fixed multiple epochs

commit-0 · wenting-zhao · Dec 7, 2024 · Dec 8, 2024 · Dec 8, 2024 · Dec 8, 2024
commit 177a5a8e1795719663ef8ce77642bcd558f311b0
diff --git a/examples/star/star.py b/examples/star/star.py
@@ -1,5 +1,6 @@
 """Main STaR Loop"""
 
+from copy import deepcopy
 from datasets import Dataset, DatasetDict, load_dataset
 from inference import generate_predictions
 from train import train
@@ -21,13 +22,13 @@ def main():
         ds[split] = ds[split].add_column(name="text", column=texts)
 
     model_name = args.model_name_or_path
-    ds["train"] = ds["train"].select(range(10))
+    output_dir = deepcopy(args.output_dir)
     for i in range(args.iteration):
         # sample
         all_samples = generate_predictions(
             model_name, ds["train"], args.temperature, args.n
         )
-        ds["train"].add_column(name="sample", column=all_samples).to_json(f"{args.output_dir}/data/samples-iter{i}.json")
+        ds["train"].add_column(name="sample", column=all_samples).to_json(f"{output_dir}/data/samples-iter{i}.json")
         assert len(ds["train"]) == len(all_samples)
 
         # verify and construct the training set
@@ -43,10 +44,10 @@ def main():
                     passed_examples.append(example)
                     break
         raw_datasets = DatasetDict({"train": Dataset.from_list(passed_examples), "validation": ds["validation"]})
-        raw_datasets["train"].to_json(f"{args.output_dir}/data/verified-samples-iter{i}.json")
+        raw_datasets["train"].to_json(f"{output_dir}/data/verified-samples-iter{i}.json")
 
         # train
-        args.output_dir = f"{args.output_dir}/models-iter{i}"
+        args.output_dir = f"{output_dir}/models-iter{i}"
         train(raw_datasets, model_name, args)
         model_name = args.output_dir
 

diff --git a/examples/star/train.py b/examples/star/train.py
@@ -254,10 +254,6 @@ def tokenize_function(examples):
     # The trackers initializes automatically on the main process.
     if args.with_tracking:
         experiment_config = vars(args)
-        # TensorBoard cannot log Enums, need the raw value
-        experiment_config["lr_scheduler_type"] = experiment_config[
-            "lr_scheduler_type"
-        ].value
         accelerator.init_trackers("clm_no_trainer", experiment_config)
 
     # Train!
@@ -407,8 +403,6 @@ def tokenize_function(examples):
             with open(os.path.join(args.output_dir, "all_results.json"), "w") as f:
                 json.dump({"perplexity": perplexity}, f)
     cleanup(model)
-    #cleanup(optimizer)
-    #cleanup(lr_scheduler)
 
 
 if __name__ == "__main__":