質問編集履歴
3
書式の修正
test
CHANGED
|
File without changes
|
test
CHANGED
|
@@ -2,8 +2,6 @@
|
|
|
2
2
|
|
|
3
3
|
PyTorchによるSSDを用いた物体検出の訓練について
|
|
4
4
|
|
|
5
|
-
|
|
6
|
-
|
|
7
5
|
「作りながら学ぶ PyTorchによる発展ディープラーニング」
|
|
8
6
|
|
|
9
7
|
の第2章 「2-7_SSD_training.ipynb」
|
|
@@ -402,7 +400,9 @@
|
|
|
402
400
|
|
|
403
401
|
### 試したこと
|
|
404
402
|
|
|
403
|
+
シングルGPUでは問題なく動かすことができますが、net = nn.DataParallel(net)を実行すると上記のようなエラー文が出てしまいます。
|
|
404
|
+
|
|
405
|
-
|
|
405
|
+
エラー文をインターネットで検索して出てきた解決方法は一通り行ってみました。
|
|
406
406
|
|
|
407
407
|
|
|
408
408
|
|
2
書式の改善
test
CHANGED
|
@@ -1 +1 @@
|
|
|
1
|
-
|
|
1
|
+
SSDでマルチGPUで学習しようとするとき、ギャザー関数がCPUテンソルに含まれていないとエラーが表示される
|
test
CHANGED
|
File without changes
|
1
エラーメッセージの不足を追加
test
CHANGED
|
File without changes
|
test
CHANGED
|
@@ -18,6 +18,130 @@
|
|
|
18
18
|
|
|
19
19
|
```
|
|
20
20
|
|
|
21
|
+
AssertionError Traceback (most recent call last)
|
|
22
|
+
|
|
23
|
+
<ipython-input-34-56fa4f8d86af> in <module>
|
|
24
|
+
|
|
25
|
+
1 # 学習・検証を実行する
|
|
26
|
+
|
|
27
|
+
2 num_epochs= 10
|
|
28
|
+
|
|
29
|
+
----> 3 train_model(net, dataloaders_dict, criterion, optimizer, num_epochs=num_epochs)
|
|
30
|
+
|
|
31
|
+
|
|
32
|
+
|
|
33
|
+
<ipython-input-33-645d91cb3a1e> in train_model(net, dataloaders_dict, criterion, optimizer, num_epochs)
|
|
34
|
+
|
|
35
|
+
60 with torch.set_grad_enabled(phase == 'train'):
|
|
36
|
+
|
|
37
|
+
61 # 順伝搬（forward）計算
|
|
38
|
+
|
|
39
|
+
---> 62 outputs = net(images)
|
|
40
|
+
|
|
41
|
+
63
|
|
42
|
+
|
|
43
|
+
64 # 損失の計算
|
|
44
|
+
|
|
45
|
+
|
|
46
|
+
|
|
47
|
+
~/anaconda3/envs/vgg/lib/python3.6/site-packages/torch/nn/modules/module.py in _call_impl(self, *input, **kwargs)
|
|
48
|
+
|
|
49
|
+
725 result = self._slow_forward(*input, **kwargs)
|
|
50
|
+
|
|
51
|
+
726 else:
|
|
52
|
+
|
|
53
|
+
--> 727 result = self.forward(*input, **kwargs)
|
|
54
|
+
|
|
55
|
+
728 for hook in itertools.chain(
|
|
56
|
+
|
|
57
|
+
729 _global_forward_hooks.values(),
|
|
58
|
+
|
|
59
|
+
|
|
60
|
+
|
|
61
|
+
~/anaconda3/envs/vgg/lib/python3.6/site-packages/torch/nn/parallel/data_parallel.py in forward(self, *inputs, **kwargs)
|
|
62
|
+
|
|
63
|
+
160 replicas = self.replicate(self.module, self.device_ids[:len(inputs)])
|
|
64
|
+
|
|
65
|
+
161 outputs = self.parallel_apply(replicas, inputs, kwargs)
|
|
66
|
+
|
|
67
|
+
--> 162 return self.gather(outputs, self.output_device)
|
|
68
|
+
|
|
69
|
+
163
|
|
70
|
+
|
|
71
|
+
164 def replicate(self, module, device_ids):
|
|
72
|
+
|
|
73
|
+
|
|
74
|
+
|
|
75
|
+
~/anaconda3/envs/vgg/lib/python3.6/site-packages/torch/nn/parallel/data_parallel.py in gather(self, outputs, output_device)
|
|
76
|
+
|
|
77
|
+
172
|
|
78
|
+
|
|
79
|
+
173 def gather(self, outputs, output_device):
|
|
80
|
+
|
|
81
|
+
--> 174 return gather(outputs, output_device, dim=self.dim)
|
|
82
|
+
|
|
83
|
+
175
|
|
84
|
+
|
|
85
|
+
176
|
|
86
|
+
|
|
87
|
+
|
|
88
|
+
|
|
89
|
+
~/anaconda3/envs/vgg/lib/python3.6/site-packages/torch/nn/parallel/scatter_gather.py in gather(outputs, target_device, dim)
|
|
90
|
+
|
|
91
|
+
66 # Setting the function to None clears the refcycle.
