Upload folder using huggingface_hub

content-vec-best/.gitattributes

@@ -0,0 +1,34 @@
+*.7z filter=lfs diff=lfs merge=lfs -text
+*.arrow filter=lfs diff=lfs merge=lfs -text
+*.bin filter=lfs diff=lfs merge=lfs -text
+*.bz2 filter=lfs diff=lfs merge=lfs -text
+*.ckpt filter=lfs diff=lfs merge=lfs -text
+*.ftz filter=lfs diff=lfs merge=lfs -text
+*.gz filter=lfs diff=lfs merge=lfs -text
+*.h5 filter=lfs diff=lfs merge=lfs -text
+*.joblib filter=lfs diff=lfs merge=lfs -text
+*.lfs.* filter=lfs diff=lfs merge=lfs -text
+*.mlmodel filter=lfs diff=lfs merge=lfs -text
+*.model filter=lfs diff=lfs merge=lfs -text
+*.msgpack filter=lfs diff=lfs merge=lfs -text
+*.npy filter=lfs diff=lfs merge=lfs -text
+*.npz filter=lfs diff=lfs merge=lfs -text
+*.onnx filter=lfs diff=lfs merge=lfs -text
+*.ot filter=lfs diff=lfs merge=lfs -text
+*.parquet filter=lfs diff=lfs merge=lfs -text
+*.pb filter=lfs diff=lfs merge=lfs -text
+*.pickle filter=lfs diff=lfs merge=lfs -text
+*.pkl filter=lfs diff=lfs merge=lfs -text
+*.pt filter=lfs diff=lfs merge=lfs -text
+*.pth filter=lfs diff=lfs merge=lfs -text
+*.rar filter=lfs diff=lfs merge=lfs -text
+*.safetensors filter=lfs diff=lfs merge=lfs -text
+saved_model/**/* filter=lfs diff=lfs merge=lfs -text
+*.tar.* filter=lfs diff=lfs merge=lfs -text
+*.tflite filter=lfs diff=lfs merge=lfs -text
+*.tgz filter=lfs diff=lfs merge=lfs -text
+*.wasm filter=lfs diff=lfs merge=lfs -text
+*.xz filter=lfs diff=lfs merge=lfs -text
+*.zip filter=lfs diff=lfs merge=lfs -text
+*.zst filter=lfs diff=lfs merge=lfs -text
+*tfevents* filter=lfs diff=lfs merge=lfs -text

content-vec-best/.gitignore

@@ -0,0 +1 @@
+content-vec-best-legacy-500.pt

content-vec-best/README.md

@@ -0,0 +1,33 @@
+---
+license: mit
+---
+
+# Content Vec Best
+Official Repo: [ContentVec](https://github.com/auspicious3000/contentvec)  
+This repo brings fairseq ContentVec model to HuggingFace Transformers.
+
+## How to use
+To use this model, you need to define
+```python
+class HubertModelWithFinalProj(HubertModel):
+    def __init__(self, config):
+        super().__init__(config)
+
+        # The final projection layer is only used for backward compatibility.
+        # Following https://github.com/auspicious3000/contentvec/issues/6
+        # Remove this layer is necessary to achieve the desired outcome.
+        self.final_proj = nn.Linear(config.hidden_size, config.classifier_proj_size)
+```
+
+and then load the model with
+```python
+model = HubertModelWithFinalProj.from_pretrained("lengyue233/content-vec-best")
+
+x = model(audio)["last_hidden_state"]
+```
+
+## How to convert
+You need to download the ContentVec_legacy model from the official repo, and then run
+```bash
+python convert.py
+```

