def _test_sparse_conv2d_correctness(self,
xsize,
bsize,
ksize,
strides,
padding,
dtype=tf.float32):
np.random.RandomState(0)
x = tf.constant(np.random.uniform(-1, 1, xsize), dtype=dtype)
w = tf.constant(np.random.uniform(-1, 1, ksize), dtype=dtype)
mask = np.random.uniform(-5, 1, xsize[:-1])
mask = tf.constant((mask > 0).astype(np.float32), dtype=dtype)
mask_out = tf.nn.max_pool(
tf.expand_dims(mask, 3), [1, ksize[0], ksize[1], 1], strides, padding)
y_exp = mask_conv2d(x, w, mask, strides, padding)
ind_blk = convert_mask_to_block_indices(mask, bsize, ksize, strides, padding, 0.)
y_act = sparse_conv2d(x, w, ind_blk, strides, padding)
with self.test_session():
y_exp_val = y_exp.eval()
y_act_val = y_act.eval()
mask_val = mask_out.eval()
mask_val = np.tile(mask_val, [1, 1, 1, ksize[-1]])
y_exp_val = y_exp_val * mask_val
y_act_val = y_act_val * mask_val
y_exp_val = y_exp_val[mask_val > 0.]
y_act_val = y_act_val[mask_val > 0.]
np.testing.assert_allclose(y_act_val, y_exp_val, rtol=1e-2, atol=0.1)
def _test_sparse_conv2d_matmul_correctness(self,
xsize,
ksize,
padding,
dtype=tf.float32):
np.random.RandomState(0)
strides = [1, 1, 1, 1] # Only [1, 1, 1, 1] is supported currently.
# Use block size to be the same with kernel size makes it the same with matrix multiplication.
bsize = [1, ksize[0], ksize[1], 1]
x = tf.constant(np.random.uniform(-1, 1, xsize), dtype=dtype)
w = tf.constant(np.random.uniform(-1, 1, ksize), dtype=dtype)
mask = np.random.uniform(-5, 1, xsize[:-1])
mask = tf.constant((mask > 0).astype(np.float32), dtype=dtype)
mask_out = tf.nn.max_pool(tf.expand_dims(mask, 3),
[1, ksize[0], ksize[1], 1], strides, padding)
y_exp = mask_conv2d(x, w, mask, strides, padding)
ind_blk = convert_mask_to_block_indices(mask, bsize, ksize, strides,
padding, 0.)
y_act = sparse_conv2d_matmul(x, w, ind_blk, strides, padding)
with self.test_session():
y_exp_val = y_exp.eval()
y_act_val = y_act.eval()
mask_val = mask_out.eval()
mask_val = np.tile(mask_val, [1, 1, 1, ksize[-1]])
y_exp_val = y_exp_val * mask_val
y_act_val = y_act_val * mask_val
y_exp_val = y_exp_val[mask_val > 0.]
y_act_val = y_act_val[mask_val > 0.]
np.testing.assert_allclose(y_act_val,
y_exp_val,
rtol=1e-2,
atol=1e-5)
def _test_sparse_conv2d_with_mask(self, mask, bsize, ksize, strides, padding, y_exp):
mask_ = tf.constant(mask)
ind_blk = convert_mask_to_block_indices(mask_, bsize, ksize, strides, padding, 0.)
x = tf.constant(np.ones([1, mask.shape[1], mask.shape[2], 1], dtype=np.float32))
w = tf.constant(np.ones(ksize, dtype=np.float32))
y = sparse_conv2d(x, w, ind_blk, strides, padding)
with self.test_session():
y_act = y.eval()
self.assertEqual(y_act.size, y_exp.size)
np.testing.assert_array_equal(y_act.reshape(y_exp.shape), y_exp)
def _sparse_conv2d_with_mask(x, w, strides, padding, mask, bsize, tol):
"""Sparse conv 2d with mask."""
ksize = [int(ss) for ss in w.get_shape()]
ind_blk = convert_mask_to_block_indices(mask,
bsize,
ksize,
strides,
padding,
tol,
avgpool=True)
y = sparse_conv2d(x, w, ind_blk, strides, padding)
return y
def gather_tf(x, mask, bsize, ksize, strides, padding):
blk_indices = convert_mask_to_block_indices(mask, bsize, ksize, strides,
padding, 0.0)
blk_shape = tf.shape(blk_indices)
blk_indices_ = tf.reshape(blk_indices, [-1, 3])
# Calculate the block strides.
bstrides = _calc_block_strides(blk_shape, ksize, strides)
# Pad input.
x_ = _pad_input(x,
ksize,
strides,
padding,
bsize=[1, blk_shape[1], blk_shape[2], 1],
bstrides=bstrides)
# Gather patches.
p = tf.gather_nd(x_, blk_indices_)
# Reshape patches.
p = tf.reshape(p, [blk_shape[0], blk_shape[1], blk_shape[2], -1])
return p, blk_indices