From f5b35ba23deb160e7d1d0d2646adaf3081070f82 Mon Sep 17 00:00:00 2001
From: 嘉渊 <wangjiaming.wjm@alibaba-inc.com>
Date: 星期三, 17 五月 2023 15:28:19 +0800
Subject: [PATCH] update repo
---
funasr/utils/prepare_data.py | 3
/dev/null | 209 ----------------------------------------------------
egs/librispeech/conformer/run.sh | 2
egs/librispeech_100h/conformer/run.sh | 2
4 files changed, 4 insertions(+), 212 deletions(-)
diff --git a/egs/librispeech/conformer/run.sh b/egs/librispeech/conformer/run.sh
index 2e34cbf..b942dd2 100755
--- a/egs/librispeech/conformer/run.sh
+++ b/egs/librispeech/conformer/run.sh
@@ -55,7 +55,7 @@
inference_config=conf/decode_asr_transformer.yaml
#inference_config=conf/decode_asr_transformer_beam60_ctc0.3.yaml
-inference_asr_model=valid.acc.ave_10best.pth
+inference_asr_model=valid.acc.ave_10best.pb
# you can set gpu num for decoding here
gpuid_list=$CUDA_VISIBLE_DEVICES # set gpus for decoding, the same as training stage by default
diff --git a/egs/librispeech_100h/conformer/run.sh b/egs/librispeech_100h/conformer/run.sh
index da7a597..d1a20bc 100755
--- a/egs/librispeech_100h/conformer/run.sh
+++ b/egs/librispeech_100h/conformer/run.sh
@@ -55,7 +55,7 @@
inference_config=conf/decode_asr_transformer.yaml
#inference_config=conf/decode_asr_transformer_beam60_ctc0.3.yaml
-inference_asr_model=valid.acc.ave_10best.pth
+inference_asr_model=valid.acc.ave_10best.pb
# you can set gpu num for decoding here
gpuid_list=$CUDA_VISIBLE_DEVICES # set gpus for decoding, the same as training stage by default
diff --git a/funasr/utils/prepare_data.py b/funasr/utils/prepare_data.py
index f61e501..36795b4 100644
--- a/funasr/utils/prepare_data.py
+++ b/funasr/utils/prepare_data.py
@@ -207,10 +207,11 @@
data_names = args.dataset_conf.get("data_names", "speech,text").split(",")
data_types = args.dataset_conf.get("data_types", "sound,text").split(",")
file_names = args.data_file_names.split(",")
+ print("data_names: {}, data_types: {}, file_names: {}".format(data_names, data_types, file_names))
assert len(data_names) == len(data_types) == len(file_names)
if args.dataset_type == "small":
args.train_shape_file = [os.path.join(args.data_dir, args.train_set, "{}_shape".format(data_names[0]))]
- args.valid_shape_file = [os.path.join(args.data_dir, args.valid_set, "{}}_shape".format(data_names[0]))]
+ args.valid_shape_file = [os.path.join(args.data_dir, args.valid_set, "{}_shape".format(data_names[0]))]
args.train_data_path_and_name_and_type, args.valid_data_path_and_name_and_type = [], []
for file_name, data_name, data_type in zip(file_names, data_names, data_types):
args.train_data_path_and_name_and_type.append(
diff --git a/funasr/utils/prepare_data.py.bak b/funasr/utils/prepare_data.py.bak
deleted file mode 100644
index 3f55170..0000000
--- a/funasr/utils/prepare_data.py.bak
+++ /dev/null
@@ -1,209 +0,0 @@
-import logging
-import os
-import shutil
-from multiprocessing import Pool
-
-import numpy as np
-import torch.distributed as dist
-import torchaudio
-
-
-def filter_wav_text(data_dir, dataset):
- wav_file = os.path.join(data_dir, dataset, "wav.scp")
- text_file = os.path.join(data_dir, dataset, "text")
- with open(wav_file) as f_wav, open(text_file) as f_text:
- wav_lines = f_wav.readlines()
- text_lines = f_text.readlines()
- os.rename(wav_file, "{}.bak".format(wav_file))
- os.rename(text_file, "{}.bak".format(text_file))
- wav_dict = {}
- for line in wav_lines:
- parts = line.strip().split()
- if len(parts) < 2:
- continue
- wav_dict[parts[0]] = parts[1]
- text_dict = {}
- for line in text_lines:
- parts = line.strip().split()
- if len(parts) < 2:
- continue
- text_dict[parts[0]] = " ".join(parts[1:])
- filter_count = 0
- with open(wav_file, "w") as f_wav, open(text_file, "w") as f_text:
- for sample_name, wav_path in wav_dict.items():
- if sample_name in text_dict.keys():
- f_wav.write(sample_name + " " + wav_path + "\n")
- f_text.write(sample_name + " " + text_dict[sample_name] + "\n")
- else:
- filter_count += 1
- logging.info("{}/{} samples in {} are filtered because of the mismatch between wav.scp and text".
