Set up BirdClef

.gitignore

@@ -149,3 +149,12 @@ checklink/cookies.txt
 
 # Quarto
 .quarto
+
+# Data folder
+data/
+
+# wandb folder
+wandb/
+
+# weight files
+**.pth

README.md

@@ -1,4 +1,4 @@
-# dm-4-cat
+# birdclef_2023
 
 <!-- WARNING: THIS FILE WAS AUTOGENERATED! DO NOT EDIT! -->
 
@@ -8,7 +8,7 @@ documentation.
 ## Install
 
 ``` sh
-pip install dm_4_cat
+pip install birdclef_2023
 ```
 
 ## How to use

birdclef/_modidx.py

@@ -0,0 +1,39 @@
+# Autogenerated by nbdev
+
+d = { 'settings': { 'branch': 'master',
+                'doc_baseurl': '/birdclef_2023',
+                'doc_host': 'https://Chavelanda.github.io',
+                'git_url': 'https://github.com/Chavelanda/birdclef_2023',
+                'lib_path': 'birdclef'},
+  'syms': { 'birdclef.dataset': { 'birdclef.dataset.BirdClef': ('dataset.html#birdclef', 'birdclef/dataset.py'),
+                                  'birdclef.dataset.BirdClef.__getitem__': ('dataset.html#birdclef.__getitem__', 'birdclef/dataset.py'),
+                                  'birdclef.dataset.BirdClef.__init__': ('dataset.html#birdclef.__init__', 'birdclef/dataset.py'),
+                                  'birdclef.dataset.BirdClef.__len__': ('dataset.html#birdclef.__len__', 'birdclef/dataset.py'),
+                                  'birdclef.dataset.MyPipeline': ('dataset.html#mypipeline', 'birdclef/dataset.py'),
+                                  'birdclef.dataset.MyPipeline.__init__': ('dataset.html#mypipeline.__init__', 'birdclef/dataset.py'),
+                                  'birdclef.dataset.MyPipeline.forward': ('dataset.html#mypipeline.forward', 'birdclef/dataset.py'),
+                                  'birdclef.dataset.get_dataloader': ('dataset.html#get_dataloader', 'birdclef/dataset.py'),
+                                  'birdclef.dataset.get_dataset': ('dataset.html#get_dataset', 'birdclef/dataset.py')},
+            'birdclef.experiment': {},
+            'birdclef.network': { 'birdclef.network.EfficientNetV2': ('network.html#efficientnetv2', 'birdclef/network.py'),
+                                  'birdclef.network.EfficientNetV2.__init__': ( 'network.html#efficientnetv2.__init__',
+                                                                                'birdclef/network.py'),
+                                  'birdclef.network.EfficientNetV2.forward': ('network.html#efficientnetv2.forward', 'birdclef/network.py'),
+                                  'birdclef.network.get_model': ('network.html#get_model', 'birdclef/network.py')},
+            'birdclef.preprocessing': {'birdclef.preprocessing.foo': ('preprocessing.html#foo', 'birdclef/preprocessing.py')},
+            'birdclef.trainer': { 'birdclef.trainer.log_weights': ('trainer.html#log_weights', 'birdclef/trainer.py'),
+                                  'birdclef.trainer.train': ('trainer.html#train', 'birdclef/trainer.py'),
+                                  'birdclef.trainer.train_one_epoch': ('trainer.html#train_one_epoch', 'birdclef/trainer.py'),
+                                  'birdclef.trainer.validate_model': ('trainer.html#validate_model', 'birdclef/trainer.py')},
+            'birdclef.training_utils': { 'birdclef.training_utils.compute_metrics': ( 'training_utils.html#compute_metrics',
+                                                                                      'birdclef/training_utils.py'),
+                                         'birdclef.training_utils.get_loss_func': ( 'training_utils.html#get_loss_func',
+                                                                                    'birdclef/training_utils.py'),
+                                         'birdclef.training_utils.get_optimizer': ( 'training_utils.html#get_optimizer',
+                                                                                    'birdclef/training_utils.py')},
+            'birdclef.utils': { 'birdclef.utils.mel_to_wave': ('utils.html#mel_to_wave', 'birdclef/utils.py'),
+                                'birdclef.utils.plot_audio': ('utils.html#plot_audio', 'birdclef/utils.py'),
+                                'birdclef.utils.plot_fbank': ('utils.html#plot_fbank', 'birdclef/utils.py'),
+                                'birdclef.utils.plot_specgram': ('utils.html#plot_specgram', 'birdclef/utils.py'),
+                                'birdclef.utils.plot_spectrogram': ('utils.html#plot_spectrogram', 'birdclef/utils.py'),
+                                'birdclef.utils.plot_waveform': ('utils.html#plot_waveform', 'birdclef/utils.py')}}}

