Skip to content

Commit

Permalink
add computational complexity to analyzer
Browse files Browse the repository at this point in the history
  • Loading branch information
@dgschwend
dgschwend committed Mar 21, 2016
1 parent da254cb commit fdd4bc9
Show file tree
Hide file tree
Showing 2 changed files with 187 additions and 84 deletions.
130 changes: 85 additions & 45 deletions assets/js/netscope.js
View file
Original file line number Diff line number Diff line change
Expand Up @@ -5,22 +5,31 @@ module.exports = Analyzer = (function() {
function Analyzer() {}

Analyzer.prototype.analyze = function(net) {
var d, failed, i, isglobal, j, k, kernel, kernel_h, kernel_w, layertype, len, len1, len2, mode, n, numout, p, pad, pad_h, pad_w, parent, ref, ref1, ref10, ref11, ref12, ref13, ref14, ref15, ref16, ref17, ref18, ref2, ref3, ref4, ref5, ref6, ref7, ref8, ref9, size, stride, stride_h, stride_w;
var d, failed, i, isglobal, j, k, kernel, kernel_h, kernel_w, layertype, len, len1, len2, mode, n, num_inputs, num_ops, numout, p, pad, pad_h, pad_w, params, parent, pooltype, ref, ref1, ref10, ref11, ref12, ref13, ref14, ref15, ref16, ref17, ref18, ref19, ref2, ref20, ref3, ref4, ref5, ref6, ref7, ref8, ref9, shape, size, stride, stride_h, stride_w, trivial_layers;
ref = net.nodes;
for (i = 0, len = ref.length; i < len; i++) {
n = ref[i];
d = n.analysis;
d.wIn = d.hIn = d.wOut = d.hOut = 0;
d.chIn = d.chOut = 0;
layertype = n.type.toUpperCase();
d.comp = {
macc: 0,
comp: 0,
add: 0,
div: 0,
exp: 0
};
d.mem = 0;
layertype = n.type.toLowerCase();
parent = (ref1 = n.parents[0]) != null ? ref1.analysis : void 0;
switch (layertype) {
case "DATA":
if ((n.attribs.input_param != null)) {
d.chIn = n.attribs.input_param.shape.dim[1];
d.wIn = n.attribs.input_param.shape.dim[2];
d.hIn = n.attribs.input_param.shape.dim[3];
} else if ((((ref2 = n.attribs.transform_param) != null ? ref2.crop_size : void 0) != null)) {
case "data":
if (((ref2 = n.attribs.input_param) != null ? ref2.shape : void 0) != null) {
shape = n.attribs.input_param.shape;
d.chIn = shape.dim[1];
d.wIn = shape.dim[2];
d.hIn = shape.dim[3];
} else if (((ref3 = n.attribs.transform_param) != null ? ref3.crop_size : void 0) != null) {
d.wIn = d.hIn = n.attribs.transform_param.crop_size;
d.chIn = 3;
} else {
Expand All @@ -31,101 +40,131 @@ module.exports = Analyzer = (function() {
d.hOut = d.hIn;
d.chOut = d.chIn;
break;
case "CONVOLUTION":
kernel_w = (ref3 = n.attribs.convolution_param.kernel_w) != null ? ref3 : n.attribs.convolution_param.kernel_size;
kernel_h = (ref4 = n.attribs.convolution_param.kernel_h) != null ? ref4 : n.attribs.convolution_param.kernel_size;
stride_w = (ref5 = n.attribs.convolution_param.stride_w) != null ? ref5 : (ref6 = n.attribs.convolution_param.stride) != null ? ref6 : 1;