- format(filter_count, len(wav_lines), dataset))
-
-
-def wav2num_frame(wav_path, frontend_conf):
- waveform, sampling_rate = torchaudio.load(wav_path)
- n_frames = (waveform.shape[1] * 1000.0) / (sampling_rate * frontend_conf["frame_shift"] * frontend_conf["lfr_n"])
- feature_dim = frontend_conf["n_mels"] * frontend_conf["lfr_m"]
- return n_frames, feature_dim
-
-
-def calc_shape_core(root_path, args, idx):
- wav_scp_file = os.path.join(root_path, "wav.scp.{}".format(idx))
- shape_file = os.path.join(root_path, "speech_shape.{}".format(idx))
- with open(wav_scp_file) as f:
- lines = f.readlines()
- frontend_conf = args.frontend_conf
- dataset_conf = args.dataset_conf
- speech_length_min = dataset_conf.speech_length_min if hasattr(dataset_conf, "speech_length_min") else -1
- speech_length_max = dataset_conf.speech_length_max if hasattr(dataset_conf, "speech_length_max") else -1
- with open(shape_file, "w") as f:
- for line in lines:
- sample_name, wav_path = line.strip().split()
- n_frames, feature_dim = wav2num_frame(wav_path, frontend_conf)
- write_flag = True
- if n_frames > 0 and speech_length_min > 0:
- write_flag = n_frames >= speech_length_min
- if n_frames > 0 and speech_length_max > 0:
- write_flag = n_frames <= speech_length_max
- if write_flag:
- f.write("{} {},{}\n".format(sample_name, str(int(np.ceil(n_frames))), str(int(feature_dim))))
- f.flush()
-
-
-def calc_shape(args, dataset, nj=64):
- shape_path = os.path.join(args.data_dir, dataset, "speech_shape")
- if os.path.exists(shape_path):
- logging.info('Shape file for small dataset already exists.')
- return
-
- split_shape_path = os.path.join(args.data_dir, dataset, "shape_files")
- if os.path.exists(split_shape_path):
- shutil.rmtree(split_shape_path)
- os.mkdir(split_shape_path)
-
- # split
- wav_scp_file = os.path.join(args.data_dir, dataset, "wav.scp")
- with open(wav_scp_file) as f:
- lines = f.readlines()
- num_lines = len(lines)
- num_job_lines = num_lines // nj
- start = 0
- for i in range(nj):
- end = start + num_job_lines
- file = os.path.join(split_shape_path, "wav.scp.{}".format(str(i + 1)))
- with open(file, "w") as f:
- if i == nj - 1:
- f.writelines(lines[start:])
- else:
- f.writelines(lines[start:end])
- start = end
-
- p = Pool(nj)
- for i in range(nj):
- p.apply_async(calc_shape_core, args=(split_shape_path, args, str(i + 1)))
- logging.info("Generating shape files, please wait a few minutes...")
- p.close()
- p.join()
-
- # combine
- with open(shape_path, "w") as f:
- for i in range(nj):
- job_file = os.path.join(split_shape_path, "speech_shape.{}".format(str(i + 1)))
- with open(job_file) as job_f:
- lines = job_f.readlines()
- f.writelines(lines)
- logging.info('Generating shape files done.')
-
-
-def generate_data_list(data_dir, dataset, nj=64):
- list_file = os.path.join(data_dir, dataset, "data.list")
- if os.path.exists(list_file):
- logging.info('Data list for large dataset already exists.')
