Add encoder part of whisper large v3 as an audio encoder model. (#9894)

Not useful yet but some models use it.

comfy/audio_encoders/audio_encoders.py

@@ -1,4 +1,5 @@
 from .wav2vec2 import Wav2Vec2Model
+from .whisper import WhisperLargeV3
 import comfy.model_management
 import comfy.ops
 import comfy.utils
@@ -11,13 +12,18 @@ class AudioEncoderModel():
         self.load_device = comfy.model_management.text_encoder_device()
         offload_device = comfy.model_management.text_encoder_offload_device()
         self.dtype = comfy.model_management.text_encoder_dtype(self.load_device)
+        model_type = config.pop("model_type")
         model_config = dict(config)
         model_config.update({
             "dtype": self.dtype,
             "device": offload_device,
             "operations": comfy.ops.manual_cast
         })
-        self.model = Wav2Vec2Model(**model_config)
+
+        if model_type == "wav2vec2":
+            self.model = Wav2Vec2Model(**model_config)
+        elif model_type == "whisper3":
+            self.model = WhisperLargeV3(**model_config)
         self.model.eval()
         self.patcher = comfy.model_patcher.ModelPatcher(self.model, load_device=self.load_device, offload_device=offload_device)
         self.model_sample_rate = 16000
@@ -40,33 +46,45 @@ class AudioEncoderModel():
 
 def load_audio_encoder_from_sd(sd, prefix=""):
     sd = comfy.utils.state_dict_prefix_replace(sd, {"wav2vec2.": ""})
-    embed_dim = sd["encoder.layer_norm.bias"].shape[0]
-    if embed_dim == 1024:# large
-        config = {
-            "embed_dim": 1024,
-            "num_heads": 16,
-            "num_layers": 24,
-            "conv_norm": True,
-            "conv_bias": True,
-            "do_normalize": True,
-            "do_stable_layer_norm": True
+    if "encoder.layer_norm.bias" in sd: #wav2vec2
+        embed_dim = sd["encoder.layer_norm.bias"].shape[0]
+        if embed_dim == 1024:# large
+            config = {
+                "model_type": "wav2vec2",
+                "embed_dim": 1024,
+                "num_heads": 16,
+                "num_layers": 24,
+                "conv_norm": True,
+                "conv_bias": True,
+                "do_normalize": True,
+                "do_stable_layer_norm": True
+                }
+        elif embed_dim == 768: # base
+            config = {
+                "model_type": "wav2vec2",
+                "embed_dim": 768,
+                "num_heads": 12,
+                "num_layers": 12,
+                "conv_norm": False,
+                "conv_bias": False,
+                "do_normalize": False, # chinese-wav2vec2-base has this False
+                "do_stable_layer_norm": False
             }
-    elif embed_dim == 768: # base
+        else:
+            raise RuntimeError("ERROR: audio encoder file is invalid or unsupported embed_dim: {}".format(embed_dim))
+    elif "model.encoder.embed_positions.weight" in sd:
+        sd = comfy.utils.state_dict_prefix_replace(sd, {"model.": ""})
         config = {
-            "embed_dim": 768,
-            "num_heads": 12,
-            "num_layers": 12,
-            "conv_norm": False,
-            "conv_bias": False,
-            "do_normalize": False, # chinese-wav2vec2-base has this False
-            "do_stable_layer_norm": False
+            "model_type": "whisper3",
         }
     else:
-        raise RuntimeError("ERROR: audio encoder file is invalid or unsupported embed_dim: {}".format(embed_dim))
+        raise RuntimeError("ERROR: audio encoder not supported.")
 
     audio_encoder = AudioEncoderModel(config)
     m, u = audio_encoder.load_sd(sd)
     if len(m) > 0:
         logging.warning("missing audio encoder: {}".format(m))
+    if len(u) > 0:
+        logging.warning("unexpected audio encoder: {}".format(u))
 
