Different results in train/validation and predicition

[cisi17]

Hello,

I have implemented my own neural network with my own dataset.

Firstly. I train and test my network using transfer learning. I have achieved 95% train/test accuaracy:

  /usr/local/micaffe1/bin/caffe train --solver=/home/ubuntu/example/modelcsl/solver_csl.prototxt --weights /home/ubuntu/example/bvlc_reference_caffenet.caffemodel 2>&1 | tee /ubuntu/example/modelcsl/resultados_TL.log

But when I am going to do a prediction, the precision is over 70%:

  net = caffe.Net('/home/marcelo/ejemplosCaffe1/legImgLaser/modelcsl/caffenet_deploy_csl.prototxt',
            '/home/marcelo/ejemplosCaffe1/legImgLaser/modelcsl_iter_500.caffemodel',
            caffe.TEST)

......

   net.blobs['data'].data[...] = transformer.preprocess('data', img)
   out = net.forward()
   pred_probas = out['prob']

.......

I think that the deploy.prototxt and the train_test.prototxt have got the same net. And the preprocess is the same.

Could you help me?

Thank you.