Added experimental guided setup

Reviewed-on: mrq/ai-voice-cloning#57

.gitignore

@@ -1,7 +1,8 @@
 # ignores user files
-/tortoise-venv/
+/venv/
-/tortoise/voices/
+/voices/*
-/models/
+/models/*
+/training/*
 /config/*
 
 # Byte-compiled / optimized / DLL files

.gitmodules

@@ -0,0 +1,6 @@
+[submodule "tortoise-tts"]
+	path = tortoise-tts
+	url = https://git.ecker.tech/mrq/tortoise-tts
+[submodule "dlas"]
+	path = dlas
+	url = https://git.ecker.tech/mrq/DL-Art-School

README.md

@@ -16,4 +16,19 @@ Please consult [the wiki](https://git.ecker.tech/mrq/ai-voice-cloning/wiki) for
 
 ## Bug Reporting
 
-If you run into any problems, please refer to the [issues you may encounter](https://git.ecker.tech/mrq/ai-voice-cloning/wiki/Issues) wiki page first. Please don't hesitate to submit an issue.
+If you run into any problems, please refer to the [issues you may encounter](https://git.ecker.tech/mrq/ai-voice-cloning/wiki/Issues) wiki page first. Please don't hesitate to submit an issue.
+
+## Changelogs
+
+Below will be a rather-loose changelogss, as I don't think I have a way to chronicle them outside of commit messages:
+
+### `2023.02.22`
+
+* greatly reduced VRAM consumption through the use of [TimDettmers/bitsandbytes](https://github.com/TimDettmers/bitsandbytes)
+* cleaned up section of code that handled parsing output from training script
+* added button to reconnect to the training script's output (sometimes skips a line to update, but it's better than nothing)
+* actually update submodules from the update script (somehow forgot to pass `--remote`)
+
+### `Before 2023.02.22`
+
+Refer to commit logs.

dlas

@@ -0,0 +1 @@
+Subproject commit 6eb7ebf847cf2e4761536391de841dc4209d1e63

models/.template.yaml

@@ -1,21 +1,21 @@
 name: ${name}
 model: extensibletrainer
 scale: 1
-gpu_ids: [0] # <-- unless you have multiple gpus, use this
+gpu_ids: [0] # Superfluous, redundant, unnecessary, the way you launch the training script will set this
-start_step: -1
+start_step: 0
-checkpointing_enabled: true  # <-- Gradient checkpointing. Enable for huge GPU memory savings. Disable for distributed training.
+checkpointing_enabled: true 
-fp16: false # might want to check this out
+fp16: ${float16}
-wandb: false  # <-- enable to log to wandb. tensorboard logging is always enabled.
+wandb: false 
 use_tb_logger: true
 
 datasets:
   train:
     name: ${dataset_name}
-    n_workers: 8 # idk what this does
+    n_workers: ${workers}
-    batch_size: ${batch_size} # This leads to ~16GB of vram usage on my 3090.
+    batch_size: ${batch_size}
     mode: paired_voice_audio
     path: ${dataset_path}
-    fetcher_mode: ['lj'] # CHANGEME if your dataset isn't in LJSpeech format
+    fetcher_mode: ['lj']
     phase: train
     max_wav_length: 255995
     max_text_length: 200
@@ -26,14 +26,14 @@ datasets:
     use_bpe_tokenizer: True
     tokenizer_vocab: ./models/tortoise/bpe_lowercase_asr_256.json
     load_aligned_codes: False
-  val:
+  val: # I really do not care about validation right now
     name: ${validation_name}
     n_workers: 1
-    batch_size: 32 # this could be higher probably
+    batch_size: 1
     mode: paired_voice_audio
     path: ${validation_path}
     fetcher_mode: ['lj']
-    phase: val # might be broken idk
+    phase: val
     max_wav_length: 255995
     max_text_length: 200
     sample_rate: 22050
@@ -47,18 +47,18 @@ datasets:
 steps:        
   gpt_train:
     training: gpt
-    loss_log_buffer: 500 # no idea what this does
+    loss_log_buffer: 500
 
     # Generally follows the recipe from the DALLE paper.
     optimizer: adamw # this should be adamw_zero if you're using distributed training
     optimizer_params:
-      lr: !!float ${learning_rate} # CHANGEME: this was originally 1e-4; I reduced it to 1e-5 because it's fine-tuning, but **you should experiment with this value**
+      lr: !!float ${learning_rate} # originally: 1e-4
       weight_decay: !!float 1e-2
       beta1: 0.9
       beta2: 0.96
     clip_grad_eps: 4
 
-    injectors:  # TODO: replace this entire sequence with the GptVoiceLatentInjector
+    injectors:
       paired_to_mel:
         type: torch_mel_spectrogram
         mel_norm_file: ./models/tortoise/clips_mel_norms.pth
@@ -74,7 +74,7 @@ steps:
         type: discrete_token
         in: paired_mel
         out: paired_mel_codes
-        dvae_config: "./models/tortoise/train_diffusion_vocoder_22k_level.yml" # EXTREMELY IMPORTANT
+        dvae_config: "./models/tortoise/train_diffusion_vocoder_22k_level.yml"
       paired_fwd_text:
         type: generator
         generator: gpt
@@ -83,7 +83,7 @@ steps:
     losses:
       text_ce:
         type: direct
-        weight: .01
+        weight: ${text_ce_lr_weight}
         key: loss_text_ce
       mel_ce:
         type: direct
@@ -95,12 +95,12 @@ networks:
     type: generator 
     which_model_G: unified_voice2 # none of the unified_voice*.py files actually match the tortoise inference code... 4 and 3 have "alignment_head" (wtf is that?), 2 lacks the types=1 parameter.
     kwargs:
-      layers: 30 # WAS 8
+      layers: 30 # originally: 8
-      model_dim: 1024 # WAS 512
+      model_dim: 1024 # originally: 512
-      heads: 16 # WAS 8
+      heads: 16 # originally: 8
-      max_text_tokens: 402 # WAS 120
+      max_text_tokens: 402 # originally: 120
-      max_mel_tokens: 604 # WAS 250
+      max_mel_tokens: 604 # originally: 250
-      max_conditioning_inputs: 2 # WAS 1
+      max_conditioning_inputs: 2 # originally: 1
       mel_length_compression: 1024
       number_text_tokens: 256 # supposed to be 255 for newer unified_voice files 
       number_mel_codes: 8194
@@ -114,19 +114,20 @@ networks:
       #only_alignment_head: False  # uv3/4
 
 path:
-  pretrain_model_gpt: './models/tortoise/autoregressive.pth' # CHANGEME: copy this from tortoise cache
+  ${pretrain_model_gpt} 
   strict_load: true
-  #resume_state: ./models/tortoise/train_imgnet_vqvae_stage1/training_state/0.state   # <-- Set this to resume from a previous training state.
+  ${resume_state}
 
-# afaik all units here are measured in **steps** (i.e. one batch of batch_size is 1 unit)
+train:
-train: # CHANGEME: ALL OF THESE PARAMETERS SHOULD BE EXPERIMENTED WITH
   niter: ${iterations}
   warmup_iter: -1
-  mega_batch_factor: 4    # <-- Gradient accumulation factor. If you are running OOM, increase this to [2,4,8].
+  mega_batch_factor: ${gradient_accumulation_size}
-  val_freq: 500
+  val_freq: ${iterations}
+
+  ema_enabled: false # I really don't think EMA matters
 
   default_lr_scheme: MultiStepLR
-  gen_lr_steps: [500, 1000, 1400, 1800] #[50000, 100000, 140000, 180000]
+  gen_lr_steps: ${gen_lr_steps} #[50000, 100000, 140000, 180000]
   lr_gamma: 0.5
 
 eval:
@@ -140,7 +141,7 @@ eval:
 
 logger: 
   print_freq: ${print_rate}
-  save_checkpoint_freq: ${save_rate} # CHANGEME: especially you should increase this it's really slow
+  save_checkpoint_freq: ${save_rate}
   visuals: [gen, mel]
   visual_debug_rate: ${print_rate}
   is_mel_spectrogram: true

