modified my_custom_dataset and augs

edited the my_custom_dataset.py file so ssl1.py can now load data using,
rather than the opso functions i added to ssl1 previously.

I then edited my_custom_dataset to use the augmentations I used for
bomb fishing.

code/simclr-pytorch-reefs/models/my_custom_dataset.py

@@ -0,0 +1,95 @@
+import os
+import json
+import librosa
+import numpy as np
+import torch
+from torch.utils.data import Dataset
+import torch.nn.functional as F
+
+# for my transformations
+#import librosa
+from audiomentations import Compose, AddGaussianNoise, PitchShift, TimeStretch, ClippingDistortion, Gain, SevenBandParametricEQ
+
+
+def resize_mel_spectrogram(mel_spec, desired_shape=(224, 224)):
+    # Convert the 2D Mel spectrogram to 4D tensor (batch, channels, height, width)
+    mel_spec_tensor = torch.tensor(mel_spec).unsqueeze(0).unsqueeze(0)
+    # Resize
+    resized_mel_spec = F.interpolate(mel_spec_tensor, size=desired_shape, mode='bilinear', align_corners=False)
+    return resized_mel_spec.squeeze(0).squeeze(0).numpy()
+
+# augmentation
+augment_raw_audio = Compose(
+    [
+        AddGaussianNoise(min_amplitude=0.0001, max_amplitude=0.0005, p=1), # good
+        PitchShift(min_semitones=-2, max_semitones=12, p=0.5), #set values so it doesnt shift too low, rmeoving bomb signal
+        TimeStretch(p = 0.5), # defaults are fine
+        ClippingDistortion(0, 5, p = 0.5), # tested params to make sure its good
+        Gain(-10, 5, p = 0.5), # defaults are fine
+        # throws an error, so i commented it out
+        #SevenBandParametricEQ(-12, 12, p = 0.5)
+    ]
+)
+
+# Modify the load_audio_and_get_mel_spectrogram function:
+def load_audio_and_get_mel_spectrogram(filename, sr=8000, n_mels=128, n_fft=1024, hop_length=64, win_length=512):
+    y, _ = librosa.load(filename, sr=sr)
+    augmented_signal = augment_raw_audio(y, sr)
+
+    mel_spectrogram = librosa.feature.melspectrogram(y=augmented_signal, sr=sr, n_mels=n_mels, n_fft=n_fft, hop_length=hop_length, win_length=win_length)
+    mel_spectrogram_resized = resize_mel_spectrogram(mel_spectrogram)
+    return mel_spectrogram_resized
+
+
+
+class CTDataset(Dataset):
+
+    def __init__(self, cfg, split, transform):
+        '''
+            Constructor. Here, we collect and index the dataset inputs and labels.
+        '''
+        #if split == 'unlabeled':
+         #   print('This will not work unless you change the getitem function to have no labels for the unlabeled set') 
+        self.data_root = cfg['data_path']
+        self.split = split
+        self.transform = transform
+#
+
+        # index data from JSON file
+        self.data = []
+        with open(cfg['json_path'], 'r') as f:
+            json_data = json.load(f)
+            for sublist in json_data.values():
+                for entry in sublist:
+                    #print(entry)
+
+                    if entry["data_type"] == split:
+                        path = entry["file_name"]
+                        label = entry["class"]
+                        self.data.append((path, label))
+
+    def __len__(self):
+        '''
+            Returns the length of the dataset.
+        '''
+        return len(self.data)
+
+    def __getitem__(self, idx):
+        '''
+            Returns a single data point at given idx.
+            Here's where we actually load the audio and get the Mel spectrogram.
+        '''
+        audio_path, label = self.data[idx]
+
+        # load audio and get Mel spectrogram
+        mel_spectrogram = load_audio_and_get_mel_spectrogram(os.path.join(self.data_root, audio_path))
+
+        # make 3 dimensions, so shape goes from [x, y] to [3, x, y]
+        mel_spectrogram_tensor = torch.tensor(mel_spectrogram).unsqueeze(0).repeat(3, 1, 1).float()
+
+        # the old transform call, its now ditched
+        #if self.transform:
+         #   mel_spectrogram_tensor = self.transform(mel_spectrogram_tensor)
+
+        # return the objects, label is commented out for now
+        return mel_spectrogram_tensor#, label

code/simclr-pytorch-reefs/models/ssl1.py

@@ -27,13 +27,14 @@
 import pandas as pd
 from opensoundscape.preprocess.preprocessors import SpectrogramPreprocessor
 from opensoundscape.ml.datasets import AudioFileDataset
-# helper function for displaying a sample as an image
+# helper function for displaying a sample as an image 
 from opensoundscape.preprocess.utils import show_tensor, show_tensor_grid
 from opensoundscape import Action
 from opensoundscape.spectrogram import MelSpectrogram
 
 
 
+
 ##################
 def _melspec_linear_to_db(melspec):
 
