|
|
069b27570f
|
set option to set training masking ratio (I don't think for tts a fixed masking ratio is beneficial since the magic of the AR+NAR is being able to still reference the prior sequence of tokens for predicting things)
|
2024-11-17 17:04:07 -06:00 |
|
|
|
23fdba0c98
|
tweaks and changes
|
2024-11-16 15:49:06 -06:00 |
|
|
|
39096f8ff3
|
redid loss calculation to be cleaner, and position ID generation, and other things (I might need to train the NAR-len from scratch and not resume from an existing checkpoint.........)
|
2024-11-14 22:17:47 -06:00 |
|
|
|
ef05c951ff
|
adjust fp16 loss scaling since I fried a model overnight when it hit 8K scale
|
2024-11-14 09:23:52 -06:00 |
|
|
|
910033343c
|
overhauled how the right resp level / classifier gets picked to avoid cringemath
|
2024-11-13 13:31:17 -06:00 |
|
|
|
269648605e
|
move NAR-len rvq level 0 to separate embedding
|
2024-11-13 11:38:58 -06:00 |
|
|
|
2f56696506
|
overhauled inference/sampler kwargs to stop being a bloated mess
|
2024-11-11 20:21:16 -06:00 |
|
|
|
cf9df71f2c
|
use homwbrewed caching system for dataloader paths / durations (I'm pretty sure I am now triggering OOM killers with my entire dataset used)
|
2024-11-11 16:32:08 -06:00 |
|
|
|
a748e223ce
|
tweaks
|
2024-11-11 12:40:41 -06:00 |
|
|
|
9cb0b6901b
|
unified nar.py into ar_nar.py
|
2024-11-10 12:19:48 -06:00 |
|
|
|
c6a38693a2
|
This better work
|
2024-11-09 18:04:59 -06:00 |
|
|
|
bcabde3454
|
more notes
|
2024-11-06 13:51:28 -06:00 |
|
|
|
c83670c38c
|
Windows specific fixes (to-do: find libespeak-ng.dll automatically because it cannot be trusted to do it by default)
|
2024-11-03 19:19:15 -06:00 |
|
|
|
d229725c76
|
more adjustments (adjustments of early-exit entropy/varentropy thresholds, default rep pen being 1.5, experimental refine-on-stop, etc.)
|
2024-11-03 18:31:28 -06:00 |
|
|
|
fb8faa295b
|
actually float16(+AMP) and layerskip is bad and will kill the model......
|
2024-11-01 18:36:44 -05:00 |
|
|
|
edf1e66bf9
|
layerskip_r=6 fries the model so hard the loss is sub-1...
|
2024-11-01 17:06:07 -05:00 |
|
|
|
9b6c57bc57
|
third time's the charm (for some reason it escaped me that I should treat early exit loss as an aux_loss to be used with the normal loss, as if I was training a MoE's router)
|
2024-11-01 12:50:37 -05:00 |
|
|
|
76ebef45dc
|
off-by-one...
|
2024-10-31 13:24:48 -05:00 |
|
|
|
a22534e8f4
|
layer skip training implemented (need to gut the inferencing from the repo, and to actually see if the model can benefit from this)
|
2024-10-30 20:05:45 -05:00 |
|
|
|
4049f51ba9
|
added option to load lora directly from the model file itself with --lora
|
2024-10-26 00:13:10 -05:00 |
|
|
|
ccf71dc1b6
|
added option to load from a model state dict directly instead of a yaml (to-do: do this for LoRAs too), automatically download the default model if none is provided
|
2024-10-25 22:15:15 -05:00 |
|
|
|
a96f5aee32
|
adjusted how i want to pass eval kwargs
|
2024-10-25 20:38:09 -05:00 |
|
|
|
8920e5e86b
|
actually have beam_width in the webUI work
|
2024-10-22 22:06:22 -05:00 |
|
|
|
8eb9a4056b
|
modified default arguments (ar temp = 0 and rep pen = 1.125 seems to be stable, at least given the few things i tested), do not pass top k/top p/min p to NAR even though technically none of those things should matter when greedy sampling
|
2024-10-22 18:12:39 -05:00 |
|
|
|
c8f31db1de
|
default to greedy sample AR (i should probably test this more but it seems to pass my harvard sentences and tongue twisters)
|
2024-10-18 16:58:56 -05:00 |
|
|
|
07f4935a75
|
more tweaks
|
2024-10-18 13:19:36 -05:00 |
|
|
|
0dfab973e7
|
oops
|
2024-10-18 09:40:06 -05:00 |
|
|
|
75b90be325
|
cleaned up unused config flags, allow less strict yaml by pruning missing keys, renamed some dataset configs to be more unified
|
