About run_cifar10

[MModerato]

After we execute the command >> run_cifar10, we get the >> MAP = 0.897373
Is this the result after reranking ? I mean the Fine-level Search.
@kevinlin311tw

[kevinlin311tw]

It's coarse-level search. In MNIST and CIFAR10 experiments, we evaluate the performance with only binary codes in order to fairly compare with other hashing approaches.

[MModerato]

name: "KevinNet_CIFAR10"
layers {
layer {
    name: "data"
    type: "data"
    source: "cifar10_train_leveldb"
    meanfile: "../../data/ilsvrc12/imagenet_mean.binaryproto"
    batchsize: 32
    cropsize: 227
    mirror: true
    det_context_pad: 16
    det_crop_mode: "warp"
    det_fg_threshold: 0.5
    det_bg_threshold: 0.5
    det_fg_fraction: 0.25
  }
  top: "data"
  top: "label"
}

In the train_CIFAR10_48.prototxt, why cropsize is 227? While the size of image in cifar is 32*32.

And when I used Netscope to view train_CIFAR10_48.prototxt, the error was reported:

Error Encountered
Uncaught TypeError: Cannot read property 'replace' of undefined


Warning
Can't infer network data shapes. Can't infer output shape of the 'undefined' layer of type 'undefined'. 
TypeError: Cannot read property 'split' of undefined

Thank you. @kevinlin311tw

[kevinlin311tw]

We first resize images to 256x256, and then center-crop 227x227 as network input.

…

On 2017年7月28日, at 上午2:17, 111Moderato ***@***.***> wrote: name: "KevinNet_CIFAR10" layers { layer { name: "data" type: "data" source: "cifar10_train_leveldb" meanfile: "../../data/ilsvrc12/imagenet_mean.binaryproto" batchsize: 32 cropsize: 227 mirror: true det_context_pad: 16 det_crop_mode: "warp" det_fg_threshold: 0.5 det_bg_threshold: 0.5 det_fg_fraction: 0.25 } top: "data" top: "label" } In the train_CIFAR10_48.prototxt, why cropsize is 227? Is the size of cifar 32*32 ? — You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub, or mute the thread.

[MModerato]

Do you have an experiment on NUS-WIDE ? @kevinlin311tw

[kevinlin311tw]

Yes. You can take a look the extension of this work here: https://arxiv.org/abs/1507.00101v2 Our workshop version can be seen as our proposed SSDH with the hyper parameters: alpha = 1, beta = 0, gamma = 0. 2017-09-05 22:02 GMT-07:00 111Moderato <notifications@github.com>:

…

Do you have an experiment on NUS-WIDE ? @kevinlin311tw <https://github.com/kevinlin311tw> — You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub <#22 (comment)>, or mute the thread <https://github.com/notifications/unsubscribe-auth/AJgtMLoJ5dPSNUOnpX0R3ARU-3yxQBmMks5sfid_gaJpZM4OhMvK> .

-- Best regards, 林可昀 Kevin Lin

[MModerato]

Uh.. Is it correct for me to use SigmoidCrossEntropyLoss on multi-label dataset? @kevinlin311tw

[kevinlin311tw]

Oh.. I forgot that NUS-WIDE is a multi-label dataset.. Yes. You are right. Since this workshop version is trained with softmax loss, we cannot directly apply it on multi-label dataset. In this case, since the label is the n-hot vector, we should use the loss function you mentioned. 2017-09-07 10:02 GMT-07:00 111Moderato <notifications@github.com>:

…

Uh.. Is it correct for me to use SigmoidCrossEntropyLoss on multi-label dataset? @kevinlin311tw <https://github.com/kevinlin311tw> — You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub <#22 (comment)>, or mute the thread <https://github.com/notifications/unsubscribe-auth/AJgtMPYJSnq4dfm_th_70-dnObyZFLRyks5sgCGrgaJpZM4OhMvK> .

-- Best regards, 林可昀 Kevin Lin

[lshignal]

@111Moderato I also meet this mistick. Do you fix it?
Error Encountered
Uncaught TypeError: Cannot read property 'replace' of undefined

Warning
Can't infer network data shapes. Can't infer output shape of the 'undefined' layer of type 'undefined'.
TypeError: Cannot read property 'split' of undefined

[MModerato]

I forgot whether I had met the mistake. I'm using another version of Caffe, so I can run it properly after the prototxt changes

[lshignal]

@111Moderato Thanks a lot!
However, this error is occurred when using Netscope to view train_CIFAR10_48.prototxt. You asked this question, but author does not answered :)

[mudi-1125]

The baidu link of model and image resource is invalid, can anyone share it?