|
|
92
|
+
|
|
93
|
+
67 try:
|
|
94
|
+
|
|
95
|
+
---> 68 res = gather_map(outputs)
|
|
96
|
+
|
|
97
|
+
69 finally:
|
|
98
|
+
|
|
99
|
+
70 gather_map = None
|
|
100
|
+
|
|
101
|
+
|
|
102
|
+
|
|
103
|
+
~/anaconda3/envs/vgg/lib/python3.6/site-packages/torch/nn/parallel/scatter_gather.py in gather_map(outputs)
|
|
104
|
+
|
|
105
|
+
61 return type(out)(((k, gather_map([d[k] for d in outputs]))
|
|
106
|
+
|
|
107
|
+
62 for k in out))
|
|
108
|
+
|
|
109
|
+
---> 63 return type(out)(map(gather_map, zip(*outputs)))
|
|
110
|
+
|
|
111
|
+
64
|
|
112
|
+
|
|
113
|
+
65 # Recursive function calls like this create reference cycles.
|
|
114
|
+
|
|
115
|
+
|
|
116
|
+
|
|
117
|
+
~/anaconda3/envs/vgg/lib/python3.6/site-packages/torch/nn/parallel/scatter_gather.py in gather_map(outputs)
|
|
118
|
+
|
|
119
|
+
53 out = outputs[0]
|
|
120
|
+
|
|
121
|
+
54 if isinstance(out, torch.Tensor):
|
|
122
|
+
|
|
123
|
+
---> 55 return Gather.apply(target_device, dim, *outputs)
|
|
124
|
+
|
|
125
|
+
56 if out is None:
|
|
126
|
+
|
|
127
|
+
57 return None
|
|
128
|
+
|
|
129
|
+
|
|
130
|
+
|
|
131
|
+
~/anaconda3/envs/vgg/lib/python3.6/site-packages/torch/nn/parallel/_functions.py in forward(ctx, target_device, dim, *inputs)
|
|
132
|
+
|
|
133
|
+
54 def forward(ctx, target_device, dim, *inputs):
|
|
134
|
+
|
|
135
|
+
55 assert all(map(lambda i: i.device.type != 'cpu', inputs)), (
|
|
136
|
+
|
|
137
|
+
---> 56 'Gather function not implemented for CPU tensors'
|
|
138
|
+
|
|
139
|
+
57 )
|
|
140
|
+
|
|
141
|
+
58 target_device = _get_device_index(target_device, True)
|
|
142
|
+
|
|
143
|
+
|
|
144
|
+
|
|
21
145
|
AssertionError: Gather function not implemented for CPU tensors
|
|
22
146
|
|
|
23
147
|
|
|
@@ -32,220 +156,6 @@
|
|
|
32
156
|
|
|
33
157
|
```python
|
|
34
158
|
|
|
35
|
-
# パッケージのimport
|
|
36
|
-
|
|
37
|
-
import os.path as osp
|
|
38
|
-
|
|
39
|
-
import random
|
|
40
|
-
|
|
41
|
-
import time
|
|
42
|
-
|
|
43
|
-
|
|
44
|
-
|
|
45
|
-
import cv2
|
|
46
|
-
|
|
47
|
-
import numpy as np
|
|
48
|
-
|
|
49
|
-
import pandas as pd
|
|
50
|
-
|
|
51
|
-
import torch
|
|
52
|
-
|
|
53
|
-
import torch.nn as nn
|
|
54
|
-
|
|
55
|
-
import torch.nn.init as init
|
|
56
|
-
|
|
57
|
-
import torch.optim as optim
|
|
58
|
-
|
|
59
|
-
import torch.utils.data as data
|
|
60
|
-
|
|
61
|
-
|
|
62
|
-
|
|
63
|
-
# 乱数のシードを設定
|
|
64
|
-
|
|
65
|
-
torch.manual_seed(1234)
|
|
66
|
-
|
|
67
|
-
np.random.seed(1234)
|
|
68
|
-
|
|
69
|
-
random.seed(1234)
|
|
70
|
-
|
|
71
|
-
|
|
72
|
-
|
|
73
|
-
from utils.ssd_model import make_datapath_list, VOCDataset, DataTransform, Anno_xml2list, od_collate_fn
|
|
74
|
-
|
|
75
|
-
|
|
76
|
-
|
|
77
|
-
|
|
78
|
-
|
|
79
|
-
# ファイルパスのリストを取得
|
|
80
|
-
|
|
81
|
-
rootpath = "./data/VOCdevkit/VOC2012/"
|
|
82
|
-
|
|
83
|
-
train_img_list, train_anno_list, val_img_list, val_anno_list = make_datapath_list(
|
|
84
|
-
|
|
85
|
-
rootpath)
|
|
86
|
-
|
|
87
|
-
|
|
88
|
-
|
|
89
|
-
# Datasetを作成
|
|
90
|
-
|
|
91
|
-
voc_classes = ['aeroplane', 'bicycle', 'bird', 'boat',
|
|
92
|
-
|
|
93
|
-
'bottle', 'bus', 'car', 'cat', 'chair',
|
|
94
|
-
|
|
95
|
-
'cow', 'diningtable', 'dog', 'horse',
|
|
96
|
-
|
|
97
|
-
'motorbike', 'person', 'pottedplant',
|
|
98
|
-
|
|
99
|
-
'sheep', 'sofa', 'train', 'tvmonitor']
|
|
100
|
-
|
|
101
|
-
color_mean = (104, 117, 123) # (BGR)の色の平均値
|
|
102
|
-
|
|
103
|
-
input_size = 300 # 画像のinputサイズを300×300にする
|
|
104
|
-
|
|
105
|
-
|
|
106
|
-
|
|
107
|
-
train_dataset = VOCDataset(train_img_list, train_anno_list, phase="train", transform=DataTransform(
|
|
108
|
-
|
|
109
|
-
input_size, color_mean), transform_anno=Anno_xml2list(voc_classes))
|
|
110
|
-
|
|
111
|
-
|
|
112
|
-
|
|
113
|
-
val_dataset = VOCDataset(val_img_list, val_anno_list, phase="val", transform=DataTransform(
|
|
114
|
-
|
|
115
|
-
input_size, color_mean), transform_anno=Anno_xml2list(voc_classes))
|
|
116
|
-
|
|
117
|
-
|
|
118
|
-
|
|
119
|
-
|
|
120
|
-
|
|
121
|
-
# DataLoaderを作成する
|
|
122
|
-
|
|
123
|
-
batch_size = 64
|
|
124
|
-
|
|
125
|
-
|
|
126
|
-
|
|
127
|
-
train_dataloader = data.DataLoader(
|
|
128
|
-
|
|
129
|
-
train_dataset, batch_size=batch_size, shuffle=True, collate_fn=od_collate_fn)
|
|
130
|
-
|
|
131
|
-
|
|
132
|
-
|
|
133
|
-
val_dataloader = data.DataLoader(
|
|
134
|
-
|
|
135
|
-
val_dataset, batch_size=batch_size, shuffle=False, collate_fn=od_collate_fn)
|
|
136
|
-
|
|
137
|
-
|
|
138
|
-
|
|
139
|
-
# 辞書オブジェクトにまとめる
|
|
140
|
-
|
|
141
|
-
dataloaders_dict = {"train": train_dataloader, "val": val_dataloader}
|
|
142
|
-
|
|
143
|
-
|
|
144
|
-
|
|
145
|
-
from utils.ssd_model import SSD
|
|
146
|
-
|
|
147
|
-
|
|
148
|
-
|
|
149
|
-
# SSD300の設定
|
|
150
|
-
|
|
151
|
-
ssd_cfg = {
|
|
152
|
-
|
|
153
|
-
'num_classes': 21, # 背景クラスを含めた合計クラス数
|
|
154
|
-
|
|
155
|
-
'input_size': 300, # 画像の入力サイズ
|
|
156
|
-
|
|
157
|
-
'bbox_aspect_num': [4, 6, 6, 6, 4, 4], # 出力するDBoxのアスペクト比の種類
|
|
158
|
-
|
|
159
|
-
'feature_maps': [38, 19, 10, 5, 3, 1], # 各sourceの画像サイズ
|
|
160
|
-
|
|
161
|
-
'steps': [8, 16, 32, 64, 100, 300], # DBOXの大きさを決める
|
|
162
|
-
|
|
163
|
-
'min_sizes': [30, 60, 111, 162, 213, 264], # DBOXの大きさを決める
|
|
164
|
-
|
|
165
|
-
'max_sizes': [60, 111, 162, 213, 264, 315], # DBOXの大きさを決める
|
|
166
|
-
|
|
167
|
-
'aspect_ratios': [[2], [2, 3], [2, 3], [2, 3], [2], [2]],
|
|
168
|
-
|
|
169
|
-
}
|
|
170
|
-
|
|
171
|
-
|
|
172
|
-
|
|
173
|
-
# SSDネットワークモデル
|
|
174
|
-
|
|
175
|
-
net = SSD(phase="train", cfg=ssd_cfg)
|
|
176
|
-
|
|
177
|
-
|
|
178
|
-
|
|
179
|
-
# SSDの初期の重みを設定
|
|
180
|
-
|
|
181
|
-
# ssdのvgg部分に重みをロードする
|
|
182
|
-
|
|
183
|
-
vgg_weights = torch.load('./weights/vgg16_reducedfc.pth')
|
|
184
|
-
|
|
185
|
-
net.vgg.load_state_dict(vgg_weights)
|
|
186
|
-
|
|
187
|
-
|
|
188
|
-
|
|
189
|
-
# ssdのその他のネットワークの重みはHeの初期値で初期化
|
|
190
|
-
|
|
191
|
-
|
|
192
|
-
|
|
193
|
-
|
|
194
|
-
|
|
195
|
-
def weights_init(m):
|
|
196
|
-
|
|
197
|
-
if isinstance(m, nn.Conv2d):
|
|
198
|
-
|
|
199
|
-
init.kaiming_normal_(m.weight.data)
|
|
200
|
-
|
|
201
|
-
if m.bias is not None: # バイアス項がある場合
|
|
202
|
-
|
|
203
|
-
nn.init.constant_(m.bias, 0.0)
|
|
204
|
-
|
|
205
|
-
|
|
206
|
-
|
|
207
|
-
|
|
208
|
-
|
|
209
|
-
# Heの初期値を適用
|
|
210
|
-
|
|
211
|
-
net.extras.apply(weights_init)
|
|
212
|
-
|
|
213
|
-
net.loc.apply(weights_init)
|
|
214
|
-
|
|
215
|
-
net.conf.apply(weights_init)
|
|
216
|
-
|
|
217
|
-
|
|
218
|
-
|
|
219
|
-
# GPUが使えるかを確認
|
|
220
|
-
|
|
221
|
-
device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
|
|
222
|
-
|
|
223
|
-
print("使用デバイス：", device)
|
|
224
|
-
|
|
225
|
-
|
|
226
|
-
|
|
227
|
-
print('ネットワーク設定完了：学習済みの重みをロードしました')
|
|
228
|
-
|
|
229
|
-
|
|
230
|
-
|
|
231
|
-
from utils.ssd_model import MultiBoxLoss
|
|
232
|
-
|
|
233
|
-
|
|
234
|
-
|
|
235
|
-
# 損失関数の設定
|
|
236
|
-
|
|
237
|
-
criterion = MultiBoxLoss(jaccard_thresh=0.5, neg_pos=3, device=device)
|
|
238
|
-
|
|
239
|
-
|
|
240
|
-
|
|
241
|
-
# 最適化手法の設定
|
|
242
|
-
|
|
243
|
-
optimizer = optim.SGD(net.parameters(), lr=1e-3,
|
|
244
|
-
|
|
245
|
-
momentum=0.9, weight_decay=5e-4)
|
|
246
|
-
|
|
247
|
-
|
|
248
|
-
|
|
249
159
|
# モデルを学習させる関数を作成
|
|
250
160
|
|
|
251
161
|
|