def resnet50_with_cam(
num_classes,
cam_layer,
state_dict=image_net_state_dict(),
):
return resnet50(num_classes=num_classes,
state_dict=state_dict,
resnet_class=BGCamResnet50,
cam_layer=cam_layer)
wandb.log(
common.merge(log(is_train=True, saver=train_saver),
log(is_train=False, saver=validation_saver)))
if __name__ == '__main__':
parser = default_parser(run_name='BinaryLabel Resnet50',
tags=['binary', 'train', 'baseline'])
parser.add_argument('--label',
type=str,
choices=common.LABELS,
required=False,
default='pigment_network',
help='Label for train')
args = parser.parse_args()
model = resnet50(num_classes=1, state_dict=image_net_state_dict())
common.init_wandb(config=args, name=args.name, tags=args.tags)
args = vars(args)
args['only_labels'] = [args['label']]
exp = BinaryLabelExp(model=model, output_labels=['model_output'], **args)
wandb.config.update({'seed': exp.seed}, allow_val_change=True)
exp.run()
log_summary_common()
def main(cfg: RunConfig):
torch.set_num_threads(4)
images_dir = resource_path(config.prop(f'images{cfg.images_size}'),
strict=True)
args, name, tags = get_wandb_starup(cfg)
print('Got args', args)
print('Got name', name)
print('Got tags', tags)
device = resolve_device(cfg.device)
print(f'Got device {device}')
cl_criterion = resolve_criterion(cfg.label, cfg.cl_criterion, device)
a_criterion = resolve_criterion(cfg.label, cfg.attention_criterion, device)
ds_index = DsIndex(images_dir=images_dir, masks_dir=config.prop('masks'))
full_ds = FullDataset(ds_index=ds_index,
image_size=cfg.images_size,
labels=[cfg.label])
samples = json.load(open(resource_path(config.prop('sample_indices'))))
if cfg.train_on_test:
samples[PHASE_TRAIN] = samples[PHASE_TEST]
for l in LABELS:
samples['balance'][l][PHASE_TRAIN] = samples['balance'][l][
PHASE_TEST]
if cfg.balanced:
samples[PHASE_TRAIN] = samples['balance'][cfg.label][PHASE_TRAIN]
print('Using balance samples')
for p in PHASES:
print(len(samples[p]), p, 'samples',
'balanced' if p == PHASE_TRAIN else '')
if cfg.neg_percent != 100:
samples[PHASE_TRAIN] = samples[f'neg_{cfg.neg_percent}'][
cfg.label][PHASE_TRAIN]
datasets = {}
for p in [PHASE_TRAIN, PHASE_TEST, PHASE_VALIDATION]:
datasets[p] = BinaryDataset(full_ds,
sample_ids=samples[p],
label=cfg.label)
model = resnet50_with_cam(num_classes=1,
state_dict=image_net_state_dict(),
cam_layer=cfg.cam_layer)
optimizer = Adam(model.parameters(), lr=cfg.lr)
scheduler = CosineAnnealingLR(optimizer, T_max=5, eta_min=0.005)
if cfg.no_scheduler:
scheduler = None
suspector = ImageSuspector(max_samples_by_label=5)
wandb_help.create_run(args, name, tags=tags, offline=False)
run_experiment(
datasets,
cl_criterion,
a_criterion,
model.to(device),
optimizer,
scheduler,
device,
suspector,
cfg,
)
suspector.log_to_wandb(full_ds)
wandb_help.log_summary_best()
for ds in datasets.values():
ds.clean()
def main():
"""
Input arguments
usage: binary_gradcam_label_train.py [-h] [--images-dir IMAGES_DIR] [--use_cpu] [--name NAME] [--image-resize IMAGE_RESIZE] [--tags TAGS [TAGS ...]] [--epochs EPOCHS]
[--lr LR] [--batch_size BATCH_SIZE] [--seed SEED] [--no-warmup] [--no-scheduler]
[--label {negative_network,streaks,milia_like_cyst,globules,pigment_network}] [--cam_layer CAM_LAYER] [--lambda LOSS_LAMBDA]
[--no-attention-loss] [--when-present] [--ae-loss-bce]
optional arguments:
-h, --help show this help message and exit
--images-dir IMAGES_DIR
Location of train images, default value present in config.py (default: /Users/ifkbhit/itmo/diploma/resources/isic_train/)
--use_cpu Use cpu only (default: False)
--name NAME WanDB run name (default: Resnet50)
--image-resize IMAGE_RESIZE
Image resize (default: 256)
--tags TAGS [TAGS ...]
