Update README.md
Browse files
README.md
CHANGED
|
@@ -1,4 +1,84 @@
|
|
| 1 |
---
|
| 2 |
license: cc-by-4.0
|
| 3 |
---
|
| 4 |
-
Coming Soon
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
---
|
| 2 |
license: cc-by-4.0
|
| 3 |
---
|
| 4 |
+
Coming Soon
|
| 5 |
+
|
| 6 |
+
|
| 7 |
+
|
| 8 |
+
This script is for merging tokenized speech datasets stored in memmap format. The input datasets can be combined to form larger training datasets.
|
| 9 |
+
```python
|
| 10 |
+
import numpy as np
|
| 11 |
+
import os
|
| 12 |
+
|
| 13 |
+
def merge_memmap_datasets(dataset_dirs, output_dir):
|
| 14 |
+
# Ensure the output directory exists
|
| 15 |
+
os.makedirs(output_dir, exist_ok=True)
|
| 16 |
+
|
| 17 |
+
# Dataset splits to be merged
|
| 18 |
+
splits = ['train', 'val']
|
| 19 |
+
|
| 20 |
+
for split in splits:
|
| 21 |
+
shapes = []
|
| 22 |
+
seq_len = None
|
| 23 |
+
total_samples = 0
|
| 24 |
+
|
| 25 |
+
# Collect shapes of all datasets and check sequence length consistency
|
| 26 |
+
for dataset_dir in dataset_dirs:
|
| 27 |
+
shape_path = os.path.join(dataset_dir, f'{split}_input_ids_shape.npy')
|
| 28 |
+
if not os.path.exists(shape_path):
|
| 29 |
+
print(f"Warning: {split}_input_ids_shape.npy not found in {dataset_dir}, skipping this dataset.")
|
| 30 |
+
continue
|
| 31 |
+
shape = np.load(shape_path)
|
| 32 |
+
print(f"Loaded shape of {split} data from {dataset_dir}: {shape}")
|
| 33 |
+
shape = tuple(shape)
|
| 34 |
+
shapes.append((dataset_dir, shape))
|
| 35 |
+
total_samples += shape[0]
|
| 36 |
+
if seq_len is None:
|
| 37 |
+
seq_len = shape[1]
|
| 38 |
+
elif seq_len != shape[1]:
|
| 39 |
+
print(f"Error: Sequence length mismatch in {split} data from {dataset_dir}.")
|
| 40 |
+
return
|
| 41 |
+
|
| 42 |
+
if total_samples == 0:
|
| 43 |
+
print(f"Error: No valid {split} data found for merging.")
|
| 44 |
+
continue
|
| 45 |
+
|
| 46 |
+
new_shape = (total_samples, seq_len)
|
| 47 |
+
|
| 48 |
+
# Create new memmap file
|
| 49 |
+
output_memmap_path = os.path.join(output_dir, f'{split}_input_ids.memmap')
|
| 50 |
+
output_memmap = np.memmap(
|
| 51 |
+
output_memmap_path, dtype='int32', mode='w+', shape=new_shape
|
| 52 |
+
)
|
| 53 |
+
|
| 54 |
+
# Copy data from each dataset to the new memmap file
|
| 55 |
+
start_idx = 0
|
| 56 |
+
for dataset_dir, shape in shapes:
|
| 57 |
+
memmap_path = os.path.join(dataset_dir, f'{split}_input_ids.memmap')
|
| 58 |
+
data = np.memmap(
|
| 59 |
+
memmap_path, dtype='int32', mode='r', shape=shape
|
| 60 |
+
)
|
| 61 |
+
end_idx = start_idx + shape[0]
|
| 62 |
+
output_memmap[start_idx:end_idx, :] = data[:]
|
| 63 |
+
print(f"Merged {split} data from {dataset_dir} into positions {start_idx}:{end_idx}")
|
| 64 |
+
start_idx = end_idx
|
| 65 |
+
del data # Free memory
|
| 66 |
+
|
| 67 |
+
# Delete temporary variable and flush data to disk
|
| 68 |
+
del output_memmap
|
| 69 |
+
|
| 70 |
+
# Save the new shape file
|
| 71 |
+
np.save(os.path.join(output_dir, f'{split}_input_ids_shape.npy'), new_shape)
|
| 72 |
+
|
| 73 |
+
print(f"Completed merging {split} data. New shape: {new_shape}")
|
| 74 |
+
|
| 75 |
+
if __name__ == "__main__":
|
| 76 |
+
dataset_dirs = [
|
| 77 |
+
'libriheavy_tts_1',
|
| 78 |
+
'libriheavy_tts_2',
|
| 79 |
+
'libriheavy_tts_3',
|
| 80 |
+
'libriheavy_tts_4'
|
| 81 |
+
]
|
| 82 |
+
output_dir = 'libriheavy_tts_all'
|
| 83 |
+
merge_memmap_datasets(dataset_dirs, output_dir)
|
| 84 |
+
```
|