notebook.ipynb

@@ -40,15 +40,23 @@
          "source":[
             "!git clone https://git.ecker.tech/mrq/ai-voice-cloning/\n",
             "%cd ai-voice-cloning\n",
+            "\n",
+            "!git submodule init\n",
+            "!git submodule update --remote\n",
+            "\n",
             "# TODO: fix venvs working for subprocess.Popen calling a bash script\n",
             "#!apt install python3.8-venv\n",
             "#!python -m venv venv\n",
             "#!source ./venv/bin/activate\n",
-            "!git clone https://git.ecker.tech/mrq/DL-Art-School dlas\n",
+            "\n",
             "!python -m pip install --upgrade pip\n",
             "!pip install torch torchvision torchaudio --extra-index-url https://download.pytorch.org/whl/cu116\n",
             "!python -m pip install -r ./dlas/requirements.txt\n",
-            "!python -m pip install -r ./requirements.txt"
+            "!python -m pip install -r ./tortoise-tts/requirements.txt\n",
+            "!python -m pip install -r ./requirements.txt\n",
+            "!python -m pip install -e ./tortoise-tts/\n",
+            "\n",
+            "!rm ./tortoise-tts/{main,webui}.py"
          ]
       },
       {
@@ -64,13 +72,8 @@
          "cell_type":"code",
          "source":[
             "# for my debugging purposes\n",
-            "%cd /content/ai-voice-cloning/dlas\n",
+            "%cd /content/ai-voice-cloning/\n",
-            "!git reset --hard HEAD\n",
+            "!./update.sh"
-            "!git pull\n",
-            "%cd ..\n",
-            "!git reset --hard HEAD\n",
-            "!git pull\n",
-            "#exit()"
          ],
          "metadata":{
             "id":"3DktoOXSHmtw"
@@ -93,21 +96,18 @@
          "cell_type":"code",
          "source":[
             "# only run once, this will save all userdata to your Drive\n",
+            "# it shouldn't delete through symlinks, but you never know\n",
+            "\n",
             "from google.colab import drive\n",
-            "drive.mount('/content/drive')\n",
             "\n",
             "%cd /content/ai-voice-cloning\n",
-            "!rm -r ./training\n",
+            "drive.flush_and_unmount()\n",
-            "!rm -r ./results\n",
+            "!rm -r ./{training,results,voices,config}\n",
-            "!rm -r ./voices\n",
+            "drive.mount('/content/drive')\n",
-            "\n",
+            "!mkdir /content/drive/MyDrive/ai-voice-cloning/\n",
-            "!mkdir /content/drive/MyDrive/training/\n",
+            "!mv /content/drive/MyDrive/{training,results,voices,config} /content/drive/MyDrive/ai-voice-cloning\n",
-            "!mkdir /content/drive/MyDrive/results/\n",
+            "!mkdir /content/drive/MyDrive/ai-voice-cloning/{training,results,voices,config}\n",
-            "!mkdir /content/drive/MyDrive/voices/\n",
+            "!ln -s /content/drive/MyDrive/ai-voice-cloning/{training,results,voices,config} ./"
-            "\n",
-            "!ln -s /content/drive/MyDrive/training/\n",
-            "!ln -s /content/drive/MyDrive/results/\n",
-            "!ln -s /content/drive/MyDrive/voices/"
          ],
          "metadata":{
             "id":"SGt9gyvubveT"
@@ -120,7 +120,7 @@
       {
          "cell_type":"markdown",
          "source":[
-            "## Running"
+            "## Running (Inlined)"
          ],
          "metadata":{
             "id":"o1gkfw3B3JSk"
@@ -136,7 +136,10 @@
             "import sys\n",
             "\n",
             "sys.argv = [\"\"]\n",
-            "sys.path.append('./src/')\n",
+            "if './src/' not in sys.path:\n",
+            "\tsys.path.append('./src/')\n",
+            "if './tortoise-tts/' not in sys.path:\n",
+            "\tsys.path.append('./tortoise-tts/')\n",
             "\n",
             "if 'TORTOISE_MODELS_DIR' not in os.environ:\n",
             "\tos.environ['TORTOISE_MODELS_DIR'] = os.path.realpath(os.path.join(os.getcwd(), './models/tortoise/'))\n",
@@ -144,18 +147,18 @@
             "if 'TRANSFORMERS_CACHE' not in os.environ:\n",
             "\tos.environ['TRANSFORMERS_CACHE'] = os.path.realpath(os.path.join(os.getcwd(), './models/transformers/'))\n",
             "\n",
-            "from utils import *\n",
+            "import utils\n",
-            "from webui import *\n",
+            "import webui\n",
             "\n",
-            "args = setup_args()\n",
+            "args = utils.setup_args()\n",
-            "\n",
+            "ui = webui.setup_gradio()\n",
-            "webui = setup_gradio()\n",
             "# Be very, very sure to check \"Defer TTS Load\" in Settings, then restart, before you start training\n",
             "# You'll crash the runtime if you don't\n",
             "if not args.defer_tts_load:\n",
-            "\ttts = setup_tortoise()\n",
+            "\tutils.setup_tortoise()\n",
-            "webui.launch(share=True, prevent_thread_lock=True, height=1000)\n",
+            "\n",
-            "webui.block_thread()"
+            "ui.launch(share=True, prevent_thread_lock=True, height=1000)\n",
+            "ui.block_thread()"
          ],
          "metadata":{
             "id":"c_EQZLTA19c7"
@@ -165,6 +168,52 @@
             
          ]
       },
+      {
+         "cell_type":"markdown",
+         "source":[
+            "## Running (non-inlined)"
+         ],
+         "metadata":{
+            "id":"EM3iNqgJF6Be"
+         }
+      },
+      {
+         "cell_type":"code",
+         "source":[
+            "%cd /content/ai-voice-cloning/\n",
+            "!./start.sh"
+         ],
+         "metadata":{
+            "id":"QRA8jF3cF-YJ"
+         },
+         "execution_count":null,
+         "outputs":[
+            
+         ]
+      },
+      {
+         "cell_type":"markdown",
+         "source":[
+            "# Restart Runtime"
+         ],
+         "metadata":{
+            "id":"vH9KU7SMGDxb"
+         }
+      },
+      {
+         "cell_type":"code",
+         "source":[
+            "!rm ./tortoise-tts/{main,webui}.py\n",
+            "exit()"
+         ],
+         "metadata":{
+            "id":"EWeyUPvgGDX5"
+         },
+         "execution_count":null,
+         "outputs":[
+            
+         ]
+      },
       {
          "cell_type":"markdown",
          "source":[
@@ -201,6 +250,7 @@
       {
          "cell_type":"code",
          "source":[
+            "# if you're not using drive mounting\n",
             "%cd /content/ai-voice-cloning\n",
             "!apt install -y p7zip-full\n",
             "from datetime import datetime\n",

requirements.txt

@@ -1,6 +1,3 @@
-git+https://git.ecker.tech/mrq/tortoise-tts.git
-# git+https://git.ecker.tech/mrq/DL-Art-School.git
-git+https://github.com/openai/whisper.git
 more-itertools
 ffmpeg-python
 gradio

setup-cuda.bat

@@ -1,8 +1,18 @@
+git submodule init
+git submodule update --remote
+
 python -m venv venv
 call .\venv\Scripts\activate.bat
 python -m pip install --upgrade pip
-python -m pip install torch torchvision torchaudio --extra-index-url https://download.pytorch.org/whl/cu116
+python -m pip install torch torchvision torchaudio --extra-index-url https://download.pytorch.org/whl/cu117
-python -m pip install -r ./requirements.txt
+python -m pip install -r .\requirements.txt
-.\setup-training.bat
+python -m pip install -r .\tortoise-tts\requirements.txt
-deactivate
+python -m pip install -e .\tortoise-tts\
-pause
+python -m pip install -r .\dlas\requirements.txt
+
+xcopy .\dlas\bitsandbytes_windows\* .\venv\Lib\site-packages\bitsandbytes\. /Y
+xcopy .\dlas\bitsandbytes_windows\cuda_setup\* .\venv\Lib\site-packages\bitsandbytes\cuda_setup\. /Y
+xcopy .\dlas\bitsandbytes_windows\nn\* .\venv\Lib\site-packages\bitsandbytes\nn\. /Y
+
+pause
+deactivate

