文章作者：Tyan
博客：noahsnail.com | CSDN | 简书
在使用Caffe模型进行图像分类时，通常使用的图像大小为224 * 224，如果改变输入图像的大小，有时候会碰到Check failed: shape[i] >= 0 (-1 vs. 0)错误。
- 背景
使用ResNet-50进行图像分类训练，输入图像的大小为160 *160。
- 错误
I1013 10:53:03.395526 76959 net.cpp:129] Top shape: 16 2048 5 5 (819200)
I1013 10:53:03.395531 76959 net.cpp:137] Memory required for data: 1889075264
I1013 10:53:03.395539 76959 layer_factory.hpp:77] Creating layer res5c_relu
I1013 10:53:03.395553 76959 net.cpp:84] Creating Layer res5c_relu
I1013 10:53:03.395567 76959 net.cpp:406] res5c_relu <- res5c
I1013 10:53:03.395594 76959 net.cpp:367] res5c_relu -> res5c (in-place)
I1013 10:53:03.395841 76959 net.cpp:122] Setting up res5c_relu
I1013 10:53:03.395855 76959 net.cpp:129] Top shape: 16 2048 5 5 (819200)
I1013 10:53:03.395861 76959 net.cpp:137] Memory required for data: 1892352064
I1013 10:53:03.395869 76959 layer_factory.hpp:77] Creating layer pool5
I1013 10:53:03.395900 76959 net.cpp:84] Creating Layer pool5
I1013 10:53:03.395908 76959 net.cpp:406] pool5 <- res5c
I1013 10:53:03.395931 76959 net.cpp:380] pool5 -> pool5
F1013 10:53:03.396807 76959 blob.cpp:32] Check failed: shape[i] >= 0 (-1 vs. 0)
*** Check failure stack trace: ***
@ 0x7fa7eef89e6d (unknown)
@ 0x7fa7eef8bced (unknown)
@ 0x7fa7eef89a5c (unknown)
@ 0x7fa7eef8c63e (unknown)
@ 0x7fa7f6141829 caffe::Blob<>::Reshape()
@ 0x7fa7f6141581 caffe::Blob<>::Reshape()
@ 0x7fa7f6200958 caffe::PoolingLayer<>::Reshape()
@ 0x7fa7f61b14d8 caffe::CuDNNPoolingLayer<>::Reshape()
@ 0x7fa7f6167d29 caffe::Layer<>::SetUp()
@ 0x7fa7f6240213 caffe::Net<>::Init()
@ 0x7fa7f623e7ea caffe::Net<>::Net()
@ 0x7fa7f62711e1 caffe::Solver<>::InitTrainNet()
@ 0x7fa7f6270a75 caffe::Solver<>::Init()
@ 0x7fa7f62705a7 caffe::Solver<>::Solver()
@ 0x7fa7f627dc66 caffe::SGDSolver<>::SGDSolver()
@ 0x7fa7f6287ab1 caffe::Creator_SGDSolver<>()
@ 0x4215f5 caffe::SolverRegistry<>::CreateSolver()
@ 0x41bdce train()
@ 0x41e167 main
@ 0x7fa7e20eeb35 __libc_start_main
@ 0x41aa69 (unknown)
- 解决方案
错误日志中可以看出问题出在pool5层，它的输入为Top shape: 16 2048 5 5 (819200)，即16 * 2048 * 5 * 5，16是batch_size，2048为channel数量，即kernel个数，feature map大小为5 * 5，在prototxt中找到pool5层，如下：
layer {
bottom: "res5c"
top: "pool5"
name: "pool5"
type: "Pooling"
pooling_param {
kernel_size: 7
stride: 1
pool: AVE
}
}
kernel size为7，比5大，因此将7改为5，重新训练，没问题。