birdclef/dataset.py

@@ -0,0 +1,176 @@
+# AUTOGENERATED! DO NOT EDIT! File to edit: ../nbs/02_dataset.ipynb.
+
+# %% auto 0
+__all__ = ['dir', 'simple_classes', 'train_metadata_simple', 'val_metadata_simple', 'test_metadata_simple', 'dataset_dict',
+           'MyPipeline', 'BirdClef', 'get_dataset', 'get_dataloader']
+
+# %% ../nbs/02_dataset.ipynb 3
+from IPython.display import Audio
+import pandas as pd
+from sklearn.preprocessing import LabelBinarizer
+
+import torch
+from torch.utils.data import Dataset, DataLoader
+import torchaudio
+
+from .utils import DATA_DIR, AUDIO_DATA_DIR, mel_to_wave, plot_audio, plot_spectrogram
+
+# %% ../nbs/02_dataset.ipynb 6
+# Define custom feature extraction pipeline.
+#
+# 1. Check for sample rate and resample
+# 2. Waveform Augmenations
+# 3. Convert to mel-scale
+# 4. Mel Augmenations
+# 5. Check for lenght and stretch shorter videos
+
+
+class MyPipeline(torch.nn.Module):
+    def __init__(
+        self,
+        c_length = 10,
+        sample_rate=32000,
+        f_min = 40,
+        f_max = 15000,
+        n_fft=2048,
+        n_mels=128,
+        hop_length = 512,
+        power = 2.0
+    ):
+        super().__init__()
+
+        self.c_length = c_length * 62.6 #626 sono 10 secondi
+        self.sample_rate = sample_rate
+        self.melspec = torchaudio.transforms.MelSpectrogram(sample_rate=self.sample_rate, n_fft=n_fft, hop_length=hop_length, n_mels=n_mels, f_min=f_min, f_max=f_max, power=power)
+        self.amptodb = torchaudio.transforms.AmplitudeToDB()
+        self.stretch = torchaudio.transforms.TimeStretch(hop_length=hop_length, n_freq=128)
+
+        #Augmentations
+        # self.maskingFreq =  torchaudio.transforms.FrequencyMasking(freq_mask_param=30)
+        # self.maskingTime = torchaudio.transforms.TimeMasking(time_mask_param=30)
+        # self.noiser = torchaudio.transforms.AddNoise()
+        # self.pitchShift = torchaudio.transforms.PitchShift(resample_freq, 4)
+
+
+    def forward(self, filename):
+        # 0 Load the File
+        waveform, sample_rate = torchaudio.load(filename, frame_offset=0, num_frames=320000)
+
+        # 1 Check for the sample rate and eventually resample to 32k
+        if sample_rate != self.sample_rate:
+           print("Wrong sample rate: resampling audio")
+           resampler = torchaudio.transforms.Resample(orig_freq=sample_rate, new_freq=self.sample_rate)
+           waveform = resampler(waveform)
+
+        # # 2 Waveform Augmentation
+        # #2.1 White noise
+        # if random.randint(0,1) < 0.3:
+        #   noise = torch.rand(1, 320000)
+        #   noise = (noise - 0.5) * 0.2
+        #   snr_dbs = torch.tensor([random.randint(2,8)])
+        #   waveform = self.noiser(waveform, noise, snr_dbs)
+
+        # # 2.2 Pitch Shift
+        # if random.randint(0,1) < 1:
+        # if True:
+        #   waveform = self.pitchShift(waveform)
+
+
+        # 3 Convert to mel-scale
+        mel = self.melspec(waveform)
+        mel = self.amptodb(mel)
+
+        # 4 Mel Augmenations
+        # 4.1 Frequency Masking
+        # if True:
+        #   mel = self.maskingFreq(mel)
+        # # 4.2 Time Masking
+        # if True:
+        #   mel = self.maskingTime(mel)
+
+
+        # 5 Check for the length and stretch it to 10s, it is a transformation used to regularize the length of the data
+        if mel.shape[2] < self.c_length:
+          print("Audio too short: stretching it.")
+          replay_rate =  mel.shape[2]/self.c_length
+          #print(f"replay rate {replay_rate}%")
