Add RNNLM recipe

recipes/KsponSpeech/LM/README.md

@@ -12,13 +12,15 @@ Also, set data_folder in the yaml file to the result of ksponspeech_prepare.py.
 Run the following to start training the language model.
 
 ```bash
-python train.py hparams/transformer.yaml
+python train.py hparams/transformer.yaml # transformerLM
+python train.py hparams/RNNLM.yaml # RNNLM
 ```
 # Results
 
 | Release | hyperparams file | eval clean loss | eval other loss | Model link | GPUs |Training time|
 |:----:|:----:|:----:|:----:|:----:|:----:|:----:|
 |01-23-23|transformer.yaml|4.40|4.67|[Dropbox](https://www.dropbox.com/sh/egv5bdn8b5i45eo/AAB7a8gFt2FqbnO4yhL6DQ8na?dl=0)|1xA100 80GB|17 hours 2 mins|
+|04-16-24|RNNLM.yaml|4.59|4.94|[Dropbox](https://www.dropbox.com/sh/egv5bdn8b5i45eo/AAB7a8gFt2FqbnO4yhL6DQ8na?dl=0)|1xA100 80GB|50 mins|
 
 # About SpeechBrain
 - Website: https://speechbrain.github.io/

recipes/KsponSpeech/LM/hparams/RNNLM.yaml

@@ -0,0 +1,117 @@
+# ############################################################################
+# Model: RNNLM of E2E ASR
+# Tokens: unigram
+# losses: NLL
+# Training: KsponSpeech 965.2h transcript
+# Authors: Ju-Chieh Chou 2020, Jianyuan Zhong 2021, Dong Won Kim 2024
+# ############################################################################
+
+# Seed needs to be set at top of yaml, before objects with parameters are made
+seed: 2226
+__set_seed: !apply:torch.manual_seed [!ref <seed>]
+output_folder: !ref results/RNN/<seed>
+save_folder: !ref <output_folder>/save
+train_log: !ref <output_folder>/train_log.txt
+
+# Data files
+# The data_folder is needed because we train the LM on the training
+# Data files
+data_folder: !PLACEHOLDER # e.g., /path/to/KsponSpeech
+train_csv: !ref <data_folder>/train.csv
+valid_csv: !ref <data_folder>/dev.csv
+test_csv:
+ - !ref <data_folder>/eval_clean.csv
+ - !ref <data_folder>/eval_other.csv
+
+# Tokenizer model
+tokenizer_file: ddwkim/asr-conformer-transducer-rnnlm-ksponspeech/tokenizer.ckpt
+
+####################### Training Parameters ####################################
+number_of_epochs: 20
+batch_size: 128
+lr: 0.001
+weight_decay: 0.01
+num_workers: 10
+grad_accumulation_factor: 1 # Gradient accumulation to simulate large batch training
+ckpt_interval_minutes: 15 # save checkpoint every N min
+
+# Dataloader options
+train_dataloader_opts:
+ batch_size: !ref <batch_size>
+ shuffle: True
+ num_workers: !ref <num_workers>
+
+valid_dataloader_opts:
+ batch_size: 1
+
+test_dataloader_opts:
+ batch_size: 1
+
+####################### Model Parameters #######################################
+emb_size: 256
+activation: !name:torch.nn.LeakyReLU
+dropout: 0.0
+rnn_layers: 6
+rnn_neurons: 512
+dnn_blocks: 1
+dnn_neurons: 256
+
+# Outputs
+output_neurons: 5000
+# blank_index: 0
+bos_index: 1
+eos_index: 2
+# pad_index: 0
+
+
+# Functions
+model: !new:speechbrain.lobes.models.RNNLM.RNNLM
+ output_neurons: !ref <output_neurons>
+ embedding_dim: !ref <emb_size>
+ activation: !ref <activation>
+ dropout: !ref <dropout>
+ rnn_layers: !ref <rnn_layers>
+ rnn_neurons: !ref <rnn_neurons>
+ dnn_blocks: !ref <dnn_blocks>
+ dnn_neurons: !ref <dnn_neurons>
+
+modules:
+ model: !ref <model>
+
+checkpointer: !new:speechbrain.utils.checkpoints.Checkpointer
+ checkpoints_dir: !ref <save_folder>
+ recoverables:
+ model: !ref <model>
+ scheduler: !ref <lr_annealing>
+ counter: !ref <epoch_counter>
+
+log_softmax: !new:speechbrain.nnet.activations.Softmax
+ apply_log: True
+
+optimizer: !name:torch.optim.AdamW
+ lr: !ref <lr>
+ betas: (0.9, 0.998)
+ eps: 0.000000001
+ weight_decay: !ref <weight_decay>
+
+lr_annealing: !new:speechbrain.nnet.schedulers.NoamScheduler
+ lr_initial: !ref <lr>
+ n_warmup_steps: 2000
+ model_size: 1
+
+epoch_counter: !new:speechbrain.utils.epoch_loop.EpochCounter
+ limit: !ref <number_of_epochs>
+
+compute_cost: !name:speechbrain.nnet.losses.nll_loss
+
+train_logger: !new:speechbrain.utils.train_logger.FileTrainLogger
+ save_file: !ref <train_log>
+
+tokenizer: !new:sentencepiece.SentencePieceProcessor
+
+pretrainer: !new:speechbrain.utils.parameter_transfer.Pretrainer
+ collect_in: !ref <save_folder>
+ loadables:
+ tokenizer: !ref <tokenizer>
+ paths:
+ tokenizer: !ref <tokenizer_file>

