gen_samples.py

''' Generate images and shapes using pretrained network pickle.
Code adapted from following paper
"Efficient Geometry-aware 3D Generative Adversarial Networks."
See LICENSES/LICENSE_EG3D for original license.
'''

import os
import re
from typing import List, Optional, Tuple, Union

import click
import dnnlib
import numpy as np
import PIL.Image
import torch
from tqdm import tqdm
import mrcfile


import legacy
from camera_utils import LookAtPoseSampler, FOV_to_intrinsics
from torch_utils import misc
from training.triplane import TriPlaneGenerator


#----------------------------------------------------------------------------

def parse_range(s: Union[str, List]) -> List[int]:
    '''Parse a comma separated list of numbers or ranges and return a list of ints.

    Example: '1,2,5-10' returns [1, 2, 5, 6, 7]
    '''
    if isinstance(s, list): return s
    ranges = []
    range_re = re.compile(r'^(\d+)-(\d+)$')
    for p in s.split(','):
        m = range_re.match(p)
        if m:
            ranges.extend(range(int(m.group(1)), int(m.group(2))+1))
        else:
            ranges.append(int(p))
    return ranges

#----------------------------------------------------------------------------

def parse_vec2(s: Union[str, Tuple[float, float]]) -> Tuple[float, float]:
    '''Parse a floating point 2-vector of syntax 'a,b'.

    Example:
        '0,1' returns (0,1)
    '''
    if isinstance(s, tuple): return s
    parts = s.split(',')
    if len(parts) == 2:
        return (float(parts[0]), float(parts[1]))
    raise ValueError(f'cannot parse 2-vector {s}')

#----------------------------------------------------------------------------

def make_transform(translate: Tuple[float,float], angle: float):
    m = np.eye(3)
    s = np.sin(angle/360.0*np.pi*2)
    c = np.cos(angle/360.0*np.pi*2)
    m[0][0] = c
    m[0][1] = s
    m[0][2] = translate[0]
    m[1][0] = -s
    m[1][1] = c
    m[1][2] = translate[1]
    return m

#----------------------------------------------------------------------------

def create_samples(N=256, voxel_origin=[0, 0, 0], cube_length=2.0):
    # NOTE: the voxel_origin is actually the (bottom, left, down) corner, not the middle
    voxel_origin = np.array(voxel_origin) - cube_length/2
    voxel_size = cube_length / (N - 1)

    overall_index = torch.arange(0, N ** 3, 1, out=torch.LongTensor())
    samples = torch.zeros(N ** 3, 3)

    # transform first 3 columns
    # to be the x, y, z index
    samples[:, 2] = overall_index % N
    samples[:, 1] = (overall_index.float() / N) % N
    samples[:, 0] = ((overall_index.float() / N) / N) % N

    # transform first 3 columns
    # to be the x, y, z coordinate
    samples[:, 0] = (samples[:, 0] * voxel_size) + voxel_origin[2]
    samples[:, 1] = (samples[:, 1] * voxel_size) + voxel_origin[1]
    samples[:, 2] = (samples[:, 2] * voxel_size) + voxel_origin[0]

    num_samples = N ** 3

    return samples.unsqueeze(0), voxel_origin, voxel_size

#----------------------------------------------------------------------------

@click.command()
@click.option('--network', 'network_pkl', help='Network pickle filename', required=True)
@click.option('--seeds', type=parse_range, help='List of random seeds (e.g., \'0,1,4-6\')', required=True)
@click.option('--trunc', 'truncation_psi', type=float, help='Truncation psi', default=0.7, show_default=True)
@click.option('--trunc-cutoff', 'truncation_cutoff', type=int, help='Truncation cutoff', default=14, show_default=True)
@click.option('--class', 'class_idx', type=int, help='Class label (unconditional if not specified)')
@click.option('--outdir', help='Where to save the output images', type=str, required=True, metavar='DIR')
@click.option('--shapes', help='Export shapes as .mrc files viewable in ChimeraX', type=bool, required=False, metavar='BOOL', default=False, show_default=True)
@click.option('--shape-res', help='', type=int, required=False, metavar='int', default=512, show_default=True)
@click.option('--fov-deg', help='Field of View of camera in degrees', type=int, required=False, metavar='float', default=18.837, show_default=True)
@click.option('--shape-format', help='Shape Format', type=click.Choice(['.mrc', '.ply']), default='.mrc')
@click.option('--reload_modules', help='Overload persistent modules?', type=bool, required=False, metavar='BOOL', default=False, show_default=True)
@click.option('--pose_cond', type=int, help='camera conditioned pose angle', default=90, show_default=True)

