Fix bugs in criteo.py that leads to NaN problem

The original script simply added 3 to the target value before taking the log. This led to the issue that in data preprocessing, if there was a value of -3, it would result in a value of -inf. This problem was mentioned in the issue facebookresearch/dlrm#363 (comment). I changed the preprocessing operation to dense_np -= dense_np.min() - 2 in the tsv_to_npys function, and correctly handled the Criteo Kaggle dataset.

torchrec/datasets/criteo.py

@@ -258,7 +258,7 @@ def row_mapper(row: List[str]) -> Tuple[List[int], List[int], int]:
         del labels
 
         # Log is expensive to compute at runtime.
-        dense_np -= (dense_np.min() - 2)
+        dense_np -= dense_np.min() - 2
         dense_np = np.log(dense_np, dtype=np.float32)
 
         # To be consistent with dense and sparse.