Add files using upload-large-folder tool

dasheng_embedding_train/checkpoints/best.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:7f05d8aebb46e3d35195aa863db58b7ea73512476238d45ac86267d727384561
+size 1104559107

dasheng_embedding_train/checkpoints/best_checkpoints.json

@@ -0,0 +1,17 @@
+[
+  {
+    "path": "step=72500-val_eer=0.051587.pt",
+    "step": 72500,
+    "metric": 0.051587301587301626
+  },
+  {
+    "path": "step=73500-val_eer=0.051587.pt",
+    "step": 73500,
+    "metric": 0.051587301587301626
+  },
+  {
+    "path": "step=75500-val_eer=0.051587.pt",
+    "step": 75500,
+    "metric": 0.051587301587301626
+  }
+]

dasheng_embedding_train/checkpoints/last.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:081cb1f45101cd82f6edb70db5f9c123197fc93eb81eb1e0d02e7a714c4709b7
+size 1104559107

dasheng_embedding_train/checkpoints/step=72500-val_eer=0.051587.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:dde78340d2720512d309de1e1e85d85e6dd4210095d46b755f88d33538201625
+size 1104607195

dasheng_embedding_train/checkpoints/step=73500-val_eer=0.051587.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d6572b458cca52dcca1f88ee30f18e705c07ded8ba43ad4d78add9d7b357a159
+size 1104607195

dasheng_embedding_train/checkpoints/step=75500-val_eer=0.051587.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:6b660c4e4ebabf623ca0e7c7c25638bb9f6ecd1337ed22e5f288ad3921e11bf7
+size 1104607195