     return audio_encoder

comfy/audio_encoders/whisper.py

@@ -0,0 +1,186 @@
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+import torchaudio
+from typing import Optional
+from comfy.ldm.modules.attention import optimized_attention_masked
+import comfy.ops
+
+class WhisperFeatureExtractor(nn.Module):
+    def __init__(self, n_mels=128, device=None):
+        super().__init__()
+        self.sample_rate = 16000
+        self.n_fft = 400
+        self.hop_length = 160
+        self.n_mels = n_mels
+        self.chunk_length = 30
+        self.n_samples = 480000
+
+        self.mel_spectrogram = torchaudio.transforms.MelSpectrogram(
+            sample_rate=self.sample_rate,
+            n_fft=self.n_fft,
+            hop_length=self.hop_length,
+            n_mels=self.n_mels,
+            f_min=0,
+            f_max=8000,
+            norm="slaney",
+            mel_scale="slaney",
+        ).to(device)
+
+    def __call__(self, audio):
+        audio = torch.mean(audio, dim=1)
+        batch_size = audio.shape[0]
+        processed_audio = []
+
+        for i in range(batch_size):
+            aud = audio[i]
+            if aud.shape[0] > self.n_samples:
+                aud = aud[:self.n_samples]
+            elif aud.shape[0] < self.n_samples:
+                aud = F.pad(aud, (0, self.n_samples - aud.shape[0]))
+            processed_audio.append(aud)
+
+        audio = torch.stack(processed_audio)
+
+        mel_spec = self.mel_spectrogram(audio.to(self.mel_spectrogram.spectrogram.window.device))[:, :, :-1].to(audio.device)
+
+        log_mel_spec = torch.clamp(mel_spec, min=1e-10).log10()
+        log_mel_spec = torch.maximum(log_mel_spec, log_mel_spec.max() - 8.0)
+        log_mel_spec = (log_mel_spec + 4.0) / 4.0
+
+        return log_mel_spec
+
+
+class MultiHeadAttention(nn.Module):
+    def __init__(self, d_model: int, n_heads: int, dtype=None, device=None, operations=None):
+        super().__init__()
+        assert d_model % n_heads == 0
+
+        self.d_model = d_model
+        self.n_heads = n_heads
+        self.d_k = d_model // n_heads
+
+        self.q_proj = operations.Linear(d_model, d_model, dtype=dtype, device=device)
+        self.k_proj = operations.Linear(d_model, d_model, bias=False, dtype=dtype, device=device)
+        self.v_proj = operations.Linear(d_model, d_model, dtype=dtype, device=device)
+        self.out_proj = operations.Linear(d_model, d_model, dtype=dtype, device=device)
+
+    def forward(
+        self,
+        query: torch.Tensor,
+        key: torch.Tensor,
+        value: torch.Tensor,
+        mask: Optional[torch.Tensor] = None,
+    ) -> torch.Tensor:
+        batch_size, seq_len, _ = query.shape
+
+        q = self.q_proj(query)
+        k = self.k_proj(key)
+        v = self.v_proj(value)
+
+        attn_output = optimized_attention_masked(q, k, v, self.n_heads, mask)
+        attn_output = self.out_proj(attn_output)
+
+        return attn_output
+
+
+class EncoderLayer(nn.Module):
+    def __init__(self, d_model: int, n_heads: int, d_ff: int, dtype=None, device=None, operations=None):
+        super().__init__()
+
+        self.self_attn = MultiHeadAttention(d_model, n_heads, dtype=dtype, device=device, operations=operations)
+        self.self_attn_layer_norm = operations.LayerNorm(d_model, dtype=dtype, device=device)
+
+        self.fc1 = operations.Linear(d_model, d_ff, dtype=dtype, device=device)
+        self.fc2 = operations.Linear(d_ff, d_model, dtype=dtype, device=device)
+        self.final_layer_norm = operations.LayerNorm(d_model, dtype=dtype, device=device)
+
+    def forward(
+        self,
+        x: torch.Tensor,
+        attention_mask: Optional[torch.Tensor] = None
+    ) -> torch.Tensor:
+        residual = x
+        x = self.self_attn_layer_norm(x)
+        x = self.self_attn(x, x, x, attention_mask)
+        x = residual + x
+
+        residual = x
+        x = self.final_layer_norm(x)
+        x = self.fc1(x)
+        x = F.gelu(x)
+        x = self.fc2(x)
+        x = residual + x
+
+        return x
+
+
+class AudioEncoder(nn.Module):
+    def __init__(
+        self,
+        n_mels: int = 128,
+        n_ctx: int = 1500,
+        n_state: int = 1280,
+        n_head: int = 20,
+        n_layer: int = 32,
+        dtype=None,
+        device=None,
+        operations=None
+    ):
+        super().__init__()
+
+        self.conv1 = operations.Conv1d(n_mels, n_state, kernel_size=3, padding=1, dtype=dtype, device=device)
+        self.conv2 = operations.Conv1d(n_state, n_state, kernel_size=3, stride=2, padding=1, dtype=dtype, device=device)
+
+        self.embed_positions = operations.Embedding(n_ctx, n_state, dtype=dtype, device=device)
+
+        self.layers = nn.ModuleList([
+            EncoderLayer(n_state, n_head, n_state * 4, dtype=dtype, device=device, operations=operations)
+            for _ in range(n_layer)
+        ])
+
+        self.layer_norm = operations.LayerNorm(n_state, dtype=dtype, device=device)
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        x = F.gelu(self.conv1(x))
+        x = F.gelu(self.conv2(x))
+
+        x = x.transpose(1, 2)
+
+        x = x + comfy.ops.cast_to_input(self.embed_positions.weight[:, :x.shape[1]], x)
+
+        all_x = ()
+        for layer in self.layers:
+            all_x += (x,)
+            x = layer(x)
+
+        x = self.layer_norm(x)
+        all_x += (x,)
+        return x, all_x
+
+
+class WhisperLargeV3(nn.Module):
+    def __init__(
+        self,
+        n_mels: int = 128,
+        n_audio_ctx: int = 1500,
+        n_audio_state: int = 1280,
+        n_audio_head: int = 20,
+        n_audio_layer: int = 32,
+        dtype=None,
+        device=None,
+        operations=None
+    ):
+        super().__init__()
+
+        self.feature_extractor = WhisperFeatureExtractor(n_mels=n_mels, device=device)
+
+        self.encoder = AudioEncoder(
+            n_mels, n_audio_ctx, n_audio_state, n_audio_head, n_audio_layer,
+            dtype=dtype, device=device, operations=operations
+        )
+
+    def forward(self, audio):
+        mel = self.feature_extractor(audio)
+        x, all_x = self.encoder(mel)
+        return x, all_x