From 2982236cb8aa26f6cd42c17130a60fb4b4d732d0 Mon Sep 17 00:00:00 2001 From: Christian Rauch Date: Fri, 11 Apr 2025 15:24:51 +0200 Subject: [PATCH 1/6] update package dependencies --- environment.yml | 14 +++++++------- 1 file changed, 7 insertions(+), 7 deletions(-) diff --git a/environment.yml b/environment.yml index 286d6b8..a1dcb09 100644 --- a/environment.yml +++ b/environment.yml @@ -5,13 +5,12 @@ channels: - defaults dependencies: - open-clip-torch - - cudatoolkit=11.6 - - plyfile=0.8.1 - - python=3.7.13 - - pip=22.3.1 - - pytorch=1.12.1 - - torchaudio=0.12.1 - - torchvision=0.13.1 + - plyfile + - python + - pip + - pytorch + - torchaudio + - torchvision - tqdm - opencv - tensorboard @@ -23,3 +22,4 @@ dependencies: - submodules/segment-anything-langsplat - submodules/langsplat-rasterization - submodules/simple-knn + - nvidia-cuda-runtime-cu12 From b425665837a3d2305da9050c6a9e7aec7a6303f0 Mon Sep 17 00:00:00 2001 From: Christian Rauch Date: Fri, 11 Apr 2025 15:27:36 +0200 Subject: [PATCH 2/6] update simple-knn from upstream --- .gitmodules | 2 +- submodules/simple-knn | 2 +- 2 files changed, 2 insertions(+), 2 deletions(-) diff --git a/.gitmodules b/.gitmodules index fbc9382..cc1aa3e 100644 --- a/.gitmodules +++ b/.gitmodules @@ -3,7 +3,7 @@ url = https://github.com/minghanqin/segment-anything-langsplat [submodule "submodules/simple-knn"] path = submodules/simple-knn - url = https://gitlab.inria.fr/bkerbl/simple-knn.git + url = https://github.com/camenduru/simple-knn.git [submodule "submodules/langsplat-rasterization"] path = submodules/langsplat-rasterization url = https://github.com/minghanqin/langsplat-rasterization diff --git a/submodules/simple-knn b/submodules/simple-knn index 44f7642..a2a3ef4 160000 --- a/submodules/simple-knn +++ b/submodules/simple-knn @@ -1 +1 @@ -Subproject commit 44f764299fa305faf6ec5ebd99939e0508331503 +Subproject commit a2a3ef44fa0b9f7b2b5145bd9610029dbec13aa2 From 31898a918d1a05c27dab3f713e5fec797d5b42ac Mon Sep 17 00:00:00 2001 From: Christian Rauch Date: Fri, 11 Apr 2025 15:27:58 +0200 Subject: [PATCH 3/6] fix training script --- process.sh | 10 ++++++---- 1 file changed, 6 insertions(+), 4 deletions(-) mode change 100644 => 100755 process.sh diff --git a/process.sh b/process.sh old mode 100644 new mode 100755 index cf7eefe..ce726f2 --- a/process.sh +++ b/process.sh @@ -1,24 +1,26 @@ #!/bin/bash # get the language feature of the scene -python preprocess.py --dataset_name $dataset_path +python preprocess.py --dataset_path $dataset_path # train the autoencoder cd autoencoder -python train.py --dataset_path $dataset_path --encoder_dims 256 128 64 32 3 --decoder_dims 16 32 64 128 256 256 512 --lr 0.0007 --dataset_name ae_ckpt +python train.py --dataset_path $dataset_path --encoder_dims 256 128 64 32 3 --decoder_dims 16 32 64 128 256 256 512 --lr 0.0007 --dataset_name $dataset_name # e.g. python train.py --dataset_path ../data/sofa --encoder_dims 256 128 64 32 3 --decoder_dims 16 32 64 128 256 256 512 --lr 0.0007 --dataset_name sofa # get the 3-dims language feature of the scene -python test.py --dataset_name $dataset_path --dataset_name $dataset_name +python test.py --dataset_path $dataset_path --dataset_name $dataset_name # e.g. python test.py --dataset_path ../data/sofa --dataset_name sofa +cd .. + # ATTENTION: Before you train the LangSplat, please follow https://github.com/graphdeco-inria/gaussian-splatting # to train the RGB 3D Gaussian Splatting model. # put the path of your RGB model after '--start_checkpoint' for