2024-10-17 17:06:48 -05:00 |
|
|
|
f88097ccf6
|
add config option to set the rate of sampling randomly vs similar speakers during training
|
2024-10-16 14:27:58 -05:00 |
|
|
|
04e983b86b
|
modified demo page to be more modular with demoing comparisons, actually provide a path to use modified naive attention, entropix sampling is not tied to an experimental yaml flag now
|
2024-10-12 11:27:55 -05:00 |
|
|
|
bef43a0c18
|
added experimental entropix sampling support
|
2024-10-11 21:18:26 -05:00 |
|
|
|
f24547ad4e
|
add top_k sampling / offset for prompt similar utterance sampling
|
2024-09-26 16:26:40 -05:00 |
|
|
|
c5e9142863
|
added option to retokenize phonemes for hdf5 (to save having to remake my hdf5 file)
|
2024-09-21 13:08:01 -05:00 |
|
|
|
fe241f6a99
|
support for wildcard in training/validation/noise dataset array (to-do: a better way to query between metadata folder and data folder)
|
2024-09-18 21:34:43 -05:00 |
|
|
|
ebac1db16c
|
maybe final tweaks, I really needed to unify my json read/write and orjson is proven to be fast enough for me to try and rely on it more
|
2024-09-17 22:57:04 -05:00 |
|
|
|
56f25f7a9b
|
more stuff for similar-speaker prompt sampling (to-do: actually test if this works...)
|
2024-09-16 23:10:29 -05:00 |
|
|
|
54203c059d
|
validated rep pen for STT (sometimes needed to wrangle the model)
|
2024-09-08 08:30:30 -05:00 |
|
|
|
5d66a7db52
|
webui cleanup, more tweaks, default to safetensors in config
|
2024-09-07 21:45:05 -05:00 |
|
|
|
54547b74d8
|
experimental implementation of STT (need to actually test on a model, test trainer seems to work)
|
2024-09-05 20:43:20 -05:00 |
|
|
|
32287710a2
|
moved prints to use logger, edited readme (fused_attn doesnt seem stable for training)
|
2024-08-29 13:27:16 -05:00 |
|
|
|
ed373957e2
|
maybe not
|
2024-08-09 11:38:08 -05:00 |
|
|
|
c658a7b440
|
make loss scaling opt-in rather than automatically determined (because it seems a DAC-based model really doesnt like loss scaling)
|
2024-08-09 10:51:36 -05:00 |
|
|
|
eac353cd0b
|
busy work and cleanup while I wait for 1TB of audio to quantize... again.
|
2024-08-06 20:23:33 -05:00 |
|
|
|
2cb465018b
|
implicitly load either normal pickled weights or safetensors on loading the model
|
2024-08-03 23:34:18 -05:00 |
|
|
|
c09133d00f
|
added safetensors support (with metadata) and feed whatever torch.load/torch.save into it
|
2024-08-03 23:15:20 -05:00 |
|
|
|
6a733eb2ed
|
changed torch.Tensor().to(device, dtype) to just torch.tensor(..., device, dtype) because it's been bothering my autism that I'm creating tensors then converting rather than creating with the right device/dtype, some 'optimization' to compile the model but it doesnt seem to do anything useful
|
2024-08-03 22:10:21 -05:00 |
|
|
|
11fa3da665
|
some cleanup, fixed the wrapper attention to explicitly use other sdpa backends
|
2024-08-03 19:51:00 -05:00 |
|
|
|
c9ec6b28ef
|
it actually wasn't working because Engines.__init__() automatically moves the entire module to the requested device, which was being called after offloading the model in the test trainer (and it seems I cant do it without injecting a bunch of shit in modeling_llama.py)
|
2024-08-01 20:56:28 -05:00 |
|
|
|
b4c895114c
|
naive model offloading support (handles automatically splitting parts of the model to requested device per memory constraints, either inferred or requested in the yaml, input tensors are automatically migrated to the right device, it SEEMS to work for training under the test trainer when split between GPU and CPU) (this was specifically only because that Flux imagegen model released so I can test it there)
|
2024-08-01 20:12:06 -05:00 |
|
|
|
387358bc8a
|
fixes for the NAR-len model, and documentation some config options, and a better way to handle resizing modules on state_dict load
