Fix mutable proj_out weight in the Attention layer (huggingface#73)

anton-l · web-flow · commit d9316bf8bc74 · 2022-07-04T12:36:37.000+02:00
* Catch unused params in DDP

* Fix proj_out, add test
diff --git a/examples/train_unconditional.py b/examples/train_unconditional.py
@@ -4,7 +4,7 @@
 import torch
 import torch.nn.functional as F
 
-from accelerate import Accelerator
+from accelerate import Accelerator, DistributedDataParallelKwargs
 from accelerate.logging import get_logger
 from datasets import load_dataset
 from diffusers import DDIMPipeline, DDIMScheduler, UNetModel
@@ -27,8 +27,14 @@
 
 
 def main(args):
+    ddp_unused_params = DistributedDataParallelKwargs(find_unused_parameters=True)
     logging_dir = os.path.join(args.output_dir, args.logging_dir)
-    accelerator = Accelerator(mixed_precision=args.mixed_precision, log_with="tensorboard", logging_dir=logging_dir)
+    accelerator = Accelerator(
+        mixed_precision=args.mixed_precision,
+        log_with="tensorboard",
+        logging_dir=logging_dir,
+        kwargs_handlers=[ddp_unused_params],
+    )
 
     model = UNetModel(
         attn_resolutions=(16,),
diff --git a/src/diffusers/models/attention.py b/src/diffusers/models/attention.py
@@ -70,7 +70,7 @@ def __init__(
         if encoder_channels is not None:
             self.encoder_kv = nn.Conv1d(encoder_channels, channels * 2, 1)
 
-        self.proj_out = zero_module(nn.Conv1d(channels, channels, 1))
+        self.proj = zero_module(nn.Conv1d(channels, channels, 1))
 
         self.overwrite_qkv = overwrite_qkv
         if overwrite_qkv:
@@ -108,15 +108,15 @@ def set_weights(self, module):
             proj_out.weight.data = module.proj_out.weight.data[:, :, :, 0]
             proj_out.bias.data = module.proj_out.bias.data
 
-            self.proj_out = proj_out
+            self.proj = proj_out
         elif self.overwrite_linear:
             self.qkv.weight.data = torch.concat(
                 [self.NIN_0.W.data.T, self.NIN_1.W.data.T, self.NIN_2.W.data.T], dim=0
             )[:, :, None]
             self.qkv.bias.data = torch.concat([self.NIN_0.b.data, self.NIN_1.b.data, self.NIN_2.b.data], dim=0)
 
-            self.proj_out.weight.data = self.NIN_3.W.data.T[:, :, None]
-            self.proj_out.bias.data = self.NIN_3.b.data
+            self.proj.weight.data = self.NIN_3.W.data.T[:, :, None]
+            self.proj.bias.data = self.NIN_3.b.data
 
             self.norm.weight.data = self.GroupNorm_0.weight.data
             self.norm.bias.data = self.GroupNorm_0.bias.data
@@ -149,7 +149,7 @@ def forward(self, x, encoder_out=None):
         a = torch.einsum("bts,bcs->bct", weight, v)
         h = a.reshape(bs, -1, length)
 
-        h = self.proj_out(h)
+        h = self.proj(h)
         h = h.reshape(b, c, *spatial)
 
         result = x + h
diff --git a/src/diffusers/training_utils.py b/src/diffusers/training_utils.py
@@ -30,7 +30,7 @@ def __init__(
             min_value (float): The minimum EMA decay rate. Default: 0.
         """
 
-        self.averaged_model = copy.deepcopy(model)
+        self.averaged_model = copy.deepcopy(model).eval()
         self.averaged_model.requires_grad_(False)
 
         self.update_after_step = update_after_step
diff --git a/tests/test_modeling_utils.py b/tests/test_modeling_utils.py
@@ -52,6 +52,7 @@
 from diffusers.pipeline_utils import DiffusionPipeline
 from diffusers.pipelines.bddm.pipeline_bddm import DiffWave
 from diffusers.testing_utils import floats_tensor, slow, torch_device
+from diffusers.training_utils import EMAModel
 
 
 torch.backends.cuda.matmul.allow_tf32 = False
@@ -197,6 +198,20 @@ def test_training(self):
         loss = torch.nn.functional.mse_loss(output, noise)
         loss.backward()
 
+    def test_ema_training(self):
+        init_dict, inputs_dict = self.prepare_init_args_and_inputs_for_common()
+
+        model = self.model_class(**init_dict)
+        model.to(torch_device)
+        model.train()
+        ema_model = EMAModel(model, device=torch_device)
+
+        output = model(**inputs_dict)
+        noise = torch.randn((inputs_dict["x"].shape[0],) + self.output_shape).to(torch_device)
+        loss = torch.nn.functional.mse_loss(output, noise)
+        loss.backward()
+        ema_model.step(model)
+
 
 class UnetModelTests(ModelTesterMixin, unittest.TestCase):
     model_class = UNetModel