Additional tags to run (default: ['binary', 'train'])
--epochs EPOCHS Num epochs (default: 20)
--lr LR Learning rate (default: 0.001)
--batch_size BATCH_SIZE
Batch size (default: 10)
--seed SEED Int or 'none', or 'rand'. 'none' - do nothing with seed, 'rand' - generate random seed (default: rand)
--no-warmup Disable cache warmup (default: False)
--no-scheduler Don't use scheduler (default: False)
--label {negative_network,streaks,milia_like_cyst,globules,pigment_network}
Label for train (default: pigment_network)
--cam_layer CAM_LAYER
Layer for gradcam extractions (default: layer4.2)
--lambda LOSS_LAMBDA Coefficient for loss: classification_loss + lambda * attention_loss (default: 1)
--no-attention-loss Run without attention loss calculation (default: False)
--when-present Calculate attention loss only if sample has attribute (default: False)
--ae-loss-bce Use bce as attention loss (default: False)
"""
parser = default_parser(run_name='Resnet50', tags=['binary', 'train'])
parser.add_argument('--label',
type=str,
choices=common.LABELS,
required=False,
default='pigment_network',
help='Label for train')
parser.add_argument('--cam_layer',
type=str,
required=False,
default='layer4.2',
help='Layer for gradcam extractions')
parser.add_argument(
'--lambda',
type=int,
required=False,
default=1,
dest='loss_lambda',
help=
'Coefficient for loss: classification_loss + lambda * attention_loss')
parser.add_argument('--no-attention-loss',
action='store_true',
required=False,
help='Run without attention loss calculation')
parser.add_argument(
'--when-present',
action='store_true',
required=False,
help='Calculate attention loss only if sample has attribute')
parser.add_argument('--ae-loss-bce',
required=False,
action='store_true',
help='Use bce as attention loss')
args = parser.parse_args()
if args.no_attention_loss:
run_type = "baseline"
else:
run_type = "gradcam"
run_name = f'[{args.label}][{run_type}] {args.name}; lambda={args.loss_lambda}; lr={args.lr}'
bgcam_resnet = resnet50_with_cam(num_classes=1,
state_dict=image_net_state_dict(),
cam_layer=args.cam_layer)
tags = build_tags(args, run_type)
common.init_wandb(config=args, name=run_name, tags=tags)
d_args = vars(args)
# обучение происходит по одному аттрибуту
d_args['only_labels'] = [args.label]
exp = BinaryGradCamLabelExp(model=bgcam_resnet,
output_labels=['model_output', 'ac'],
**d_args)
# Использовать BCE, вместо MSE
if args.ae_loss_bce:
exp.criterion_e = nn.BCEWithLogitsLoss()
# Не считать attention loss, т.е. запустить baseline
exp.calculate_attention_error_loss = not args.no_attention_loss
# веса для классификационной BCEWithLogitsLoss
pos_weights = torch.tensor([BCE_WEIGHTS[args.label]])
exp.pos_weights = pos_weights
print("Got pos_weights =", pos_weights)
# не считать attention loss, если нет дерм. признака на фото(маска пустая)
exp.when_present = args.when_present
# номера примеров из датасета, на которых стоит сохранить визуализацию
exp.suspect_attention = {
# for train phase
True: [
13,
16,
17,
19,
20, # with label
11,
14,
23,
24,
26 # without label
],
# for validation phase
False: [
49,
51,
55,
56,
63, # with label
40,
45,
92,
99,
110 # without label
]
}
# лямбда параметр для attention loss
exp.loss_lambda = args.loss_lambda
wandb.config.update({
'seed': exp.seed,
}, allow_val_change=True)
# запуск обучения
exp.run()
print('Logging summary')
# логгирование результатов в wandb summary section
log_summary_common()
print('Logging images')
# логгирование визуализации
to_pil = transforms.ToPILImage()
for sample_key, epochs_maps in exp.suspects.items():
images = []
for epoch in sorted(epochs_maps.keys()):
if epoch == -1:
images.append(
wandb.Image(to_pil(epochs_maps[epoch]),
caption='Orig_mask'))
continue
data = epochs_maps[epoch]
img_t = data['map'].unsqueeze(0)
img_t = F.interpolate(img_t, (exp.image_resize, exp.image_resize),
mode='bilinear',
align_corners=True)
img_t = img_t[0]
capt = f'epoch_{epoch + 1}_{"correct" if data["correct"] else "incorrect"}'
images.append(wandb.Image(to_pil(img_t), caption=capt))
wandb.log({sample_key: images})