trigger actions

setup.py

@@ -107,6 +107,7 @@
     "deepspeed>=0.9.3",
     "diffusers",
     "dill<0.3.5",
+    "einops",
     "evaluate>=0.2.0",
     "faiss-cpu",
     "fastapi",

src/transformers/__init__.py

@@ -2153,10 +2153,10 @@
     )
     _import_structure["models.encodec"].extend(
         [
+            "EncodecDiscriminator",
+            "EncodecDiscriminatorConfig",
             "EncodecModel",
             "EncodecPreTrainedModel",
-            "EncodecDiscriminatorConfig",
-            "EncodecDiscriminator",
         ]
     )
     _import_structure["models.encoder_decoder"].append("EncoderDecoderModel")

src/transformers/dependency_versions_table.py

@@ -14,6 +14,7 @@
     "deepspeed": "deepspeed>=0.9.3",
     "diffusers": "diffusers",
     "dill": "dill<0.3.5",
+    "einops": "einops",
     "evaluate": "evaluate>=0.2.0",
     "faiss-cpu": "faiss-cpu",
     "fastapi": "fastapi",

src/transformers/models/encodec/__init__.py

@@ -24,10 +24,10 @@
     "configuration_encodec": ["EncodecConfig"],
     "feature_extraction_encodec": ["EncodecFeatureExtractor"],
     "loss_encodec": [
+        "Balancer",
         "compute_discriminator_loss",
-        "compute_generator_adv_loss",
         "compute_feature_matching_loss",
-        "Balancer",
+        "compute_generator_adv_loss",
     ],
 }
 

src/transformers/models/encodec/modeling_encodec.py

@@ -1141,7 +1141,7 @@ def compute_loss(self, real_audio, fake_audio):
             g_adv_loss += -fake_logit.mean()
         g_adv_loss /= self.num_discriminators
 
-        # Feature matching loss
+        # feature matching loss
         fm_loss = 0
         for real_feat, fake_feat in zip(real_features, fake_features):
             for real_f, fake_f in zip(real_feat, fake_feat):

tests/models/encodec/test_modeling_encodec.py

@@ -33,7 +33,7 @@
     compute_feature_matching_loss,
     compute_generator_adv_loss,
 )
-from transformers.testing_utils import is_torch_available, require_torch, slow, torch_device
+from transformers.testing_utils import is_torch_available, require_torch, require_torchaudio, slow, torch_device
 
 from ...test_configuration_common import ConfigTester
 from ...test_modeling_common import ModelTesterMixin, _config_zero_init, floats_tensor, ids_tensor
@@ -183,6 +183,7 @@ def test_balancer_basic(self):
         assert torch.allclose(x.grad, torch.tensor(0.0)), x.grad
 
     @slow
+    @require_torchaudio
     def test_training_with_discriminator(self):
         model_id = "facebook/encodec_24khz"
         model = EncodecModel.from_pretrained(model_id).to(torch_device)
@@ -311,6 +312,7 @@ def test_training_with_discriminator(self):
             print(f"Total generator loss (before balancing): {total_gen_loss:.4f}\n")
 
     @slow
+    @require_torchaudio
     def test_reconstruction_loss(self):
         model_id = "facebook/encodec_24khz"
         model = EncodecModel.from_pretrained(model_id).to(torch_device)
@@ -354,6 +356,7 @@ def test_reconstruction_loss(self):
             print(f"Spectrogram MAE: {spec_mae.item()}")
 
     @slow
+    @require_torchaudio
     def test_gradients_exist(self):
         model_id = "facebook/encodec_24khz"
         model = EncodecModel.from_pretrained(model_id).to(torch_device)