nemo_experiments/Speech_To_Text_Finetuning/git-info.log

commit hash: 81461d2fe9b52e0a6fe293e9f0b172e1564aebee
diff --git a/runTrain.sh b/runTrain.sh
index 98322c8..92cce48 100755
--- a/runTrain.sh
+++ b/runTrain.sh
@@ -6,10 +6,8 @@ export TORCH_FORCE_NO_WEIGHTS_ONLY_LOAD=1
 NUMBA_CUDA_USE_NVIDIA_BINDING=1 HYDRA_FULL_ERROR=1 python ./speech_to_text_finetune.py \
   --config-path="." \
   --config-name="speech_to_text_finetune" \
-  trainer.accumulate_grad_batches=16 \
-  trainer.devices=1 \
   trainer.accelerator=gpu \
   +trainer.limit_train_batches=null \
   exp_manager.create_wandb_logger=True \
-  exp_manager.wandb_logger_kwargs.name=parakeet_v3_finetune_fixed \
-  exp_manager.wandb_logger_kwargs.project=parakeet-hebrew-asr 
\ No newline at end of file
+  exp_manager.wandb_logger_kwargs.name=parakeet_v3_finetune_Phase2 \
+  exp_manager.wandb_logger_kwargs.project=parakeet-hebrew-asr 
diff --git a/speech_to_text_finetune.yaml b/speech_to_text_finetune.yaml
index f71ec04..8f102ae 100644
--- a/speech_to_text_finetune.yaml
+++ b/speech_to_text_finetune.yaml
@@ -1,7 +1,6 @@
 name: "Speech_To_Text_Finetuning"
 
-init_from_pretrained_model: nvidia/parakeet-tdt-0.6b-v3
-
+init_from_nemo_model: ./Parakeet-RNNT-XXL-1.1b_merged_universal_spe8.5k_1.0.nemo
 model:
   sample_rate: 16000
   activation_checkpointing: true
@@ -9,29 +8,29 @@ model:
     shar_path: hf://buckets/Tyl3rDrden/IvritKnessetRecordingsData
     use_lhotse: true
     is_tarred: true
-    batch_size: 1
+    batch_size: 2
     allow_missing_data: true
     shuffle: false
     num_workers: 8
     pin_memory: true
 
   validation_ds:
-    manifest_path: null
+    manifest_filepath: null
     shar_path: hf://buckets/Tyl3rDrden/parkaeetValidation
     use_lhotse: true
     is_tarred: true
-    batch_size: 1
+    batch_size: 5
     allow_missing_data: true
     shuffle: false
     num_workers: 8
     pin_memory: true
 
   test_ds:
-    manifest_path: null
+    manifest_filepath: null
     shar_path: hf://buckets/Tyl3rDrden/parkaeetTest
     use_lhotse: true
     is_tarred: true
-    batch_size: 1
+    batch_size: 9
     allow_missing_data: true
     shuffle: false
     num_workers: 8
@@ -42,9 +41,9 @@ model:
     labels: null
 
   tokenizer:
-    update_tokenizer: true
-    dir: ./tokenizer_spe_bpe_v32768_pad_bos_eos
-    type: bpe
+    update_tokenizer: false
+    #dir: ./tokenizer_spe_bpe_v32768_pad_bos_eos
+    #type: bpe
 
   spec_augment:
     _target_: nemo.collections.asr.modules.SpectrogramAugmentation
@@ -69,18 +68,19 @@ trainer:
   devices: -1
   num_nodes: 1
   max_epochs: -1
-  max_steps: 1000000
+  max_steps: 50000
   val_check_interval: 2000
   accelerator: auto
   strategy:
     _target_: lightning.pytorch.strategies.DDPStrategy
     gradient_as_bucket_view: true
-  accumulate_grad_batches: 16
+  accumulate_grad_batches: 32
   gradient_clip_val: 0.0
-  precision: 16-mixed
-  log_every_n_steps: 500
+  precision: 32
+  log_every_n_steps: 30
   enable_progress_bar: True
   num_sanity_val_steps: 0
+  limit_val_batches: 100
   check_val_every_n_epoch: 1
   use_distributed_sampler: false
   sync_batchnorm: true
@@ -96,7 +96,7 @@ exp_manager:
   checkpoint_callback_params:
     monitor: "val_wer"       # The metric to track
     mode: "min"              # We want to MINIMIZE the error rate
-    save_top_k: 5            # Keep the 5 best versions
+    save_top_k: 2            # Keep the 5 best versions
     every_n_epochs: 1
     always_save_nemo: True
     save_best_model: True    # Ensures the best overall is saved as a .nemo file
@@ -106,4 +106,4 @@ exp_manager:
   wandb_logger_kwargs:
     name: runV2.0
     log_model: True
-    project: ParkaeetRealExperiment
\ No newline at end of file
+    project: ParkaeetRealExperiment