@@ -131,7 +132,24 @@ def samplers(self):
 
     def prepare_data(self):
         ##self.trainset = dataset
-        ##train_transform, test_transform = self.transforms() ### comment out later??
+        train_transform = self.transforms()#, test_transform = self.transforms() ### comment out later??
+
+        if self.hparams.data == 'ROV':
+            # not used with new dataset issue fix but can leave in for now
+            cfg = {'data_path': '/mnt/ssd-cluster/ben/data/full_dataset/', #############################
+                'json_path': '/home/ben/reef-audio-representation-learning/data/dataset.json'}#### tarun : for pretraining self.trainset is the unlabeled dataset.
+            print(f'Dataset path:')
+            print(cfg['data_path'])
+
+            self.trainset = CTDataset(cfg, split='train_data', transform=train_transform)
+            #### tarun : for eval or finetuning,self.trainset is the 10percent train dataset
+            #self.trainset = CTDataset(cfg, split='train', transform=train_transform)
+            #self.testset = CTDataset(cfg, split='val', transform=test_transform)
+
+        else:
+            raise NotImplementedError
+        print(f'Number of training data samples: {len(self.trainset)}')
+
         # print('The following train transform is used:\n', train_transform)
         # print('The following test transform is used:\n', test_transform)
         # if self.hparams.data == 'cifar':
@@ -142,84 +160,97 @@ def prepare_data(self):
         #     valdir = os.path.join(self.IMAGENET_PATH, 'val')
         #     self.trainset = datasets.ImageFolder(traindir, transform=train_transform)
         #     self.testset = datasets.ImageFolder(valdir, transform=test_transform)
-        if self.hparams.data == 'ROV':
+
+
+
+
+ ##############################################     
+
+        # if self.hparams.data == 'ROV':
 
-            # not used with new dataset issue fix but can leave in for now
-            cfg = {'dataset_path': '/home/ben/data/full_dataset/', #############################
-                'json_path': '/home/ben/data/dataset.json'}
+        #     # not used with new dataset issue fix but can leave in for now
+        #     cfg = {'dataset_path': '/home/ben/data/full_dataset/', #############################
+        #         'json_path': '/home/ben/data/dataset.json'}
 
-            #cfg = {'data_root':'/root/all_ROV_crops_with_unknown/all_ROV_crops_with_unknown', 'train_label_file':'../10_percent_train_with_unknown.csv', 'val_label_file':'../5_percent_val_with_unknown.csv', 'test_label_file':'../10_percent_test_with_unknown.csv', 'unlabeled_file':'../75_percent_unlabeled_with_unknown.csv'}
-            #### tarun : for pretraining self.trainset is the unlabeled dataset.
+        #     #cfg = {'data_root':'/root/all_ROV_crops_with_unknown/all_ROV_crops_with_unknown', 'train_label_file':'../10_percent_train_with_unknown.csv', 'val_label_file':'../5_percent_val_with_unknown.csv', 'test_label_file':'../10_percent_test_with_unknown.csv', 'unlabeled_file':'../75_percent_unlabeled_with_unknown.csv'}
+        #     #### tarun : for pretraining self.trainset is the unlabeled dataset.
 
 
-            #self.trainset = CTDataset(**cfg)#, split='unlabeled', transform=train_transform) ##################################
-            #################################################
-            # fixing dataset issue we add a load of load of code here to preprocess data and make train.dataset
-            # Load the JSON data from the file
-            json_path = '/home/ben/reef-audio-representation-learning/data/dataset.json'
-            dataset_path = '/mnt/ssd-cluster/ben/data/full_dataset/'
-            with open(json_path, "r") as file:
-                data = json.load(file)
+        #     #self.trainset = CTDataset(**cfg)#, split='unlabeled', transform=train_transform) ##################################
+
+
+        #     #################################################
+        #     # fixing dataset issue we add a load of load of code here to preprocess data and make train.dataset
+        #     # Load the JSON data from the file
+        #     json_path = '/home/ben/reef-audio-representation-learning/data/dataset.json'
+        #     dataset_path = '/mnt/ssd-cluster/ben/data/full_dataset/'
+        #     with open(json_path, "r") as file:
+        #         data = json.load(file)
 
-            # Extract the list of dictionaries from the "audio" key
-            audio_data = data.get("audio", [])
+        #     # Extract the list of dictionaries from the "audio" key
+        #     audio_data = data.get("audio", [])
 
-            # Filter the list to only include entries where data_type = "train_data"
-            data = [entry for entry in audio_data if entry.get("data_type") == "train_data"]
+        #     # Filter the list to only include entries where data_type = "train_data"
+        #     data = [entry for entry in audio_data if entry.get("data_type") == "train_data"]
 
-            # Convert the filtered list into a DataFrame
-            #df = pd.DataFrame(self.data)
-            df = pd.DataFrame(data)#[:32]) to rig dataset size for testing
+        #     # Convert the filtered list into a DataFrame
+        #     #df = pd.DataFrame(self.data)
+        #     df = pd.DataFrame(data)#[:32]) to rig dataset size for testing
 
