add a simple example for using matrix multiplication layer

Author: simbaforrest

examples/matrix_multiplication/py_data_layer.py

@@ -0,0 +1,26 @@
+import caffe
+import numpy as np
+
+class LineDataLayer(caffe.Layer):
+    def setup(self, bottom, top):
+        self.batch_size = 10;
+        top[0].reshape(self.batch_size, 2, 1)
+        top[1].reshape(self.batch_size)
+        
+        self.W = np.array([2,5], dtype=np.float32)
+        self.b = -7.0
+        print("W=\n"+str(self.W))
+        print("b="+str(self.b))
+    
+    def forward(self, bottom, top):
+        for b in xrange(self.batch_size):
+            x = np.random.rand(2).astype(np.float32)
+            y = self.W.dot(x)+self.b
+            top[0].data[b,...]=x.reshape(2,1)
+            top[1].data[b,...]=y
+    
+    def reshape(self, bottom, top):
+        pass
+    
+    def backward(self, top, propagate_down, bottom):
+        pass

examples/matrix_multiplication/solve_ip.ptt

@@ -0,0 +1,15 @@
+train_net: "train_line_fitting_ip.ptt"
+lr_policy: "poly"
+base_lr: 1.0e-1
+power: 1
+gamma: 0.1
+iter_size: 10
+stepsize: 2000
+average_loss: 20
+display: 1
+max_iter: 5000
+momentum: 0.9
+weight_decay: 0
+snapshot: 5000
+snapshot_prefix: "train_line_fitting_ip"
+solver_mode: GPU

examples/matrix_multiplication/solve_ip.py

@@ -0,0 +1,16 @@
+import numpy as np
+import sys
+import caffe
+
+# init
+caffe.set_mode_gpu()
+caffe.set_device(0)
+
+solver = caffe.SGDSolver("solve_ip.ptt")
+print("W_0=\n")
+print(np.array(solver.net.params['ip_WXpb'][0].data))
+solver.solve()
+print("W=\n")
+print(np.array(solver.net.params['ip_WXpb'][0].data))
+print("b=\n")
+print(np.array(solver.net.params['ip_WXpb'][1].data))

examples/matrix_multiplication/solve_matmul.ptt

@@ -0,0 +1,15 @@
+train_net: "train_line_fitting_matmul.ptt"
+lr_policy: "poly"
+base_lr: 1.0e-1
+power: 1
+gamma: 0.1
+iter_size: 10
+stepsize: 2000
+average_loss: 20
+display: 1
+max_iter: 5000
+momentum: 0.9
+weight_decay: 0
+snapshot: 5000
+snapshot_prefix: "train_line_fitting_matmul"
+solver_mode: GPU

examples/matrix_multiplication/solve_matmul.py

@@ -0,0 +1,17 @@
+import numpy as np
+import sys
+import caffe
+
+# init
+caffe.set_mode_gpu()
+caffe.set_device(0)
+
+solver = caffe.SGDSolver("solve_matmul.ptt")
+solver.net.blobs['W'].data[...]=np.random.randn(1,2)
+print("W_0=\n")
+print(np.array(solver.net.blobs['W'].data))
+solver.solve()
+print("W=\n")
+print(np.array(solver.net.blobs['W'].data))
+print("b=\n")
+print(solver.net.params["add_WX_b"][0].data)

examples/matrix_multiplication/train_line_fitting_ip.ptt

@@ -0,0 +1,36 @@
+layer {
+  name: "data"
+  type: "Python"
+  top: "X"
+  top: "Y"
+  python_param {
+    module: "py_data_layer"
+    layer: "LineDataLayer"
+  }
+}
+layer {
+  name: "ip_WXpb"
+  type: "InnerProduct"
+  param { decay_mult: 0 }
+  param { decay_mult: 0 }
+  inner_product_param {
+    num_output: 1
+    weight_filler {
+      type: "gaussian"
+      std: 0.01
+    }
+    bias_filler {
+      type: "constant"
+      value: 0
+    }
+  }
+  bottom: "X"
+  top: "WXpb"  # 10x1x1
+}
+layer {
+  name: "loss_L2"
+  type: "EuclideanLoss"
+  bottom: "WXpb"
+  bottom: "Y"
+  top: "loss"
+}

examples/matrix_multiplication/train_line_fitting_matmul.ptt

@@ -0,0 +1,40 @@
+layer {
+  name: "data"
+  type: "Python"
+  top: "X"
+  top: "Y"
+  python_param {
+    module: "py_data_layer"
+    layer: "LineDataLayer"
+  }
+}
+layer {
+  name: "param_W"
+  type: "Parameter"
+  top: "W"
+  parameter_param { shape: { dim: 1 dim: 2 } }
+}
+layer {
+  name: "matmul_W_X"
+  type: "MatrixMultiplication"
+  bottom: "W"
+  bottom: "X"
+  top: "WX"  # 10x1x1
+}
+layer {
+  name: "add_WX_b"
+  type: "Bias"
+  bottom: "WX"
+  top: "WXpb"
+  bias_param {
+    num_axes: 0
+  }
+  param { decay_mult: 0 }
+}
+layer {
+  name: "loss_L2"
+  type: "EuclideanLoss"
+  bottom: "WXpb"
+  bottom: "Y"
+  top: "loss"
+}