setup-cuda.sh

@@ -1,8 +1,17 @@
 #!/bin/bash
-python -m venv venv
+# get local dependencies
+git submodule init
+git submodule update --remote
+# setup venv
+python3 -m venv venv
 source ./venv/bin/activate
-python -m pip install --upgrade pip
+python3 -m pip install --upgrade pip # just to be safe
-pip install torch torchvision torchaudio --extra-index-url https://download.pytorch.org/whl/cu116
+# CUDA
-python -m pip install -r ./requirements.txt
+pip3 install torch torchvision torchaudio --extra-index-url https://download.pytorch.org/whl/cu117
-./setup-training.sh
+# install requirements
-deactivate
+python3 -m pip install -r ./requirements.txt # install local requirements
+python3 -m pip install -r ./tortoise-tts/requirements.txt # install TorToiSe requirements
+python3 -m pip install -e ./tortoise-tts/ # install TorToiSe
+python3 -m pip install -r ./dlas/requirements.txt # instal DLAS requirements, last, because whisperx will break a dependency here
+
+deactivate

setup-directml.bat

@@ -1,8 +1,14 @@
+git submodule init
+git submodule update --remote
+
 python -m venv venv
 call .\venv\Scripts\activate.bat
 python -m pip install --upgrade pip
-python -m pip install torch torchvision torchaudio torch-directml==0.1.13.1.dev230119
+python -m pip install torch torchvision torchaudio torch-directml
-python -m pip install -r ./requirements.txt
+python -m pip install -r .\requirements.txt
-.\setup-training.bat
+python -m pip install -r .\tortoise-tts\requirements.txt
-deactivate
+python -m pip install -e .\tortoise-tts\
-pause
+python -m pip install -r .\dlas\requirements.txt
+
+pause
+deactivate

setup-guided.sh

@@ -0,0 +1,54 @@
+#!/bin/bash
+
+if ! command -v git &> /dev/null; then
+	if [[ "$(read -e -p 'Could not find git. Continue? [y/N]> '; echo $REPLY)" != [Yy]* ]]; then exit 1; fi
+else
+	printf "git - ok\n"
+fi
+
+# TODO: This could be more precise. e.g. checking for python3, then python, etc.
+if ! command -v python &> /dev/null; then
+	if [[ "$(read -e -p 'Could not find python. Continue? [y/N]> '; echo $REPLY)" != [Yy]* ]]; then exit 1; fi
+else
+	python -c 'import sys; sys.stderr.write("Wrong python version.\n") if sys.version_info.major != 3 else sys.stderr.write("Python 3 - ok\n")'
+fi
+
+printf "Which GPU brand do you have?\n"
+gpus=(nvidia amd)
+
+gpu=""
+while [ "$gpu" = "" ]; do
+	select gpu in $(printf '%s\n' ${gpus[@]}); do break; done
+done
+
+if [ $gpu = "nvidia" ]; then
+	./setup-cuda.sh
+elif [ $gpu = "amd" ]; then
+	./setup-rocm.sh
+fi
+
+source ./venv/bin/activate
+
+printf "Which Whisper backend would you like to use?\n"
+whisper_backends=("openai/whisper" "m-bain/whisperx" "lightmare/whispercpp")
+
+whisper_backend=""
+while [ "$whisper_backend" = "" ]; do
+	select whisper_backend in $(printf '%s\n' ${whisper_backends[@]}); do break; done
+done
+
+if [ $whisper_backend = "openai/whisper" ]; then
+	python -m pip install git+https://github.com/openai/whisper.git
+elif [ $whisper_backend = "m-bain/whisperx" ]; then
+	python -m pip install git+https://github.com/m-bain/whisperx.git
+elif [ $whisper_backend = "lightmare/whispercpp" ]; then
+	# This depends on SemVer
+	# Git > v2.18 for `--sort`
+	# Git > v2.4 for `versionsort.suffix`
+	# For older versions:
+	# git ls-remote --refs --tags https://git.ecker.tech/lightmare/whispercpp.py | cut --delimiter='/' --fields=3 | tr '-' '~' | sort --version-sort | tail --lines=1
+	WHISPERCPP_LATEST=$(git -c 'versionsort.suffix=-' ls-remote --exit-code --refs --sort='version:refname' --tags https://git.ecker.tech/lightmare/whispercpp.py '*.*.*' | tail -n 1 | cut --delimiter='/' --fields=3)
+	python -m pip install git+https://git.ecker.tech/lightmare/whispercpp.py@$WHISPERCPP_LATEST
+fi
+
+deactivate

setup-rocm-bnb.sh

@@ -0,0 +1,8 @@
+#!/bin/bash
+source ./venv/bin/activate
+git clone https://git.ecker.tech/mrq/bitsandbytes-rocm
+cd bitsandbytes-rocm
+make hip
+CUDA_VERSION=gfx1030 python setup.py install # assumes you're using a 6XXX series card
+python3 -m bitsandbytes # to validate it works
+cd ..

setup-rocm.sh

@@ -1,9 +1,19 @@
 #!/bin/bash
-python -m venv venv
+# get local dependencies
+git submodule init
+git submodule update --remote
+# setup venv
+python3 -m venv venv
 source ./venv/bin/activate
-python -m pip install --upgrade pip
+python3 -m pip install --upgrade pip # just to be safe
 # ROCM
-pip install torch torchvision torchaudio --extra-index-url https://download.pytorch.org/whl/rocm5.1.1 # 5.2 does not work for me desu
+pip3 install torch torchvision torchaudio --extra-index-url https://download.pytorch.org/whl/rocm5.1.1 # 5.2 does not work for me desu
-python -m pip install -r ./requirements.txt
+# install requirements
-./setup-training.sh
+python3 -m pip install -r ./requirements.txt # install local requirements
-deactivate
+python3 -m pip install -r ./tortoise-tts/requirements.txt # install TorToiSe requirements
+python3 -m pip install -e ./tortoise-tts/ # install TorToiSe
+python3 -m pip install -r ./dlas/requirements.txt # instal DLAS requirements
+# swap to ROCm version of BitsAndBytes
+pip3 uninstall -y bitsandbytes
+./setup-rocm-bnb.sh
+deactivate

setup-training.bat

@@ -1,2 +0,0 @@
-git clone https://git.ecker.tech/mrq/DL-Art-School dlas
-python -m pip install -r .\dlas\requirements.txt

setup-training.sh

@@ -1,3 +0,0 @@
-#!/bin/bash
-git clone https://git.ecker.tech/mrq/DL-Art-School dlas
-python -m pip install -r ./dlas/requirements.txt

src/main.py

@@ -1,14 +1,14 @@
 import os
 
-from utils import *
-from webui import *
-
 if 'TORTOISE_MODELS_DIR' not in os.environ:
 	os.environ['TORTOISE_MODELS_DIR'] = os.path.realpath(os.path.join(os.getcwd(), './models/tortoise/'))
 
 if 'TRANSFORMERS_CACHE' not in os.environ:
 	os.environ['TRANSFORMERS_CACHE'] = os.path.realpath(os.path.join(os.getcwd(), './models/transformers/'))
 
+from utils import *
+from webui import *
+
 if __name__ == "__main__":
 	args = setup_args()
 

src/train.py

@@ -1,8 +1,34 @@
-import torch
-import argparse
-
 import os
 import sys
+import argparse
+import yaml
+
+"""
+if 'BITSANDBYTES_OVERRIDE_LINEAR' not in os.environ:
+    os.environ['BITSANDBYTES_OVERRIDE_LINEAR'] = '0'
+if 'BITSANDBYTES_OVERRIDE_EMBEDDING' not in os.environ:
+    os.environ['BITSANDBYTES_OVERRIDE_EMBEDDING'] = '1'
+if 'BITSANDBYTES_OVERRIDE_ADAM' not in os.environ:
+    os.environ['BITSANDBYTES_OVERRIDE_ADAM'] = '1'
+if 'BITSANDBYTES_OVERRIDE_ADAMW' not in os.environ:
+    os.environ['BITSANDBYTES_OVERRIDE_ADAMW'] = '1'
+"""
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser()
+    parser.add_argument('-opt', type=str, help='Path to option YAML file.', default='../options/train_vit_latent.yml', nargs='+') # ugh
+    parser.add_argument('--launcher', choices=['none', 'pytorch'], default='none', help='job launcher')
+    args = parser.parse_args()
+    args.opt = " ".join(args.opt) # absolutely disgusting
+
+    with open(args.opt, 'r') as file:
+        opt_config = yaml.safe_load(file)
+
+    if "ext" in opt_config and "bitsandbytes" in opt_config["ext"] and not opt_config["ext"]["bitsandbytes"]:
+        os.environ['BITSANDBYTES_OVERRIDE_LINEAR'] = '0'
+        os.environ['BITSANDBYTES_OVERRIDE_EMBEDDING'] = '0'
+        os.environ['BITSANDBYTES_OVERRIDE_ADAM'] = '0'
+        os.environ['BITSANDBYTES_OVERRIDE_ADAMW'] = '0'
 
