神经网络学习工具
目前就职于人工智能企业，日常工作内容集中在模型在不同平台的适配、模型转换、量化等方面，虽然与模型相关，但是对于模型结构、原理的研究不够深入，希望能够在闲暇时间积累一些关于模型结构和原理方面的知识，跟进一下比较前沿的进展。但在此过程中，发现在：
1、知识获取；
2、代码复现
中都存在不少的障碍，主要障碍是：
1、没有足够时间去跟进最新的进展（日常工作中这种需求也不强）；
2、各个框架（尤其pytorch）的可视化支持不够；
3、不够熟悉相关框架的常用函数；
针对这三个问题，有以下解决思路：问题3，只能靠文档和阅读别人代码来解决；问题1目前跟进https://www.paperswithcode.com/即可；针对问题2，总结了一个可行的学习路径。
1、在有可视化需求的场合尽量使用tensorflow，可以很方便的看到模型结构，方法见后文；
2、pytorch可视化小工具：https://codechina.csdn.net/lvsolo/pytorch-summary
对于已经有复现代码的论文，尽量找到tf的实现，在tensorflow中可视化：
from tensorflow.python.framework import graph_util
output_graph_def = graph_util.convert_variables_to_constants(retinanet.sess,retinanet.sess.graph_def,output_node_names = \
['regressor/conv2d_4/BiasAdd','regressor/conv2d_9/BiasAdd'])
with tf.gfile.GFile('retina.pb', "wb") as f:
f.write(output_graph_def.SerializeToString())
保存下的pb文件可以用netron打开，唯一需要预先获得的参数是output_node_names，在对论文研究后可以在代码中找到输出节点，
print(output_node.name)
即可获得输出节点名称。
pytorch可视化目前可用的工具：https://codechina.csdn.net/lvsolo/pytorch-summary
目前可以有以下类似的tensor shape输出，暂时没有对模型连接结构的输出：
----------------------------------------------------------------
Layer (type) Output Shape Param #
================================================================
Conv2d-1 [-1, 64, 112, 112] 9,408
BatchNorm2d-2 [-1, 64, 112, 112] 128
ReLU-3 [-1, 64, 112, 112] 0
MaxPool2d-4 [-1, 64, 56, 56] 0
Conv2d-5 [-1, 64, 56, 56] 36,928
BatchNorm2d-6 [-1, 64, 56, 56] 128
ReLU-7 [-1, 64, 56, 56] 0
BasicBlock-8 [-1, 64, 56, 56] 0
Conv2d-9 [-1, 64, 56, 56] 36,928
BatchNorm2d-10 [-1, 64, 56, 56] 128
ReLU-11 [-1, 64, 56, 56] 0
BasicBlock-12 [-1, 64, 56, 56] 0
Conv2d-13 [-1, 128, 28, 28] 73,856
BatchNorm2d-14 [-1, 128, 28, 28] 256
Conv2d-15 [-1, 128, 28, 28] 8,192
BatchNorm2d-16 [-1, 128, 28, 28] 256
ReLU-17 [-1, 128, 28, 28] 0
BasicBlock-18 [-1, 128, 28, 28] 0
Conv2d-19 [-1, 128, 28, 28] 147,584
BatchNorm2d-20 [-1, 128, 28, 28] 256
ReLU-21 [-1, 128, 28, 28] 0
BasicBlock-22 [-1, 128, 28, 28] 0
Conv2d-23 [-1, 256, 14, 14] 295,168
BatchNorm2d-24 [-1, 256, 14, 14] 512
Conv2d-25 [-1, 256, 14, 14] 32,768
BatchNorm2d-26 [-1, 256, 14, 14] 512
ReLU-27 [-1, 256, 14, 14] 0
BasicBlock-28 [-1, 256, 14, 14] 0
Conv2d-29 [-1, 256, 14, 14] 590,080
BatchNorm2d-30 [-1, 256, 14, 14] 512
ReLU-31 [-1, 256, 14, 14] 0
BasicBlock-32 [-1, 256, 14, 14] 0
Conv2d-33 [-1, 512, 7, 7] 1,180,160
BatchNorm2d-34 [-1, 512, 7, 7] 1,024
