An intriguing failing of convolutional neural networks and the CoordConv solution
An intriguing failing of convolutional neural networks and the CoordConv solution
NeurIPS 2018
2019-10-10 15:01:48
Paper: https://arxiv.org/pdf/1807.03247.pdf
Official TensorFlow Code: https://github.com/uber-research/coordconv
Unofficial PyTorch Code: https://github.com/walsvid/CoordConv
机器之心：卷积神经网络「失陷」，CoordConv 来填坑: https://zhuanlan.zhihu.com/p/39665894
Uber提出CoordConv：解决普通CNN坐标变换问题: https://zhuanlan.zhihu.com/p/39919038
要拯救CNN的CoordConv受嘲讽，翻译个坐标还用训练? https://zhuanlan.zhihu.com/p/39841356
1. 给定 feature map and 坐标（x, y）如何生成对应的 relative CoordinateMap？
The following code is from: [ICCV19] AdaptIS: Adaptive Instance Selection Network [Github]
def get_instances_maps(self, F, points, adaptive_input, controller_input):
if isinstance(points, mx.nd.NDArray):
self.num_points = points.shape[1] if getattr(self.controller_net, 'return_map', False):
w = self.eqf(controller_input, points)
else:
w = self.eqf(controller_input, points)
w = self.controller_net(w) points = F.reshape(points, shape=(-1, 2))
x = F.repeat(adaptive_input, self.num_points, axis=0)
x = self.add_coord_features(x, points) x = self.block0(x)
x = self.adain(x, w)
x = self.block1(x) return x
class AppendCoordFeatures(gluon.HybridBlock):
def __init__(self, norm_radius, append_dist=True, spatial_scale=1.0):
super(AppendCoordFeatures, self).__init__()
self.xs = None
self.spatial_scale = spatial_scale
self.norm_radius = norm_radius
self.append_dist = append_dist def _ctx_kwarg(self, x):
if isinstance(x, mx.nd.NDArray):
return {"ctx": x.context}
return {} def get_coord_features(self, F, points, rows, cols, batch_size, **ctx_kwarg):
row_array = F.arange(start=0, stop=rows, step=1, **ctx_kwarg)
col_array = F.arange(start=0, stop=cols, step=1, **ctx_kwarg)
coord_rows = F.repeat(F.reshape(row_array, (1, 1, rows, 1)), repeats=cols, axis=3)
coord_cols = F.repeat(F.reshape(col_array, (1, 1, 1, cols)), repeats=rows, axis=2) coord_rows = F.repeat(coord_rows, repeats=batch_size, axis=0)
coord_cols = F.repeat(coord_cols, repeats=batch_size, axis=0) coords = F.concat(coord_rows, coord_cols, dim=1) add_xy = F.reshape(points * self.spatial_scale, shape=(0, 0, 1))
add_xy = F.reshape(F.repeat(add_xy, rows * cols, axis=2),
shape=(0, 0, rows, cols)) coords = (coords - add_xy) / (self.norm_radius * self.spatial_scale)
if self.append_dist:
dist = F.sqrt(F.sum(F.square(coords), axis=1, keepdims=1))
coord_features = F.concat(coords, dist, dim=1)
else:
coord_features = coords coord_features = F.clip(coord_features, a_min=-1, a_max=1)
return coord_features def hybrid_forward(self, F, x, coords):
if isinstance(x, mx.nd.NDArray):
self.xs = x.shape batch_size, rows, cols = self.xs[0], self.xs[2], self.xs[3]
coord_features = self.get_coord_features(F, coords, rows, cols, batch_size, **self._ctx_kwarg(x)) return F.concat(coord_features, x, dim=1)
def get_coord_features(self, F, points, rows, cols, batch_size, **ctx_kwarg):
