reticle_placement/train_egnn.py at main · cylinbao/reticle_placement · 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
import torch
import torch.nn as nn
import torch.optim as optim
import torch.nn.functional as F
import dgl
import json
from utils import *
from model import EdgeGNN

def gen_feat(nnode):
    types = np.ones(n_tot, dtype=int)
    slices = np.ones(n_tot, dtype=int)
    x = np.random.randint(32, size=n_tot)
    y = np.random.randint(32, size=n_tot)

    feat = np.array([types, slices, x, y]).T
    return feat

def loc_to_x_y(grid, loc):
    width, length = grid
    x = (loc % width)
    y = int(loc/width)
    return x, y

def locs_to_coords(grid, locs_data):
    x_data, y_data = [], []

    for locs in locs_data:
        x_arr, y_arr = [], []
        for loc in locs:
            x, y = loc_to_x_y(grid, loc)
            x_arr.append(x)
            y_arr.append(y)
        x_data.append(x_arr)
        y_data.append(y_arr)

    return (x_data, y_data)

def gen_feat(num_data, num_node, coords):
    types = np.ones(num_node, dtype=int)
    slices = np.ones(num_node, dtype=int)

    feat = []
    for i in range(num_data):
        x = coords[0][i]
        y = coords[1][i]
        h = np.array([types, slices, x, y]).T
        feat.append(h)
    feat = np.array(feat)

    return feat

def prepare_dataset(data):
    num_node = data['nnode']
    grid = data['grid']
    edges = data['edges']
    plcs = data['placements']

    num_edge = len(edges[0])
    num_data = len(plcs)

    graph = dgl.graph((torch.tensor(edges[0]), torch.tensor(edges[1])))

    cost_labels = []
    locs = []
    for d in plcs.items():
        cost_labels.append(d[1]['cost']/num_edge)
        # cost_labels.append(d[1]['cost'])
        locs.append(d[1]['locs'])

    coords = locs_to_coords(grid, locs)
    feat = gen_feat(num_data, num_node, coords)

    return (num_data, grid, graph, cost_labels, locs, feat)

def train(num_epoch, lr, num_layer, num_feat, num_hidden, data):
    model = EdgeGNN(num_layer, num_feat, num_hidden, F.relu)

    num_data, grid, g, cost_labels, locs, feat = prepare_dataset(data)

    g = dgl.add_reverse_edges(g)
    g = dgl.add_self_loop(g)

    labels = torch.tensor(cost_labels, dtype=torch.float)
    feat = torch.tensor(feat, dtype=torch.float)

    loss_fn = nn.MSELoss()
    optimizer = optim.SGD(model.parameters(), lr=lr)

    min_mean_loss = 99999999
    for e in range(num_epoch):
        with torch.autograd.set_detect_anomaly(True):
            loss_arr = []
            for i in range(num_data):
                model.train()
                p = model(g, feat[i])
                loss = loss_fn(p[0], labels[i])

                optimizer.zero_grad()
                loss.backward()
                optimizer.step()

                loss_arr.append(loss.item())
                # print("Epoch: {:d}, Iteration: {:d}, Predict: {:.3f}, Label: {:.3f}, MSE Loss: {:.3f}".format(e, i, p.item(), labels[i].item(), loss))

            mean_loss = np.mean(np.array(loss_arr))
            print("Epoch: {:d}, Mean MSE Loss: {:.3f}".format(e, mean_loss))
            if mean_loss < min_mean_loss:
                min_mean_loss = mean_loss

    print("Min Mean MSE Loss: {:.3f}".format(min_mean_loss))

    return

if __name__ == '__main__':
    num_epoch = 100
    num_layer = 7
    num_feat = 4
    num_hidden = 32
    lr = 0.00001

    path = "./programs/mac"
    prog = "mac4"

    f = open('./dataset/' + prog + '_dataset.json',)
    data = json.load(f)

    train(num_epoch, lr, num_layer, num_feat, num_hidden, data)