Skip to content

Instantly share code, notes, and snippets.

@Max-Ryujin

Max-Ryujin/experiments.py Secret

Created October 31, 2023 09:46
Show Gist options
  • Save Max-Ryujin/690647f79773d6cd8338c524be039040 to your computer and use it in GitHub Desktop.
Save Max-Ryujin/690647f79773d6cd8338c524be039040 to your computer and use it in GitHub Desktop.
Download ZIP
Config for the setup. The relevant function is run_scf_baseline_big()
Raw
experiments.py
import copy
import os.path
from sisyphus import tk, gs
from sisyphus.delayed_ops import DelayedFormat
from i6_core.meta.system import CorpusObject
from i6_core.lexicon.modification import AddEowPhonemesToLexiconJob
from i6_core.returnn.config import CodeWrapper
from i6_core.recognition import Hub5ScoreJob
from i6_experiments.common.datasets.switchboard.corpus_eval import get_hub5e00
from i6_experiments.common.setups.rasr.util import RasrDataInput
from i6_experiments.users.berger.recipe.lexicon.modification import DeleteEmptyOrthJob, MakeBlankLexiconJob
from i6_experiments.users.vieting.tools.report import Report
# TODO: run_gmm_system_from_common might be copied here for stability
from i6_experiments.users.vieting.experiments.switchboard.hybrid.feat.experiments import run_gmm_system_from_common
from i6_experiments.users.vieting.experiments.switchboard.ctc.feat.transducer_system_v2 import (
TransducerSystem,
ReturnnConfigs,
ScorerInfo,
SearchTypes,
)
from .baseline_args import get_nn_args as get_nn_args_baseline
from .data import get_corpus_data_inputs_oggzip # TODO: might be copied here for stability
from .default_tools import RASR_BINARY_PATH, RETURNN_ROOT, RETURNN_EXE, SCTK_BINARY_PATH
def get_datasets(**kwargs):
gmm_system = run_gmm_system_from_common()
# TODO: get oggzip independent of GMM system
# noinspection PyTypeChecker
(
nn_train_data_inputs,
nn_cv_data_inputs,
nn_devtrain_data_inputs,
nn_dev_data_inputs,
nn_test_data_inputs,
train_corpus_path,
traincv_segments,
) = get_corpus_data_inputs_oggzip(
gmm_system,
partition_epoch={"train": 6, "dev": 1},
returnn_root=RETURNN_ROOT,
returnn_python_exe=RETURNN_EXE,
**kwargs,
)
returnn_datasets = {
"train":{
"class": "MultiProcDataset",
"dataset": nn_train_data_inputs["switchboard.train"].get_data_dict()["datasets"]["ogg"],
"num_workers": 2,
"buffer_size": 5,
},
"dev": nn_cv_data_inputs["switchboard.cv"].get_data_dict()["datasets"]["ogg"],
"eval_datasets": {
"devtrain": nn_devtrain_data_inputs["switchboard.devtrain"].get_data_dict()["datasets"]["ogg"],
},
}
lexicon = gmm_system.crp["switchboard"].lexicon_config.file
lexicon = DeleteEmptyOrthJob(lexicon).out_lexicon
rasr_loss_lexicon = MakeBlankLexiconJob(lexicon).out_lexicon
nonword_phones = ["[LAUGHTER]", "[NOISE]", "[VOCALIZEDNOISE]"]
recog_lexicon = AddEowPhonemesToLexiconJob(rasr_loss_lexicon, nonword_phones=nonword_phones).out_lexicon
rasr_loss_corpus = train_corpus_path
rasr_loss_segments = traincv_segments
hub5e00 = get_hub5e00()
corpus_object = CorpusObject()
corpus_object.corpus_file = hub5e00.bliss_corpus
corpus_object.audio_format = nn_dev_data_inputs["hub5e00"].crp.audio_format
corpus_object.duration = nn_dev_data_inputs["hub5e00"].crp.corpus_duration
dev_corpora = {
"hub5e00": RasrDataInput(
corpus_object=corpus_object,
lexicon={
"filename": recog_lexicon,
"normalize_pronunciation": False,
"add_all": True,
"add_from_lexicon": False,
},
lm={"filename": nn_dev_data_inputs["hub5e00"].crp.language_model_config.file, "type": "ARPA"},
stm=hub5e00.stm,
glm=hub5e00.glm,
)
}
return returnn_datasets, rasr_loss_corpus, rasr_loss_segments, rasr_loss_lexicon, dev_corpora
def run_test_mel():
gs.ALIAS_AND_OUTPUT_SUBDIR = "experiments/switchboard/ctc/feat/"
(
returnn_datasets,
rasr_loss_corpus_path,
rasr_loss_corpus_segments,
rasr_loss_lexicon_path,
dev_corpora,
) = get_datasets()
returnn_args = {
"batch_size": 10000,
"rasr_binary_path": RASR_BINARY_PATH,
"rasr_loss_corpus_path": rasr_loss_corpus_path,
"rasr_loss_corpus_segments": rasr_loss_corpus_segments,
"rasr_loss_lexicon_path": rasr_loss_lexicon_path,
"datasets": returnn_datasets,
}
feature_args = {"class": "LogMelNetwork", "wavenorm": True, "frame_size": 200, "frame_shift": 80, "fft_size": 256}
returnn_datasets_laplace25 = copy.deepcopy(returnn_datasets)
returnn_datasets_laplace25["train"]["seq_ordering"] = "laplace:.25"
nn_args, report_args_collection = get_nn_args_baseline(
nn_base_args={
# "lgm80_conf-simon": dict(
# returnn_args={"conformer_type": "simon", **returnn_args},
# feature_args=feature_args,
# ),
"lgm80_conf-wei_old-lr": dict(
returnn_args={"conformer_type": "wei", **returnn_args},
feature_args=feature_args,
report_args={"architecture": "conf-wei", "lr": "default"},
),
"lgm80_conf-wei_old-lr-4e-4": dict(
returnn_args={"conformer_type": "wei", **returnn_args},
feature_args=feature_args,
lr_args={"peak_lr": 4e-4},
report_args={"architecture": "conf-wei", "lr": "default_peak_4e-4"},
),
"lgm80_conf-wei": dict( # matches original lr schedule from wei
returnn_args={"conformer_type": "wei", **returnn_args},
feature_args=feature_args,
lr_args={
"peak_lr": 4e-4,
"start_lr": 1.325e-05,
"end_lr": 1e-5,
"increase_epochs": 119,
