cleanup, more

rwth-i6 · Dec 9, 2024 · a5d353d · a5d353d
1 parent f3ad5a5
commit a5d353d
Showing 1 changed file with 5 additions and 48 deletions.
diff --git a/users/zeyer/experiments/exp2024_04_23_baselines/claix2023.py b/users/zeyer/experiments/exp2024_04_23_baselines/claix2023.py
@@ -210,10 +210,6 @@ def py():
         for cr_ctc in [None, {"cr_loss_scale": 0.2}, {"cr_loss_scale": 0.5}, {"cr_loss_scale": 1.0}]:
             # TODO also adapt specaug for CR...
             use_cr_ctc = cr_ctc is not None
-            if use_cr_ctc:
-                cr_ctc: Dict[str, Any]
-                cr_ctc = cr_ctc.copy()
-                cr_ctc["use_fixed_ctc_grad"] = "v2"
             name = f"crLoss{cr_ctc['cr_loss_scale']}-" if use_cr_ctc else ""
             if opts.get("time_downsampling"):
                 name += f"time{opts['time_downsampling']}-"
@@ -266,7 +262,7 @@ def py():
                     # purely used for training
                     "aux_attention_decoder": rf.build_dict(TransformerDecoder, num_layers=6),
                     **(cr_ctc if use_cr_ctc else {}),
-                    **({"aed_loss_bug_fix": True} if use_cr_ctc else {}),
+                    **({"use_fixed_ctc_grad": "v2", "aed_loss_bug_fix": True} if use_cr_ctc else {}),
                     "max_seq_length_default_target": None,
                     # Note on max seq len stats: Before, when we used max_seq_length_default_target=75 with bpe10k,
                     # out of 281241 seqs in train, we removed only 71 seqs.
@@ -349,48 +345,6 @@ def py():
             "num_enc_layers": 12,
             "out_blank_separated": True,
         },
-        config_updates={
-            **_get_cfg_lrlin_oclr_by_bs_nep_v3(150_000, 100, batch_size_factor=_batch_size_factor),
-            "optimizer.weight_decay": 1e-2,
-            "max_seq_length_default_target": None,
-            # Note on max seq len stats: Before, when we used max_seq_length_default_target=75 with bpe10k,
-            # out of 281241 seqs in train, we removed only 71 seqs.
-            # With max seq len 19.5 secs on the audio, we also remove exactly 71 seqs.
-            "max_seq_length_default_input": 19.5 * _raw_sample_rate,
-            "__train_audio_preprocess": speed_pert_librosa_config,
-            "speed_pert_discrete_values": [0.7, 0.8, 0.9, 1.0, 1.1],
-            "aux_attention_decoder": rf.build_dict(TransformerDecoder, num_layers=6),  # purely used for training
-        },
-        post_config_updates={"log_grad_norm": True, "__multi_proc_dataset_opts": {"num_workers": 25}},
-        vocab="spm512",
-        train_vocab_opts={"other_opts": {"class": "SamplingBytePairEncoding", "breadth_prob": 0.01}},
-        dataset_train_opts={"train_epoch_split": 1, "train_epoch_wise_filter": None},
-        # avoid OOM
-        env_updates={"PYTORCH_CUDA_ALLOC_CONF": "expandable_segments:True"},
-    )
-
-    ctc_train_exp(
-        "time4-n12-spm512-blankSep-auxAED-b150k-ctcFixGrad",
-        config_96gb_bf16_accgrad1,
-        model_config={
-            "enc_input_layer": rf.build_dict(
-                ConformerConvSubsample,
-                out_dims=[32, 64, 64],
-                filter_sizes=[(3, 3), (3, 3), (3, 3)],
-                pool_sizes=[(1, 2)],
-                strides=[(1, 1), (2, 1), (2, 1)],
-            ),
-            "enc_conformer_layer": rf.build_dict(
-                ConformerEncoderLayer,
-                ff=rf.build_dict(
-                    ConformerPositionwiseFeedForward, activation=rf.build_dict(rf.relu_square), with_bias=False
-                ),
-                num_heads=8,
-            ),
-            "feature_batch_norm": True,
-            "num_enc_layers": 12,
-            "out_blank_separated": True,
-        },
         config_updates={
             **_get_cfg_lrlin_oclr_by_bs_nep_v3(150_000, 100, batch_size_factor=_batch_size_factor),
             "optimizer.weight_decay": 1e-2,
@@ -506,6 +460,7 @@ def py():
                 "ctc_am_scale": am_scale,
                 "ctc_prior_scale": prior_scale,
                 "ctc_prior_type": prior_type,
+                "use_fixed_ctc_grad": "v2",
             },
             post_config_updates={"log_grad_norm": True, "__multi_proc_dataset_opts": {"num_workers": 25}},
             vocab="spm512",
@@ -589,6 +544,7 @@ def py():
             "__train_audio_preprocess": speed_pert_librosa_config,
             "speed_pert_discrete_values": [0.7, 0.8, 0.9, 1.0, 1.1],
             "aux_attention_decoder": rf.build_dict(TransformerDecoder, num_layers=6),  # purely used for training
+            "use_fixed_ctc_grad": "v2",
         },
         post_config_updates={"log_grad_norm": True, "__multi_proc_dataset_opts": {"num_workers": 25}},
         vocab="spm10k",
@@ -640,6 +596,7 @@ def py():
                 "ctc_am_scale": am_scale,
                 "ctc_prior_scale": prior_scale,
                 "ctc_prior_type": prior_type,
+                "use_fixed_ctc_grad": "v2",
             },
             post_config_updates={"log_grad_norm": True, "__multi_proc_dataset_opts": {"num_workers": 25}},
             vocab="spm512",
@@ -878,7 +835,7 @@ def py():
                 "speed_pert_discrete_values": [0.7, 0.8, 0.9, 1.0, 1.1],
                 "aux_attention_decoder": rf.build_dict(TransformerDecoder, num_layers=6),  # purely used for training
                 **(cr_ctc if use_cr_ctc else {}),
-                **({"aed_loss_bug_fix": True} if use_cr_ctc else {}),
+                **({"use_fixed_ctc_grad": "v2", "aed_loss_bug_fix": True} if use_cr_ctc else {}),
             },
             config_deletes=["aux_loss_layers"],
             post_config_updates={"log_grad_norm": True, "__multi_proc_dataset_opts": {"num_workers": 25}},