Fix depreciated commands

examples/glue/README.md

@@ -91,6 +91,16 @@ This reduces computation time to about 30 minutes on an A100 (80GB) GPU.
 ----------------------------------------------------------------------------------------------------------------------------------
 ```
 
+```bash
+python analyze.py --dataset_name sst2 \
+    --query_batch_size 175 \
+    --train_batch_size 128 \
+    --checkpoint_dir ./checkpoints \
+    --factor_strategy ekfac \
+    --use_half_precision \
+    --query_gradient_rank 32
+```
+
 ## Counterfactual Evaluation
 
 Let's evaluate the impact of removing top positively influential training examples on query misclassification. 

examples/glue/analyze.py

@@ -188,8 +188,8 @@ def main():
         scores_name += "_half"
     rank = args.query_gradient_rank if args.query_gradient_rank != -1 else None
     if rank is not None:
-        score_args.query_gradient_rank = rank
-        score_args.num_query_gradient_accumulations = 10
+        score_args.query_gradient_low_rank = rank
+        score_args.query_gradient_accumulation_steps = 10
         scores_name += f"_qlr{rank}"
     analyzer.compute_pairwise_scores(
         score_args=score_args,

examples/imagenet/README.md

@@ -16,8 +16,9 @@ To compute pairwise influence scores on 1000 query data points using the `ekfac`
 ```bash
 python analyze.py --dataset_dir PATH_TO_IMAGENET \
     --query_gradient_rank -1 \
+    --factor_batch_size 512 \
     --query_batch_size 100 \
-    --train_batch_size 300 \
+    --train_batch_size 256 \
     --factor_strategy ekfac
 ```
 

examples/imagenet/analyze.py

@@ -34,6 +34,12 @@ def parse_args():
         default=-1,
         help="Rank for the low-rank query gradient approximation.",
     )
+    parser.add_argument(
+        "--factor_batch_size",
+        type=int,
+        default=512,
+        help="Batch size for computing factors.",
+    )
     parser.add_argument(
         "--query_batch_size",
         type=int,
@@ -85,7 +91,7 @@ def main():
     )
     # Configure parameters for DataLoader.
     dataloader_kwargs = DataLoaderKwargs(
-        num_workers=4, pin_memory=True,
+        num_workers=4,
     )
     analyzer.set_dataloader_kwargs(dataloader_kwargs)
 
@@ -98,7 +104,7 @@ def main():
     analyzer.fit_all_factors(
         factors_name=factors_name,
         dataset=train_dataset,
-        per_device_batch_size=None,
+        per_device_batch_size=args.factor_batch_size,
         factor_args=factor_args,
         overwrite_output_dir=False,
     )

examples/imagenet/ddp_analyze.py

@@ -106,7 +106,7 @@ def main():
     )
     # Configure parameters for DataLoader.
     dataloader_kwargs = DataLoaderKwargs(
-        num_workers=4, pin_memory=True,
+        num_workers=4,
     )
     analyzer.set_dataloader_kwargs(dataloader_kwargs)