Age model fails with error

[jolibrain]

Age model fails with error

[beniz]

Let me understand, you are trying to use the Caffe model from gillevi with dd, is that correct ?

I should be able to help, but more information is needed: all prototxt files typically and service + train API calls so that I can reproduce.

[Qwasser]

Yes, I am trying to use pretrained caffe model. It can be downloaded from here http://www.openu.ac.il/home/hassner/projects/cnn_agegender/cnn_age_gender_models_and_data.0.0.2.zip
To use the model I created "models/age" folder and put there deploy.prototxt (renamed deploy_age) and age.caffemodel files from the archive. Then I had created service using following code:

model = {'repository': '/home/ubuntu/models/age''}
parameters_input = {'connector':'image',
                            'width': 256,
                            'height': 256}
parameters_mllib = {'nclasses': 8}
parameters_output = {}

dd.put_service("age",
                       model,
                       "age classification",
                       'caffe',
                       parameters_input,
                       parameters_mllib,
                       parameters_output)

and the query is

curl -X POST "http://localhost:8080/predict" -d "{\"service\":\"age\",\"parameters\":{\"input\":{\"width\":256,\"height\":256},\"output\":{\"best\":3}},\"data\":[\"http://i.ytimg.com/vi/0vxOhd4qlnA/maxresdefault.jpg\"]}"

I think the problem is in deploy.prototxt and it should be converted somehow

[beniz]

Yes you need a memorydata layer at the top, this is so dd can pass data to the net. Look at the deploy file from templates/caffe/googlenet as an example. I will try to reproduce as well on my spare time, but please let me know how it goes.

[Qwasser]

Added following text to deploy.prototxt:

layers  {
  name: "caffenet"
  type: MEMORY_DATA
  top: "data"
  top: "label"
  memory_data_param {
    batch_size: 32
    channels: 3
    height: 256
    width: 256
  }
}

And everything worked :) Should it be in the documentation?

[beniz]

Ah got it to work in parallel to you. The key is to convert the Caffe model from the old format to the new one, and then update the deploy.prototxt file as well.

I've put a model tarball for the 8 classes age model here: http://www.deepdetect.com/models/age_model.tar.bz2
It is ready to use, though I can't report on the accuracy.

It could be added to the 'model' page http://www.deepdetect.com/applications/model/

[Qwasser]

Have you run into a trouble mean.binaryproto has incorrect dimensions?

[beniz]

I believe dimension is 227x227.

[beniz]

Btw, the predict call does not yet read the mean.binaryproto file. I will commit this shortly. From my short tests, it does affect the results significantly on this dataset.

[beniz]

Do you know what the age classes correspond to by any chance ? (0 to what range, etc..). Thanks.

OK. got it, put a file named corresp.txt in your model repository with the following in it:

0 0_2
1 4_6
2 8_13
3 15_20
4 25_32
5 38_43
6 48_53
7 60+

This is deducted from http://www.openu.ac.il/home/hassner/projects/cnn_agegender/CNN_AgeGenderEstimation.pdf and I may be wrong.

[Qwasser]

Yes, I just wanted to send them to you. I copied classes from notebook that comes with modela and they are the same as in your list.

[beniz]

OK good! I was just surprised that the example they give in their tarball, example_image.jpg seems to be reported as 60+ (0.53) and 0_2 (0.35). My own picture however is unfortunately not reported as younger...

[Qwasser]

Have you tried it in notebook from archive? Works fine there (even without mean substraction).

[beniz]

Somehow it works better for me without the mean substraction. I've investigated and failed to understand why. It's possible the mean.binaryproto is not in the exact same format as the ones produced by the Caffe compute_mean. The Caffe classification.cpp example uses a planar 32-bit float BGR or grayscale format.
Closing this since it appears to be working fine with dd anyways, can be reopened if the mean file issue is a problem in practice.

[beniz]

Age model added to http://www.deepdetect.com/applications/model/