# (Pdb) points, rows, cols, batch_size
# ([[61. 71.]] <NDArray 1x2 @gpu(0)>, 96, 96, 1)
# row_array and col_array:
# [ 0. 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.]
# <NDArray 96 @gpu(0)>
# (Pdb) coord_rows
# [[[[ 0. 0. 0. ... 0. 0. 0.]
# [ 1. 1. 1. ... 1. 1. 1.]
# [ 2. 2. 2. ... 2. 2. 2.]
# ...
# [93. 93. 93. ... 93. 93. 93.]
# [94. 94. 94. ... 94. 94. 94.]
# [95. 95. 95. ... 95. 95. 95.]]]]
# <NDArray 1x1x96x96 @gpu(0)>
# (Pdb) coord_cols
# [[[[ 0. 1. 2. ... 93. 94. 95.]
# [ 0. 1. 2. ... 93. 94. 95.]
# [ 0. 1. 2. ... 93. 94. 95.]
# ...
# [ 0. 1. 2. ... 93. 94. 95.]
# [ 0. 1. 2. ... 93. 94. 95.]
# [ 0. 1. 2. ... 93. 94. 95.]]]]
# <NDArray 1x1x96x96 @gpu(0)>
# (Pdb) add_xy
# [[[[61. 61. 61. ... 61. 61. 61.]
# [61. 61. 61. ... 61. 61. 61.]
# [61. 61. 61. ... 61. 61. 61.]
# ...
# [61. 61. 61. ... 61. 61. 61.]
# [61. 61. 61. ... 61. 61. 61.]
# [61. 61. 61. ... 61. 61. 61.]]
# [[71. 71. 71. ... 71. 71. 71.]
# [71. 71. 71. ... 71. 71. 71.]
# [71. 71. 71. ... 71. 71. 71.]
# ...
# [71. 71. 71. ... 71. 71. 71.]
# [71. 71. 71. ... 71. 71. 71.]
# [71. 71. 71. ... 71. 71. 71.]]]]
# <NDArray 1x2x96x96 @gpu(0)>
# (Pdb) if self.append_dist, then coord_features is:
# [[[[-1. -1. -1. ... -1. -1.
# -1. ]
# [-1. -1. -1. ... -1. -1.
# -1. ]
# [-1. -1. -1. ... -1. -1.
# -1. ]
# ...
# [ 0.7619048 0.7619048 0.7619048 ... 0.7619048 0.7619048
# 0.7619048 ]
# [ 0.78571427 0.78571427 0.78571427 ... 0.78571427 0.78571427
# 0.78571427]
# [ 0.8095238 0.8095238 0.8095238 ... 0.8095238 0.8095238
# 0.8095238 ]]
# [[-1. -1. -1. ... 0.52380955 0.54761904
# 0.5714286 ]
# [-1. -1. -1. ... 0.52380955 0.54761904
# 0.5714286 ]
# [-1. -1. -1. ... 0.52380955 0.54761904
# 0.5714286 ]
# ...
# [-1. -1. -1. ... 0.52380955 0.54761904
# 0.5714286 ]
# [-1. -1. -1. ... 0.52380955 0.54761904
# 0.5714286 ]
# [-1. -1. -1. ... 0.52380955 0.54761904
# 0.5714286 ]]
# [[ 1. 1. 1. ... 1. 1.
# 1. ]
# [ 1. 1. 1. ... 1. 1.
# 1. ]
# [ 1. 1. 1. ... 1. 1.
# 1. ]
# ...
# [ 1. 1. 1. ... 0.9245947 0.9382886
# 0.95238096]
# [ 1. 1. 1. ... 0.944311 0.9577231
# 0.9715336 ]
# [ 1. 1. 1. ... 0.96421224 0.97735125
# 0.99088824]]]]
# <NDArray 1x3x96x96 @gpu(0)>
pdb.set_trace()
row_array = F.arange(start=0, stop=rows, step=1, **ctx_kwarg) ## (96,)
col_array = F.arange(start=0, stop=cols, step=1, **ctx_kwarg) ## (96,)
coord_rows = F.repeat(F.reshape(row_array, (1, 1, rows, 1)), repeats=cols, axis=3)
coord_cols = F.repeat(F.reshape(col_array, (1, 1, 1, cols)), repeats=rows, axis=2)
coord_rows = F.repeat(coord_rows, repeats=batch_size, axis=0)
coord_cols = F.repeat(coord_cols, repeats=batch_size, axis=0)
coords = F.concat(coord_rows, coord_cols, dim=1) ## (1, 2, 96, 96)
add_xy = F.reshape(points * self.spatial_scale, shape=(0, 0, 1)) ## [[[61.] [71.]]] <NDArray 1x2x1 @gpu(0)>
add_xy = F.reshape(F.repeat(add_xy, rows * cols, axis=2), shape=(0, 0, rows, cols))
## self.norm_radius: 42
coords = (coords - add_xy) / (self.norm_radius * self.spatial_scale) ## <NDArray 1x2x96x96 @gpu(0)>
if self.append_dist:
dist = F.sqrt(F.sum(F.square(coords), axis=1, keepdims=1)) ## <NDArray 1x1x96x96 @gpu(0)>
coord_features = F.concat(coords, dist, dim=1)
else:
coord_features = coords
coord_features = F.clip(coord_features, a_min=-1, a_max=1)
return coord_features
I also write one PyTorch version according to the MXNet version:
class AddCoords(nn.Module):
def __init__(self, ):
super().__init__()
def forward(self, input_tensor, points):
_, x_dim, y_dim = input_tensor.size()
batch_size = 1
xx_channel = torch.arange(x_dim).repeat(1, y_dim, 1) ## torch.Size([1, 9, 9])
yy_channel = torch.arange(y_dim).repeat(1, x_dim, 1).transpose(1, 2) ## torch.Size([1, 9, 9])
xx_channel = xx_channel.repeat(batch_size, 1, 1, 1).transpose(2, 3)
yy_channel = yy_channel.repeat(batch_size, 1, 1, 1).transpose(2, 3)
coords = torch.cat((xx_channel, yy_channel), dim=1) ## torch.Size([20, 2, 9, 9])
coords = coords.type(torch.FloatTensor)
add_xy = torch.reshape(points, (1, 2, 1)) ## torch.Size([1, 2, 1])
add_xy_ = add_xy.repeat(1, 1, x_dim * y_dim) ## torch.Size([1, 2, 81])
add_xy_ = torch.reshape(add_xy_, (1, 2, x_dim, y_dim)) ## torch.Size([1, 2, 9, 9])
add_xy_ = add_xy_.type(torch.FloatTensor)
coords = (coords - add_xy_) ## torch.Size([1, 2, 9, 9])
coord_features = np.clip(np.array(coords), -1, 1) ## (1, 2, 9, 9)
coord_features = torch.from_numpy(coord_features).cuda()
return coord_features
An intriguing failing of convolutional neural networks and the CoordConv solution的更多相关文章
- Understanding the Effective Receptive Field in Deep Convolutional Neural Networks
Understanding the Effective Receptive Field in Deep Convolutional Neural Networks 理解深度卷积神经网络中的有效感受野 ...
