def validate(self):
model.eval()
with torch.no_grad():
iou_epoch = 0.
iou_NN_epoch = 0.
num_batches = 0
loss_epoch = 0.
# Validation loop
for i, data in enumerate(tqdm(dataloader_val), 0):
# data creation
tgt_odms = data['odms'].to(args.device)
tgt_voxels = data['voxels'].to(args.device)
inp_voxels = down_sample(tgt_voxels)
inp_odms = []
for voxel in inp_voxels:
inp_odms.append(kal.rep.voxel.extract_odms(voxel).unsqueeze(0))
inp_odms = torch.cat(inp_odms)
# inference
initial_odms = upsample_omd(inp_odms)*2
distance = 30 - initial_odms
pred_odms_update = model(inp_odms)
pred_odms_update = pred_odms_update * distance
pred_odms = initial_odms + pred_odms_update
# losses
loss = loss_fn(pred_odms, tgt_odms)
loss_epoch += float(loss.item())
NN_pred = up_sample(inp_voxels)
iou_NN = kal.metrics.voxel.iou(NN_pred.contiguous(), tgt_voxels)
iou_NN_epoch += iou_NN
pred_odms = pred_odms.int()
pred_voxels = []
for odms, voxel_NN in zip(pred_odms, NN_pred):
pred_voxels.append(kal.rep.voxel.project_odms(odms,voxel =voxel_NN, votes = 2).unsqueeze(0))
pred_voxels = torch.cat(pred_voxels)
iou = kal.metrics.voxel.iou(pred_voxels.contiguous(), tgt_voxels)
iou_epoch += iou
# logging
num_batches += 1
if i % args.print_every == 0:
out_iou = iou_epoch.item() / float(num_batches)
out_iou_NN = iou_NN_epoch.item() / float(num_batches)
tqdm.write(f'[VAL] Epoch {self.cur_epoch:03d}, Batch {i:03d}: IoU: {out_iou}, Iou Base: {out_iou_NN}')
out_iou = iou_epoch.item() / float(num_batches)
out_iou_NN = iou_NN_epoch.item() / float(num_batches)
tqdm.write(f'[VAL Total] Epoch {self.cur_epoch:03d}, Batch {i:03d}: IoU: {out_iou}, Iou Base: {out_iou_NN}')
loss_epoch = loss_epoch / num_batches
self.val_loss.append(out_iou)
def train(self):
loss_epoch = 0.
num_batches = 0
diff = 0
model.train()
# Train loop
for i, data in enumerate(tqdm(dataloader_train), 0):
optimizer.zero_grad()
# data creation
tgt_odms = data['odms'].to(args.device)
tgt_voxels = data['voxels'].to(args.device)
inp_voxels = down_sample(tgt_voxels)
inp_odms = []
for voxel in inp_voxels:
inp_odms.append(kal.rep.voxel.extract_odms(voxel).unsqueeze(0))
inp_odms = torch.cat(inp_odms)
# inference
initial_odms = upsample_omd(inp_odms)*2
distance = 30 - initial_odms
pred_odms_update = model(inp_odms)
pred_odms_update = pred_odms_update * distance
pred_odms = initial_odms + pred_odms_update
pred_odms = initial_odms + pred_odms_update
# losses
loss = loss_fn(pred_odms, tgt_odms)
loss.backward()
loss_epoch += float(loss.item())
# logging
num_batches += 1
if i % args.print_every == 0:
tqdm.write(f'[TRAIN] Epoch {self.cur_epoch:03d}, Batch {i:03d}: Loss: {float(loss.item())}')
optimizer.step()
loss_epoch = loss_epoch / num_batches
self.train_loss.append(loss_epoch)
self.cur_epoch += 1
def train(self):
loss_epoch = 0.
num_batches = 0
diff = 0
model.train()
# Train loop
for i, data in enumerate(tqdm(dataloader_train), 0):
optimizer.zero_grad()
# data creation
tgt_odms = data['odms_128'].to(args.device)
inp_odms = data['odms_32'].to(args.device)
# inference
initial_odms = upsample_omd(inp_odms) * 4
distance = 128 - initial_odms
pred_odms_update = model(inp_odms)
pred_odms_update = pred_odms_update * distance
pred_odms = initial_odms + pred_odms_update
# losses
loss = loss_fn(pred_odms, tgt_odms)
loss.backward()
loss_epoch += float(loss.item())
# logging
num_batches += 1
if i % args.print_every == 0:
tqdm.write(
f'[TRAIN] Epoch {self.cur_epoch:03d}, Batch {i:03d}: Loss: {float(loss.item())}'
)
optimizer.step()
loss_epoch = loss_epoch / num_batches
self.train_loss.append(loss_epoch)
self.cur_epoch += 1
model_res.eval()
model_occ.eval()
with torch.no_grad():
for data in tqdm(dataloader_val):
tgt_odms = data['odms'].to(args.device)
tgt_voxels = data['voxels'].to(args.device)
inp_voxels = down_sample(tgt_voxels)
inp_odms = []
for voxel in inp_voxels:
inp_odms.append(kal.rep.voxel.extract_odms(voxel).unsqueeze(0))
inp_odms = torch.cat(inp_odms)
# inference res
initial_odms = upsample_omd(inp_odms) * 2
distance = 30 - initial_odms
pred_odms_update = model_res(inp_odms)
pred_odms_update = pred_odms_update * distance
pred_odms_res = initial_odms + pred_odms_update
# inference occ
pred_odms_occ = model_occ(inp_odms)
# combine
pred_odms_res = pred_odms_res.int()
ones = pred_odms_occ > .5
zeros = pred_odms_occ <= .5
pred_odms_occ[ones] = pred_odms_occ.shape[-1]
pred_odms_occ[zeros] = 0