pointraft/train.py at main · pieterblok/pointraft · 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
149
## thanks to: https://colab.research.google.com/drive/1D45E5bUK3gQ40YpZo65ozs7hg5l-eo_U?usp=sharing#scrollTo=hUbRw_BhLuXr

import warnings
warnings.filterwarnings("ignore")

import os
from pathlib import Path
import random
import numpy as np

import torch
import torch.nn as nn
import torch.optim as optim
import torch_geometric.transforms as T
from torch_geometric.loader import DataLoader, ImbalancedSampler
from torch_geometric.typing import WITH_TORCH_CLUSTER

from data import *
from utils import *
from models import *
model_dict = {'dgcnn': DGCNN,
              'pointraft': PointRAFT,
              'pointnet': PointNet,
              'pointnet2': PointNet2,
              'pointtransformer': PointTransformer,
              'randlanet': RandLANet}


if not WITH_TORCH_CLUSTER:
    quit("This code requires 'torch-cluster'")


if __name__ == '__main__':
    torch.manual_seed(133)
    random.seed(133)
    np.random.seed(133)


    ## File paths
    current_file = Path(__file__).resolve()
    project_root = current_file.parent
    datafolder = os.path.join(project_root, 'data', '3DPotatoTwin')

    data_root = os.path.join(datafolder, '1_rgbd', '2_pcd')
    splits_csv = os.path.join(datafolder, 'splits.csv')
    target_csv = os.path.join(datafolder, 'ground_truth.csv')

    weightsfolder = os.path.join(project_root, 'weights')
    os.makedirs(weightsfolder, exist_ok=True)


    ## Data preprocessing and augmentation
    pre_transform = T.Center()
    transform = T.Compose([T.RandomJitter(0.0005),
                           T.RandomRotate(2, axis=0),
                           T.RandomRotate(2, axis=1),
                           T.RandomRotate(2, axis=2),
                           T.RandomFlip(axis=0, p=0.5),
                           T.RandomFlip(axis=1, p=0.5),
                           T.RandomShear(0.2)])


    ## Create/load InMemoryDatasets
    ## Please remove the "processed" folder in your data_root if you want to redo the data augmentation!
    train_dataset = PointCloudDataset(data_root,
                                      splits_csv,
                                      target_csv,
                                      target_col="weight_g_inctack",
                                      class_col="weight_class",
                                      split="train",
                                      conveyor_depth={'2023': 0.345, '2024': 0.460, '2025': 0.465},
                                      search_col="growing_season",
                                      max_height=0.08,
                                      pre_transform=pre_transform,
                                      transform=transform,
                                      num_points=1024,
                                      apply_augmentation=True)

    val_dataset = PointCloudDataset(data_root,
                                    splits_csv,
                                    target_csv,
                                    target_col="weight_g_inctack",
                                    split="val",
                                    conveyor_depth={'2023': 0.345, '2024': 0.460, '2025': 0.465},
                                    search_col="growing_season",
                                    max_height=0.08,
                                    pre_transform=pre_transform,
                                    transform=transform,
                                    num_points=1024,
                                    apply_augmentation=False)


    ## Dataloaders
    train_sampler = ImbalancedSampler(train_dataset.cls)
    train_loader = DataLoader(train_dataset, batch_size=32, sampler=train_sampler, num_workers=6)
    val_loader = DataLoader(val_dataset, batch_size=32, shuffle=False, num_workers=6)


    ## Initialize the training parameters
    device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
    model_name = 'pointraft'
    visualize = False
    model = model_dict[model_name]().to(device)

    criterion = nn.SmoothL1Loss(beta=20)
    optimizer = optim.Adam(model.parameters(), lr=0.001, weight_decay=0.0001)
    scheduler = optim.lr_scheduler.ExponentialLR(optimizer, gamma=0.97)

    if visualize:
        visualize_batch(train_loader)
        visualize_augmentation(train_loader, pre_transform)


    ## Training
    best_loss = float('inf')
    for epoch in range(1, 51):
        model.train()
        train_loss = 0
        for train_data in train_loader:
            train_data = train_data.to(device)
            optimizer.zero_grad()
            pred = model(train_data)
            target = train_data.y
            loss = criterion(pred, target)
            loss.backward()
            optimizer.step()
            train_loss += loss.item()

        train_loss /= len(train_loader)

        model.eval()
        val_loss = 0
        for val_data in val_loader:
            val_data = val_data.to(device)
            with torch.no_grad():
                pred = model(val_data)
                target = val_data.y
                loss = criterion(pred, target)
                val_loss  += loss.item()

        val_loss /= len(val_loader)

        print(f'Epoch: {epoch:03d}, Train Loss: {train_loss:.5f}, Val Loss: {val_loss:.5f}')
        if val_loss < best_loss:
            best_loss = val_loss
            torch.save(model.state_dict(), os.path.join(weightsfolder, model_name + '.pth'))
            print('Saved best model!')
        scheduler.step()