Closed roedoejet closed 2 months ago
semanticdiff-com[bot]
commented
2 months ago Review changes with SemanticDiff.
Analyzed 2 of 3 files.
| Filename | Status | |
|---|---|---|
| :heavy_check_mark: | everyvoice/cli.py | Analyzed |
| :heavy_check_mark: | everyvoice/model/vocoder/HiFiGAN_iSTFT_lightning | Analyzed |
| :grey_question: | .github/workflows/test.yml | Unsupported file format |
github-actions[bot]
commented
2 months ago CLI load time: 0:00.24
Pull Request HEAD: 96bc4d6a7e6eeb19dccd7138054d2a2848188516
Imports that take more than 0.1 s:
import time: self [us] | cumulative | imported package
codecov[bot]
commented
2 months ago All modified and coverable lines are covered by tests :white_check_mark:
Project coverage is 77.90%. Comparing base (
eb460a2) to head (96bc4d6). Report is 9 commits behind head on main.
:umbrella: View full report in Codecov by Sentry.
:loudspeaker: Have feedback on the report? Share it here.
joanise
commented
2 months ago Hum, something is broken with our pre-commit CI workaround: ecd5478d6d ran isort and broke what was previously correctly isorted.
joanise
commented
2 months ago CI failed due to #555. Fixed by excluding submodules from pre-commit run in CI.
This PR now also fixes #555
roedoejet
commented
2 months ago CI failed due to #555. Fixed by excluding submodules from pre-commit run in CI.
This PR now also fixes #555
Thanks @joanise ! yes, I saw this last week but couldn't see what was happening. thanks for digging into this!
joanise
commented
2 months ago The CI commit is now separate in #557, we can remove it from this branch once that PR is merged, just rebasing on the (future) updated main should do.
marctessier
commented
2 months ago Running some tests.
Wondering in the help message if we should add that the path to the vocoder model is needed as an argument?
everyvoice export --help
Usage: everyvoice export [OPTIONS] COMMAND [ARGS]...
┏━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━┓
┃ Export Help ┃
┗━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━┛
• spec-to-wav --- You can export your spec-to-wav model to a much smaller format for synthesis. Advanced: this will export only the
generator, leaving the weights of the discriminators behind.
╭─ Options ────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────╮
│ --help -h Show this message and exit. │
╰──────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────╯
╭─ Commands ───────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────╮
│ spec-to-wav Export and optimize a spec-to-wav model for inference │
╰──────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────╯
(EveryVoice) [U20-GPSC7]:$ everyvoice export spec-to-wav
╭───────────────────────────────────────────────────── Traceback (most recent call last) ──────────────────────────────────────────────────────╮
│ /gpfs/fs5/nrc/nrc-fs1/ict/others/u/tes001/TxT2SPEECH/EveryVoice/everyvoice/model/vocoder/HiFiGAN_iSTFT_lightning/hfgl/cli.py:94 in export │
│ │
│ 91 │ │
│ 92 │ from .utils import sizeof_fmt │
│ 93 │ │
│ ❱ 94 │ orig_size = sizeof_fmt(os.path.getsize(model_path)) │
│ 95 │ vocoder_ckpt = torch.load(model_path, map_location=torch.device("cpu")) │
│ 96 │ for k in list(vocoder_ckpt["state_dict"].keys()): │
│ 97 │ │ if not k.startswith("generator"): │
│ │
│ /home/tes001/u/TxT2SPEECH/miniconda3_u20/envs/EveryVoice/lib/python3.10/genericpath.py:50 in getsize │
│ │
│ 47 │
│ 48 def getsize(filename): │
│ 49 │ """Return the size of a file, reported by os.stat().""" │
│ ❱ 50 │ return os.stat(filename).st_size │
│ 51 │
│ 52 │
│ 53 def getmtime(filename): │
╰──────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────╯
TypeError: stat: path should be string, bytes, os.PathLike or integer, not NoneType
I am running a bit more test but looks good up to now :-) I think I was expecting the exported vocoder be placed in the "./logs_and_checkpoints/VocoderExperiment/base/checkpoints/" but I see it was exported to the location where I ran the script. That should also be good and easier I think.
everyvoice export spec-to-wav ./logs_and_checkpoints/VocoderExperiment/base/checkpoints/last.ckpt
2024-09-23 11:40:07.137 | INFO | everyvoice.model.vocoder.HiFiGAN_iSTFT_lightning.hfgl.cli:export:105 - Checkpoint saved at 'exported.ckpt'. Reduced size from 969.4MiB to 53.3MiB. This checkpoint will only be usable for inference/synthesis, and not for training.Running some tests.
