Add config.yaml so the Hub counts downloads

README.md

@@ -24,7 +24,7 @@ Code, training and evaluation: <https://github.com/JusperLee/Falcon>. Paper: *Ef
 |---|---|---|---|
 | `falcon-big.pt` | `studio/configs/FalconBig.yml` | `SpeakerContrastiveV2` | 12.35 M |
 
-Total parameters include the frozen 11.2 M-parameter FRCNN lip-reading encoder. The file holds only `state_dict` (`audio_model.*`, `video_model.*`), the architecture name and the training epoch; load it with `torch.load(..., weights_only=True)`.
+Total parameters include the frozen 11.2 M-parameter FRCNN lip-reading encoder. The file holds only `state_dict` (`audio_model.*`, `video_model.*`), the architecture name and the training epoch; load it with `torch.load(..., weights_only=True)`. `config.yaml` is a copy of `studio/configs/FalconBig.yml` from the code repo; the Hub counts downloads by this file, and `FalconRuntime` requests it on every load.
 
 ## Usage
 

config.yaml

@@ -0,0 +1,166 @@
+audionet:
+  audionet_config:
+    audio_bn_params:
+      is2d: true
+      kernel_size: 1
+      out_chan: 256
+      pre_act_type: ReLU
+      pre_norm_type: cLNhw
+    audio_params:
+      act_type: PReLU
+      causal: true
+      hid_chan: 96
+      is2d: true
+      kernel_size: 5
+      layers:
+        layer_1:
+          bidirectional: true
+          dim: 4
+          hid_chan: 48
+          kernel_size: 8
+          num_layers: 1
+          rnn_type: LSTM
+          stride: 1
+        layer_2:
+          bidirectional: false
+          dim: 3
+          hid_chan: 96
+          kernel_size: 8
+          num_layers: 1
+          rnn_type: LSTM
+          stride: 1
+        layer_3:
+          act_type: PReLU
+          causal: true
+          dim: 3
+          hid_chan: 4
+          n_head: 4
+          norm_type: LayerNormalization4D
+      n_freqs: 129
+      norm_type: cLNhw
+      repeats: 2
+      shared: true
+      stride: 2
+      upsampling_depth: 4
+    causal: true
+    enc_dec_params:
+      act_type: null
+      bias: false
+      causal: true
+      hop_length: 128
+      kernel_size: 3
+      norm_type: null
+      out_chan: 256
+      stride: 1
+      win: 256
+    fusion_params:
+      fusion_shared: true
+      fusion_type: NeuroSCNetFusion
+      is2d: true
+      kernel_size: 4
+    mask_generation_params:
+      RI_split: true
+      is2d: true
+      mask_act: ReLU
+      mask_generator_type: MaskGenerator
+    n_src: 1
+    pretrained_vout_chan: 512
+    spk_embed_dim: 128
+    video_bn_params:
+      kernel_size: -1
+    video_params:
+      act_type: PReLU
+      causal: true
+      hid_chan: 64
+      is2d: false
+      kernel_size: 3
+      layers:
+        layer_1:
+          causal: true
+          dropout: 0.1
+          kernel_size: 3
+          n_head: 8
+          norm_type: cLN
+      norm_type: BatchNorm1d
+      repeats: 1
+      shared: true
+      stride: 2
+  audionet_name: SpeakerContrastiveV2
+datamodule:
+  data_config:
+    batch_size: 4
+    n_src: 1
+    normalize_audio: false
+    num_workers: 8
+    persistent_workers: true
+    pin_memory: true
+    sample_rate: 16000
+    segment: 2.0
+    test_dir: ${DATA_ROOT}/tt
+    train_dir: ${DATA_ROOT}/tr
+    use_near: true
+    valid_dir: ${DATA_ROOT}/cv
+  data_name: AVSpeechDataModuleWithNear
+exp:
+  exp_name: big_1_1000_imgswd
+loss:
+  train:
+    config:
+      pit_from: pw_mtx
+      threshold_byloss: true
+    loss_func: PITLossWrapper
+    sdr_type: pairwise_neg_snr
+  val:
+    config:
+      pit_from: pw_mtx
+      threshold_byloss: false
+    loss_func: PITLossWrapper
+    sdr_type: pairwise_neg_sisdr
+main_args:
+  exp_dir: ${ROOT}/Experiments/checkpoint/big_1_1000_imgswd
+optimizer:
+  lr: 0.001
+  optim_name: adamw
+  weight_decay: 0.01
+options:
+  ckpt_path: ''
+  conf_dir: configs/FalconBig.yml
+  help: null
+scheduler:
+  sche_config:
+    factor: 0.5
+    patience: 15
+  sche_name: ReduceLROnPlateau
+spk_contrast:
+  alpha_neg: 1.0
+  alpha_pos: 1.0
+  beta_anneal_epochs: 200
+  beta_final: 1.0
+  beta_init: 0.0
+  loss_weight: 1.0
+  margin: 0.5
+training:
+  early_stop:
+    mode: min
+    monitor: val_loss/dataloader_idx_0
+    patience: 30
+    verbose: true
+  epochs: 1000
+  gpus:
+  - 0
+  - 1
+  - 2
+  - 3
+  - 4
+  - 5
+  - 6
+  - 7
+  parallel: ddp
+  system: AudioVisualLightningModuleNASSv2
+videonet:
+  videonet_config:
+    backbone_type: resnet
+    pretrain: pretrain_zoo/frcnn_128_512.backbone.pth.tar
+    relu_type: prelu
+    width_mult: 1.0
+  videonet_name: FRCNNVideoModel