def test_sample(sample, compute_metrics=True):
model.eval()
imgL, imgR, disp_gt = sample['left'], sample['right'], sample['disparity']
imgL = imgL.cuda()
imgR = imgR.cuda()
disp_gt = disp_gt.cuda()
disp_ests = model(imgL, imgR)
mask = (disp_gt < args.maxdisp) & (disp_gt > 0)
loss = model_loss(disp_ests, disp_gt, mask)
scalar_outputs = {"loss": loss}
image_outputs = {"disp_est": disp_ests, "disp_gt": disp_gt, "imgL": imgL, "imgR": imgR}
scalar_outputs["D1"] = [D1_metric(disp_est, disp_gt, mask) for disp_est in disp_ests]
scalar_outputs["EPE"] = [EPE_metric(disp_est, disp_gt, mask) for disp_est in disp_ests]
scalar_outputs["Thres1"] = [Thres_metric(disp_est, disp_gt, mask, 1.0) for disp_est in disp_ests]
scalar_outputs["Thres2"] = [Thres_metric(disp_est, disp_gt, mask, 2.0) for disp_est in disp_ests]
scalar_outputs["Thres3"] = [Thres_metric(disp_est, disp_gt, mask, 3.0) for disp_est in disp_ests]
if compute_metrics:
image_outputs["errormap"] = [disp_error_image_func()(disp_est, disp_gt) for disp_est in disp_ests]
return tensor2float(loss), tensor2float(scalar_outputs), image_outputs
def train_sample(sample, compute_metrics=False):
model.train()
imgL, imgR, disp_gt = sample['left'], sample['right'], sample['disparity']
imgL = imgL.cuda()
imgR = imgR.cuda()
disp_gt = disp_gt.cuda()
optimizer.zero_grad()
disp_ests = model(imgL, imgR)
mask = (disp_gt < args.maxdisp) & (disp_gt > 0)
loss = model_loss(disp_ests, disp_gt, mask)
scalar_outputs = {"loss": loss}
image_outputs = {"disp_est": disp_ests, "disp_gt": disp_gt, "imgL": imgL, "imgR": imgR}
if compute_metrics:
with torch.no_grad():
image_outputs["errormap"] = [disp_error_image_func()(disp_est, disp_gt) for disp_est in disp_ests]
scalar_outputs["EPE"] = [EPE_metric(disp_est, disp_gt, mask) for disp_est in disp_ests]
scalar_outputs["D1"] = [D1_metric(disp_est, disp_gt, mask) for disp_est in disp_ests]
# scalar_outputs["Thres1"] = [Thres_metric(disp_est, disp_gt, mask, 1.0) for disp_est in disp_ests]
# scalar_outputs["Thres2"] = [Thres_metric(disp_est, disp_gt, mask, 2.0) for disp_est in disp_ests]
# scalar_outputs["Thres3"] = [Thres_metric(disp_est, disp_gt, mask, 3.0) for disp_est in disp_ests]
loss.backward()
optimizer.step()
return tensor2float(loss), tensor2float(scalar_outputs), image_outputs
# 辞書型変数にまとめる
dataloaders_dict = {"train": train_dataloader, "val": val_dataloader}
# 動作の確認
batch_iterator = iter(dataloaders_dict["val"]) # イタレータに変換
images, targets = next(batch_iterator) # 1番目の要素を取り出す
print(images.size())
print("batch len is ", len(targets))
print(targets[0].shape) # ミニバッチのサイズのリスト、各要素は[n, 5]、nは物体数
# set model
model = fc_model([400, 200, 100, 50, 100, 200, 300, 400])
print(model)
# define loss
criterion = model_loss()
# define optimizer. use SGD here.
optimizer = optim.SGD(model.parameters(),
lr=args.lr,
momentum=0.9,
weight_decay=0.0001)
######### start training ##########
# enable GPUs if any.
device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
torch.backends.cudnn.benchmark = True
model.to(device)
model.train()