more

users/zeyer/datasets/librispeech.py

@@ -73,7 +73,7 @@ def _get_spm_vocab(
     dim_str = str(dim)
     if isinstance(dim, str):
         # Not sure if power-of-two or just multiple-of-64, but 10240 has more 2s in it (2048*5) than 10048.
-        dim = {"20k": 20_480, "10k": 10_240, "5k": 5_120, "4k": 4_096, "1k": 1_024}[dim]
+        dim = {"20k": 20_480, "10k": 10_240, "5k": 5_120, "4k": 4_096, "1k": 1_024, "512": 512, "128": 128}[dim]
     assert isinstance(dim, int) and dim >= 10
 
     # https://github.com/google/sentencepiece/blob/master/doc/options.md

users/zeyer/experiments/exp2024_04_23_baselines/ctc.py

@@ -397,7 +397,14 @@ def py():
         ("spm4k", "spm", 0.7),
         ("spm4k", "bpe", 0.01),
         ("spm1k", None, None),  # 6.07
+        ("spm1k", "spm", 0.7),
         ("spm1k", "bpe", 0.01),  # 6.13 (but dev-clean,test-* are better than no sampling)
+        ("spm_bpe1k", None, None),
+        ("spm_bpe1k", "bpe", 0.01),
+        ("spm512", None, None),
+        ("spm512", "bpe", 0.01),
+        ("spm128", None, None),
+        ("spm128", "bpe", 0.01),
     ]:
         train_exp(
             f"v6-relPosAttDef-bhv20-11gb-f32-bs15k-accgrad1-mgpu4-pavg100"