stride_h = (ref7 = n.attribs.convolution_param.stride_h) != null ? ref7 : (ref8 = n.attribs.convolution_param.stride) != null ? ref8 : 1;
pad_w = (ref9 = n.attribs.convolution_param.pad_w) != null ? ref9 : (ref10 = n.attribs.convolution_param.pad) != null ? ref10 : 0;
pad_h = (ref11 = n.attribs.convolution_param.pad_h) != null ? ref11 : (ref12 = n.attribs.convolution_param.pad) != null ? ref12 : 0;
numout = n.attribs.convolution_param.num_output;
case "convolution":
params = n.attribs.convolution_param;
kernel_w = (ref4 = params.kernel_w) != null ? ref4 : params.kernel_size;
kernel_h = (ref5 = params.kernel_h) != null ? ref5 : params.kernel_size;
stride_w = (ref6 = params.stride_w) != null ? ref6 : (ref7 = params.stride) != null ? ref7 : 1;
stride_h = (ref8 = params.stride_h) != null ? ref8 : (ref9 = params.stride) != null ? ref9 : 1;
pad_w = (ref10 = params.pad_w) != null ? ref10 : (ref11 = params.pad) != null ? ref11 : 0;
pad_h = (ref12 = params.pad_h) != null ? ref12 : (ref13 = params.pad) != null ? ref13 : 0;
numout = params.num_output;
d.wIn = parent.wOut;
d.hIn = parent.hOut;
d.wOut = Math.floor((d.wIn + 2 * pad_w - kernel_w) / stride_w) + 1;
d.hOut = Math.floor((d.hIn + 2 * pad_h - kernel_h) / stride_h) + 1;
d.chIn = parent.chOut;
d.chOut = numout;
d.comp.macc = (kernel_w * kernel_h) * (d.wOut * d.hOut) * d.chIn * d.chOut;
break;
case "INNERPRODUCT":
case "INNER_PRODUCT":
case "innerproduct":
case "inner_product":
numout = n.attribs.inner_product_param.num_output;
d.wIn = parent.wOut;
d.hIn = parent.hOut;
d.chIn = parent.chOut;
d.wOut = 1;
d.hOut = 1;
d.chIn = parent.chOut;
d.chOut = numout;
d.comp.macc = (d.wIn * d.hIn) * d.chIn * d.chOut;
break;
case "POOLING":
kernel = n.attribs.pooling_param.kernel_size;
stride = (ref13 = n.attribs.pooling_param.stride) != null ? ref13 : 1;
pad = (ref14 = n.attribs.pooling_param.pad) != null ? ref14 : 0;
isglobal = (ref15 = n.attribs.pooling_param.global_pooling) != null ? ref15 : 0;
case "pooling":
params = n.attribs.pooling_param;
kernel = params.kernel_size;
stride = (ref14 = params.stride) != null ? ref14 : 1;
pad = (ref15 = params.pad) != null ? ref15 : 0;
isglobal = (ref16 = params.global_pooling) != null ? ref16 : 0;
pooltype = ((ref17 = params.pool) != null ? ref17 : 'MAX').toUpperCase();
d.wIn = parent.wOut;
d.hIn = parent.hOut;
d.chIn = parent.chOut;
d.chOut = d.chIn;
d.wOut = Math.ceil((d.wIn + 2 * pad - kernel) / stride) + 1;
d.hOut = Math.ceil((d.hIn + 2 * pad - kernel) / stride) + 1;
if (isglobal) {
d.wOut = d.hOut = 1;
}
num_ops = isglobal ? (d.wIn * d.hIn) * d.chIn : (d.wOut * d.hOut) * kernel * kernel * d.chOut;
if (pooltype === 'MAX') {
d.comp.comp = num_ops;
} else if (pooltype === 'AVE') {
d.comp.add = num_ops;
} else {
d.wOut = Math.ceil((d.wIn + 2 * pad - kernel) / stride) + 1;
d.hOut = Math.ceil((d.hIn + 2 * pad - kernel) / stride) + 1;
