mnist network result is different from Movidius's

[idata]

Hi,

Now I am working the mnist network realized with Movidius stick, but I have gotten the different result from pycaffe's.

python code is the following for setting;

network_blob='../networks/mnist/graph' #compile succeeded

dim=(28,28)

with open(network_blob, mode='rb') as f:

blob = f.read()graph = device.AllocateGraph(blob)graph.SetGraphOption(mvnc.GraphOption.ITERATIONS, 1)iterations = graph.GetGraphOption(mvnc.GraphOption.ITERATIONS)img = cv2.imread('../images/two-0.jpg')img = cv2.resize(img,dim)inp = (255 - img.astype(numpy.float16)) / 255graph.LoadTensor(inp, 'user object')output, userobj = graph.GetResult()print(output)

Here is the prototxt for the network;

name: "LeNet"

input: "data"input_shape { dim: 1 dim: 1 dim: 28 dim: 28}layer { name: "conv1" type: "Convolution" bottom: "data" top: "conv1" param { lr_mult: 1 } param { lr_mult: 2 } convolution_param { num_output: 20 kernel_size: 5 stride: 1 weight_filler { type: "xavier" } bias_filler { type: "constant" } }}layer { name: "pool1" type: "Pooling" bottom: "conv1" top: "pool1" pooling_param { pool: MAX kernel_size: 2 stride: 2 }}layer { name: "conv2" type: "Convolution" bottom: "pool1" top: "conv2" param { lr_mult: 1 } param { lr_mult: 2 } convolution_param { num_output: 50 kernel_size: 5 stride: 1 weight_filler { type: "xavier" } bias_filler { type: "constant" } }}layer { name: "pool2" type: "Pooling" bottom: "conv2" top: "pool2" pooling_param { pool: MAX kernel_size: 2 stride: 2 }}layer { name: "ip1" type: "InnerProduct" bottom: "pool2" top: "ip1" param { lr_mult: 1 } param { lr_mult: 2 } inner_product_param { num_output: 500 weight_filler { type: "xavier" } bias_filler { type: "constant" } }}layer { name: "relu1" type: "ReLU" bottom: "ip1" top: "ip1"}layer { name: "ip2" type: "InnerProduct" bottom: "ip1" top: "ip2" param { lr_mult: 1 } param { lr_mult: 2 } inner_product_param { num_output: 10 weight_filler { type: "xavier" } bias_filler { type: "constant" } }}layer { name: "prob" type: "Softmax" bottom: "ip2" top: "prob"}

"inp" values from movidius code are almost the same as pycaffe's which is read by "a" in following code;

#!/usr/bin/env python

import numpy as np

import pandas as pdimport osimport argparseimport time

import caffe

net = caffe.Classifier('lenet.prototxt','lenet_iter_10000.caffemodel', image_dims = (28, 28))

caffe._caffe.set_mode_cpu()net.transformer.set_raw_scale('data', 1)

a=1 - caffe.io.load_image('three-0.jpg', color = False, ) #maximum 1

print(a)

scores = net.predict([a], oversample=False)

print (scores)

If somebody can give the good advises to me, I can send the trained parameter.

Could somebody teach the reason why I got the almost different result from pycaffe?

Thanks

[idata]

@KKK Regarding your results from LeNet, I was just wondering what your results were and what your expected results are. If you can provide the caffemodel file, that would be great also. Thanks.

[idata]

@KKK Additionally, the Movidius SDK toolkit has an example called top5_over_a_dataset.py which uses mnist and LeNet. It may be helpful to check it out.