=None,
yield_epoch_info=True)
model = PiledUnet(
n_nets=3,
in_channels=1,
out_channels=[1, 1, 1],
filter_sizes_down=(((16, 32), (32, 64), (64, 128)),
((16, 32), (32, 64), (64, 128)), ((16, 32), (32, 64),
(64, 128))),
filter_sizes_bottleneck=((128, 256), (128, 256), (128, 256)),
filter_sizes_up=(((256, 128, 128), (128, 64, 64), (64, 32, 32)),
((256, 128, 128), (128, 64, 64), (64, 32, 32)),
((256, 128, 128), (128, 64, 64), (64, 32, 32))),
batch_norm=True,
output_activation='sigmoid')
model.cuda()
summary(model, (1, 64, 64, 64))
if not os.path.exists(results_folder):
os.mkdir(results_folder)
train_model_with_generators(
model,
train_gen,
import numpy as np
from pytorch.pytorch_tools.run_models import predict_model_from_h5_parallel_generator
from glob import glob
experiment_name = 'unet3d_200311_00_membranes'
results_folder = os.path.join(
'/g/schwab/hennies/phd_project/image_analysis/autoseg/cnn_3d_devel',
'unet3d_200311_pytorch_for_membranes', experiment_name)
model = PiledUnet(
n_nets=3,
in_channels=1,
out_channels=[1, 1, 1],
filter_sizes_down=(((16, 32), (32, 64), (64, 128)),
((16, 32), (32, 64), (64, 128)), ((16, 32), (32, 64),
(64, 128))),
filter_sizes_bottleneck=((128, 256), (128, 256), (128, 256)),
filter_sizes_up=(((256, 128, 128), (128, 64, 64), (64, 32, 32)),
((256, 128, 128), (128, 64, 64), (64, 32, 32)),
((256, 128, 128), (128, 64, 64), (64, 32, 32))),
batch_norm=True,
output_activation='sigmoid',
predict=True)
model.cuda()
summary(model, (1, 64, 64, 64))
model.load_state_dict(t.load(os.path.join(results_folder, 'model_0073.h5')))
if not os.path.exists(results_folder):
os.mkdir(results_folder)
aug_dict_preprocessing = dict(smooth_output_sigma=0.5)