- Deep learning_CNN_Review：A Survey of the Recent Architectures of Deep Convolutional Neural Networks——2019
CNN综述文章 的翻译 [2019 CVPR] A Survey of the Recent Architectures of Deep Convolutional Neural Networks 翻 ...
- tensorfolw配置过程中遇到的一些问题及其解决过程的记录（配置SqueezeDet: Unified, Small, Low Power Fully Convolutional Neural Networks for Real-Time Object Detection for Autonomous Driving）
今天看到一篇关于检测的论文<SqueezeDet: Unified, Small, Low Power Fully Convolutional Neural Networks for Real- ...
- Notes on Convolutional Neural Networks
这是Jake Bouvrie在2006年写的关于CNN的训练原理,虽然文献老了点,不过对理解经典CNN的训练过程还是很有帮助的.该作者是剑桥的研究认知科学的.翻译如有不对之处,还望告知,我好及时改正, ...
- 《ImageNet Classification with Deep Convolutional Neural Networks》 剖析
<ImageNet Classification with Deep Convolutional Neural Networks> 剖析 CNN 领域的经典之作, 作者训练了一个面向数量为 ...
- 卷积神经网络CNN(Convolutional Neural Networks)没有原理只有实现
零.说明: 本文的所有代码均可在 DML 找到,欢迎点星星. 注.CNN的这份代码非常慢,基本上没有实际使用的可能,所以我只是发出来,代表我还是实践过而已 一.引入: CNN这个模型实在是有些年份了, ...
- A Beginner's Guide To Understanding Convolutional Neural Networks(转)
A Beginner's Guide To Understanding Convolutional Neural Networks Introduction Convolutional neural ...
- 阅读笔记 The Impact of Imbalanced Training Data for Convolutional Neural Networks [DegreeProject2015] 数据分析型
The Impact of Imbalanced Training Data for Convolutional Neural Networks Paulina Hensman and David M ...
- 读convolutional Neural Networks Applied to House Numbers Digit Classification 的收获。
本文以下内容来自读论文以后认为有价值的地方,论文来自:convolutional Neural Networks Applied to House Numbers Digit Classificati ...
随机推荐
- 记录下登录Xcode 时出现 please sign in with an app-specific....
今天做到准备提交到APP store时,要用xcode提交,在xcode->open developer tool->application loader时,要登录自己的ID,我登录时输入 ...
- CSS 基础样式
文本 p{ font-family:Cambria, "Hoefler Text", "Liberation Serif", Times, "Time ...
- Django 之 cookie & session
Cookie的由来 大家都知道HTTP协议是无状态的. 无状态的意思是每次请求都是独立的,它的执行情况和结果与前面的请求和之后的请求都无直接关系,它不会受前面的请求响应情况直接影响,也不会直接影响后面 ...
- spring cloud微服务实战教程/pdf/视频/百度云资源
资源站:http://www.supan.vip 点击进入直接查找资源: http://www.supan.vip/spring%20cloud微服务实战 <Spring Cloud微服务实战& ...
- Linux运维技术之scp命令使用
命令格式: scp -P 8031 -r 路径1 路径2 命令解释, 其中-P中的P一定要大写,小写不行(这个也是问题的关键,这个大小写耽搁了好长时间) -P 2424表示更改SSH端口后的端口,如 ...
- Java Mail 异常
Java Mail 异常 java.lang.NoClassDefFoundError: com/sun/mail/util/LineInputStream 解决方案: 把Jar包换一个比较高的版本就 ...
- 201671010425邱世妍 团队评审&课程总结
实验十四 团队项目评审&课程学习总结 项目 内容 这个作业属于哪个课程 http://www.cnblogs.com/nwnu-daizh/ 这个作业的要求在哪里 https://www.cn ...
- live555流媒体框架介绍
LIVE555 Streaming Media This code forms a set of C++ libraries for multimedia streaming, using open ...
- sublime中文解决
1.写好文件sublime_imfix.c并保存. #include <gtk/gtkimcontext.h> void gtk_im_context_set_client_window ...
- 用mysql实现类似于oracle dblink的功能
用mysql实现类似于oracle dblink的功能 首先看看有没有federated 引擎. mysql> show engines; +------------+----------+ ...