Duplicate from hance-ai/audiomae

Co-authored-by: Daesoo Lee <dslee2601@users.noreply.huggingface.co>

.gitattributes

@@ -0,0 +1,36 @@
+*.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
+*.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
+sanity_check_result_audiomae.png filter=lfs diff=lfs merge=lfs -text

.gitignore

@@ -0,0 +1,2 @@
+__pycache__/
+token.txt

README.md

@@ -0,0 +1,51 @@
+---
+license: mit
+tags:
+- AudioMAE
+- PyTorch
+---
+
+# AudioMAE
+This model card provides an easy-to-use API for a *pretrained encoder of AudioMAE* [1] whose weights are from [its original reposotiry](https://github.com/facebookresearch/AudioMAE).
+The provided model is specifically designed to easily obtain learned representations given an input audio file.
+The resulting representation $z$ has a dimension of $(d, h, w)$ for a single audio file, where $d$, $h$, and $w$ denote a latent dimension size, latent frequency dim, and latent temporal dim, respectively. 
+[2] indicates that both frequency and temporal dimensional semantics are preserved in $z$.
+
+# Dependency
+See `requirements.txt`
+
+
+# Usage
+```python
+from transformers import AutoModel
+
+device = 'cpu'  # 'cpu' or 'cuda'
+model = AutoModel.from_pretrained("hance-ai/audiomae", trust_remote_code=True).to(device)  # load the pretrained model
+z = model('path/audio_fname.wav')  # (768, 8, 64) = (latent_dim_size, latent_freq_dim, latent_temporal_dim)
+```
+
+Depending on a task, a different pooling strategy should be facilitated. 
+For instance, a global average pooling can be used for a classification task. [2] uses an adaptive pooling.
+
+⚠️ AudioMAE accepts audio with maximum length of 10s (as described in [1]). Any audio longer than 10s will be clipped to 10s, meaning the excess beyond 10s will be discarded.
+
+
+# Sanity Check Result
+In the following, a spectrogram of an input audio and corresponding $z$ are visualized. 
+The input audio is 10s, containing baby coughing, hiccuping, and adult sneezing.
+The latent dimension size of $z$ is reduced to 8 using PCA for visualization.
+
+<p align="center">
+<img src=".fig/sanity_check_result_audiomae.png" alt="" width=100%>
+</p>
+
+The result shows that the presence of labeled sound is clearly captured in the 3rd principal component (PC). 
+While the baby coughing and hiccuping sounds are not so distinugisable up to the 5th PC, they are in the 6th PC.
+This result briefly shows the effectiveness of the pretrained AudioMAE.
+
+
+# References
+
+[1] Huang, Po-Yao, et al. "Masked autoencoders that listen." Advances in Neural Information Processing Systems 35 (2022): 28708-28720.
+
+[2] Liu, Haohe, et al. "Audioldm 2: Learning holistic audio generation with self-supervised pretraining." IEEE/ACM Transactions on Audio, Speech, and Language Processing (2024).

config.json

@@ -0,0 +1,18 @@
+{
+  "architectures": [
+    "PretrainedAudioMAEEncoder"
+  ],
+  "auto_map": {
+    "AutoConfig": "model.AudioMAEConfig",
+    "AutoModel": "model.PretrainedAudioMAEEncoder"
+  },
+  "img_size": [
+    1024,
+    128
+  ],
+  "in_chans": 1,
+  "model_type": "audiomae",
+  "num_classes": 0,
+  "torch_dtype": "float32",
+  "transformers_version": "4.44.0"
+}

