File size: 7,807 Bytes
3a0062c
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
"""
Implementation of YOLOv3 architecture
"""

import pytorch_lightning as pl
import torch
import torch.nn as nn
import torch.optim as optim
from torch.optim.lr_scheduler import OneCycleLR


from . import config
from .loss import YoloLoss

model_config = [
    (32, 3, 1),
    (64, 3, 2),
    ["B", 1],
    (128, 3, 2),
    ["B", 2],
    (256, 3, 2),
    ["B", 8],
    (512, 3, 2),
    ["B", 8],
    (1024, 3, 2),
    ["B", 4],   # darknet 53 ends here

    (512, 1, 1),
    (1024, 3, 1),
    "S",
    
    (256, 1, 1),
    "U",
    (256, 1, 1),
    (512, 3, 1),
    "S",
    
    (128, 1, 1),
    "U",
    (128, 1, 1),
    (256, 3, 1),
    "S"
]

class CNNBlock(pl.LightningModule):
    def __init__(self, in_channels, out_channels, bn_act=True, **kwargs):
        super().__init__()
        self.conv = nn.Conv2d(in_channels, out_channels, bias=not bn_act, **kwargs)
        self.bn = nn.BatchNorm2d(out_channels)
        self.leaky = nn.LeakyReLU(0.1)
        self.use_bn_act = bn_act

    def forward(self, x):
        if self.use_bn_act:
            return self.leaky(self.bn((self.conv(x))))
        else:
            return self.conv(x)

class ResidualBlock(pl.LightningModule):
    def __init__(self, channels, use_residual=True, num_repeats=1):
        super().__init__()
        self.layers = nn.ModuleList()
        for repeat in range(num_repeats):
            self.layers += [
                nn.Sequential(
                    CNNBlock(channels, channels//2, kernel_size=1),
                    CNNBlock(channels//2, channels, kernel_size=3, padding=1)
                )
            ]
        self.use_residual = use_residual
        self.num_repeats = num_repeats

    def forward(self, x):
        for layer in self.layers:
            if self.use_residual:
                x = x + layer(x)
            else:
                x = layer(x)

        return x
    
class ScalePrediction(pl.LightningModule):
    def __init__(self, in_channels, num_classes):
        super().__init__()
        self.pred = nn.Sequential(
            CNNBlock(in_channels, 2 * in_channels, kernel_size=3, padding=1),
            CNNBlock(2 * in_channels, (num_classes + 5) * 3, kernel_size=1, bn_act=False)
        )
        self.num_classes = num_classes

    def forward(self, x):
        return (
            self.pred(x).
            reshape(x.shape[0], 3, self.num_classes + 5, x.shape[2], x.shape[3]).
            permute(0, 1, 3, 4, 2)
        )

class YOLOv3(pl.LightningModule):
    def __init__(self, in_channels=3, num_classes=20):
        super().__init__()
        self.num_classes = num_classes
        self.in_channels = in_channels
        self.layers = self._create_conv_layers()

        self.scaled_anchors = (
            torch.tensor(config.ANCHORS) * torch.tensor(config.S).unsqueeze(1).unsqueeze(1).repeat(1, 3, 2)         #  ?
        ).to(config.DEVICE)

        self.learning_rate = config.LEARNING_RATE
        self.weight_decay = config.WEIGHT_DECAY
        self.best_lr = 1e-3  ## ?

    def forward(self, x):  # ?
        outputs = []    # for each scale
        route_connections = []
        for layer in self.layers:
            if isinstance(layer, ScalePrediction):
                outputs.append(layer(x))
                continue

            x = layer(x)

            if isinstance(layer, ResidualBlock) and layer.num_repeats == 8:
                route_connections.append(x)
            elif isinstance(layer, nn.Upsample):
                x = torch.cat([x, route_connections[-1]], dim=1)
                route_connections.pop()

        return outputs
    
    def _create_conv_layers(self):
        layers = nn.ModuleList()
        in_channels = self.in_channels

        for module in model_config:
            if isinstance(module, tuple):
                out_channels, kernel_size, stride = module
                layers.append(
                    CNNBlock(in_channels, out_channels, kernel_size=kernel_size, stride=stride, padding=1 if kernel_size==3 else 0)
                )
                in_channels = out_channels

            elif isinstance(module, list):
                num_repeats = module[1]
                layers.append(
                    ResidualBlock(in_channels, num_repeats=num_repeats)
                )
            elif isinstance(module, str):
                if module == "S":
                    layers += [
                        ResidualBlock(in_channels, use_residual=False, num_repeats=1),
                        CNNBlock(in_channels, in_channels//2, kernel_size=1),
                        ScalePrediction(in_channels//2, num_classes=self.num_classes)
                    ]
                    in_channels = in_channels // 2

                elif module == "U":
                    layers.append(nn.Upsample(scale_factor=2))
                    in_channels = in_channels * 3

        return layers
    

    def yololoss(self):
        return YoloLoss()
    
    def training_step(self, batch, batch_idx):
        x, y = batch
        y0, y1, y2 = y[0], y[1], y[2]
        out = self.forward(x)
        # print(out[0].shape, y0.shape)

        loss = (                                                    # ?
            self.yololoss()(out[0], y0, self.scaled_anchors[0])
            + self.yololoss()(out[1], y1, self.scaled_anchors[1])
            + self.yololoss()(out[2], y2, self.scaled_anchors[2])
        )
        
        self.log(
            "train_loss", loss, prog_bar=True, logger=True, on_step=True, on_epoch=True 
        )
        return loss

    def test_step(self, batch, batch_idx):
        x, y = batch
        y0, y1, y2 = y[0], y[1], y[2]
        out = self.forward(x)

        loss = (
            self.yololoss()(out[0], y0, self.scaled_anchors[0])
            + self.yololoss()(out[1], y1, self.scaled_anchors[1])
            + self.yololoss()(out[2], y2, self.scaled_anchors[2])
        )

        self.log(
            "test_loss", loss, prog_bar=True, logger=True, on_step=True, on_epoch=True 
        )

        return loss
    
    def on_train_epoch_end(self) -> None:
        print(
            f"Epoch: {self.current_epoch}, Loss: {self.trainer.callback_metrics['train_loss_epoch']}"
        )

    def on_test_epoch_end(self) -> None:
        print(
            f"Epoch: {self.current_epoch}, Loss: {self.trainer.callback_metrics['test_loss_epoch']}"
        )
    
    def configure_optimizers(self):
        optimizer = optim.Adam(
            self.parameters(), lr=self.learning_rate, weight_decay=self.weight_decay
        )

        scheduler = OneCycleLR(
            optimizer,
            max_lr=self.best_lr,
            steps_per_epoch=len(self.trainer.datamodule.train_dataloader()),
            epochs=config.NUM_EPOCHS,
            pct_start=8 / config.NUM_EPOCHS,
            div_factor=100,
            three_phase=False,
            final_div_factor=100,
            anneal_strategy="linear"
        )

        return [optimizer], [{"scheduler": scheduler, "interval": "step", "frequency": 1}]

    def on_train_end(self) -> None:
        torch.save(self.state_dict(), config.MODEL_STATE_DICT_PATH)

if __name__ == "main":
    num_classes = 20
    IMAGE_SIZE = 416
    model = YOLOv3(num_classes=num_classes)
    x = torch.randn((2, 3, IMAGE_SIZE, IMAGE_SIZE))
    out = model(x)
    assert model(x)[0].shape == (
        2,
        3,
        IMAGE_SIZE // 32, 
        IMAGE_SIZE // 32,
        num_classes + 5
    )
    assert model(x)[1].shape == (
        2,
        3,
        IMAGE_SIZE // 16, 
        IMAGE_SIZE // 16,
        num_classes + 5
    )
    assert model(x)[2].shape == (
        2,
        3,
        IMAGE_SIZE // 8, 
        IMAGE_SIZE // 8,
        num_classes + 5
    )
    print("Image size compatibility check passed!")