A mean and std computation and use question

[RobinWitch]

Hello, what a good work!
While I have a question when I retrain rvqvae, it quite diffuse me a lot.
Why in this place we need to recalculate std?

momask-codes/data/t2m_dataset.py

Lines 41 to 68 in d4d3412

	if opt.is_train:
	# root_rot_velocity (B, seq_len, 1)
	std[0:1] = std[0:1] / opt.feat_bias
	# root_linear_velocity (B, seq_len, 2)
	std[1:3] = std[1:3] / opt.feat_bias
	# root_y (B, seq_len, 1)
	std[3:4] = std[3:4] / opt.feat_bias
	# ric_data (B, seq_len, (joint_num - 1)*3)
	std[4: 4 + (joints_num - 1) * 3] = std[4: 4 + (joints_num - 1) * 3] / 1.0
	# rot_data (B, seq_len, (joint_num - 1)*6)
	std[4 + (joints_num - 1) * 3: 4 + (joints_num - 1) * 9] = std[4 + (joints_num - 1) * 3: 4 + (
	joints_num - 1) * 9] / 1.0
	# local_velocity (B, seq_len, joint_num*3)
	std[4 + (joints_num - 1) * 9: 4 + (joints_num - 1) * 9 + joints_num * 3] = std[
	4 + (joints_num - 1) * 9: 4 + (
	joints_num - 1) * 9 + joints_num * 3] / 1.0
	# foot contact (B, seq_len, 4)
	std[4 + (joints_num - 1) * 9 + joints_num * 3:] = std[
	4 + (
	joints_num - 1) * 9 + joints_num * 3:] / opt.feat_bias
	assert 4 + (joints_num - 1) * 9 + joints_num * 3 + 4 == mean.shape[-1]
	np.save(pjoin(opt.meta_dir, 'mean.npy'), mean)
	np.save(pjoin(opt.meta_dir, 'std.npy'), std)
	self.mean = mean
	self.std = std
	print("Total number of motions {}, snippets {}".format(len(self.data), self.cumsum[-1]))

[EricGuo5513]

Hi, here we are calculating group-based std and also amplifying the importance of certain motion features.

…

On Wed, Mar 27, 2024 at 9:13 PM RobinWitch ***@***.***> wrote: Hello, what a good work! While I have a question when I retrain rvqvae, it quite diffuse me a lot. Why in this place we need to recalculate std? https://github.com/EricGuo5513/momask-codes/blob/d4d34120ea01e3fda09944c3c030e8befe8c19e2/data/t2m_dataset.py#L43 — Reply to this email directly, view it on GitHub <#33>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AKRYNBYUESAOME5XNSJEQJ3Y2ODF3AVCNFSM6AAAAABFMAUNMKVHI2DSMVQWIX3LMV43ASLTON2WKOZSGIYTEMRWGIZTIMQ> . You are receiving this because you are subscribed to this thread.Message ID: ***@***.***>

[RobinWitch]

In intuitive, I think amplifying the importance of certain motion features we should plus opt.feat_bias instead of dividing? The default value of opt.feat_bias is 5.