Instructions to use arda-argmax/sortformer-finetuned with libraries, inference providers, notebooks, and local apps. Follow these links to get started.
- Libraries
- NeMo
How to use arda-argmax/sortformer-finetuned with NeMo:
# tag did not correspond to a valid NeMo domain.
- Notebooks
- Google Colab
- Kaggle
Init run 6spk-v0.6-fastmss-prepared-shar-bs12-a100-80gb-2gpu-fullset-v4-1780546132
Browse files
6spk-v0.6-fastmss-prepared-shar-bs12-a100-80gb-2gpu-fullset-v4-1780546132/README.md
ADDED
|
@@ -0,0 +1,26 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
---
|
| 2 |
+
library_name: nemo
|
| 3 |
+
tags:
|
| 4 |
+
- sortformer
|
| 5 |
+
- speaker-diarization
|
| 6 |
+
- automatic-speech-recognition
|
| 7 |
+
---
|
| 8 |
+
|
| 9 |
+
# 6spk-v0.6-fastmss-prepared-shar-bs12-a100-80gb-2gpu-fullset-v4-1780546132
|
| 10 |
+
|
| 11 |
+
Sortformer fine-tuning run, hosted under [arda-argmax/sortformer-finetuned](https://huggingface.co/arda-argmax/sortformer-finetuned).
|
| 12 |
+
|
| 13 |
+
## Best checkpoint
|
| 14 |
+
|
| 15 |
+
- **monitor**: `val_der` (min)
|
| 16 |
+
|
| 17 |
+
Files:
|
| 18 |
+
- `best.nemo` — full NeMo checkpoint (config + weights), restorable via `SortformerEncLabelModel.restore_from('best.nemo')`
|
| 19 |
+
- `best.json` — machine-readable metadata for the best checkpoint
|
| 20 |
+
- `checkpoints/step_*_epoch_*_*.nemo` — archived copies of every checkpoint that became the best during training
|
| 21 |
+
- `config.yaml` — Hydra config the run was launched with
|
| 22 |
+
|
| 23 |
+
## Run info
|
| 24 |
+
|
| 25 |
+
- started: `2026-06-04T05:05:06Z`
|
| 26 |
+
- live dashboard: [arda-argmax/sortformer-trackio-v2](arda-argmax/sortformer-trackio-v2)
|
6spk-v0.6-fastmss-prepared-shar-bs12-a100-80gb-2gpu-fullset-v4-1780546132/config.yaml
ADDED
|
@@ -0,0 +1,114 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
seed: 42
|
| 2 |
+
force_exit_on_completion: true
|
| 3 |
+
model:
|
| 4 |
+
init_checkpoint: arda-argmax/diar_streaming_sortformer_6spk-init
|
| 5 |
+
num_spks: 6
|
| 6 |
+
data:
|
| 7 |
+
cuts_path: null
|
| 8 |
+
dataset_hf_repo: null
|
| 9 |
+
split: null
|
| 10 |
+
cuts_filename: null
|
| 11 |
+
preprocess: true
|
| 12 |
+
audio_root: null
|
| 13 |
+
session_len_sec: 120
|
| 14 |
+
sample_rate: 16000
|
| 15 |
+
batch_size: 12
|
| 16 |
+
num_workers: 4
|
| 17 |
+
shuffle: true
|
| 18 |
+
prepared:
|
| 19 |
+
enabled: true
|
| 20 |
+
uri: gs://fastmss-synth/sortformer/fastmss_synth/v0.6_rttm_segment_opus96k
|
| 21 |
+
local_dir: /root/fastmss_prepared_shar
|
| 22 |
+
root_dir: null
|
| 23 |
+
train_dir: null
|
| 24 |
+
val_dir: null
|
| 25 |
+
train_path: train
|
| 26 |
+
val_path: val
|
| 27 |
+
sync: true
|
| 28 |
+
gcloud_bin: gcloud
|
| 29 |
+
max_train_cuts: null
|
| 30 |
+
max_val_cuts: null
|
| 31 |
+
num_train_samples: 2996000
|
| 32 |
