def main(flags):
nn_utils.set_gpu(GPU)
# define network
model = pspnet.PSPNet(flags.num_classes, flags.tile_size, suffix=flags.model_suffix, learn_rate=flags.learning_rate,
decay_step=flags.decay_step, decay_rate=flags.decay_rate, epochs=flags.epochs,
batch_size=flags.batch_size, weight_decay=flags.weight_decay, momentum=flags.momentum)
cm_train = cityscapes_reader.CollectionMakerCityscapes(flags.data_dir, flags.rgb_type, flags.gt_type, 'train',
flags.rgb_ext, flags.gt_ext, ['png', 'png'],
clc_name='{}_train'.format(flags.ds_name),
force_run=flags.force_run)
cm_valid = cityscapes_reader.CollectionMakerCityscapes(flags.data_dir, flags.rgb_type, flags.gt_type, 'val',
flags.rgb_ext, flags.gt_ext, ['png', 'png'],
clc_name='{}_valid'.format(flags.ds_name),
force_run=flags.force_run)
cm_train.print_meta_data()
cm_valid.print_meta_data()
resize_func = lambda img: resize_image(img, flags.tile_size)
train_init_op, valid_init_op, reader_op = dataReaderSegmentation.DataReaderSegmentationTrainValid(
flags.tile_size, cm_train.meta_data['file_list'], cm_valid.meta_data['file_list'],
flags.batch_size, cm_train.meta_data['chan_mean'], aug_func=[reader_utils.image_flipping_hori,
reader_utils.image_scaling_with_label],
random=True, has_gt=True, gt_dim=1, include_gt=True, valid_mult=flags.val_mult, global_func=resize_func)\
.read_op()
feature, label = reader_op
model.create_graph(feature)
model.load_resnet(flags.res_dir)
model.compile(feature, label, flags.n_train, flags.n_valid, flags.tile_size, ersaPath.PATH['model'],
par_dir=flags.model_par_dir, val_mult=flags.val_mult, loss_type='xent')
train_hook = hook.ValueSummaryHook(flags.verb_step, [model.loss, model.lr_op],
value_names=['train_loss', 'learning_rate'], print_val=[0])
model_save_hook = hook.ModelSaveHook(model.get_epoch_step()*flags.save_epoch, model.ckdir)
valid_loss_hook = hook.ValueSummaryHook(model.get_epoch_step(), [model.loss, model.loss_iou],
value_names=['valid_loss', 'valid_mIoU'], log_time=True,
run_time=model.n_valid, iou_pos=1)
image_hook = hook.ImageValidSummaryHook(model.input_size, model.get_epoch_step(), feature, label, model.output,
cityscapes_labels.image_summary, img_mean=cm_train.meta_data['chan_mean'])
start_time = time.time()
model.train(train_hooks=[train_hook, model_save_hook], valid_hooks=[valid_loss_hook, image_hook],
train_init=train_init_op, valid_init=valid_init_op)
print('Duration: {:.3f}'.format((time.time() - start_time)/3600))
def main(flags):
nn_utils.set_gpu(GPU)
# define network
model = stn.STN(flags.num_classes,
flags.tile_size,
suffix=flags.model_suffix,
learn_rate=flags.learning_rate,
decay_step=flags.decay_step,
decay_rate=flags.decay_rate,
epochs=flags.epochs,
batch_size=flags.batch_size)
cm_train = cityscapes_reader.CollectionMakerCityscapes(
flags.data_dir,
flags.rgb_type,
flags.gt_type,
'train',
flags.rgb_ext,
flags.gt_ext, ['png', 'png'],
clc_name='{}_train'.format(flags.ds_name),
force_run=flags.force_run)
cm_valid = cityscapes_reader.CollectionMakerCityscapes(
flags.data_dir,
flags.rgb_type,
flags.gt_type,
'val',
flags.rgb_ext,
flags.gt_ext, ['png', 'png'],
clc_name='{}_valid'.format(flags.ds_name),
force_run=flags.force_run)
cm_train.print_meta_data()
cm_valid.print_meta_data()
train_pred_dir = r'/home/lab/Documents/bohao/data/deeplab_model/vis_train/raw_segmentation_results'
valid_pred_dir = r'/home/lab/Documents/bohao/data/deeplab_model/vis/raw_segmentation_results'
file_list_train = get_image_list(flags.data_dir, train_pred_dir, 'train')
file_list_valid = get_image_list(flags.data_dir, valid_pred_dir, 'val')
