About the SPP layer

[WanliOuyang]

TypeError: Expected binary or unicode string , got None
in line: spp = tf.reshape(max_pool, [num_sample, -1])
How to resolve it ?

[peace195]

Which tensorflow version do you use?

You should use tensorflow version 1.4.0 or higher.

[WanliOuyang]

ok， My tensorflow version 1.1.0. maybe is lower.

[WanliOuyang]

How should the code be modified if it runs on a relatively low version of tensorflow?
Thanks

[peace195]

You can search the structure of tf.reshape function in tensorflow 1.1.0 then modify my code using this structure.

Anyway, you should upgrade your tensorflow to newest version and prepare the dataset carefully.

[WanliOuyang]

Is the model able to ascend to the pool of 3D data? for example, videos

[peace195]

Only 2D data.

Best regards,

Binh Do

[WanliOuyang]

@peace195 When I use a fixed batch size value(such as batch_size=4), the code can run, but when training, the precision of the training keeps fluctuating on a very small number(such as the mean accuracy is 0.16). Why is this?

[peace195]

@WanliOuyang I'm so sorry because I was busy these days. Did you solve that problem? I'm very appreciated that you might help me to fix it. I think you can try with large batch size and small learning rate.

[CXSunshine]

I also want to use this to deal with 3D data.I am confused with the input of the SPP（such as the size of the imag）

[hbb21st]

Hi, many thanks for your code, I have a small question about the way deal with your x = tf.placeholder('float', shape = x_train.get_shape()), as TRAINING_ITERATIONS going on, the leaking mem happens and memory usage quickly increase, how you overcome it?

[world2025]

Hi,thanks for your code,do you use the multi_scale image inputs in your experiment?

[peace195]

Yes, I do.

…

On Fri, Jul 20, 2018, 2:58 PM tkone2018 ***@***.***> wrote: Hi,thanks for your code,do you use the multi_scale image inputs in your experiment? — You are receiving this because you modified the open/close state. Reply to this email directly, view it on GitHub <#2 (comment)>, or mute the thread <https://github.com/notifications/unsubscribe-auth/AIlTsi68UCra5wTHH_l5fHKwB9N6gMeSks5uIdQOgaJpZM4SzAis> .

[world2025]

Ok,thank you.I want to use like this network to realize the image retrieval ,but I don't know how to build my own network ,can you help me ? Thank you very much.

[peace195]

Sorry about my lack of help. But I have a suggestion that you should use Pytorch with batch_size = 1 for using multi-size images as inputs.

…

On Fri, Jul 20, 2018, 3:14 PM tkone2018 ***@***.***> wrote: Ok,thank you.I want to use like this network to realize the image retrieval ,but I don't know how to build my own network ,can you help me ? Thank you very much. — You are receiving this because you modified the open/close state. Reply to this email directly, view it on GitHub <#2 (comment)>, or mute the thread <https://github.com/notifications/unsubscribe-auth/AIlTsjg7VxvKCdcfaNK5j9k9fbjvVKK1ks5uIdedgaJpZM4SzAis> .

[world2025]

Ok, thank you

[world2025]

Hello， h_strd =math.floor(float(previous_conv_size[0]) / out_pool_size[i])
w_strd = math.floor(float(previous_conv_size[1]) / out_pool_size[i])
the stride size should be math.floor not math.ceil acoording to the Spatial Pyramid Pooling in Deep Convolutional Networks for Visual Recognition paper.Maybe you think math.ceil is also right.

[ProjectDimlight]

Hello, I suggest that the shape of the tf.nn.max_pool in the for-loop should be:
ksize = [1, h_size, w_size, 1]
instead of :
ksize = [1, h_size, h_size, 1]
As the input is always in the shape of a square, this might not be a big issue, but somehow confusing?

[zjz5250]

hello，can the input data hava different shape？

[peace195]

@zjz5250 zj Yes.

[zjz5250]

&nbsp; Thanks for&nbsp; your replay！！！ &nbsp; &nbsp;&nbsp;&nbsp;&nbsp; but&nbsp; I mean that&nbsp; can I&nbsp;&nbsp; train the model&nbsp;using data with different shape？ &nbsp;&nbsp;&nbsp;&nbsp; when&nbsp; build the&nbsp; network model，the&nbsp;&nbsp;input data shape&nbsp; must be&nbsp; specified，isn't it？&nbsp;

…

------------------&nbsp;原始邮件&nbsp;------------------ 发件人:&nbsp;"Binh Do"<[email protected]&gt;; 发送时间:&nbsp;2020年1月6日(星期一) 上午10:41 收件人:&nbsp;"peace195/sppnet"<[email protected]&gt;; 抄送:&nbsp;"JunZ翟"<[email protected]&gt;;"Mention"<[email protected]&gt;; 主题:&nbsp;Re: [peace195/sppnet] About the SPP layer (#2) @zjz5250 zj Yes. — You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub, or unsubscribe.