def evaluate(data_source):
model.eval()
PreNN, PreNNT, PreNNDR = 0, 0, 0
with torch.no_grad():
for i_batch, sample_batched in enumerate(data_source, 1):
batch = parse_batch(sample_batched, device)
TPNN, PNN, TPNNT, PNNT, TPNNDR, PNNDR = eval_model(
model.module if mgpu else model,
batch,
cfg.MODEL.DES_THRSH,
cfg.MODEL.COO_THRSH,
)
PreNN += TPNN / PNN
PreNNT += TPNNT / PNNT
PreNNDR += TPNNDR / PNNDR
length = len(data_source)
PreNN, PreNNT, PreNNDR = (PreNN / length, PreNNT / length,
PreNNDR / length)
meanms = (PreNN + PreNNT + PreNNDR) / 3
checkpoint_name = (
f"NN_{PreNN:.3f}_NNT_{PreNNT:.3f}_NNDR_{PreNNDR:.3f}_MeanMS_{meanms:.3f}"
)
return checkpoint_name, meanms
model = pretrainedmodels.__dict__[model_name](num_classes=1000,
pretrained='imagenet')
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
def initialize_model(model, feature_extract):
num_ftrs = model.last_linear.in_features
model.last_linear = nn.Linear(num_ftrs, num_classes)
return model
image_loader_dict = eval_utils.get_image_loader(train_list, val_list, model,
label_num, num_classes,
batch_size)
model_ft = initialize_model(model, feature_extract)
#model_ft = nn.DataParallel(model_ft)
model_ft = model_ft.to(device)
model_ft.load_state_dict(torch.load(model_path))
best_model = eval_utils.eval_model(num_classes,
model_name,
label_map,
model_ft,
image_loader_dict,
device=device,
is_inception=(model_name == "inception"))