WaveStitch/training_utils.py at master · adis98/WaveStitch · 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
146
147
148
import pandas as pd
import torch

from data_utils import datasets, CyclicEncoder, Preprocessor
from copy import deepcopy
import argparse
import numpy as np
import torch.optim as optim
from torch import nn, from_numpy
from torch.utils.data import Dataset, DataLoader
from TSImputers.SSSDS4Imputer import SSSDS4Imputer, SSSDS4Weaver, SSSDS4ImputerClassic
from TSImputers.TimeGAN import TimeGAN


class MyDataset(Dataset):
    def __init__(self, inputs):
        self.inputs = inputs

    def __len__(self):
        return len(self.inputs)

    def __getitem__(self, idx):
        return self.inputs[idx]


def fetchModel(in_features, out_features, args):
    model = None

    if args.backbone.lower() == 's4':
        model = SSSDS4Imputer(in_features, args.res_channels, args.skip_channels,
                              out_features, args.num_res_layers, args.diff_step_embed_in,
                              args.diff_step_embed_mid, args.diff_step_embed_out,
                              args.s4_lmax, args.s4_dstate, args.s4_dropout,
                              args.s4_bidirectional, args.s4_layernorm)

    elif args.backbone.lower() == 's4classic':
        model = SSSDS4ImputerClassic(in_features, args.res_channels, args.skip_channels,
                              out_features, args.num_res_layers, args.diff_step_embed_in,
                              args.diff_step_embed_mid, args.diff_step_embed_out,
                              args.s4_lmax, args.s4_dstate, args.s4_dropout,
                              args.s4_bidirectional, args.s4_layernorm)
    elif args.backbone.lower() == 's4weaver':
        model = SSSDS4Weaver(in_features, args.res_channels, args.skip_channels, out_features,
                             args.num_res_layers, args.diff_step_embed_in,
                             args.diff_step_embed_mid, args.diff_step_embed_out,
                             args.s4_lmax, args.s4_dstate, args.s4_dropout,
                             args.s4_bidirectional, args.s4_layernorm)

    elif args.backbone.lower() == 'timegan':
        model = TimeGAN(args)
    return model


def fetchDiffusionConfig(args):
    betas = np.linspace(args.beta_0, args.beta_T, args.timesteps).reshape((-1, 1))
    alphas = 1 - betas
    alpha_bars = np.cumprod(alphas).reshape((-1, 1))
    diffusion_config = {'betas': from_numpy(betas).float(), 'alpha_bars': from_numpy(alpha_bars).float(),
                        'alphas': from_numpy(alphas).float(), 'T': args.timesteps}
    return diffusion_config


