From a84a00eecfd3b230155a228c4be683c5b1d689a0 Mon Sep 17 00:00:00 2001 From: Hui Zhang Date: Sat, 8 May 2021 09:01:46 +0000 Subject: [PATCH] filter test, since librispeech will cause OOM, but all test wer will be worse, since mismatch train with test --- deepspeech/exps/deepspeech2/model.py | 12 ++++++------ deepspeech/exps/u2/model.py | 12 ++++++------ deepspeech/training/trainer.py | 1 + 3 files changed, 13 insertions(+), 12 deletions(-) diff --git a/deepspeech/exps/deepspeech2/model.py b/deepspeech/exps/deepspeech2/model.py index a163ef4c..b1ff5c59 100644 --- a/deepspeech/exps/deepspeech2/model.py +++ b/deepspeech/exps/deepspeech2/model.py @@ -323,12 +323,12 @@ class DeepSpeech2Tester(DeepSpeech2Trainer): config.data.manifest = config.data.test_manifest config.data.keep_transcription_text = True config.data.augmentation_config = "" - config.data.min_input_len = 0.0 # second - config.data.max_input_len = float('inf') # second - config.data.min_output_len = 0.0 # tokens - config.data.max_output_len = float('inf') # tokens - config.data.min_output_input_ratio = 0.00 - config.data.max_output_input_ratio = float('inf') + # config.data.min_input_len = 0.0 # second + # config.data.max_input_len = float('inf') # second + # config.data.min_output_len = 0.0 # tokens + # config.data.max_output_len = float('inf') # tokens + # config.data.min_output_input_ratio = 0.00 + # config.data.max_output_input_ratio = float('inf') test_dataset = ManifestDataset.from_config(config) # return text ord id diff --git a/deepspeech/exps/u2/model.py b/deepspeech/exps/u2/model.py index 7a409d98..0bcd1d22 100644 --- a/deepspeech/exps/u2/model.py +++ b/deepspeech/exps/u2/model.py @@ -254,12 +254,12 @@ class U2Trainer(Trainer): config.data.manifest = config.data.test_manifest config.data.keep_transcription_text = True config.data.augmentation_config = "" - config.data.min_input_len = 0.0 # second - config.data.max_input_len = float('inf') # second - config.data.min_output_len = 0.0 # tokens - config.data.max_output_len = float('inf') # tokens - config.data.min_output_input_ratio = 0.00 - config.data.max_output_input_ratio = float('inf') + # config.data.min_input_len = 0.0 # second + # config.data.max_input_len = float('inf') # second + # config.data.min_output_len = 0.0 # tokens + # config.data.max_output_len = float('inf') # tokens + # config.data.min_output_input_ratio = 0.00 + # config.data.max_output_input_ratio = float('inf') test_dataset = ManifestDataset.from_config(config) # return text ord id self.test_loader = DataLoader( diff --git a/deepspeech/training/trainer.py b/deepspeech/training/trainer.py index 220914f3..e630febb 100644 --- a/deepspeech/training/trainer.py +++ b/deepspeech/training/trainer.py @@ -264,6 +264,7 @@ class Trainer(): @mp_tools.rank_zero_only def destory(self): + """Close visualizer to avoid hanging after training""" # https://github.com/pytorch/fairseq/issues/2357 if self.visualizer: self.visualizer.close() -- GitLab