 # this is some massive kludge that only works if it's called from a shell and not an import/PIP package
 # it's smart-yet-irritating module-model loader breaks when trying to load something specifically when not from a shell
@@ -19,6 +45,7 @@ sys.path.insert(0, './dlas/')
 # don't even really bother trying to get DLAS PIP'd
 # without kludge, it'll have to be accessible as `codes` and not `dlas`
 
+import torch
 from codes import train as tr
 from utils import util, options as option
 
@@ -44,7 +71,7 @@ def train(yaml, launcher='none'):
         print('Disabled distributed training.')
     else:
         opt['dist'] = True
-        init_dist('nccl')
+        tr.init_dist('nccl')
         trainer.world_size = torch.distributed.get_world_size()
         trainer.rank = torch.distributed.get_rank()
         torch.cuda.set_device(torch.distributed.get_rank())
@@ -53,9 +80,14 @@ def train(yaml, launcher='none'):
     trainer.do_training()
 
 if __name__ == "__main__":
-    parser = argparse.ArgumentParser()
+    # simple check because I'm brain damaged and forgot I can't modify what a module exports by simply changing the booleans that decide what it exports after the fact
-    parser.add_argument('-opt', type=str, help='Path to option YAML file.', default='../options/train_vit_latent.yml')
+    try:
-    parser.add_argument('--launcher', choices=['none', 'pytorch'], default='none', help='job launcher')
+        import torch_intermediary
-    args = parser.parse_args()
+        if torch_intermediary.OVERRIDE_ADAM:
+            print("Using BitsAndBytes ADAMW optimizations")
+        else:
+            print("NOT using BitsAndBytes ADAMW optimizations")
+    except Exception as e:
+        pass
 
     train(args.opt, args.launcher)

src/webui.py

@@ -16,6 +16,7 @@ from datetime import datetime
 
 import tortoise.api
 from tortoise.utils.audio import get_voice_dir, get_voices
+from tortoise.utils.device import get_device_count
 
 from utils import *
 
@@ -45,35 +46,40 @@ def run_generation(
 	experimental_checkboxes,
 	progress=gr.Progress(track_tqdm=True)
 ):
+	if not text:
+		raise gr.Error("Please provide text.")
+	if not voice:
+		raise gr.Error("Please provide a voice.")
+
 	try:
 		sample, outputs, stats = generate(
-			text,
+			text=text,
-			delimiter,
+			delimiter=delimiter,
-			emotion,
+			emotion=emotion,
-			prompt,
+			prompt=prompt,
-			voice,
+			voice=voice,
-			mic_audio,
+			mic_audio=mic_audio,
-			voice_latents_chunks,
+			voice_latents_chunks=voice_latents_chunks,
-			seed,
+			seed=seed,
-			candidates,
+			candidates=candidates,
-			num_autoregressive_samples,
+			num_autoregressive_samples=num_autoregressive_samples,
-			diffusion_iterations,
+			diffusion_iterations=diffusion_iterations,
-			temperature,
+			temperature=temperature,
-			diffusion_sampler,
+			diffusion_sampler=diffusion_sampler,
-			breathing_room,
+			breathing_room=breathing_room,
-			cvvp_weight,
+			cvvp_weight=cvvp_weight,
-			top_p,
+			top_p=top_p,
-			diffusion_temperature,
+			diffusion_temperature=diffusion_temperature,
-			length_penalty,
+			length_penalty=length_penalty,
-			repetition_penalty,
+			repetition_penalty=repetition_penalty,
-			cond_free_k,
+			cond_free_k=cond_free_k,
-			experimental_checkboxes,
+			experimental_checkboxes=experimental_checkboxes,
-			progress
+			progress=progress
 		)
 	except Exception as e:
 		message = str(e)
 		if message == "Kill signal detected":
-			reload_tts()
+			unload_tts()
 
 		raise gr.Error(message)
 	
@@ -82,31 +88,9 @@ def run_generation(
 		outputs[0],
 		gr.update(value=sample, visible=sample is not None),
 		gr.update(choices=outputs, value=outputs[0], visible=len(outputs) > 1, interactive=True),
-		gr.update(visible=len(outputs) > 1),
 		gr.update(value=stats, visible=True),
 	)
 
-def compute_latents(voice, voice_latents_chunks, progress=gr.Progress(track_tqdm=True)):
-	global tts
-	global args
-
-	if not tts:
-		raise Exception("TTS is uninitialized or still initializing...")
-
-	voice_samples, conditioning_latents = load_voice(voice, load_latents=False)
-
-	if voice_samples is None:
-		return
-
-	conditioning_latents = tts.get_conditioning_latents(voice_samples, return_mels=not args.latents_lean_and_mean, progress=progress, slices=voice_latents_chunks, force_cpu=args.force_cpu_for_conditioning_latents)
-
-	if len(conditioning_latents) == 4:
-		conditioning_latents = (conditioning_latents[0], conditioning_latents[1], conditioning_latents[2], None)
-			
-	torch.save(conditioning_latents, f'{get_voice_dir()}/{voice}/cond_latents.pth')
-
-	return voice
-
 def update_presets(value):
 	PRESETS = {
 		'Ultra Fast': {'num_autoregressive_samples': 16, 'diffusion_iterations': 30, 'cond_free': False},
@@ -146,6 +130,9 @@ history_headers = {
 	"Rep Pen": "repetition_penalty",
 	"Cond-Free K": "cond_free_k",
 	"Time": "time",
+	"Datetime": "datetime",
+	"Model": "model",
+	"Model Hash": "model_hash",
 }
 
 def history_view_results( voice ):
@@ -164,7 +151,7 @@ def history_view_results( voice ):
 		for k in history_headers:
 			v = file
 			if k != "Name":
-				v = metadata[headers[k]]
+				v = metadata[history_headers[k]] if history_headers[k] in metadata else '?'
 			values.append(v)
 
 
@@ -193,15 +180,117 @@ def read_generate_settings_proxy(file, saveAs='.temp'):
 
 	return (
 		gr.update(value=j, visible=j is not None),
-		gr.update(visible=j is not None),
 		gr.update(value=latents, visible=latents is not None),
-		None if j is None else j['voice']
+		None if j is None else j['voice'],
+		gr.update(visible=j is not None),
 	)
 
-def prepare_dataset_proxy( voice, language, progress=gr.Progress(track_tqdm=True) ):
+def prepare_dataset_proxy( voice, language, skip_existings, progress=gr.Progress(track_tqdm=True) ):
-	return prepare_dataset( get_voices(load_latents=False)[voice], outdir=f"./training/{voice}/", language=language, progress=progress )
+	return prepare_dataset( get_voices(load_latents=False)[voice], outdir=f"./training/{voice}/", language=language, skip_existings=skip_existings, progress=progress )
 
