Medix-AI/coqui-tts
Medix-AI
/
coqui-tts
mirror of https://github.com/coqui-ai/TTS.git
1
0
Fork 0
Code Issues Packages Projects Releases Wiki Activity

compute and config update with new attention entropy loss

This commit is contained in:
Eren Golge 2019-10-15 15:05:42 +02:00
parent c1f598b5d0
commit 77f5fd0584
3 changed files with 21 additions and 8 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

9
.compute
View File

@ -4,16 +4,13 @@ yes | apt-get install ffmpeg
yes | apt-get install espeak yes | apt-get install espeak
yes | apt-get install tmux yes | apt-get install tmux
yes | apt-get install zsh yes | apt-get install zsh
# pip3 install https://download.pytorch.org/whl/cu100/torch-1.1.0-cp37-cp37m-linux_x86_64.whl pip3 install https://download.pytorch.org/whl/cu100/torch-1.3.0%2Bcu100-cp36-cp36m-linux_x86_64.whl
# wget https://www.dropbox.com/s/m8waow6b3ydpf6h/MozillaDataset.tar.gz?dl=0 -O /data/rw/home/mozilla.tar
wget https://www.dropbox.com/s/wqn5v3wkktw9lmo/install.sh?dl=0 -O install.sh
sudo sh install.sh sudo sh install.sh
pip install pytorch==1.3.0+cu100
python3 setup.py develop python3 setup.py develop
# cp -R ${USER_DIR}/GermanData ../tmp/
# cp -R /data/ro/shared/data/keithito/LJSpeech-1.1/ ../tmp/
# python3 distribute.py --config_path config.json --data_path /data/ro/shared/data/keithito/LJSpeech-1.1/ # python3 distribute.py --config_path config.json --data_path /data/ro/shared/data/keithito/LJSpeech-1.1/
# cp -R ${USER_DIR}/Mozilla_22050 ../tmp/ # cp -R ${USER_DIR}/Mozilla_22050 ../tmp/
# python3 distribute.py --config_path config_tacotron_gst.json --data_path ../tmp/Mozilla_22050/ # python3 distribute.py --config_path config_tacotron_gst.json --data_path ../tmp/Mozilla_22050/
# python3 distribute.py --config_path config.json --data_path /data/rw/home/LibriTTS/train-clean-360 # python3 distribute.py --config_path config.json --data_path /data/rw/home/LibriTTS/train-clean-360
python3 distribute.py --config_path config.json # python3 distribute.py --config_path config.json
while true; do sleep 1000000; done while true; do sleep 1000000; done

4
config.json
View File

@ -1,6 +1,6 @@
{ {
"run_name": "ljspeech", "run_name": "ljspeech",
"run_description": "Tacotron2 ljspeech release training", "run_description": "Tacotron ljspeech release training",
"audio":{ "audio":{
// Audio processing parameters // Audio processing parameters
@ -31,7 +31,7 @@
"reinit_layers": [], "reinit_layers": [],
"model": "Tacotron2", // one of the model in models/ "model": "Tacotron", // one of the model in models/
"grad_clip": 1, // upper limit for gradients for clipping. "grad_clip": 1, // upper limit for gradients for clipping.
"epochs": 1000, // total number of epochs to train. "epochs": 1000, // total number of epochs to train.
"lr": 0.0001, // Initial learning rate. If Noam decay is active, maximum learning rate. "lr": 0.0001, // Initial learning rate. If Noam decay is active, maximum learning rate.

16
layers/losses.py
View File

@ -1,3 +1,5 @@
import numpy as np
import torch
from torch import nn from torch import nn
from torch.nn import functional from torch.nn import functional
from TTS.utils.generic_utils import sequence_mask from TTS.utils.generic_utils import sequence_mask
@ -53,3 +55,17 @@ class MSELossMasked(nn.Module):
x * mask, target * mask, reduction="sum") x * mask, target * mask, reduction="sum")
loss = loss / mask.sum() loss = loss / mask.sum()
return loss return loss
class AttentionEntropyLoss(nn.Module):
def forward(self, align):
"""
Forces attention to be more decisive by penalizing
soft attention weights
TODO: arguments
TODO: unit_test
"""
entropy = torch.distributions.Categorical(probs=align).entropy()
loss = (entropy / np.log(align.shape[1])).mean()
return loss