CERBERUS/trainer.py at master · cscribano/CERBERUS · 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
150
151
152
153
154
155
156
157
158
159
160
161
162
163
# -*- coding: utf-8 -*-
# ---------------------

from conf import Conf

import torch
import numpy as np
import cv2

import pytorch_lightning as pl
import torchvision as tv
from torch.utils.data import DataLoader
from pytorch_lightning import loggers as pl_loggers

from torch.optim import *
from utils.lr_scheduler import *

from models import CerberusModel
from models.losses import MultiTaskLoss
from dataset import MultitaskDataset, ignore_collate


class PL_trainable(pl.LightningModule):
    def __init__(self, cnf):
        super().__init__()

        self.cnf = cnf
        self.backbone = CerberusModel(cnf)
        self.criterion = MultiTaskLoss(cnf)

        self.plot_images = 10

    def forward(self, img):
        pred = self.backbone(img)
        return pred

    def training_step(self, batch, batch_idx):

        img, targets = batch
        preds = self.forward(img)

        # Loss
        loss, loss_detail = self.criterion(preds, targets)

        # single scheduler
        sch = self.lr_schedulers()
        sch.step()

        lr = sch.get_last_lr()[0]
        self.log('train_loss', loss, on_step=True, on_epoch=False)
        self.log('lr', lr, on_step=True, on_epoch=False)
        for k, v in loss_detail.items():
            self.log(f'train_{k}_loss', v, on_step=True, on_epoch=False)

        return loss

    def validation_step(self, batch, batch_idx):
        # Inference
        img, targets = batch
        preds = self.forward(img)

        # Loss
        loss, loss_detail = self.criterion(preds, targets)

        # plot
        if self.plot_images > 0:
            true, pred = [], []

            if self.cnf.base.get("object_det", False):
                true.append(targets["obj_det"]["heatmaps"])
                pred.append(preds["obj_det"]["heatmaps"])

            if self.cnf.base.get("lane_det", False):
                true.append(targets["lane_det"]["heatmaps"])
                pred.append(preds["lane_det"]["heatmaps"])

            true = torch.cat(true, dim=1)
            pred = torch.cat(pred, dim=1)
            img_resize = torch_input_img(img[0].cpu().detach())
            hm_true = torch_heatmap_img(true[0].cpu().detach())
            hm_pred = torch_heatmap_img(pred[0].cpu().detach())
            grid = torch.stack([img_resize, hm_true, hm_pred], dim=0)

            grid = tv.utils.make_grid(grid.float())
            self.logger.experiment.add_image(tag=f'results_{self.plot_images}',
                                             img_tensor=grid, global_step=self.global_step)
            self.plot_images -= 1

        # Log
        self.log('val_loss', loss, on_step=False, on_epoch=True)

        for k, v in loss_detail.items():
            self.log(f'val_{k}_loss', v, on_step=False, on_epoch=True)

        return loss


    def test_step(self, batch, batch_idx):
        pass

    def validation_epoch_end(self, outputs) -> None:
        self.plot_images = 10

    def configure_optimizers(self):
        optimizer = eval(self.cnf.optimizer.name)(self.parameters(), **self.cnf.optimizer.args)

        if self.cnf.lr_scheduler.get("name", None) is not None:
            scheduler = eval(self.cnf.lr_scheduler.name)(optimizer, **self.cnf.lr_scheduler.args)
            return [optimizer], [scheduler]
        return [optimizer]

    def on_validation_epoch_end(self):
        torch.save(self.backbone.state_dict(), f'{self.cnf.exp_log_path}/last.pth')


def torch_heatmap_img(heatmap):
    hm_show, _ = torch.max(heatmap, dim=0)
    hm_show = hm_show.numpy() * 255
    hm_show = hm_show.astype(np.uint8)
    hm_show = cv2.applyColorMap(hm_show, cv2.COLORMAP_JET)
    hm_show = cv2.cvtColor(hm_show, cv2.COLOR_BGR2RGB)
    hm_show = cv2.resize(hm_show, (640, 480)) / 255

    return torch.from_numpy(hm_show).permute(2, 0, 1)


def torch_input_img(img):
    invTrans = tv.transforms.Compose([tv.transforms.Normalize(mean=[0., 0., 0.],
                                                        std=[1 / 0.229, 1 / 0.224, 1 / 0.225]),
                                   tv.transforms.Normalize(mean=[-0.485, -0.456, -0.406],
                                                        std=[1., 1., 1.]),
                                   ])
    img = invTrans(img)
    img = tv.transforms.Resize((480, 640))(img)
    return img

def trainer_run(cnf):
    # type: (Conf) -> None

    # ------------
    # data
    # ------------
    trainset = MultitaskDataset(cnf, mode="train")
    valset = MultitaskDataset(cnf, mode="val")

    collate_fn = ignore_collate(["centers", "offsets", "keypoints", "occlusion", "quant_offsets"])
    train_loader = DataLoader(trainset, collate_fn=collate_fn, **cnf.dataset.train_dataset.loader_args)
    val_loader = DataLoader(valset, collate_fn=collate_fn, **cnf.dataset.val_dataset.loader_args)

    # ------------
    # model
    # ------------
    model = PL_trainable(cnf)

    # ------------
    # training
    # ------------
    gpus = [0]
    tb_logger = pl_loggers.TensorBoardLogger(save_dir=cnf.exp_log_path, name="", version="")

    trainer = pl.Trainer(default_root_dir=cnf.exp_log_path, logger=tb_logger,
                         max_epochs=cnf.epochs, gpus=gpus)
    trainer.fit(model, train_loader, val_loader)