-def save_training_settings_proxy( iterations, batch_size, learning_rate, print_rate, save_rate, voice ):
+def optimize_training_settings_proxy( *args, **kwargs ):
+	tup = optimize_training_settings(*args, **kwargs)
+
+	return (
+		gr.update(value=tup[0]),
+		gr.update(value=tup[1]),
+		gr.update(value=tup[2]),
+		gr.update(value=tup[3]),
+		gr.update(value=tup[4]),
+		gr.update(value=tup[5]),
+		gr.update(value=tup[6]),
+		gr.update(value=tup[7]),
+		"\n".join(tup[8])
+	)
+
+def import_training_settings_proxy( voice ):
+	indir = f'./training/{voice}/'
+	outdir = f'./training/{voice}-finetune/'
+
+	in_config_path = f"{indir}/train.yaml"
+	out_config_path = None
+	out_configs = []
+	if os.path.isdir(outdir):
+		out_configs = sorted([d[:-5] for d in os.listdir(outdir) if d[-5:] == ".yaml" ])
+	if len(out_configs) > 0:
+		out_config_path = f'{outdir}/{out_configs[-1]}.yaml'
+
+	config_path = out_config_path if out_config_path else in_config_path
+
+	messages = []
+	with open(config_path, 'r') as file:
+		config = yaml.safe_load(file)
+		messages.append(f"Importing from: {config_path}")
+
+	dataset_path = f"./training/{voice}/train.txt"
+	with open(dataset_path, 'r', encoding="utf-8") as f:
+		lines = len(f.readlines())
+		messages.append(f"Basing epoch size to {lines} lines")
+
+	batch_size = config['datasets']['train']['batch_size']
+	gradient_accumulation_size = config['train']['mega_batch_factor']
+
+	iterations = config['train']['niter']
+	steps_per_iteration = int(lines / batch_size)
+	epochs = int(iterations / steps_per_iteration)
+
+
+	learning_rate = config['steps']['gpt_train']['optimizer_params']['lr']
+	text_ce_lr_weight = config['steps']['gpt_train']['losses']['text_ce']['weight']
+	learning_rate_schedule = [ int(x / steps_per_iteration) for x in config['train']['gen_lr_steps'] ]
+
+
+	print_rate = int(config['logger']['print_freq'] / steps_per_iteration)
+	save_rate = int(config['logger']['save_checkpoint_freq'] / steps_per_iteration)
+
+	statedir = f'{outdir}/training_state/' # NOOO STOP MIXING YOUR CASES
+	resumes = []
+	resume_path = None
+	source_model = None
+	
+	if "pretrain_model_gpt" in config['path']:
+		source_model = config['path']['pretrain_model_gpt']
+	elif "resume_state" in config['path']:
+		resume_path = config['path']['resume_state']
+
+
+	if os.path.isdir(statedir):
+		resumes = sorted([int(d[:-6]) for d in os.listdir(statedir) if d[-6:] == ".state" ])
+
+	if len(resumes) > 0:
+		resume_path = f'{statedir}/{resumes[-1]}.state'
+		messages.append(f"Latest resume found: {resume_path}")
+
+
+	half_p = config['fp16']
+	bnb = True
+
+	if "ext" in config and "bitsandbytes" in config["ext"]:
+		bnb = config["ext"]["bitsandbytes"]
+
+	workers = config['datasets']['train']['n_workers']
+
+	messages = "\n".join(messages)
+
+	return (
+		epochs,
+		learning_rate,
+		text_ce_lr_weight,
+		learning_rate_schedule,
+		batch_size,
+		gradient_accumulation_size,
+		print_rate,
+		save_rate,
+		resume_path,
+		half_p,
+		bnb,
+		workers,
+		source_model,
+		messages
+	)
+
+
+def save_training_settings_proxy( epochs, learning_rate, text_ce_lr_weight, learning_rate_schedule, batch_size, gradient_accumulation_size, print_rate, save_rate, resume_path, half_p, bnb, workers, source_model, voice ):
 	name = f"{voice}-finetune"
 	dataset_name = f"{voice}-train"
 	dataset_path = f"./training/{voice}/train.txt"
@@ -211,17 +300,47 @@ def save_training_settings_proxy( iterations, batch_size, learning_rate, print_r
 	with open(dataset_path, 'r', encoding="utf-8") as f:
 		lines = len(f.readlines())
 
-	if batch_size > lines:
+	messages = []
-		print("Batch size is larger than your dataset, clamping...")
-		batch_size = lines
 
-	out_name = f"{voice}/train.yaml"
+	iterations = calc_iterations(epochs=epochs, lines=lines, batch_size=batch_size)
+	messages.append(f"For {epochs} epochs with {lines} lines, iterating for {iterations} steps")
 
-	return save_training_settings(iterations, batch_size, learning_rate, print_rate, save_rate, name, dataset_name, dataset_path, validation_name, validation_path, out_name )
+	print_rate = int(print_rate * iterations / epochs)
+	save_rate = int(save_rate * iterations / epochs)
+
+	if not learning_rate_schedule:
+		learning_rate_schedule = EPOCH_SCHEDULE
+	elif isinstance(learning_rate_schedule,str):
+		learning_rate_schedule = json.loads(learning_rate_schedule)
+
+	learning_rate_schedule = schedule_learning_rate( iterations / epochs, learning_rate_schedule )
+
+	messages.append(save_training_settings(
+		iterations=iterations,
+		batch_size=batch_size,
+		learning_rate=learning_rate,
+		text_ce_lr_weight=text_ce_lr_weight,
+		learning_rate_schedule=learning_rate_schedule,
+		gradient_accumulation_size=gradient_accumulation_size,
+		print_rate=print_rate,
+		save_rate=save_rate,
+		name=name,
+		dataset_name=dataset_name,
+		dataset_path=dataset_path,
+		validation_name=validation_name,
+		validation_path=validation_path,
+		output_name=f"{voice}/train.yaml",
+		resume_path=resume_path,
+		half_p=half_p,
+		bnb=bnb,
+		workers=workers,
+		source_model=source_model,
+	))
+	return "\n".join(messages)
 
 def update_voices():
 	return (
-		gr.Dropdown.update(choices=get_voice_list()),
+		gr.Dropdown.update(choices=get_voice_list(append_defaults=True)),
 		gr.Dropdown.update(choices=get_voice_list()),
 		gr.Dropdown.update(choices=get_voice_list("./results/")),
 	)
@@ -229,14 +348,6 @@ def update_voices():
 def history_copy_settings( voice, file ):
 	return import_generate_settings( f"./results/{voice}/{file}" )
 
-def update_model_settings( autoregressive_model, whisper_model ):
-	if args.autoregressive_model != autoregressive_model:
-		update_autoregressive_model(autoregressive_model)
-
-	args.whisper_model = whisper_model
-
-	save_args_settings()
-
 def setup_gradio():
 	global args
 	global ui
@@ -257,41 +368,60 @@ def setup_gradio():
 	if args.models_from_local_only:
 		os.environ['TRANSFORMERS_OFFLINE']='1'
 
+	voice_list_with_defaults = get_voice_list(append_defaults=True)
+	voice_list = get_voice_list()
+	result_voices = get_voice_list("./results/")
+	autoregressive_models = get_autoregressive_models()
+	dataset_list = get_dataset_list()
+
 	with gr.Blocks() as ui:
 		with gr.Tab("Generate"):
 			with gr.Row():
 				with gr.Column():
-					text = gr.Textbox(lines=4, label="Prompt")
+					text = gr.Textbox(lines=4, label="Input Prompt")
 			with gr.Row():
 				with gr.Column():
 					delimiter = gr.Textbox(lines=1, label="Line Delimiter", placeholder="\\n")
 
-					emotion = gr.Radio( ["Happy", "Sad", "Angry", "Disgusted", "Arrogant", "Custom"], value="Custom", label="Emotion", type="value", interactive=True )
+					emotion = gr.Radio( ["Happy", "Sad", "Angry", "Disgusted", "Arrogant", "Custom", "None"], value="None", label="Emotion", type="value", interactive=True )
-					prompt = gr.Textbox(lines=1, label="Custom Emotion + Prompt (if selected)")
+					prompt = gr.Textbox(lines=1, label="Custom Emotion")
-					voice = gr.Dropdown(get_voice_list(), label="Voice", type="value")
+					voice = gr.Dropdown(choices=voice_list_with_defaults, label="Voice", type="value", value=voice_list_with_defaults[0]) # it'd be very cash money if gradio was able to default to the first value in the list without this shit
-					mic_audio = gr.Audio( label="Microphone Source", source="microphone", type="filepath" )
+					mic_audio = gr.Audio( label="Microphone Source", source="microphone", type="filepath", visible=False )
-					refresh_voices = gr.Button(value="Refresh Voice List")
+					voice_latents_chunks = gr.Slider(label="Voice Chunks", minimum=1, maximum=128, value=1, step=1)
-					voice_latents_chunks = gr.Slider(label="Voice Chunks", minimum=1, maximum=64, value=1, step=1)
+					with gr.Row():
-					recompute_voice_latents = gr.Button(value="(Re)Compute Voice Latents")
+						refresh_voices = gr.Button(value="Refresh Voice List")
+						recompute_voice_latents = gr.Button(value="(Re)Compute Voice Latents")
+
+					voice.change(
+						fn=update_baseline_for_latents_chunks,
+						inputs=voice,
+						outputs=voice_latents_chunks
+					)
+					voice.change(
+						fn=lambda value: gr.update(visible=value == "microphone"),
+						inputs=voice,
+						outputs=mic_audio,
+					)
 				with gr.Column():
 					candidates = gr.Slider(value=1, minimum=1, maximum=6, step=1, label="Candidates")
 					seed = gr.Number(value=0, precision=0, label="Seed")
 