content-vec-best/config.json

@@ -0,0 +1,71 @@
+{
+  "activation_dropout": 0.1,
+  "apply_spec_augment": true,
+  "architectures": [
+    "HubertModelWithFinalProj"
+  ],
+  "attention_dropout": 0.1,
+  "bos_token_id": 1,
+  "classifier_proj_size": 256,
+  "conv_bias": false,
+  "conv_dim": [
+    512,
+    512,
+    512,
+    512,
+    512,
+    512,
+    512
+  ],
+  "conv_kernel": [
+    10,
+    3,
+    3,
+    3,
+    3,
+    2,
+    2
+  ],
+  "conv_stride": [
+    5,
+    2,
+    2,
+    2,
+    2,
+    2,
+    2
+  ],
+  "ctc_loss_reduction": "sum",
+  "ctc_zero_infinity": false,
+  "do_stable_layer_norm": false,
+  "eos_token_id": 2,
+  "feat_extract_activation": "gelu",
+  "feat_extract_norm": "group",
+  "feat_proj_dropout": 0.0,
+  "feat_proj_layer_norm": true,
+  "final_dropout": 0.1,
+  "hidden_act": "gelu",
+  "hidden_dropout": 0.1,
+  "hidden_size": 768,
+  "initializer_range": 0.02,
+  "intermediate_size": 3072,
+  "layer_norm_eps": 1e-05,
+  "layerdrop": 0.1,
+  "mask_feature_length": 10,
+  "mask_feature_min_masks": 0,
+  "mask_feature_prob": 0.0,
+  "mask_time_length": 10,
+  "mask_time_min_masks": 2,
+  "mask_time_prob": 0.05,
+  "model_type": "hubert",
+  "num_attention_heads": 12,
+  "num_conv_pos_embedding_groups": 16,
+  "num_conv_pos_embeddings": 128,
+  "num_feat_extract_layers": 7,
+  "num_hidden_layers": 12,
+  "pad_token_id": 0,
+  "torch_dtype": "float32",
+  "transformers_version": "4.27.3",
+  "use_weighted_layer_sum": false,
+  "vocab_size": 32
+}