tests/recipes/KsponSpeech.csv

@@ -1,4 +1,7 @@
 Task,Dataset,Script_file,Hparam_file,Data_prep_file,Readme_file,Result_url,HF_repo,test_debug_flags,test_debug_checks,performance
 ASR,KsponSpeech,recipes/KsponSpeech/ASR/transformer/train.py,recipes/KsponSpeech/ASR/transformer/hparams/conformer_medium.yaml,recipes/KsponSpeech/ASR/transformer/ksponspeech_prepare.py,recipes/KsponSpeech/ASR/transformer/README.md,https://www.dropbox.com/sh/uibokbz83o8ybv3/AACtO5U7mUbu_XhtcoOphAjza?dl=0,https://huggingface.co/speechbrain/asr-conformer-transformerlm-ksponspeech,--data_folder=tests/samples/ASR/ --train_csv=tests/samples/annotation/ASR_train.csv --valid_csv=tests/samples/annotation/ASR_train.csv --test_csv=[tests/samples/annotation/ASR_train.csv] --number_of_epochs=2 --skip_prep=True --num_encoder_layers=3 --num_decoder_layers=3,"file_exists=[wer_ASR_train.txt,train_log.txt,log.txt,env.log,train.py,hyperparams.yaml,save/tokenizer.ckpt,save/lm.ckpt]",Test-clean-WER=20.78% Test-others-WER=25.73%
+ASR,KsponSpeech,recipes/KsponSpeech/ASR/transformer/train.py,recipes/KsponSpeech/ASR/transformer/hparams/conformer_small.yaml,recipes/KsponSpeech/ASR/transformer/ksponspeech_prepare.py,recipes/KsponSpeech/ASR/transformer/README.md,https://www.dropbox.com/sh/uibokbz83o8ybv3/AACtO5U7mUbu_XhtcoOphAjza?dl=0,https://huggingface.co/dddwkim/asr-conformer-small-transformerlm-ksponspeech,--data_folder=tests/samples/ASR/ --train_csv=tests/samples/annotation/ASR_train.csv --valid_csv=tests/samples/annotation/ASR_train.csv --test_csv=[tests/samples/annotation/ASR_train.csv] --number_of_epochs=2 --skip_prep=True --num_encoder_layers=3 --num_decoder_layers=3,"file_exists=[wer_ASR_train.txt,train_log.txt,log.txt,env.log,train.py,hyperparams.yaml,save/tokenizer.ckpt,save/lm.ckpt]",Test-clean-WER=21.78% Test-others-WER=26.48%
+ASR,KsponSpeech,recipes/KsponSpeech/ASR/transformer/train.py,recipes/KsponSpeech/ASR/transformer/hparams/brachformer_medium.yaml,recipes/KsponSpeech/ASR/transformer/ksponspeech_prepare.py,recipes/KsponSpeech/ASR/transformer/README.md,https://www.dropbox.com/sh/uibokbz83o8ybv3/AACtO5U7mUbu_XhtcoOphAjza?dl=0,https://huggingface.co/dddwkim/asr-branchformer-transformerlm-ksponspeech,--data_folder=tests/samples/ASR/ --train_csv=tests/samples/annotation/ASR_train.csv --valid_csv=tests/samples/annotation/ASR_train.csv --test_csv=[tests/samples/annotation/ASR_train.csv] --number_of_epochs=2 --skip_prep=True --num_encoder_layers=3 --num_decoder_layers=3,"file_exists=[wer_ASR_train.txt,train_log.txt,log.txt,env.log,train.py,hyperparams.yaml,save/tokenizer.ckpt,save/lm.ckpt]",Test-clean-WER=21.01% Test-others-WER=25.68%
 LM,KsponSpeech,recipes/KsponSpeech/LM/train.py,recipes/KsponSpeech/LM/hparams/transformer.yaml,recipes/KsponSpeech/LM/ksponspeech_prepare.py,recipes/KsponSpeech/LM/README.md,https://www.dropbox.com/sh/egv5bdn8b5i45eo/AAB7a8gFt2FqbnO4yhL6DQ8na?dl=0,,--data_folder=tests/samples/ASR/ --train_csv=tests/samples/annotation/ASR_train.csv --valid_csv=tests/samples/annotation/ASR_train.csv --test_csv=[tests/samples/annotation/ASR_train.csv] --number_of_epochs=2 --d_model=120 --d_ffn=96,"file_exists=[train_log.txt,log.txt,env.log,train.py,hyperparams.yaml,save/tokenizer.ckpt]",
+LM,KsponSpeech,recipes/KsponSpeech/LM/train.py,recipes/KsponSpeech/LM/hparams/RNNLM.yaml,recipes/KsponSpeech/LM/ksponspeech_prepare.py,recipes/KsponSpeech/LM/README.md,https://www.dropbox.com/sh/egv5bdn8b5i45eo/AAB7a8gFt2FqbnO4yhL6DQ8na?dl=0,,--data_folder=tests/samples/ASR/ --train_csv=tests/samples/annotation/ASR_train.csv --valid_csv=tests/samples/annotation/ASR_train.csv --test_csv=[tests/samples/annotation/ASR_train.csv] --number_of_epochs=2 --emb_size=48 --rnn_neurons=96 --dnn_neurons=48,"file_exists=[train_log.txt,log.txt,env.log,train.py,hyperparams.yaml,save/tokenizer.ckpt]",
 Tokenizer,KsponSpeech,recipes/KsponSpeech/Tokenizer/train.py,recipes/KsponSpeech/Tokenizer/hparams/5K_unigram_subword_bpe.yaml,recipes/KsponSpeech/Tokenizer/ksponspeech_prepare.py,recipes/KsponSpeech/Tokenizer/README.md,https://www.dropbox.com/sh/prnqt09e7xpc1kr/AAB-HkfUazPifn7kXnKnAJSga?dl=0,,--data_folder=tests/samples/ASR/ --train_csv=tests/samples/annotation/ASR_train.csv --valid_csv=tests/samples/annotation/ASR_train.csv --skip_prep=True --token_output=23,"file_exists=[23_unigram.model,23_unigram.vocab,log.txt,ASR_train.txt,env.log,train.py,hyperparams.yaml]",