"peak_epochs": 1,
"decrease_epochs": 120,
"final_epochs": 0,
},
report_args={"architecture": "conf-wei", "lr": "wei_peak_4e-4"},
),
"lgm80_conf-wei2": dict( # almost matches original lr schedule from wei
returnn_args={"conformer_type": "wei", **returnn_args},
feature_args=feature_args,
lr_args={
"peak_lr": 4e-4,
"start_lr": 1.325e-05,
"end_lr": 1e-5,
"increase_epochs": 119,
"peak_epochs": 2,
"decrease_epochs": 119,
"final_epochs": 0,
},
report_args={"architecture": "conf-wei", "lr": "wei_peak_4e-4"},
),
"lgm80_conf-wei-oldspecaug": dict( # specaugment as in wei's setup
returnn_args={"conformer_type": "wei", "specaug_old": {}, **returnn_args},
feature_args=feature_args,
lr_args={
"peak_lr": 4e-4,
"start_lr": 1.325e-05,
"end_lr": 1e-5,
"increase_epochs": 119,
"peak_epochs": 2,
"decrease_epochs": 119,
"final_epochs": 0,
},
report_args={"architecture": "conf-wei", "lr": "wei_peak_4e-4", "specaug": "wei"},
),
"lgm80_conf-wei-oldspecaug2": dict( # specaugment as in wei's setup but double feature dim due to log Mel
returnn_args={"conformer_type": "wei", "specaug_old": {"max_feature": 8}, **returnn_args},
feature_args=feature_args,
lr_args={
"peak_lr": 4e-4,
"start_lr": 1.325e-05,
"end_lr": 1e-5,
"increase_epochs": 119,
"peak_epochs": 2,
"decrease_epochs": 119,
"final_epochs": 0,
},
report_args={"architecture": "conf-wei", "lr": "wei_peak_4e-4", "specaug": "wei_adapt_80dim"},
),
"lgm80_conf-wei-oldspecaug-bs3200step": dict(
returnn_args={
"conformer_type": "wei",
"specaug_old": {},
**returnn_args,
"batch_size": {"data": 514635},
},
feature_args=feature_args,
lr_args={
"peak_lr": 4e-4,
"start_lr": 1.325e-05,
"end_lr": 1e-5,
"increase_epochs": 119,
"peak_epochs": 2,
"decrease_epochs": 119,
"final_epochs": 0,
},
report_args={"architecture": "conf-wei", "lr": "wei_peak_4e-4", "specaug": "wei"},
),
"lgm80_conf-wei-oldspecaug2-bs3200step": dict(
returnn_args={
"conformer_type": "wei",
"specaug_old": {"max_feature": 8},
**returnn_args,
"batch_size": {"data": 514635},
},
feature_args=feature_args,
lr_args={
"peak_lr": 4e-4,
"start_lr": 1.325e-05,
"end_lr": 1e-5,
"increase_epochs": 119,
"peak_epochs": 2,
"decrease_epochs": 119,
"final_epochs": 0,
},
report_args={"architecture": "conf-wei", "lr": "wei_peak_4e-4", "specaug": "wei_adapt_80dim"},
),
"lgm80_conf-wei-oldspecaug-laplace25": dict(
returnn_args={
"conformer_type": "wei",
"specaug_old": {},
**returnn_args,
"datasets": returnn_datasets_laplace25,
},
feature_args=feature_args,
lr_args={
"peak_lr": 4e-4,
"start_lr": 1.325e-05,
"end_lr": 1e-5,
"increase_epochs": 119,
"peak_epochs": 2,
"decrease_epochs": 119,
"final_epochs": 0,
},
report_args={"architecture": "conf-wei", "lr": "wei_peak_4e-4", "specaug": "wei"},
),
"lgm80_conf-wei-oldspecaug2-lrv1": dict(
returnn_args={"conformer_type": "wei", "specaug_old": {"max_feature": 8}, **returnn_args},
feature_args=feature_args,
lr_args={
"peak_lr": 2 * 4e-4,
"start_lr": 2 * 1.325e-05,
"end_lr": 2 * 1e-5,
"increase_epochs": 119,
"peak_epochs": 2,
"decrease_epochs": 119,
"final_epochs": 0,
},
report_args={"architecture": "conf-wei", "lr": "wei_peak_8e-4", "specaug": "wei_adapt_80dim"},
),
# "lgm80_conf-wei2-nadam": dict( # does not work well
# returnn_args={
# "conformer_type": "wei",
# "extra_args": {"optimizer": None, "optimizer_epsilon": 1e-8, "nadam": True},
# **returnn_args,
# },
# feature_args=feature_args,
# lr_args={
# "peak_lr": 4e-4, "start_lr": 1.325e-05, "end_lr": 1e-5,
# "increase_epochs": 119, "peak_epochs": 2, "decrease_epochs": 119, "final_epochs": 0,
# },
# ),
# "gt40_pe_conf-wei_old-lr": dict( # does not converge
# returnn_args={"conformer_type": "wei", **returnn_args},
# feature_args={
# "class": "GammatoneNetwork", "sample_rate": 8000, "freq_max": 3800., "output_dim": 40,
# "preemphasis": 1.0,
# },
# ),
# "acc2_scf750_conf-wei_old-lr": dict( # does not converge
# returnn_args={
# "conformer_type": "wei",
# **returnn_args,
# "batch_size": 5000,
# "extra_args": {"accum_grad_multiple_step": 2},
# },
# feature_args={"class": "ScfNetwork", "size_tf": 256 // 2, "stride_tf": 10 // 2},
# ),
},
num_epochs=300,
prefix="conformer_bs10k_",
)
returnn_configs = {}
for exp in nn_args.returnn_training_configs:
prior_config = copy.deepcopy(nn_args.returnn_training_configs[exp])
prior_config.config["batch_size"] = prior_config.config["batch_size"]["data"]
assert isinstance(prior_config.config["batch_size"], int)
returnn_configs[exp] = ReturnnConfigs(
train_config=nn_args.returnn_training_configs[exp],
prior_config=prior_config,
recog_configs={"recog": nn_args.returnn_recognition_configs[exp]},
)
recog_args = {
"lm_scales": [0.7],
"prior_scales": [0.3, 0.5],
"epochs": [300],
"lookahead_options": {"lm_lookahead_scale": 0.7},
"label_scorer_args": {
"use_prior": True,
"extra_args": {"blank_label_index": 0},
},
"label_tree_args": {"skip_silence": True},
"search_parameters": {
"allow-blank-label": True,
"allow-label-loop": True,
"allow-label-recombination": True,
"allow-word-end-recombination": True,
"create-lattice": True,
"label-pruning": 11.2,
"label-pruning-limit": 100000,
"word-end-pruning": 0.5,
"word-end-pruning-limit": 10000,
},
}
score_info = ScorerInfo()
