This version of DistributedDataParallel is designed to be used in conjunction with the multiproc.py launcher included with this example. It a

Given a list of dict, each of which contains torch.Tensor or a list of torch.Tensor, we concat along `dim` and create a dict by concat each o

Clean the meshes in mesh_filenames. Note that the actual cleaning will only be executed by rank=0. The rest of the ranks should wait for

Compute the mean squared error between arr and ref. The average is taken over the d_shape. Args: arr: (*b_shape, *d_shape)

Compute the receptive fields of each layers. Returns: list of the length of each layer's receptive field.

Get a coordinate frame from z and y vector. z will be used directly as the z axis. y will be made orthogonal to y and used as the y axis.

Detach each element in x, regardless if it is a tensor or nested list of tensors.

Drop features related to camera pose used to capture the point

Returns exp(t*S) of a 3*3 skew symmetric matrix to get a rotation matrix.

Return a point cloud object from o3d_pcd Args: o3d_pcd: (n,) Returns: (b=1, n)

Generate uv coordinates ([0, w), [0, h)) of the patches centered at patch_center. Args: patch_center: (*b, 2) the center

Generate 3D point coordinates and related rgb feature Args: rgb_image: (h, w, 3) rgb depth_image: (h, w) depth, along z dire

Create o3d meshes of camera frames

Get the clean the meshes' filenames in mesh_filenames. Args: mesh_filenames: a list containing the meshes to be cleaned.

Given a (*, 3) vector v, return a (*, 3, 3) cross_product matrix [v]_x, such that for all (3,) vector u, we have v * u = [v]_x * u. Vx =

Compute the grid index given xyz_w. Args: points: (*, n, 3) grid_size: (*, 3) long. number of grid ce

Return the initial mean of each gaussian Args: batch_size: Batch size device:

Return V^-1 = I - 0.5 * S + 1/theta^2 * (1 - theta * sin(theta) / (2 * (1 - cos(theta)))) * S^2

get the pointer to the point index set

Returns t * V(t, S), where V(t, S) = I + (1 - cos(t * theta)) / (t * theta)^2 * t * S + (t * theta - sin(t * theta)) / (t * theta)^3

given setting (train/test), generate a list of obj file names in the tex dataset Args: setting: train/test num_classes: numbe

Generate all-zero hidden states that can be used to initialize the model or as a template. Args: batch_size:

insert a point representing inf at n=0

Returns H^-1 given RigidMotion H

Load mesh files as o3d meshes, randomly combine meshes to form more complex meshes. Args: mesh_filenames: list of f

Load mesh files as o3d meshes. Args: mesh_filenames: list of filenames Returns: list of o3d meshes

Return the log(R) where R is a rotation matrix. So it returns a skew symmetric matrix.

Compute the image coordinates of the 3D points in the world. Args: xyz_w: (*b_shape, *n_shape, 3) the points in world co

Render the point cloud using surfel rasterization. Args: camera: camera (b, q) point_size:

Register variable to be loaded during retraining.

Given a mesh file (as ground truth), 1) sample point cloud using RGBD cameras, or directly sample from mesh 2) render the point cloud fro

Render given rgbd images.

Args: bucket_boundaries: int (number of bins) or a list (containing the edges in ascending order, excluding two o

Returns a dictionary that can be saved or load.

Given x y z index, change to the linear index. (matlab's sub2ind) Args: idx: (*, n, 3) long size: (

Set all the nn.modules to train mode.

Uniformly sample more cameras from the current ones. Currently, we do not support gradient (though nothing stops it theoretically).

Unscale the gradient of the parameter controlled by the optimizer. Should only be called once per iteration.

r"""Pad, upsample, FIR filter, and downsample a batch of 2D images. Accepts a batch of 2D images of the shape `[majorDim, minorDim(inC), inH, inW

Args: dim_x: int Dimension of the past input sequence xs dim_c: int Dimension of the

Create Variational RNN. Args: param_graves: Please see the documentation of :py:class:`ParamGraves`.

Content encoder: some conv1D -> bidirectional LSTM Args: dim_input: dimension of the input n

A LSTM with additional linear layers at the top. Args: dim_input: input dimension of the lstm

Construct the sub-networks. Overview of the model: attn_rnn (lstm_cell_layers): input: x_{t-1}, last_attn_context

Construct the sub-networks. Overview of the model: (1) A :py:class:`NetworkGraves` that handles all the rendering, or P(pt | zt,

Create the folders to store codes, checkpoints, etc.

Create a logger that prints to both console and a file. Args: log_file: File name to save the log. None: not save

Divide data into buckets based on seq_lens.

Figure out which device (i.e., gpu or cpu) to use.

Flatten dense tensors into a contiguous 1D buffer. Assume tensors are of same dense type. Since inputs are dense, the resulting tensor will be

Return the values in losses_name in a dictionary.

Return the values in losses_name in a dictionary.

Load depth map (z in camera coordinate (x to right, y to down, z to far)) and surface normal in world coordinate (after bump map).

Create a thread to run tensorboard.

Given camera poses, return the rasterized RGBD images. Args: camera: (b, q) already on device

Read the index file and returns the uids and filenames.

Read the info file if existed, else set everything to None.

Register scalar loss so that it will be gathered by :py:`_gather_loss` Args: var_name: loss name (str) o

Register output so that it will be reset by `reset_outputs` Args: var_name: var name (str) or list of va

Save all the samples of the target_chunk_idx to the memmap. Overview: 1. Read all samples in the target chunk into a list and save t

Send all base model and nn.modules to device.

Set the defaults for attributes.

Manually assign camera eye: list of (3,) where the cameras are (before global transform) up: None (assume to be (0,1

Set all the nn.modules to train mode.

Setup the networks in the model for distributed computation. It automactically gather all modules.

View a flat buffer using the sizes of tensors. Assume that tensors are of same dense type, and that flat is given by _flatten_dense_tensors. A

Write the info to info_filename.

Read the samples from the dataset. Args: dataset: dataset to be added ratio: ratio of

Add multiple scalars. The tag of the scalars will be main_tag/tag. Note: The scalars are plotted in separated figures.

Additional things to do before getting dataloader.

Additional things to do before training starts.

Aggregate a list of PointersectRecord of the same shape. Note that it will set many attributes to None.