SUMIN/HuggingFace_transformer
1
0
Fork 0
Code Issues Pull Requests Actions 7 Packages Projects Releases Wiki Activity

Fix bug when requesting input normalization with EnCodec (#34756)

Browse Source 
* EnCodec: unsqueeze padding mask

* add test for normalization
This commit is contained in:
Francesco Cariaggi
2025-01-07 11:50:02 +01:00
committed by GitHub
parent 96bf3d6cc5
commit f408d55448
2 changed files with 20 additions and 5 deletions
Download Patch File Download Diff File Expand all files Collapse all files

2
src/transformers/models/encodec/modeling_encodec.py
View File

@@ -576,7 +576,7 @@ class EncodecModel(EncodecPreTrainedModel):
scale = None scale = None
if self.config.normalize: if self.config.normalize:
# if the padding is non zero # if the padding is non zero
input_values = input_values * padding_mask input_values = input_values * padding_mask.unsqueeze(1)
mono = torch.sum(input_values, 1, keepdim=True) / input_values.shape[1] mono = torch.sum(input_values, 1, keepdim=True) / input_values.shape[1]
scale = mono.pow(2).mean(dim=-1, keepdim=True).sqrt() + 1e-8 scale = mono.pow(2).mean(dim=-1, keepdim=True).sqrt() + 1e-8
input_values = input_values / scale input_values = input_values / scale

23
tests/models/encodec/test_modeling_encodec.py
View File

@@ -39,7 +39,7 @@ from ...test_pipeline_mixin import PipelineTesterMixin
if is_torch_available(): if is_torch_available():
import torch import torch
from transformers import EncodecModel from transformers import EncodecFeatureExtractor, EncodecModel
def prepare_inputs_dict( def prepare_inputs_dict(
@@ -111,6 +111,19 @@ class EncodecModelTester:
return config, inputs_dict return config, inputs_dict
def prepare_config_and_inputs_for_normalization(self):
input_values = floats_tensor([self.batch_size, self.num_channels, self.intermediate_size], scale=1.0)
config = self.get_config()
config.normalize = True
processor = EncodecFeatureExtractor(feature_size=config.audio_channels, sampling_rate=config.sampling_rate)
input_values = list(input_values.cpu().numpy())
inputs_dict = processor(
input_values, sampling_rate=config.sampling_rate, padding=True, return_tensors="pt"
).to(torch_device)
return config, inputs_dict
def get_config(self): def get_config(self):
return EncodecConfig( return EncodecConfig(
audio_channels=self.num_channels, audio_channels=self.num_channels,
@@ -125,9 +138,7 @@ class EncodecModelTester:
def create_and_check_model_forward(self, config, inputs_dict): def create_and_check_model_forward(self, config, inputs_dict):
model = EncodecModel(config=config).to(torch_device).eval() model = EncodecModel(config=config).to(torch_device).eval()
result = model(**inputs_dict)
input_values = inputs_dict["input_values"]
result = model(input_values)
self.parent.assertEqual( self.parent.assertEqual(
result.audio_values.shape, (self.batch_size, self.num_channels, self.intermediate_size) result.audio_values.shape, (self.batch_size, self.num_channels, self.intermediate_size)
) )
@@ -435,6 +446,10 @@ class EncodecModelTest(ModelTesterMixin, PipelineTesterMixin, unittest.TestCase)
config.use_conv_shortcut = False config.use_conv_shortcut = False
self.model_tester.create_and_check_model_forward(config, inputs_dict) self.model_tester.create_and_check_model_forward(config, inputs_dict)
def test_model_forward_with_normalization(self):
config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_normalization()
self.model_tester.create_and_check_model_forward(config, inputs_dict)
def normalize(arr): def normalize(arr):
norm = np.linalg.norm(arr) norm = np.linalg.norm(arr)