score_info.ref_file = dev_corpora["hub5e00"].stm
score_info.job_type = Hub5ScoreJob
score_info.score_kwargs = {"glm": dev_corpora["hub5e00"].glm, "sctk_binary_path": SCTK_BINARY_PATH}
ctc_nn_system = TransducerSystem(
returnn_root=RETURNN_ROOT,
returnn_python_exe=RETURNN_EXE,
rasr_binary_path=RASR_BINARY_PATH,
require_native_lstm=False,
)
ctc_nn_system.init_system(
returnn_configs=returnn_configs,
dev_keys=["hub5e00"],
corpus_data=dev_corpora,
am_args={
"state_tying": "monophone",
"states_per_phone": 1,
"tdp_transition": (0, 0, 0, 0),
"tdp_silence": (0, 0, 0, 0),
"phon_history_length": 0,
"phon_future_length": 0,
},
scorer_info=score_info,
report=Report(
columns_start=["train_name"],
columns_end=["lm_scale", "prior_scale", "sub", "del", "ins", "wer"],
),
)
ctc_nn_system.crp["hub5e00"].acoustic_model_config.allophones.add_from_lexicon = False
ctc_nn_system.crp["hub5e00"].acoustic_model_config.allophones.add_all = True
ctc_nn_system.crp["hub5e00"].acoustic_model_config.allophones.add_from_file = tk.Path(
"/u/vieting/setups/swb/20230406_feat/dependencies/allophones_blank",
hash_overwrite="SWB_ALLOPHONE_FILE_WEI_BLANK",
cached=True,
)
ctc_nn_system.run_train_step(nn_args.training_args)
ctc_nn_system.run_dev_recog_step(recog_args=recog_args, report_args=report_args_collection)
ctc_nn_system.run_recogs_for_corpora( # test for single model with larger prior_scales
["hub5e00"],
"conformer_bs10k_lgm80_conf-wei_old-lr-4e-4",
search_type=SearchTypes.GenericSeq2SeqSearchJob,
report_args=report_args_collection,
**{**recog_args, "prior_scales": [0.4, 0.6, 0.7, 0.9]},
)
ctc_nn_system.run_recogs_for_corpora( # test for single model with larger prior_scales
["hub5e00"],
"conformer_bs10k_lgm80_conf-wei_old-lr",
search_type=SearchTypes.GenericSeq2SeqSearchJob,
report_args=report_args_collection,
**{**recog_args, "prior_scales": [0.5], "epochs": [260]},
)
# test blank penalty as we have more deletions than insertions
ctc_nn_system_blank_penalty = copy.deepcopy(ctc_nn_system)
exp_name = "conformer_bs10k_lgm80_conf-wei_old-lr-4e-4"
recog_config = ctc_nn_system_blank_penalty.returnn_configs[exp_name].recog_configs.pop("recog")
for blank_penalty in [0.2, 0.3, 0.5, 0.8]:
config = copy.deepcopy(recog_config)
blank_index = 0
num_outputs = 88
config.config["network"]["output_blank_penalty"] = {
"class": "eval",
"from": "output",
"is_output_layer": True,
"eval": f"source(0) - tf.expand_dims("
f"tf.one_hot([{blank_index}], {num_outputs}, on_value={blank_penalty}, dtype=tf.float32), axis=0)",
}
ctc_nn_system_blank_penalty.returnn_configs[exp_name].recog_configs[
f"recog_blank-penalty-{blank_penalty}"
] = config
ctc_nn_system_blank_penalty.run_recogs_for_corpora(
["hub5e00"],
exp_name,
search_type=SearchTypes.GenericSeq2SeqSearchJob,
report_args=report_args_collection,
extra_name=f"_blank-pen-{blank_penalty}",
tf_flow_args={"output_layer_name": "output_blank_penalty"},
**{**recog_args, "prior_scales": [0.5]},
)
# same lm as in wei's setup, results indicate that this is not better (if any, slightly worse)
ctc_nn_system_wei_lm = copy.deepcopy(ctc_nn_system)
for train_name in list(ctc_nn_system_wei_lm.returnn_configs.keys()):
if train_name not in ["conformer_bs10k_lgm80_conf-wei2", "conformer_bs10k_lgm80_conf-wei_old-lr-4e-4"]:
# only use some trainings
ctc_nn_system_wei_lm.returnn_configs.pop(train_name)
report_args_wei_lm = copy.deepcopy(report_args_collection)
for name in report_args_wei_lm:
report_args_wei_lm[name]["lm"] = "wei"
wei_lm = tk.Path(
"/u/vieting/setups/swb/20230406_feat/dependencies/zoltan_4gram.gz",
hash_overwrite="ZOLTAN_SWB_LM_4GRAM",
cached=True,
)
ctc_nn_system_wei_lm.corpus_data["hub5e00"].lm["filename"] = wei_lm
ctc_nn_system_wei_lm.crp["hub5e00"].language_model_config.file = wei_lm
ctc_nn_system_wei_lm.run_dev_recog_step(recog_args=recog_args, extra_name="_lm-wei", report_args=report_args_wei_lm)
# same lexicon as in wei's setup, results indicate that this is not better (less del, but overall slightly worse)
ctc_nn_system_wei_lex = copy.deepcopy(ctc_nn_system)
for train_name in list(ctc_nn_system_wei_lex.returnn_configs.keys()):
if train_name not in ["conformer_bs10k_lgm80_conf-wei2", "conformer_bs10k_lgm80_conf-wei_old-lr-4e-4"]:
# only use some trainings
ctc_nn_system_wei_lex.returnn_configs.pop(train_name)
report_args_wei_lex = copy.deepcopy(report_args_collection)
for name in report_args_wei_lex:
report_args_wei_lex[name]["lex"] = "wei"
wei_lex = tk.Path(
"/u/vieting/setups/swb/20230406_feat/dependencies/lexicon_wei_blank.xml",
hash_overwrite="WEI_SWB_LEX",
cached=True,
)
ctc_nn_system_wei_lex.corpus_data["hub5e00"].lexicon["filename"] = wei_lex
ctc_nn_system_wei_lex.run_dev_recog_step(
recog_args=recog_args, extra_name="_lex-wei", report_args=report_args_wei_lex
)
# longer training to compensate for fewer steps per epoch
feature_args_wave_norm = {
"class": "LogMelNetwork",
"wave_norm": True,
"frame_size": 200,
"frame_shift": 80,
"fft_size": 256,
}
nn_args, report_args_collection = get_nn_args_baseline(
nn_base_args={
"lgm80_conf-wei-oldspecaug-e450v1": dict(
returnn_args={"conformer_type": "wei", "specaug_old": {}, **returnn_args},
