From 2f595bee1cc540fd22b10fb8f92bc5f9485a14b9 Mon Sep 17 00:00:00 2001
From: wangbixing123 <wangbixing@huawei.com>
Date: Sat, 18 Jun 2022 11:04:42 +0800
Subject: [PATCH] fix bug for ecapa_tdnn
---
.../audio/ecapa_tdnn/ecapa-tdnn_config.yaml | 18 +++++++++---------
official/audio/ecapa_tdnn/train.py | 4 +++-
2 files changed, 12 insertions(+), 10 deletions(-)
diff --git a/official/audio/ecapa_tdnn/ecapa-tdnn_config.yaml b/official/audio/ecapa_tdnn/ecapa-tdnn_config.yaml
index 131cd29fe..381b62ba1 100644
--- a/official/audio/ecapa_tdnn/ecapa-tdnn_config.yaml
+++ b/official/audio/ecapa_tdnn/ecapa-tdnn_config.yaml
@@ -28,33 +28,33 @@ run_distribute: 0
# ==============================================================================
# options
-in_channels: 80 # input channel size, same as the dim of fbank feature
+in_channels: 80 # input channel size, same as the dim of fbank feature
channels: 1024 # channel size of middle layer feature map
base_lrate: 0.000001 # base learning rate of cyclic LR
max_lrate: 0.0001 # max learning rate of cyclic LR
momentum: 0.95 # weight decay for optimizer
-weightDecay: 0.000002 # momentum for optimizer
-num_epochs: 2 # training epoch
+weight_decay: 0.000002 # momentum for optimizer
+num_epochs: 3 # training epoch
minibatch_size: 192 # batch size
emb_size: 192 # embedding dim
step_size: 65000 # steps to achieve max learning rate cyclic LR
class_num: 7205 # speaker num pf voxceleb1&2
pre_trained: False # if pre-trained model exist
-train_data_path: "/Temp/abc000/feat_train/" # path to fbank training data
+train_data_path: "/Temp/abc/feat_train/" # path to fbank training data
keep_checkpoint_max: 20
checkpoint_path: "train_ecapa_vox2_full-2_664204.ckpt" # path to pre-trained model
ckpt_save_dir: "./ckpt/" # path to store train model
# eval
-eval_data_path: "/home/abc000/feat_eval/" # path to eval fbank data
-veri_file_path: /home/abc000/feat_eval/veri_test_bleeched.txt # trials
+eval_data_path: "/Temp/abc/feat_eval/" # path to eval fbank data
+veri_file_path: /Temp/abc/feat_eval/veri_test_bleeched.txt # trials
cut_wav: false # cut wav to 3s (cut wav to 3s, same as train data)
-model_path: "ckpt/train_ecapa_vox2_full-2_664204.ckpt" # path of eval model
-train_norm_path: "/home/abc000/feat_norm/" # fbank data for norm
+model_path: "/Temp/abc/ckpt/train_ecapa_vox12-1_660204.ckpt" # path of eval model
+train_norm_path: "/Temp/abc/feat_norm/" # fbank data for norm
score_norm: "s-norm" # if do norm, uncomment this two line
cohort_size: 20000 # max number of utts to do norm
-npy_file_path: './npys/' # dir to save intermediate result
+npy_file_path: '/Temp/abc/npys/' # dir to save intermediate result
# # export option
exp_ckpt_file: './train_ecapa_vox2_full-2_664204.ckpt'
diff --git a/official/audio/ecapa_tdnn/train.py b/official/audio/ecapa_tdnn/train.py
index c28e18f6d..955325d2e 100644
--- a/official/audio/ecapa_tdnn/train.py
+++ b/official/audio/ecapa_tdnn/train.py
@@ -50,7 +50,9 @@ def create_dataset(cfg, data_home, shuffle=False):
"""
dataset_generator = DatasetGenerator(data_home)
- distributed_sampler = DistributedSampler(len(dataset_generator), cfg.group_size, cfg.rank, shuffle=True)
+ distributed_sampler = None
+ if cfg.run_distribute:
+ distributed_sampler = DistributedSampler(len(dataset_generator), cfg.group_size, cfg.rank, shuffle=True)
vox2_ds = ds.GeneratorDataset(dataset_generator, ["data", "label"], shuffle=shuffle, sampler=distributed_sampler)
cnt = int(len(dataset_generator) / cfg.group_size)
return vox2_ds, cnt
--
GitLab