- return
- split_path = os.path.join(data_dir, dataset, "split")
- if os.path.exists(split_path):
- shutil.rmtree(split_path)
- os.mkdir(split_path)
-
- with open(os.path.join(data_dir, dataset, "wav.scp")) as f_wav:
- wav_lines = f_wav.readlines()
- with open(os.path.join(data_dir, dataset, "text")) as f_text:
- text_lines = f_text.readlines()
- num_lines = len(wav_lines)
- num_job_lines = num_lines // nj
- start = 0
- for i in range(nj):
- end = start + num_job_lines
- split_path_nj = os.path.join(split_path, str(i + 1))
- os.mkdir(split_path_nj)
- wav_file = os.path.join(split_path_nj, "wav.scp")
- text_file = os.path.join(split_path_nj, "text")
- with open(wav_file, "w") as fw, open(text_file, "w") as ft:
- if i == nj - 1:
- fw.writelines(wav_lines[start:])
- ft.writelines(text_lines[start:])
- else:
- fw.writelines(wav_lines[start:end])
- ft.writelines(text_lines[start:end])
- start = end
-
- with open(list_file, "w") as f_data:
- for i in range(nj):
- wav_path = os.path.join(split_path, str(i + 1), "wav.scp")
- text_path = os.path.join(split_path, str(i + 1), "text")
- f_data.write(wav_path + " " + text_path + "\n")
-
-
-def prepare_data(args, distributed_option):
- distributed = distributed_option.distributed
- if not distributed or distributed_option.dist_rank == 0:
- filter_wav_text(args.data_dir, args.train_set)
- filter_wav_text(args.data_dir, args.valid_set)
-
- if args.dataset_type == "small":
- calc_shape(args, args.train_set)
- calc_shape(args, args.valid_set)
-
- if args.dataset_type == "large":
- generate_data_list(args.data_dir, args.train_set)
- generate_data_list(args.data_dir, args.valid_set)
-
- if args.dataset_type == "small":
- args.train_shape_file = [os.path.join(args.data_dir, args.train_set, "speech_shape")]
- args.valid_shape_file = [os.path.join(args.data_dir, args.valid_set, "speech_shape")]
- data_names = args.dataset_conf.get("data_names", "speech,text").split(",")
- data_types = args.dataset_conf.get("data_types", "sound,text").split(",")
- args.train_data_path_and_name_and_type = [
- ["{}/{}/wav.scp".format(args.data_dir, args.train_set), data_names[0], data_types[0]],
- ["{}/{}/text".format(args.data_dir, args.train_set), data_names[1], data_types[1]]
- ]
- args.valid_data_path_and_name_and_type = [
- ["{}/{}/wav.scp".format(args.data_dir, args.valid_set), data_names[0], data_types[0]],
- ["{}/{}/text".format(args.data_dir, args.valid_set), data_names[1], data_types[1]]
- ]
- if args.embed_path is not None:
- args.train_data_path_and_name_and_type.append(
- [os.path.join(args.embed_path, "embeds", args.train_set, "embeds.scp"), "embed", "kaldi_ark"])
- args.valid_data_path_and_name_and_type.append(
- [os.path.join(args.embed_path, "embeds", args.valid_set, "embeds.scp"), "embed", "kaldi_ark"])
- else:
- args.train_data_file = os.path.join(args.data_dir, args.train_set, "data.list")
- args.valid_data_file = os.path.join(args.data_dir, args.valid_set, "data.list")
- if args.embed_path is not None:
- if not distributed or distributed_option.dist_rank == 0:
- for d in [args.train_set, args.valid_set]:
- file = os.path.join(args.data_dir, d, "data.list")
- with open(file) as f:
- lines = f.readlines()
- out_file = os.path.join(args.data_dir, d, "data_with_embed.list")
- with open(out_file, "w") as out_f:
- for line in lines:
- parts = line.strip().split()
- idx = parts[0].split("/")[-2]
- embed_file = os.path.join(args.embed_path, "embeds", args.valid_set, "ark",
- "embeds.{}.ark".format(idx))
- out_f.write(parts[0] + " " + parts[1] + " " + embed_file + "\n")
- args.train_data_file = os.path.join(args.data_dir, args.train_set, "data_with_embed.list")
- args.valid_data_file = os.path.join(args.data_dir, args.valid_set, "data_with_embed.list")
- if distributed:
- dist.barrier()
--
Gitblit v1.9.1