Upload finetuned ChunkFormer with resume checkpoint epoch=1 val_loss=192.815601

FINETUNE_RESUME_README.md

@@ -0,0 +1,17 @@
+# ChunkFormer fine-tuned checkpoint
+
+This repo contains:
+
+- `pytorch_model.bin`: model weights for inference / loading model.
+- `resume_checkpoint/training_checkpoint.pt`: full checkpoint for resuming training.
+- `resume_checkpoint/model_state_dict.pt`: model state dict only.
+- `resume_checkpoint/optimizer.pt`: optimizer state, if available.
+- `resume_checkpoint/scheduler.pt`: scheduler state, if available.
+- `resume_checkpoint/scaler.pt`: AMP GradScaler state, if available.
+- `best_checkpoints.json`: validation loss metadata.
+
+Best epoch: 1
+Best val_loss: 192.81560051995353
+Train loss: inf
+
+Base model: khanhld/chunkformer-ctc-large-vie

best_checkpoints.json

@@ -0,0 +1,8 @@
+[
+  {
+    "epoch": 1,
+    "val_loss": 192.81560051995353,
+    "train_loss": Infinity,
+    "path": "/kaggle/working/chunkformer-ctc-large-vie-finetune/epoch_1_val_loss_192.8156.pt"
+  }
+]

history.json

@@ -0,0 +1,9 @@
+[
+  {
+    "epoch": 1,
+    "train_loss": 326.93824258185526,
+    "val_loss": 192.81560051995353,
+    "lr": 1.0515468292711804e-07,
+    "encoder_trainable": false
+  }
+]

resume_checkpoint/best_checkpoints.json

@@ -0,0 +1,8 @@
+[
+  {
+    "epoch": 1,
+    "val_loss": 192.81560051995353,
+    "train_loss": Infinity,
+    "path": "/kaggle/working/chunkformer-ctc-large-vie-finetune/epoch_1_val_loss_192.8156.pt"
+  }
+]

resume_checkpoint/history.json

@@ -0,0 +1,9 @@
+[
+  {
+    "epoch": 1,
+    "train_loss": 326.93824258185526,
+    "val_loss": 192.81560051995353,
+    "lr": 1.0515468292711804e-07,
+    "encoder_trainable": false
+  }
+]

resume_checkpoint/model_state_dict.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c2280877d48f58d07ad9576af0be61499fdeb807678d9473b50cad58e55d4c43
+size 595483284

resume_checkpoint/optimizer.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:145fa8b792a413b1c3fb509b27ac429040314b6cd4f25e7bae1da77bba91fc75
+size 287970059

resume_checkpoint/scaler.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:294e0d71ff4bf6c9a330f45f3200b5eb1db565d454c054091a0d21922053a55e
+size 1383

resume_checkpoint/scheduler.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:03775b05773d0e09256e22a3ef3692425bdf8073bd7a3fb9e193db74e54a1260
+size 1401

resume_checkpoint/trainer_state.json

@@ -0,0 +1,41 @@
+{
+  "resume_type": "chunkformer_finetune_resume",
+  "base_repo": "khanhld/chunkformer-ctc-large-vie",
+  "repo_id": "efrainmain/chunkformer-ctc-vie-medical",
+  "best_checkpoint_path": "/kaggle/working/chunkformer-ctc-large-vie-finetune/epoch_1_val_loss_192.8156.pt",
+  "checkpoint_file_used": "/kaggle/working/chunkformer-ctc-large-vie-finetune/epoch_1_val_loss_192.8156.pt",
+  "epoch": 1,
+  "best_epoch": 1,
+  "best_val_loss": 192.81560051995353,
+  "val_loss": 192.81560051995353,
+  "train_loss": Infinity,
+  "has_optimizer_state": true,
+  "has_scheduler_state": true,
+  "has_scaler_state": true,
+  "config": {
+    "model_name": "khanhld/chunkformer-ctc-large-vie",
+    "dataset_root": "/kaggle/input/datasets/lqucng/vietmed-vimedcss-dataset",
+    "train_parquet": "/kaggle/input/datasets/lqucng/vietmed-vimedcss-dataset/train.parquet",
+    "validation_parquet": "/kaggle/input/datasets/lqucng/vietmed-vimedcss-dataset/validation.parquet",
+    "test_parquet": "/kaggle/input/datasets/lqucng/vietmed-vimedcss-dataset/test.parquet",
+    "preprocess_root": "/kaggle/working/chunkformer_preprocessed",
+    "output_dir": "/kaggle/working/chunkformer-ctc-large-vie-finetune",
+    "batch_size": 8,
+    "num_epochs": 1,
+    "learning_rate": 1e-05,
+    "weight_decay": 0.01,
+    "scheduler_type": "warmup_cosine",
+    "warmup_steps": 500,
+    "min_learning_rate": 1e-07,
+    "freeze_encoder_epochs": 2,
+    "patience": 3,
+    "keep_best": 2,
+    "max_grad_norm": 5.0,
+    "num_workers": 0,
+    "device": "cuda",
+    "use_amp": true,
+    "skip_bad_samples": true,
+    "max_train_batches": null,
+    "max_eval_batches": null
+  }
+}

resume_checkpoint/training_checkpoint.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:e2d252e5d611ddeffabfb89f1f618b2f651429e40209f85264c01d8eeb93be0d
+size 883464805

trainer_state.json

@@ -0,0 +1,41 @@
+{
+  "resume_type": "chunkformer_finetune_resume",
+  "base_repo": "khanhld/chunkformer-ctc-large-vie",
+  "repo_id": "efrainmain/chunkformer-ctc-vie-medical",
+  "best_checkpoint_path": "/kaggle/working/chunkformer-ctc-large-vie-finetune/epoch_1_val_loss_192.8156.pt",
+  "checkpoint_file_used": "/kaggle/working/chunkformer-ctc-large-vie-finetune/epoch_1_val_loss_192.8156.pt",
+  "epoch": 1,
+  "best_epoch": 1,
+  "best_val_loss": 192.81560051995353,
+  "val_loss": 192.81560051995353,
+  "train_loss": Infinity,
+  "has_optimizer_state": true,
+  "has_scheduler_state": true,
+  "has_scaler_state": true,
+  "config": {
+    "model_name": "khanhld/chunkformer-ctc-large-vie",
+    "dataset_root": "/kaggle/input/datasets/lqucng/vietmed-vimedcss-dataset",
+    "train_parquet": "/kaggle/input/datasets/lqucng/vietmed-vimedcss-dataset/train.parquet",
+    "validation_parquet": "/kaggle/input/datasets/lqucng/vietmed-vimedcss-dataset/validation.parquet",
+    "test_parquet": "/kaggle/input/datasets/lqucng/vietmed-vimedcss-dataset/test.parquet",
+    "preprocess_root": "/kaggle/working/chunkformer_preprocessed",
+    "output_dir": "/kaggle/working/chunkformer-ctc-large-vie-finetune",
+    "batch_size": 8,
+    "num_epochs": 1,
+    "learning_rate": 1e-05,
+    "weight_decay": 0.01,
+    "scheduler_type": "warmup_cosine",
+    "warmup_steps": 500,
+    "min_learning_rate": 1e-07,
+    "freeze_encoder_epochs": 2,
+    "patience": 3,
+    "keep_best": 2,
+    "max_grad_norm": 5.0,
+    "num_workers": 0,
+    "device": "cuda",
+    "use_amp": true,
+    "skip_bad_samples": true,
+    "max_train_batches": null,
+    "max_eval_batches": null
+  }
+}