def run(self):
model = TopkNet(self.cfg).to(Global.DEVICE)
Global.load_weight(model, self.W_PATH)
model.eval()
future_coords_offsets_pd = []
timestamps = []
agent_ids = []
confs = []
with torch.no_grad():
dataiter = tqdm(self.test_dataloader)
for data in dataiter:
inputs = data["image"].to(Global.DEVICE)
images = []
samples_means, means, mixture_weights = model(inputs)
fit_outputs = torch.stack([mean for mean in means], dim=1)
fit_confidences = torch.stack(
[mixture_weight for mixture_weight in mixture_weights],
dim=1).squeeze()
outputs = torch.zeros(fit_outputs.size(0), fit_outputs.size(1),
fit_outputs.size(2), 2).to(Global.DEVICE)
conf = []
one_hot_el = torch.eye(3, 3)
for i in range(fit_confidences.size(0)):
idx = torch.argmax(fit_confidences[i]).item()
conf.append(one_hot_el[idx])
outputs[:, 0] = Utils.map_writer_from_image_to_world(
data, fit_outputs[:, 0, :, :], self.cfg)
outputs[:, 1] = Utils.map_writer_from_image_to_world(
data, fit_outputs[:, 1, :, :], self.cfg)
outputs[:, 2] = Utils.map_writer_from_image_to_world(
data, fit_outputs[:, 2, :, :], self.cfg)
future_coords_offsets_pd.append(outputs.cpu().numpy().copy())
timestamps.append(data["timestamp"].numpy().copy())
agent_ids.append(data["track_id"].numpy().copy())
confs.append(fit_confidences.cpu().numpy().copy())
if (len(confs) == 10): break
write_pred_csv(f'{Global.MULTI_MODE_SUBMISSION}',
timestamps=np.concatenate(timestamps),
track_ids=np.concatenate(agent_ids),
coords=np.concatenate(future_coords_offsets_pd),
confs=np.concatenate(confs))
def run(self):
model = TopkNet(self.cfg).to(Global.DEVICE)
Global.load_weight(model, self.W_PATH)
optimizer = optim.Adam(model.parameters(), lr=1e-6)
progress_bar = tqdm(range(self.cfg["train_params"]["max_num_steps"]))
losses_train = []
prelosses_train = []
straight_it = iter(self.straight_train_dataloader)
for itr in progress_bar:
data = next(straight_it)
model.train()
torch.set_grad_enabled(True)
inputs = data["image"].to(Global.DEVICE)
targets = Utils.map_writer_from_world_to_image(data, self.cfg).to(
Global.DEVICE)
target_availabilities = data["target_availabilities"].to(
Global.DEVICE)
samples_means, means, mixture_weights = model(inputs)
fit_outputs = torch.stack([mean for mean in means], dim=1)
fit_confidences = torch.stack(
[mixture_weight for mixture_weight in mixture_weights],
dim=1).squeeze()
if (itr <= 0):
if (itr <= 20):
loss = Loss.sampling_loss(targets, samples_means,
target_availabilities)
elif (itr <= 40):
loss = Loss.sampling_loss(targets, samples_means,
target_availabilities,
"epe-top-n", 40)
elif (itr <= 60):
loss = Loss.sampling_loss(targets, samples_means,
target_availabilities,
"epe-top-n", 20)
elif (itr <= 80):
loss = Loss.sampling_loss(targets, samples_means,
target_availabilities,
"epe-top-n", 10)
elif (itr <= 100):
loss = Loss.sampling_loss(targets, samples_means,
target_availabilities,
"epe-top-n", 5)
elif (itr <= 120):
loss = Loss.sampling_loss(targets, samples_means,
target_availabilities, "epe")
for i, param in enumerate(model.parameters()):
if i >= 68:
param.requires_grad = False
elif (itr <= 200):
loss, pre_loss = Loss.fitting_loss(targets, fit_outputs,
fit_confidences,
target_availabilities)
for i, param in enumerate(model.parameters()):
if i >= 68:
param.requires_grad = True
elif i < 68:
param.requires_grad = False
elif (itr <= 300):
loss, pre_loss = Loss.fitting_loss(targets, fit_outputs,
fit_confidences,
target_availabilities)
for i, param in enumerate(model.parameters()):
param.requires_grad = True
optimizer.zero_grad()
loss.backward()
optimizer.step()
losses_train.append(loss.item())
if (itr <= 100):
progress_bar.set_description(
f"loss: {loss.item()} loss(avg): {np.mean(losses_train[-100:])}"
)
else:
prelosses_train.append(pre_loss.item())
progress_bar.set_description(
f"pre_loss: {pre_loss.item()} loss(avg): {np.mean(prelosses_train[-100:])}"
)
if (itr + 1) % self.cfg['train_params'][
'checkpoint_every_n_steps'] == 0 and not Global.DEBUG:
torch.save(model.state_dict(),
f"./output/param/topk_model_state_{itr}.pth")
