Inherited from the V2XSimSeg dataset, now support multi timestamp loading output padded_voxel_points_next are for other timestamp

Inherited from the V2XSimDet dataset, now support multi timestamp training output padded_voxel_points_next are for other timestamp

Decay the learning rate with half-cycle cosine after warmup

x_ts: [bxa x time_stamp, L, D] NOTE: the returned mask and ids_restore has a different dimensionality from the random masking

x_ts: [bxa x time_stamp, L, D] NOTE: the returned mask and ids_restore has a different dimensionality from the random masking

x_ts: [bxa x time_stamp, L, D] NOTE: the returned mask and ids_restore has a different dimensionality from the random masking

x_ts: [BxAxts, L, D] mask: [BxA, ts, L] -> this is real L ids_restore: [BxA, L] -> this is real L restore, add tempor

x_ts: [BxAxts, L, D] mask: [BxA, ts, L] -> this is real L ids_restore: [BxA, L] -> this is real L restore, add tempor

x_ts: [BxAxts, L, D] mask: [BxA, ts, L] -> this is real L ids_restore: [BxA, L] -> this is real L restore, add tempor

Random masking for each data of each timestamp. Maintain indices for decoder to fuse and possibly fill in mask tokens x_ts: [

Random masking for each data of each timestamp. Maintain indices for decoder to fuse and possibly fill in mask tokens x_ts: [

Random masking for each data of each timestamp. Maintain indices for decoder to fuse and possibly fill in mask tokens x_ts: [

Reverse masking with fillings from other timestamp, using mask and torch where x_ts: [BxAxts, L, D] mask: [BxAxts, L]

Reverse masking with fillings from other timestamp, using mask and torch where x_ts: [BxAxts, L, D] mask: [BxAxts, L] -> full

Reverse masking with fillings from other timestamp, using mask and torch where x_ts: [BxAxts, L, D] mask: [BxAxts, L] -> full

mask x1 with mask_ratio mask x2 with 1 - mask ratio so that they are complementary x1, x2: same size [N, D, L]

Modified from the original forward, make fusion happen

Encoder encodes each timestamp alone Decoder fuses multi timestamp together

Encoder encodes each timestamp alone Decoder fuses multi timestamp together

Encoder encodes each timestamp alone Decoder fuses multi timestamp together

x1: [bxa, C, H, W] x_next: [bxa, ts-1, C, H, W] beq_next_frames

x1: [bxa, C, H, W] x_next: [bxa, ts-1, C, H, W] beq_next_frames

x1: [bxa, C, H, W] x_next: [bxa, ts-1, C, H, W] beq_next_frames

x1: [bxa, C, H, W] x_next: [bxa, ts-1, C, H, W] beq_next_frames

overwrite the original forward_loss, now calculate loss on the entire image

overwrite the original forward_loss, now calculate loss on the entire image

eval_mask_lidar is only to ingore unknown voxels in groundtruth