Why does the different result come out between Caffe and CNTK with Resnet??

[YiKo]

I'm trying to use the caffe.

but i have got a problem.

when i tested with CNTK, i got a result of 99% Accuracy.

but when i tested with Caffe, i get a 27% Accuracy.

both of that are tested with same images.

1. OS : windows 7

2. file format is gray scale bmp file.

3. image size : 224 x 224

4. Train/Test data (2 classes)

 1) Train : 2,000 samples

 2) Test : 4,000 samples

[CNTK-Information]

### CNTK-SGD ###

    SGD=[ 

        epochSize=0 

        minibatchSize=16 

       learningRatesPerMB=0.1*15:0.05*15:0.01*15:0.001 

       momentumPerMB=0.9 

       maxEpochs=60 

       gradUpdateType="None" 

       L2RegWeight=0.0001 

       dropoutRate=0 

        

       ParallelTrain=[ 

           parallelizationMethod="DataParallelSGD" 

           distributedMBReading="true" 

           parallelizationStartEpoch=1 

           DataParallelSGD=[ 

               gradientBits=32 

           ] 

       ] 

   ] 

i have 3,000 data for training

and set the max Epochs 60,

3,000 * 60 => 180,000 data is used for training

so I set the caffe parameter like below,

max_iter : 3000 and batch size 64

3,000 * 60 => 180,000 data is used for training

and base_lr: 0.1 (such as CNTK's starting learningRatesPerMB=0.1*15:0.05*15:0.01*15:0.001)

but... 

[Caffe-Information]

#### Caffe-solver ####

display: 30

base_lr: 0.1

lr_policy: "multistep"

stepvalue: 3000

stepvalue: 6000

stepvalue: 8000

stepvalue: 10000

gamma: 0.5

max_iter: 11250

momentum: 0.9

weight_decay: 0.0001

snapshot: 1000

solver_mode: GPU

#### Caffe-prototxt ####

image_data_param {

        source: "Train.txt" <- list of bmp files (gray scale)

        batch_size: 16

is_color: false

shuffle: true

}

what should i do for getting similar result to the CNTK.