def train(self):
num_epochs = self.hyperparams.epochs
start_epoch = self.hyperparams.start_epoch if hasattr(
self.hyperparams, 'start_epoch') else 0
for epoch in range(start_epoch, num_epochs):
if hasattr(self.conf, 'epoch'):
self.conf.epoch.value = epoch
print(self.conf.epoch)
self.model.train()
try:
self.conf.lr_scheduler.step()
print(
f'Elapsed time: {datetime.datetime.now() - self.experiment_start}'
)
for group in self.optimizer.param_groups:
print('LR: {:.4e}'.format(group['lr']))
eval_epoch = ((epoch % self.conf.eval_each == 0)
or (epoch == num_epochs - 1)) # and (epoch > 0)
self.model.criterion.step_counter = 0
print(f'Epoch: {epoch} / {num_epochs - 1}')
if eval_epoch and not self.args.dry:
print("Experiment dir: %s" % self.experiment_dir)
batch_iterator = iter(enumerate(self.loader_train))
start_t = perf_counter()
for step, batch in batch_iterator:
self.optimizer.zero_grad()
loss = self.model.loss(batch)
loss.backward()
self.optimizer.step()
if step % 80 == 0 and step > 0:
curr_t = perf_counter()
print(
f'{(step * self.conf.batch_size) / (curr_t - start_t):.2f}fps'
)
if not self.args.dry:
store(self.model, self.store_path, 'model')
store(self.optimizer, self.store_path, 'optimizer')
if eval_epoch and self.args.eval:
print('Evaluating model')
iou, per_class_iou = evaluate_semseg(
self.model, self.loader_val,
self.dataset_val.class_info)
self.validation_ious += [iou]
if self.args.eval_train:
print('Evaluating train')
evaluate_semseg(self.model, self.loader_train,
self.dataset_train.class_info)
if iou > self.best_iou:
self.best_iou = iou
self.best_iou_epoch = epoch
if not self.args.dry:
copy(self.store_path.format('model'),
self.store_path.format('model_best'))
print(
f'Best mIoU: {self.best_iou:.2f}% (epoch {self.best_iou_epoch})'
)
except KeyboardInterrupt:
break
def eval(config='configs/pyramid.py'):
conf_path = Path(config)
conf = import_module(config)
class_info = conf.dataset_val.class_info
model = conf.model.cuda()
for loader, name in conf.eval_loaders:
iou_acc = evaluate_semseg(model,
loader,
class_info,
observers=conf.eval_observers)
print(f'{name}: {iou_acc:.2f}')
# Note: we are expecting only one loader (val)
return iou_acc
import argparse
from pathlib import Path
import importlib.util
from evaluation import evaluate_semseg
def import_module(path):
spec = importlib.util.spec_from_file_location("module", path)
module = importlib.util.module_from_spec(spec)
spec.loader.exec_module(module)
return module
parser = argparse.ArgumentParser(description='Detector train')
parser.add_argument('config', type=str, help='Path to configuration .py file')
parser.add_argument('--profile', dest='profile', action='store_true', help='Profile one forward pass')
if __name__ == '__main__':
args = parser.parse_args()
conf_path = Path(args.config)
conf = import_module(args.config)
class_info = conf.dataset_val.class_info
model = conf.model.cuda()
for loader, name in conf.eval_loaders:
iou, per_class_iou = evaluate_semseg(model, loader, class_info, observers=conf.eval_observers)
print(f'{name}: {iou:.2f}')
BasicBlock = BasicBlock34
else:
BasicBlock = BasicBlock_SN
store_dir = Path(args.store_dir)
store_stats_path = store_dir / "params_mious.txt"
os.makedirs(str(store_dir), exist_ok=True)
store_data(store_stats_path, "Params", "mIoU", "GMACs", "FPS")
pruning_percentages = conf.pruning_percentages
print("PRUNING PERCENTAGES: ", pruning_percentages)
pruning_indices = reset_pruning_indices(conf.model, BasicBlock)
print("INITIAL TESTING:")
print("MODEL:\n", conf.model)
iou, per_class_iou = evaluate_semseg(conf.model.cuda(), conf.loader_val,
conf.dataset_val.class_info)
params, macs, fps = measure_gmacs_fps()
store_data(store_stats_path, params, iou, macs, fps)
for i in range(10):
if i != 0:
iou, per_class_iou = evaluate_semseg(conf.model.cuda(),
conf.loader_val,
conf.dataset_val.class_info)
conf.model.to("cpu")
if conf.prune_mode == "rewind":
model, prune_stats, _ = prune_with_rewinding()
elif conf.prune_mode == "reset":
model, prune_stats = prune_with_resetting()
else: