[GH-ISSUE #2850] PART 2: Critical Argument Reversals, NumPy 1.24+ Security Breaks, and Dynamic Mel-Spectrogram Shape Mismatches in Python 3.12 Branch #2050

New issue

Open

opened 2026-10-20 10:14:10 +00:00 by bsakaguchi · 0 comments

bsakaguchi commented

2026-10-20 10:14:10 +00:00

Owner

Originally created by @LebedevIV on GitHub (Aug 24, 2026).
Original GitHub issue: https://github.com/RVC-Project/Retrieval-based-Voice-Conversion-WebUI/issues/2850

Bug 7: Complete Positional Argument Mismatch inside `train/dataset/extract_f0.py`

Problem: When invoking pitch extraction directly via the CLI module execution (python -m train.dataset.extract_f0), the updated internal argument parser (sys.argv) completely reverses the expected input layout. Instead of reading standard pipeline indices, line 20 checks mode = sys.argv[1].lower() against hardcoded environment strings ("cpu", "cuda", "dml"). If a user passes an extraction method string first (e.g., rmvpe), it falls back to a fatal crash: ValueError: Unsupported F0 extraction mode: rmvpe. Conversely, if paths are passed sequentially, it parses the folder directory string as the hardware backend, throwing: ValueError: Unsupported F0 extraction mode: d:/git/rvc/logs/k_dedu.
File affected: train/dataset/extract_f0.py
Solution required: Standardize the argument parser structure across both WebUI wrapper calls and direct standalone module initializations. Enforce explicit argparse flags instead of relying on fragile, shifting sys.argv absolute array positioning.

Bug 8: NumPy 1.24+ Validation Failure (`allow_pickle=False`) in `train/data_utils.py`

Problem: Recent security patches in NumPy 1.24+ deprecate and block the deserialization of object arrays via np.load() by default to mitigate remote code execution risks (ValueError: Cannot load file containing pickled data when allow_pickle=False). However, RVC’s feature extraction sub-scripts (extract_hubert_feature.py) continue to generate hidden context .npy voice arrays utilizing object serialization protocols. As a result, the DataLoader completely halts training at the very first step of Epoch 1.
File affected: train/data_utils.py
Solution provided: Modify the label compilation arrays inside get_labels() to explicitly override the strict global security defaults since these feature weights are computed locally and are entirely deterministic:

def get_labels(self, phone, pitch, pitchf):
    phone = np.load(phone, allow_pickle=True)
    phone = np.repeat(phone, 2, axis=0)
    pitch = np.load(pitch, allow_pickle=True)
    pitchf = np.load(pitchf, allow_pickle=True)

Bug 9: Fatal `RuntimeError` Due to HiFi-GAN Vocoder Mel-Spectrogram Frame Alignment Dropping

Problem: When a dataset consists of short sliced audio packets (e.g., 3-second segments produced by the automated audio-slicer), the random target cropping length can drop below or fall out of mathematical alignment with the vocoder's network layer chain. The generator upsamples temporal data features using a strict multiplier sequence (12 * 10 * 2 * 2 = 480). If the internal segment_size window doesn't map cleanly to the remaining audio chunk borders, the target size and input size mismatch, triggering a fatal tensor crash during L1 loss calculation:
RuntimeError: The size of tensor a (120) must match the size of tensor b (480) at non-singleton dimension 2
File affected: train/train.py
Solution provided: Implement a resilient frame-level matrix shape balancer immediately prior to executing the F.l1_loss loss evaluation module to clip padding edge discrepancies dynamically on the fly:

# --- ARCHITECTURAL TENSOR SIZE BALANCER FIX ---
if y_mel.size(-1) != y_hat_mel.size(-1):
    min_frames = min(y_mel.size(-1), y_hat_mel.size(-1))
    y_mel = y_mel[:, :, :min_frames]
    y_hat_mel = y_hat_mel[:, :, :min_frames]
loss_mel = F.l1_loss(y_mel, y_hat_mel) * hps.train.c_mel
# -----------------------------------------------

Bug 10: Broken Output Directory Fallback in Dataset Preprocessing

Problem: The custom destination folder argument passed to train/preprocess.py from the WebUI is systematically ignored during dataset slicing routines. The script hard-routes all sliced audio chunks directly into the project's base fallback directory (./logs/k_dedu/0_gt_wavs), completely disregarding user-specified experiment directory tags (e.g., k_dedu_v3). This overwrites existing index footprints and creates background file conflicts.
File affected: train/preprocess.py

Bug 11: Missing Native UTF-8 Byte Order Mark (BOM) Sanitization

Problem: When files like filelist.txt or config.json are parsed on Windows environments, common local shell text outputs or editors prepend a non-visual 3-byte BOM marker (п»ї) to the initialization stream. The framework's input parsing methods read this header as text data, prepending it to file path lookups and throwing: OSError: [WinError 123] Invalid filename syntax: 'п»їD:\\git\\...'.
File affected: tools/file_io.py / train/utils.py
Solution suggested: Enforce utf-8-sig encoding globally across all text-reading functions to automatically strip neoteric Windows structural byte artifacts.

FEATURE REQUEST / SUGGESTION FOR EXTRACT SCRIPTS

Add a Force Regenerate Cache flag to extract_f0.py and extract_hubert_feature.py. Currently, the tools check for file completion globally based on simple directory string presence. If the naming convention shifts (e.g., creating 0_1.wav.npy instead of 0_1.npy), the scripts write a success state log and skip extraction completely, rendering empty data packets and forcing users to manually flush cache folders via system shells.

Originally created by @LebedevIV on GitHub (Aug 24, 2026). Original GitHub issue: https://github.com/RVC-Project/Retrieval-based-Voice-Conversion-WebUI/issues/2850 ## Bug 7: Complete Positional Argument Mismatch inside `train/dataset/extract_f0.py` * **Problem:** When invoking pitch extraction directly via the CLI module execution (`python -m train.dataset.extract_f0`), the updated internal argument parser (`sys.argv`) completely reverses the expected input layout. Instead of reading standard pipeline indices, line 20 checks `mode = sys.argv[1].lower()` against hardcoded environment strings (`"cpu"`, `"cuda"`, `"dml"`). If a user passes an extraction method string first (e.g., `rmvpe`), it falls back to a fatal crash: `ValueError: Unsupported F0 extraction mode: rmvpe`. Conversely, if paths are passed sequentially, it parses the folder directory string as the hardware backend, throwing: `ValueError: Unsupported F0 extraction mode: d:/git/rvc/logs/k_dedu`. * **File affected:** `train/dataset/extract_f0.py` * **Solution required:** Standardize the argument parser structure across both WebUI wrapper calls and direct standalone module initializations. Enforce explicit `argparse` flags instead of relying on fragile, shifting `sys.argv` absolute array positioning. --- ## Bug 8: NumPy 1.24+ Validation Failure (`allow_pickle=False`) in `train/data_utils.py` * **Problem:** Recent security patches in NumPy 1.24+ deprecate and block the deserialization of object arrays via `np.load()` by default to mitigate remote code execution risks (`ValueError: Cannot load file containing pickled data when allow_pickle=False`). However, RVC’s feature extraction sub-scripts (`extract_hubert_feature.py`) continue to generate hidden context `.npy` voice arrays utilizing object serialization protocols. As a result, the `DataLoader` completely halts training at the very first step of Epoch 1. * **File affected:** `train/data_utils.py` * **Solution provided:** Modify the label compilation arrays inside `get_labels()` to explicitly override the strict global security defaults since these feature weights are computed locally and are entirely deterministic: ```python def get_labels(self, phone, pitch, pitchf): phone = np.load(phone, allow_pickle=True) phone = np.repeat(phone, 2, axis=0) pitch = np.load(pitch, allow_pickle=True) pitchf = np.load(pitchf, allow_pickle=True) ``` --- ## Bug 9: Fatal `RuntimeError` Due to HiFi-GAN Vocoder Mel-Spectrogram Frame Alignment Dropping * **Problem:** When a dataset consists of short sliced audio packets (e.g., 3-second segments produced by the automated `audio-slicer`), the random target cropping length can drop below or fall out of mathematical alignment with the vocoder's network layer chain. The generator upsamples temporal data features using a strict multiplier sequence (`12 * 10 * 2 * 2 = 480`). If the internal `segment_size` window doesn't map cleanly to the remaining audio chunk borders, the target size and input size mismatch, triggering a fatal tensor crash during L1 loss calculation: `RuntimeError: The size of tensor a (120) must match the size of tensor b (480) at non-singleton dimension 2` * **File affected:** `train/train.py` * **Solution provided:** Implement a resilient frame-level matrix shape balancer immediately prior to executing the `F.l1_loss` loss evaluation module to clip padding edge discrepancies dynamically on the fly: ```python # --- ARCHITECTURAL TENSOR SIZE BALANCER FIX --- if y_mel.size(-1) != y_hat_mel.size(-1): min_frames = min(y_mel.size(-1), y_hat_mel.size(-1)) y_mel = y_mel[:, :, :min_frames] y_hat_mel = y_hat_mel[:, :, :min_frames] loss_mel = F.l1_loss(y_mel, y_hat_mel) * hps.train.c_mel # ----------------------------------------------- ``` --- ## Bug 10: Broken Output Directory Fallback in Dataset Preprocessing * **Problem:** The custom destination folder argument passed to `train/preprocess.py` from the WebUI is systematically ignored during dataset slicing routines. The script hard-routes all sliced audio chunks directly into the project's base fallback directory (`./logs/k_dedu/0_gt_wavs`), completely disregarding user-specified experiment directory tags (e.g., `k_dedu_v3`). This overwrites existing index footprints and creates background file conflicts. * **File affected:** `train/preprocess.py` --- ## Bug 11: Missing Native UTF-8 Byte Order Mark (BOM) Sanitization * **Problem:** When files like `filelist.txt` or `config.json` are parsed on Windows environments, common local shell text outputs or editors prepend a non-visual 3-byte BOM marker (`п»ї`) to the initialization stream. The framework's input parsing methods read this header as text data, prepending it to file path lookups and throwing: `OSError: [WinError 123] Invalid filename syntax: 'п»їD:\\git\\...'`. * **File affected:** `tools/file_io.py` / `train/utils.py` * **Solution suggested:** Enforce `utf-8-sig` encoding globally across all text-reading functions to automatically strip neoteric Windows structural byte artifacts. --- ### FEATURE REQUEST / SUGGESTION FOR EXTRACT SCRIPTS Add a **Force Regenerate Cache** flag to `extract_f0.py` and `extract_hubert_feature.py`. Currently, the tools check for file completion globally based on simple directory string presence. If the naming convention shifts (e.g., creating `0_1.wav.npy` instead of `0_1.npy`), the scripts write a success state log and skip extraction completely, rendering empty data packets and forcing users to manually flush cache folders via system shells.

bsakaguchi added the

following up

label

2026-10-20 10:14:10 +00:00