CNN唯一开源FPGA实现

## Where to get You can download the fixed-point CNN weights, input test vectors and output golden reference data for AlexNet/VGG-16 from the following links to run and evaluate the project. * Baidu's Cloud Drive: ``` http://pan.baidu.com/s/1jIl6qkm ``` * GoogleDrive: ``` https://drive.google.com/open?id=0B3srpZY5rHcASEhSSmh3Tm5LclU ``` ## Fixed-Point CNN models The weights are quantilized with 8-bit precisions, and the model quantization is performed such that there is only 1% loss in top-1/5 accuracy for all models. In quantization, we assume the fixed-point weight is presented as *N * 2^-m*, where *N* is a fixed-point integer with *n-bit* word length, and *m* denotes the fractional bits of the quantized weight. Therefore, we use the pair of integers *(n,m)* as the quantization parameters. In the following tables, we report the quantization parameters used in this project for different CNN models. * AlexNet(CaffeNet) <table> <tr> <td>LayerName</td> <td>Input</td> <td>Output</td> <td>Weight</td> </tr> <tr> <td>conv1</td> <td>8,0</td> <td>8,-4</td> <td>8,8</td> </tr> <tr> <td>relu1</td> <td>8.-4</td> <td>8,-4</td> <td></td> </tr> <tr> <td>lrn1</td> <td>8,-4</td> <td>8,0</td> <td></td> </tr> <tr> <td>pool1</td> <td>8,0</td> <td>8,0</td> <td></td> </tr> <tr> <td>conv2</td> <td>8,0</td> <td>8,-2</td> <td>8,8</td> </tr> <tr> <td>relu2</td> <td>8,-2</td> <td>8,-2</td> <td></td> </tr> <tr> <td>lrn2</td> <td>8,-2</td> <td>8,0</td> <td></td> </tr> <tr> <td>pool2</td> <td>8,0</td> <td>8,0</td> <td></td> </tr> <tr> <td>conv3</td> <td>8,0</td> <td>8,-1</td> <td>8,8</td> </tr> <tr> <td>relu3</td> <td>8,-1</td> <td>8,-1</td> <td></td> </tr> <tr> <td>conv4</td> <td>8,-1</td> <td>8,-1</td> <td>8,8</td> </tr> <tr> <td>relu4</td> <td>8,-1</td> <td>8,-1</td> <td></td> </tr> <tr> <td>conv5</td> <td>8,-1</td> <td>8,-1</td> <td>8,8</td> </tr> <tr> <td>relu5</td> <td>8,-1</td> <td>8,-1</td> <td></td> </tr> <tr> <td>pool5</td> <td>8,-1</td> <td>8,-1</td> <td></td> </tr> <tr> <td>fc6</td> <td>8,-1</td> <td>8,0</td> <td>8,11</td> </tr> <tr> <td>relu6</td> <td>8,0</td> <td>8,0</td> <td></td> </tr> <tr> <td>drop6</td> <td>8,0</td> <td>8,0</td> <td></td> </tr> <tr> <td>fc7</td> <td>8,0</td> <td>8,2</td> <td>8,10</td> </tr> <tr> <td>relu7</td> <td>8,2</td> <td>8,2</td> <td></td> </tr> <tr> <td>drop7</td> <td>8,2</td> <td>8,2</td> <td></td> </tr> <tr> <td>fc8</td> <td>8,2</td> <td>8,2</td> <td>8,10</td> </tr> </table> * VGG-16 <table> <tr> <td>LayerName</td> <td>Input</td> <td>Output</td> <td>Weight</td> </tr> <tr> <td>conv1_1</td> <td>8,0</td> <td>8,-2</td> <td>8,7</td> </tr> <tr> <td>relu1_1</td> <td></td> <td></td> <td></td> </tr> <tr> <td>conv1_2</td> <td>8,-2</td> <td>8,-5</td> <td>8,8</td> </tr> <tr> <td>relu1_2</td> <td></td> <td></td> <td></td> </tr> <tr> <td>conv2_1</td> <td>8,-5</td> <td>8,-5</td> <td>8,8</td> </tr> <tr> <td>relu2_1</td> <td></td> <td></td> <td></td> </tr> <tr> <td>conv2_2</td> <td>8,-5</td> <td>8,-6</td> <td>8,8</td> </tr> <tr> <td>relu2_2</td> <td></td> <td></td> <td></td> </tr> <tr> <td>conv3_1</td> <td>8,-6</td> <td>8,-7</td> <td>8,7</td> </tr> <tr> <td>relu3_1</td> <td></td> <td></td> <td></td> </tr> <tr> <td>conv3_2</td> <td>8,-7</td> <td>8,-7</td> <td>8,8</td> </tr> <tr> <td>relu3_2</td> <td></td> <td></td> <td></td> </tr> <tr> <td>conv3_3</td> <td>8,-7</td> <td>8,-7</td> <td>8,8</td> </tr> <tr> <td>relu3_3</td> <td></td> <td></td> <td></td> </tr> <tr> <td>conv4_1</td> <td>8,-7</td> <td>8,-6</td> <td>8,8</td> </tr> <tr> <td>relu4_1</td> <td></td> <td></td> <td></td> </tr> <tr> <td>conv4_2</td> <td>8,-6</td> <td>8,-5</td> <td>8,8</td> </tr> <tr> <td>relu4_2</td> <td></td> <td></td> <td></td> </tr> <tr> <td>conv4_3</td> <td>8,-5</td> <td>8,-5</td> <td>8,8</td> </tr> <tr> <td>relu4_3</td> <td></td> <td></td> <td></td> </tr> <tr> <td>conv5_1</td> <td>8,-5</td> <td>8,-4</td> <td>8,9</td> </tr> <tr> <td>relu5_1</td> <td></td> <td></td> <td></td> </tr> <tr> <td>conv5_2</td> <td>8,-4</td> <td>8,-3</td> <td>8,9</td> </tr> <tr> <td>relu5_2</td> <td></td> <td></td> <td></td> </tr> <tr> <td>conv5_3</td> <td>8,-3</td> <td>8,-2</td> <td>8,8</td> </tr> <tr> <td>relu5_3</td> <td></td> <td></td> <td></td> </tr> <tr> <td>fc6</td> <td>8,-2</td> <td>8,0</td> <td>4,8</td> </tr> <tr> <td>fc7</td> <td>8,0</td> <td>8,2</td> <td>4,7</td> </tr> <tr> <td>fc8</td> <td>8,2</td> <td>8,2</td> <td>4,7</td> </tr> </table> ## How to prepare the CNN models for PipeCNN * Install [Caffe](http://caffe.berkeleyvision.org/) and use the following matlab script to extract the CNN model (assuming we are dealing with caffenet) ``` caffe.set_mode_cpu(); model = './models/bvlc_reference_caffenet/deploy.prototxt'; weights = './models/bvlc_reference_caffenet.caffemodel'; net = caffe.Net(model, weights, 'test'); netparams = {{net.params('conv1',1).get_data(),net.params('conv1',2).get_data()}, ... {net.params('conv2',1).get_data(),net.params('conv2',2).get_data()}, ... {net.params('conv3',1).get_data(),net.params('conv3',2).get_data()}, ... {net.params('conv4',1).get_data(),net.params('conv4',2).get_data()}, ... {net.params('conv5',1).get_data(),net.params('conv5',2).get_data()}, ... {net.params('fc6',1).get_data(),net.params('fc6',2).get_data()}, ... {net.params('fc7',1).get_data(),net.params('fc7',2).get_data()}, ... {net.params('fc8',1).get_data(),net.params('fc8',2).get_data()}}; ``` * Quantize the extracted mode (stored in netparams) to fixed-point weight and bias. Matlab's Fixed-point toolbox should be installed. The word length and fractional bit length are set according to the tables in previous sections. ``` WeightWidth = [ 8; 8; 8; 8; 8; 8; 8; 8]; WeightFrac = [ 8; 8; 8; 8; 8; 11; 10; 10]; MathType = fimath('RoundingMethod', 'Nearest', 'OverflowAction', 'Saturate', 'ProductMode', 'FullPrecision', 'SumMode', 'FullPrecision'); for i=1:8 WeightType{i} = numerictype('Signed',1, 'WordLength', WeightWidth(i), 'FractionLength', WeightFrac(i)); weight{i} = fi(netParams.netparams{i}{1}, WeightType{i}, MathType); bias{i} = fi(netParams.netparams{i}{2}, WeightType{i}, MathType); end ``` * Combine and store the weights and bias as a single binary file. ``` fid = fopen('weights.dat', 'w'); for i=1:8 fwrite(fid, storedInteger(weight{i}), 'int8'); fwrite(fid, storedInteger(bias{i}), 'int8'); end fclose(fid); ``` ## Notes: Remember to change the corresponding path