Spaces:
Sleeping
Sleeping
Download src/utils.py from pacomesimon/vibeLabel: direct link, hf CLI and curl.
- Browser
- Download file 973 Bytes
-
https://huggingface.co/spaces/pacomesimon/vibeLabel/resolve/main/src/utils.py
- Command line
-
hf download hf://spaces/pacomesimon/vibeLabel/src/utils.py
-
curl -L -o utils.py https://huggingface.co/spaces/pacomesimon/vibeLabel/resolve/main/src/utils.py
973 Bytes
| import os | |
| def batch_iterable(iterable, batch_size): | |
| """Yield successive batches from iterable""" | |
| it = list(iterable).copy() | |
| for i in range(0, len(it), batch_size): | |
| yield it[i:i + batch_size] | |
| def zip_folder(folder_path): | |
| """ | |
| Reads all .txt files in the folder and returns them as a list of dictionaries. | |
| Wrapper name preserved from original code, though it returns content list. | |
| """ | |
| if (len(folder_path) == 0) or (not os.path.exists(folder_path)): | |
| return None | |
| # Original code had references to creating a csv/zip but commented out. | |
| # We only reproduce the active logic. | |
| data = [] | |
| for fname in os.listdir(folder_path): | |
| if fname.lower().endswith(".txt"): | |
| fpath = os.path.join(folder_path, fname) | |
| with open(fpath, "r", encoding="utf-8", errors="ignore") as f: | |
| content = f.read() | |
| data.append({"filename": fname, "content": content}) | |
| return data | |