You can not select more than 25 topics Topics must start with a letter or number, can include dashes ('-') and can be up to 35 characters long.
PaddleSpeech/examples/aishell/asr0/conf/deepspeech2.yaml

68 lines
1.8 KiB

Support paddle 2.x (#538) * 2.x model * model test pass * fix data * fix soundfile with flac support * one thread dataloader test pass * export feasture size add trainer and utils add setup model and dataloader update travis using Bionic dist * add venv; test under venv * fix unittest; train and valid * add train and config * add config and train script * fix ctc cuda memcopy error * fix imports * fix train valid log * fix dataset batch shuffle shift start from 1 fix rank_zero_only decreator error close tensorboard when train over add decoding config and code * test process can run * test with decoding * test and infer with decoding * fix infer * fix ctc loss lr schedule sortagrad logger * aishell egs * refactor train add aishell egs * fix dataset batch shuffle and add batch sampler log print model parameter * fix model and ctc * sequence_mask make all inputs zeros, which cause grad be zero, this is a bug of LessThanOp add grad clip by global norm add model train test notebook * ctc loss remove run prefix using ord value as text id * using unk when training compute_loss need text ids ord id using in test mode, which compute wer/cer * fix tester * add lr_deacy refactor code * fix tools * fix ci add tune fix gru model bugs add dataset and model test * fix decoding * refactor repo fix decoding * fix musan and rir dataset * refactor io, loss, conv, rnn, gradclip, model, utils * fix ci and import * refactor model add export jit model * add deploy bin and test it * rm uselss egs * add layer tools * refactor socket server new model from pretrain * remve useless * fix instability loss and grad nan or inf for librispeech training * fix sampler * fix libri train.sh * fix doc * add license on cpp * fix doc * fix libri script * fix install * clip 5 wer 7.39, clip 400 wer 7.54, 1.8 clip 400 baseline 7.49
4 years ago
# https://yaml.org/type/float.html
###########################################
# Data #
###########################################
train_manifest: data/manifest.train
dev_manifest: data/manifest.dev
test_manifest: data/manifest.test
min_input_len: 0.0
max_input_len: 27.0 # second
min_output_len: 0.0
max_output_len: .inf
min_output_input_ratio: 0.00
max_output_input_ratio: .inf
###########################################
# Dataloader #
###########################################
vocab_filepath: data/lang_char/vocab.txt
spm_model_prefix: ''
unit_type: 'char'
preprocess_config: conf/preprocess.yaml
feat_dim: 161
stride_ms: 10.0
window_ms: 25.0
sortagrad: 0 # Feed samples from shortest to longest ; -1: enabled for all epochs, 0: disabled, other: enabled for 'other' epochs
batch_size: 64
maxlen_in: 512 # if input length > maxlen-in, batchsize is automatically reduced
maxlen_out: 150 # if output length > maxlen-out, batchsize is automatically reduced
minibatches: 0 # for debug
batch_count: auto
batch_bins: 0
batch_frames_in: 0
batch_frames_out: 0
batch_frames_inout: 0
num_workers: 8
subsampling_factor: 1
num_encs: 1
E2E/Streaming Transformer/Conformer ASR (#578) * add cmvn and label smoothing loss layer * add layer for transformer * add glu and conformer conv * add torch compatiable hack, mask funcs * not hack size since it exists * add test; attention * add attention, common utils, hack paddle * add audio utils * conformer batch padding mask bug fix #223 * fix typo, python infer fix rnn mem opt name error and batchnorm1d, will be available at 2.0.2 * fix ci * fix ci * add encoder * refactor egs * add decoder * refactor ctc, add ctc align, refactor ckpt, add warmup lr scheduler, cmvn utils * refactor docs * add fix * fix readme * fix bugs, refactor collator, add pad_sequence, fix ckpt bugs * fix docstring * refactor data feed order * add u2 model * refactor cmvn, test * add utils * add u2 config * fix bugs * fix bugs * fix autograd maybe has problem when using inplace operation * refactor data, build vocab; add format data * fix text featurizer * refactor build vocab * add fbank, refactor feature of speech * refactor audio feat * refactor data preprare * refactor data * model init from config * add u2 bins * flake8 * can train * fix bugs, add coverage, add scripts * test can run * fix data * speed perturb with sox * add spec aug * fix for train * fix train logitc * fix logger * log valid loss, time dataset process * using np for speed perturb, remove some debug log of grad clip * fix logger * fix build vocab * fix logger name * using module logger as default * fix * fix install * reorder imports * fix board logger * fix logger * kaldi fbank and mfcc * fix cmvn and print prarams * fix add_eos_sos and cmvn * fix cmvn compute * fix logger and cmvn * fix subsampling, label smoothing loss, remove useless * add notebook test * fix log * fix tb logger * multi gpu valid * fix log * fix log * fix config * fix compute cmvn, need paddle 2.1 * add cmvn notebook * fix layer tools * fix compute cmvn * add rtf * fix decoding * fix layer tools * fix log, add avg script * more avg and test info * fix dataset pickle problem; using 2.1 paddle; num_workers can > 0; ckpt save in exp dir;fix setup.sh; * add vimrc * refactor tiny script, add transformer and stream conf * spm demo; librisppech scripts and confs * fix log * add librispeech scripts * refactor data pipe; fix conf; fix u2 default params * fix bugs * refactor aishell scripts * fix test * fix cmvn * fix s0 scripts * fix ds2 scripts and bugs * fix dev & test dataset filter * fix dataset filter * filter dev * fix ckpt path * filter test, since librispeech will cause OOM, but all test wer will be worse, since mismatch train with test * add comment * add syllable doc * fix ds2 configs * add doc * add pypinyin tools * fix decoder using blank_id=0 * mmseg with pybind11 * format code
4 years ago
############################################
# Network Architecture #
############################################
num_conv_layers: 2
num_rnn_layers: 5
rnn_layer_size: 1024
rnn_direction: bidirect # [forward, bidirect]
num_fc_layers: 0
fc_layers_size_list: -1,
use_gru: False
blank_id: 0
###########################################
# Training #
###########################################
n_epoch: 50
accum_grad: 1
lr: 5.0e-4
lr_decay: 0.93
weight_decay: 1.0e-6
global_grad_clip: 3.0
dist_sampler: False
log_interval: 1
checkpoint:
kbest_n: 50
latest_n: 5