feature: implement training

.gitignore

@@ -167,3 +167,6 @@ cython_debug/
 # Project Specific
 *.mp3
 results/
+train/
+data/
+tmp/

commonvoice.py

@@ -0,0 +1,21 @@
+import pandas as pd
+from umsc import UgMultiScriptConverter
+from tqdm import tqdm
+
+source_script = 'UAS'
+target_script = 'ULS'
+converter = UgMultiScriptConverter(source_script, target_script)
+
+df = pd.read_csv('./data/commonvoice/validated.tsv', sep='\t')
+
+new_df = pd.DataFrame(columns=['path', 'script'])
+
+qbar = tqdm(total=len(df))
+for index, row in df.iterrows():
+    qbar.update(1)
+    # Skip too bad samples
+    if row['up_votes'] < row['down_votes']:
+        continue
+    new_df.loc[index] = [row['path'], converter(row['sentence'])]
+
+new_df.to_csv('./data/training/commonvoice_train.csv', index=False)

data.py

@@ -1,3 +1,4 @@
+import pandas as pd
 import torch
 from torch.utils.data import Dataset
 from torch.utils.data import DataLoader
@@ -38,7 +39,7 @@ def addnoise(audio):
 
 def randomstretch(audio):
     factor = random.uniform(0.8, 1.2)
-    audio = librosa.core.resample(audio,sample_rate,sample_rate*factor)
+    audio = librosa.core.resample(audio, orig_sr=sample_rate, target_sr=sample_rate*factor)
     return audio
 
 #def spec_augment(feat, T=70, F=15, time_mask_num=1, freq_mask_num=1):
@@ -104,18 +105,19 @@ def melfuture(wav_path, augument = False):
 class SpeechDataset(Dataset):
     def __init__(self, index_path, augumentation = False):
         self.Raw = False
-        with open(index_path,encoding='utf_8_sig') as f:
+        self.idx = []
-            lines = f.readlines()
         
-        self.idx  = []
+        df = pd.read_csv(index_path)
-        for x in lines:
+        for _, row in df.iterrows():
-            item = x.strip().split("\t")
+            path = row['path']
-            if os.path.exists(item[0]):
+            sentence = row['script']
-                line = []
+            if not os.path.exists(path):
-                line.append(item[0])
+                continue
-                char_indx = uyghur_latin.encode(item[1])
+            line = []
-                line.append(char_indx)
+            line.append(path)
-                self.idx.append(line)
+            char_indx = uyghur_latin.encode(sentence)
+            line.append(char_indx)
+            self.idx.append(line)
 
         self.augument = augumentation
     

train.py

@@ -1,5 +1,6 @@
 import numpy as np
 import os
+import pandas as pd
 import torch
 import torch.nn.functional as F
 
@@ -128,25 +129,59 @@ def train(model, train_loader):
 
 
 if __name__ == "__main__":
-    device = "mps"
+    device = "cuda"
+    
+    training_data_path = './data/training/'
     
     os.makedirs('./results',exist_ok=True)
+    os.makedirs('./tmp',exist_ok=True)
     
-    train_file = 'thuyg20_train.csv'
+    # Read the folder train, list all the files in it
-    test_file  = 'thuyg20_test.csv'
+    files = os.listdir(training_data_path)
+    # Filter out the files that are not csv/tsv files
+    files = [f for f in files if f.endswith('.csv') or f.endswith('.tsv')]
+    # Find out the file with `_train` and `_test`
+    train_files = [f for f in files if '_train' in f]
+    test_files  = [f for f in files if '_test' in f]
+
+    # Read all train files, combine their `path` and `script` columns
+    train_df = pd.DataFrame()
+    for f in train_files:
+        if f.endswith('.tsv'):
+            df = pd.read_csv(training_data_path + f, sep='\t')
+        else:
+            df = pd.read_csv(training_data_path + f)
+        
+        train_df = pd.concat([train_df, df])
+    # Read all test files, combine their `path` and `script` columns
+    test_df = pd.DataFrame()
+    for f in test_files:
+        if f.endswith('.tsv'):
+            df = pd.read_csv(training_data_path + f, sep='\t')
+        else:
+            df = pd.read_csv(training_data_path + f)
+        test_df = pd.concat([test_df, df])
+    
+    # Save the combined dataframe to csv
+    train_df.to_csv('./tmp/train.csv', index=False)
+    test_df.to_csv('./tmp/test.csv', index=False)
+
+    train_file = './tmp/train.csv'
+    test_file  = './tmp/test.csv'
     
     train_set    = SpeechDataset(train_file, augumentation=True)
-    train_loader = SpeechDataLoader(train_set,num_workers=4, pin_memory = True, shuffle=True, batch_size=20)
+    train_loader = SpeechDataLoader(train_set, num_workers=8, pin_memory = True, shuffle=True, batch_size=64)
 
     validation_set = SpeechDataset(test_file, augumentation=False)
-    validation_loader = SpeechDataLoader(validation_set,num_workers=4, pin_memory = True, shuffle=True, batch_size=20)
+    validation_loader = SpeechDataLoader(validation_set, num_workers=8, pin_memory = True, shuffle=True, batch_size=64)
 
     print("="*50)
+    
     msg =  f"        Training Set: {train_file}, {len(train_set)} samples" + "\n" 
     msg += f"      Validation Set: {test_file}, {len(validation_set)} samples" + "\n"
     msg += f"         Vocab Size : {uyghur_latin.vocab_size}" 
-
     print(msg)
+    
     model = UModel(num_features_input = featurelen) 
 
     print("="*50)