|
|
@@ -11,12 +11,13 @@ cntk.ModelFactory = class {
|
|
|
match(context, host) {
|
|
|
var extension = context.identifier.split('.').pop().toLowerCase();
|
|
|
var buffer = null;
|
|
|
- if (extension == 'model') {
|
|
|
+ if (extension == 'model' || extension == 'cmf' || extension == 'dnn') {
|
|
|
buffer = context.buffer;
|
|
|
+ // Filter MXNet Model server ZIP files
|
|
|
if (buffer && buffer.length > 2 && buffer[0] == 0x50 && buffer[1] == 0x4B) {
|
|
|
return false;
|
|
|
}
|
|
|
- // Filter PyTorch models published with incorrect .h5 file extension.
|
|
|
+ // Filter PyTorch models published with incorrect file extension.
|
|
|
var torch = [ 0x8a, 0x0a, 0x6c, 0xfc, 0x9c, 0x46, 0xf9, 0x20, 0x6a, 0xa8, 0x50, 0x19 ];
|
|
|
if (buffer && buffer.length > torch.length + 2 &&
|
|
|
buffer[0] == 0x80 && buffer[1] > 0x00 && buffer[1] < 0x05) {
|
|
|
@@ -28,6 +29,7 @@ cntk.ModelFactory = class {
|
|
|
}
|
|
|
if (extension == 'cntk') {
|
|
|
buffer = context.buffer;
|
|
|
+ // Check protobuf data to exclude .cntk text files
|
|
|
if (buffer && buffer.length > 3 && buffer[0] == 0x08 && buffer[1] < 0x10 && buffer[2] == 0x12) {
|
|
|
return true;
|
|
|
}
|
|
|
@@ -420,14 +422,27 @@ cntk.Node = class {
|
|
|
}
|
|
|
}
|
|
|
else {
|
|
|
- this._operator = this._metadata.getOperatorName(obj.op);
|
|
|
- if (this._operator == null) {
|
|
|
- this._operator = obj.op.toString();
|
|
|
- }
|
|
|
+ if (!obj.hasOwnProperty('op')) {
|
|
|
+ this._operator = obj.type;
|
|
|
+ // if ( === false
|
|
|
+ if (obj.user_defined_state) {
|
|
|
+ Object.keys(obj.user_defined_state).forEach((key) => {
|
|
|
+ this._attributes.push(new cntk.Attribute(this._metadata, this._operator, key, obj.user_defined_state[key]));
|
|
|
+ });
|
|
|
+ }
|
|
|
+ }
|
|
|
+ else {
|
|
|
+ this._operator = this._metadata.getOperatorName(obj.op);
|
|
|
+ if (this._operator == null) {
|
|
|
+ this._operator = obj.op ? obj.op.toString() : '?';
|
|
|
+ }
|
|
|
+ }
|
|
|
+ }
|
|
|
+ if (obj.attributes) {
|
|
|
+ Object.keys(obj.attributes).forEach((key) => {
|
|
|
+ this._attributes.push(new cntk.Attribute(this._metadata, this._operator, key, obj.attributes[key]));
|
|
|
+ });
|
|
|
}
|
|
|
- Object.keys(obj.attributes).forEach((key) => {
|
|
|
- this._attributes.push(new cntk.Attribute(this._metadata, this._operator, key, obj.attributes[key]));
|
|
|
- });
|
|
|
obj.inputs.forEach((input) => {
|
|
|
var connection = connections[input];
|
|
|
if (connection) {
|
|
|
@@ -830,7 +845,7 @@ cntk.Metadata = class {
|
|
|
|
|
|
getOperatorName(code) {
|
|
|
// cntk/Source/CNTKv2LibraryDll/API/Internals/PrimitiveOpType.h
|
|
|
- return this._operatorMap[code.toString()] || null;
|
|