 					preset = gr.Radio( ["Ultra Fast", "Fast", "Standard", "High Quality"], label="Preset", type="value" )
-					num_autoregressive_samples = gr.Slider(value=128, minimum=0, maximum=512, step=1, label="Samples")
+					num_autoregressive_samples = gr.Slider(value=128, minimum=2, maximum=512, step=1, label="Samples")
 					diffusion_iterations = gr.Slider(value=128, minimum=0, maximum=512, step=1, label="Iterations")
 
 					temperature = gr.Slider(value=0.2, minimum=0, maximum=1, step=0.1, label="Temperature")
-					breathing_room = gr.Slider(value=8, minimum=1, maximum=32, step=1, label="Pause Size")
-					diffusion_sampler = gr.Radio(
-						["P", "DDIM"], # + ["K_Euler_A", "DPM++2M"],
-						value="P", label="Diffusion Samplers", type="value" )
 					show_experimental_settings = gr.Checkbox(label="Show Experimental Settings")
 					reset_generation_settings_button = gr.Button(value="Reset to Default")
 				with gr.Column(visible=False) as col:
 					experimental_column = col
 
 					experimental_checkboxes = gr.CheckboxGroup(["Half Precision", "Conditioning-Free"], value=["Conditioning-Free"], label="Experimental Flags")
+					breathing_room = gr.Slider(value=8, minimum=1, maximum=32, step=1, label="Pause Size")
+					diffusion_sampler = gr.Radio(
+						["P", "DDIM"], # + ["K_Euler_A", "DPM++2M"],
+						value="DDIM", label="Diffusion Samplers", type="value"
+					)
 					cvvp_weight = gr.Slider(value=0, minimum=0, maximum=1, label="CVVP Weight")
 					top_p = gr.Slider(value=0.8, minimum=0, maximum=1, label="Top P")
 					diffusion_temperature = gr.Slider(value=1.0, minimum=0, maximum=1, label="Diffusion Temperature")
@@ -299,25 +429,34 @@ def setup_gradio():
 					repetition_penalty = gr.Slider(value=2.0, minimum=0, maximum=8, label="Repetition Penalty")
 					cond_free_k = gr.Slider(value=2.0, minimum=0, maximum=4, label="Conditioning-Free K")
 				with gr.Column():
-					submit = gr.Button(value="Generate")
+					with gr.Row():
-					stop = gr.Button(value="Stop")
+						submit = gr.Button(value="Generate")
+						stop = gr.Button(value="Stop")
 
 					generation_results = gr.Dataframe(label="Results", headers=["Seed", "Time"], visible=False)
 					source_sample = gr.Audio(label="Source Sample", visible=False)
 					output_audio = gr.Audio(label="Output")
-					candidates_list = gr.Dropdown(label="Candidates", type="value", visible=False)
+					candidates_list = gr.Dropdown(label="Candidates", type="value", visible=False, choices=[""], value="")
-					output_pick = gr.Button(value="Select Candidate", visible=False)
+
+					def change_candidate( val ):
+						if not val:
+							return
+						return val
+
+					candidates_list.change(
+						fn=change_candidate,
+						inputs=candidates_list,
+						outputs=output_audio,
+					)
 		with gr.Tab("History"):
 			with gr.Row():
 				with gr.Column():
 					history_info = gr.Dataframe(label="Results", headers=list(history_headers.keys()))
 			with gr.Row():
 				with gr.Column():
-					history_voices = gr.Dropdown(choices=get_voice_list("./results/"), label="Voice", type="value")
+					history_voices = gr.Dropdown(choices=result_voices, label="Voice", type="value", value=result_voices[0] if len(result_voices) > 0 else "")
-					history_view_results_button = gr.Button(value="View Files")
 				with gr.Column():
-					history_results_list = gr.Dropdown(label="Results",type="value", interactive=True)
+					history_results_list = gr.Dropdown(label="Results",type="value", interactive=True, value="")
-					history_view_result_button = gr.Button(value="View File")
 				with gr.Column():
 					history_audio = gr.Audio()
 					history_copy_settings_button = gr.Button(value="Copy Settings")
@@ -327,17 +466,20 @@ def setup_gradio():
 					audio_in = gr.Files(type="file", label="Audio Input", file_types=["audio"])
 					import_voice_name = gr.Textbox(label="Voice Name")
 					import_voice_button = gr.Button(value="Import Voice")
-				with gr.Column():
+				with gr.Column(visible=False) as col:
-					metadata_out = gr.JSON(label="Audio Metadata", visible=False)
+					utilities_metadata_column = col
-					copy_button = gr.Button(value="Copy Settings", visible=False)
+
-					latents_out = gr.File(type="binary", label="Voice Latents", visible=False)
+					metadata_out = gr.JSON(label="Audio Metadata")
+					copy_button = gr.Button(value="Copy Settings")
+					latents_out = gr.File(type="binary", label="Voice Latents")
 		with gr.Tab("Training"):
 			with gr.Tab("Prepare Dataset"):
 				with gr.Row():
 					with gr.Column():
 						dataset_settings = [
-							gr.Dropdown( get_voice_list(), label="Dataset Source", type="value" ),
+							gr.Dropdown( choices=voice_list, label="Dataset Source", type="value", value=voice_list[0] if len(voice_list) > 0 else "" ),
-							gr.Textbox(label="Language", placeholder="English")
+							gr.Textbox(label="Language", value="en"),
+							gr.Checkbox(label="Skip Already Transcribed", value=False)
 						]
 						prepare_dataset_button = gr.Button(value="Prepare")
 					with gr.Column():
@@ -346,38 +488,78 @@ def setup_gradio():
 				with gr.Row():
 					with gr.Column():
 						training_settings = [
-							gr.Slider(label="Iterations", minimum=0, maximum=5000, value=500),
+							gr.Number(label="Epochs", value=500, precision=0),
-							gr.Slider(label="Batch Size", minimum=2, maximum=128, value=64),
-							gr.Slider(label="Learning Rate", value=1e-5, minimum=0, maximum=1e-4, step=1e-6),
-							gr.Number(label="Print Frequency", value=50),
-							gr.Number(label="Save Frequency", value=50),
 						]
-						dataset_list = gr.Dropdown( get_dataset_list(), label="Dataset", type="value" )
+						with gr.Row():
+							with gr.Column():
+								training_settings = training_settings + [
+									gr.Slider(label="Learning Rate", value=1e-5, minimum=0, maximum=1e-4, step=1e-6),
+									gr.Slider(label="Text_CE LR Ratio", value=0.01, minimum=0, maximum=1),
+								]
+							training_settings = training_settings + [
+								gr.Textbox(label="Learning Rate Schedule", placeholder=str(EPOCH_SCHEDULE)),
+							]
+						with gr.Row():
+							training_settings = training_settings + [
+								gr.Number(label="Batch Size", value=128, precision=0),
+								gr.Number(label="Gradient Accumulation Size", value=4, precision=0),
+							]
+						with gr.Row():
+							training_settings = training_settings + [
+								gr.Number(label="Print Frequency (in epochs)", value=5, precision=0),
+								gr.Number(label="Save Frequency (in epochs)", value=5, precision=0),
+							]
 						training_settings = training_settings + [
-							dataset_list
+							gr.Textbox(label="Resume State Path", placeholder="./training/${voice}-finetune/training_state/${last_state}.state"),
 						]
-						refresh_dataset_list = gr.Button(value="Refresh Dataset List")
+
-						"""
+						with gr.Row():
-						training_settings = training_settings + [
+							training_halfp = gr.Checkbox(label="Half Precision", value=args.training_default_halfp)
-							gr.Textbox(label="Training Name", placeholder="finetune"),
+							training_bnb = gr.Checkbox(label="BitsAndBytes", value=args.training_default_bnb)
-							gr.Textbox(label="Dataset Name", placeholder="finetune"),
+
-							gr.Textbox(label="Dataset Path", placeholder="./training/finetune/train.txt"),
+						training_workers = gr.Number(label="Worker Processes", value=2, precision=0)
-							gr.Textbox(label="Validation Name", placeholder="finetune"),
+
-							gr.Textbox(label="Validation Path", placeholder="./training/finetune/train.txt"),
+						source_model = gr.Dropdown( choices=autoregressive_models, label="Source Model", type="value", value=autoregressive_models[0] )
-						]
+						dataset_list_dropdown = gr.Dropdown( choices=dataset_list, label="Dataset", type="value", value=dataset_list[0] if len(dataset_list) else ""  )
-						"""
+						training_settings = training_settings + [ training_halfp, training_bnb, training_workers, source_model, dataset_list_dropdown ]
+
+						with gr.Row():
+							refresh_dataset_list = gr.Button(value="Refresh Dataset List")
+							import_dataset_button = gr.Button(value="Reuse/Import Dataset")
 					with gr.Column():
 						save_yaml_output = gr.TextArea(label="Console Output", interactive=False, max_lines=8)
-						save_yaml_button = gr.Button(value="Save Training Configuration")
+						with gr.Row():
+							optimize_yaml_button = gr.Button(value="Validate Training Configuration")
+							save_yaml_button = gr.Button(value="Save Training Configuration")
 			with gr.Tab("Run Training"):
 				with gr.Row():
 					with gr.Column():
 						training_configs = gr.Dropdown(label="Training Configuration", choices=get_training_list())
-						refresh_configs = gr.Button(value="Refresh Configurations")
+						with gr.Row():
-						start_training_button = gr.Button(value="Train")
+							refresh_configs = gr.Button(value="Refresh Configurations")
-						stop_training_button = gr.Button(value="Stop")
+						
+						training_loss_graph = gr.LinePlot(label="Training Metrics",
+							x="step",
+							y="value",
+							title="Training Metrics",
+							color="type",
+							tooltip=['step', 'value', 'type'],
+							width=600,
+							height=350,
+						)
+						view_losses = gr.Button(value="View Losses")
+
 					with gr.Column():
 						training_output = gr.TextArea(label="Console Output", interactive=False, max_lines=8)
+						verbose_training = gr.Checkbox(label="Verbose Console Output", value=True)
+						
+						with gr.Row():
+							training_keep_x_past_datasets = gr.Slider(label="Keep X Previous States", minimum=0, maximum=8, value=0, step=1)
+							training_gpu_count = gr.Number(label="GPUs", value=get_device_count())
+						with gr.Row():
+							start_training_button = gr.Button(value="Train")
+							stop_training_button = gr.Button(value="Stop")
+							reconnect_training_button = gr.Button(value="Reconnect")
 		with gr.Tab("Settings"):
 			with gr.Row():
 				exec_inputs = []
@@ -390,31 +572,59 @@ def setup_gradio():
 						gr.Checkbox(label="Low VRAM", value=args.low_vram),
 						gr.Checkbox(label="Embed Output Metadata", value=args.embed_output_metadata),
 						gr.Checkbox(label="Slimmer Computed Latents", value=args.latents_lean_and_mean),
-						gr.Checkbox(label="Voice Fixer", value=args.voice_fixer),
+						gr.Checkbox(label="Use Voice Fixer on Generated Output", value=args.voice_fixer),
 						gr.Checkbox(label="Use CUDA for Voice Fixer", value=args.voice_fixer_use_cuda),
 						gr.Checkbox(label="Force CPU for Conditioning Latents", value=args.force_cpu_for_conditioning_latents),
-						gr.Checkbox(label="Defer TTS Load", value=args.defer_tts_load),
+						gr.Checkbox(label="Do Not Load TTS On Startup", value=args.defer_tts_load),
+						gr.Checkbox(label="Delete Non-Final Output", value=args.prune_nonfinal_outputs),
+						gr.Checkbox(label="Use BigVGAN Vocoder", value=args.use_bigvgan_vocoder),
 						gr.Textbox(label="Device Override", value=args.device_override),
 					]
 				with gr.Column():
 					exec_inputs = exec_inputs + [
 						gr.Number(label="Sample Batch Size", precision=0, value=args.sample_batch_size),
-						gr.Number(label="Concurrency Count", precision=0, value=args.concurrency_count),
+						gr.Number(label="Gradio Concurrency Count", precision=0, value=args.concurrency_count),
-						gr.Number(label="Ouptut Sample Rate", precision=0, value=args.output_sample_rate),
+						gr.Number(label="Auto-Calculate Voice Chunk Duration (in seconds)", precision=0, value=args.autocalculate_voice_chunk_duration_size),
-						gr.Slider(label="Ouptut Volume", minimum=0, maximum=2, value=args.output_volume),
+						gr.Slider(label="Output Volume", minimum=0, maximum=2, value=args.output_volume),
 					]
 					
