Getting different accuracies using different caffe classes(98.65 vs 98.1 vs 98.20)

[Frank Darius]

When I train and test my model using Caffe's command line interface, I get e.g. `98.65%` whereas when I myself write code(given below) to calculate accuracy I get e.g `98.1%` using `Caffe.Net`.  

Everything is straight forward and I have no idea what is causing the issue.   

I also tried using `Caffe.Classifier` and its `predict` method, and yet get another lesser accuracy(i.e. `98.20%`!)      

Here is the snippet of code I wrote:       

import sys
import caffe
import numpy as np
import lmdb
import argparse
from collections import defaultdict
from sklearn.metrics import classification_report
from sklearn.metrics import confusion_matrix
import matplotlib.pyplot as plt
import itertools
from sklearn.metrics import roc_curve, auc
import random

   
if __name__ == "__main__":
    parser = argparse.ArgumentParser()
    parser.add_argument('--proto', help='path to the network prototxt file(deploy)', type=str, required=True)
    parser.add_argument('--model', help='path to your caffemodel file', type=str, required=True)
    parser.add_argument('--mean', help='path to the mean file(.binaryproto)', type=str, required=True)
    #group = parser.add_mutually_exclusive_group(required=True)
    parser.add_argument('--db_type', help='lmdb or leveldb', type=str, required=True)
    parser.add_argument('--db_path', help='path to your lmdb/leveldb dataset', type=str, required=True)
    args = parser.parse_args()
   
    predicted_lables=[]
    true_labels = []
    misclassified =[]
    class_names = ['unsafe','safe']
    count=0
    correct = 0
    batch=[]
    plabe_ls=[]
    batch_size = 50
    cropx = 224
    cropy = 224
    i = 0
    multi_crop = False
    use_caffe_classifier = True
   
    caffe.set_mode_gpu()
    # Extract mean from the mean image file
    mean_blobproto_new = caffe.proto.caffe_pb2.BlobProto()
    f = open(args.mean, 'rb')
    mean_blobproto_new.ParseFromString(f.read())
    mean_image = caffe.io.blobproto_to_array(mean_blobproto_new)
    f.close()


    net = caffe.Classifier(args.proto, args.model,
                           mean = mean_image[0].mean(1).mean(1),
                           image_dims = (224, 224))
   
    net1 = caffe.Net(args.proto, args.model, caffe.TEST)
    net1.blobs['data'].reshape(batch_size, 3,224, 224)
    data_blob_shape = net1.blobs['data'].data.shape


    #check and see if its lmdb or leveldb
    if(args.db_type.lower() == 'lmdb'):
        lmdb_env = lmdb.open(args.db_path)
        lmdb_txn = lmdb_env.begin()
        lmdb_cursor = lmdb_txn.cursor()
        for key, value in lmdb_cursor:
            count += 1
            datum = caffe.proto.caffe_pb2.Datum()
            datum.ParseFromString(value)
            label = int(datum.label)
            image = caffe.io.datum_to_array(datum).astype(np.float32)
            #key,image,label
            #buffer n image
            if(count % 5000 == 0):          
                print('{0} samples processed so far'.format(count))
           
            if(i < batch_size):
                i+=1
                inf= key,image,label
                batch.append(inf)
                #print(key)                
            if(i >= batch_size):
                #process n image
                ims=[]              
                for x in range(len(batch)):
                    img = batch[x][1]
                    #img has c,w,h shape! its already gone through transpose and channel swap when it was being saved into lmdb!
                    #Method III : use center crop just like caffe does in test time
                    if (use_caffe_classifier != True):
                        #center crop
                        c,w,h = img.shape
                        startx = h//2 - cropx//2
                        starty = w//2 - cropy//2
                        img = img[:, startx:startx + cropx, starty:starty + cropy]                  
                        #transpose the image so we can subtract from mean
                        img = img.transpose(2,1,0)
                        img -= mean_image[0].mean(1).mean(1)
                        #transpose back to the original state
                        img = img.transpose(2,1,0)
                        ims.append(img)
                    else:
                        ims.append(img.transpose(2,1,0))    
                       
                if (use_caffe_classifier != True):
                    net1.blobs['data'].data[...] = ims[:]
                    out_1 = net1.forward()
                    plabe_ls = out_1['pred']                                                                                
                else:
                    out_1 = net.predict(np.asarray(ims), oversample=multi_crop)
                    plabe_ls = out_1    
                   
                plbl = np.asarray(plabe_ls)
                plbl = plbl.argmax(axis=1)
                for j in range(len(batch)):
                    if (plbl[j] == batch[j][2]):
                        correct+=1
                    else:
                        misclassified.append(batch[j][0])
                       
                    predicted_lables.append(plbl[j])        
                    true_labels.append(batch[j][2])
                batch.clear()
                i = 0              
                                 
           
    sys.stdout.write("\rAccuracy: %.2f%%" % (100.*correct/count))
    sys.stdout.flush()
    print(", %i/%i corrects" % (correct, count))  

What is causing this difference in accuracies?   

More information :

I am using Python3.5 on Windows.  

I read images from a `lmdb` dataset.  

The images have `256x256` and center cropped with the size `224x224`.  

It is fine tuned on `GoogleNet`.         

For the `Caffe.predict` to work well I had to change `classify.py`   

In training, I just use `Caffe` defaults, such as random crops at training and center crop at test-time.  

Changes:   

changed line 35 to:  

     self.transformer.set_transpose(in_, (2, 1, 0))

and line 99 to : 

    predictions = predictions.reshape((len(predictions) // 10, 10, -1))