feature_args=feature_args,
lr_args={
"peak_lr": 4e-4,
"start_lr": 1.325e-05,
"end_lr": 1e-5,
"increase_epochs": 180,
"decrease_epochs": 180,
"final_epochs": 0,
},
report_args={"architecture": "conf-wei", "lr": "wei_peak_4e-4_e450_cycle360", "specaug": "wei"},
),
"lgm80_conf-wei-oldspecaug2-e450v1": dict(
returnn_args={"conformer_type": "wei", "specaug_old": {"max_feature": 8}, **returnn_args},
feature_args=feature_args,
lr_args={
"peak_lr": 4e-4,
"start_lr": 1.325e-05,
"end_lr": 1e-5,
"increase_epochs": 180,
"decrease_epochs": 180,
"final_epochs": 0,
},
report_args={
"architecture": "conf-wei",
"lr": "wei_peak_4e-4_e450_cycle360",
"specaug": "wei_adapt_80dim",
},
),
"lgm80_conf-wei-oldspecaug2-e450v2": dict(
returnn_args={"conformer_type": "wei", "specaug_old": {"max_feature": 8}, **returnn_args},
feature_args=feature_args,
lr_args={
"peak_lr": 4e-4,
"start_lr": 1.325e-05,
"end_lr": 1e-5,
"increase_epochs": 160,
"decrease_epochs": 160,
"final_epochs": 0,
},
report_args={
"architecture": "conf-wei",
"lr": "wei_peak_4e-4_e450_320cycle",
"specaug": "wei_adapt_80dim",
},
),
"lgm80_conf-wei-oldspecaug2-e450v3": dict(
returnn_args={"conformer_type": "wei", "specaug_old": {"max_feature": 8}, **returnn_args},
feature_args=feature_args,
lr_args={
"peak_lr": 4e-4,
"start_lr": 1.325e-05,
"end_lr": 1e-5,
"increase_epochs": 200,
"decrease_epochs": 200,
"final_epochs": 0,
},
report_args={
"architecture": "conf-wei",
"lr": "wei_peak_4e-4_e450_400_cycle",
"specaug": "wei_adapt_80dim",
},
),
"lgm80_conf-wei-oldspecaug2-e450v1-wavenorm": dict(
returnn_args={"conformer_type": "wei", "specaug_old": {"max_feature": 8}, **returnn_args},
feature_args=feature_args_wave_norm,
lr_args={
"peak_lr": 4e-4,
"start_lr": 1.325e-05,
"end_lr": 1e-5,
"increase_epochs": 180,
"decrease_epochs": 180,
"final_epochs": 0,
},
report_args={
"architecture": "conf-wei",
"lr": "wei_peak_4e-4_e450_cycle360",
"specaug": "wei_adapt_80dim",
"wave_norm": "True",
},
),
},
num_epochs=450,
prefix="conformer_bs10k_",
)
returnn_configs = {}
for exp in nn_args.returnn_training_configs:
prior_config = copy.deepcopy(nn_args.returnn_training_configs[exp])
prior_config.config["batch_size"] = prior_config.config["batch_size"]["data"]
assert isinstance(prior_config.config["batch_size"], int)
returnn_configs[exp] = ReturnnConfigs(
train_config=nn_args.returnn_training_configs[exp],
prior_config=prior_config,
recog_configs={"recog": nn_args.returnn_recognition_configs[exp]},
)
recog_args_e450 = copy.deepcopy(recog_args)
recog_args_e450["epochs"] = [300, 400, 450, "best"]
ctc_nn_system_e450 = copy.deepcopy(ctc_nn_system)
ctc_nn_system_e450.returnn_configs = returnn_configs
ctc_nn_system_e450.run_train_step(nn_args.training_args)
ctc_nn_system_e450.run_dev_recog_step(recog_args=recog_args_e450, report_args=report_args_collection)
report = Report.merge_reports(
[
ctc_nn_system.report,
ctc_nn_system_e450.report,
ctc_nn_system_blank_penalty.report,
ctc_nn_system_wei_lm.report,
ctc_nn_system_wei_lex.report,
]
)
report.delete_redundant_columns()
report.delete_redundant_rows()
tk.register_report(
os.path.join(gs.ALIAS_AND_OUTPUT_SUBDIR, "report.csv"),
values=report.get_values(),
template=report.get_template(),
)
def run_nn_args(nn_args, report_args_collection, report_name, dev_corpora):
returnn_configs = {}
for exp in nn_args.returnn_training_configs:
prior_config = copy.deepcopy(nn_args.returnn_training_configs[exp])
prior_config.config["batch_size"] = prior_config.config["batch_size"]["data"]
assert isinstance(prior_config.config["batch_size"], int)
returnn_configs[exp] = ReturnnConfigs(
train_config=nn_args.returnn_training_configs[exp],
prior_config=prior_config,
recog_configs={"recog": nn_args.returnn_recognition_configs[exp]},
)
recog_args = {
"lm_scales": [0.7],
"prior_scales": [0.3, 0.5],
"epochs": [300, 400, 450, "best"],
"lookahead_options": {"lm_lookahead_scale": 0.7},
"label_scorer_args": {
"use_prior": True,
"extra_args": {"blank_label_index": 0},
},
"label_tree_args": {"skip_silence": True},
"search_parameters": {
"allow-blank-label": True,
"allow-label-loop": True,
"allow-label-recombination": True,
"allow-word-end-recombination": True,
"create-lattice": True,
"label-pruning": 11.2,
"label-pruning-limit": 100000,
"word-end-pruning": 0.5,
"word-end-pruning-limit": 10000,
},
}
score_info = ScorerInfo()
score_info.ref_file = dev_corpora["hub5e00"].stm
score_info.job_type = Hub5ScoreJob
score_info.score_kwargs = {"glm": dev_corpora["hub5e00"].glm, "sctk_binary_path": SCTK_BINARY_PATH}
ctc_nn_system = TransducerSystem(
returnn_root=RETURNN_ROOT,
returnn_python_exe=RETURNN_EXE,
rasr_binary_path=RASR_BINARY_PATH,
require_native_lstm=False,
)
ctc_nn_system.init_system(
returnn_configs=returnn_configs,
dev_keys=["hub5e00"],
corpus_data=dev_corpora,
am_args={
"state_tying": "monophone",
"states_per_phone": 1,
"tdp_transition": (0, 0, 0, 0),
"tdp_silence": (0, 0, 0, 0),
"phon_history_length": 0,
"phon_future_length": 0,
},
scorer_info=score_info,
report=Report(
columns_start=["train_name"],
columns_end=["lm_scale", "prior_scale", "sub", "del", "ins", "wer"],
),
)
