These are chat archives for beniz/deepdetect

28th
Jun 2016
sazsazsaz
@sazsazsaz
Jun 28 2016 05:43
First time here, congrats on the impressive work. I have a question about image search. Is there any way to make image search solely with curl commands instead of python? I tried the code @beniz paste here but it gives me internal error.
Emmanuel Benazera
@beniz
Jun 28 2016 06:41
No, as you need to access the index. You'll have to hack your own calls for this.
sazsazsaz
@sazsazsaz
Jun 28 2016 11:19
Is there any example curl command for indexing?
Emmanuel Benazera
@beniz
Jun 28 2016 11:35
@sazsazsaz read the API and doc carefully, and look at the similarity search demo code. There's no index in DD, this is handled by a package called annoy in the similarity search demo code.
sazsazsaz
@sazsazsaz
Jun 28 2016 11:36
Thank you, I will read the doc more carrefully. I am using Docker, is it possible to this with docker?
Emmanuel Benazera
@beniz
Jun 28 2016 11:41
the search similarity demo is independent from where and how you use dd
it is making call to the dd server, and it can run into docker
seeb0h
@seeb0h
Jun 28 2016 16:08
Hi! I am trying deepdetect. Love the idea.
But.... I can't setup a service using a already trained LeNet over MNIST model
Here is my try (in your provided docker) : https://gist.github.com/seeb0h/784a1d15da25f968ce8cb4aa305d3cec
I guess it is stupid. But I miss the point . Any help welcome..
Emmanuel Benazera
@beniz
Jun 28 2016 16:18
@seeb0h check the documentation and API carefully, there's no LeNet template. If you want to play with MNIST you need to import the .prototxt files along with the pre-trained weights, and make sure the prototxtfiles use MemoryDataLayer. All in all, I'd strongly recommend to just follow one of the step by step tutorials instead. Unless you have a specific need for MNIST.
seeb0h
@seeb0h
Jun 28 2016 18:24
OK I misunderstood. I did not find the template list, so I guessed this parameter was the name in the prototxt (got from here )
Actually I need to do a quick test of digit classification. Using a service like DeepDetect could help me a lot for software integration. So I tried to use an already trained LeNet model to skip the training step and go straigth to the predict.
seeb0h
@seeb0h
Jun 28 2016 18:56
OK (bis). I find the template list in the API. I am unfamiliar with Caffe. But actually, the net I want is a convnet template with the right layers configured through your API. Am I right ?
If so, and if I set a parameters_mllib with the right layers, is there possible to use the pretrained model I put a ../models/lenet-mnist ?