gotp/models.py at main · minsikseo-cdl/gotp · 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
from torch import nn
from torch.optim.lr_scheduler import ReduceLROnPlateau
from torch_geometric.loader import DataLoader
from pytorch_lightning import LightningModule
import matplotlib.pyplot as plt
from matplotlib.tri import Triangulation
from utils import load_sequenced_dataset, CosineAnnealingWarmUpRestarts, load_mesh
from networks import GOTPNet


class GOTPModel(LightningModule):
    def __init__(self,
                 hidden_size, num_layers, depth, ratio=0.5, heads=1, num_seq=1,
                 lr=1e-3, T_0=100, T_mult=1.2, T_up=1, gamma=0.8,
                 batch_size=1, schedule='cosine'):
        super().__init__()
        self.save_hyperparameters()

        self.hidden_size = hidden_size
        self.depth = depth
        self.num_seq = num_seq
        self.lr = lr
        self.T_0 = T_0
        self.T_mult = T_mult
        self.T_up = T_up
        self.gamma = gamma
        self.schedule = schedule
        self.batch_size = batch_size

        self.net = GOTPNet(
            input_size=6, hidden_size=hidden_size, output_size=1,
            num_layers=num_layers, depth=depth, ratio=ratio, heads=heads)

        self.L = nn.SmoothL1Loss(beta=0.2)

        self.fig, self.axs = plt.subplots(2, num_seq, dpi=200)

    def forward(self, graphs):
        zk = [graphs.x[:, :1]]
        for i in range(self.num_seq):
            zk.append(self.net(
                torch.cat([zk[i].detach(), graphs.u, graphs.s], dim=1),
                graphs.edge_index, graphs.batch))
        return torch.cat(zk, dim=1)

    def configure_optimizers(self):
        if self.schedule == 'cosine':
            optimizer = torch.optim.Adam(self.parameters(), lr=1e-9)
            scheduler = CosineAnnealingWarmUpRestarts(
                optimizer, T_0=self.T_0, T_mult=self.T_mult, T_up=self.T_up,
                eta_max=self.lr, gamma=self.gamma)
            return [optimizer], [scheduler]
        elif self.schedule == 'reduce':
            optimizer = torch.optim.Adam(self.parameters(), lr=self.lr)
            scheduler = ReduceLROnPlateau(optimizer, factor=self.gamma)
            return {
                'optimizer': optimizer,
                'lr_scheduler': {
                    'scheduler': scheduler,
                    'monitor': 'loss/val'
                }
            }

    def training_step(self, graphs, *args):
        Z = self(graphs)[:, 1:]
        loss = self.L(Z, graphs.x[:, 1:])
        self.log('loss/train', loss.item(), on_step=False, on_epoch=True, batch_size=self.batch_size)
        metric = torch.mean(torch.abs(Z[:, -1].detach().cpu() - graphs.x[:, -1].cpu()))
        self.log('metric/train', metric.item(), on_step=False, on_epoch=True, batch_size=self.batch_size)
        torch.cuda.empty_cache()
        return {
            'loss': loss,
            'Y': graphs.x[:, 1:].cpu()[graphs.batch == 0],
            'Z': Z.detach().cpu()[graphs.batch == 0],
            'prob': graphs[0].prob,
            'idx': graphs[0].idx
        }

    def training_epoch_end(self, outputs):
        Y = outputs[0]['Y']
        Z = outputs[0]['Z']
        prob = outputs[0]['prob']
        idx = outputs[0]['idx']
        self.visualization(Y, Z, prob, idx)
        self.logger.experiment.add_figure(
            'Training monitor', self.fig, self.current_epoch
        )

    @torch.no_grad()
    def validation_step(self, graphs, *args):
        Z = self(graphs)[:, 1:]
        loss = self.L(Z, graphs.x[:, 1:])
        self.log('loss/val', loss.item(), on_step=False, on_epoch=True, batch_size=self.batch_size)
        metric = torch.mean(torch.abs(Z[:, -1].detach().cpu() - graphs.x[:, -1].cpu()))
        self.log('metric/val', metric.item(), on_step=False, on_epoch=True, batch_size=self.batch_size)
        torch.cuda.empty_cache()
        return {
            'loss': loss,
            'Y': graphs.x[:, 1:].cpu()[graphs.batch == 0],
            'Z': Z.detach().cpu()[graphs.batch == 0],
            'prob': graphs[0].prob,
            'idx': graphs[0].idx
        }

    def validation_epoch_end(self, outputs):
        Y = outputs[0]['Y']
        Z = outputs[0]['Z']
        prob = outputs[0]['prob']
        idx = outputs[0]['idx']
        self.visualization(Y, Z, prob, idx)
        self.logger.experiment.add_figure(
            'Validation monitor', self.fig, self.current_epoch
        )

    def visualization(self, Y, Z, prob, idx):
        pos, face = load_mesh(prob, idx)
        T = Triangulation(x=pos[0], y=pos[1], triangles=face)
        for ax in self.axs.ravel():
            ax.cla()
        if self.num_seq > 1:
            for i in range(self.num_seq):
                self.axs[0, i].tricontourf(T, Y[:, i], vmin=0, vmax=1)
                self.axs[1, i].tricontourf(T, Z[:, i], vmin=0, vmax=1)
        else:
            self.axs[0].tricontourf(T, Y[:, 0], vmin=0, vmax=1)
            self.axs[1].tricontourf(T, Z[:, 0], vmin=0, vmax=1)
        for ax in self.axs.ravel():
            ax.set_aspect('equal')
            ax.axes.get_xaxis().set_visible(False)
            ax.axes.get_yaxis().set_visible(False)
        self.fig.tight_layout()


if __name__ == '__main__':
    num_seq_ = 5
    data_list = load_sequenced_dataset(['clever'], num_seq=num_seq_)
    loader = DataLoader(data_list, batch_size=8)
    data = next(iter(loader))

    model = GOTPModel(hidden_size=64, num_layers=2, depth=2, heads=2, num_seq=num_seq_)
    model.cuda(0)
    z = model(data.cuda(0))
    print(z)