content-vec-best/convert.py

@@ -0,0 +1,150 @@
+import torch
+from torch import nn
+from transformers import HubertConfig, HubertModel
+import logging
+
+# Ignore fairseq's logger
+logging.getLogger("fairseq").setLevel(logging.WARNING)
+logging.getLogger("torch.distributed.nn.jit.instantiator").setLevel(logging.WARNING)
+
+from fairseq import checkpoint_utils
+
+models, _, _ = checkpoint_utils.load_model_ensemble_and_task(
+    ["content-vec-best-legacy-500.pt"], suffix=""
+)
+model = models[0]
+model.eval()
+model.eval()
+
+
+class HubertModelWithFinalProj(HubertModel):
+    def __init__(self, config):
+        super().__init__(config)
+
+        self.final_proj = nn.Linear(config.hidden_size, config.classifier_proj_size)
+
+
+# Default Config
+hubert = HubertModelWithFinalProj(HubertConfig())
+
+# huggingface: fairseq
+mapping = {
+    "masked_spec_embed": "mask_emb",
+    "encoder.layer_norm.bias": "encoder.layer_norm.bias",
+    "encoder.layer_norm.weight": "encoder.layer_norm.weight",
+    "encoder.pos_conv_embed.conv.bias": "encoder.pos_conv.0.bias",
+    "encoder.pos_conv_embed.conv.weight_g": "encoder.pos_conv.0.weight_g",
+    "encoder.pos_conv_embed.conv.weight_v": "encoder.pos_conv.0.weight_v",
+    "feature_projection.layer_norm.bias": "layer_norm.bias",
+    "feature_projection.layer_norm.weight": "layer_norm.weight",
+    "feature_projection.projection.bias": "post_extract_proj.bias",
+    "feature_projection.projection.weight": "post_extract_proj.weight",
+    "final_proj.bias": "final_proj.bias",
+    "final_proj.weight": "final_proj.weight",
+}
+
+# Convert encoder
+for layer in range(12):
+    for j in ["q", "k", "v"]:
+        mapping[
+            f"encoder.layers.{layer}.attention.{j}_proj.weight"
+        ] = f"encoder.layers.{layer}.self_attn.{j}_proj.weight"
+        mapping[
+            f"encoder.layers.{layer}.attention.{j}_proj.bias"
+        ] = f"encoder.layers.{layer}.self_attn.{j}_proj.bias"
+
+    mapping[
+        f"encoder.layers.{layer}.final_layer_norm.bias"
+    ] = f"encoder.layers.{layer}.final_layer_norm.bias"
+    mapping[
+        f"encoder.layers.{layer}.final_layer_norm.weight"
+    ] = f"encoder.layers.{layer}.final_layer_norm.weight"
+
+    mapping[
+        f"encoder.layers.{layer}.layer_norm.bias"
+    ] = f"encoder.layers.{layer}.self_attn_layer_norm.bias"
+    mapping[
+        f"encoder.layers.{layer}.layer_norm.weight"
+    ] = f"encoder.layers.{layer}.self_attn_layer_norm.weight"
+
+    mapping[
+        f"encoder.layers.{layer}.attention.out_proj.bias"
+    ] = f"encoder.layers.{layer}.self_attn.out_proj.bias"
+    mapping[
+        f"encoder.layers.{layer}.attention.out_proj.weight"
+    ] = f"encoder.layers.{layer}.self_attn.out_proj.weight"
+
+    mapping[
+        f"encoder.layers.{layer}.feed_forward.intermediate_dense.bias"
+    ] = f"encoder.layers.{layer}.fc1.bias"
+    mapping[
+        f"encoder.layers.{layer}.feed_forward.intermediate_dense.weight"
+    ] = f"encoder.layers.{layer}.fc1.weight"
+
+    mapping[
+        f"encoder.layers.{layer}.feed_forward.output_dense.bias"
+    ] = f"encoder.layers.{layer}.fc2.bias"
+    mapping[
+        f"encoder.layers.{layer}.feed_forward.output_dense.weight"
+    ] = f"encoder.layers.{layer}.fc2.weight"
+
+# Convert Conv Layers
+for layer in range(7):
+    mapping[
+        f"feature_extractor.conv_layers.{layer}.conv.weight"
+    ] = f"feature_extractor.conv_layers.{layer}.0.weight"
+
+    if layer != 0:
+        continue
+
+    mapping[
+        f"feature_extractor.conv_layers.{layer}.layer_norm.weight"
+    ] = f"feature_extractor.conv_layers.{layer}.2.weight"
+    mapping[
+        f"feature_extractor.conv_layers.{layer}.layer_norm.bias"
+    ] = f"feature_extractor.conv_layers.{layer}.2.bias"
+
+hf_keys = set(hubert.state_dict().keys())
+fair_keys = set(model.state_dict().keys())
+
+hf_keys -= set(mapping.keys())
+fair_keys -= set(mapping.values())
+
+for i, j in zip(sorted(hf_keys), sorted(fair_keys)):
+    print(i, j)
+
+print(hf_keys, fair_keys)
+print(len(hf_keys), len(fair_keys))
+
+# try loading the weights
+new_state_dict = {}
+for k, v in mapping.items():
+    new_state_dict[k] = model.state_dict()[v]
+
+x = hubert.load_state_dict(new_state_dict, strict=False)
+print(x)
+hubert.eval()
+
+with torch.no_grad():
+    new_input = torch.randn(1, 16384)
+
+    result1 = hubert(new_input, output_hidden_states=True)["hidden_states"][9]
+    result1 = hubert.final_proj(result1)
+
+    result2 = model.extract_features(
+        **{
+            "source": new_input,
+            "padding_mask": torch.zeros(1, 16384, dtype=torch.bool),
+            # "features_only": True,
+            "output_layer": 9,
+        }
+    )[0]
+    result2 = model.final_proj(result2)
+
+    assert torch.allclose(result1, result2, atol=1e-3)
+
+print("Sanity check passed")
+
+# Save huggingface model
+hubert.save_pretrained(".")
+print("Saved model")

content-vec-best/pytorch_model.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d8dd400e054ddf4e6be75dab5a2549db748cc99e756a097c496c099f65a4854e
+size 378342945