Conv2d-35 [-1, 512, 7, 7] 131,072
BatchNorm2d-36 [-1, 512, 7, 7] 1,024
ReLU-37 [-1, 512, 7, 7] 0
BasicBlock-38 [-1, 512, 7, 7] 0
Conv2d-39 [-1, 512, 7, 7] 2,359,808
BatchNorm2d-40 [-1, 512, 7, 7] 1,024
ReLU-41 [-1, 512, 7, 7] 0
BasicBlock-42 [-1, 512, 7, 7] 0
ResNet-43 [[-1, 128, 28, 28], [-1, 256, 14, 14], [-1, 512, 7, 7]] 0
Conv2d-44 [-1, 256, 7, 7] 131,328
Conv2d-45 [-1, 256, 7, 7] 590,080
Conv2d-46 [-1, 256, 14, 14] 65,792
Conv2d-47 [-1, 256, 14, 14] 590,080
Conv2d-48 [-1, 256, 28, 28] 33,024
Conv2d-49 [-1, 256, 28, 28] 590,080
Conv2d-50 [-1, 256, 4, 4] 1,179,904
ReLU-51 [-1, 256, 4, 4] 0
Conv2d-52 [-1, 256, 2, 2] 590,080
FeaturePyramidNet-53 [[-1, 256, 28, 28], [-1, 256, 14, 14], [-1, 256, 7, 7], [-1, 256, 4, 4], [-1, 256, 2, 2]] 0
Anchors-54 [-1, 9441, 4] 0
Conv2d-55 [-1, 256, 28, 28] 590,080
ReLU-56 [-1, 256, 28, 28] 0
Conv2d-57 [-1, 256, 28, 28] 590,080
ReLU-58 [-1, 256, 28, 28] 0
Conv2d-59 [-1, 256, 28, 28] 590,080
ReLU-60 [-1, 256, 28, 28] 0
Conv2d-61 [-1, 256, 28, 28] 590,080
ReLU-62 [-1, 256, 28, 28] 0
Conv2d-63 [-1, 18, 28, 28] 41,490
Sigmoid-64 [-1, 18, 28, 28] 0
ClassificationNet-65 [-1, 7056, 2] 0
Conv2d-66 [-1, 256, 14, 14] 590,080
ReLU-67 [-1, 256, 14, 14] 0
Conv2d-68 [-1, 256, 14, 14] 590,080
ReLU-69 [-1, 256, 14, 14] 0
Conv2d-70 [-1, 256, 14, 14] 590,080
ReLU-71 [-1, 256, 14, 14] 0
Conv2d-72 [-1, 256, 14, 14] 590,080
ReLU-73 [-1, 256, 14, 14] 0
Conv2d-74 [-1, 18, 14, 14] 41,490
Sigmoid-75 [-1, 18, 14, 14] 0
ClassificationNet-76 [-1, 1764, 2] 0
Conv2d-77 [-1, 256, 7, 7] 590,080
ReLU-78 [-1, 256, 7, 7] 0
Conv2d-79 [-1, 256, 7, 7] 590,080
ReLU-80 [-1, 256, 7, 7] 0
Conv2d-81 [-1, 256, 7, 7] 590,080
ReLU-82 [-1, 256, 7, 7] 0
Conv2d-83 [-1, 256, 7, 7] 590,080
ReLU-84 [-1, 256, 7, 7] 0
Conv2d-85 [-1, 18, 7, 7] 41,490
Sigmoid-86 [-1, 18, 7, 7] 0
ClassificationNet-87 [-1, 441, 2] 0
Conv2d-88 [-1, 256, 4, 4] 590,080
ReLU-89 [-1, 256, 4, 4] 0
Conv2d-90 [-1, 256, 4, 4] 590,080
ReLU-91 [-1, 256, 4, 4] 0
Conv2d-92 [-1, 256, 4, 4] 590,080
ReLU-93 [-1, 256, 4, 4] 0
Conv2d-94 [-1, 256, 4, 4] 590,080
ReLU-95 [-1, 256, 4, 4] 0
Conv2d-96 [-1, 18, 4, 4] 41,490
Sigmoid-97 [-1, 18, 4, 4] 0
ClassificationNet-98 [-1, 144, 2] 0
Conv2d-99 [-1, 256, 2, 2] 590,080
ReLU-100 [-1, 256, 2, 2] 0
Conv2d-101 [-1, 256, 2, 2] 590,080
ReLU-102 [-1, 256, 2, 2] 0
Conv2d-103 [-1, 256, 2, 2] 590,080