ctc_nn_system.crp["hub5e00"].acoustic_model_config.allophones.add_from_lexicon = False
ctc_nn_system.crp["hub5e00"].acoustic_model_config.allophones.add_all = True
ctc_nn_system.crp["hub5e00"].acoustic_model_config.allophones.add_from_file = tk.Path(
"/u/vieting/setups/swb/20230406_feat/dependencies/allophones_blank",
hash_overwrite="SWB_ALLOPHONE_FILE_WEI_BLANK",
cached=True,
)
ctc_nn_system.run_train_step(nn_args.training_args)
ctc_nn_system.run_dev_recog_step(recog_args=recog_args, report_args=report_args_collection)
assert ctc_nn_system.report is not None
report = ctc_nn_system.report
report.delete_redundant_columns()
report.delete_redundant_rows()
tk.register_report(
os.path.join(gs.ALIAS_AND_OUTPUT_SUBDIR, report_name),
values=report.get_values(),
template=report.get_template(),
)
def run_mel_baseline():
gs.ALIAS_AND_OUTPUT_SUBDIR = "experiments/switchboard/ctc/feat/"
(
returnn_datasets,
rasr_loss_corpus_path,
rasr_loss_corpus_segments,
rasr_loss_lexicon_path,
dev_corpora,
) = get_datasets()
returnn_args = {
"batch_size": 5000,
"rasr_binary_path": RASR_BINARY_PATH,
"rasr_loss_corpus_path": rasr_loss_corpus_path,
"rasr_loss_corpus_segments": rasr_loss_corpus_segments,
"rasr_loss_lexicon_path": rasr_loss_lexicon_path,
"datasets": returnn_datasets,
}
feature_args = {"class": "LogMelNetwork", "wave_norm": True, "frame_size": 200, "frame_shift": 80, "fft_size": 256}
nn_args, report_args_collection = get_nn_args_baseline(
nn_base_args={
"lgm80_baseline": dict(
returnn_args={"conformer_type": "wei", "specaug_old": {"max_feature": 8}, **returnn_args},
feature_args=feature_args,
lr_args={
"peak_lr": 4e-4,
"start_lr": 1.325e-05,
"end_lr": 1e-5,
"increase_epochs": 180,
"decrease_epochs": 180,
"final_epochs": 0,
},
report_args={
"architecture": "conf-wei",
"lr": "wei_peak_4e-4_e450_cycle360",
"specaug": "wei_adapt_80dim",
"wave_norm": "True",
},
),
},
num_epochs=450,
prefix="conformer_bs5k_",
)
run_nn_args(nn_args, report_args_collection, "report_mel_5K_baseline.csv", dev_corpora)
def run_scf_baseline():
gs.ALIAS_AND_OUTPUT_SUBDIR = "experiments/switchboard/ctc/feat/"
(
returnn_datasets,
rasr_loss_corpus_path,
rasr_loss_corpus_segments,
rasr_loss_lexicon_path,
dev_corpora,
) = get_datasets()
returnn_args = {
"batch_size": 5000,
"rasr_binary_path": RASR_BINARY_PATH,
"rasr_loss_corpus_path": rasr_loss_corpus_path,
"rasr_loss_corpus_segments": rasr_loss_corpus_segments,
"rasr_loss_lexicon_path": rasr_loss_lexicon_path,
"datasets": returnn_datasets,
}
feature_args = {"class": "ScfNetwork", "size_tf": 256 // 2, "stride_tf": 10 // 2}
nn_args, report_args_collection = get_nn_args_baseline(
nn_base_args={
"scf_baseline": dict(
returnn_args={"conformer_type": "wei", "specaug_old": {"max_feature": 8}, **returnn_args},
feature_args=feature_args,
lr_args={
"peak_lr": 4e-4,
"start_lr": 1.325e-05,
"end_lr": 1e-5,
"increase_epochs": 180,
"decrease_epochs": 180,
"final_epochs": 0,
},
report_args={
"architecture": "conf-wei",
"lr": "wei_peak_4e-4_e450_cycle360",
"specaug": "wei_adapt_80dim",
"wave_norm": "True",
},
),
},
num_epochs=450,
prefix="conformer_bs5k_",
)
run_nn_args(nn_args, report_args_collection, "report_scf_baseline.csv", dev_corpora)
def run_scf_baseline_big():
gs.ALIAS_AND_OUTPUT_SUBDIR = "experiments/switchboard/ctc/feat/"
(
returnn_datasets,
rasr_loss_corpus_path,
rasr_loss_corpus_segments,
rasr_loss_lexicon_path,
dev_corpora,
) = get_datasets()
returnn_args = {
"batch_size": 5000,
"rasr_binary_path": RASR_BINARY_PATH,
"rasr_loss_corpus_path": rasr_loss_corpus_path,
"rasr_loss_corpus_segments": rasr_loss_corpus_segments,
"rasr_loss_lexicon_path": rasr_loss_lexicon_path,
"datasets": returnn_datasets,
"extra_args": {
"watch_memory": True,
},
}
feature_args = {"class": "ScfNetwork", "size_tf": 256 // 2, "stride_tf": 10 // 2}
nn_args, report_args_collection = get_nn_args_baseline(
nn_base_args={
"scf_baseline_big": dict(
returnn_args={"conformer_type": "wei", "specaug_old": {"max_feature": 8}, **returnn_args},
feature_args=feature_args,
lr_args={
"peak_lr": 4e-4,
"start_lr": 1.325e-05,
"end_lr": 1e-5,
"increase_epochs": 180,
"decrease_epochs": 180,
"final_epochs": 0,
},
report_args={
"architecture": "conf-wei",
"lr": "wei_peak_4e-4_e450_cycle360",
"specaug": "wei_adapt_80dim",
"wave_norm": "True",
},
),
},
num_epochs=450,
prefix="conformer_bs5k_",
)
run_nn_args(nn_args, report_args_collection, "report_scf_baseline_big.csv", dev_corpora)
def run_mel_audio_perturbation():
gs.ALIAS_AND_OUTPUT_SUBDIR = "experiments/switchboard/ctc/feat/"
(
returnn_datasets,
rasr_loss_corpus_path,
rasr_loss_corpus_segments,
rasr_loss_lexicon_path,
dev_corpora,
) = get_datasets(pre_process=CodeWrapper("audio_perturb_runner.run"))
returnn_args = {
"batch_size": 5000,
"rasr_binary_path": RASR_BINARY_PATH,
"rasr_loss_corpus_path": rasr_loss_corpus_path,
"rasr_loss_corpus_segments": rasr_loss_corpus_segments,
"rasr_loss_lexicon_path": rasr_loss_lexicon_path,
"datasets": returnn_datasets,
"audio_perturbation": True,
}
nn_base_args = {}
feature_args = {"class": "LogMelNetwork", "wave_norm": True, "frame_size": 200, "frame_shift": 80, "fft_size": 256}
speeds = [
{"prob": 0.6, "minimum": 0.9, "maximum": 1.1},