-            # Convert the list of dictionaries (which is the value of the main dictionary) into a DataFrame
-            #df = pd.DataFrame(data[list(data.keys())[0]])
-            #self.data = {k: v for k, v in data.items() if v.get("data_type") == "train_data"}
-            #df = pd.DataFrame(self.data[list(self.data.keys())[0]])
+        #     # Convert the list of dictionaries (which is the value of the main dictionary) into a DataFrame
+        #     #df = pd.DataFrame(data[list(data.keys())[0]])
+        #     #self.data = {k: v for k, v in data.items() if v.get("data_type") == "train_data"}
+        #     #df = pd.DataFrame(self.data[list(self.data.keys())[0]])
 
 
-            # Create a dataframe with just file_path and a class column (req for AudioFileDataset)
-            transformed_df = df[['file_name', 'class']].copy()
+        #     # Create a dataframe with just file_path and a class column (req for AudioFileDataset)
+        #     transformed_df = df[['file_name', 'class']].copy()
 
-            # rename 'file_name' column to 'file'
-            transformed_df.rename(columns={'file_name': 'file'}, inplace=True)
+        #     # rename 'file_name' column to 'file'
+        #     transformed_df.rename(columns={'file_name': 'file'}, inplace=True)
 
-            # set file to be the index for AudioFileDataset
-            transformed_df.set_index('file', inplace=True)
+        #     # set file to be the index for AudioFileDataset
+        #     transformed_df.set_index('file', inplace=True)
 
-            # set all classes to 1 as AudioFileDataset requires class
-            transformed_df['class'] = 1
+        #     # set all classes to 1 as AudioFileDataset requires class
+        #     transformed_df['class'] = 1
 
-            # append dataset_path to start of file_name column
-            transformed_df.index = dataset_path + transformed_df.index
-            #transformed_df.head() # for notebook
+        #     # append dataset_path to start of file_name column
+        #     transformed_df.index = dataset_path + transformed_df.index
+        #     #transformed_df.head() # for notebook
 
-            # initialize the preprocessor (forget what this does?)
-            pre = SpectrogramPreprocessor(sample_duration=1.92)
+        #     # initialize the preprocessor (forget what this does?)
+        #     pre = SpectrogramPreprocessor(sample_duration=1.92)
 
-            # initialize the dataset
-            dataset = AudioFileDataset(transformed_df, pre)
+        #     # initialize the dataset
+        #     dataset = AudioFileDataset(transformed_df, pre)
 
-            # change the bandpass from the default to 8kHz
-            dataset.preprocessor.pipeline.bandpass.set(min_f=0,max_f=8000)
+        #     # change the bandpass from the default to 8kHz
+        #     dataset.preprocessor.pipeline.bandpass.set(min_f=0,max_f=8000)
 
-            melspec_action = Action(_my_melspec)
-            melspec_bandpass_action = Action(MelSpectrogram.bandpass, min_f=0, max_f=8000)
+        #     melspec_action = Action(_my_melspec)
+        #     melspec_bandpass_action = Action(MelSpectrogram.bandpass, min_f=0, max_f=8000)
 
-            dataset.preprocessor.pipeline['to_spec'] = melspec_action
-            dataset.preprocessor.pipeline['bandpass'] = melspec_bandpass_action
-            dataset.bypass_augmentations = True ### added to stop augs
-            print(f'Total number of train samples found: {len(dataset)}')
-            #print(dataset[0].shape)
-            self.trainset = dataset
+        #     dataset.preprocessor.pipeline['to_spec'] = melspec_action
+        #     dataset.preprocessor.pipeline['bandpass'] = melspec_bandpass_action
+        #     dataset.bypass_augmentations = True ### added to stop augs
+        #     print(f'Total number of train samples found: {len(dataset)}')
+        #     #print(dataset[0].shape)
+        #     self.trainset = dataset
 
 
             ##############################################
 
             #### tarun : for eval or finetuning,self.trainset is the 10percent train dataset
             #self.trainset = CTDataset(cfg, split='train', transform=train_transform)
             ##self.testset = CTDataset(cfg, split='val', transform=test_transform)
-        else:
-            raise NotImplementedError
+        # else:
+        #     raise NotImplementedError
+####################################
+
+
+
+
 
     def dataloaders(self, iters=None):
         train_batch_sampler = self.samplers()#, test_batch_sampler = self.samplers()
@@ -401,16 +432,19 @@ def transforms(self):
             from utils.datautils import GaussianBlur
 
             im_size = 224
+            # could put my transformation code in here? should i??
+
+            # note in my_custom_dataset.py this is currently written over, so these transforms arent used
             train_transform = transforms.Compose([
                 transforms.RandomResizedCrop(
                     im_size,
                     scale=(self.hparams.scale_lower, 1.0),
                     interpolation=PIL.Image.BICUBIC,
                 ),
-               ## transforms.RandomHorizontalFlip(0.5),
+               transforms.RandomHorizontalFlip(0.5),
                ## datautils.get_color_distortion(s=self.hparams.color_dist_s),
                ## transforms.ToTensor(),
-               GaussianBlur(im_size // 10, 0.5),
+               #GaussianBlur(im_size // 10, 0.5),
                ## datautils.Clip(),
             ])
             ##test_transform = train_transform