Caffe net.predict() outputs random results (GoogleNet)

Question

I used pretrained GoogleNet from https://github.com/BVLC/caffe/tree/master/models/bvlc_googlenet and finetuned it with my own data (~ 100k images, 101 classes). After one day training I achieved 62% in top-1 and 85% in top-5 classification and try to use this network to predict several images.

I just followed example from https://github.com/BVLC/caffe/blob/master/examples/classification.ipynb,

Here is my Python code:

import caffe
import numpy as np


caffe_root = './caffe'


MODEL_FILE = 'caffe/models/bvlc_googlenet/deploy.prototxt'
PRETRAINED = 'caffe/models/bvlc_googlenet/bvlc_googlenet_iter_200000.caffemodel'

caffe.set_mode_gpu()

net = caffe.Classifier(MODEL_FILE, PRETRAINED,
               mean=np.load('ilsvrc_2012_mean.npy').mean(1).mean(1),
               channel_swap=(2,1,0),
               raw_scale=255,
               image_dims=(224, 224))

def caffe_predict(path):
        input_image = caffe.io.load_image(path)
        print path
        print input_image
        prediction = net.predict([input_image])


        print prediction
        print "----------"

        print 'prediction shape:', prediction[0].shape
        print 'predicted class:', prediction[0].argmax()


        proba = prediction[0][prediction[0].argmax()]
        ind = prediction[0].argsort()[-5:][::-1] # top-5 predictions


        return prediction[0].argmax(), proba, ind

In my deploy.prototxt I changed the last layer only to predict my 101 classes.

layer {
  name: "loss3/classifier"
  type: "InnerProduct"
  bottom: "pool5/7x7_s1"
  top: "loss3/classifier"
  param {
    lr_mult: 1
    decay_mult: 1
  }
  param {
    lr_mult: 2
    decay_mult: 0
  }
  inner_product_param {
    num_output: 101
    weight_filler {
      type: "xavier"
    }
    bias_filler {
      type: "constant"
      value: 0
    }
  }
}
layer {
  name: "prob"
  type: "Softmax"
  bottom: "loss3/classifier"
  top: "prob"
}

Here is the distribution of softmax output:

[[ 0.01106235  0.00343131  0.00807581  0.01530041  0.01077161  0.0081002
   0.00989228  0.00972753  0.00429183  0.01377776  0.02028225  0.01209726
   0.01318955  0.00669979  0.00720005  0.00838189  0.00335461  0.01461464
   0.01485041  0.00543212  0.00400191  0.0084842   0.02134697  0.02500303
   0.00561895  0.00776423  0.02176422  0.00752334  0.0116104   0.01328687
   0.00517187  0.02234021  0.00727272  0.02380056  0.01210031  0.00582192
   0.00729601  0.00832637  0.00819836  0.00520551  0.00625274  0.00426603
   0.01210176  0.00571806  0.00646495  0.01589645  0.00642173  0.00805364
   0.00364388  0.01553882  0.01549598  0.01824486  0.00483241  0.01231962
   0.00545738  0.0101487   0.0040346   0.01066607  0.01328133  0.01027429
   0.01581303  0.01199994  0.00371804  0.01241552  0.00831448  0.00789811
   0.00456275  0.00504562  0.00424598  0.01309276  0.0079432   0.0140427
   0.00487625  0.02614347  0.00603372  0.00892296  0.00924052  0.00712763
   0.01101298  0.00716757  0.01019373  0.01234141  0.00905332  0.0040798
   0.00846442  0.00924353  0.00709366  0.01535406  0.00653238  0.01083806
   0.01168014  0.02076091  0.00542234  0.01246306  0.00704035  0.00529556
   0.00751443  0.00797437  0.00408798  0.00891858  0.00444583]]

It seems just like random distribution with no sense.

Thank you for any help or hint and best regards, Alex

Answer 1

The solution is really simple: I just forgot to rename the last layer in deploy file:

layer {
  name: "loss3/classifier"
  type: "InnerProduct"
  bottom: "pool5/7x7_s1"
  top: "loss3/classifier"
  param {
    lr_mult: 1
    decay_mult: 1
  }

Answer 2

Please check the image transformation you are using - is it the same in training and test-time?

AFAIK bvlc_googlenet subtract image mean with one value per channel, while your python classifier uses different mean mean=np.load('ilsvrc_2012_mean.npy').mean(1).mean(1). This might cause the net to be unable to classify your inputs correctly.