resize_func = lambda img: resize_image(img, flags.tile_size)
train_init_op, valid_init_op, reader_op = DataReaderSegmentationTrainValid(
flags.tile_size, file_list_train, file_list_valid,
flags.batch_size, cm_train.meta_data['chan_mean'], aug_func=[reader_utils.image_flipping_hori],
random=True, has_gt=True, gt_dim=1, include_gt=True, valid_mult=flags.val_mult, global_func=resize_func)\
.read_op()
feature, label, pred = reader_op
train_init_op_valid, _, reader_op = DataReaderSegmentationTrainValid(
flags.tile_size, file_list_valid, file_list_train,
flags.batch_size, cm_valid.meta_data['chan_mean'], aug_func=[reader_utils.image_flipping_hori],
random=True, has_gt=True, gt_dim=1, include_gt=True, valid_mult=flags.val_mult, global_func=resize_func) \
.read_op()
_, _, pred_valid = reader_op
model.create_graph(pred, feature_valid=pred_valid, rgb=feature)
model.compile(pred,
label,
flags.n_train,
flags.n_valid,
flags.tile_size,
ersaPath.PATH['model'],
par_dir=flags.model_par_dir,
val_mult=flags.val_mult,
loss_type='xent')
train_hook = hook.ValueSummaryHook(
flags.verb_step, [
model.loss, model.g_loss, model.d_loss, model.lr_op[0],
model.lr_op[1], model.lr_op[2]
],
value_names=['seg_loss', 'g_loss', 'd_loss', 'lr_seg', 'lr_g', 'lr_d'],
print_val=[0, 1, 2])
model_save_hook = hook.ModelSaveHook(
model.get_epoch_step() * flags.save_epoch, model.ckdir)
valid_loss_hook = hook.ValueSummaryHookIters(
model.get_epoch_step(), [model.loss_xent, model.loss_iou],
value_names=['valid_loss', 'valid_mIoU'],
log_time=True,
run_time=model.n_valid)
image_hook = hook.ImageValidSummaryHook(
model.input_size,
model.get_epoch_step(),
feature,
label,
model.refine,
cityscapes_labels.image_summary,
img_mean=cm_train.meta_data['chan_mean'])
start_time = time.time()
model.train(train_hooks=[train_hook, model_save_hook],
valid_hooks=[valid_loss_hook, image_hook],
train_init=[train_init_op, train_init_op_valid],
valid_init=valid_init_op)
print('Duration: {:.3f}'.format((time.time() - start_time) / 3600))
def main(flags):
nn_utils.set_gpu(GPU)
# define network
model = deeplab.DeepLab(flags.num_classes,
flags.tile_size,
batch_size=flags.batch_size)
cm_train = cityscapes_reader.CollectionMakerCityscapes(
flags.data_dir,
flags.rgb_type,
flags.gt_type,
'train',
flags.rgb_ext,
flags.gt_ext, ['png', 'png'],
clc_name='{}_train'.format(flags.ds_name),
force_run=False)
cm_test = cityscapes_reader.CollectionMakerCityscapes(
flags.data_dir,
flags.rgb_type,
flags.gt_type,
'val',
flags.rgb_ext,
flags.gt_ext, ['png', 'png'],
clc_name='{}_valid'.format(flags.ds_name),
force_run=False)
cm_test.print_meta_data()
resize_func_train = lambda img: skimage.transform.resize(
img, flags.tile_size, mode='reflect')
resize_func_test = lambda img: skimage.transform.resize(
img,
cm_test.meta_data['tile_dim'],
order=0,
preserve_range=True,
mode='reflect')
init_op, reader_op = dataReaderSegmentation.DataReaderSegmentation(
flags.tile_size,
cm_test.meta_data['file_list'],
batch_size=flags.batch_size,
random=False,
chan_mean=cm_train.meta_data['chan_mean'],
is_train=False,
has_gt=True,
gt_dim=1,
include_gt=True,
global_func=resize_func_train).read_op()
estimator = nn_processor.NNEstimatorSegmentScene(
model,
cm_test.meta_data['file_list'],
flags.res_dir,
init_op,
reader_op,
ds_name='city_scapes',
save_result_parent_dir='Cityscapes',
gpu=flags.GPU,
score_result=True,
split_char='.',
post_func=resize_func_test,
save_func=make_general_id_map,
ignore_label=(-1, 255))
estimator.run(force_run=flags.force_run)