-					autoregressive_model_dropdown = gr.Dropdown(get_autoregressive_models(), label="Autoregressive Model", value=args.autoregressive_model)
+					autoregressive_model_dropdown = gr.Dropdown(choices=autoregressive_models, label="Autoregressive Model", value=args.autoregressive_model if args.autoregressive_model else autoregressive_models[0])
-					whisper_model_dropdown = gr.Dropdown(["tiny", "tiny.en", "base", "base.en", "small", "small.en", "medium", "medium.en", "large"], label="Whisper Model", value=args.whisper_model)
+					
-					save_settings_button = gr.Button(value="Save Settings")
+					whisper_backend = gr.Dropdown(WHISPER_BACKENDS, label="Whisper Backends", value=args.whisper_backend)
+					whisper_model_dropdown = gr.Dropdown(WHISPER_MODELS, label="Whisper Model", value=args.whisper_model)
+					
+					exec_inputs = exec_inputs + [ autoregressive_model_dropdown, whisper_backend, whisper_model_dropdown, training_halfp, training_bnb ]
 
-					gr.Button(value="Check for Updates").click(check_for_updates)
+					with gr.Row():
-					gr.Button(value="(Re)Load TTS").click(reload_tts)
+						autoregressive_models_update_button = gr.Button(value="Refresh Model List")
+						gr.Button(value="Check for Updates").click(check_for_updates)
+						gr.Button(value="(Re)Load TTS").click(
+							reload_tts,
+							inputs=autoregressive_model_dropdown,
+							outputs=None
+						)
+						# kill_button = gr.Button(value="Close UI")
+
+					def update_model_list_proxy( val ):
+						autoregressive_models = get_autoregressive_models()
+						if val not in autoregressive_models:
+							val = autoregressive_models[0]
+						return gr.update( choices=autoregressive_models, value=val )
+
+					autoregressive_models_update_button.click(
+						update_model_list_proxy,
+						inputs=autoregressive_model_dropdown,
+						outputs=autoregressive_model_dropdown,
+					)
 
 				for i in exec_inputs:
 					i.change( fn=update_args, inputs=exec_inputs )
-
+				
-		# console_output = gr.TextArea(label="Console Output", interactive=False, max_lines=8)
+				autoregressive_model_dropdown.change(
+					fn=update_autoregressive_model,
+					inputs=autoregressive_model_dropdown,
+					outputs=None
+				)
 
 		input_settings = [
 			text,
@@ -440,7 +650,7 @@ def setup_gradio():
 			experimental_checkboxes,
 		]
 
-		history_view_results_button.click(
+		history_voices.change(
 			fn=history_view_results,
 			inputs=history_voices,
 			outputs=[
@@ -448,7 +658,7 @@ def setup_gradio():
 				history_results_list,
 			]
 		)
-		history_view_result_button.click(
+		history_results_list.change(
 			fn=lambda voice, file: f"./results/{voice}/{file}",
 			inputs=[
 				history_voices,
@@ -461,9 +671,9 @@ def setup_gradio():
 			inputs=audio_in,
 			outputs=[
 				metadata_out,
-				copy_button,
 				latents_out,
-				import_voice_name
+				import_voice_name,
+				utilities_metadata_column,
 			]
 		)
 
@@ -496,9 +706,10 @@ def setup_gradio():
 			outputs=voice,
 		)
 		
-		prompt.change(fn=lambda value: gr.update(value="Custom"),
+		emotion.change(
-			inputs=prompt,
+			fn=lambda value: gr.update(visible=value == "Custom"),
-			outputs=emotion
+			inputs=emotion,
+			outputs=prompt
 		)
 		mic_audio.change(fn=lambda value: gr.update(value="microphone"),
 			inputs=mic_audio,
@@ -514,20 +725,15 @@ def setup_gradio():
 			]
 		)
 
