End of training

all_results.json

@@ -1,17 +1,32 @@
 {
     "epoch": 1.0,
     "eval_loss": 1.6191972494125366,
+    "eval_pretrained_loss": 1.6191972494125366,
+    "eval_pretrained_runtime": 59.7459,
+    "eval_pretrained_samples_per_second": 0.067,
+    "eval_pretrained_steps_per_second": 0.033,
+    "eval_pretrained_wer": 153.2258064516129,
     "eval_runtime": 56.3363,
     "eval_samples_per_second": 0.071,
     "eval_steps_per_second": 0.036,
     "eval_wer": 153.2258064516129,
+    "test_finetuned_loss": 1.7568330764770508,
+    "test_finetuned_runtime": 39.6579,
+    "test_finetuned_samples_per_second": 0.101,
+    "test_finetuned_steps_per_second": 0.05,
+    "test_finetuned_wer": 138.5964912280702,
     "test_loss": 1.7568330764770508,
+    "test_pretrained_loss": 1.7568330764770508,
+    "test_pretrained_runtime": 42.5376,
+    "test_pretrained_samples_per_second": 0.094,
+    "test_pretrained_steps_per_second": 0.047,
+    "test_pretrained_wer": 138.5964912280702,
     "test_runtime": 37.8582,
     "test_samples_per_second": 0.106,
     "test_steps_per_second": 0.053,
     "test_wer": 138.5964912280702,
     "train_loss": 1.4339025020599365,
-    "train_runtime": 108.1566,
-    "train_samples_per_second": 0.074,
-    "train_steps_per_second": 0.018
+    "train_runtime": 102.2429,
+    "train_samples_per_second": 0.078,
+    "train_steps_per_second": 0.02
 }

eval_pretrained_results.json

@@ -0,0 +1,7 @@
+{
+    "eval_pretrained_loss": 1.6191972494125366,
+    "eval_pretrained_runtime": 59.7459,
+    "eval_pretrained_samples_per_second": 0.067,
+    "eval_pretrained_steps_per_second": 0.033,
+    "eval_pretrained_wer": 153.2258064516129
+}

huggingface_training.py

@@ -322,55 +322,55 @@ I hope you haven't left yet. If you have, bad for you, as we are ready for train
 As Whisper is a pretrained model ready to be used off-the-shelf, it is advisable to evaluate it before training on both the validation and test sets. Let's make sure we make no harm to it.
 """
 
-# eval_metrics = trainer.evaluate(
-#     eval_dataset=preprocessed_dataset["validation"],
-#     metric_key_prefix="eval",
-#     max_length=448,
-#     num_beams=1,
-#     # gen_kwargs={"key": value}  to provide additional generation specific arguments by keyword
-# )
+eval_metrics = trainer.evaluate(
+    eval_dataset=preprocessed_dataset["validation"],
+    metric_key_prefix="eval_pretrained",
+    max_length=448,
+    num_beams=1,
+    # gen_kwargs={"key": value}  to provide additional generation specific arguments by keyword
+)
 
-# trainer.log_metrics("eval", eval_metrics)
-# trainer.save_metrics("eval", eval_metrics)
-# print(eval_metrics)
+trainer.log_metrics("eval_pretrained", eval_metrics)
+trainer.save_metrics("eval_pretrained", eval_metrics)
+print(eval_metrics)
 
-# test_metrics = trainer.evaluate(
-#     eval_dataset=preprocessed_dataset["test"],
-#     metric_key_prefix="test",
-#     max_length=448,
-#     num_beams=1,
-#     # gen_kwargs={"key": value}  to provide additional generation specific arguments by keyword
-# )
+test_metrics = trainer.evaluate(
+    eval_dataset=preprocessed_dataset["test"],
+    metric_key_prefix="test_pretrained",
+    max_length=448,
+    num_beams=1,
+    # gen_kwargs={"key": value}  to provide additional generation specific arguments by keyword
+)
 
-# trainer.log_metrics("test", test_metrics)
-# trainer.save_metrics("test", test_metrics)
-# print(test_metrics)
+trainer.log_metrics("test_pretrained", test_metrics)
+trainer.save_metrics("test_pretrained", test_metrics)
+print(test_metrics)
 
-# train_result = trainer.train()
-# trainer.save_model()
+train_result = trainer.train()
+trainer.save_model()
 
