xref: /aosp_15_r20/external/libopus/dnn/torch/fwgan/inference.py (revision a58d3d2adb790c104798cd88c8a3aff4fa8b82cc)
1import os
2import time
3import torch
4import numpy as np
5from scipy import signal as si
6from scipy.io import wavfile
7import argparse
8
9from models import model_dict
10
11parser = argparse.ArgumentParser()
12parser.add_argument('model', choices=['fwgan400', 'fwgan500'], help='model name')
13parser.add_argument('weightfile', type=str, help='weight file')
14parser.add_argument('input', type=str, help='input: feature file or folder with feature files')
15parser.add_argument('output', type=str, help='output: wav file name or folder name, depending on input')
16
17
18########################### Signal Processing Layers ###########################
19
20def preemphasis(x, coef= -0.85):
21
22    return si.lfilter(np.array([1.0, coef]), np.array([1.0]), x).astype('float32')
23
24def deemphasis(x, coef= -0.85):
25
26    return si.lfilter(np.array([1.0]), np.array([1.0, coef]), x).astype('float32')
27
28gamma = 0.92
29weighting_vector = np.array([gamma**i for i in range(16,0,-1)])
30
31
32def lpc_synthesis_one_frame(frame, filt, buffer, weighting_vector=np.ones(16)):
33
34    out = np.zeros_like(frame)
35
36    filt = np.flip(filt)
37
38    inp = frame[:]
39
40
41    for i in range(0, inp.shape[0]):
42
43        s = inp[i] - np.dot(buffer*weighting_vector, filt)
44
45        buffer[0] = s
46
47        buffer = np.roll(buffer, -1)
48
49        out[i] = s
50
51    return out
52
53def inverse_perceptual_weighting (pw_signal, filters, weighting_vector):
54
55    #inverse perceptual weighting= H_preemph / W(z/gamma)
56
57    pw_signal = preemphasis(pw_signal)
58
59    signal = np.zeros_like(pw_signal)
60    buffer = np.zeros(16)
61    num_frames = pw_signal.shape[0] //160
62    assert num_frames == filters.shape[0]
63
64    for frame_idx in range(0, num_frames):
65
66        in_frame = pw_signal[frame_idx*160: (frame_idx+1)*160][:]
67        out_sig_frame = lpc_synthesis_one_frame(in_frame, filters[frame_idx, :], buffer, weighting_vector)
68        signal[frame_idx*160: (frame_idx+1)*160] = out_sig_frame[:]
69        buffer[:] = out_sig_frame[-16:]
70
71    return signal
72
73
74def process_item(generator, feature_filename, output_filename, verbose=False):
75
76    feat = np.memmap(feature_filename, dtype='float32', mode='r')
77
78    num_feat_frames = len(feat) // 36
79    feat = np.reshape(feat, (num_feat_frames, 36))
80
81    bfcc = np.copy(feat[:, :18])
82    corr = np.copy(feat[:, 19:20]) + 0.5
83    bfcc_with_corr =  torch.from_numpy(np.hstack((bfcc, corr))).type(torch.FloatTensor).unsqueeze(0)#.to(device)
84
85    period = torch.from_numpy((0.1 + 50 * np.copy(feat[:, 18:19]) + 100)\
86                            .astype('int32')).type(torch.long).view(1,-1)#.to(device)
87
88    lpc_filters = np.copy(feat[:, -16:])
89
90    start_time = time.time()
91    x1 = generator(period, bfcc_with_corr, torch.zeros(1,320)) #this means the vocoder runs in complete synthesis mode with zero history audio frames
92    end_time = time.time()
93    total_time = end_time - start_time
94    x1 = x1.squeeze(1).squeeze(0).detach().cpu().numpy()
95    gen_seconds = len(x1)/16000
96    out = deemphasis(inverse_perceptual_weighting(x1, lpc_filters, weighting_vector))
97    if verbose:
98        print(f"Took {total_time:.3f}s to generate {len(x1)}  samples ({gen_seconds}s) -> {gen_seconds/total_time:.2f}x real time")
99
100    out = np.clip(np.round(2**15 * out), -2**15, 2**15 -1).astype(np.int16)
101    wavfile.write(output_filename, 16000, out)
102
103
104########################### The inference loop over folder containing lpcnet feature files #################################
105if __name__ == "__main__":
106
107    args = parser.parse_args()
108
109    generator = model_dict[args.model]()
110
111
112    #Load the FWGAN500Hz Checkpoint
113    saved_gen= torch.load(args.weightfile, map_location='cpu')
114    generator.load_state_dict(saved_gen)
115
116    #this is just to remove the weight_norm from the model layers as it's no longer needed
117    def _remove_weight_norm(m):
118        try:
119            torch.nn.utils.remove_weight_norm(m)
120        except ValueError:  # this module didn't have weight norm
121            return
122    generator.apply(_remove_weight_norm)
123
124    #enable inference mode
125    generator = generator.eval()
126
127    print('Successfully loaded the generator model ... start generation:')
128
129    if os.path.isdir(args.input):
130
131        os.makedirs(args.output, exist_ok=True)
132
133        for fn in os.listdir(args.input):
134            print(f"processing input {fn}...")
135            feature_filename = os.path.join(args.input, fn)
136            output_filename = os.path.join(args.output, os.path.splitext(fn)[0] + f"_{args.model}.wav")
137            process_item(generator, feature_filename, output_filename)
138    else:
139        process_item(generator, args.input, args.output)
140
141    print("Finished!")