initial commit

.gitattributes

@@ -32,3 +32,4 @@ saved_model/**/* 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
+examples/*.wav filter=lfs diff=lfs merge=lfs -text

app.py

@@ -0,0 +1,168 @@
+# AUTOGENERATED! DO NOT EDIT! File to edit: app.ipynb.
+
+# %% auto 0
+__all__ = ['data', 'audios', 'metadata', 'to_consider', 'processed_metadata', 'repo_id', 'learner', 'categories', 'title',
+           'description', 'mic', 'label', 'examples', 'intf', 'process_audio_exists', 'load_x', 'load_label_tfm',
+           'classify_audio']
+
+# %% app.ipynb 1
+import torch
+import gradio as gr
+from gradio import CSVLogger
+from fastai.vision.all import *
+import torchaudio
+import torchaudio.transforms as T
+import warnings
+from huggingface_hub import from_pretrained_fastai
+
+# %% app.ipynb 2
+warnings.filterwarnings("ignore")
+
+# %% app.ipynb 3
+def process_audio_exists(audio):
+    slice_name = audio.name
+    
+    # check if slice name exists in new metadata file
+    row = processed_metadata.loc[processed_metadata['slice_file_name'] == slice_name].index.any()
+    
+    return row
+
+# %% app.ipynb 4
+data = Path('examples')
+audios = get_files(data, extensions='.wav')
+
+metadata = pd.read_csv('UrbanSound8K.csv')
+to_consider = ['siren', 'street_music', 'children_playing', 'dog_bark', 'car_horn']
+processed_metadata = metadata.loc[metadata['class'].isin(to_consider)]
+processed_metadata.loc[processed_metadata['class'] == 'siren', 'classID'] = 4
+processed_metadata.loc[processed_metadata['class'] == 'street_music', 'classID'] = 0
+
+# %% app.ipynb 5
+class load_x(Transform):
+    def __init__(self):
+        self.sr = 44100
+        self.max_ms = 4000
+        self.channels = 2
+        # self.transform = transform
+    def rechannel(self, waveform, sr):
+        if (waveform.shape[0] == self.channels):
+            # no rechanneling needed
+            return waveform, sr
+    
+        if (self.channels==1):
+            # converting stereo to mono
+            # by selecting the first channel
+            new_waveform = waveform[:1,:]
+            
+        elif (self.channels==2):
+            # converting mono to stereo
+            # by duplicating the first channel
+            new_waveform = torch.cat([waveform, waveform])
+        return new_waveform, sr
+    
+    def resample(self, waveform, sr):
+        if (sr==self.sr):
+            # no resampling needed
+            return waveform, sr
+
+        num_channels = waveform.shape[0]
+
+        # resample first channel
+        new_waveform = torchaudio.transforms.Resample(sr, self.sr)(waveform[:1,:])
+        if (num_channels) > 1:
+            # resample second channel and merge the two
+            re_two = torchaudio.transforms.Resample(sr, self.sr)(waveform[1:,:])
+            new_waveform = torch.cat([new_waveform, re_two])
+
+        return (new_waveform, self.sr)
+    
+    def pad_trunc(self, waveform, sr):
+        num_channels, num_frames = waveform.shape
+        max_len = sr//1000 * self.max_ms
+
+        if (num_frames>max_len):
+          # truncate signal to given length
+          waveform = waveform[:,:max_len]
+
+        else:
+            # get padding lengths for beginning and end
+            begin_ln = random.randint(0, max_len-num_frames)
+            end_ln = max_len - num_frames - begin_ln
+
+            # pad the audio with zeros
+            pad_begin = torch.zeros((num_channels, begin_ln))
+            pad_end = torch.zeros((num_channels, end_ln))
+
+            waveform = torch.cat((pad_begin, waveform, pad_end), 1)
+
+        return (waveform, sr)
+    
+    def mel_specgram(self, waveform, sr):
+        mel_tfm = T.MelSpectrogram(
+            sample_rate=sr,
+            n_fft=1024,
+            win_length=None,
+            hop_length=512,
+            center=True,
+            pad_mode="reflect",
+            power=2.0,
+            norm="slaney",
+            onesided=True,
+            n_mels=128,
+            mel_scale="htk")
+        
+        spec = mel_tfm(waveform)
+        
+        waveform = torchaudio.transforms.AmplitudeToDB(top_db=80)(spec)
+        
+        return waveform, sr
+        
+    
+    def encodes(self, x):
+        waveform, sr = torchaudio.load(x)
+        waveform, sr = self.resample(waveform, sr)
+        waveform, sr = self.pad_trunc(waveform, sr)
+        waveform, sr = self.rechannel(waveform, sr)
+        waveform, sr = self.mel_specgram(waveform, sr)
+        return waveform
+
+    
+class load_label_tfm(Transform):
+    def __init__(self, metadata=processed_metadata): self.metadata = metadata
+    def encodes(self, x): 
+        return self.metadata.loc[self.metadata['slice_file_name'] == x.name]['class'].item()
+
+# %% app.ipynb 6
+repo_id = "Jimmie/urban8k"
+
+learner = from_pretrained_fastai(repo_id)
+
+# %% app.ipynb 14
+categories = tuple(learner.dls.vocab)
+
+def classify_audio(audio):
+    # use Path to open audio
+    audio_path = Path(audio)
+    pred,idx,probs = learner.predict(audio_path)
+    return dict(zip(categories, map(float, probs)))
+
+# %% app.ipynb 16
+title = "Environmental Sound Classification"
+
+description = """
+This demo showcases how AI can be used to recognize environmental sounds. It focuses specifically on 5 classes: car_horn, children_playing, dog_bark, siren and street music
+ 
+
+When uploading audio, make sure it is in .wav format and is less than 4 seconds long.
+
+Enjoy!
+"""
+mic = gr.Audio(source='upload', type="filepath", label='Upload Audio File here')
+label = gr.outputs.Label()
+examples = list(data.ls())
+
+intf = gr.Interface(fn=classify_audio, inputs=mic, outputs=label, examples=examples,
+                    title=title, description=description,  cache_examples=False, 
+                    auto_submit_duration=5)
+
+intf.launch(inline=False)

examples/car_horn.wav

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:390a545a53dfe52f47a92876691eb40e64d1240c8885be7f72df3654b8fe70f8
+size 705644

examples/children_playing.wav

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:0a01b7b6f6e9d51a57a7abf1128518c68631f3c7095736f0364479c813e07ab8
+size 768044

examples/dog_bark.wav

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:fefae5223783da73b535df8815dea61a285f444f0770228c9d9ec8ea5a2e65c7
+size 705644

examples/siren.wav

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c7b54c8d4a92dbd21fdbe5ba3027a289fe2c4f636d14bacf7205b07543e26f78
+size 768044

examples/street_music.wav

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d976eeb884ede8c4c731bf616e197a40a7a9ecef47b9005e2a1f6acaec8888c3
+size 1152080

requirements.txt

@@ -0,0 +1,2 @@
+fastai<=2.7.11
+torchaudio<=0.13.1