-		output_pick.click(
-			lambda x: x,
-			inputs=candidates_list,
-			outputs=output_audio,
-		)
-
 		submit.click(
-			lambda: (gr.update(visible=False), gr.update(visible=False), gr.update(visible=False), gr.update(visible=False)),
+			lambda: (gr.update(visible=False), gr.update(visible=False), gr.update(visible=False)),
-			outputs=[source_sample, candidates_list, output_pick, generation_results],
+			outputs=[source_sample, candidates_list, generation_results],
 		)
 
 		submit_event = submit.click(run_generation,
 			inputs=input_settings,
-			outputs=[output_audio, source_sample, candidates_list, output_pick, generation_results],
+			outputs=[output_audio, source_sample, candidates_list, generation_results],
+			api_name="generate",
 		)
 
 
@@ -556,13 +762,45 @@ def setup_gradio():
 			outputs=training_configs
 		)
 		start_training_button.click(run_training,
-			inputs=training_configs,
+			inputs=[
-			outputs=training_output #console_output
+				training_configs,
+				verbose_training,
+				training_gpu_count,
+				training_keep_x_past_datasets,
+			],
+			outputs=[
+				training_output,
+			],
 		)
+		training_output.change(
+			fn=update_training_dataplot,
+			inputs=None,
+			outputs=[
+				training_loss_graph,
+			],
+			show_progress=False,
+		)
+
+		view_losses.click(
+			fn=update_training_dataplot,
+			inputs=[
+				training_configs
+			],
+			outputs=[
+				training_loss_graph,
+			],
+		)
+
 		stop_training_button.click(stop_training,
 			inputs=None,
 			outputs=training_output #console_output
 		)
+		reconnect_training_button.click(reconnect_training,
+			inputs=[
+				verbose_training,
+			],
+			outputs=training_output #console_output
+		)
 		prepare_dataset_button.click(
 			prepare_dataset_proxy,
 			inputs=dataset_settings,
@@ -571,20 +809,32 @@ def setup_gradio():
 		refresh_dataset_list.click(
 			lambda: gr.update(choices=get_dataset_list()),
 			inputs=None,
-			outputs=dataset_list,
+			outputs=dataset_list_dropdown,
+		)
+		optimize_yaml_button.click(optimize_training_settings_proxy,
+			inputs=training_settings,
+			outputs=training_settings[1:9] + [save_yaml_output] #console_output
+		)
+		import_dataset_button.click(import_training_settings_proxy,
+			inputs=dataset_list_dropdown,
+			outputs=training_settings[:13] + [save_yaml_output] #console_output
 		)
 		save_yaml_button.click(save_training_settings_proxy,
 			inputs=training_settings,
 			outputs=save_yaml_output #console_output
 		)
 
-		save_settings_button.click(update_model_settings,
+		"""
-			inputs=[
+		def kill_process():
-				autoregressive_model_dropdown,
+			ui.close()
-				whisper_model_dropdown,
+			exit()
-			],
+
+		kill_button.click(
+			kill_process,
+			inputs=None,
 			outputs=None
 		)
+		"""
 
 		if os.path.isfile('./config/generate.json'):
 			ui.load(import_generate_settings, inputs=None, outputs=input_settings)
@@ -592,7 +842,7 @@ def setup_gradio():
 		if args.check_for_updates:
 			ui.load(check_for_updates)
 
-		stop.click(fn=cancel_generate, inputs=None, outputs=None, cancels=[submit_event])
+		stop.click(fn=cancel_generate, inputs=None, outputs=None)
 
 
 	ui.queue(concurrency_count=args.concurrency_count)

start.bat

@@ -1,5 +1,4 @@
 call .\venv\Scripts\activate.bat
 set PATH=.\bin\;%PATH%
-python .\src\main.py
+python .\src\main.py %*
-deactivate
 pause

start.sh

@@ -1,4 +1,6 @@
 #!/bin/bash
+if [ ! -d "venv" ]; then ./setup-guided.sh; fi
+
 source ./venv/bin/activate
-python3 ./src/main.py
+python3 ./src/main.py "$@"
 deactivate

tortoise-tts

@@ -0,0 +1 @@
+Subproject commit 6fcd8c604f066e4e346da522bd14e6670395025f

train.bat

@@ -1,4 +1,4 @@
 call .\venv\Scripts\activate.bat
 python ./src/train.py -opt "%1"
-deactivate
+pause
-pause
+deactivate

train.ipynb

@@ -1,69 +0,0 @@
-{
-	"nbformat":4,
-	"nbformat_minor":0,
-	"metadata":{
-		"colab":{
-			"private_outputs":true,
-			"provenance":[
-				
-			]
-		},
-		"kernelspec":{
-			"name":"python3",
-			"display_name":"Python 3"
-		},
-		"language_info":{
-			"name":"python"
-		},
-		"accelerator":"GPU",
-		"gpuClass":"standard"
-	},
-	"cells":[
-		{
-			"cell_type":"code",
-			"execution_count":null,
-			"metadata":{
-				"id":"AaKpV3rCI3Eo"
-			},
-			"outputs":[
-				
-			],
-			"source":[
-				"!git clone https://git.ecker.tech/mrq/DL-Art-School\n",
-				"%cd DL-Art-School\n",
-				"!rm -r experiments\n",
-				"!pip install -r requirements.txt"
-			]
-		},
-		{
-			"cell_type":"code",
-			"source":[
-				"from google.colab import drive\n",
-				"drive.mount('/content/drive')",
-				"%cd /content/DL-Art-School/\n",
-				"!ln -s /content/drive/MyDrive/experiments/\n",
-			],
-			"metadata":{
-				"id":"8eV92cjGI4XL"
-			},
-			"execution_count":null,
-			"outputs":[
-				
-			]
-		},
-		{
-			"cell_type":"code",
-			"source":[
-				"%cd /content/DL-Art-School/\n",
-				"!python ./codes/train.py -opt ./experiments/ar.yml"
-			],
-			"metadata":{
-				"id":"7lcRGqglX2FC"
-			},
-			"execution_count":null,
-			"outputs":[
-				
-			]
-		}
-	]
-}

train.sh

@@ -1,4 +1,13 @@
 #!/bin/bash
 source ./venv/bin/activate
-python3 ./src/train.py -opt "$1"
+
+GPUS=$1
+CONFIG=$2
+PORT=1234
+
+if (( $GPUS > 1 )); then
+	torchrun --nproc_per_node=$GPUS --master_port=$PORT ./src/train.py -opt "$CONFIG" --launcher=pytorch
+else
+	python3 ./src/train.py -opt "$CONFIG"
+fi
 deactivate

update-force.bat

@@ -1,3 +1,15 @@
 git fetch --all
 git reset --hard origin/master
-call .\update.bat
+call .\update.bat
+
+python -m venv venv
+call .\venv\Scripts\activate.bat
+
+python -m pip install --upgrade pip
+python -m pip install -U -r .\requirements.txt
+python -m pip install -U -r .\tortoise-tts\requirements.txt
+python -m pip install -U -e .\tortoise-tts 
+python -m pip install -U -r .\dlas\requirements.txt
+
+pause
+deactivate

update-force.sh

@@ -1,4 +1,17 @@
 #!/bin/bash
 git fetch --all
 git reset --hard origin/master
-./update.sh
+
+./update.sh
+
+# force install requirements
+python3 -m venv venv
+source ./venv/bin/activate
+
+python3 -m pip install --upgrade pip
+python3 -m pip install -r ./requirements.txt
+python3 -m pip install -r ./tortoise-tts/requirements.txt
+python3 -m pip install -e ./tortoise-tts 
+python3 -m pip install -r ./dlas/requirements.txt
+
+deactivate

update.bat

@@ -1,13 +1,2 @@
 git pull
-python -m venv venv
+git submodule update --remote
-call .\venv\Scripts\activate.bat
-python -m pip install --upgrade pip
-python -m pip install -r ./requirements.txt
-python -m pip install -r ./dlas/requirements.txt
-
-cd dlas
-git pull
-cd ..
-
-deactivate
-pause

update.sh

@@ -1,13 +1,8 @@
 #!/bin/bash
 git pull
-python -m venv venv
+git submodule update --remote
-source ./venv/bin/activate
-python -m pip install --upgrade pip
-python -m pip install -r ./requirements.txt
-python -m pip install -r ./dlas/requirements.txt
 
-cd dlas
+if python -m pip show whispercpp &>/dev/null; then python -m pip install -U git+https://git.ecker.tech/lightmare/whispercpp.py; fi
-git pull
+if python -m pip show whisperx &>/dev/null; then python -m pip install -U git+https://github.com/m-bain/whisperx.git; fi
-cd ..
 
 deactivate