update test with slice

tests/models/dac/test_modeling_dac.py

@@ -457,12 +457,11 @@ def test_integration_16khz(self):
 
     def test_integration_24khz(self):
         expected_rmse = 0.0039
-
-        expected_encoder_sums_dict = {
-            "loss": 30.0128,
-            "quantized_representation": 0.0104,
-            "audio_codes": 518.2788,
-            "projected_latents": -0.0051,
+
+        expected_encoder_output_dict = {
+            "quantized_representation": torch.tensor([0.9807, 2.8212, 5.2514, 2.7241, 1.0426]),
+            "audio_codes": torch.tensor([919, 919, 234, 777, 234]),
+            "projected_latents": torch.tensor([-4.7822, -5.0046, -4.5574, -5.0363, -5.4271]),
         }
         librispeech_dummy = load_dataset("hf-internal-testing/librispeech_asr_dummy", "clean", split="validation")
 
@@ -483,12 +482,15 @@ def test_integration_24khz(self):
 
         with torch.no_grad():
             encoder_outputs = model.encode(inputs["input_values"])
-
-            expected_encoder_sums = torch.tensor(list(expected_encoder_sums_dict.values()), dtype=torch.float32)
-            encoder_outputs_mean = torch.tensor([v.float().mean().cpu().item() for v in encoder_outputs.to_tuple()])
-
-            # make sure audio encoded codes are correct
-            self.assertTrue(torch.allclose(encoder_outputs_mean, expected_encoder_sums, atol=1e-3))
+
+            expected_quantized_representation = encoder_outputs['quantized_representation'][0,0,:5].cpu()
+            expected_audio_codes = encoder_outputs['audio_codes'][0,0,:5].cpu()
+            expected_projected_latents = encoder_outputs['projected_latents'][0,0,:5].cpu()
+
+            # make sure values are correct for audios slices
+            self.assertTrue(torch.allclose(expected_quantized_representation, expected_encoder_output_dict['quantized_representation'], atol=1e-3))
+            self.assertTrue(torch.allclose(expected_audio_codes, expected_encoder_output_dict['audio_codes'], atol=1e-3))
+            self.assertTrue(torch.allclose(expected_projected_latents, expected_encoder_output_dict['projected_latents'], atol=1e-3))
 
             _, quantized_representation, _, _ = encoder_outputs.to_tuple()
             input_values_dec = model.decode(quantized_representation)[0]