{"prob": 0.6, "minimum": 0.8, "maximum": 1.2},
{"prob": 0.6, "minimum": 0.7, "maximum": 1.3},
{"prob": 0.5, "minimum": 0.9, "maximum": 1.1},
{"prob": 0.5, "minimum": 0.8, "maximum": 1.2},
{"prob": 0.5, "minimum": 0.7, "maximum": 1.3},
{"prob": 0.4, "minimum": 0.9, "maximum": 1.1},
{"prob": 0.4, "minimum": 0.8, "maximum": 1.2},
{"prob": 0.4, "minimum": 0.7, "maximum": 1.3},
]
tempos = [
{"prob": 0.4, "minimum": 0.9, "maximum": 1.1},
{"prob": 0.4, "minimum": 0.8, "maximum": 1.2},
{"prob": 0.4, "minimum": 0.7, "maximum": 1.3},
{"prob": 0.5, "minimum": 0.9, "maximum": 1.1},
{"prob": 0.5, "minimum": 0.8, "maximum": 1.2},
{"prob": 0.5, "minimum": 0.7, "maximum": 1.3},
{"prob": 0.6, "minimum": 0.9, "maximum": 1.1},
{"prob": 0.6, "minimum": 0.8, "maximum": 1.2},
{"prob": 0.6, "minimum": 0.7, "maximum": 1.3},
]
preemphases = [
{"prob": 0.9, "minimum": 0.9, "maximum": 1.0},
{"prob": 0.9, "minimum": 0.8, "maximum": 1.0},
{"prob": 0.8, "minimum": 0.9, "maximum": 1.0},
{"prob": 0.8, "minimum": 0.8, "maximum": 1.0},
]
codecs = [
{"encoding": "ULAW", "prob": 0.4},
{"encoding": "ULAW", "prob": 0.6},
]
non_linearities = [
{"prob": 0.4, "alpha": 0.2},
{"prob": 0.4, "alpha": 0.4},
{"prob": 0.6, "alpha": 0.2},
{"prob": 0.6, "alpha": 0.4},
]
for speed in speeds:
key_suffix = []
#build key suffix
key_suffix.append("speed"+ f"{speed['prob']}_{speed['minimum']}_{speed['maximum']}")
key = f"log_mel_conf-wei-oldspecaug-audio_perturbation_{'_'.join(key_suffix)}"
audio_perturb_args = {}
audio_perturb_args["speed"] = speed
report_args={
"architecture": "conf-wei",
"lr": "wei_peak_4e-4_e450_cycle360",
"specaug": "wei_adapt_80dim",
"wave_norm": "True",
}
report_args["speed"] = f"{speed['prob']}_{speed['minimum']}_{speed['maximum']}"
nn_base_args[key] = dict(
returnn_args={
"extra_args": {
"audio_perturb_args": audio_perturb_args,
"audio_perturb_runner": CodeWrapper("WaveformPerturbation(**audio_perturb_args)"),
},
**returnn_args,},
feature_args=feature_args,
lr_args={
"peak_lr": 4e-4,
"start_lr": 1.325e-05,
"end_lr": 1e-5,
"increase_epochs": 180,
"decrease_epochs": 180,
"final_epochs": 0,
},
report_args=report_args,
)
for tempo in tempos:
key_suffix = []
key_suffix.append("tempo"+ f"{tempo['prob']}_{tempo['minimum']}_{tempo['maximum']}")
key = f"log_mel_conf-wei-oldspecaug-audio_perturbation_{'_'.join(key_suffix)}"
audio_perturb_args = {}
audio_perturb_args["tempo"] = tempo
report_args={
"architecture": "conf-wei",
"lr": "wei_peak_4e-4_e450_cycle360",
"specaug": "wei_adapt_80dim",
"wave_norm": "True",
}
report_args["tempo"] = f"{tempo['prob']}_{tempo['minimum']}_{tempo['maximum']}"
nn_base_args[key] = dict(
returnn_args={
"extra_args": {
"audio_perturb_args": audio_perturb_args,
"audio_perturb_runner": CodeWrapper("WaveformPerturbation(**audio_perturb_args)"),
},
**returnn_args,},
feature_args=feature_args,
lr_args={
"peak_lr": 4e-4,
"start_lr": 1.325e-05,
"end_lr": 1e-5,
"increase_epochs": 180,
"decrease_epochs": 180,
"final_epochs": 0,
},
report_args=report_args,
)
for preemphasis in preemphases:
key_suffix = []
key_suffix.append("preemphasis"+ f"{preemphasis['prob']}_{preemphasis['minimum']}_{preemphasis['maximum']}")
key = f"log_mel_conf-wei-oldspecaug-audio_perturbation_{'_'.join(key_suffix)}"
audio_perturb_args = {}
audio_perturb_args["preemphasis"] = preemphasis
report_args={
"architecture": "conf-wei",
"lr": "wei_peak_4e-4_e450_cycle360",
"specaug": "wei_adapt_80dim",
"wave_norm": "True",
}
report_args[
"preemphasis"
] = f"{preemphasis['prob']}_{preemphasis['minimum']}_{preemphasis['maximum']}"
nn_base_args[key] = dict(
returnn_args={
"extra_args": {
"audio_perturb_args": audio_perturb_args,
"audio_perturb_runner": CodeWrapper("WaveformPerturbation(**audio_perturb_args)"),
},
**returnn_args,},
feature_args=feature_args,
lr_args={
"peak_lr": 4e-4,
"start_lr": 1.325e-05,
"end_lr": 1e-5,
"increase_epochs": 180,
"decrease_epochs": 180,
"final_epochs": 0,
},
report_args=report_args,
)
for codec in codecs:
key_suffix = []
key_suffix.append("codec" + f"wav_{codec['encoding'].lower()}_{codec['prob']}")
key = f"log_mel_conf-wei-oldspecaug-audio_perturbation_{'_'.join(key_suffix)}"
audio_perturb_args = {}
audio_perturb_args["codecs"] = [codec]
report_args={
"architecture": "conf-wei",
"lr": "wei_peak_4e-4_e450_cycle360",
"specaug": "wei_adapt_80dim",
"wave_norm": "True",
}
report_args["codec"] = f"wav_{codec['encoding'].lower()}_{codec['prob']}"
nn_base_args[key] = dict(
returnn_args={
"extra_args": {
"audio_perturb_args": audio_perturb_args,
"audio_perturb_runner": CodeWrapper("WaveformPerturbation(**audio_perturb_args)"),
},
**returnn_args,},
feature_args=feature_args,
lr_args={
"peak_lr": 4e-4,
"start_lr": 1.325e-05,
"end_lr": 1e-5,
"increase_epochs": 180,
"decrease_epochs": 180,
"final_epochs": 0,
},
report_args=report_args,
)
for non_linearity in non_linearities:
key_suffix = []
key_suffix.append("non_linearity" + f"{non_linearity['prob']}_{non_linearity['alpha']}")
key = f"log_mel_conf-wei-oldspecaug-audio_perturbation_{'_'.join(key_suffix)}"
audio_perturb_args = {}
audio_perturb_args["non_linearities"] = [non_linearity]
report_args={