|
2024-07-31 20:35:09 -05:00 |
|
|
|
52d13b321f
|
I rather have it default to non-strict loading instead so I can clean up YAMLs
|
2024-07-30 22:24:38 -05:00 |
|
|
|
07f8e2ad06
|
added option to set the causal size (how many tokens to sample per AR step), but requires the model to be trained for this (which explains why recurrent chunk sampling just doesn't work for the retnet tests, obvious in hindsight)
|
2024-07-30 20:53:51 -05:00 |
|
|
|
682e4387dc
|
oops (fixed proms being erased from a config oversight)
|
2024-07-25 12:39:57 -05:00 |
|
|
|
1acb0e9c84
|
added experimental training setting to perform token dropout to MAYBE compensate for errors from the preceding RVQ level (two types: token error offset, token dropout embedding replace)
|
2024-07-24 19:35:17 -05:00 |
|
|
|
75b04686f8
|
added prom-less training / inferencing, some other things
|
2024-07-22 19:36:07 -05:00 |
|
|
|
e19aa643a6
|
cleaned up demo page creation, added option to pass in RVQ level sampling distribution for training
|
2024-07-21 19:12:03 -05:00 |
|
|
|
d53038a9e4
|
actually have split classifiers working
|
2024-07-19 15:33:31 -05:00 |
|
|
|
83a0954f85
|
fixes for re-introducing SpeechX tasks (need to actually validate if these all do the right things)
|
2024-07-18 17:16:32 -05:00 |
|
|
|
bccbb77a1a
|
added option to either naively concat codes to concat audio waveforms (prior behavior) or to decode => concat => encode instead (although this only currently happens for prom sampling if an utternace is too small)
|
2024-07-18 16:48:41 -05:00 |
|
|
|
97e768601c
|
re-introducing SpeechX tasks (need to validate them all, everything works with base tts anyways)
|
2024-07-18 16:16:14 -05:00 |
|
|
|
22fe53508c
|
added experimental disjointed position IDs (because I *think* this might help because technically a sequence is made up of several parts, and the position embeddings shouldn't be unified)
|
2024-07-16 19:52:41 -05:00 |
|
|
|
fe0f235335
|
mechanism to store the model config inside the weights and load them, some other things to allow LoRA training on the RetNet (gradient checkpointing will gripe about inputs not having require_grad and nothing seems to remedy it)
|
2024-07-16 18:23:13 -05:00 |
|
|
|
3acc54df22
|
allow loading a different model within the web ui (apparently I did not have the web UI in the documentation)
|
2024-07-15 19:59:48 -05:00 |
|
|
|
7b210d9738
|
sanity cleanup
|
2024-07-04 15:58:08 -05:00 |
|
|
|
1ecf2793f4
|
(commented-out) support for facebookresearch/AudioDec, but support really didn't wow me (so I commented it out until I figure out why my output audio is super crusty with AudioDec)
|
2024-07-04 15:40:51 -05:00 |
|
|
|
f770467eb3
|
stuff
|
2024-07-01 18:13:29 -05:00 |
|
|
|
312a8e3ead
|
add shuffle to samplers that can support it
|
2024-06-30 11:36:46 -05:00 |
|
|
|
396af541c5
|
ugh
|
2024-06-30 11:11:58 -05:00 |
|
|
|
dced595391
|
more cleanup
|
2024-06-30 11:00:12 -05:00 |
|
|
|
bc2a6fa756
|
sanity cleanup: moved experimental features under its own thing
|
2024-06-30 10:37:33 -05:00 |
|
|
|
2808f881c8
|
cleaned up subjugated audio embedding into a flag, flag can also have it include the original, underlying embedding as well (it seems to do better when set to inclusive)
|
2024-06-29 21:46:35 -05:00 |
|
|
|
ec5eaebcbc
|
experimental method of using DACs quantizer ""embeddings"" to see if it helps with model quality
|
2024-06-29 19:46:11 -05:00 |
|
|
|
83075c1505
|
sort duration buckets to ensure that paths sorted-by-duration are actually sorted by duration (because i didnt know that python dicts can have non-strings as keys), added batching samples based on total duration to ensure best training throughput
|
2024-06-28 22:28:54 -05:00 |
|
|
|
8fffb94964
|
backport fix from tortoise_tts with local trainer + loading state when training lora
