Commit 3d66da98 authored by Pierre Champion's avatar Pierre Champion
Browse files

fix half_resnet34 training

parent 35f3a0a0
Pipeline #644 canceled with stages
......@@ -10,11 +10,12 @@ Authors: Anthony Larcher & .....
```sh
git clone https://git-lium.univ-lemans.fr/speaker/sidekit
cd sidekit
# you might need to adjust $CUDAROOT in ./install.sh
# to match your cuda config. default /usr/local/cuda
# you might need to adjust $CUDAROOT in ./install.sh to match your cuda config. default /usr/local/cuda
./install.sh
```
---
## Usage
#### Extract x-vector for kaldi-like wav.scp
......@@ -79,6 +80,7 @@ compute_spk_cosine.py ./trials ./utt2spk x-vector-trials.scp ./x-vector-enrolls.
compute_metrics.py -k ./trials -s cosine_score.txt
```
---
#### For Python
......
......@@ -27,7 +27,7 @@ python ./local/dataprep_aug.py --from ./data/musan_split --make-csv-augment-nois
```
### Train from scratch
Multiple x-vector architectures are implemented, each of them has their own `train_<model_type>.sh` script.
Multiple x-vector architectures are implemented, each of them has their own `train_<model_type>.sh` script.
Example:
```bash
......@@ -35,13 +35,15 @@ Example:
```
During training, logs will be put under `logs/<model_type>` and checkpoints will be placed under `model_<model_type>/`.
---
### Results
#### `train_vox2_wavlm_ecapa_circle.sh`
Training data Voxceleb 2
Training time on 3xRTX8000: less than 8 hours
Last Test EER `0.756917179336569` %
Last Test EER with AS-norm `0.736604252853012` %
validation_ratio is set to zero (no train/val split, training on the full dataset)
Training data Voxceleb 2
Training time on 3xRTX8000: less than 8 hours
Last Test EER `0.756917179336569` %
Last Test EER with AS-norm `0.736604252853012` %
validation_ratio is set to zero (no train/val split, training on the full dataset)
---
......@@ -10,7 +10,7 @@ sample_rate: 16000
validation_ratio: 0.02
batch_size: 512
batch_size: 192
# Training set
train:
......@@ -20,7 +20,7 @@ train:
sampler:
examples_per_speaker: 1
samples_per_speaker: 100
samples_per_speaker: 192
augmentation_replica: 1
transform_number: 1
......
......@@ -31,7 +31,7 @@ scheduler:
# Evaluation options
compute_test_eer: false
log_interval: 10
log_interval: 461
validation_frequency: 1
# Save options
......
......@@ -1178,7 +1178,7 @@ def get_loaders(dataset_opts, training_opts, model_opts, local_rank=0):
samples_per_speaker = 1
if training_opts["multi_gpu"]:
# assert dataset_opts["batch_size"] % torch.cuda.device_count() == 0
assert dataset_opts["batch_size"] % torch.cuda.device_count() == 0
assert dataset_opts["batch_size"] % samples_per_speaker == 0
training_set = SideSet(dataset_opts,
......
Supports Markdown
0% or .
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment