Skip to content

Commit 6ad1a0c

Browse files
JackTemakiJackTemaki
committed
push update for ls standalone
1 parent 89a3991 commit 6ad1a0c

File tree

4 files changed

+74
-67
lines changed

4 files changed

+74
-67
lines changed

‎users/rossenbach/experiments/librispeech/standalone_2023/ctc_bpe/exp_ls100_1023_base.py‎

Lines changed: 63 additions & 59 deletions
Original file line numberDiff line numberDiff line change
@@ -204,64 +204,68 @@ def run_exp(ft_name, datasets, train_args, search_args=None, with_prior=False, n
204204

205205

206206
# from here on onwards, use default AdamW with same OCLR
207-
train_args_adamw_02 = {
208-
"config": {
209-
"optimizer": {"class": "adamw", "epsilon": 1e-16, "weight_decay": 1e-2},
210-
"learning_rates": list(np.linspace(1e-5, 1e-3, 150)) + list(np.linspace(1e-3, 1e-6, 150)),
211-
#############
212-
"batch_size": 200 * 16000,
213-
"max_seq_length": {"audio_features": 35 * 16000},
214-
"accum_grad_multiple_step": 2,
215-
},
216-
}
217-
218-
model_config_smaller = ModelConfig(
219-
feature_extraction_config=fe_config,
220-
frontend_config=frontend_config,
221-
specaug_config=specaug_config,
222-
label_target_size=vocab_size_without_blank,
223-
conformer_size=384,
224-
num_layers=12,
225-
num_heads=4,
226-
ff_dim=384,
227-
att_weights_dropout=0.2,
228-
conv_dropout=0.2,
229-
ff_dropout=0.2,
230-
mhsa_dropout=0.2,
231-
conv_kernel_size=9,
232-
final_dropout=0.2,
233-
specauc_start_epoch=1,
234-
)
207+
# train_args_adamw_02 = {
208+
# "config": {
209+
# "optimizer": {"class": "adamw", "epsilon": 1e-16, "weight_decay": 1e-2},
210+
# "learning_rates": list(np.linspace(1e-5, 1e-3, 150)) + list(np.linspace(1e-3, 1e-6, 150)),
211+
# #############
212+
# "batch_size": 200 * 16000,
213+
# "max_seq_length": {"audio_features": 35 * 16000},
214+
# "accum_grad_multiple_step": 2,
215+
# },
216+
# }
235217

236-
train_args = {
237-
**copy.deepcopy(train_args_adamw_02),
238-
"network_module": "ctc.conformer_1023.i6modelsV1_VGG4LayerActFrontendV1_v6",
239-
"net_args": {"model_config_dict": asdict(model_config_smaller)},
240-
}
241-
for lm_weight in [1.6, 1.8, 2.0, 2.2]:
242-
for prior_scale in [0.3, 0.5]:
243-
search_args = {
244-
**default_search_args,
245-
"lm_weight": lm_weight,
246-
"prior_scale": prior_scale,
247-
}
248-
run_exp(
249-
prefix_name + "conformer_1023/i6modelsV1_VGG4LayerActFrontendV1_v6_peaknorm_smaller_decay1e-2/lm%.1f_prior%.2f_bs1024_th14" % (
250-
lm_weight, prior_scale),
251-
datasets=train_data, train_args=train_args, search_args=search_args, with_prior=True)
218+
# model_config_smaller = ModelConfig(
219+
# feature_extraction_config=fe_config,
220+
# frontend_config=frontend_config,
221+
# specaug_config=specaug_config,
222+
# label_target_size=vocab_size_without_blank,
223+
# conformer_size=384,
224+
# num_layers=12,
225+
# num_heads=4,
226+
# ff_dim=384,
227+
# att_weights_dropout=0.2,
228+
# conv_dropout=0.2,
229+
# ff_dropout=0.2,
230+
# mhsa_dropout=0.2,
231+
# conv_kernel_size=9,
232+
# final_dropout=0.2,
233+
# specauc_start_epoch=1,
234+
# )
235+
#
236+
# train_args = {
237+
# **copy.deepcopy(train_args_adamw_02),
238+
# "network_module": "ctc.conformer_1023.i6modelsV1_VGG4LayerActFrontendV1_v6",
239+
# "net_args": {"model_config_dict": asdict(model_config_smaller)},
240+
# }
252241

