problem in test result

[LegendBC]

Thanks for your great work! And I haved trained your model on pascal2007, but the result in test shows no class only background. I haven't changed anything else except for the function get_data(input_path). Can you give me some advice? thanks!

[LegendBC]

It's of great help that you give a more specific readme about your environment configuration.

[wusonglin-ali]

It's of great help that you give a more specific readme about your environment configuration.

你好，你现在解决这个问题了吗

[hsl8856]

Thanks for your great work! And I haved trained your model on pascal2007, but the result in test shows no class only background. I haven't changed anything else except for the function get_data(input_path). Can you give me some advice? thanks!

The test results are all background questions.,. Have you solved it? What is the final convergence result of your training loss function(RPN,CLass)? thanks

[hlq1025]

I also encounter this problem .Does anyone know how to solve this problem?

[cfeeney5]

There is a bug in the PSRoiAlignPooling function in detector.py in the latest commit. It will work if the RoiPoolingConv function is used instead.

[Threeki1019]

There is a bug in the PSRoiAlignPooling function in detector.py in the latest commit. It will work if the RoiPoolingConv function is used instead.

hi! can you tell me how to use RoiPoolingConv function

[cfeeney5]

@Threeki1019 I didn't fix it I just used RoiPoolingConv function instead

[Threeki1019]

yeah but i test.py also only bg

------------------ Original ------------------ From: cfeeney5 <[email protected]> Date: Fri,Feb 21,2020 3:09 AM To: mohhao/TF-Keras-ThunderNet <[email protected]> Cc: XP <[email protected]>, Mention <[email protected]> Subject: Re: [mohhao/TF-Keras-ThunderNet] problem in test result (#2)

@Threeki1019 I didn't fix it I just used RoiPoolingConv function instead

— You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub, or unsubscribe.

[Threeki1019]

if use RoiPoolingConv function ，you get the correct results ？

------------------ Original ------------------ From: cfeeney5 <[email protected]> Date: Fri,Feb 21,2020 3:09 AM To: mohhao/TF-Keras-ThunderNet <[email protected]> Cc: XP <[email protected]>, Mention <[email protected]> Subject: Re: [mohhao/TF-Keras-ThunderNet] problem in test result (#2)

@Threeki1019 I didn't fix it I just used RoiPoolingConv function instead

— You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub, or unsubscribe.

[cfeeney5]

Hi, I changed over ROIPoolingConv and I could train the network but the results were only OK. The code suggests initializes the weights from VGG-16 but I dont think it is correct because the layer names do not match. How many images are in your training set? The bounding box threshold in this commit is 0 (see https://github.com/mohhao/TF-Keras-ThunderNet/blob/d85ccf9f44b9470ff8459ba8c47f2ec6cc6a3e4a/test.py#L130 ) which should be 0.5 or so.

[Threeki1019]

I used your advice, but my training result loss has been decreasing, but the ACC remains the same. Train on voc2007

------------------ Original ------------------ From: cfeeney5 <[email protected]> Date: Fri,Feb 21,2020 9:55 PM To: mohhao/TF-Keras-ThunderNet <[email protected]> Cc: XP <[email protected]>, Mention <[email protected]> Subject: Re: [mohhao/TF-Keras-ThunderNet] problem in test result (#2)

Hi, I changed over ROIPoolingConv and I could train the network but the results were only OK. The code suggests initializes the weights from VGG-16 but I dont think it is correct because the layer names do not match. How many images are in your training set? The bounding box threshold in this commit is 0 (see https://github.com/mohhao/TF-Keras-ThunderNet/blob/d85ccf9f44b9470ff8459ba8c47f2ec6cc6a3e4a/test.py#L130 ) which should be 0.5 or so.

— You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub, or unsubscribe.

[Threeki1019]

i used the vgg16weight，but i think it didnt work。

------------------ Original ------------------ From: cfeeney5 <[email protected]> Date: Fri,Feb 21,2020 9:55 PM To: mohhao/TF-Keras-ThunderNet <[email protected]> Cc: XP <[email protected]>, Mention <[email protected]> Subject: Re: [mohhao/TF-Keras-ThunderNet] problem in test result (#2)

Hi, I changed over ROIPoolingConv and I could train the network but the results were only OK. The code suggests initializes the weights from VGG-16 but I dont think it is correct because the layer names do not match. How many images are in your training set? The bounding box threshold in this commit is 0 (see https://github.com/mohhao/TF-Keras-ThunderNet/blob/d85ccf9f44b9470ff8459ba8c47f2ec6cc6a3e4a/test.py#L130 ) which should be 0.5 or so.

— You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub, or unsubscribe.