minor BCO fixes (#1923)

* checkpointing BCO UDM classifier * kto_config remove unused parameters * BCO fix loading * kto_config remove unused parameters * kto_config remove unused parameters --------- Co-authored-by: Clara Luise Pohland <[email protected]> Co-authored-by: Kashif Rasul <[email protected]>
huggingface · Aug 14, 2024 · c1b272f · c1b272f
1 parent f05f63c
commit c1b272f
Show file tree

Hide file tree

Showing 3 changed files with 15 additions and 37 deletions.
diff --git a/tests/test_trainers_args.py b/tests/test_trainers_args.py
@@ -197,10 +197,6 @@ def test_kto(self):
                 model_init_kwargs={"trust_remote_code": True},
                 ref_model_init_kwargs={"trust_remote_code": True},
                 dataset_num_proc=4,
-                loss_type="bco",
-                prompt_sample_size=512,
-                min_density_ratio=0.2,
-                max_density_ratio=20.0,
             )
             trainer = KTOTrainer(model="gpt2", ref_model="gpt2", args=args, train_dataset=dataset, tokenizer=tokenizer)
             self.assertEqual(trainer.args.max_length, 256)
@@ -218,10 +214,6 @@ def test_kto(self):
             self.assertEqual(trainer.args.model_init_kwargs, {"trust_remote_code": True})
             self.assertEqual(trainer.args.ref_model_init_kwargs, {"trust_remote_code": True})
             self.assertEqual(trainer.args.dataset_num_proc, 4)
-            self.assertEqual(trainer.args.loss_type, "bco")
-            self.assertEqual(trainer.args.prompt_sample_size, 512)
-            self.assertEqual(trainer.args.min_density_ratio, 0.2)
-            self.assertEqual(trainer.args.max_density_ratio, 20.0)
 
     def test_online_dpo(self):
         tokenizer = AutoTokenizer.from_pretrained("gpt2")

diff --git a/trl/trainer/bco_trainer.py b/trl/trainer/bco_trainer.py
@@ -72,7 +72,8 @@
 if TYPE_CHECKING:
     from transformers import PreTrainedModel, PreTrainedTokenizer
 
-RUNNING_NAME = "running.pt"
+RUNNING_NAME = "running.json"
+CLF_NAME = "clf.pt"
 
 
 def _tokenize(
@@ -822,6 +823,9 @@ def _save_optimizer_and_scheduler(self, output_dir):
 
         self.running.save_to_json(os.path.join(output_dir, RUNNING_NAME))
 
+        if self.match_underlying_distribution:
+            torch.save(self.clf.get_params(), os.path.join(output_dir, CLF_NAME))
+
     def _load_optimizer_and_scheduler(self, checkpoint):
         super()._load_optimizer_and_scheduler(checkpoint)
 
@@ -831,7 +835,15 @@ def _load_optimizer_and_scheduler(self, checkpoint):
         running_file = os.path.join(checkpoint, RUNNING_NAME)
         if not os.path.isfile(running_file):
             warnings.warn(f"Missing file {running_file}. Will use a new running delta value for BCO loss calculation")
-        self.running = RunningMoments.load_from_json(self.accelerator, running_file)
+        else:
+            self.running = RunningMoments.load_from_json(self.accelerator, running_file)
+
+        if self.match_underlying_distribution:
+            clf_file = os.path.join(checkpoint, CLF_NAME)
+            if not os.path.isfile(running_file):
+                warnings.warn(f"Missing file {clf_file}. Will use a new UDM classifier for BCO loss calculation")
+            else:
+                self.clf.set_params(**torch.load(clf_file, weights_only=True, map_location="cpu"))
 
     @contextmanager
     def null_ref_context(self):

diff --git a/trl/trainer/kto_config.py b/trl/trainer/kto_config.py
@@ -12,12 +12,10 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 from dataclasses import dataclass
-from typing import Dict, Literal, Optional
+from typing import Dict, Optional
 
 from transformers import TrainingArguments
 
-from ..import_utils import is_sklearn_available
-
 
 @dataclass
 class KTOConfig(TrainingArguments):
@@ -60,14 +58,6 @@ class KTOConfig(TrainingArguments):
             Dict of Optional kwargs to pass when instantiating the ref model from a string.
         dataset_num_proc: (`Optional[int]`, *optional*, defaults to `None`):
             Number of processes to use for processing the datasets.
-        loss_type: (`Literal["kto", "bco"]`, *optional*):
-            The type of loss to use. Either `"kto"` the default KTO loss, `"bco"` loss from [BCO](https://huggingface.co/papers/2404.04656) paper.
-        prompt_sample_size: (`int`, defaults to 1024):
-            Number of prompts that are fed to density ratio classifier.
-        min_density_ratio: (`float`, defaults to 0.5):
-            The minimum value of the density ratio. The estimated density ratio is clamped to this value.
-        max_density_ratio: (`float`, defaults to 10.0):
-            The maximum value of the density ratio. The estimated density ratio is clamped to this value.
     """
 
     max_length: Optional[int] = None
@@ -92,19 +82,3 @@ class KTOConfig(TrainingArguments):
     model_init_kwargs: Optional[Dict] = None
     ref_model_init_kwargs: Optional[Dict] = None
     dataset_num_proc: Optional[int] = None
-
-    loss_type: Literal["kto", "bco"] = "kto"
-
-    # BCO config
-    prompt_sample_size: int = 1024
-    min_density_ratio: float = 0.5
-    max_density_ratio: float = 10.0
-
-    def __post_init__(self):
-        super().__post_init__()
-
-        if self.loss_type == "bco" and not is_sklearn_available():
-            raise ImportError(
-                "You need to install scikit-learn to use loss_type='bco' "
-                "You can install it with `pip install scikit-learn`."
-            )