"architecture": "conf-wei",
"lr": "wei_peak_4e-4_e450_cycle360",
"specaug": "wei_adapt_80dim",
"wave_norm": "True",
}
report_args["non_linearity"] = f"{non_linearity['prob']}_{non_linearity['alpha']}"
nn_base_args[key] = dict(
returnn_args={
"extra_args": {
"audio_perturb_args": audio_perturb_args,
"audio_perturb_runner": CodeWrapper("WaveformPerturbation(**audio_perturb_args)"),
},
**returnn_args,},
feature_args=feature_args,
lr_args={
"peak_lr": 4e-4,
"start_lr": 1.325e-05,
"end_lr": 1e-5,
"increase_epochs": 180,
"decrease_epochs": 180,
"final_epochs": 0,
},
report_args=report_args,
)
nn_args, report_args_collection = get_nn_args_baseline(
nn_base_args=nn_base_args,
num_epochs=450,
prefix="conformer_bs5k_audio_perturbation_",
)
run_nn_args(nn_args, report_args_collection, "report_log_mel_audio_perturbation.csv", dev_corpora)
def run_scf_audio_perturbation():
gs.ALIAS_AND_OUTPUT_SUBDIR = "experiments/switchboard/ctc/feat/"
(
returnn_datasets,
rasr_loss_corpus_path,
rasr_loss_corpus_segments,
rasr_loss_lexicon_path,
dev_corpora,
) = get_datasets(pre_process=CodeWrapper("audio_perturb_runner.run"))
returnn_args = {
"batch_size": 5000,
"rasr_binary_path": RASR_BINARY_PATH,
"rasr_loss_corpus_path": rasr_loss_corpus_path,
"rasr_loss_corpus_segments": rasr_loss_corpus_segments,
"rasr_loss_lexicon_path": rasr_loss_lexicon_path,
"datasets": returnn_datasets,
"audio_perturbation": True,
}
nn_base_args = {}
feature_args = {"class": "ScfNetwork", "size_tf": 256 // 2, "stride_tf": 10 // 2}
speeds = [
{"prob": 0.6, "minimum": 0.9, "maximum": 1.1},
{"prob": 0.6, "minimum": 0.8, "maximum": 1.2},
{"prob": 0.6, "minimum": 0.7, "maximum": 1.3},
{"prob": 0.5, "minimum": 0.9, "maximum": 1.1},
{"prob": 0.5, "minimum": 0.8, "maximum": 1.2},
{"prob": 0.5, "minimum": 0.7, "maximum": 1.3},
{"prob": 0.4, "minimum": 0.9, "maximum": 1.1},
{"prob": 0.4, "minimum": 0.8, "maximum": 1.2},
{"prob": 0.4, "minimum": 0.7, "maximum": 1.3},
]
tempos = [
{"prob": 0.4, "minimum": 0.9, "maximum": 1.1},
{"prob": 0.4, "minimum": 0.8, "maximum": 1.2},
{"prob": 0.4, "minimum": 0.7, "maximum": 1.3},
{"prob": 0.5, "minimum": 0.9, "maximum": 1.1},
{"prob": 0.5, "minimum": 0.8, "maximum": 1.2},
{"prob": 0.5, "minimum": 0.7, "maximum": 1.3},
{"prob": 0.6, "minimum": 0.9, "maximum": 1.1},
{"prob": 0.6, "minimum": 0.8, "maximum": 1.2},
{"prob": 0.6, "minimum": 0.7, "maximum": 1.3},
]
preemphases = [
{"prob": 0.9, "minimum": 0.9, "maximum": 1.0},
{"prob": 0.9, "minimum": 0.8, "maximum": 1.0},
{"prob": 0.8, "minimum": 0.9, "maximum": 1.0},
{"prob": 0.8, "minimum": 0.8, "maximum": 1.0},
]
codecs = [
{"encoding": "ULAW", "prob": 0.4},
{"encoding": "ULAW", "prob": 0.6},
]
non_linearities = [
{"prob": 0.4, "alpha": 0.2},
{"prob": 0.4, "alpha": 0.4},
{"prob": 0.6, "alpha": 0.2},
{"prob": 0.6, "alpha": 0.4},
]
for speed in speeds:
key_suffix = []
#build key suffix
key_suffix.append("speed"+ f"{speed['prob']}_{speed['minimum']}_{speed['maximum']}")
key = f"scf_conf-wei-oldspecaug-audio_perturbation_{'_'.join(key_suffix)}"
audio_perturb_args = {}
audio_perturb_args["speed"] = speed
report_args={
"architecture": "conf-wei",
"lr": "wei_peak_4e-4_e450_cycle360",
"specaug": "wei_adapt_80dim",
"wave_norm": "True",
}
report_args["speed"] = f"{speed['prob']}_{speed['minimum']}_{speed['maximum']}"
nn_base_args[key] = dict(
returnn_args={
"extra_args": {
"audio_perturb_args": audio_perturb_args,
"audio_perturb_runner": CodeWrapper("WaveformPerturbation(**audio_perturb_args)"),
},
**returnn_args,},
feature_args=feature_args,
lr_args={
"peak_lr": 4e-4,
"start_lr": 1.325e-05,
"end_lr": 1e-5,
"increase_epochs": 180,
"decrease_epochs": 180,
"final_epochs": 0,
},
report_args=report_args,
)
for tempo in tempos:
key_suffix = []
key_suffix.append("tempo"+ f"{tempo['prob']}_{tempo['minimum']}_{tempo['maximum']}")
key = f"scf_conf-wei-oldspecaug-audio_perturbation_{'_'.join(key_suffix)}"
audio_perturb_args = {}
audio_perturb_args["tempo"] = tempo
report_args={
"architecture": "conf-wei",
"lr": "wei_peak_4e-4_e450_cycle360",
"specaug": "wei_adapt_80dim",
"wave_norm": "True",
}
report_args["tempo"] = f"{tempo['prob']}_{tempo['minimum']}_{tempo['maximum']}"
nn_base_args[key] = dict(
returnn_args={
"extra_args": {
"audio_perturb_args": audio_perturb_args,
"audio_perturb_runner": CodeWrapper("WaveformPerturbation(**audio_perturb_args)"),
},
**returnn_args,},
feature_args=feature_args,
lr_args={
"peak_lr": 4e-4,
"start_lr": 1.325e-05,
"end_lr": 1e-5,
"increase_epochs": 180,
"decrease_epochs": 180,
"final_epochs": 0,
},
report_args=report_args,
)
for preemphasis in preemphases:
key_suffix = []
key_suffix.append("preemphasis"+ f"{preemphasis['prob']}_{preemphasis['minimum']}_{preemphasis['maximum']}")
key = f"scf_conf-wei-oldspecaug-audio_perturbation_{'_'.join(key_suffix)}"
audio_perturb_args = {}
audio_perturb_args["preemphasis"] = preemphasis
report_args={
"architecture": "conf-wei",
"lr": "wei_peak_4e-4_e450_cycle360",
"specaug": "wei_adapt_80dim",
"wave_norm": "True",
}
report_args[
"preemphasis"
] = f"{preemphasis['prob']}_{preemphasis['minimum']}_{preemphasis['maximum']}"
nn_base_args[key] = dict(
