def testGain(self): shape = (3, 3, 10, 10) for dtype in [dtypes.float32, dtypes.float64]: init1 = init_ops.convolutional_orthogonal_2d(seed=1, dtype=dtype) init2 = init_ops.convolutional_orthogonal_2d(gain=3.14, seed=1, dtype=dtype) with self.test_session(graph=ops.Graph(), use_gpu=True): t1 = init1(shape).eval() t2 = init2(shape).eval() return np.allclose(t1, t2 / 3.14, rtol=1e-15, atol=1e-15)
def testGain(self): shape = (3, 3, 10, 10) for dtype in [dtypes.float32, dtypes.float64]: init1 = init_ops.convolutional_orthogonal_2d(seed=1, dtype=dtype) init2 = init_ops.convolutional_orthogonal_2d(gain=3.14, seed=1, dtype=dtype) with self.session(graph=ops.Graph(), use_gpu=True): t1 = init1(shape).eval() t2 = init2(shape).eval() self.assertAllClose(t1, t2 / 3.14)
def testShapesValues(self): def circular_pad(input_, width, kernel_size): """Pad input_ for computing (circular) convolution. Args: input_: the input tensor width: the width of the tensor. kernel_size: the kernel size of the filter. Returns: a tensor whose width is (width + kernel_size - 1). """ beg = kernel_size // 2 end = kernel_size - 1 - beg tmp_up = array_ops.slice(input_, [0, width - beg, 0, 0], [-1, beg, width, -1]) tmp_down = array_ops.slice(input_, [0, 0, 0, 0], [-1, end, width, -1]) tmp = array_ops.concat([tmp_up, input_, tmp_down], 1) new_width = width + kernel_size - 1 tmp_left = array_ops.slice(tmp, [0, 0, width - beg, 0], [-1, new_width, beg, -1]) tmp_right = array_ops.slice(tmp, [0, 0, 0, 0], [-1, new_width, end, -1]) final = array_ops.concat([tmp_left, tmp, tmp_right], 2) return final cout = 45 shape = [64, 28, 28, 32] outputs_shape = shape[0:-1] + [cout] dtype = dtypes.float32 tol = 1e-3 gain = 3.14 # Check orthogonality/isometry by computing the ratio between # the 2-norms of the inputs and ouputs. for kernel_size in [[1, 1], [2, 2], [3, 3], [4, 4], [5, 5]]: convolution = convolutional.conv2d inputs = random_ops.random_normal(shape, dtype=dtype) inputs_2norm = linalg_ops.norm(inputs) input_with_circular_pad = circular_pad(inputs, shape[1], kernel_size[0]) outputs = convolution( input_with_circular_pad, padding="valid", filters=cout, kernel_size=kernel_size, use_bias=False, kernel_initializer=init_ops.convolutional_orthogonal_2d(gain=gain)) outputs_2norm = linalg_ops.norm(outputs) my_ops = variables.global_variables_initializer() with self.test_session(use_gpu=True) as sess: sess.run(my_ops) # Check the shape of the outputs t = outputs.eval() self.assertAllEqual(t.shape, outputs_shape) # Check isometry of the orthogonal kernel. self.assertAllClose( sess.run(inputs_2norm)/np.sqrt(np.prod(shape)), sess.run(outputs_2norm)/(np.sqrt(np.prod(shape))*np.sqrt(gain)), rtol=tol, atol=tol)
def testInvalidShape(self): init1 = init_ops.convolutional_orthogonal_2d() with self.test_session(graph=ops.Graph(), use_gpu=True): self.assertRaises(ValueError, init1, shape=[3, 3, 6, 5])
def testDuplicatedInitializer(self): init = init_ops.convolutional_orthogonal_2d() self.assertFalse(duplicated_initializer(self, init, 1, (3, 3, 10, 10)))
def testInitializerDifferent(self): for dtype in [dtypes.float32, dtypes.float64]: init1 = init_ops.convolutional_orthogonal_2d(seed=1, dtype=dtype) init2 = init_ops.convolutional_orthogonal_2d(seed=2, dtype=dtype) self.assertFalse(identicaltest(self, init1, init2, (3, 3, 10, 10)))