Update resampling

feature_extractor.py

@@ -6,13 +6,13 @@ from transformers import AutoTokenizer, Wav2Vec2ForCTC
 import torch
 import numpy as np
 import glob
+import librosa
 import numpy
 import os.path
 
 processor = AutoTokenizer.from_pretrained("facebook/wav2vec2-large-960h-lv60")
-
 model = Wav2Vec2ForCTC.from_pretrained("facebook/wav2vec2-large-960h-lv60")
-
+new_sample_rate = 16000
 # Dementia path
 # /home/bmoell/data/media.talkbank.org/dementia/English/Pitt
 # cookie dementia /home/bmoell/data/media.talkbank.org/dementia/English/Pitt/Dementia/cookie
@@ -35,28 +35,33 @@ def feature_extractor(path):
         if not os.path.isfile(wav_file + ".wav2vec2.pt"):      
             get_wav2vecembeddings_from_audiofile(wav_file)
 
+
+def change_sample_rate(y, sample_rate, new_sample_rate):
+    value = librosa.resample(y, sample_rate, new_sample_rate)
+    return value
+
 def get_wav2vecembeddings_from_audiofile(wav_file):
     print("the file is", wav_file)
     speech, sample_rate = sf.read(wav_file)
-    input_values = processor(wav_file, return_tensors="pt", padding=True) # there is no truncation param anymore
+     # change sample rate to 16 000 hertz
+    resampled = change_sample_rate(speech, sample_rate, new_sample_rate)
+    print("the speech is", speech)
+    input_values = processor(resampled, return_tensors="pt", padding=True, sampling_rate=new_sample_rate) # there is no truncation param anymore
     print("input values", input_values)
-
-    file_info = os.stat(wav_file)
-    file_size = file_info.st_size
-    print("the size is", file_size)
-
-    if file_size > 250:
-        with torch.no_grad():
-            encoded_states = model(
-                input_values=input_values["input_ids"], 
-                attention_mask=input_values["attention_mask"], 
-                output_hidden_states=True
-            )
-
-            last_hidden_state = encoded_states.hidden_states[-1] # The last hidden-state is the first element of the output tuple
-            print("getting wav2vec2 embeddings")
-            print(last_hidden_state)
-            torch.save(last_hidden_state, wav_file + '.wav2vec2.pt')
+    # import pdb
+    # pdb.set_trace()
+
+    with torch.no_grad():
+        encoded_states = model(
+            **input_values, 
+            # attention_mask=input_values["attention_mask"], 
+            output_hidden_states=True
+        )
+
+        last_hidden_state = encoded_states.hidden_states[-1] # The last hidden-state is the first element of the output tuple
+        print("getting wav2vec2 embeddings")
+        print(last_hidden_state)
+        torch.save(last_hidden_state, wav_file + '.wav2vec2.pt')
 
     
 

readme.MD

@@ -6,7 +6,7 @@ train
 # Important readmes
 https://github.com/huggingface/transformers/tree/f42a0abf4bd765ad08e14b347a3acbe9fade31b9/examples/research_projects/jax-projects/wav2vec2
 
-# path to files
+# path to file
 # cookie control 
 data/media.talkbank.org/dementia/English/Pitt/Control/cookie