+
streaming:
|
| 33 |
+
enabled: false
|
| 34 |
+
repo: arda-argmax/simchoir-parquet
|
| 35 |
+
subset: v0.1
|
| 36 |
+
label_source: legacy
|
| 37 |
+
cache_materialization: stream
|
| 38 |
+
download_dir: /tmp/sortformer_hf_parquet_chunks
|
| 39 |
+
materialization_workers: 1
|
| 40 |
+
subset_prefix: null
|
| 41 |
+
train_subsets: null
|
| 42 |
+
val_subsets: null
|
| 43 |
+
train_split: train
|
| 44 |
+
val_split: val
|
| 45 |
+
train_shuffle_buffer: 0
|
| 46 |
+
sampler_shuffle_buffer: 1000
|
| 47 |
+
num_train_samples: 800
|
| 48 |
+
train_cache_size: -1
|
| 49 |
+
cache_storage: disk
|
| 50 |
+
cache_dir: /tmp/sortformer_streaming_cache
|
| 51 |
+
val_cache_size: -1
|
| 52 |
+
val_replicate_across_ranks: false
|
| 53 |
+
bucketing:
|
| 54 |
+
enabled: false
|
| 55 |
+
max_duration_sec: null
|
| 56 |
+
num_buckets: 10
|
| 57 |
+
buffer_size: 256
|
| 58 |
+
quadratic_duration: 30.0
|
| 59 |
+
duration_bins: []
|
| 60 |
+
bin_estimate_samples: 10000
|
| 61 |
+
optim:
|
| 62 |
+
learning_rate: 5.0e-05
|
| 63 |
+
weight_decay: 0.0
|
| 64 |
+
warmup_steps: 50
|
| 65 |
+
min_lr: 1.0e-06
|
| 66 |
+
trainer:
|
| 67 |
+
accelerator: gpu
|
| 68 |
+
devices: auto
|
| 69 |
+
precision: bf16-mixed
|
| 70 |
+
fallback: false
|
| 71 |
+
num_steps: -1
|
| 72 |
+
num_epochs: 3
|
| 73 |
+
val_check_interval: 0.05
|
| 74 |
+
check_val_every_n_epoch: 1
|
| 75 |
+
log_every_n_steps: 1
|
| 76 |
+
limit_val_batches: 1.0
|
| 77 |
+
num_sanity_val_steps: 0
|
| 78 |
+
accumulate_grad_batches: 1
|
| 79 |
+
strategy: auto
|
| 80 |
+
regularization:
|
| 81 |
+
disable_all_dropout: false
|
| 82 |
+
causal_attn_rate: null
|
| 83 |
+
scores_add_rnd: null
|
| 84 |
+
output:
|
| 85 |
+
final_checkpoint_save_path: /app/training_runs/1780546132/final.nemo
|
| 86 |
+
log_dir: /app/training_runs/1780546132/logs
|
| 87 |
+
run_name: 6spk-v0.6-fastmss-prepared-shar-bs12-a100-80gb-2gpu-fullset-v4-1780546132
|
| 88 |
+
tracking:
|
| 89 |
+
enabled: true
|
| 90 |
+
project: sortformer-robustness
|
| 91 |
+
run_name: 6spk-v0.6-fastmss-prepared-shar-bs12-a100-80gb-2gpu-fullset-v4-1780546132
|
| 92 |
+
group: null
|
| 93 |
+
space_id: null
|
| 94 |
+
auto_log_gpu: true
|
| 95 |
+
gpu_log_interval: 10.0
|
| 96 |
+
console_log:
|
| 97 |
+
enabled: false
|
| 98 |
+
max_chars: 200000
|
| 99 |
+
every_n_train_steps: 0
|
| 100 |
+
val_der:
|
| 101 |
+
enabled: true
|
| 102 |
+
collar: 0.0
|
| 103 |
+
skip_overlap: false
|
| 104 |
+
threshold: 0.5
|
| 105 |
+
max_samples: null
|
| 106 |
+
checkpointing:
|
| 107 |
+
enabled: true
|
| 108 |
+
repo: arda-argmax/sortformer-finetuned
|
| 109 |
+
run_subdir: 6spk-v0.6-fastmss-prepared-shar-bs12-a100-80gb-2gpu-fullset-v4-1780546132
|
| 110 |
+
monitor: val_der
|
| 111 |
+
mode: min
|
| 112 |
+
private: true
|
| 113 |
+
local_dir: null
|
| 114 |
+
flush_timeout_sec: 1800
|