search

bastiaanv / Yolov2-tiny-tf-NCS

Yolov2-tiny Tensorflow implementation for Movidius Neural Compute Stick
MIT License
5 stars 2 forks source link

Training on Tiny Yolo V2 #1

Open izesaon opened 5 years ago

izesaon commented 5 years ago

Hi,

Thanks for the great repository. I am currently building a model to be retrained for my own dataset and wanted to clarify how I should go about retraining it using your repository if the support is provided?

bastiaanv commented 5 years ago

Hi,

Thank you for your interest in my repo!

As it is now, I do not support retraining in my repo using a automatic system or something like that. You can, however, eddit the weights file to something that would fit better on your needs.

I assume you want a system, in which you can put in a lot of images and it would adjusts it's weights and biases. In the YoloV2 system, you have 2 systems build in: 1. Image classification 2. Object location system. I have made, some time ago, code that would make backprop possible for the image classification, but I have not been succesfull in creating the same for the object location system...

So as it stands, it is not possible to retrain tge network using your own images/dataset. It is possible to use a different weights file that fits the networks shape and your needs

Op di 11 dec. 2018 08:42 schreef Cindy <notifications@github.com:

Hi,

Thanks for the great repository. I am currently building a model to be retrained for my own dataset and wanted to clarify how I should go about retraining it using your repository if the support is provided?

— You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub https://github.com/bastiaanv/Yolov2-tiny-tf-NCS/issues/1, or mute the thread https://github.com/notifications/unsubscribe-auth/ADzZXKmHpalWPe3zXtrCpaVc5WaKXg6nks5u32H2gaJpZM4ZM1KV .

izesaon commented 5 years ago

Hi,

Does this mean that the weights I generated from retraining yolo tiny v2 on another dataset could be used for your repository?

On 11 Dec 2018, at 10:58 PM, Bastiaan Verhaar notifications@github.com wrote:

Hi,

Thank you for your interest in my repo!

As it is now, I do not support retraining in my repo using a automatic system or something like that. You can, however, eddit the weights file to something that would fit better on your needs.

I assume you want a system, in which you can put in a lot of images and it would adjusts it's weights and biases. In the YoloV2 system, you have 2 systems build in: 1. Image classification 2. Object location system. I have made, some time ago, code that would make backprop possible for the image classification, but I have not been succesfull in creating the same for the object location system...

So as it stands, it is not possible to retrain tge network using your own images/dataset. It is possible to use a different weights file that fits the networks shape and your needs

Op di 11 dec. 2018 08:42 schreef Cindy <notifications@github.com:

Hi,

Thanks for the great repository. I am currently building a model to be retrained for my own dataset and wanted to clarify how I should go about retraining it using your repository if the support is provided?

— You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub https://github.com/bastiaanv/Yolov2-tiny-tf-NCS/issues/1, or mute the thread https://github.com/notifications/unsubscribe-auth/ADzZXKmHpalWPe3zXtrCpaVc5WaKXg6nks5u32H2gaJpZM4ZM1KV .

— You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub https://github.com/bastiaanv/Yolov2-tiny-tf-NCS/issues/1#issuecomment-446232033, or mute the thread https://github.com/notifications/unsubscribe-auth/AP3QrKQ7yoLMsYcJqCWuFjn-gF65Vc80ks5u38gfgaJpZM4ZM1KV.

bastiaanv commented 5 years ago

Yes it can! I pick the one on the yolo website, but theory ever file generated using the yolov2 tiny network can be loaded into the repo

Op di 11 dec. 2018 16:40 schreef Cindy <notifications@github.com:

Hi,

Does this mean that the weights I generated from retraining yolo tiny v2 on another dataset could be used for your repository?

On 11 Dec 2018, at 10:58 PM, Bastiaan Verhaar notifications@github.com wrote:

Hi,

Thank you for your interest in my repo!

As it is now, I do not support retraining in my repo using a automatic system or something like that. You can, however, eddit the weights file to something that would fit better on your needs.

I assume you want a system, in which you can put in a lot of images and it would adjusts it's weights and biases. In the YoloV2 system, you have 2 systems build in: 1. Image classification 2. Object location system. I have made, some time ago, code that would make backprop possible for the image classification, but I have not been succesfull in creating the same for the object location system...

So as it stands, it is not possible to retrain tge network using your own images/dataset. It is possible to use a different weights file that fits the networks shape and your needs

Op di 11 dec. 2018 08:42 schreef Cindy <notifications@github.com:

Hi,

Thanks for the great repository. I am currently building a model to be retrained for my own dataset and wanted to clarify how I should go about retraining it using your repository if the support is provided?