dasheng_embedding_train/config.yaml

@@ -0,0 +1,144 @@
+name: dasheng_embedding_train
+seed: 42
+num_workers: 16
+batch_size: 16
+eval_batch_size: 16
+model:
+  audio_backbone: dasheng
+  encoder_type: base
+  audio_encoder: dasheng
+  target_sr: 16000
+  mono: true
+  dasheng_name: dasheng_base
+  dasheng_pretrained: true
+  dasheng_path: null
+  dasheng_model_args: {}
+  dasheng_preprocessor: {}
+  dasheng_encoder: {}
+  use_ecapa: false
+  ecapa_preprocessor:
+    mode: 16k
+    freeze_pretrained: true
+  embedding_dim: 768
+  train_ds:
+    s3_base_path: https://d2j287p0ytux1o.cloudfront.net
+    datasets:
+    - path: sv_datasets/voxceleb2_dev_full/data.parquet
+      ratio: 1
+    aws_profile: test_user
+    cache_dir: /workspace/cache
+    enable_audio_cache: true
+    audio_cache_dir: /workspace/cache/audio
+    max_query_audio_length: 30.0
+    max_doc_audio_length: 30.0
+    lmdb_augment:
+      enabled: true
+      aug_prob: 1
+      reverb_lmdb_path: /workspace/ALM/aug_dataset/rirs/lmdb
+      noise_lmdb_path: /workspace/ALM/aug_dataset/musan/lmdb
+    num_workers: 16
+    batch_size: 16
+    shuffle: true
+    drop_last: true
+    pin_memory: true
+    prefetch_factor: 2
+    persistent_workers: true
+  validation_ds:
+    s3_base_path: https://d2j287p0ytux1o.cloudfront.net
+    aws_profile: test_user
+    cache_dir: /workspace/cache
+    enable_audio_cache: true
+    audio_cache_dir: /workspace/cache/audio
+    datasets:
+    - name: voxceleb_hard_original
+      path: sv_benchmarks/voxceleb1_hard_500_original/data.parquet
+      ratio: 1
+    - name: voxceleb1_hard_500_500ms_one_way
+      path: sv_benchmarks/voxceleb1_hard_500_500ms_one_way/data.parquet
+      ratio: 1
+    - name: voxceleb1_hard_500_500ms_two_way
+      path: sv_benchmarks/voxceleb1_hard_500_500ms_two_way/data.parquet
+      ratio: 1
+    - name: voxceleb1_hard_500_1000ms_one_way
+      path: sv_benchmarks/voxceleb1_hard_500_1000ms_one_way/data.parquet
+      ratio: 1
+    - name: voxceleb1_hard_500_1000ms_two_way
+      path: sv_benchmarks/voxceleb1_hard_500_1000ms_two_way/data.parquet
+      ratio: 1
+    - name: voxceleb1_hard_500_1500ms_one_way
+      path: sv_benchmarks/voxceleb1_hard_500_1500ms_one_way/data.parquet
+      ratio: 1
+    - name: voxceleb1_hard_500_1500ms_two_way
+      path: sv_benchmarks/voxceleb1_hard_500_1500ms_two_way/data.parquet
+      ratio: 1
+    - name: voxceleb1_hard_500_2000ms_one_way
+      path: sv_benchmarks/voxceleb1_hard_500_2000ms_one_way/data.parquet
+      ratio: 1
+    - name: voxceleb1_hard_500_2000ms_two_way
+      path: sv_benchmarks/voxceleb1_hard_500_2000ms_two_way/data.parquet
+      ratio: 1
+    - name: voxceleb1_hard_500_2500ms_one_way
+      path: sv_benchmarks/voxceleb1_hard_500_2500ms_one_way/data.parquet
+      ratio: 1
+    - name: voxceleb1_hard_500_2500ms_two_way
+      path: sv_benchmarks/voxceleb1_hard_500_2500ms_two_way/data.parquet
+      ratio: 1
+    - name: voxceleb1_hard_500_stitch_100ms_1500ms
+      path: sv_benchmarks/voxceleb1_hard_500_stitch_100ms_1500ms/data.parquet
+      ratio: 1
+    - name: voxceleb1_hard_500_stitch_500ms_2500ms
+      path: sv_benchmarks/voxceleb1_hard_500_stitch_500ms_2500ms/data.parquet
+      ratio: 1
+    - name: alm_benchmark_pairs
+      path: sv_benchmarks/alm_benchmark_pairs/data.parquet
+      ratio: 1
+    eval_batch_size: 16
+    eval_max_query_audio_length: null
+    eval_max_doc_audio_length: null
+    num_workers: 16
+    shuffle: false
+    drop_last: false
+    pin_memory: true
+    prefetch_factor: 2
+    persistent_workers: true
+  logit_scale: 4.25
+  loss:
+    type: aamsoftmax
+    params:
+      n_class: 5994
+      m: 0.2
+      s: 30.0
+      embedding_dim: 768
+  lr: 1.0e-05
+  optim:
+    name: adamw
+    lr: 1.0e-05
+    betas:
+    - 0.9
+    - 0.999
+    weight_decay: 0.01
+    sched:
+      name: InverseSquareRootAnnealing
+      warmup_steps: 500
+      min_lr: 1.0e-06
+trainer:
+  precision: bf16-mixed
+  gradient_accumulation_steps: 4
+  max_steps: 1000000
+  log_every_n_steps: 20
+  val_check_interval: 500
+  max_grad_norm: 1.0
+  num_sanity_val_steps: 0
+  limit_val_batches: null
+exp_manager:
+  exp_dir: ./outputs
+  name: dasheng_embedding_train
+  resume_if_exists: true
+  resume_from_checkpoint: null
+  save_last: true
+  create_checkpoint_callback: true
+  checkpoint_callback_params:
+    monitor: val_eer
+    mode: min
+    save_top_k: 3
+    every_n_train_steps: 1000

dasheng_embedding_train/tensorboard/events.out.tfevents.1776947142.fd77baabd680.137401.0

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:9102691c01ba22e5a85b143e9b767d06747cfdc1518a1ff15d3cfbb099ce9e7d
+size 983163