summer of 2016错误及解决汇总(1)

最新推荐文章于 2020-03-22 00:48:12 发布

转载最新推荐文章于 2020-03-22 00:48:12 发布 · 143 阅读

0 ·

CC 4.0 BY-SA版权

原文链接：http://www.cnblogs.com/chanyn/p/5796407.html

文章标签：

#人工智能 #python

本文针对AlexNet训练过程中遇到的问题进行详细分析，包括prototxt文件格式错误导致的网络读取失败、plt.imshow()使用不当引发的TypeError以及训练时网络无法收敛等问题。通过参数可视化发现卷积层参数异常，并提出解决方案。

读入Net时发生：
Error parsing text-format caffe.NetParameter: 64:7: Message type "caffe.LayerParameter" has no field named "layer".
Reson：
prototxt文件格式出错，line 64 缺少‘}’

python Visualize param：
=>plt.imshow()
TypeError: Invalid dimensions for image data
Reason：
plt.imshow() needs a 2D array (or a 3D array with the third dimension being of shape 3 or 4)!

# Assuming your array/image has the variable name: array

import numpy as np
np.asarray(array).ndim 
# If the result is 2 you can always use it with imshow

# If the result is 3 you need to check the third dimension:
np.asarray(array).shape[2] # Must be 3 or 4 for RGB or RGBA images

Train a reformed AlexNet，keep conv layer input size，transform fc layers to conv which kernel size is 1*1.But loss is shocking and Net cannot converge.

By visualizing each Layer parameters and Blob，find conv6/conv7(kernel = 1*1) parameters are all zero!

Reason:

conv6/conv7 in train_val.prototxt cannot initialize.

convolution_param {
num_output: 1
bias_term: false
kernel_size: 16
stride: 8
pad: 4

weight_filler {
type: "gaussian"
std: 0.01
}
bias_filler {
type: "constant"
value: 0.1
}
}

Or:利用已有的caffemodel fine-tuning

在训练脚本中加上：-weights=models/bvlc_alexnet/bvlc_alexnet.caffemodel

转载于:https://www.cnblogs.com/chanyn/p/5796407.html