onerror("Unknown pooling type " + pooltype);
}
d.chOut = d.chIn = parent.chOut;
break;
case "BATCHNORM":
case "batchnorm":
d.wIn = parent.wOut;
d.hIn = parent.hOut;
d.wOut = d.wIn;
d.hOut = d.hIn;
d.chOut = d.chIn = parent.chOut;
d.comp.add = d.wIn * d.hIn * d.chIn;
d.comp.div = d.wIn * d.hIn * d.chIn;
break;
case "LRN":
mode = (ref16 = n.attribs.lrn_param.norm_region) != null ? ref16 : 'ACROSS_CHANNELS';
case "lrn":
mode = (ref18 = n.attribs.lrn_param.norm_region) != null ? ref18 : 'ACROSS_CHANNELS';
size = n.attribs.lrn_param.local_size;
d.wIn = parent.wOut;
d.hIn = parent.hOut;
d.wOut = d.wIn;
d.hOut = d.hIn;
d.chOut = d.chIn = parent.chOut;
num_inputs = d.wIn * d.hIn * d.chIn;
d.comp.macc = num_inputs * size;
d.comp.add = num_inputs;
d.comp.exp = num_inputs;
d.comp.div = num_inputs * 2;
break;
case "CONCAT":
case "concat":
d.wIn = parent.wOut;
d.hIn = parent.hOut;
d.wOut = d.wIn;
d.hOut = d.hIn;
ref17 = n.parents;
for (j = 0, len1 = ref17.length; j < len1; j++) {
p = ref17[j];
ref19 = n.parents;
for (j = 0, len1 = ref19.length; j < len1; j++) {
p = ref19[j];
d.chIn += p.analysis.chOut;
}
d.chOut = d.chIn;
ref18 = n.parents;
for (k = 0, len2 = ref18.length; k < len2; k++) {
p = ref18[k];
ref20 = n.parents;
for (k = 0, len2 = ref20.length; k < len2; k++) {
p = ref20[k];
failed = failed || (p.analysis.wOut !== d.wIn || p.analysis.hOut !== d.hIn);
}
if (failed) {
window.onerror('CONCAT: input dimensions dont agree!');
}
break;
case "RELU":
case "DROPOUT":
case "SOFTMAX":
case "SOFTMAXWITHLOSS":
case "SOFTMAX_LOSS":
case "relu":
case "dropout":
d.wIn = parent.wOut;
d.hIn = parent.hOut;
d.wOut = d.wIn;
d.hOut = d.hIn;
d.chOut = d.chIn = parent.chOut;
d.comp.comp = d.wIn * d.hIn * d.chIn;
break;
case "softmax":
case "softmaxwithloss":
case "softmax_loss":
d.wIn = parent.wOut;
d.hIn = parent.hOut;
d.wOut = d.wIn;
d.hOut = d.hIn;
d.chOut = d.chIn = parent.chOut;
d.comp.exp = d.wIn * d.hIn * d.chIn;
d.comp.add = d.wIn * d.hIn * d.chIn;
d.comp.div = d.wIn * d.hIn * d.chIn;
break;
case "FLATTEN":
case "flatten":
d.wIn = parent.wOut;
d.hIn = parent.hOut;
d.chIn = parent.chOut;
d.wOut = d.hOut = 1;
d.chOut = d.chIn * d.wIn * d.hIn;
break;
case "IMPLICIT":
case "implicit":
d.wIn = d.hIn = 0;
d.chIn = 0;
d.wOut = d.hOut = 0;
Expand All @@ -136,7 +175,8 @@ module.exports = Analyzer = (function() {
console.log(n);
debugger;
}
if (layertype !== "RELU" && layertype !== "SOFTMAX" && layertype !== "SOFTMAXWITHLOSS" && layertype !== "SOFTMAX_LOSS") {
trivial_layers = ["relu", "softmax", "softmaxwithloss", "softmax_loss", "dropout", "concat"];
if ($.inArray(layertype, trivial_layers) === -1) {
_.extend(n.attribs, {
analysis: {
"in": d.chIn + 'ch ⋅ ' + d.wIn + '×' + d.hIn,
Expand Down
Loading

0 comments on commit fdd4bc9

Please sign in to comment.