added a guard for hypernet training that will stop early if weights are getting no gradients

This commit is contained in:
AUTOMATIC 2022-10-22 14:48:43 +03:00
parent 1cd3ed7def
commit 7fd90128eb

View File

@ -310,6 +310,8 @@ def train_hypernetwork(hypernetwork_name, learn_rate, batch_size, data_root, log
scheduler = LearnRateScheduler(learn_rate, steps, ititial_step) scheduler = LearnRateScheduler(learn_rate, steps, ititial_step)
optimizer = torch.optim.AdamW(weights, lr=scheduler.learn_rate) optimizer = torch.optim.AdamW(weights, lr=scheduler.learn_rate)
steps_without_grad = 0
pbar = tqdm.tqdm(enumerate(ds), total=steps - ititial_step) pbar = tqdm.tqdm(enumerate(ds), total=steps - ititial_step)
for i, entries in pbar: for i, entries in pbar:
hypernetwork.step = i + ititial_step hypernetwork.step = i + ititial_step
@ -332,8 +334,17 @@ def train_hypernetwork(hypernetwork_name, learn_rate, batch_size, data_root, log
losses[hypernetwork.step % losses.shape[0]] = loss.item() losses[hypernetwork.step % losses.shape[0]] = loss.item()
optimizer.zero_grad() optimizer.zero_grad()
weights[0].grad = None
loss.backward() loss.backward()
if weights[0].grad is None:
steps_without_grad += 1
else:
steps_without_grad = 0
assert steps_without_grad < 10, 'no gradient found for the trained weight after backward() for 10 steps in a row; this is a bug; training cannot continue'
optimizer.step() optimizer.step()
mean_loss = losses.mean() mean_loss = losses.mean()
if torch.isnan(mean_loss): if torch.isnan(mean_loss):
raise RuntimeError("Loss diverged.") raise RuntimeError("Loss diverged.")