def load_checkpoint(name):
check = torch.load("../models/" + model, map_location="cuda")
loss = check["val_best_loss"]
model = se_resnet34(num_classes=2, multi_output=True).to(device)
model = densenet121(if_selayer=True).to(device)
model.load_state_dict(check["model"])
model = model.to(device)
model.eval()
return model
size = args.size
cutmix_prob = 0.5
mixup_prob = 0.1
cutout_prob = 0.5
random_erasing_prob = 0.5
print("Running device: ", device)
print("batchsize: ", batchsize)
efficientnet_name = ""
efficientnets = ["efficientnet-b" + str(i) for i in range(8)]
if args.model == "resnet34":
# model = se_resnet34(num_classes=2).to(device)
model = se_resnet34(num_classes=2, multi_output=True).to(device)
elif args.model == "bengali_resnet34":
model = model_bengali.se_resnet34(num_classes=2,
multi_output=True).to(device)
elif args.model == "bengali_resnext50":
model = model_bengali.se_resnext50_32x4d(num_classes=2,
multi_output=True).to(device)
elif args.model == "resnet152":
model = se_resnet152(num_classes=2, multi_output=True).to(device)
elif args.model == "resnext50":
model = se_resnext50_32x4d(num_classes=2, multi_output=True).to(device)
elif args.model == "resnext101":
model = se_resnext101_32x8d(num_classes=2, multi_output=True).to(device)
elif args.model == "densenet":
model = densenet121(if_selayer=True).to(device)
elif args.model == "inception_v3":
transforms = transforms.Compose(
[transforms.ToPILImage(mode=None),
transforms.ToTensor()])
# load
test_all = load_test_df()
test_dataset = BengalDataset(df=test_all,
transform=transforms,
test_dataset_flag=True)
test_loader = torch.utils.data.DataLoader(test_dataset,
batch_size=batchsize,
shuffle=False)
model = se_resnet34(num_classes=2).to(device)
loss_fn = torch.nn.CrossEntropyLoss()
# model load
fn = "../models/model.dat"
checkpoint = torch.load(fn)
model.load_state_dict(checkpoint["model"])
model = model.to(device)
print("model loaded.")
model.eval()
pred1, pred2, pred3 = [], [], []
with torch.no_grad():
print("start inference...")
for idx, (inputs, _, _, _) in tqdm(enumerate(test_loader),
lr = args.lr
device = "cuda:0" if torch.cuda.is_available() else "cpu"
height = 137
width = 236
size = 128
cutmix_prob = 0.1
mixup_prob = 0.1
cutout_prob = 0.5
random_erasing_prob = 0.5
print("Running device: ", device)
# train model for label2
if args.model == "resnet34":
# model = se_resnet34(num_classes=2).to(device)
model = se_resnet34(num_classes=168, multi_output=False).to(device)
elif args.model == "resnet152":
model = se_resnet152(num_classes=168, multi_output=False).to(device)
elif args.model == "densenet":
mode2 = densenet121(if_selayer=True).to(device)
elif args.model == "resnext50":
model = se_resnext50_32x4d(num_classes=168, multi_output=True).to(device)
elif args.model == "resnext101":
model = se_resnext101_32x8d(num_classes=168, multi_output=True).to(device)
else:
raise ValueError()
# train_all = load_train_df()
_, vowels, graphemes, consonants = load_pickle_images()
imgs = np.asarray(pd.read_pickle(os.path.join(data_folder,
"cropped_imgs.pkl")))
seed = args.seed
epoch_num = args.epoch
batchsize = args.batchsize
lr = args.lr
device = "cuda:0" if torch.cuda.is_available() else "cpu"
height = 137
width = 236
cutmix_prob = 0.1
mixup_prob = 0.1
print("Running device: ", device)
# train models for each label
if args.model == "resnet34":
# model = se_resnet34(num_classes=2).to(device)
model1 = se_resnet34(num_classes=11, multi_output=False).to(device)
model2 = se_resnet34(num_classes=168, multi_output=False).to(device)
model3 = se_resnet34(num_classes=7, multi_output=False).to(device)
elif args.model == "resnet152":
model1 = se_resnet152(num_classes=11, multi_output=False).to(device)
model2 = se_resnet152(num_classes=168, multi_output=False).to(device)
model3 = se_resnet152(num_classes=7, multi_output=False).to(device)
elif args.model == "densenet":
model = densenet121(if_selayer=True).to(device)
mode2 = densenet121(if_selayer=True).to(device)
mode3 = densenet121(if_selayer=True).to(device)
else:
raise ValueError()
# train_all = load_train_df()
_, vowels, graphemes, consonants = load_pickle_images()