test model card

run_speech_recognition_seq2seq_streaming.py

@@ -834,13 +834,13 @@ def main():
         "tags": "whisper-event",
     }
     if data_args.dataset_train_name is not None:
-        dataset_name = list(set(data_args.dataset_train_name.split(",")))
-        kwargs["dataset_tags"] = dataset_name
+        dataset_names = list(data_args.dataset_train_name.split(","))
+        kwargs["dataset_tags"] = dataset_names
         if data_args.dataset_train_config_name is not None:
-            dataset_config_name = list(set(data_args.dataset_train_config_name.split(",")))
-            kwargs["dataset"] = f"{dataset_name} {dataset_config_name}"
+            dataset_config_names = list(data_args.dataset_train_config_name.split(","))
+            kwargs["dataset"] = [f"{ds_name} {ds_cfg_name}" for ds_name, ds_cfg_name in zip(dataset_names, dataset_config_names)]
         else:
-            kwargs["dataset"] = dataset_name
+            kwargs["dataset"] = dataset_names
         # if "common_voice" in data_args.dataset_name:
         #     kwargs["language"] = data_args.dataset_config_name[:2]
         if data_args.language_train is not None:
@@ -849,24 +849,35 @@ def main():
         if model_args.model_index_name is not None:
             kwargs["model_name"] = model_args.model_index_name
 
+    logger.info("*** Training stats written ***")
+    logger.info(json.dumps(kwargs, indent=4))
+
     # Training complete notification
+    logger.info("*** Training and eval complete ***")
     logger.info(SENDING_NOTIFICATION)
-    notify_me(recipient=RECIPIENT_ADDRESS, message=json.dumps(kwargs, indent=4))
-    logger.info("*** Training complete!!! ***")
-
+    with open(os.path.join(training_args.output_dir, "train_results.json"), "r") as f:
+            train_results = json.load(f)
+    with open(os.path.join(training_args.output_dir, "eval_results.json"), "r") as f:
+            eval_results = json.load(f)
+    notify_me(recipient=RECIPIENT_ADDRESS,
+              message=f"Training complete! {train_results = } {eval_results = }")
+    
 
     if training_args.push_to_hub:
         logger.info("*** Pushing to hub ***")
         trainer.push_to_hub(**kwargs)
         logger.info("*** Pushed to hub ***")
         logger.info(SENDING_NOTIFICATION)
-        notify_me(recipient=RECIPIENT_ADDRESS, message="Model pushed to hub")
     else:
         logger.info("*** Creating model card ***")
         trainer.create_model_card(**kwargs)
         logger.info("*** Model card created ***")
         logger.info(SENDING_NOTIFICATION)
-        notify_me(recipient=RECIPIENT_ADDRESS, message="Model card created")
+
+    with open(os.path.join(training_args.output_dir, "README.md"), "r") as f:
+        readme = f.read()
+    notify_me(recipient=RECIPIENT_ADDRESS,
+              message=f"Model pushed to hub! {readme = }")
 
     return results
 

test_run_nordic.sh

@@ -11,10 +11,10 @@ python $1run_speech_recognition_seq2seq_streaming.py \
 	--model_index_name="Whisper Tiny Nordic" \
 	--max_train_samples="64" \
 	--max_eval_samples="32" \
-	--max_steps="8" \
+	--max_steps="1" \
 	--output_dir="./" \
-	--per_device_train_batch_size="8" \
-	--per_device_eval_batch_size="4" \
+	--per_device_train_batch_size="1" \
+	--per_device_eval_batch_size="1" \
 	--logging_steps="25" \
 	--learning_rate="1e-5" \
 	--warmup_steps="500" \
@@ -35,8 +35,8 @@ python $1run_speech_recognition_seq2seq_streaming.py \
 	--overwrite_output_dir \
 	--do_train \
 	--do_eval \
+	--fp16 \
 	--predict_with_generate \
 	--do_normalize_eval \
 	--streaming \
-	--use_auth_token \
-	--push_to_hub
+	--use_auth_token