Commit History
-
Re-export ONNX with the fixed graph, plus an EP-agnostic optimization pass
Verified signatureReplaces graphs exported from the previous modelling code. Same voice, same weights. Fixes carried over from the Python side: - No Expand broadcasting word_to_phoneme against the BERT feature rows. That node enforced the phoneme/BERT-token contract at runtime and surfaced any mismatch as `invalid expand shape`; the contract is now checked in the processor with the offending sentence in the message. - Relative-position padding is no longer frozen at the traced sequence length. The previous export was only numerically correct at the length it was calibrated with, which shifted attention and therefore predicted durations at every other length. - The stochastic duration predictor is gone from the graph (this checkpoint runs it at ratio 0, so it was computed and multiplied by zero), taking its splines and NonZero indexing with it. - The flow's attention is masked, so a batched call no longer leaks across utterances. Exported from a checkpoint trimmed by optimize_for_inference.py (BERT lay…3efd258 -
Ship the pinyin-to-phoneme table as a tokenizer asset
Verified signature`from_pretrained` only downloads files the tokenizer declares, so `data/zh_g2p.json` was never fetched and Mandarin g2p failed for anyone who had not cloned the whole repo. This is the same file at the path `vocab_files_names` resolves; `data/zh_g2p.json` stays for older revisions of the code.a0b6b53 -
Fix transformers v5 compatibility and rework BERT feature expansion
Verified signatureModelling code only; weights are unchanged. - bert_features: replace repeat_interleave(tensor repeats) with a cumsum/gather expansion. Bit-identical output, but no ONNX Loop/Expand over word_to_phoneme, and the phoneme/BERT-token contract is now checked in the processor with a readable error. - Fix attention relative-position embeddings freezing their padding at the traced sequence length, which made ONNX exports wrong at every other length. - Fix a crash on any text containing a space (the SP phoneme collided with a special token and the phoneme string was re-split). - Rewrite the Chinese number normalizer: numbers containing a zero were misread (100 -> 一百零十) and '0' raised IndexError. Adds a digit-by-digit style. - Restore transformers v5 compatibility (removed APIs, processor construction, pad_token_id) and fix trust_remote_code loading for AutoModel/AutoProcessor. - Skip the stochastic duration predictor when its ratio is 0, and the prior draw when noise_scale is 0. - Mask the flo…a85a126
-
Create processor_config.json
Verified signaturecdcafe3
-
Update README.md
Verified signature2f1301a
-
Fix error when load from latest transformers
Verified signatured965a2f
-
Update processor_config.json
Verified signature8731e9c
-
Update processor_config.json
Verified signature16e04f2
-
Fix add blank
Verified signature9ba62f0 -
Upload model_quantized.onnx
Verified signature1900559
-
Correct example
Verified signaturec4a5682 -
Update modeling_bert_vits2.py
Verified signature510ec41 -
Update modeling_bert_vits2.py
Verified signaturec206106 -
Update modeling_bert_vits2.py
Verified signature625ffa6 -
Update modeling_bert_vits2.py
Verified signature0612c39 -
Create data/zh_g2p.json
Verified signature052cb38 -
load g2p map from repo
Verified signaturea45adc6 -
Update tokenizer_config.json
Verified signature4f27953 -
Update processing_bert_vits2.py
Verified signature23bb9d9 -
Update modeling_bert_vits2.py
Verified signature3c62bd5 -
Update modeling_bert_vits2.py
Verified signature913c804 -
Update tokenizer_config.json
Verified signature621c9ce -
Update processor_config.json
Verified signaturee353091 -
Update config.json
Verified signature5903979 -
Update config.json
Verified signature5e4855c
-
Update README.md
Verified signature94a7184 -
Upload custom-code python model
Verified signature7b9818b -
Delete onnx/model-quantized.onnx
Verified signatured282b16 -
Update model_quantized.onnx
Verified signature3a509e6 -
Upload preprocessor_config.json
Verified signatureccbde49
-
Upload folder using huggingface_hub
Verified signature83262a0 -
Create README.md
Verified signaturef866902 -
initial commit
Verified signatureab0eb08