+          mel = self.stretch(mel, replay_rate)
+          mel = mel[:,:,0:626]
+          #print(f"stretched shape {stretched.shape}")
+
+        return mel.float()
+
+# %% ../nbs/02_dataset.ipynb 9
+class BirdClef(Dataset):
+
+    def __init__(self, metadata=None, classes=None):
+
+        self.metadata = metadata
+        self.classes = classes
+
+        self.length = len(self.metadata)
+
+        binarizer = LabelBinarizer()
+        binarizer.fit(self.classes)
+
+        self.labels = binarizer.transform(metadata.primary_label)
+
+        # Initialize a pipeline
+        self.pipeline = MyPipeline()
+
+    def __len__(self):
+        return self.length
+
+    def __getitem__(self, idx):
+        filename = AUDIO_DATA_DIR + self.metadata['filename'][idx]
+        mel_spectrogram = self.pipeline(filename)
+
+        label = self.labels[idx]
+        label = torch.from_numpy(label).float()
+
+        return mel_spectrogram, label
+
+# %% ../nbs/02_dataset.ipynb 13
+dir = DATA_DIR
+try:
+    train_metadata_base = pd.read_csv(dir + 'base/train_metadata.csv')
+    val_metadata_base = pd.read_csv(dir + 'base/val_metadata.csv')
+    test_metadata_base = pd.read_csv(dir + 'base/test_metadata.csv')
+except FileNotFoundError:
+    dir = 'data/'
+    train_metadata_base = pd.read_csv(dir + 'base/train_metadata.csv')
+    val_metadata_base = pd.read_csv(dir + 'base/val_metadata.csv')
+    test_metadata_base = pd.read_csv(dir + 'base/test_metadata.csv')
+
+simple_classes = ['thrnig1', 'wlwwar', 'barswa']
+train_metadata_simple = train_metadata_base.loc[train_metadata_base.primary_label.isin(simple_classes)].reset_index()
+val_metadata_simple = val_metadata_base.loc[val_metadata_base.primary_label.isin(simple_classes)].reset_index()
+test_metadata_simple = test_metadata_base.loc[test_metadata_base.primary_label.isin(simple_classes)].reset_index()
+
+# %% ../nbs/02_dataset.ipynb 14
+dataset_dict = {
+            'train_base': (BirdClef, {'metadata': train_metadata_base, 'classes': train_metadata_base.primary_label}),
+            'val_base': (BirdClef, {'metadata': val_metadata_base, 'classes': train_metadata_base.primary_label}),
+            'test_base': (BirdClef, {'metadata': test_metadata_base, 'classes': train_metadata_base.primary_label}),
+
+            'train_simple': (BirdClef, {'metadata': train_metadata_simple, 'classes': train_metadata_simple.primary_label}),
+            'val_simple': (BirdClef, {'metadata': val_metadata_simple, 'classes': train_metadata_simple.primary_label}),
+            'test_simple': (BirdClef, {'metadata': test_metadata_simple, 'classes': train_metadata_simple.primary_label})
+        }
+
+# %% ../nbs/02_dataset.ipynb 15
+def get_dataset(dataset_key:str        # A key of the dataset dictionary
+                )->Dataset:         # Pytorch dataset
+    "A getter method to retrieve the wanted dataset."
+    assert dataset_key in dataset_dict, f'{dataset_key} is not an existing dataset, choose one from {dataset_dict.keys()}.'
+    ds_class, kwargs = dataset_dict[dataset_key]
+    return ds_class(**kwargs)
+
+# %% ../nbs/02_dataset.ipynb 19
+def get_dataloader(dataset_key:str,            # The key to access the dataset
+                dataloader_kwargs:dict={}      # The optional parameters for a pytorch dataloader
+                )->DataLoader:              # Pytorch dataloader
+    "A function to get a dataloader from a specific dataset"
+    dataset = get_dataset(dataset_key)
+
+    return DataLoader(dataset, **dataloader_kwargs)

