Feature Space Duplication in Implementation

[Leerw]

The paper says that you expand the number of points in the feature space, but from your implementation, if I got correct, the feature is concatenated with uniformly sampled 2d grids. seems that the paper's demonstration is different with your implementation, so, how about the performance of simply expand the number of points in the feature space? Is it worse than the way in your implementation?

Dis-PU/Common/ops.py

Lines 1152 to 1199 in ee9a74d

	def duplicate_up(pc, feature,is_training, up_ratio=4, use_bn=False,bn_decay=False,scope="up_shuffle_layer4", atten=False,edge=False):
	with tf.variable_scope(scope,reuse=tf.AUTO_REUSE):
	B, N, C = feature.get_shape()
	net = tf.expand_dims(feature,axis=2)
	# if up_ratio<2:
	# grid = gen_1d_grid(up_ratio)
	# else:
	grid = gen_grid(up_ratio) #[1,R,2]
	grid = tf.tile(tf.expand_dims(grid, 0), [tf.shape(net)[0], 1,tf.shape(net)[1]]) # [B,R,2*N]
	#grid = tf.tile(tf.expand_dims(grid, 0), [tf.shape(net)[0], tf.shape(net)[1], 1]) # [B,N*R,2]
	# [B,N*R,1,2]
	grid = tf.reshape(grid, [tf.shape(net)[0], -1, 1, 2])
	#grid = tf.expand_dims(grid, axis=2)
	net = tf.tile(net, [1, up_ratio, 1, 1])
	net = tf.concat([net, grid], axis=-1)
	if atten:
	net = attention_unit(net, is_training=is_training)
	if edge:
	net = tf.squeeze(net, axis=2)
	net = EdgeConv(net, 256, k=16, is_training=is_training, bn=use_bn, bn_decay=bn_decay,
	scope='shuffle_layer_0')
	net = EdgeConv(net, 128, k=16, is_training=is_training, bn=use_bn, bn_decay=bn_decay,
	scope='shuffle_layer_1')
	else:
	net = conv2d(net, 256, [1, 1],
	padding='VALID', stride=[1, 1],
	bn=False, is_training=is_training,
	scope='conv1', bn_decay=bn_decay)
	net = conv2d(net, 128, [1, 1],
	padding='VALID', stride=[1, 1],
	bn=False, is_training=is_training,
	scope='conv2', bn_decay=bn_decay)
	net = tf.squeeze(net,axis=2)
	return net

[ququuy]

I think these codes correspond to section 3.2 of the paper:
"Hence, in the feature expansion unit, we adopt the commonly-used expansion operation by duplicating FP with r copies and concatenating with a regular 2D grid to obtain FE."

[anshen666]

I think these codes correspond to section 3.2 of the paper: "Hence, in the feature expansion unit, we adopt the commonly-used expansion operation by duplicating FP with r copies and concatenating with a regular 2D grid to obtain FE."
请问方便留个联系方式吗？这个项目一直没跑通，请问一定要在ubuntu环境下才可以跑通吗？