level in 1 2 3 do - python train.py -s $dataset_path -m output/${casename} --start_checkpoint $dataset_path/$casename/chkpnt30000.pth --feature_level ${level} + python train.py -s $dataset_path -m output/${casename} --start_checkpoint $dataset_path/output/$casename/chkpnt30000.pth --feature_level ${level} # e.g. python train.py -s data/sofa -m output/sofa --start_checkpoint data/sofa/sofa/chkpnt30000.pth --feature_level 3 done From b51d7bc4268073d3566d749473c0c6c2cd2ae701 Mon Sep 17 00:00:00 2001 From: Christian Rauch Date: Fri, 11 Apr 2025 15:29:28 +0200 Subject: [PATCH 4/6] set script fail on any command failure --- process.sh | 2 ++ 1 file changed, 2 insertions(+) diff --git a/process.sh b/process.sh index ce726f2..f4ad78d 100755 --- a/process.sh +++ b/process.sh @@ -1,5 +1,7 @@ #!/bin/bash +set -e + # get the language feature of the scene python preprocess.py --dataset_path $dataset_path From abd87884661175d0767e80a8e2c881ed808278ba Mon Sep 17 00:00:00 2001 From: Christian Rauch Date: Fri, 11 Apr 2025 17:18:59 +0200 Subject: [PATCH 5/6] load checkpoint with "weights_only=False" --- autoencoder/test.py | 2 +- eval/evaluate_iou_loc.py | 2 +- render.py | 2 +- train.py | 2 +- 4 files changed, 4 insertions(+), 4 deletions(-) diff --git a/autoencoder/test.py b/autoencoder/test.py index 33ffcea..69259a5 100644 --- a/autoencoder/test.py +++ b/autoencoder/test.py @@ -42,7 +42,7 @@ shutil.copy(source_path, target_path) - checkpoint = torch.load(ckpt_path) + checkpoint = torch.load(ckpt_path, weights_only=False) train_dataset = Autoencoder_dataset(data_dir) test_loader = DataLoader( diff --git a/eval/evaluate_iou_loc.py b/eval/evaluate_iou_loc.py index e9d50f8..9ff5dee 100644 --- a/eval/evaluate_iou_loc.py +++ b/eval/evaluate_iou_loc.py @@ -238,7 +238,7 @@ def evaluate(feat_dir, output_path, ae_ckpt_path, json_folder, mask_thresh, enco # instantiate autoencoder and openclip clip_model = OpenCLIPNetwork(device) - checkpoint = torch.load(ae_ckpt_path, map_location=device) + checkpoint = torch.load(ae_ckpt_path, map_location=device, weights_only=False) model = Autoencoder(encoder_hidden_dims, decoder_hidden_dims).to(device) model.load_state_dict(checkpoint) model.eval() diff --git a/render.py b/render.py index 97f0be2..9d3cdde 100644 --- a/render.py +++ b/render.py @@ -56,7 +56,7 @@ def render_sets(dataset : ModelParams, iteration : int, pipeline : PipelineParam 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) + (model_params, first_iter) = torch.load(checkpoint, weights_only=False) gaussians.restore(model_params, args, mode='test') bg_color = [1,1,1] if dataset.white_background else [0, 0, 0] diff --git a/train.py b/train.py index b6f6a4e..d84c883 100644 --- a/train.py +++ b/train.py @@ -42,7 +42,7 @@ def training(dataset, opt, pipe, testing_iterations, saving_iterations, checkpoi if not checkpoint: raise ValueError("checkpoint missing!!!!!") if checkpoint: - (model_params, first_iter) = torch.load(checkpoint) + (model_params, first_iter) = torch.load(checkpoint, weights_only=False) if len(model_params) == 12 and opt.include_feature: first_iter = 0 gaussians.restore(model_params, opt) From 8bc930792d5d4b6cf4b3c96e3946bda2593e8522 Mon Sep 17 00:00:00 2001 From: Christian Rauch Date: Fri, 11 Apr 2025 17:19:36 +0200 Subject: [PATCH 6/6] remove trailing whitespace --- autoencoder/test.py | 16 ++++++------ eval/evaluate_iou_loc.py | 54 ++++++++++++++++++++-------------------- render.py | 12 ++++----- train.py | 28 ++++++++++----------- 4 files changed, 55 