birdclef/experiment.py

@@ -0,0 +1,9 @@
+# AUTOGENERATED! DO NOT EDIT! File to edit: ../nbs/06_experiment.ipynb.
+
+# %% auto 0
+__all__ = []
+
+# %% ../nbs/06_experiment.ipynb 4
+import wandb
+
+from .trainer import train

birdclef/network.py

@@ -0,0 +1,56 @@
+# AUTOGENERATED! DO NOT EDIT! File to edit: ../nbs/03_network.ipynb.
+
+# %% auto 0
+__all__ = ['model_dict', 'EfficientNetV2', 'get_model']
+
+# %% ../nbs/03_network.ipynb 3
+from typing import Union, BinaryIO, IO
+from os import PathLike
+
+import torch
+import torchvision
+from torch.nn import Module
+
+from .dataset import get_dataloader
+
+
+# %% ../nbs/03_network.ipynb 5
+class EfficientNetV2(torch.nn.Module):
+    def __init__(self, num_classes=264, size='s'):
+        super().__init__()
+
+        if size=='s':
+            self.efficientnet_v2 = torchvision.models.efficientnet_v2_s(weights=None, progress=True, num_classes=num_classes)
+        elif size=='m':
+            self.efficientnet_v2 = torchvision.models.efficientnet_v2_m(weights=None, progress=True, num_classes=num_classes)
+        else:
+            self.efficientnet_v2 = torchvision.models.efficientnet_v2_l(weights=None, progress=True, num_classes=num_classes)
+
+        self.init_conv = torch.nn.Conv2d(1, 3, (3,3), padding="same")
+        #self.sigmoid = torch.nn.functional.sigmoid
+
+    def forward(self, x):
+        x = self.init_conv(x)
+        x = self.efficientnet_v2(x)
+
+        return x
+
+# %% ../nbs/03_network.ipynb 10
+model_dict = {
+        'simple_efficient_net_v2_s': (EfficientNetV2, {'num_classes': 3}),
+        'efficient_net_v2_s': (EfficientNetV2, {})
+        }
+
+def get_model(model_key:str, # A key of the model dictionary
+              weights_path:Union[str, PathLike, BinaryIO, IO[bytes]] = None   # A file-like object to the model weights
+              )->Module:      # A pytorch model
+    "A getter method to retrieve the wanted (possibly pretrained) model"
+    assert model_key in model_dict, f'{model_key} is not an existing network, choose one from {model_dict.keys()}.'
+
+    net_class, kwargs = model_dict[model_key]
+    model = net_class(**kwargs)
+
+    if weights_path is not None:
+        model.load_state_dict(torch.load(weights_path))
+
+    return model