def main():
args = parse_args()
model_path = pjoin(cfg.checkpoints_folder, args.dataset, args.model)
params_file, params_dict = parse_model(model_path)
ctx = mx.gpu()
train_dataset, val_dataset, eval_metric = get_dataset(args.dataset)
net = get_model(params_dict['base_model'],
pretrained=True,
norm_layer=gluon.nn.BatchNorm)
net.reset_class(classes=train_dataset.classes)
net.initialize(force_reinit=True)
net.load_parameters(pjoin(model_path, params_file))
async_net = net
if args.type.lower() == 'ssd':
train_data, val_data = get_ssd_dataloader(
async_net,
train_dataset,
val_dataset,
params_dict['data_shape'],
params_dict['batch_size'],
args.num_workers,
bilateral_kernel_size=args.bilateral_kernel_size,
sigma_vals=args.sigma_vals,
grayscale=args.grayscale)
def main():
args = parse_args()
if args.ctx == 'gpu':
ctx = mx.gpu()
elif args.ctx == 'cpu':
ctx = mx.cpu()
else:
raise ValueError(f'ctx {args.ctx} invalid.')
dataset = args.dataset
model = args.model
_, val_ds, _ = get_dataset(dataset)
print(f'Prediction {dataset} dataset ({len(val_ds)} examples) with {model} model.')
net, trans = load_model(args.model_type, model, dataset)
if args.benchmark:
times = benchmark(net, val_ds, trans, ctx=ctx)
folder = pjoin(cfg.gen_data_folder, 'benchmark')
try:
os.makedirs(folder)
except OSError:
pass
with open(pjoin(folder, f'{model}_{dataset}_{args.ctx}.pkl'), 'wb') as f:
pickle.dump(times, f)
else:
preds, labels = get_predictions(net, val_ds, trans, ctx=ctx)
folder = pjoin(cfg.gen_data_folder, 'predictions')
try:
os.makedirs(folder)
except OSError:
pass
with open(pjoin(folder, f'{model}_{dataset}.pkl'), 'wb') as f:
pickle.dump({'preds': preds, 'labels': labels}, f)
def test_train_test_different(self):
for dataset_name in cfg.dataset_names:
trn_ds, val_ds, _ = get_dataset(dataset_name)
for fn1 in trn_ds.fns:
for fn2 in val_ds.fns:
self.assertNotEqual(fn1, fn2)
def test_split(self):
for dataset_name in cfg.dataset_names:
if 'mixed' not in dataset_name:
trn_ds, val_ds, _ = get_dataset(dataset_name)
trn, val = len(trn_ds), len(val_ds)
self.assertAlmostEqual(trn/(trn + val), 0.7, places=2)
from copy import deepcopy as copy
from easydict import EasyDict as edict
from sklearn import metrics
from sklearn.metrics import classification_report
from traindet import get_class_preds, build_confusion_matrix, calc_detection
import seaborn as sns
sns.set_style('whitegrid')
# %%
article_root = pjoin(cfg.data_folder, 'article')
out_folder = pjoin(article_root, 'tables')
outimg_folder = pjoin(article_root, 'figures')
#%%
_, val_ds_real, _ = get_dataset('real')
_, val_ds, _ = get_dataset('synth_small_bg')
print('Validation set sizes:', len(val_ds), len(val_ds_real))
# %%
models = 'ssd_default,yolo_default,frcnn_default'.split(',')
datasets = 'real,synth_small_printer,synth_small_nobg'.split(',')
datasets_form = ['Real', 'Synth', 'Synth No Bg.']
models_form = ['FRCNN', 'SSD', 'YOLOv3']
# %%
data = load_predictions('ssd_default,yolo_default,frcnn_default',
'real,synth_small_printer,synth_small_nobg')
save_params(net, logger, best_map, current_map, epoch, args.save_interval, args.save_prefix)
end_epoch_time = time.time()
logger.info('Epoch time {:.3f}'.format(end_epoch_time - start_epoch_time))
end_train_time = time.time()
logger.info('Train time {:.3f}'.format(end_train_time - start_train_time))
if __name__ == '__main__':
args = parse_args()
# fix seed for mxnet, numpy and python builtin random generator.
gutils.random.seed(args.seed)
# training contexts
ctx = [mx.gpu(int(i)) for i in args.gpus.split(',') if i.strip()]
ctx = ctx if ctx else [mx.cpu()]
train_dataset, val_dataset, eval_metric = get_dataset(args.dataset, args.mixup)
args.num_samples = len(train_dataset)
if args.transfer:
net_name = f'transfer_{args.data_shape}_{args.base_model}'
else:
net_name = f'scratch_{args.data_shape}_{args.base_model}'
args.save_prefix += net_name
if args.transfer:
net = get_model(args.base_model, pretrained=True)
else:
net = get_model(args.base_model, pretrained_base=True)
net.initialize()
if __name__ == '__main__':
import sys
sys.setrecursionlimit(1100)
args = parse_args()
# fix seed for mxnet, numpy and python builtin random generator.
gutils.random.seed(args.seed)
# training contexts
ctx = [mx.gpu(int(i)) for i in args.gpus.split(',') if i.strip()]
ctx = ctx if ctx else [mx.cpu()]
args.batch_size = len(ctx) # 1 batch per device
# training data
train_dataset, val_dataset, eval_metric = get_dataset(args.dataset,
args.mixup,
tclass=args.tclass)
if args.transfer:
net_name = f'transfer_{args.base_model}'
else:
net_name = f'scratch_{args.base_model}'
args.save_prefix += net_name
if args.transfer:
net = get_model(args.base_model, pretrained=True, ctx=ctx)
else:
net = get_model(args.base_model, pretrained_base=True, ctx=ctx)
net.reset_class(classes=train_dataset.classes)