|
2024-06-25 13:41:29 -05:00 |
|
|
|
62a53eed64
|
fixed deducing tokenizer path, added option to default to naive tokenizer (for old models, like ar+nar-retnet-8)
|
2024-06-18 22:11:14 -05:00 |
|
|
|
8a986eb480
|
load exported LoRA weights if exists (to-do: make a better LoRA loading mechanism)
|
2024-06-18 21:45:46 -05:00 |
|
|
|
7cfb78fa64
|
enable LoRA for targetted RVQ levels (to experiment with, seems to help)
|
2024-06-17 21:45:03 -05:00 |
|
|
|
1d159b1476
|
updated export routine to split LoRA weights from the state dict (should work with deepspeed)
|
2024-06-17 13:28:18 -05:00 |
|
|
|
bd0bc10ec0
|
added LoRA policy to decide what layer of the model gets adapted based on simple inclusion/exclusion terms
|
2024-06-17 13:05:06 -05:00 |
|
|
|
45a39fb79f
|
very rudimentary lora support (no deepspeed support, tested training and saving but not loading yet)
|
2024-06-17 00:09:16 -05:00 |
|
|
|
b3b67f34ac
|
added option to sort paths by durations to better group equally lengthed sequences together (and there was maybe a logic error from creating the samplers and then interleave-reordering paths, desyncing them, maybe)
|
2024-06-13 22:37:34 -05:00 |
|
|
|
65a8960305
|
option to split classifier per-level instead of sharing one (at this point I'm just scrambling to try and cope with training a DAC model, the NAR is being a pain)
|
2024-06-11 22:28:59 -05:00 |
|
|
|
a7a6e0ac76
|
validated that inferencing works, changed some defaults (NAR benefits from greedy sampling)
|
2024-06-09 17:11:38 -05:00 |
|
|
|
132a02c48b
|
sanity cleanup, backup config yaml for each log file
|
2024-06-09 11:22:52 -05:00 |
|
|
|
58fb0a84db
|
added experimental NAR only model (inferences text length, need more experimenting), AudioEmbedding logic cleanup (I still think it's being done wrong)
|
2024-06-08 15:42:02 -05:00 |
|
|
|
e35a91c67a
|
ugh
|
2024-06-07 21:56:14 -05:00 |
|
|
|
eafa622be2
|
I forgot the actual reason I was cleaning things up was to re-include prom loss calculation (I realized the reason I did this was because of an prom embedding oversight, it seems to work now)
|
2024-06-07 20:29:25 -05:00 |
|
|
|
da8242d086
|
finally got around to removing omegaconf
|
2024-06-07 20:23:53 -05:00 |
|
|
|
b2194b859a
|
re-added loading multiple models because I'm now entertaining having split AR/NAR models again (and need a way to load both at once)
|
2024-06-06 09:48:43 -05:00 |
|
|
|
4073656293
|
oops
|
2024-06-05 20:53:10 -05:00 |
|
|
|
48cd1054f9
|
madness
|
2024-06-04 23:48:51 -05:00 |
|
|
|
406ff7bbe1
|
re-implemented config.model.interleave for the HF-compat experimental method
|
2024-06-04 14:19:52 -05:00 |
|
|
|
934672252b
|
feverish cleanup
|
2024-06-03 21:28:49 -05:00 |
|
|
|
c1fcd889d5
|
reverted automatically disabling split loss calc, since it seems that it's actually cacling loss on prom causes the oddities, maybe
|
2024-06-01 12:34:59 -05:00 |
|
|
|
31785f4eeb
|
actually don't default to compute split losses, test bitnet model doesn't seem to be doing things right (despite debug printouts showing theyre roughly the same logit/loss sequences, could just be bitnet linears being not up to par on actual models)
|
2024-06-01 09:12:51 -05:00 |
|
|
|
e9c87060df
|
oops
|
2024-05-31 22:22:28 -05:00 |
|
|
|
b482ca19ff
|
added model config option to set KV head count for MQA/GQA instead of MHA for llama-based models (i think its very negligible both ways on such a small model size)
|
2024-05-31 19:32:37 -05:00 |
|
|
|
da473295b7
|
better way to compute per-segment losses
|
2024-05-28 19:29:54 -05:00 |
|
|
|
5af6f41c94
|
added loss calcs against prom (requires the right settings for not shit results, disabled by default)
|
2024-05-27 08:43:00 -05:00 |
|
|
|
ddbacde0d1
|
DAC just doesn't work well enough......
|
2024-05-25 11:07:52 -05:00 |
|
