-
Notifications
You must be signed in to change notification settings - Fork 11
New issue
Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.
By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.
Already on GitHub? Sign in to your account
Training files for DNN-based Tau-Ids #1
Conversation
A new Pull Request was created by @mbluj for branch master. @cmsbuild, @smuzaffar, @gudrutis, @mrodozov can you please review it and eventually sign? Thanks. external issue cms-sw/cmsdist#4455 |
@mbluj |
@slava77, do you refer to a particular training file (the biggest one?) or all of them? Regardless, this question is beyond my knowledge, so I redirect it to authors of the discriminants: @kandrosov, @ocolegrove could you comment please? |
@slava77 I can answer about DeepTauId/deepTau_2017v1_20L1024N.pb (the biggest file): yes, this is a model, which is optimized for inference. |
I had in mind all of them (there are just 3 with somewhat similar size). |
Right now we can probably do with just one of the two graphs (for DeepPF),
but the code as written does require both.
…On Fri, Oct 26, 2018 at 4:14 PM Slava Krutelyov ***@***.***> wrote:
@slava77 <https://github.com/slava77>, do you refer to a particular
training file (the biggest one?) or all of them? Regardless, this question
is beyond my knowledge, so I redirect it to authors of the discriminants:
@kandrosov <https://github.com/kandrosov>, @ocolegrove
<https://github.com/ocolegrove> could you comment please?
I had in mind all of them (there are just 3 with somewhat similar size).
—
You are receiving this because you were mentioned.
Reply to this email directly, view it on GitHub
<#1 (comment)>,
or mute the thread
<https://github.com/notifications/unsubscribe-auth/ANs5Z4F3JQHv3dvZ9m_DKKSQ5g0THcb5ks5uoxjAgaJpZM4X8N8S>
.
|
Hi Owen,
On Fri, Oct 26, 2018 at 4:16 PM ocolegrove ***@***.***> wrote:
Right now we can probably do with just one of the two graphs (for DeepPF),
but the code as written does require both.
yes, right. However the main question is about size of the files and in
particular if the files contain a model optimized/minimized for inference
or a training model (usually bigger).
|
I will investigate if the model size could be reduced
…On Fri, Oct 26, 2018 at 4:29 PM mbluj ***@***.***> wrote:
Hi Owen,
On Fri, Oct 26, 2018 at 4:16 PM ocolegrove ***@***.***>
wrote:
> Right now we can probably do with just one of the two graphs (for
DeepPF),
> but the code as written does require both.
yes, right. However the main question is about size of the files and in
particular if the files contain a model optimized/minimized for inference
or a training model (usually bigger).
—
You are receiving this because you were mentioned.
Reply to this email directly, view it on GitHub
<#1 (comment)>,
or mute the thread
<https://github.com/notifications/unsubscribe-auth/ANs5Z6e5WfE3cEYI0SLgjYzIcpiM1K5Kks5uoxxQgaJpZM4X8N8S>
.
|
please get in touch with BTV developers @riga @mverzett @pablodecm may be able to check quickly. |
Hi @ocolegrove , |
Thanks Marcel! I will investigate this on Monday
…On Fri, Oct 26, 2018 at 7:49 PM Marcel R. ***@***.***> wrote:
Hi @ocolegrove <https://github.com/ocolegrove> ,
in case you want to convert your model to a constant graph which reduces
the size quite a lot, I summarized the procedure here:
https://gitlab.cern.ch/mrieger/CMSSW-DNN#constant-graphs
—
You are receiving this because you were mentioned.
Reply to this email directly, view it on GitHub
<#1 (comment)>,
or mute the thread
<https://github.com/notifications/unsubscribe-auth/ANs5Z2E6TOpQe87ad6cGmsLtZ-f8M58Cks5uo0sigaJpZM4X8N8S>
.
|
Added training files with quantized weights
Updated DPF new training files quantized, compatible with 94X release
- first layer not quantized; - quantizing only kernels.
Custom quantization for deepTauId.
Update training files and labeling
Quantized DNN
Pull request #1 was updated. external issue cms-sw/cmsdist#4455 |
The most recent update replaces original training files by their quantized versions with size reduced by a factor of ~4. |
@fabiocos These are the training files needed by #25016, and therefore they must be made available to it. |
+externals |
merge |
cmsdist PR cms-sw/cmsdist#4554 includes this data package. |
Hello, |
As title says: this PR adds initial versions of training files for DNN-based Tau-Ids. Related to cms-sw/cmssw#25016