fix unittest warning for python3 and add submodule numpy-groupies

Author: wkcn

.gitmodules

@@ -0,0 +1,3 @@
+[submodule "thirdparty/numpy-groupies"]
+	path = thirdparty/numpy-groupies
+	url = https://github.com/ml31415/numpy-groupies

README.md

@@ -23,7 +23,13 @@ It's **an efficent Python-DNN Implementation used numpy mainly**, and it's aimed
 	Mobula is implemented by only Python. You can modify the code easily to implement what you want.
 
 ## How to install it?
+- Git:
+```bash
+git clone https://github.com/wkcn/mobula --recursive
 ```
+
+- pip:
+```bash
 pip install mobula
 ```
 

mobula/layers/Concat.py

@@ -11,6 +11,7 @@ def reshape(self):
             ss = x.shape[self.axis]
             self.slices[i][self.axis] = slice(last, last + ss)
             last += ss
+        self.slices = [tuple(s) for s in self.slices]
 
         shp = list(self.X[0].shape)
         shp[self.axis] = last

mobula/layers/Crop.py

@@ -16,7 +16,7 @@ def reshape(self):
             assert len(self.offset) == n, "len(Crop.offset) must be the number of dims which will be cropped"
             self.axes = [slice(None)] * self.axis + [slice(o, o + self.R.Y.shape[i + self.axis]) for i, o in enumerate(self.offset)]
     def forward(self):
-        self.Y = self.X[self.axes]
+        self.Y = self.X[tuple(self.axes)]
     def backward(self):
         self.dX = np.zeros(self.X.shape)
-        self.dX[self.axes] = self.dY
+        self.dX[tuple(self.axes)] = self.dY

mobula/layers/Slice.py

@@ -19,6 +19,7 @@ def reshape(self):
             self.Y[i] = self.X[s]
         s = slice(last, None)
         self.slices[-1][self.axis] = s
+        self.slices = [tuple(s) for s in self.slices]
         self.Y[-1] = self.X[s] 
     def forward(self):
         self.Y = [self.X[s] for s in self.slices]

mobula/layers/__init__.py

@@ -3,7 +3,9 @@
 
 PATH = os.path.dirname(__file__)
 sys.path.append(PATH)
-sys.path.append(PATH + "/../../thirdparty")
+thirdparties = ['numpy-groupies']
+for name in thirdparties:
+    sys.path.append(os.path.join(PATH, '../../thirdparty/', name))
 
 # Data Layer
 from .Data import *

tests/test_layers/test_concat.py

@@ -20,7 +20,7 @@ def go_concat(axis):
         def G(s):
             w = [slice(None) for _ in range(4)]
             w[axis] = s
-            return w
+            return tuple(w)
         ssa = G(sa)
         ssb = G(sb)
         ssc = G(sc)

tests/test_layers/test_crop.py

@@ -16,11 +16,11 @@ def test_crop():
             l = L.Crop([x1, x2], offset = offset, axis = axis)
             l.reshape()
             l.forward()
-            assert np.allclose(l.Y, X1[w]) 
+            assert np.allclose(l.Y, X1[tuple(w)])
             l.dY = np.random.random(l.Y.shape)
             l.backward()
             tmp = np.zeros(X1.shape)
-            tmp[w] = l.dY
+            tmp[tuple(w)] = l.dY
             assert np.allclose(l.dX, tmp)
 
 def test_crop2():

tests/test_layers/test_slice.py

@@ -45,8 +45,8 @@ def go_slice1(axis):
         axes2 = [slice(None)] * 4
         axes1[axis] = slice(None, slice_point) 
         axes2[axis] = slice(slice_point, None) 
-        assert np.allclose(y1.Y, X[axes1])
-        assert np.allclose(y2.Y, X[axes2])
+        assert np.allclose(y1.Y, X[tuple(axes1)])
+        assert np.allclose(y2.Y, X[tuple(axes2)])
 
         y1.dY = np.random.random(y1.Y.shape)
         y2.dY = np.random.random(y2.Y.shape)
@@ -64,9 +64,9 @@ def go_slice2(axis):
         axes1[axis] = slice(None, 3) 
         axes2[axis] = slice(3, 5) 
         axes3[axis] = slice(5, None) 
-        assert np.allclose(y1.Y, X[axes1])
-        assert np.allclose(y2.Y, X[axes2])
-        assert np.allclose(y3.Y, X[axes3])
+        assert np.allclose(y1.Y, X[tuple(axes1)])
+        assert np.allclose(y2.Y, X[tuple(axes2)])
+        assert np.allclose(y3.Y, X[tuple(axes3)])
 
 
         y1.dY = np.random.random(y1.Y.shape)

tests/test_layers/test_softmax.py

@@ -22,7 +22,7 @@ def test_softmax():
         axes[axis] = np.newaxis
         pu = [1] * 4
         pu[axis] = a.shape[axis]
-        s = np.tile(su[axes], pu)
+        s = np.tile(su[tuple(axes)], pu)
 
         # softmax forward
         assert np.allclose(y, exp / s)