def generate_images(
    network_pkl: str,
    seeds: List[int],
    truncation_psi: float,
    truncation_cutoff: int,
    outdir: str,
    shapes: bool,
    shape_res: int,
    fov_deg: float,
    shape_format: str,
    class_idx: Optional[int],
    reload_modules: bool,
    pose_cond: int,
):
    """Generate images using pretrained network pickle.

    Examples:

    \b
    # Generate an image using pre-trained model.
    python gen_samples.py --outdir=out --trunc=0.7 --shapes=true --seeds=0-3 \
        --network models/easy-khair-180-gpc0.8-trans10-025000.pkl
    """

    print('Loading networks from "%s"...' % network_pkl)
    device = torch.device('cuda:1')
    with dnnlib.util.open_url(network_pkl) as f:
        G = legacy.load_network_pkl(f)['G_ema'].to(device) # type: ignore

    # Specify reload_modules=True if you want code modifications to take effect; otherwise uses pickled code
    if reload_modules:
        print("Reloading Modules!")
        G_new = TriPlaneGenerator(*G.init_args, **G.init_kwargs).eval().requires_grad_(False).to(device)
        misc.copy_params_and_buffers(G, G_new, require_all=True)
        G_new.neural_rendering_resolution = G.neural_rendering_resolution
        G_new.rendering_kwargs = G.rendering_kwargs
        G = G_new

    network_pkl = os.path.basename(network_pkl)
    outdir = os.path.join(outdir, os.path.splitext(network_pkl)[0] + '_' + str(pose_cond))
    os.makedirs(outdir, exist_ok=True)

    pose_cond_rad = pose_cond/180*np.pi

    intrinsics = FOV_to_intrinsics(fov_deg, device=device)

    # Generate images.
    for seed_idx, seed in enumerate(seeds):
        print('Generating image for seed %d (%d/%d) ...' % (seed, seed_idx, len(seeds)))
        # cond camera settings
        cam_pivot = torch.tensor([0, 0, 0], device=device)
        cam_radius = G.rendering_kwargs.get('avg_camera_radius', 2.7)
        conditioning_cam2world_pose = LookAtPoseSampler.sample(pose_cond_rad, np.pi/2, cam_pivot, radius=cam_radius, device=device)
        conditioning_params = torch.cat([conditioning_cam2world_pose.reshape(-1, 16), intrinsics.reshape(-1, 9)], 1)
        
        # z and w
        z = torch.from_numpy(np.random.RandomState(seed).randn(1, G.z_dim)).to(device)
        
        # z1 = torch.from_numpy(np.random.RandomState(44).randn(1, G.z_dim)).to(device)
        # ws_list = []
        # ws_list.append(G.mapping(torch.from_numpy(np.random.RandomState(0).randn(1, G.z_dim)).to(device), conditioning_params, truncation_psi=truncation_psi, truncation_cutoff=truncation_cutoff))
        # ws_list.append(G.mapping(torch.from_numpy(np.random.RandomState(0).randn(1, G.z_dim)).to(device), conditioning_params, truncation_psi=truncation_psi, truncation_cutoff=truncation_cutoff))
        # ws_list.append(G.mapping(z1, conditioning_params, truncation_psi=truncation_psi, truncation_cutoff=truncation_cutoff))
 
        imgs = []
        angle_p = -0.2
        # for angle_y, angle_p in [(2.1, angle_p), (1.05, angle_p), (0, angle_p), (-1.05, angle_p), (-2.1, angle_p)]:
        for idx, angles in enumerate([(0, angle_p), (-45/180*np.pi, angle_p), (-90/180*np.pi, angle_p), (-135/180*np.pi, angle_p), (-np.pi, angle_p)]):
            angle_y = angles[0]
            angle_p = angles[1]
            # rand camera setting
            cam2world_pose = LookAtPoseSampler.sample(np.pi/2 + angle_y, np.pi/2 + angle_p, cam_pivot, radius=cam_radius, device=device)
            camera_params = torch.cat([cam2world_pose.reshape(-1, 16), intrinsics.reshape(-1, 9)], 1)

            ws = G.mapping(z, conditioning_params, truncation_psi=truncation_psi, truncation_cutoff=truncation_cutoff)
            
            # img = G.synthesis(ws, camera_params, ws_bcg = ws_list[idx])['image']
            img = G.synthesis(ws, camera_params)['image']
            img = (img.permute(0, 2, 3, 1) * 127.5 + 128).clamp(0, 255).to(torch.uint8)
            imgs.append(img)

        img = torch.cat(imgs, dim=2)

        PIL.Image.fromarray(img[0].cpu().numpy(), 'RGB').save(f'{outdir}/seed{seed:04d}.png')

        if shapes:
            # extract a shape.mrc with marching cubes. You can view the .mrc file using ChimeraX from UCSF.
            max_batch=1000000

            samples, voxel_origin, voxel_size = create_samples(N=shape_res, voxel_origin=[0, 0, 0], cube_length=G.rendering_kwargs['box_warp'] * 1)#.reshape(1, -1, 3)
            samples = samples.to(z.device)
            sigmas = torch.zeros((samples.shape[0], samples.shape[1], 1), device=z.device)
            transformed_ray_directions_expanded = torch.zeros((samples.shape[0], max_batch, 3), device=z.device)
            transformed_ray_directions_expanded[..., -1] = -1

            head = 0
            with tqdm(total = samples.shape[1]) as pbar:
                with torch.no_grad():
                    while head < samples.shape[1]:
                        torch.manual_seed(0)
                        sigma = G.sample(samples[:, head:head+max_batch], transformed_ray_directions_expanded[:, :samples.shape[1]-head], z, conditioning_params, truncation_psi=truncation_psi, truncation_cutoff=truncation_cutoff, noise_mode='const')['sigma']
                        sigmas[:, head:head+max_batch] = sigma
                        head += max_batch
                        pbar.update(max_batch)

            sigmas = sigmas.reshape((shape_res, shape_res, shape_res)).cpu().numpy()
            sigmas = np.flip(sigmas, 0)

            # Trim the border of the extracted cube
            pad = int(30 * shape_res / 256)
            pad_value = -1000
            sigmas[:pad] = pad_value
            sigmas[-pad:] = pad_value
            sigmas[:, :pad] = pad_value
            sigmas[:, -pad:] = pad_value
            sigmas[:, :, :pad] = pad_value
            sigmas[:, :, -pad:] = pad_value

            if shape_format == '.ply':
                from shape_utils import convert_sdf_samples_to_ply
                convert_sdf_samples_to_ply(np.transpose(sigmas, (2, 1, 0)), [0, 0, 0], 1, os.path.join(outdir, f'seed{seed:04d}.ply'), level=10)
            elif shape_format == '.mrc': # output mrc
                with mrcfile.new_mmap(os.path.join(outdir, f'seed{seed:04d}.mrc'), overwrite=True, shape=sigmas.shape, mrc_mode=2) as mrc:
                    mrc.data[:] = sigmas


#----------------------------------------------------------------------------

if __name__ == "__main__":
    generate_images() # pylint: disable=no-value-for-parameter

#----------------------------------------------------------------------------