if params.dataset in ["ISIC"]:
datamgr = ISIC_few_shot.SimpleDataManager(image_size, batch_size = 64)
data_loader = datamgr.get_data_loader(aug = False )
elif params.dataset in ["EuroSAT"]:
datamgr = EuroSAT_few_shot.SimpleDataManager(image_size, batch_size = 64)
data_loader = datamgr.get_data_loader(aug = False )
elif params.dataset in ["CropDisease"]:
datamgr = CropDisease_few_shot.SimpleDataManager(image_size, batch_size = 64)
data_loader = datamgr.get_data_loader(aug = False )
elif params.dataset in ["ChestX"]:
datamgr = Chest_few_shot.SimpleDataManager(image_size, batch_size = 64)
data_loader = datamgr.get_data_loader(aug = False )
model = model_dict[params.model]()
model = model.cuda()
tmp = torch.load(modelfile)
state = tmp['state']
state_keys = list(state.keys())
for i, key in enumerate(state_keys):
if "feature." in key:
newkey = key.replace("feature.","") # an architecture model has attribute 'feature', load architecture feature to backbone by casting name from 'feature.trunk.xx' to 'trunk.xx'
state[newkey] = state.pop(key)
else:
n_query=15,
**few_shot_params)
novel_loader = datamgr.get_data_loader(num_aug=params.gen_examples)
elif params.test_dataset == "CropDisease":
print("Loading CropDisease")
datamgr = CropDisease_few_shot.SetDataManager2(image_size,
n_eposide=iter_num,
n_query=15,
**few_shot_params)
novel_loader = datamgr.get_data_loader(num_aug=params.gen_examples)
elif params.test_dataset == "ChestX":
print("Loading ChestX")
datamgr = Chest_few_shot.SetDataManager2(image_size,
n_eposide=iter_num,
n_query=15,
**few_shot_params)
novel_loader = datamgr.get_data_loader(
num_aug=params.gen_examples) ### what if aug is true???
## uncomment code below to see if code is same across loaders
#########################################################################
acc_all = []
start_epoch = params.start_epoch
stop_epoch = params.stop_epoch
print(freeze_backbone)
# replace finetine() with your own method
iter_num = 600
1, int(16 * params.test_n_way / params.train_n_way)
) #if test_n_way is smaller than train_n_way, reduce n_query to keep batch size small
train_few_shot_params = dict(n_way=params.train_n_way,
n_support=params.n_shot)
test_few_shot_params = dict(n_way=params.test_n_way,
n_support=params.n_shot)
#if params.dataset == "miniImageNet":
base_datamgr = miniImageNet_few_shot.SetDataManager(
image_size, n_query=n_query, **train_few_shot_params)
base_loader = base_datamgr.get_data_loader(aug=params.train_aug)
# use unlabeled data from these novel domains for adversarial domain adaptation
# TODO: since the data is unlabeled, we need to modify data manager / data loader
if params.dataset == "ChestX":
target_datamgr = Chest_few_shot.SetDataManager(
image_size, n_query=n_query, **train_few_shot_params)
elif params.dataset == "EuroSAT":
target_datamgr = EuroSAT_few_shot.SetDataManager(
image_size, n_query=n_query, **train_few_shot_params)
elif params.dataset == "ISIC2018":
target_datamgr = ISIC_few_shot.SetDataManager(
image_size, n_query=n_query, **train_few_shot_params)
elif params.dataset == "CropDiseases":
target_datamgr = CropDisease_few_show.SetDataManager(
image_size, n_query=n_query, **train_few_shot_params)
else:
raise ValueError('Unknown dataset')