Wondering in the help message if we should add that the path to the vocoder model is needed as an argument?
everyvoice export --help Usage: everyvoice export [OPTIONS] COMMAND [ARGS]... ┏━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━┓ ┃ Export Help ┃ ┗━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━┛ • spec-to-wav --- You can export your spec-to-wav model to a much smaller format for synthesis. Advanced: this will export only the generator, leaving the weights of the discriminators behind. ╭─ Options ────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────╮ │ --help -h Show this message and exit. │ ╰──────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────╯ ╭─ Commands ───────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────╮ │ spec-to-wav Export and optimize a spec-to-wav model for inference │ ╰──────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────╯ (EveryVoice) [U20-GPSC7]:$ everyvoice export spec-to-wav ╭───────────────────────────────────────────────────── Traceback (most recent call last) ──────────────────────────────────────────────────────╮ │ /gpfs/fs5/nrc/nrc-fs1/ict/others/u/tes001/TxT2SPEECH/EveryVoice/everyvoice/model/vocoder/HiFiGAN_iSTFT_lightning/hfgl/cli.py:94 in export │ │ │ │ 91 │ │ │ 92 │ from .utils import sizeof_fmt │ │ 93 │ │ │ ❱ 94 │ orig_size = sizeof_fmt(os.path.getsize(model_path)) │ │ 95 │ vocoder_ckpt = torch.load(model_path, map_location=torch.device("cpu")) │ │ 96 │ for k in list(vocoder_ckpt["state_dict"].keys()): │ │ 97 │ │ if not k.startswith("generator"): │ │ │ │ /home/tes001/u/TxT2SPEECH/miniconda3_u20/envs/EveryVoice/lib/python3.10/genericpath.py:50 in getsize │ │ │ │ 47 │ │ 48 def getsize(filename): │ │ 49 │ """Return the size of a file, reported by os.stat().""" │ │ ❱ 50 │ return os.stat(filename).st_size │ │ 51 │ │ 52 │ │ 53 def getmtime(filename): │ ╰──────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────╯ TypeError: stat: path should be string, bytes, os.PathLike or integer, not NoneTypeI am running a bit more test but looks good up to now :-) I think I was expecting the exported vocoder be placed in the "./logs_and_checkpoints/VocoderExperiment/base/checkpoints/" but I see it was exported to the location where I ran the script. That should also be good and easier I think.
everyvoice export spec-to-wav ./logs_and_checkpoints/VocoderExperiment/base/checkpoints/last.ckpt 2024-09-23 11:40:07.137 | INFO | everyvoice.model.vocoder.HiFiGAN_iSTFT_lightning.hfgl.cli:export:105 - Checkpoint saved at 'exported.ckpt'. Reduced size from 969.4MiB to 53.3MiB. This checkpoint will only be usable for inference/synthesis, and not for training.
Nice catch, yes the model path is required and should give a better message. I just fixed this.
allow synthesis with only the generator checkpoint
fixes https://github.com/EveryVoiceTTS/EveryVoice/issues/424
PR Goal?
This PR allows users to "export" vocoder models (
everyvoice export) which trims down the model by removing the discriminators. It also allows loading the vocoder model and inference/synthesis (not training) using only the generator. This allows us to distribute 50MB files instead of 950MB files when all people want to do is synthesize (not fine-tune/resume training)Fixes?
https://github.com/EveryVoiceTTS/EveryVoice/issues/424
Feedback sought?
sanitty
Priority?
high - alpha 4
Tests added?
Again, we do not have tests for inference unfortunately
How to test?
take an existing vocoder and run
everyvoice export <path_to_vocoder.ckpt>then try synthesizing using the generatedexported.ckptcheckpoint.Confidence?
medium-high
Version change?
yes, alpha 4
Related PRs?
https://github.com/EveryVoiceTTS/HiFiGAN_iSTFT_lightning/pull/34