capacitron_layers multi speaker bug fix #1664
Merged
Conversation
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
crashing because of dimension miss match at line no. 57 [batch, 256] vs [batch , 1, 512] enc_out = torch.cat([enc_out, speaker_embedding], dim=-1)
manmay-nakhashi
changed the title
|
multi-speaker with an embedding layer or with d-vectors? I think it'd work without squeezing with d-vectors |
|
i tried with embedding layer, let me check with d-vectors |
WeberJulian
approved these changes
Jul 21, 2022
|
@WeberJulian you mean tried this PR? |
erogol
added a commit
that referenced
this pull request
Aug 22, 2022
* Fix checkpointing GAN models (#1641) * checkpoint sae step crash fix * checkpoint save step crash fix * Update gan.py updated requested changes * crash fix * Fix the --model_name and --vocoder_name arguments need a <model_type> element (#1469) Co-authored-by: Eren Gölge <erogol@hotmail.com> * Fix Publish CI (#1597) * Try out manylinux * temporary removal of useless pipeline * remove check and use only manylinux * Try --plat-name * Add install requirements * Add back other actions * Add PR trigger * Remove conditions * Fix sythax * Roll back some changes * Add other python versions * Add test pypi upload * Add username * Add back __token__ as username * Modify name of entry to testpypi * Set it to release only * Fix version checking * Fix tokenizer for punc only (#1717) * Remove redundant config field * Fix SSIM loss * Separate loss tests * Fix BCELoss adressing #1192 * Make style * Add durations as aux input for VITS (#1694) * Add durations as aux input for VITS * Make style * Fix tts_tests * Fix test_get_aux_input * Make lint * feat: updated recipes and lr fix (#1718) - updated the recipes activating more losses for more stable training - re-enabling guided attention loss - fixed a bug about not the correct lr fetched for logging * Implement VitsAudioConfig (#1556) * Implement VitsAudioConfig * Update VITS LJSpeech recipe * Update VITS VCTK recipe * Make style * Add missing decorator * Add missing param * Make style * Update recipes * Fix test * Bug fix * Exclude tests folder * Make linter * Make style * Fix device allocation * Fix SSIM loss correction * Fix aux tests (#1753) * Set n_jobs to 1 for resample script * Delete resample test * Set n_jobs 1 in vad test * delete vad test * Revert "Delete resample test" This reverts commit bb7c846. * Remove tests with resample * Fix for FloorDiv Function Warning (#1760) * Fix for Floor Function Warning Fix for Floor Function Warning * Adding double quotes to fix formatting Adding double quotes to fix formatting * Update glow_tts.py * Update glow_tts.py * Fix type in download_vctk.sh (#1739) typo in comment * Update decoder.py (#1792) Minor comment correction. * Update requirements.txt (#1791) Support for #1775 * Update README.md (#1776) Fix typo in different and code sample * Fix & update WaveRNN vocoder model (#1749) * Fixes KeyError bug. Adding logging to dashboard. * Make pep8 compliant * Make style compliant * Still fixing style * Fix rand_segment edge case (input_len == seg_len - 1) * Update requirements.txt; inflect==5.6 (#1809) New inflect version (6.0) depends on pydantic which has some issues irrelevant to 🐸 TTS. #1808 Force inflect==5.6 (pydantic free) install to solve dependency issue. * Update README.md; download progress bar in CLI. (#1797) * Update README.md - minor PR - added model_info usage guide based on #1623 in README.md . * "added tqdm bar for model download" * Update manage.py * fixed style * fixed style * sort imports * Update wavenet.py (#1796) * Update wavenet.py Current version does not use "in_channels" argument. In glowTTS, we use normalizing flows and so "input dim" == "ouput dim" (channels and length). So, the existing code just uses hidden_channel sized tensor as input to first layer as well as outputs hidden_channel sized tensor. However, since it is a generic implementation, I believe it is better to update it for a more general use. * "in_channels -> hidden_channels" * Adjust default to be able to process longer sentences (#1835) Running `tts --text "$text" --out_path …` with a somewhat longer sentences in the text will lead to warnings like “Decoder stopped with max_decoder_steps 500” and the sentences just being cut off in the resulting WAV file. This happens quite frequently when feeding longer texts (e.g. a blog post) to `tts`. It's particular frustrating since the error is not always obvious in the output. You have to notice that there are missing parts. This is something other users seem to have run into as well [1]. This patch simply increases the maximum number of steps allowed for the tacotron decoder to fix this issue, resulting in a smoother default behavior. [1] mozilla/TTS#734 * Fix language flags generated by espeak-ng phonemizer (#1801) * fix language flags generated by espeak-ng phonemizer * Style * Updated language flag regex to consider all language codes alike * fix get_random_embeddings --> get_random_embedding (#1726) * fix get_random_embeddings --> get_random_embedding function typo leads to training crash, no such function * fix typo get_random_embedding * Introduce numpy and torch transforms (#1705) * Refactor audio processing functions * Add tests for numpy transforms * Fix imports * Fix imports2 * Implement bucketed weighted sampling for VITS (#1871) * Update capacitron_layers.py (#1664) crashing because of dimension miss match at line no. 57 [batch, 256] vs [batch , 1, 512] enc_out = torch.cat([enc_out, speaker_embedding], dim=-1) * updates to dataset analysis notebooks for compatibility with latest version of TTS (#1853) * Fix BCE loss issue (#1872) * Fix BCE loss issue * Remove import * Remove deprecated files (#1873) - samplers.py is moved - distribute.py is replaces by the 👟Trainer * Handle when no batch sampler (#1882) * Fix tune wavegrad (#1844) * fix imports in tune_wavegrad * load_config returns Coqpit object instead None * set action (store true) for flag "--use_cuda"; start to tune if module is running as the main program * fix var order in the result of batch collating * make style * make style with black and isort * Bump up to v0.8.0 * Add new DE Thorsten models (#1898) - Tacotron2-DDC - HifiGAN vocoder Co-authored-by: manmay nakhashi <manmay.nakhashi@gmail.com> Co-authored-by: camillem <camillem@users.noreply.github.com> Co-authored-by: WeberJulian <julian.weber@hotmail.fr> Co-authored-by: a-froghyar <adamfroghyar@gmail.com> Co-authored-by: ivan provalov <iprovalo@yahoo.com> Co-authored-by: Tsai Meng-Ting <sarah13680@gmail.com> Co-authored-by: p0p4k <rajiv.punmiya@gmail.com> Co-authored-by: Yuri Pourre <yuripourre@users.noreply.github.com> Co-authored-by: vanIvan <alfa1211@gmail.com> Co-authored-by: Lars Kiesow <lkiesow@uos.de> Co-authored-by: rbaraglia <baraglia.r@live.fr> Co-authored-by: jchai.me <jreus@users.noreply.github.com> Co-authored-by: Stanislav Kachnov <42406556+geth-network@users.noreply.github.com>
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
crashing for multi speaker training because of dimension miss match at line no. 57
[batch, 256] vs [batch , 1, 512]
enc_out = torch.cat([enc_out, speaker_embedding], dim=-1)