From 7fd90128eb6d1820045bfe2c2c1269661023a712 Mon Sep 17 00:00:00 2001
From: AUTOMATIC <16777216c@gmail.com>
Date: Sat, 22 Oct 2022 14:48:43 +0300
Subject: [PATCH] added a guard for hypernet training that will stop early if
 weights are getting no gradients

---
 modules/hypernetworks/hypernetwork.py | 11 +++++++++++
 1 file changed, 11 insertions(+)

diff --git a/modules/hypernetworks/hypernetwork.py b/modules/hypernetworks/hypernetwork.py
index 47d91ea5..46039a49 100644
--- a/modules/hypernetworks/hypernetwork.py
+++ b/modules/hypernetworks/hypernetwork.py
@@ -310,6 +310,8 @@ def train_hypernetwork(hypernetwork_name, learn_rate, batch_size, data_root, log
     scheduler = LearnRateScheduler(learn_rate, steps, ititial_step)
     optimizer = torch.optim.AdamW(weights, lr=scheduler.learn_rate)
 
+    steps_without_grad = 0
+
     pbar = tqdm.tqdm(enumerate(ds), total=steps - ititial_step)
     for i, entries in pbar:
         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()
 
             optimizer.zero_grad()
+            weights[0].grad = None
             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()
+
         mean_loss = losses.mean()
         if torch.isnan(mean_loss):
             raise RuntimeError("Loss diverged.")