if __name__ == "__main__":
    parser = argparse.ArgumentParser()
    parser.add_argument('-dataset', '-d', type=str,
                        help='MetroTraffic, BeijingAirQuality, AustraliaTourism, WebTraffic, StoreItems', required=True)
    parser.add_argument('-backbone', type=str, help='Transformer, Bilinear, Linear, S4', default='Transformer')
    parser.add_argument('-beta_0', type=float, default=0.0001, help='initial variance schedule')
    parser.add_argument('-beta_T', type=float, default=0.02, help='last variance schedule')
    parser.add_argument('-timesteps', '-T', type=int, default=200, help='training/inference timesteps')
    parser.add_argument('-hdim', type=int, default=128, help='hidden embedding dimension')
    parser.add_argument('-lr', type=float, default=1e-4, help='learning rate')
    parser.add_argument('-batch_size', type=int, help='batch size', default=1024)
    parser.add_argument('-epochs', type=int, default=1000, help='training epochs')
    parser.add_argument('-layers', type=int, default=4, help='number of hidden layers')
    args = parser.parse_args()
    dataset = args.dataset
    preprocessor = Preprocessor(dataset)
    cols = preprocessor.df_cleaned.columns
    hierarchical_cols = ["year", "month", "day"]
    temp = [x + '_sine' for x in hierarchical_cols]
    temp2 = [x + '_cos' for x in hierarchical_cols]
    temp.extend(temp2)
    metadata = preprocessor.df_cleaned[temp]
    # metadata.to_csv('orig_meta_metrotraffic.csv')
    # exit()
    real_metadata = deepcopy(metadata).iloc[np.random.permutation(len(metadata))]
    # real_metadata = deepcopy(metadata)  # no shuffling of samples
    device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
    training_data = real_metadata.values
    in_features = training_data.shape[1]
    diffusion_config = fetchDiffusionConfig(args)
    model = fetchModel(in_features, args).to(device)
    optimizer = optim.Adam(model.parameters(), lr=args.lr)
    criterion = nn.MSELoss()
    # model = Model(diffusion_params, backbone, model_params)
    dataset = MyDataset(from_numpy(training_data).float())
    dataloader = DataLoader(dataset, batch_size=args.batch_size)
    for epoch in range(args.epochs):
        total_loss = 0.0
        for batch in dataloader:
            timesteps = torch.randint(diffusion_config['T'], size=(batch.shape[0],))
            sigmas = torch.normal(0, 1, size=batch.shape)
            """Forward noising"""
            alpha_bars = diffusion_config['alpha_bars']
            batch_noised = torch.sqrt(alpha_bars[timesteps]) * batch + torch.sqrt(1 - alpha_bars[timesteps]) * sigmas
            batch_noised = batch_noised.to(device)
            timesteps_normalized = (timesteps / diffusion_config['T']).reshape((-1, 1))
            timesteps_normalized = timesteps_normalized.to(device)
            sigmas_predicted = model(batch_noised, timesteps_normalized)
            optimizer.zero_grad()
            sigmas = sigmas.to(device)
            loss = criterion(sigmas_predicted, sigmas)
            loss.backward()
            total_loss += loss
            optimizer.step()

        print(f'epoch: {epoch}, loss: {total_loss}')

    """Synthesis"""
    data = torch.normal(0, 1, size=training_data.shape).to(device)
    with torch.no_grad():
        for step in range(diffusion_config['T'] - 1, -1, -1):
            print(f"backward step: {step}")
            times = torch.full(size=(training_data.shape[0], 1), fill_value=step)
            times_normalized = (times / (diffusion_config['T'])).to(device)
            epsilon_pred = model(data, times_normalized)
            difference_coeff = diffusion_config['betas'][step] / torch.sqrt(1 - diffusion_config['alpha_bars'][step])
            denom = diffusion_config['alphas'][step]
            sigma = diffusion_config['betas'][step] * (1 - diffusion_config['alpha_bars'][step - 1]) / (
                    1 - diffusion_config['alpha_bars'][step])
            sigma = torch.sqrt(sigma) * torch.normal(0, 1, training_data.shape)
            sigma = sigma.to(device)
            difference_coeff = difference_coeff.to(device)
            denom = denom.to(device)
            data = (data - difference_coeff * epsilon_pred) / denom + sigma
    synth = data.cpu().numpy()
    synth_meta_dataframe = pd.DataFrame(synth, columns=metadata.columns)
    synth_meta_decoded = preprocessor.cyclicDecode(synth_meta_dataframe)
    synth_meta_decoded = synth_meta_decoded.sort_values(by=hierarchical_cols).reset_index(drop=True)
    # synth_meta_decoded.drop(columns=['Unnamed:0'], inplace=True)
    synth_meta_decoded.to_csv('synths/synth_meta_metrotraffic.csv')
    real = preprocessor.df_orig
    real = real[hierarchical_cols]
    real = real.sort_values(by=hierarchical_cols).reset_index(drop=True)
    # real.drop(columns=['Unnamed:0'], inplace=True)
    real.to_csv('synths/real_meta_metrotraffic.csv')
    print('finished')