def __getitem__()

in data/audioVisual_dataset.py [0:0]

• 20 lines of code
• 2 McCabe index (conditional complexity)

    def __getitem__(self, index):
        #load audio
        audio, audio_rate = librosa.load(self.audios[index], sr=self.opt.audio_sampling_rate, mono=False)

        #randomly get a start time for the audio segment from the 10s clip
        audio_start_time = random.uniform(0, 9.9 - self.opt.audio_length)
        audio_end_time = audio_start_time + self.opt.audio_length
        audio_start = int(audio_start_time * self.opt.audio_sampling_rate)
        audio_end = audio_start + int(self.opt.audio_length * self.opt.audio_sampling_rate)
        audio = audio[:, audio_start:audio_end]
        audio = normalize(audio)
        audio_channel1 = audio[0,:]
        audio_channel2 = audio[1,:]

        #get the frame dir path based on audio path
        path_parts = self.audios[index].strip().split('/')
        path_parts[-1] = path_parts[-1][:-4] + '.mp4'
        path_parts[-2] = 'frames'
        frame_path = '/'.join(path_parts)

        # get the closest frame to the audio segment
        #frame_index = int(round((audio_start_time + audio_end_time) / 2.0 + 0.5))  #1 frame extracted per second
        frame_index = int(round(((audio_start_time + audio_end_time) / 2.0 + 0.05) * 10))  #10 frames extracted per second
        frame = process_image(Image.open(os.path.join(frame_path, str(frame_index).zfill(6) + '.png')).convert('RGB'), self.opt.enable_data_augmentation)
        frame = self.vision_transform(frame)

        #passing the spectrogram of the difference
        audio_diff_spec = torch.FloatTensor(generate_spectrogram(audio_channel1 - audio_channel2))
        audio_mix_spec = torch.FloatTensor(generate_spectrogram(audio_channel1 + audio_channel2))

        return {'frame': frame, 'audio_diff_spec':audio_diff_spec, 'audio_mix_spec':audio_mix_spec}