traeumerAI/train_transfer.py at master · jdasam/traeumerAI · GitHub

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
import torch
import pickle
import argparse
import numpy as np
import librosa
from audio_model import SiameseNet, TransferNet, HParams
from music_embedder import extractor


def load_audio_model(args):
    # with open("hparams.dat", "rb") as f:
    #     hparams = pickle.load(f)
    hparams = HParams()
    audio_embedder = SiameseNet(hparams)
    audio_checkpoint = torch.load(args.ckpt)
    audio_embedder.load_state_dict(audio_checkpoint["state_dict"])
    return audio_embedder

def load_audio_model_and_get_embedding(audio, model_types):
    # audio_length = audio.shape[1]
    input_length, model, checkpoint_path = extractor.load_model(model_types)
    # audio = extractor.make_frames_of_batch(audio, input_length, target_fps=1/3)[:,1,:]
    audio = extractor.make_frames_of_batch(audio, input_length, target_fps=0.5).view(-1, input_length)
    # audio = extractor.make_audio_batch(audio, input_length)

    state_dict = torch.load("music_embedder/"+checkpoint_path, map_location=torch.device('cpu'))

    new_state_map = {model_key: model_key.split("model.")[1] for model_key in state_dict.get("state_dict").keys()}
    new_state_dict = {new_state_map[key]: value for (key, value) in state_dict.get("state_dict").items() if key in new_state_map.keys()}
    model.load_state_dict(new_state_dict)

    audio = audio.to('cuda')
    model = model.to('cuda')

    with torch.no_grad():
        model.eval()
        if "CPC" in model_types:
            embeddings = torch.zeros(audio.shape[0], 256)
            for i in range(0, audio.shape[0], 100):
                _, embeddings[i:i+100] =model.get_emb(audio[i:i+100])
        else:
            embeddings = model.get_emb(audio)
    return embeddings

def train(args, device):
    if "siamese" in args.model_code:
        audio_embedder = load_audio_model(args).to(device)
        embd_size = audio_embedder.conv_size
    elif "FCN037" in args.model_code:
        embd_size = 512
    elif "CPC" in args.model_code:
        embd_size = 256
    else:
        embd_size = 64
    model = TransferNet(embd_size, args.style_dim)
    style_stats = torch.load("style_latent_stat.pt")
    # style_stats = torch.load("ocean_image_stat_half_std.pt")
    model.bias = torch.nn.Parameter(style_stats['mean'].squeeze(0), requires_grad=False)
    model.std = torch.nn.Parameter(style_stats['std'].squeeze(0), requires_grad=False)
    model = model.to(device)
    learning_rate = args.learning_rate
    optimizer = torch.optim.Adam(model.parameters(), lr=learning_rate,
                                 weight_decay=args.weight_decay)
    # loss_fn = torch.nn.MSELoss()
    loss_fn = torch.nn.L1Loss()

    data = np.load(args.data_path, allow_pickle=True)
    # embs = [audio_embedder.inference_with_audio(x['audio'])[0] for x in data]
    styles = [x['style'] for x in data]
    audios = [x['audio'] for x in data]
    # audios = [librosa.core.resample(x['audio'], 44100, 16000) for x in data]


    # mels = [librosa.feature.melspectrogram(y=x, sr=16000, n_fft=512, hop_length=256, n_mels=48) for x in audios]
    # mels = torch.Tensor(mels).to(device)
    # audio_embedder = load_audio_model(args).to(device)
    # audio_embedder.eval()
    # with torch.no_grad():
    #     embs = audio_embedder.cnn.fwd_wo_pool(mels)
    # audios =torch.Tensor([librosa.core.resample(x['audio'], 44100, 16000) for x in data])
    # embs = torch.Tensor(embs).to(device)

    if "siamese" in args.model_code:
        # mel_basis = librosa.filters.mel(16000, n_fft=512, n_mels=48)
        # spec = np.asarray([librosa.stft(x, n_fft=512, hop_length=256, win_length=512, window='hann') for x in audios ])
        # mel_spec = np.dot(mel_basis, librosa.core.amplitude_to_db(np.abs(spec)).transpose(2,1,0)).transpose(2,0,1)
        # mel_spec = np.dot(mel_basis, np.abs(spec.transpose(2,1,0))).transpose(2,0,1)
        mel_spec = np.asarray([librosa.feature.melspectrogram(x, sr=16000,n_mels=48, n_fft=512, hop_length=256, win_length=512, window='hann') for x in audios ])
        mel_spec = mel_spec / 80 + 0.5

        mels = torch.Tensor(mel_spec).to(device)
        # audio_embedder = load_audio_model(args).to(device)
        audio_embedder.eval()
        with torch.no_grad():
            embs = audio_embedder.infer_mid_level(mels, max_pool=False).permute(0,2,1)
        # embs = torch.Tensor(embs).to(device)
    else:
        audios = torch.Tensor(audios)
        embs = load_audio_model_and_get_embedding(audios, model_types=args.model_code).to(device)
        embs = embs.view(audios.shape[0], -1, embs.shape[-1])

    styles = torch.Tensor(styles).to(device).unsqueeze(1).repeat(1,embs.shape[1],1)


    model.train()
    for i in range(args.epoch):
        model.zero_grad()
        transfer_style = model(embs)
        loss = loss_fn(transfer_style, styles)
        loss.backward()
        print('Step {}: Loss value is {}'.format(i, loss.item()))
        optimizer.step()

    torch.save({'iteration': i,
            'state_dict': model.state_dict(),
            'optimizer': optimizer.state_dict(),
            'learning_rate': learning_rate}, 'tf_tanh_L1_{}_it{}_lr{}.pt'.format(args.model_code, args.epoch, args.learning_rate))


if __name__ == "__main__":
    device = "cuda"

    parser = argparse.ArgumentParser(description="Generate samples from the generator")
    parser.add_argument("--style_dim",type=int, default=512,help="style_dim",)
    parser.add_argument("--epoch",type=int, default=30000,help="num training epochs")
    parser.add_argument("--learning_rate",type=float, default=1e-5)
    parser.add_argument("--weight_decay",type=float, default=1e-6)
    parser.add_argument("--model_code",type=str,
                            # default="artist_siamese")
                             default="FCN037")
    parser.add_argument(
        "--ckpt",
        type=str,
        default="checkpoint_best",
        help="path to the audio model checkpoint",
    )
    parser.add_argument(
        "--data_path",
        type=str,
        default="label_data_with_16kHz_audio.npy",
        help="label pair path",
    )
    # torch.cuda.set_device(1)
    args = parser.parse_args()
    train(args, device)