returnn_args={
"extra_args": {
"audio_perturb_args": audio_perturb_args,
"audio_perturb_runner": CodeWrapper("WaveformPerturbation(**audio_perturb_args)"),
},
**returnn_args,},
feature_args=feature_args,
lr_args={
"peak_lr": 4e-4,
"start_lr": 1.325e-05,
"end_lr": 1e-5,
"increase_epochs": 180,
"decrease_epochs": 180,
"final_epochs": 0,
},
report_args=report_args,
)
for codec in codecs:
key_suffix = []
key_suffix.append("codec" + f"wav_{codec['encoding'].lower()}_{codec['prob']}")
key = f"scf_conf-wei-oldspecaug-audio_perturbation_{'_'.join(key_suffix)}"
audio_perturb_args = {}
audio_perturb_args["codecs"] = [codec]
report_args={
"architecture": "conf-wei",
"lr": "wei_peak_4e-4_e450_cycle360",
"specaug": "wei_adapt_80dim",
"wave_norm": "True",
}
report_args["codec"] = f"wav_{codec['encoding'].lower()}_{codec['prob']}"
nn_base_args[key] = dict(
returnn_args={
"extra_args": {
"audio_perturb_args": audio_perturb_args,
"audio_perturb_runner": CodeWrapper("WaveformPerturbation(**audio_perturb_args)"),
},
**returnn_args,},
feature_args=feature_args,
lr_args={
"peak_lr": 4e-4,
"start_lr": 1.325e-05,
"end_lr": 1e-5,
"increase_epochs": 180,
"decrease_epochs": 180,
"final_epochs": 0,
},
report_args=report_args,
)
for non_linearity in non_linearities:
key_suffix = []
key_suffix.append("non_linearity" + f"{non_linearity['prob']}_{non_linearity['alpha']}")
key = f"scf_conf-wei-oldspecaug-audio_perturbation_{'_'.join(key_suffix)}"
audio_perturb_args = {}
audio_perturb_args["non_linearities"] = [non_linearity]
report_args={
"architecture": "conf-wei",
"lr": "wei_peak_4e-4_e450_cycle360",
"specaug": "wei_adapt_80dim",
"wave_norm": "True",
}
report_args["non_linearity"] = f"{non_linearity['prob']}_{non_linearity['alpha']}"
nn_base_args[key] = dict(
returnn_args={
"extra_args": {
"audio_perturb_args": audio_perturb_args,
"audio_perturb_runner": CodeWrapper("WaveformPerturbation(**audio_perturb_args)"),
},
**returnn_args,},
feature_args=feature_args,
lr_args={
"peak_lr": 4e-4,
"start_lr": 1.325e-05,
"end_lr": 1e-5,
"increase_epochs": 180,
"decrease_epochs": 180,
"final_epochs": 0,
},
report_args=report_args,
)
nn_args, report_args_collection = get_nn_args_baseline(
nn_base_args=nn_base_args,
num_epochs=450,
prefix="conformer_bs5k_audio_perturbation_",
)
run_nn_args(nn_args, report_args_collection, "report_scf_audio_perturbation.csv", dev_corpora)
def run_scf_watch_mem_test():
gs.ALIAS_AND_OUTPUT_SUBDIR = "experiments/switchboard/ctc/feat/"
(
returnn_datasets,
rasr_loss_corpus_path,
rasr_loss_corpus_segments,
rasr_loss_lexicon_path,
dev_corpora,
) = get_datasets()
returnn_args = {
"batch_size": 5000,
"rasr_binary_path": RASR_BINARY_PATH,
"rasr_loss_corpus_path": rasr_loss_corpus_path,
"rasr_loss_corpus_segments": rasr_loss_corpus_segments,
"rasr_loss_lexicon_path": rasr_loss_lexicon_path,
"datasets": returnn_datasets,
"extra_args": {
"watch_memory": True,
},
}
feature_args = {"class": "ScfNetwork", "size_tf": 256 // 2, "stride_tf": 10 // 2}
nn_args, report_args_collection = get_nn_args_baseline(
nn_base_args={
"scf_watch_mem_test": dict(
returnn_args={"conformer_type": "wei", "specaug_old": {"max_feature": 8}, **returnn_args},
feature_args=feature_args,
lr_args={
"peak_lr": 4e-4,
"start_lr": 1.325e-05,
"end_lr": 1e-5,
"increase_epochs": 180,
"decrease_epochs": 180,
"final_epochs": 0,
},
report_args={
"architecture": "conf-wei",
"lr": "wei_peak_4e-4_e450_cycle360",
"specaug": "wei_adapt_80dim",
"wave_norm": "True",
},
),
},
num_epochs=450,
prefix="conformer_bs5k_",
)
run_nn_args(nn_args, report_args_collection, "report_watch_mem_test.csv", dev_corpora)
def run_mel_watch_mem_test():
gs.ALIAS_AND_OUTPUT_SUBDIR = "experiments/switchboard/ctc/feat/"
(
returnn_datasets,
rasr_loss_corpus_path,
rasr_loss_corpus_segments,
rasr_loss_lexicon_path,
dev_corpora,
) = get_datasets()
returnn_args = {
"batch_size": 5000,
"rasr_binary_path": RASR_BINARY_PATH,
"rasr_loss_corpus_path": rasr_loss_corpus_path,
"rasr_loss_corpus_segments": rasr_loss_corpus_segments,
"rasr_loss_lexicon_path": rasr_loss_lexicon_path,
"datasets": returnn_datasets,
"extra_args": {
"watch_memory": True,
},
}
feature_args = {"class": "LogMelNetwork", "wave_norm": True, "frame_size": 200, "frame_shift": 80, "fft_size": 256}
nn_args, report_args_collection = get_nn_args_baseline(
nn_base_args={
"lgm80__watch_mem_test": dict(
returnn_args={"conformer_type": "wei", "specaug_old": {"max_feature": 8}, **returnn_args},
feature_args=feature_args,
lr_args={
"peak_lr": 4e-4,
"start_lr": 1.325e-05,
"end_lr": 1e-5,
"increase_epochs": 180,
"decrease_epochs": 180,
"final_epochs": 0,
},
report_args={
"architecture": "conf-wei",
"lr": "wei_peak_4e-4_e450_cycle360",
"specaug": "wei_adapt_80dim",
"wave_norm": "True",
},
),
},
num_epochs=450,
prefix="conformer_bs5k_",
)
run_nn_args(nn_args, report_args_collection, "report_mel_watch_mem_test", dev_corpora)
def py():
"""
called if the file is passed to sis manager, used to run all experiments (replacement for main)
"""
run_mel_baseline()
run_scf_baseline()
run_mel_audio_perturbation()
run_scf_audio_perturbation()
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment