Print an info to the shell and the log file.

Return a

Load the model saved at the filenames. It automatically gathers all the nn.modules and optimizers and load them. Everything w

Split camera (b, q) into a list of cameras (b', q'), such that b' * q' * h * w < chunk_size. Note that we only chunk q or ch

Send the model to device.

Returns cameras in the trajactory

Generate camera rays: ray_origin is at pinhole and ray directions outward from a pixel location (somewhere withing a pixel pi

Reproject RGBD pixels to world coordinate. Args: subsample: use 1 out of every `subsample` pixels

Get all parameters from individual torch.nn.Module

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

associate pidx with gidx

Return a pytorch dataset instance, which contains (total_samples, dim). :param val: constant fill in value. None: random array

Register a figure to the tensorboard. See tensorboard's documentation https://pytorch.org/docs/stable/tensorboard.html for usage.

Set the epoch number. Should be called at the beginning of every epoch, before iterator is constructed.

Args: arr: (b, n, *) bidx: batch index Returns: (n, dim)

Return a list of b o3d pcds containing xyz and rgb.

returns the 4*4 homogeneous matrix.

Load a pretrained model given a state dictionary.

Returns a dictionary that can be saved or load.

Create memmaps that store chunked of the samples. It will overwrite the working_dir. Args: samples: a sequence of

Set all the nn.modules to evaluation mode.

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

Args: type: "depth_meters": Euclidean distances (in meters) to the optical center of the camera (ray_t)

Make sure the var will be saved in the checkpoint. Note that it is just for recording purpose. They will NOT be loaded during

test plan: - given created random datasets of different sizes - create batch_preparer of different dataset ratios - c

Register a scalar. Args: tag: Name of the scalar. Can be grouped by adding '/' in it. scalar_value:

Check all attributes have the same number of points.

Flush the buffer to console and tensorboard. Note: It uses the latest epoch, global_step, and batch_idx, so it should be called e

Returns extrinsic matrices (inverse of H_c2w), shape: (b, q, 4, 4).

Return the number of pixels in each image.

Intersect the mesh with rays to get ground truth Args: ray: (b, *m_shape) Returns:

The main process.

Update parameters and learning rate.

Voxel downsampling uses a voxel grid to uniformly downsample the input point cloud. Procedure: - Points are discretized into

Args: type: "color": color image before any tone mapping has been applied "diffuse_illumination

Register variable to be loaded during retraining.

One training step. Return a dictionary that will be passed to logging.

Args: epoch: bidx: batch: input_rgbd_images: RGBDImage, (b, q, h, w),

Visualize the results computed in _step.

Close the logger and terminates the tensorboard.

Compute the distance between each point to each ray. Args: points: (*, n, 3) ray_origins: (*, m, 3)

Compute the grid index given xyz_w. Args: points: (*, n, 3) grid_size: (*, 3) long. number of grid cells in x y z. center: (*, 3) center of the grid

get the underlying table

Return the total number of samples (multiplied with their ratios).

returns the inverse of a homegeneous matrix.

Load the config file, which is a dictionary of: var_name -> Dict[key, val]. Each of the key in var_name will replace the current valu

Main rendering function that calls different methods.

Print a 1D array. The function automatically breaks lines when printing a long 1D array. Args: arr: The 1D array to prin

Make sure the var will be saved in the state_dict. Note that it is just for recording purpose. They will NOT be loaded during

Set the losses to None.

Set the outputs of the model to None.

r"""One-dimensional upsample, fir filtering, downsample. Args: input (batch, cin, seq_len): input tensor kernel (kern

Zero the gradient buffer.

Create a linear layer whose weights are initialized with a chosen method. The usage of the layer is the same as torch.nn.Linear.

r""" Args: kernel (list of float): 1D blur_kernel. pad (tuple of int): (pad_left, pad_

Args: in_features (int): input feature dimension (number of columns) out_features (int):

Create a one-dimensional convolution layer whose weights are initialized with a chosen method. The usage of the layer is the same as

Construct the batches.

Determine the filenames of the memmaps.

Register the var_name to the target (set of var names) Args: var_name: name (str) or list of names of th

Set all the nn.modules to evaluation mode.