-# metrics = train_result.metrics
-# trainer.log_metrics("train", metrics)
-# trainer.save_metrics("train", metrics)
-# trainer.save_state()
-# print(metrics)
+metrics = train_result.metrics
+trainer.log_metrics("train", metrics)
+trainer.save_metrics("train", metrics)
+trainer.save_state()
+print(metrics)
 
 # """ADD SOMETHING ABOUT THE TRAINING.
 
 # Now let's evaluate the 
 # """
 
-# final_metrics = trainer.evaluate(
-#     eval_dataset=preprocessed_dataset["test"],
-#     metric_key_prefix="test",
-#     max_length=448,
-#     num_beams=1,
-#     # gen_kwargs={"key": value}  to provide additional generation specific arguments by keyword
-# )
+final_metrics = trainer.evaluate(
+    eval_dataset=preprocessed_dataset["test"],
+    metric_key_prefix="test_finetuned",
+    max_length=448,
+    num_beams=1,
+    # gen_kwargs={"key": value}  to provide additional generation specific arguments by keyword
+)
 
-# trainer.log_metrics("test", final_metrics)
-# trainer.save_metrics("test", final_metrics)
-# print(final_metrics)
+trainer.log_metrics("test_finetuned", final_metrics)
+trainer.save_metrics("test_finetuned", final_metrics)
+print(final_metrics)
 
 # Pushing to hub during training slows down training
 # so we push it only in the end.

test_finetuned_results.json

@@ -0,0 +1,8 @@
+{
+    "epoch": 1.0,
+    "test_finetuned_loss": 1.7568330764770508,
+    "test_finetuned_runtime": 39.6579,
+    "test_finetuned_samples_per_second": 0.101,
+    "test_finetuned_steps_per_second": 0.05,
+    "test_finetuned_wer": 138.5964912280702
+}

test_pretrained_results.json

@@ -0,0 +1,7 @@
+{
+    "test_pretrained_loss": 1.7568330764770508,
+    "test_pretrained_runtime": 42.5376,
+    "test_pretrained_samples_per_second": 0.094,
+    "test_pretrained_steps_per_second": 0.047,
+    "test_pretrained_wer": 138.5964912280702
+}

train_results.json

@@ -1,7 +1,7 @@
 {
     "epoch": 1.0,
     "train_loss": 1.4339025020599365,
-    "train_runtime": 108.1566,
-    "train_samples_per_second": 0.074,
-    "train_steps_per_second": 0.018
+    "train_runtime": 102.2429,
+    "train_samples_per_second": 0.078,
+    "train_steps_per_second": 0.02
 }

trainer_state.json

@@ -16,9 +16,9 @@
     {
       "epoch": 0.5,
       "eval_loss": 1.6191972494125366,
-      "eval_runtime": 41.7522,
-      "eval_samples_per_second": 0.096,
-      "eval_steps_per_second": 0.048,
+      "eval_runtime": 38.6993,
+      "eval_samples_per_second": 0.103,
+      "eval_steps_per_second": 0.052,
       "eval_wer": 153.2258064516129,
       "step": 1
     },
@@ -31,9 +31,9 @@
     {
       "epoch": 1.0,
       "eval_loss": 1.6191972494125366,
-      "eval_runtime": 37.2469,
-      "eval_samples_per_second": 0.107,
-      "eval_steps_per_second": 0.054,
+      "eval_runtime": 39.2638,
+      "eval_samples_per_second": 0.102,
+      "eval_steps_per_second": 0.051,
       "eval_wer": 153.2258064516129,
       "step": 2
     },
@@ -42,9 +42,9 @@
       "step": 2,
       "total_flos": 196951080960000.0,
       "train_loss": 1.4339025020599365,
-      "train_runtime": 108.1566,
-      "train_samples_per_second": 0.074,
-      "train_steps_per_second": 0.018
+      "train_runtime": 102.2429,
+      "train_samples_per_second": 0.078,
+      "train_steps_per_second": 0.02
     }
   ],
   "max_steps": 2,

training_args.bin

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:081bb11364a81f07b31bca0107ccb5d62d955d95e115f206089558cf85595e34
+oid sha256:c9cf1b9241b15d95f3310024d7bec9fb6d139a94c9d760c0be51787d41a93fc3
 size 3579