import torch import time import librosa import numpy as np from ssr_eval.eval import SSR_Eval_Helper, BasicTestee from model import NGAN #from load_balanced_model import NGAN import scipy.signal from scipy.interpolate import interp1d, CubicSpline class MyTestee(BasicTestee): def __init__(self, device) -> None: super().__init__() self.device = device self.model = NGAN( batch_size=1, target_sample_rate=48000, segment_length=24000, padding='same').to(device) #self.checkpoint = torch.load('generators/w4.ckpt') #To load load balanced model self.checkpoint = torch.load('generators/n8.ckpt') #For best performing 8 channel model def cubic_upsampler(self, audio_signal, input_rate, target_rate): original_duration = audio_signal.size / input_rate original_time = np.linspace(0, original_duration, audio_signal.size) target_time = np.linspace(0, original_duration, int(original_duration * target_rate)) cubic_spline = CubicSpline(original_time, audio_signal, axis=-1) upsampled_audio = cubic_spline(target_time) return upsampled_audio def infer(self, x): x = self.cubic_upsampler(x, 8000, 48000) #for 8KHz input x = torch.tensor(x.copy(), dtype=torch.float32).to(self.device) x = x.unsqueeze(dim=0) j = [x, x] self.model.superResolver.load_state_dict(self.checkpoint) with torch.no_grad(): y = self.model.superResolver(j).squeeze(dim=0) out = self.tensor2numpy(y)[0] return out if __name__ == "__main__": if(torch.cuda.is_available()): device = "cuda" else: device="cpu" testee = MyTestee(device) #"butter", "cheby1", "ellip", "bessel" helper = SSR_Eval_Helper(testee, test_name="8x48_performance", test_data_root="test_data", input_sr=8000, output_sr=48000, evaluation_sr=48000, setting_lowpass_filtering = { "filter":["cheby", "bessel", "ellip", "butter"], "cutoff_freq": [4000], "filter_order": [8] }, save_processed_result=False, ) helper.evaluate(limit_test_nums=-1, limit_test_speaker=-1)