recursively load all tensors in a nested dictionary structure onto the gpu.

the collate function for a given index

the collate function for a given index

the collate function for a given index

the collate function for a given index

(X - X_train)*(X - X_train) = -2X*X_train + X*X + X_train*X_train Params: -- matrix1: N1 x D -- matrix2: N2 x D

parse evaluation args

compute the cosine similarity loss

load the pretrained SMPL autoencoder

create dataloader from the caption path and generated motion path

sample two list with same length.

forward for evaluation inputs: ------- batch_data: clip_token motions offsets

compute the cosine similarity loss

compute the global translation loss inputs: ------- out_motion : [bs, num_frames, num_joints + 1, 6] the

compute the global translation loss inputs: ------- out_motion : [bs, num_frames, num_joints + 1, 6] the

compute the velocity loss

compute the velocity loss

convert lengths to mask

process the list of file path. this function is for quick debug propose such that we do not need to load the entire dataset.

encode text with clip. inputs: ------- raw_text list (batch_size length) of strings with input text prom

encode text with clip. inputs: ------- raw_text list (batch_size length) of strings with input text prom

Params: -- activation: num_samples x dim_feat Returns: -- mu: dim_feat -- sigma: dim_feat x dim_feat

calculate the masked l2 loss assuming a.shape == b.shape == bs, J, Jdim, seqlen assuming mask.shape == bs, 1, 1, seqlen

calculate the masked l2 loss assuming a.shape == b.shape == bs, J, Jdim, seqlen assuming mask.shape == bs, 1, 1, seqlen

add options for data loading

add options for model structure

evaluate the model on the out fid for on epoch.

evaluate the model for one epoch.

get the animal name from the file path

get the animal name from the file path

get the end effector velocity from a motion inputs: ------- motion : [bs, num_frames, num_joints + 1, 6]

get the meta dir from the file path

get the meta dir from the file path

make all the directories that could be used in training. this could help with prevent the error of directory not found.

this function is coupled with _add_path function. if _add_path is not used, this function should not be used. make all the directories t

set the random seed for torch and numpy

set the random seed for torch and numpy

recursively load all tensors in a nested dictionary structure onto the gpu.

recursively load all tensors in a nested dictionary structure onto the gpu.

add options for data loading

add some freq related options.

add options for model structure

add options for optimization setting include the total number of steps for training.

add the path for data loading, model training, result saving

check the name of motions generated matches the number of captions.

check the generated motion match the caption

the collate function for a given index

the collate function for a given index

the collate function for a given index

the collate function for a given index

evaluate the model for one epoch.

get the end effector velocity from a motion inputs: ------- motion : [bs, num_frames, num_joints + 1, 6]

take the folder of ood caption as input.

get the end effector velocity from a motion

get the joints xyz positions.

initialize the clip to intermediate

initialize the clip to intermediate

initialize the mapping part

initialize the mapping part

initialize the semantic head

initialize the semantic head

Check if the motion's length is within the valid range.

Check if the motion's length is within the valid range.

process the list of file path

recover the position from the motion. the motion is normalized.

recover the position from the motion. the motion is normalized.

recover the position from the motion. the motion is normalized.

recover the position from the motion. the motion is normalized.

Numpy implementation of the Frechet Distance. The Frechet distance between two multivariate Gaussians X_1 ~ N(mu_1, C_1) and X_2 ~ N(mu_2, C_2

end the timer and print the time

get the poly3d object from vertices and faces.

get dataloader for smal motion and caption dataset.

get dataloader for smpl motion and caption dataset.

initialize the smal model with an initial guess of beta.

recursively load all tensors in a nested dictionary structure onto the gpu.

recursively load all tensors in a nested dictionary structure onto the gpu.

the main function for fit a initial smal model onto target mesh.

calculate the masked l2 loss assuming a.shape == b.shape == bs, J, Jdim, seqlen assuming mask.shape == bs, 1, 1, seqlen

optimize the smal model to fit the target mesh. inputs: ------- num_epoch: int

Compute pairwise Euclidean distances.

start the timer

Take the sum over all non-batch dimensions.

Take the sum over all non-batch dimensions.

Take the sum over all non-batch dimensions.