Hugging Face's logo

BUT-FIT
/
DeCRED-base

Automatic Speech Recognition
Transformers
Safetensors
English
joint_aed_ctc_speech-encoder-decoder
custom_code
Eval Results
Model card Files Community
DeCRED-base
File size: 3,195 Bytes 
5b20446
673be4f
5b20446
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
 
#!/usr/bin/bash
#SBATCH --job-name DeCRED
#SBATCH --account OPEN-28-57
#SBATCH --partition qgpu
#SBATCH --nodes=6
#SBATCH --ntasks=6
#SBATCH --ntasks-per-node=1
#SBATCH --gpus-per-node 8
#SBATCH --cpus-per-task=128
#SBATCH --time 2-00:00:00
#SBATCH --output=/mnt/proj1/open-28-58/lakoc/huggingface_asr/outputs/ebranchformer_english_medium_normalized_regularized.out

EXPERIMENT="ebranchformer_english_medium_normalized_regularized"
PROJECT="regularizations_english_corpus"
WORK_DIR="/mnt/proj1/open-28-58/lakoc/huggingface_asr"
RECIPE_DIR="${WORK_DIR}/recipes/ebranchformer_english"
EXPERIMENT_PATH="${WORK_DIR}/experiments/${EXPERIMENT}"
HF_HOME="/scratch/project/open-28-57/lakoc/huggingface_cache"

args=(
  # General training arguments
  --output_dir=$EXPERIMENT_PATH
  --per_device_train_batch_size="64"
  --per_device_eval_batch_size="8"
  --dataloader_num_workers="24"
  --num_train_epochs="100"
  --group_by_length="True"
  --bf16
  --do_train
  --do_evaluate
  --joint_decoding_during_training
  --load_best_model_at_end
  --metric_for_best_model="eval_wer"

  # Optimizer related arguments
  --optim="adamw_torch"
  --learning_rate="1e-3"
  --warmup_steps="40000"
  --early_stopping_patience="10"
  --weight_decay="1e-6"
  --max_grad_norm="0.5"
  --lsm_factor="0.1"
  --mask_unks
  --gradient_accumulation_steps="1"

  # Logging, saving and evaluation related arguments
  --report_to="wandb"
  --logging_steps="10"
  --save_strategy="epoch"
  --evaluation_strategy="epoch"
  --wandb_predictions_to_save=500
  --greater_is_better="False"
  --save_total_limit="5"

  # Data related arguments
  --max_duration_in_seconds="20.0"
  --min_duration_in_seconds="0.2"
  --length_column_name="input_len"
  --remove_unused_columns="False"
  --preprocessing_num_workers="32"
  --dataset_name="/scratch/project/open-28-57/lakoc/processed_dataset_full"
  --writer_batch_size="500"
  --test_splits wsj_test fisher_swbd_dev voxpopuli_test tedlium3_test librispeech_test.clean librispeech_test.other commonvoice_en_test fleurs_test
  --validation_slice="20%"
  --validation_slice_seed=42

  # Preprocessing related arguments
  --data_preprocessing_config="${RECIPE_DIR}/data_preprocessing.json"

  # Model related arguments
  --from_encoder_decoder_config
  --tokenizer_name="Lakoc/english_corpus_uni5000_normalized"
  --feature_extractor_name="Lakoc/log_80mel_extractor_16k"
  --base_encoder_model="Lakoc/ebranchformer_16l_512h"
  --base_decoder_model="Lakoc/gpt2_512h_8l_add_head6_04"
  --ctc_weight="0.3"
  --decoder_pos_emb_fixed
  --expect_2d_input

  # Generation related arguments
  --num_beams="1"
  --max_length="512"
  --predict_with_generate
  --decoding_ctc_weight="0"
)

export PARENT=`/bin/hostname -s`
export MPORT=13000
export CHILDREN=`scontrol show hostnames $SLURM_JOB_NODELIST | grep -v $PARENT`
export HOSTLIST="$PARENT $CHILDREN"
export WORLD_SIZE=$SLURM_NTASKS

conda deactivate
source activate loco_asr

mkdir -p $EXPERIMENT_PATH

srun --cpus-per-task $SLURM_CPUS_ON_NODE --gpus-per-task $SLURM_GPUS_ON_NODE  \
/mnt/proj1/open-28-58/lakoc/huggingface_asr/recipes/multinode_training/start_single_node_job.sh \
"${EXPERIMENT}" $PROJECT $WORK_DIR $RECIPE_DIR $HF_HOME "${args[@]}"