ReLU-104 [-1, 256, 2, 2] 0
Conv2d-105 [-1, 256, 2, 2] 590,080
ReLU-106 [-1, 256, 2, 2] 0
Conv2d-107 [-1, 18, 2, 2] 41,490
Sigmoid-108 [-1, 18, 2, 2] 0
ClassificationNet-109 [-1, 36, 2] 0
Conv2d-110 [-1, 256, 28, 28] 590,080
ReLU-111 [-1, 256, 28, 28] 0
Conv2d-112 [-1, 256, 28, 28] 590,080
ReLU-113 [-1, 256, 28, 28] 0
Conv2d-114 [-1, 256, 28, 28] 590,080
ReLU-115 [-1, 256, 28, 28] 0
Conv2d-116 [-1, 256, 28, 28] 590,080
ReLU-117 [-1, 256, 28, 28] 0
Conv2d-118 [-1, 36, 28, 28] 82,980
RegressionNet-119 [-1, 7056, 4] 0
Conv2d-120 [-1, 256, 14, 14] 590,080
ReLU-121 [-1, 256, 14, 14] 0
Conv2d-122 [-1, 256, 14, 14] 590,080
ReLU-123 [-1, 256, 14, 14] 0
Conv2d-124 [-1, 256, 14, 14] 590,080
ReLU-125 [-1, 256, 14, 14] 0
Conv2d-126 [-1, 256, 14, 14] 590,080
ReLU-127 [-1, 256, 14, 14] 0
Conv2d-128 [-1, 36, 14, 14] 82,980
RegressionNet-129 [-1, 1764, 4] 0
Conv2d-130 [-1, 256, 7, 7] 590,080
ReLU-131 [-1, 256, 7, 7] 0
Conv2d-132 [-1, 256, 7, 7] 590,080
ReLU-133 [-1, 256, 7, 7] 0
Conv2d-134 [-1, 256, 7, 7] 590,080
ReLU-135 [-1, 256, 7, 7] 0
Conv2d-136 [-1, 256, 7, 7] 590,080
ReLU-137 [-1, 256, 7, 7] 0
Conv2d-138 [-1, 36, 7, 7] 82,980
RegressionNet-139 [-1, 441, 4] 0
Conv2d-140 [-1, 256, 4, 4] 590,080
ReLU-141 [-1, 256, 4, 4] 0
Conv2d-142 [-1, 256, 4, 4] 590,080
ReLU-143 [-1, 256, 4, 4] 0
Conv2d-144 [-1, 256, 4, 4] 590,080
ReLU-145 [-1, 256, 4, 4] 0
Conv2d-146 [-1, 256, 4, 4] 590,080
ReLU-147 [-1, 256, 4, 4] 0
Conv2d-148 [-1, 36, 4, 4] 82,980
RegressionNet-149 [-1, 144, 4] 0
Conv2d-150 [-1, 256, 2, 2] 590,080
ReLU-151 [-1, 256, 2, 2] 0
Conv2d-152 [-1, 256, 2, 2] 590,080
ReLU-153 [-1, 256, 2, 2] 0
Conv2d-154 [-1, 256, 2, 2] 590,080
ReLU-155 [-1, 256, 2, 2] 0
Conv2d-156 [-1, 256, 2, 2] 590,080
ReLU-157 [-1, 256, 2, 2] 0
Conv2d-158 [-1, 36, 2, 2] 82,980
RegressionNet-159 [-1, 36, 4] 0
SubNet-160 [[-1, 9441, 2], [-1, 9441, 4]] 0
RetinaNet-161 [[-1], [-1]] 0
================================================================
Total params: 32,903,630
Trainable params: 32,903,630
Non-trainable params: 0
----------------------------------------------------------------
Input size (MB): 0.57
Forward/backward pass size (MB): 38997339272428.49
Params size (MB): 125.52
Estimated Total Size (MB): 38997338357760.00
----------------------------------------------------------------
1719
被折叠的 条评论
为什么被折叠?
到【灌水乐园】发言
