These are chat archives for beniz/deepdetect

27th
Mar 2018
but while training, I am getting output "measure" as {}
{
"status": {
"code": 200,
"msg": "OK"
},
"head": {
"method": "/train",
"job": 1,
"status": "running",
"time": 661
},
"body": {
"measure": {}
}
}
Can someone please help?
Emmanuel Benazera
@beniz
Mar 27 2018 06:10
Hi, fine-tuning works fine, let us know of the error you see maybe.
Aayush Jain
@jainaayush05
Mar 27 2018 06:11

Can't see any error - just this message after running the code in this example https://deepdetect.com/overview/examples/#finetuning-googlenet:9ecc00cd7c4c5bc9e8febc4fe00186bf

{'status': {'code': 200, 'msg': 'OK'}, 'head': {'method': '/train', 'job': 1, 'status': 'error'}, 'body': {}}

Can you please let me know how should I see detailed errors?
Emmanuel Benazera
@beniz
Mar 27 2018 06:16
Look for the error report on server side
No measure probably means your dataset is still being preprocessed
Aayush Jain
@jainaayush05
Mar 27 2018 06:17
this might be a stupid question - but how do I look at error report on server side. I am using docker
Emmanuel Benazera
@beniz
Mar 27 2018 06:19
Look at docker documentation in the docker directory of deepdetect.
Aayush Jain
@jainaayush05
Mar 27 2018 06:20
okay. Thank you
Aayush Jain
@jainaayush05
Mar 27 2018 06:32
I think so I figured out the error - one question though
I am trying to fine tune the clothing model
for that should I rename the last layer in just deploy.prototxt
or in both "googlenet.prototxt" and"deploy.txt"?
and use googlenet as template?
Emmanuel Benazera
@beniz
Mar 27 2018 06:34
It's done automatically when using the finetuning option.
Aayush Jain
@jainaayush05
Mar 27 2018 06:35
so do I need to keep both googlenet.prototxt and deploy.prototxt in my model?
Aayush Jain
@jainaayush05
Mar 27 2018 07:12

@beniz : I am stuck at this for a past few days and would really really appreciate your little guidance on this.

Basically, I am trying to fine tune your clothing.tar.bz model to train with my custom classes. These are the steps that I am taking

  1. put all the images as per the "resnet" structure. I have used your downloader to do the same.
  2. use clothing.tar.bz model as it is
  3. followed the code example here as it is: https://deepdetect.com/overview/examples/#finetuning-googlenet:9ecc00cd7c4c5bc9e8febc4fe00186bf

I have a couple of doubts here

  1. do I need to make any modifications to any of the model files? or will finetuning option will take care of any layers modifications required
  2. should I include the parameter "template" : "googlenet", as shown in the example or remove it?
Emmanuel Benazera
@beniz
Mar 27 2018 07:56
you don't need to modify the model files, the finetuning parameter tells the server to do it. you don't need the template parameter since the model and prototxt files are already there.
Aayush Jain
@jainaayush05
Mar 27 2018 08:16
Perfect! Thank you very much!
Emmanuel Benazera
@beniz
Mar 27 2018 08:27
Np
Aayush Jain
@jainaayush05
Mar 27 2018 13:04
@beniz - the clothing model have an accuracy of 0.34 and footwear has that of 0.344
Any idea why so low and how can that be improved?
So if I am fine-tuning, should I finetune my model over "Clothing" model or base "GoogleNet" to get better performance?
Emmanuel Benazera
@beniz
Mar 27 2018 16:43
up to you to try playing around with your own datasets and architectures, there are many clothing datasets around
Aayush Jain
@jainaayush05
Mar 27 2018 16:45
got it! And thank you for such an amazing platform - it makes life so simple for newbies like me. Let me know if I can contribute or evangelise in any possible way
Emmanuel Benazera
@beniz
Mar 27 2018 17:49
Thanks for the kind words :) good luck with your projects!
Aayush Jain
@jainaayush05
Mar 27 2018 19:50
if there are multiple .caffemodels in a models directory - which one is used by default?
Emmanuel Benazera
@beniz
Mar 27 2018 20:56
Latest by date