local change

.gitignore

@@ -28,3 +28,6 @@ multigpu-trained
 *-b*.prototxt
 *tmp*
 *Models*
+*.caffemodel
+*.solverstate
+

configs/test.config

@@ -1,5 +1,5 @@
 flag:		testdpBenchmark		#Flag of current experiment
-tools:		mxnet			#Tools to benchmark
+tools:		cntk			#Tools to benchmark
 experiments: #<network type>; <network name>;  <device id>; <device count>;  <batch size>;  <number of epochs>;  <epoch size>; <Learning rate>
 {
 		fc;		fcn5;		0;		1;		4096;		2;		60000;		0.05
@@ -8,19 +8,19 @@ experiments: #<network type>; <network name>;  <device id>; <device count>;  <ba
 #		fc;		fcn5;		0;		1;		1024;		2;		60000;		0.05
 #		fc;		fcn5;		0;		1;		512;	   	2;		60000;		0.05
 #		fc;		fcn5;		0;		1;		342;	   	2;		60000;		0.05
-		cnn;		alexnet;	0;		1;		1024;		2;		50000;		0.01
-		cnn;		alexnet;	-1;		20;		1024;		2;		50000;		0.01
+#		cnn;		alexnet;	0;		1;		1024;		2;		50000;		0.01
+#		cnn;		alexnet;	-1;		20;		1024;		2;		50000;		0.01
 #		cnn;		alexnet;	0;		1;		512;  	 	2;		50000;		0.01
 #		cnn;		alexnet;	0;		1;		256;	   	2;		50000;		0.01
 #		cnn;		alexnet;	0;		1;		128;	   	2;		50000;		0.01
 #		cnn;		alexnet;	0;		1;		86;   		2;		50000;		0.01
-		cnn;		resnet;		0;		1;		128;		2;		50000;		0.01
-		cnn;		resnet;		-1;		20;		128;		2;		50000;		0.01
+#		cnn;		resnet;		0;		1;		128;		2;		50000;		0.01
+#		cnn;		resnet;		-1;		20;		128;		2;		50000;		0.01
 #		cnn;		resnet;		0;		1;		64;		2;		50000;		0.01
 #		cnn;		resnet;		0;		1;		32;		2;		50000;		0.01
 #		cnn;		resnet;		0;		1;		16;		2;		50000;		0.01
 #		cnn;		resnet;		0;		1;		11;		2;		50000;		0.01
-		rnn;		lstm;		0;		1;		1024;		2;		-1;		0.01
+#		rnn;		lstm;		0;		1;		1024;		2;		-1;		0.01
 #		rnn;		lstm;		0;		1;		512;   		2;		-1;		0.01
 #		rnn;		lstm;		0;		1;		256;   		2;		-1;		0.01
 #		rnn;		lstm;		0;		1;		128;   		2;		-1;		0.01

tools/caffe/cnn/alexnet/alexnet_imagenet.prototxt

@@ -49,10 +49,11 @@ layer {
   }
   data_param {
     source: "/home/datasets/imagenet/caffe/ilsvrc12_val_lmdb"
-    batch_size: 200
+    batch_size: 50
     backend: LMDB
   }
 }
+
 layer {
   name: "conv1"
   type: "Convolution"
@@ -218,14 +219,14 @@ layer {
     num_output: 384
     pad: 1
     kernel_size: 3
-    #group: 2
+    group: 2
     weight_filler {
       type: "gaussian"
       std: 0.01
     }
     bias_filler {
       type: "constant"
-      value: 1
+      value: 0.1
     }
   }
 }
@@ -259,7 +260,7 @@ layer {
     }
     bias_filler {
       type: "constant"
-      value: 0
+      value: 0.1
     }
   }
 }
@@ -301,16 +302,16 @@ layer {
     }
     bias_filler {
       type: "constant"
-      value: 1
+      value: 0.1
     }
   }
 }
-#layer {
-#  name: "relu6"
-#  type: "ReLU"
-#  bottom: "fc6"
-#  top: "fc6"
-#}
+layer {
+  name: "relu6"
+  type: "ReLU"
+  bottom: "fc6"
+  top: "fc6"
+}
 layer {
   name: "drop6"
   type: "Dropout"
@@ -341,16 +342,16 @@ layer {
     }
     bias_filler {
       type: "constant"
-      value: 1
+      value: 0.1
     }
   }
 }
-#layer {
-#  name: "relu7"
-#  type: "ReLU"
-#  bottom: "fc7"
-#  top: "fc7"
-#}
+layer {
+  name: "relu7"
+  type: "ReLU"
+  bottom: "fc7"
+  top: "fc7"
+}
 layer {
   name: "drop7"
   type: "Dropout"

tools/caffe/cnn/alexnet/solver.prototxt

@@ -1,14 +1,14 @@
-# The train/test net protocol buffer definition
-net: "alexnet-bBATCHSIZE.prototxt"
-base_lr: 0.001
-
-lr_policy: "fixed"
-max_iter: MAXITER 
-display: 1
-solver_mode: DEVICE 
+net: "train_val.prototxt"
+test_iter: 1000
+test_interval: 1000
+base_lr: 0.01
+lr_policy: "step"
+gamma: 0.1
+stepsize: 100000
+display: 20
+max_iter: 450000
 momentum: 0.9
-
-# covering the full 10,000 testing images.
-test_iter: 100
-# Carry out testing every 2 training epochs.
-test_interval: TESTINTERVAL 
+weight_decay: 0.0005
+snapshot: 10000
+snapshot_prefix: "caffe_alexnet_train"
+solver_mode: GPU

tools/caffe/cnn/alexnet/solver_imagenet.prototxt

@@ -1,6 +1,6 @@
 net: "alexnet_imagenet.prototxt"
-test_iter: 250
-test_interval: 5000
+test_iter: 1000
+test_interval: 1000
 base_lr: 0.01
 lr_policy: "step"
 gamma: 0.1

tools/cntk/cnn/alexnet/alexnet_imagenet.cntk

@@ -107,9 +107,9 @@ Train = {
         ParallelTrain = {
             parallelizationMethod = "DataParallelSGD"
             distributedMBReading = "true"
-            parallelizationStartEpoch = 3
+            parallelizationStartEpoch = 1
             DataParallelSGD = {
-                gradientBits = 4
+                gradientBits = 1
             }
         }
 

tools/cntk/cnn/alexnet/testimagenet.sh

@@ -1,2 +1,2 @@
-mpirun -np 16 -machinefile cluster16 cntk configFile=alexnet_imagenet2.cntk deviceId=auto minibatchSize=256 maxEpochs=40 parallelTrain=true command=Train
+#mpirun -np 16 -machinefile cluster16 cntk configFile=alexnet_imagenet2.cntk deviceId=auto minibatchSize=256 maxEpochs=40 parallelTrain=true command=Train
 #cntk configFile=alexnet_imagenet2.cntk deviceId=auto minibatchSize=256 command=Test