Add customized python layer for converting tf.StridedSlice layer.

Author: vera121

python/stridedslice.py

@@ -0,0 +1,80 @@
+import caffe
+import numpy as np
+
+
+class TensorSlice(caffe.Layer):
+    """
+    Get a tensor's slicing: realize the function of the tf.strided_slice().
+
+    TO OPTIMIZE: For simplification, assume all the input tensors are 4 dimensions now.
+    """
+
+    def setup(self, bottom, top):
+        # check number of inputs and outputs
+        if len(bottom) != 1:
+            raise Exception("Only input one Tensor at a time!")
+        if len(top) != 1:
+	    raise Exception("Only output one Tensor at a time!")
+
+	params = eval(self.param_str)
+	self.begin = np.array(params["begins"])
+	self.end = np.array(params["ends"])
+	if params["strides"] != None:
+	  self.strides = np.array(params["strides"])
+	else:
+	  self.strides = np.array([1, 1, 1, 1])
+	
+	if params["beginmask"] != None:
+	  self.beginmask = int(params["beginmask"])
+	else:
+	  self.beginmask = None
+	if params["endmask"] != None:
+	  self.endmask = int(params["endmask"])
+	else:
+	  self.endmask = None
+
+	# Handles the condition where the "ends" is assigned (0, 0, 0, 0) meaning the end of the input Tensor 
+	if ((self.end==0).all() and (self.strides>0).all()):
+	  self.end = bottom[0].shape
+	
+	# According to the tf.strided_slice():
+	# If the ith bit of beginmask is set, begin[i] is ignored and the fullest possible range in that dimension is used instead;
+	# The endmask works analogously, except with the end range.
+	if self.beginmask != None:
+	  self.begin[self.beginmask] = 0
+	if self.endmask != None:
+	  self.end[self.endmask] = bottom[0].shape[self.endmask]
+	
+	# other parameters...
+
+
+    def reshape(self, bottom, top):
+        # check input dimensions
+        if bottom[0].count == 0:
+            raise Exception("Input must not be empty!")
+	
+	#top[0].reshape(*bottom[0].data.shape)		
+	#dim = len(self.begin) 
+	num = [0, 0, 0, 0]
+	for i in range(4):
+	  if self.strides[i]==0:
+	    raise Exception("Strides should never equal to 0!")
+	  else: 
+	    num[i] = (abs(self.end[i]-self.begin[i])/abs(self.strides[i]))	
+	top[0].reshape(num[0], num[1], num[2], num[3])
+
+
+    def forward(self, bottom, top):
+	#pass
+	#top[0].data[...] = bottom[0].data[:]
+	top[0].data[...] = np.zeros(top[0].shape)
+	for i in range(len(self.begin)):
+       	  top[0].data[...] = bottom[0].data[self.begin[0]:self.end[0]:self.strides[0], self.begin[1]:self.end[1]:self.strides[1], self.begin[2]:self.end[2]:self.strides[2], self.begin[3]:self.end[3]:self.strides[3]]
+
+
+    def backward(self, top, propagate_down, bottom):
+	for i in range(len(propagate_down)):
+            if not propagate_down[i]:
+                continue
+            bottom[i].diff[...] = top[i].diff[:]
+