insertions(+), 55 deletions(-) diff --git a/autoencoder/test.py b/autoencoder/test.py index 69259a5..54a495f 100644 --- a/autoencoder/test.py +++ b/autoencoder/test.py @@ -23,7 +23,7 @@ default=[16, 32, 64, 128, 256, 256, 512], ) args = parser.parse_args() - + dataset_name = args.dataset_name encoder_hidden_dims = args.encoder_dims decoder_hidden_dims = args.decoder_dims @@ -33,7 +33,7 @@ data_dir = f"{dataset_path}/language_features" output_dir = f"{dataset_path}/language_features_dim3" os.makedirs(output_dir, exist_ok=True) - + # copy the segmentation map for filename in os.listdir(data_dir): if filename.endswith("_s.npy"): @@ -46,11 +46,11 @@ train_dataset = Autoencoder_dataset(data_dir) test_loader = DataLoader( - dataset=train_dataset, + dataset=train_dataset, batch_size=256, - shuffle=False, - num_workers=16, - drop_last=False + shuffle=False, + num_workers=16, + drop_last=False ) @@ -62,7 +62,7 @@ for idx, feature in tqdm(enumerate(test_loader)): data = feature.to("cuda:0") with torch.no_grad(): - outputs = model.encode(data).to("cpu").numpy() + outputs = model.encode(data).to("cpu").numpy() if idx == 0: features = outputs else: @@ -70,7 +70,7 @@ os.makedirs(output_dir, exist_ok=True) start = 0 - + for k,v in train_dataset.data_dic.items(): path = os.path.join(output_dir, k) np.save(path, features[start:start+v]) diff --git a/eval/evaluate_iou_loc.py b/eval/evaluate_iou_loc.py index 9ff5dee..7f34e24 100644 --- a/eval/evaluate_iou_loc.py +++ b/eval/evaluate_iou_loc.py @@ -63,9 +63,9 @@ def eval_gt_lerfdata(json_folder: Union[str, Path] = None, ouput_path: Path = No img_ann = defaultdict(dict) with open(js_path, 'r') as f: gt_data = json.load(f) - + h, w = gt_data['info']['height'], gt_data['info']['width'] - idx = int(gt_data['info']['name'].split('_')[-1].split('.jpg')[0]) - 1 + idx = int(gt_data['info']['name'].split('_')[-1].split('.jpg')[0]) - 1 for prompt_data in gt_data["objects"]: label = prompt_data['category'] box = np.asarray(prompt_data['bbox']).reshape(-1) # x1y1x2y2 @@ -77,7 +77,7 @@ def eval_gt_lerfdata(json_folder: Union[str, Path] = None, ouput_path: Path = No else: img_ann[label]['bboxes'] = box img_ann[label]['mask'] = mask - + # # save for visulsization save_path = ouput_path / 'gt' / gt_data['info']['name'].split('.jpg')[0] / f'{label}.jpg' save_path.parent.mkdir(exist_ok=True, parents=True) @@ -87,12 +87,12 @@ def eval_gt_lerfdata(json_folder: Union[str, Path] = None, ouput_path: Path = No return gt_ann, (h, w), img_paths -def activate_stream(sem_map, - image, - clip_model, +def activate_stream(sem_map, + image, + clip_model, image_name: Path = None, - img_ann: Dict = None, - thresh : float = 0.5, + img_ann: Dict = None, + thresh : float = 0.5, colormap_options = None): valid_map = clip_model.get_max_across(sem_map) # 3xkx832x1264 n_head, n_prompt, h, w = valid_map.shape @@ -110,12 +110,12 @@ def activate_stream(sem_map, avg_filtered = cv2.filter2D(np_relev, -1, kernel) avg_filtered = torch.from_numpy(avg_filtered).to(valid_map.device) valid_map[i][k] = 0.5 * (avg_filtered + valid_map[i][k]) - + output_path_relev = image_name / 'heatmap' / f'{clip_model.positives[k]}_{i}' output_path_relev.parent.mkdir(exist_ok=True, parents=True) colormap_saving(valid_map[i][k].unsqueeze(-1), colormap_options, output_path_relev) - + # NOTE 与lerf一致，激活值低于0.5的认为是背景 p_i = torch.clip(valid_map[i][k] - 0.5, 0, 1).unsqueeze(-1) valid_composited = colormaps.apply_colormap(p_i / (p_i.max() + 1e-6), colormaps.ColormapOptions("turbo")) @@ -124,7 +124,7 @@ def activate_stream(sem_map, output_path_compo = image_name / 'composited' / f'{clip_model.positives[k]}_{i}' output_path_compo.parent.mkdir(exist_ok=True, parents=True) colormap_saving(valid_composited, colormap_options, output_path_compo) - + # truncate the heatmap into mask output = valid_map[i][k] output = output - torch.min(output) @@ -136,7 +136,7 @@ def activate_stream(sem_map, mask_pred = smooth(mask_pred) mask_lvl[i] = mask_pred mask_gt = img_ann[clip_model.positives[k]]['mask'].astype(np.uint8) - + # calculate iou intersection = np.sum(np.logical_and(mask_gt, mask_pred)) union = np.sum(np.logical_or(mask_gt, mask_pred)) @@ -148,10 +148,10 @@ def activate_stream(sem_map, score = valid_map[i, k].max() score_lvl[i] = score chosen_lvl = torch.argmax(score_lvl) - + chosen_iou_list.append(iou_lvl[chosen_lvl]) chosen_lvl_list.append(chosen_lvl.cpu().numpy()) - + # save for visulsization save_path = image_name / f'chosen_{clip_model.positives[k]}.png' vis_mask_save(mask_lvl[chosen_lvl], save_path) @@ -165,19 +165,19 @@ def lerf_localization(sem_map, image, clip_model, image_name, img_ann): valid_map = clip_model.get_max_across(sem_map) # 3xkx832x1264 n_head, n_prompt, h, w = valid_map.shape - + # positive prompts acc_num = 0 positives = list(img_ann.keys()) for k in range(len(positives)): select_output = valid_map[:, k] - + # NOTE 平滑后的激活值图中找最大值点 scale = 30 kernel = np.ones((scale,scale)) / (scale**2) np_relev = select_output.cpu().numpy() avg_filtered = cv2.filter2D(np_relev.transpose(1,2,0), -1, kernel) - + score_lvl = np.zeros((n_head,)) coord_lvl = [] for i in range(n_head): @@ -188,21 +188,21 @@ def lerf_localization(sem_map, image, clip_model, image_name, img_ann): selec_head = np.argmax(score_lvl) coord_final = coord_lvl[selec_head] - + for box in img_ann[positives[k]]['bboxes'].reshape(-1, 4): flag = 0 x1, y1, x2, y2 = box x_min, x_max = min(x1, x2), max(x1, x2) y_min, y_max = min(y1, y2), max(y1, y2) for cord_list in coord_final: - if (cord_list[0] >= x_min and cord_list[0] <= x_max and + if (cord_list[0] >= x_min and cord_list[0] <= x_max and cord_list[1] >= y_min and cord_list[1] <= y_max): acc_num += 1 flag = 1 break if flag != 0: break - + # NOTE 将平均后的结果与原结果相加，抑制噪声并保持激活边界清晰 avg_filtered = torch.from_numpy(avg_filtered[..., selec_head]).unsqueeze(-1).to(select_output.device) torch_relev = 0.5 * (avg_filtered + select_output[selec_head].unsqueeze(-1)) @@ -210,7 +210,7 @@ def lerf_localization(sem_map, image, clip_model, image_name, img_ann): valid_composited = colormaps.apply_colormap(p_i / (p_i.max() + 1e-6), colormaps.ColormapOptions("turbo")) mask = (torch_relev < 0.5).squeeze() valid_composited[mask, :] = image[mask, :] * 0.3 - + save_path = output_path_loca / f"{positives[k]}.png" show_result(valid_composited.cpu().numpy(), coord_final, img_ann[positives[k]]['bboxes'], save_path) @@ -248,7 +248,7 @@ def evaluate(feat_dir, output_path, ae_ckpt_path, json_folder, mask_thresh, enco for j, idx in enumerate(tqdm(eval_index_list)): image_name = Path(output_path) / f'{idx+1:0>5}' image_name.mkdir(exist_ok=True, parents=True) - + sem_feat = compressed_sem_feats[:, j, ...] sem_feat = torch.from_numpy(sem_feat).float().to(device) rgb_img = cv2.imread(image_paths[j])[..., ::-1] @@ -259,10 +259,10 @@ def evaluate(feat_dir, output_path, ae_ckpt_path, json_folder, mask_thresh, enco lvl, h, w, _ = sem_feat.shape restored_feat = model.decode(sem_feat.flatten(0, 2)) restored_feat = restored_feat.view(lvl, h, w, -1) # 3x832x1264x512 - + img_ann = gt_ann[f'{idx}'] clip_model.set_positives(list(img_ann.keys())) - + c_iou_list, c_lvl = activate_stream(restored_feat, rgb_img, clip_model, image_name, img_ann, thresh=mask_thresh, colormap_options=colormap_options) chosen_iou_all.extend(c_iou_list) @@ -290,8 +290,8 @@ def seed_everything(seed_value): np.random.seed(seed_value) torch.manual_seed(seed_value) os.environ['PYTHONHASHSEED'] = str(seed_value) - - if torch.cuda.is_available(): + + if torch.cuda.is_available(): torch.cuda.manual_seed(seed_value) torch.cuda.manual_seed_all(seed_value) torch.backends.cudnn.deterministic = True @@ -301,7 +301,7 @@ def seed_everything(seed_value): if __name__ == "__main__": seed_num = 42 seed_everything(seed_num) - + parser = ArgumentParser(description="prompt any label") parser.add_argument("--dataset_name", type=str, default=None) parser.add_argument('--feat_dir', type=str, default=None) diff --git a/render.py b/render.py index 9d3cdde..7053230 100644 --- a/render.py +++ b/render.py @@ -3,7 +3,7 @@ # GRAPHDECO research group, https://team.inria.fr/graphdeco # All rights reserved. # -# This software is free for non-commercial, research and evaluation use +# 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 @@ -39,10 +39,10 @@ def render_set(model_path, source_path, name, iteration, views, gaussians, pipel 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) @@ -50,7 +50,7 @@ def render_set(model_path, source_path, name, iteration, views, gaussians, pipel 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")) - + 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) @@ -58,7 +58,7 @@ def render_sets(dataset : ModelParams, iteration : int, pipeline : PipelineParam checkpoint = os.path.join(args.model_path, 'chkpnt30000.pth') (model_params, first_iter) = torch.load(checkpoint, weights_only=False) gaussians.restore(model_params, args, mode='test') - + bg_color = [1,1,1] if dataset.white_background else [0, 0, 0] background = torch.tensor(bg_color, dtype=torch.float32, device="cuda") @@ -70,7 +70,7 @@ def render_sets(dataset : ModelParams, iteration : int, pipeline : PipelineParam if __name__ == "__main__": # Set up command line argument parser - + parser = ArgumentParser(description="Testing script parameters") model = ModelParams(parser, sentinel=True) pipeline = PipelineParams(parser) diff --git a/train.py b/train.py index d84c883..6d0ccf6 100644 --- a/train.py +++ b/train.py @@ -3,7 +3,7 @@ # GRAPHDECO research group, https://team.inria.fr/graphdeco # All rights reserved. # -# This software is free for non-commercial, research and evaluation use +# 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 @@ -27,11 +27,11 @@ TENSORBOARD_FOUND = True except ImportError: TENSORBOARD_FOUND = False - + def training(dataset, opt, pipe, testing_iterations, saving_iterations, checkpoint_iterations, checkpoint, debug_from): - + first_iter = 0 tb_writer = prepare_output_and_logger(dataset) gaussians = GaussianModel(dataset.sh_degree) @@ -46,7 +46,7 @@ def training(dataset, opt, pipe, testing_iterations, saving_iterations, checkpoi if len(model_params) == 12 and opt.include_feature: first_iter = 0 gaussians.restore(model_params, opt) - + bg_color = [1, 1, 1] if dataset.white_background else [0, 0, 0] background = torch.tensor(bg_color, dtype=torch.float32, device="cuda") @@ -57,7 +57,7 @@ def training(dataset, opt, pipe, testing_iterations, saving_iterations, checkpoi ema_loss_for_log = 0.0 progress_bar = tqdm(range(first_iter, opt.iterations), desc="Training progress") first_iter += 1 - for iteration in range(first_iter, opt.iterations + 1): + for iteration in range(first_iter, opt.iterations + 1): if network_gui.conn == None: network_gui.try_connect() while network_gui.conn != None: @@ -85,17 +85,17 @@ def training(dataset, opt, pipe, testing_iterations, saving_iterations, checkpoi if not viewpoint_stack: viewpoint_stack = scene.getTrainCameras().copy() viewpoint_cam = viewpoint_stack.pop(randint(0, len(viewpoint_stack)-1)) - + # Render if (iteration - 1) == debug_from: pipe.debug = True render_pkg = render(viewpoint_cam, gaussians, pipe, background, opt) image, language_feature, viewspace_point_tensor, visibility_filter, radii = render_pkg["render"], render_pkg["language_feature_image"], render_pkg["viewspace_points"], render_pkg["visibility_filter"], render_pkg["radii"] - + # Loss if opt.include_feature: gt_language_feature, language_feature_mask = viewpoint_cam.get_language_feature(language_feature_dir=dataset.lf_path, feature_level=dataset.feature_level) - Ll1 = l1_loss(language_feature*language_feature_mask, gt_language_feature*language_feature_mask) + Ll1 = l1_loss(language_feature*language_feature_mask, gt_language_feature*language_feature_mask) loss = Ll1 else: gt_image = viewpoint_cam.original_image.cuda() @@ -128,7 +128,7 @@ def training(dataset, opt, pipe, testing_iterations, saving_iterations, checkpoi if iteration > opt.densify_from_iter and iteration % opt.densification_interval == 0: size_threshold = 20 if iteration > opt.opacity_reset_interval else None gaussians.densify_and_prune(opt.densify_grad_threshold, 0.005, scene.cameras_extent, size_threshold) - + if iteration % opt.opacity_reset_interval == 0 or (dataset.white_background and iteration == opt.densify_from_iter): gaussians.reset_opacity() @@ -140,15 +140,15 @@ def training(dataset, opt, pipe, testing_iterations, saving_iterations, checkpoi if (iteration in checkpoint_iterations): print("\n[ITER {}] Saving Checkpoint".format(iteration)) torch.save((gaussians.capture(opt.include_feature), iteration), scene.model_path + "/chkpnt" + str(iteration) + ".pth") - -def prepare_output_and_logger(args): + +def prepare_output_and_logger(args): if not args.model_path: if os.getenv('OAR_JOB_ID'): unique_str=os.getenv('OAR_JOB_ID') else: unique_str = str(uuid.uuid4()) args.model_path = os.path.join("./output/", unique_str[0:10]) - + # Set up output folder print("Output folder: {}".format(args.model_path)) os.makedirs(args.model_path, exist_ok = True) @@ -173,7 +173,7 @@ def training_report(tb_writer, iteration, Ll1, loss, l1_loss, elapsed, testing_i if iteration in testing_iterations: print(f'testing for iter {iteration}') torch.cuda.empty_cache() - validation_configs = ({'name': 'test', 'cameras' : scene.getTestCameras()}, + validation_configs = ({'name': 'test', 'cameras' : scene.getTestCameras()}, {'name': 'train', 'cameras' : [scene.getTrainCameras()[idx % len(scene.getTrainCameras())] for idx in range(5, 30, 5)]}) for config in validation_configs: @@ -190,7 +190,7 @@ def training_report(tb_writer, iteration, Ll1, loss, l1_loss, elapsed, testing_i l1_test += l1_loss(image, gt_image).mean().double() psnr_test += psnr(image, gt_image).mean().double() psnr_test /= len(config['cameras']) - l1_test /= len(config['cameras']) + l1_test /= len(config['cameras']) print("\n[ITER {}] Evaluating {}: L1 {} PSNR {}".format(iteration, config['name'], l1_test, psnr_test)) if tb_writer: tb_writer.add_scalar(config['name'] + '/loss_viewpoint - l1_loss', l1_test, iteration)