render.py

#
# Copyright (C) 2023, Inria
# GRAPHDECO research group, https://team.inria.fr/graphdeco
# All rights reserved.
#
# This software is free for non-commercial, research and evaluation use 
# under the terms of the LICENSE.md file.
#
# For inquiries contact  george.drettakis@inria.fr
#
import numpy as np
import torch
from scene import Scene
import os
from tqdm import tqdm
from os import makedirs
from gaussian_renderer import render
import torchvision
from utils.general_utils import safe_state
from argparse import ArgumentParser
from arguments import ModelParams, PipelineParams, get_combined_args
from gaussian_renderer import GaussianModel
from autoencoder.model import Autoencoder
from eval.openclip_encoder import OpenCLIPNetwork
import cv2

def render_set(model_path, source_path, name, iteration, views, gaussians, pipeline, background, args, ae_model, clip_model):
    render_path = os.path.join(model_path, name, "ours_{}".format(iteration), "renders")
    gts_path = os.path.join(model_path, name, "ours_{}".format(iteration), "gt")
    render_npy_path = os.path.join(model_path, name, "ours_{}".format(iteration), "renders_npy")
    gts_npy_path = os.path.join(model_path, name, "ours_{}".format(iteration), "gt_npy")

    makedirs(render_npy_path, exist_ok=True)
    makedirs(gts_npy_path, exist_ok=True)
    makedirs(render_path, exist_ok=True)
    makedirs(gts_path, exist_ok=True)

    for idx, view in enumerate(tqdm(views, desc="Rendering progress")):
        output = render(view, gaussians, pipeline, background, args, ae_model=ae_model, clip_model=clip_model)
        if idx == 0: continue
        if not args.include_feature:
            rendering = output["render"]
        else:
            rendering = output["language_feature_image"]
            
        if not args.include_feature:
            gt = view.original_image[0:3, :, :]
            
        else:
            gt, mask = view.get_language_feature(os.path.join(source_path, args.language_features_name), feature_level=args.feature_level)
        
        # if clip_model is not None and ae_model is not None:
        #     language_image = rendering.clone().permute(1, 2, 0).unsqueeze(0)
        #     lvl, h, w, _ = language_image.shape
        #     restored_feat = ae_model.decode(language_image.flatten(0, 2))
        #     restored_feat = restored_feat.view(lvl, h, w, -1)
        #     relevancy_map = clip_model.get_max_across(restored_feat) #.sum(dim=0)
        #     n_head, n_prompt, h, w = relevancy_map.shape
    
        #     for i in range(relevancy_map.shape[1]):
        #         scale = 30
        #         kernel = np.ones((scale,scale)) / (scale**2)
        #         np_relev = relevancy_map[:, i].cpu().numpy()
        #         avg_filtered = cv2.filter2D(np_relev.transpose(1,2,0), -1, kernel)
        #         h, w = avg_filtered.shape
        #         avg_filtered = avg_filtered.reshape(h, w, 1)        
        #         score = avg_filtered[..., 0].max()
        #         coord = np.nonzero(avg_filtered[..., 0] == score)
        #         cx, cy = round(coord[1][0]), round(coord[0][0])
        #         # breakpoint()
        #         # score_lvl[0] = score
        #         width = 20
        #         x1 , y1 , x2, y2 = round(max(cx - width, 0)), round(max(cy - width, 0)), round(min(w, cx + width)), round(min(h, cy + width))
        #         rendering[:, y1:y2, x1:x2] = 1.0
        #         break
        
        np.save(os.path.join(render_npy_path, '{0:05d}'.format(idx) + ".npy"),rendering.permute(1,2,0).cpu().numpy())
        np.save(os.path.join(gts_npy_path, '{0:05d}'.format(idx) + ".npy"),gt.permute(1,2,0).cpu().numpy())
        torchvision.utils.save_image(rendering, os.path.join(render_path, '{0:05d}'.format(idx) + ".png"))
        torchvision.utils.save_image(gt, os.path.join(gts_path, '{0:05d}'.format(idx) + ".png"))
        breakpoint()
        break
               
def render_sets(dataset : ModelParams, iteration : int, pipeline : PipelineParams, skip_train : bool, skip_test : bool, args):
    with torch.no_grad():
        gaussians = GaussianModel(dataset.sh_degree)
        scene = Scene(dataset, gaussians, shuffle=False)
        checkpoint = os.path.join(args.model_path, 'chkpnt30000.pth')
        (model_params, first_iter) = torch.load(checkpoint)
        gaussians.restore(model_params, args, mode='test')

        ae_model = Autoencoder([256, 128, 64, 32, 3], [16, 32, 64, 128, 256, 256, 512]).to("cuda")
        ae_model.load_state_dict(torch.load('autoencoder/ckpt/office_scene_50/best_ckpt.pth', map_location='cuda'))
        ae_model.eval()

        clip_model = OpenCLIPNetwork("cuda")
        clip_model.set_positives(["bottle", "sanitizer", "tv", "screen", "television", "chair"])
        
        bg_color = [1,1,1] if dataset.white_background else [0, 0, 0]
        background = torch.tensor(bg_color, dtype=torch.float32, device="cuda")

        if not skip_train:
             render_set(dataset.model_path, dataset.source_path, "train", scene.loaded_iter, scene.getTrainCameras(), gaussians, pipeline, background, args, ae_model, clip_model)

        if not skip_test:
             render_set(dataset.model_path, dataset.source_path, "test", scene.loaded_iter, scene.getTestCameras(), gaussians, pipeline, background, args, ae_model, clip_model)

if __name__ == "__main__":
    # Set up command line argument parser
    
    parser = ArgumentParser(description="Testing script parameters")
    model = ModelParams(parser, sentinel=True)
    pipeline = PipelineParams(parser)
    parser.add_argument("--iteration", default=-1, type=int)
    parser.add_argument("--skip_train", action="store_true")
    parser.add_argument("--skip_test", action="store_true")
    parser.add_argument("--quiet", action="store_true")
    parser.add_argument("--include_feature", action="store_true")

    args = get_combined_args(parser)
    print("Rendering " + args.model_path)

    safe_state(args.quiet)

    render_sets(model.extract(args), args.iteration, pipeline.extract(args), args.skip_train, args.skip_test, args)