feat: Change eval batch size (#128)

* Surface # of eval batches and # of eval sequences

* fix formatting

* fix print statement accidentally left in

sae_lens/training/activations_store.py

@@ -191,12 +191,12 @@ def dataloader(self) -> Iterator[Any]:
             self._dataloader = self.get_data_loader()
         return self._dataloader
 
-    def get_batch_tokens(self):
+    def get_batch_tokens(self, batch_size: int | None = None):
         """
         Streams a batch of tokens from a dataset.
         """
-
-        batch_size = self.store_batch_size
+        if not batch_size:
+            batch_size = self.store_batch_size
         context_size = self.context_size
         device = self.device
 

sae_lens/training/config.py

@@ -106,6 +106,10 @@ class LanguageModelSAERunnerConfig:
 
     dead_feature_threshold: float = 1e-8
 
+    # Evals
+    n_eval_batches: int = 10
+    n_eval_seqs: int | None = None  # useful if evals cause OOM
+
     # WANDB
     log_to_wandb: bool = True
     log_activations_store_to_wandb: bool = False

sae_lens/training/evals.py

@@ -17,6 +17,8 @@ def run_evals(
     model: HookedRootModule,
     n_training_steps: int,
     suffix: str = "",
+    n_eval_batches: int = 10,
+    n_eval_seqs: int | None = None,
 ) -> Mapping[str, Any]:
     hook_point = sparse_autoencoder.cfg.hook_point
     hook_point_layer = sparse_autoencoder.hook_point_layer
@@ -25,14 +27,15 @@ def run_evals(
         layer=hook_point_layer
     )
     ### Evals
-    eval_tokens = activation_store.get_batch_tokens()
+    eval_tokens = activation_store.get_batch_tokens(n_eval_seqs)
 
     # Get Reconstruction Score
     losses_df = recons_loss_batched(
         sparse_autoencoder,
         model,
         activation_store,
-        n_batches=10,
+        n_batches=n_eval_batches,
+        n_eval_seqs=n_eval_seqs,
     )
 
     recons_score = losses_df["score"].mean()
@@ -100,10 +103,11 @@ def recons_loss_batched(
     model: HookedRootModule,
     activation_store: ActivationsStore,
     n_batches: int = 100,
+    n_eval_seqs: int | None = None,
 ):
     losses = []
     for _ in range(n_batches):
-        batch_tokens = activation_store.get_batch_tokens()
+        batch_tokens = activation_store.get_batch_tokens(n_eval_seqs)
         score, loss, recons_loss, zero_abl_loss = get_recons_loss(
             sparse_autoencoder, model, batch_tokens
         )

sae_lens/training/lm_runner.py

@@ -85,6 +85,8 @@ def language_model_sae_runner(cfg: LanguageModelSAERunnerConfig):
         wandb_log_frequency=cfg.wandb_log_frequency,
         eval_every_n_wandb_logs=cfg.eval_every_n_wandb_logs,
         autocast=cfg.autocast,
+        n_eval_batches=cfg.n_eval_batches,
+        n_eval_seqs=cfg.n_eval_seqs,
     ).sae_group
 
     if cfg.log_to_wandb:

sae_lens/training/train_sae_on_language_model.py

@@ -188,6 +188,8 @@ def train_sae_on_language_model(
     wandb_log_frequency: int = 50,
     eval_every_n_wandb_logs: int = 100,
     autocast: bool = False,
+    n_eval_batches: int = 10,
+    n_eval_seqs: int | None = None,
 ) -> SparseAutoencoderDictionary:
     """
     @deprecated Use `train_sae_group_on_language_model` instead. This method is kept for backward compatibility.
@@ -203,6 +205,8 @@ def train_sae_on_language_model(
         wandb_log_frequency=wandb_log_frequency,
         eval_every_n_wandb_logs=eval_every_n_wandb_logs,
         autocast=autocast,
+        n_eval_batches=n_eval_batches,
+        n_eval_seqs=n_eval_seqs,
     ).sae_group
 
 
@@ -223,6 +227,8 @@ def train_sae_group_on_language_model(
     wandb_log_frequency: int = 50,
     eval_every_n_wandb_logs: int = 100,
     autocast: bool = False,
+    n_eval_batches: int = 10,
+    n_eval_seqs: int | None = None,
 ) -> TrainSAEGroupOutput:
     total_training_tokens = get_total_training_tokens(sae_group=sae_group)
     _update_sae_lens_training_version(sae_group)
@@ -325,6 +331,8 @@ def interrupt_callback(sig_num: Any, stack_frame: Any):
                                 model,
                                 training_run_state.n_training_steps,
                                 suffix=wandb_suffix,
+                                n_eval_batches=n_eval_batches,
+                                n_eval_seqs=n_eval_seqs,
                             )
                             sparse_autoencoder.train()