model.py

@@ -0,0 +1,130 @@
+from typing import Tuple
+
+import torch
+import torchaudio
+import torchaudio.transforms as transforms
+from torchaudio.compliance import kaldi
+from transformers import PretrainedConfig
+
+from einops import rearrange
+
+from timm.models.vision_transformer import VisionTransformer
+from transformers import PreTrainedModel
+
+
+# it seems like Config class and Model class should be located in the same file; otherwise, seemingly casuing an issue in model loading after pushing to HF.
+class AudioMAEConfig(PretrainedConfig):
+    model_type = "audiomae"
+
+    def __init__(self,
+                 img_size:Tuple[int,int]=(1024,128),
+                 in_chans:int=1,
+                 num_classes:int=0,
+                 **kwargs,):
+        super().__init__(**kwargs)
+        self.img_size = img_size
+        self.in_chans = in_chans
+        self.num_classes = num_classes
+
+
+class AudioMAEEncoder(VisionTransformer):
+    def __init__(self, *args, **kwargs):
+        super().__init__(*args, **kwargs)
+        """
+        - img_size of (1024, 128) = (temporal_length, n_freq_bins) is fixed, as described in the paper
+        - AudoMAE accepts a mono-channel (i.e., in_chans=1)
+        """
+        self.MEAN = -4.2677393  # written on the paper
+        self.STD = 4.5689974  # written on the paper
+
+    def load_wav_file(self, file_path:str):
+        """
+        to use this, `torchaudio` and `ffmpeg` must be installed
+        - `ffmpeg` version must be >=4.4 and <7.
+        - `ffmpeg` installation by `conda install -c conda-forge ffmpeg==6.1.1`
+        """
+        audio, sample_rate = torchaudio.load(file_path)  # audio: (n_channels, length); 
+
+        # length clip
+        audio_len = audio.shape[-1] / sample_rate
+        if audio_len > 10.0:
+            print('current audio length is:', audio_len)
+            print('[WARNING] AudioMAE only accepts audio length up to 10s. The audio frames exceeding 10s will be clipped.')
+
+        # Check if the audio has multiple channels
+        if audio.shape[0] > 1:
+            # Convert stereo audio to mono by taking the mean across channels
+            # AudioMAE accepts a mono channel.
+            audio = torch.mean(audio, dim=0, keepdim=True)
+
+        # resample the audio into 16khz
+        # AudioMAE accepts 16khz
+        if sample_rate != 16000:
+            converter = transforms.Resample(orig_freq=sample_rate, new_freq=16000)
+            audio = converter(audio)
+        return audio
+    
+    def waveform_to_melspec(self, waveform:torch.FloatTensor):
+        # Compute the Mel spectrogram using Kaldi-compatible features
+        # the parameters are chosen as described in the audioMAE paper (4.2 implementation details)
+        mel_spectrogram = kaldi.fbank(
+            waveform, 
+            num_mel_bins=128, 
+            frame_length=25.0, 
+            frame_shift=10.0, 
+            htk_compat=True, 
+            use_energy=False,
+            sample_frequency=16000, 
+            window_type='hanning',
+            dither=0.0
+        )
+        
+        # Ensure the output shape matches 1x1024x128 by padding or trimming the time dimension
+        expected_frames = 1024  # as described in the paper
+        current_frames = mel_spectrogram.shape[0]
+        if current_frames > expected_frames:
+            mel_spectrogram = mel_spectrogram[:expected_frames, :]
+        elif current_frames < expected_frames:
+            padding = expected_frames - current_frames
+            mel_spectrogram = torch.nn.functional.pad(mel_spectrogram, (0, 0,  # (left, right) for the 1st dim
+                                                                        0, padding),  # (left, right) for the 2nd dim
+                                                                        )
+            
+        # scale
+        # as in the AudioMAE implementation [REF: https://github.com/facebookresearch/AudioMAE/blob/bd60e29651285f80d32a6405082835ad26e6f19f/dataset.py#L300]
+        mel_spectrogram = (mel_spectrogram - self.MEAN) / (self.STD * 2)  # (length, n_freq_bins) = (1024, 128)
+        return mel_spectrogram
+
+    @torch.no_grad()
+    def encode(self, file_path:str, device):
+        self.eval()
+
+        waveform = self.load_wav_file(file_path)
+        melspec = self.waveform_to_melspec(waveform)  # (length, n_freq_bins) = (1024, 128)
+        melspec = melspec[None,None,:,:]  # (1, 1, length, n_freq_bins) = (1, 1, 1024, 128)
+        z = self.forward_features(melspec.to(device)).cpu()  # (b, 1+n, d); d=768
+        z = z[:,1:,:]  # (b n d); remove [CLS], the class token
+
+        b, c, w, h = melspec.shape  # w: temporal dim; h:freq dim
+        wprime = round(w / self.patch_embed.patch_size[0])  # width in the latent space
+        hprime = round(h / self.patch_embed.patch_size[1])  # height in the latent space
+
+        # reconstruct the temporal and freq dims
+        z = rearrange(z, 'b (w h) d -> b d h w', h=hprime)  # (b d h' w')
+
+        # remove the batch dim
+        z = z[0]  # (d h' w')
+        return z  # (d h' w')
+
+
+
+class PretrainedAudioMAEEncoder(PreTrainedModel):
+    config_class = AudioMAEConfig
+
+    def __init__(self, config):
+        super().__init__(config)
+        self.encoder = AudioMAEEncoder(img_size=config.img_size, in_chans=config.in_chans, num_classes=config.num_classes)
+    
+    def forward(self, file_path:str):
+        device = self.device
+        return self.encoder.encode(file_path, device)  # (d h' w')

model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c868088d7f6f9ee8c29292bfa029a552bad781e1c59abb3692762330811bf535
+size 342607672

requirements.txt

@@ -0,0 +1,4 @@
+torch==2.4.0
+torchaudio==2.4.0
+transformers==4.44.0
+timm==1.0.8