— You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub https://github.com/bastiaanv/Yolov2-tiny-tf-NCS/issues/1, or mute the thread < https://github.com/notifications/unsubscribe-auth/ADzZXKmHpalWPe3zXtrCpaVc5WaKXg6nks5u32H2gaJpZM4ZM1KV

.

— You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub < https://github.com/bastiaanv/Yolov2-tiny-tf-NCS/issues/1#issuecomment-446232033>, or mute the thread < https://github.com/notifications/unsubscribe-auth/AP3QrKQ7yoLMsYcJqCWuFjn-gF65Vc80ks5u38gfgaJpZM4ZM1KV .

— You are receiving this because you commented. Reply to this email directly, view it on GitHub https://github.com/bastiaanv/Yolov2-tiny-tf-NCS/issues/1#issuecomment-446247558, or mute the thread https://github.com/notifications/unsubscribe-auth/ADzZXAOClJMVuNvEcKhaVkGnawI3LuiAks5u39HdgaJpZM4ZM1KV .

izesaon commented 5 years ago

Hi,

Sorry I am not too sure what you mean. But I suppose that since changing the weights is obtained when I am retraining it on another dataset, your repository would come in handy for me when I do live object detection with the stick as it can identify the objects I trained it for:) The only changes to be done would be the weights, the classes and colors right? Is there any reason for the 8 convolution layers or they are part of the Tiny Yolo V2 Network?

On 11 Dec 2018, at 11:44 PM, Bastiaan Verhaar notifications@github.com wrote:

Yes it can! I pick the one on the yolo website, but theory ever file generated using the yolov2 tiny network can be loaded into the repo

Op di 11 dec. 2018 16:40 schreef Cindy <notifications@github.com:

Hi,

Does this mean that the weights I generated from retraining yolo tiny v2 on another dataset could be used for your repository?

On 11 Dec 2018, at 10:58 PM, Bastiaan Verhaar notifications@github.com wrote:

Hi,

Thank you for your interest in my repo!

As it is now, I do not support retraining in my repo using a automatic system or something like that. You can, however, eddit the weights file to something that would fit better on your needs.

I assume you want a system, in which you can put in a lot of images and it would adjusts it's weights and biases. In the YoloV2 system, you have 2 systems build in: 1. Image classification 2. Object location system. I have made, some time ago, code that would make backprop possible for the image classification, but I have not been succesfull in creating the same for the object location system...

So as it stands, it is not possible to retrain tge network using your own images/dataset. It is possible to use a different weights file that fits the networks shape and your needs

Op di 11 dec. 2018 08:42 schreef Cindy <notifications@github.com:

Hi,

Thanks for the great repository. I am currently building a model to be retrained for my own dataset and wanted to clarify how I should go about retraining it using your repository if the support is provided?

— You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub https://github.com/bastiaanv/Yolov2-tiny-tf-NCS/issues/1, or mute the thread < https://github.com/notifications/unsubscribe-auth/ADzZXKmHpalWPe3zXtrCpaVc5WaKXg6nks5u32H2gaJpZM4ZM1KV

.

— You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub < https://github.com/bastiaanv/Yolov2-tiny-tf-NCS/issues/1#issuecomment-446232033>, or mute the thread < https://github.com/notifications/unsubscribe-auth/AP3QrKQ7yoLMsYcJqCWuFjn-gF65Vc80ks5u38gfgaJpZM4ZM1KV .

— You are receiving this because you commented. Reply to this email directly, view it on GitHub https://github.com/bastiaanv/Yolov2-tiny-tf-NCS/issues/1#issuecomment-446247558, or mute the thread https://github.com/notifications/unsubscribe-auth/ADzZXAOClJMVuNvEcKhaVkGnawI3LuiAks5u39HdgaJpZM4ZM1KV .

— You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub https://github.com/bastiaanv/Yolov2-tiny-tf-NCS/issues/1#issuecomment-446249252, or mute the thread https://github.com/notifications/unsubscribe-auth/AP3QrJXmQ4EcZp8SNC94DLb7gR2gy84Kks5u39LmgaJpZM4ZM1KV.

izesaon commented 5 years ago

hi, what does

load_conv_layer_bn do? is this section of code specific to the number of classes you have? because when i loaded my pretrained weights, the kernel weights generated using this function returns nan for certain values.

izesaon commented 5 years ago

hi, which part of the code is specific to the weights using darknet? Because when I tried to load my own weights, I get a lot of nan values for prediction but for the simo23 github repository using his code, i didn't get any nan values. Sorry, just wanted to understand the code better as I do not understand every single part.