These are chat archives for beniz/deepdetect

16th
Feb 2017
sunatthegilddotcom
@sunatthegilddotcom
Feb 16 2017 00:20
@beniz. need help. I am able to do the object detection from https://deepdetect.com/tutorials/object-detector/ successfully with https://photos.wi.gcs.trstatic.net/e9hyHkaRFZdDV_jLZuTS6jcYq1eLUfiFzfl9zNavmNuoyZ-3UCX_EGg6D5TNU--V-f-z2CT8Kg0u3mF9gccUiA. But when I am using the same image with the python client. I got this error
@beniz python objdetect.py --image ./obj-de-orig.jpg --confidence-threshold 0.1
{u'status': {u'msg': u'NotFound', u'dd_msg': u'Service Not Found', u'code': 400, u'dd_code': 1002}}
error 400
@beniz change the service in the objectdetect.py

import os, sys, argparse
from dd_client import DD
import cv2

parser = argparse.ArgumentParser()
parser.add_argument("--image",help="path to image")
parser.add_argument("--confidence-threshold",help="keep detections with confidence above threshold",type=float,default=0.1)
args = parser.parse_args()

host = 'localhost'
sname = 'voc0712'

sunatthegilddotcom
@sunatthegilddotcom
Feb 16 2017 06:54
@beniz This might be useful for you to check the price difference. But Personally, I think Azure's price http://www.rightscale.com/blog/cloud-cost-analysis/aws-vs-azure-vs-google-cloud-pricing-compute-instances
is cheaper due to EA agreement. Plus it is billed by minutes vs. AWS hourly. GCP is by minutes as well.
Emmanuel Benazera
@beniz
Feb 16 2017 07:01
But GPU availability differs I think
regarding the Python script above, please PR a fix against any glitch you be found and solved if you can
I'll look at Azure marketplace again but last time I could really understand the process ;)
sunatthegilddotcom
@sunatthegilddotcom
Feb 16 2017 07:02
@beniz . Thanks. I will try to fix it. and let you know
@beniz Once you get used to AWS. It is a little hard to navigate in Azure. Not sure you have looked at GCP. Similar there. Only Cert Azure region has GPU. in US.
Emmanuel Benazera
@beniz
Feb 16 2017 07:06
all out AWS customers use the GPU instance. This is why at the moment I havent added another marketplace.
Are you saying you haven't been able to run the object detector demo ?
sunatthegilddotcom
@sunatthegilddotcom
Feb 16 2017 07:08
@beniz, Understand. I am planning to use GPU for training only, but use CPU for classification, it might be more cost effective. Do you agree?
I am able to run the object detect demo
@beniz , I have not been able to do the web demo due to the image for Openimage TF yet due to the image size. I am trying to figure out a way to run images with (224x224 and 299x299) at the same time.
Emmanuel Benazera
@beniz
Feb 16 2017 07:12
I'm not sure CPU for images is cost effective. We still have to release a benchmark Python script so that users can measure throughput easily.
sunatthegilddotcom
@sunatthegilddotcom
Feb 16 2017 07:13
@beniz I am able to run the object demo from the CURL, but not the python script
Emmanuel Benazera
@beniz
Feb 16 2017 07:13
OK please open an issue
sunatthegilddotcom
@sunatthegilddotcom
Feb 16 2017 07:13
@beniz , I will do more research before I open an issue. Thanks again
@beniz , I think CPU and GPU benchmark is a great idea
Emmanuel Benazera
@beniz
Feb 16 2017 07:14
There s a PR for an updated web demo based on the new JS client. In the meantime, I believe the size is harcoded somewhere in the JS
sunatthegilddotcom
@sunatthegilddotcom
Feb 16 2017 07:15
@beniz , that's what I thought too. but I just grep it, I did not see it. I am going to look at the code itself.
Emmanuel Benazera
@beniz
Feb 16 2017 07:37
@sunatthegilddotcom there's a default built-in, so if it's not there, you'd need to add it
we'll review this when merging the PR
sunatthegilddotcom
@sunatthegilddotcom
Feb 16 2017 07:37
@beniz, thank you very much!
Gilbow
@Gilbow
Feb 16 2017 10:33
Hi all. i'm trying to understand how build a training set.
all images in a category is in a separate folder
there is a way to 'focus' training on RoI specified in a file.
?
the file format could be "image filename, x1, y1, x2, y2".
Emmanuel Benazera
@beniz
Feb 16 2017 10:43
hi, there's no ability for training object detectors from DD at the moment. You'd need to train with SSD then serve with DD. We have our own pipeline for training detectors, along with automatic data augmentation for building robust models, but it's customer-only at the moment. The file format you are mentioning is actually the one we use.