Create fast-conformer_aed.yaml

config/fast-conformer_aed.yaml

@@ -22,14 +22,25 @@ name: "FastConformer-Transformer-MultiTask"
 
 # Note: for larger models (1B+ params) initializing from a pretrained encoder
 #  may help (or even be required to) stabilize the training.
-init_from_nemo_model: null
+
+init_from_nemo_model: 
+
+  model0:
+    path: "/home/ubuntu/NeMo_Canary/canary_results/Higurashi_ASR/checkpoints/Higurashi_ASR.nemo"
+    exclude: ["transf_decoder._embedding.token_embedding", "log_softmax.mlp.layer0"]
+
+# init_from_pretrained_model:
+#   model0:
+#     name: "nvidia/canary-180m-flash"
+#     include: ["encoder"]
+
 
 # If using example training script, below will be used to instantiate spl_tokens tokenizer.
 # Similar can be done by calling CanaryTokenizer.build_special_tokenizer(tokens, output_dir).
 # If a tokenizer exists in dir, will skip building and use already built tokenizer.
 spl_tokens:
   model_dir: ???
-  tokens: ["translate", "transcribe", "ja"]
+  tokens: ["translate", "transcribe", 'ja']
   force_rebuild: False  # Set to True to build new tokenizer each time.
 
 model:
@@ -229,7 +240,7 @@ model:
     return_best_hypothesis: true  # Returns the most probably hypothesis after beam search
 
     beam:
-      beam_size: 1
+      beam_size: 4
       len_pen: 0.0
       max_generation_delta: 50
 
@@ -253,7 +264,7 @@ model:
     sched:
       name: InverseSquareRootAnnealing
       # scheduler config override
-      warmup_steps: 2500
+      warmup_steps: 5000
       warmup_ratio: null
       min_lr: 1e-6
 
@@ -279,27 +290,52 @@ trainer:
   logger: false  # Provided by exp_manager
   use_distributed_sampler: false  # Lhotse has its own distributed sampler
 
+# exp_manager:
+#   exp_dir: null
+#   name: ${name}
+#   create_tensorboard_logger: true
+#   create_checkpoint_callback: true
+#   checkpoint_callback_params:
+#     # in case of multiple validation sets, first one is used
+#     monitor: "val_loss"
+#     mode: "min"
+#     save_top_k: 5
+#     always_save_nemo: True # saves the checkpoints as nemo files instead of PTL checkpoints
+#   # checkpoint_callback_params:
+#   #   every_n_train_steps: 2000
+#   #   every_n_epochs: null # must be set to null to use every_n_train_steps
+#   #   monitor: "val_loss" # want all checkpoints, so step + mode: max always succeeds
+#   #   mode: "min"
+#   #   save_top_k: 5 # save all models
+#   #   save_last: True
+#   #   always_save_nemo: True
+
 exp_manager:
   exp_dir: null
   name: ${name}
   create_tensorboard_logger: true
+  
   create_checkpoint_callback: true
   checkpoint_callback_params:
-    # in case of multiple validation sets, first one is used
-    monitor: "val_loss"
+    every_n_train_steps: 4990
+    every_n_epochs: null # must be set to null to use every_n_train_steps
+    monitor: "step" # want all checkpoints, so step + mode: max always succeeds
     mode: "min"
-    save_top_k: 5
-    always_save_nemo: True # saves the checkpoints as nemo files instead of PTL checkpoints
+    save_top_k: 5 # save all models
+    save_last: True
+    always_save_nemo: True
+
+  # create_checkpoint_callback: true
   # checkpoint_callback_params:
-  #   every_n_train_steps: 2000
-  #   every_n_epochs: null # must be set to null to use every_n_train_steps
-  #   monitor: "step" # want all checkpoints, so step + mode: max always succeeds
+  #   # in case of multiple validation sets, first one is used
+  #   monitor: "val_loss"
   #   mode: "min"
-  #   save_top_k: 5 # save all models
-  #   save_last: True
-  #   always_save_nemo: True
+  #   save_top_k: 5
+  #   always_save_nemo: True # saves the checkpoints as nemo files instead of PTL checkpoints
+
+
 
-  resume_from_checkpoint: "/home/ubuntu/NeMo/canary_results/canary-small/checkpoints/canary-small--val_loss=0.1680-epoch=16.ckpt" # The path to a checkpoint file to continue the training, restores the whole state including the epoch, step, LR schedulers, apex, etc.
+  resume_from_checkpoint: /home/ubuntu/NeMo_Canary/canary_results/Higurashi_ASR_v.02/checkpoints/Higurashi_ASR_v.02--step=29940.0000-epoch=1-last.ckpt # The path to a checkpoint file to continue the training, restores the whole state including the epoch, step, LR schedulers, apex, etc.
   # you need to set these two to True to continue the training
   resume_if_exists: true
   resume_ignore_no_checkpoint: true