|
+ return this._operatorMap[code] || null;
|
|
|
}
|
|
|
|
|
|
getSchema(operator) {
|
|
|
@@ -1063,7 +1078,86 @@ cntk_v1.ComputationNetwork = class {
|
|
|
this.convertRunningVariancePending = true;
|
|
|
}
|
|
|
};
|
|
|
-
|
|
|
+ op.Tanh = function(reader) {};
|
|
|
+ op.Sigmoid = function(reader) {};
|
|
|
+ op.Logistic = function(reader) {};
|
|
|
+ op.SquareError = function(reader) {};
|
|
|
+ op.ErrorPrediction = function(reader) {};
|
|
|
+ op.RowStack = function(reader) {
|
|
|
+ this.spliceDim = (reader.version >= 3) ? reader.int32() : 1;
|
|
|
+ };
|
|
|
+ op.Slice = function(reader) {
|
|
|
+ var num = 1;
|
|
|
+ if (reader.version >= 22) {
|
|
|
+ num = reader.int32();
|
|
|
+ }
|
|
|
+ this.index = [];
|
|
|
+ this.axis = [];
|
|
|
+ this.strideMultiplier = [];
|
|
|
+ for (var i = 0; i < num; i++) {
|
|
|
+ this.index.push([ [ reader.uint64(), reader.uint64() ] ]);
|
|
|
+ if (reader.version >= 3) {
|
|
|
+ this.axis.push(reader.int32());
|
|
|
+ }
|
|
|
+ if (reader.version >= 27) {
|
|
|
+ this.strideMultiplier.push(reader.int32());
|
|
|
+ }
|
|
|
+ }
|
|
|
+ };
|
|
|
+ op.PastValue = function(reader) {
|
|
|
+ this.timeStep = reader.int32();
|
|
|
+ if (reader.version > 3)
|
|
|
+ {
|
|
|
+ this.sampleLayout = new cntk_v1.TensorShape(reader, false);
|
|
|
+ }
|
|
|
+ else
|
|
|
+ {
|
|
|
+ var rows = reader.uint64();
|
|
|
+ reader.uint64();
|
|
|
+ this.sampleLayout = new cntk_v1.TensorShape([ rows ], true);
|
|
|
+ }
|
|
|
+ if (reader.version >= 2)
|
|
|
+ {
|
|
|
+ this.initialStateValue = reader.int32();
|
|
|
+ }
|
|
|
+ };
|
|
|
+ op.FutureValue = function(reader) {
|
|
|
+ this.timeStep = reader.int32();
|
|
|
+ if (reader.version > 3)
|
|
|
+ {
|
|
|
+ this.sampleLayout = new cntk_v1.TensorShape(reader, false);
|
|
|
+ }
|
|
|
+ else
|
|
|
+ {
|
|
|
+ var rows = reader.uint64();
|
|
|
+ reader.uint64();
|
|
|
+ this.sampleLayout = new cntk_v1.TensorShape([ rows ], true);
|
|
|
+ }
|
|
|
+ if (reader.version >= 2)
|
|
|
+ {
|
|
|
+ this.initialStateValue = reader.int32();
|
|
|
+ }
|
|
|
+ };
|
|
|
+ op.TransposeDimensions = function(reader) {
|
|
|
+ if (reader.version >= 3)
|
|
|
+ {
|
|
|
+ this.axis1 = reader.int32();
|
|
|
+ this.axis2 = reader.int32();
|
|
|
+ if (reader.version >= 25 && this.axis1 == 0 && this.axis2 == 0)
|
|
|
+ {
|
|
|
+ var size = reader.uint64();
|
|
|
+ this.perm = [];
|
|
|
+ for (var i = 0; i < size; i++) {
|
|
|
+ this.perm.push(reader.uint64());
|
|
|
+ }
|
|
|
+ }
|
|
|
+ }
|
|
|
+ else {
|
|
|
+ this.axis1 = 1;
|
|
|
+ this.axis2 = 2;
|
|
|
+ }
|
|
|
+ };
|
|
|
+
|
|
|
var nodes = [];
|
|
|
this.nodes = {};
|
|
|
for (var i = 0; i < numNodes; i++) {
|
Lutz Roeder