253-
model_config_smaller_start11 = copy.deepcopy(model_config_smaller)
254-
model_config_smaller_start11.specauc_start_epoch = 11
255-
train_args_start11 = copy.deepcopy(train_args)
256-
train_args_start11["net_args"]["model_config_dict"] = asdict(model_config_smaller_start11)
257-
for lm_weight in [1.6, 1.8, 2.0, 2.2]:
258-
for prior_scale in [0.3, 0.5]:
259-
search_args = {
260-
**default_search_args,
261-
"lm_weight": lm_weight,
262-
"prior_scale": prior_scale,
263-
}
264-
run_exp(
265-
prefix_name + "conformer_1023/i6modelsV1_VGG4LayerActFrontendV1_v6_peaknorm_smaller_decay1e-2_start11/lm%.1f_prior%.2f_bs1024_th14" % (
266-
lm_weight, prior_scale),
267-
datasets=train_data, train_args=train_args_start11, search_args=search_args, with_prior=True)
242+
# Diverged
243+
244+
# for lm_weight in [1.6, 1.8, 2.0, 2.2]:
245+
# for prior_scale in [0.3, 0.5]:
246+
# search_args = {
247+
# **default_search_args,
248+
# "lm_weight": lm_weight,
249+
# "prior_scale": prior_scale,
250+
# }
251+
# run_exp(
252+
# prefix_name + "conformer_1023/i6modelsV1_VGG4LayerActFrontendV1_v6_peaknorm_smaller_decay1e-2/lm%.1f_prior%.2f_bs1024_th14" % (
253+
# lm_weight, prior_scale),
254+
# datasets=train_data, train_args=train_args, search_args=search_args, with_prior=True)
255+
256+
# This one was worse than the baseline 16.5 -> 17.9
257+
# model_config_smaller_start11 = copy.deepcopy(model_config_smaller)
258+
# model_config_smaller_start11.specauc_start_epoch = 11
259+
# train_args_start11 = copy.deepcopy(train_args)
260+
# train_args_start11["net_args"]["model_config_dict"] = asdict(model_config_smaller_start11)
261+
# for lm_weight in [1.6, 1.8, 2.0, 2.2]:
262+
# for prior_scale in [0.3, 0.5]:
263+
# search_args = {
264+
# **default_search_args,
265+
# "lm_weight": lm_weight,
266+
# "prior_scale": prior_scale,
267+
# }
268+
# run_exp(
269+
# prefix_name + "conformer_1023/i6modelsV1_VGG4LayerActFrontendV1_v6_peaknorm_smaller_decay1e-2_start11/lm%.1f_prior%.2f_bs1024_th14" % (
270+
# lm_weight, prior_scale),
271+
# datasets=train_data, train_args=train_args_start11, search_args=search_args, with_prior=True)

‎users/rossenbach/experiments/librispeech/standalone_2023/data/common.py‎

Lines changed: 2 additions & 1 deletion
Original file line numberDiff line numberDiff line change
@@ -16,7 +16,8 @@
1616
from i6_experiments.users.rossenbach.common_setups.returnn.datastreams.base import Datastream
1717
from i6_experiments.users.rossenbach.datasets.librispeech import get_mixed_cv_segments
1818

19-
from returnn_common.datasets import Dataset, OggZipDataset, MetaDataset
19+
# from returnn_common.datasets import Dataset, OggZipDataset, MetaDataset
20+
from i6_experiments.common.setups.returnn.datasets import Dataset, OggZipDataset, MetaDataset
2021

2122
from ..default_tools import MINI_RETURNN_ROOT, RETURNN_EXE
2223

‎users/rossenbach/experiments/librispeech/standalone_2023/default_tools.py‎

Lines changed: 3 additions & 1 deletion
Original file line numberDiff line numberDiff line change
@@ -4,7 +4,9 @@
44

55
# python from apptainer
66
RETURNN_EXE = tk.Path("/usr/bin/python3", hash_overwrite="GENERIC_RETURNN_LAUNCHER")
7-
MINI_RETURNN_ROOT = tk.Path("/u/rossenbach/src/NoReturnn", hash_overwrite="LIBRISPEECH_DEFAULT_RETURNN_ROOT")
7+
# MINI_RETURNN_ROOT = tk.Path("/u/rossenbach/src/NoReturnn", hash_overwrite="LIBRISPEECH_DEFAULT_RETURNN_ROOT")
8+
MINI_RETURNN_ROOT = CloneGitRepositoryJob("https://github.com/JackTemaki/MiniReturnn", commit="1ccdcb77414cb062b4fe69f051238d01022e2b15").out_repository
9+
MINI_RETURNN_ROOT.hash_overwrite="LIBRISPEECH_DEFAULT_RETURNN_ROOT"
810

911
from i6_experiments.common.tools.sctk import compile_sctk
1012

‎users/rossenbach/experiments/librispeech/standalone_2023/serializer.py‎

Lines changed: 6 additions & 6 deletions
Original file line numberDiff line numberDiff line change
@@ -51,12 +51,12 @@ def get_pytorch_serializer_v3(
5151
)
5252

5353
# TODO: add flag to switch and maybe move to default tools
54-
# i6_models_repo = CloneGitRepositoryJob(
55-
# url="https://github.com/rwth-i6/i6_models",
56-
# commit="1e94a4d9d1aa48fe3ac7f60de2cd7bd3fea19c3e",
57-
# checkout_folder_name="i6_models"
58-
# ).out_repository
59-
i6_models_repo = tk.Path("/u/rossenbach/experiments/tts_asr_2023_pycharm/i6_models")
54+
i6_models_repo = CloneGitRepositoryJob(
55+
url="https://github.com/rwth-i6/i6_models",
56+
commit="933c6c13f7d6c74e5a59af0257e17c208dae9da3",
57+
checkout_folder_name="i6_models"
58+
).out_repository
59+
#i6_models_repo = tk.Path("/u/rossenbach/experiments/tts_asr_2023_pycharm/i6_models")
6060
i6_models_repo.hash_overwrite = "LIBRISPEECH_DEFAULT_I6_MODELS"
6161
i6_models = ExternalImport(import_path=i6_models_repo)
6262

0 commit comments

Comments
 (0)