#!/usr/bin/env python from __future__ import annotations import os import datetime import gradio as gr import spaces @spaces.GPU(duration=60 * 3) def run_on_gpu(input_point_cloud: gr.utils.NamedString, gen_resolution_global: int, padding_factor: float, gen_subsample_manifold_iter: int, gen_refine_iter: int): print('Started inference at {}'.format(datetime.datetime.now())) print('Inputs:', input_point_cloud, gen_resolution_global, padding_factor, gen_subsample_manifold_iter, gen_refine_iter) print('Types:', type(input_point_cloud), type(gen_resolution_global), type(padding_factor), type(gen_subsample_manifold_iter), type(gen_refine_iter)) import os import sys sys.path.append(os.path.abspath('ppsurf')) # from ppsurf.pps import cli_main import subprocess in_file = '{}'.format(input_point_cloud.name) # append 'rec' to the input file name splitext_result = os.path.splitext(in_file) out_file = splitext_result[0] + '_rec' + splitext_result[1] out_dir = os.path.dirname(out_file) model_path = 'models/ppsurf_50nn/version_0/checkpoints/last.ckpt' args = [ 'pps.py', 'predict', '-c', 'ppsurf/configs/poco.yaml', '-c', 'ppsurf/configs/ppsurf.yaml', '-c', 'ppsurf/configs/ppsurf_50nn.yaml', '--ckpt_path', model_path, '--data.init_args.in_file', in_file, '--model.init_args.results_dir', out_dir, '--trainer.logger', 'False', '--trainer.devices', '1', '--model.init_args.gen_resolution_global', str(gen_resolution_global), '--data.init_args.padding_factor', str(padding_factor), '--model.init_args.gen_subsample_manifold_iter', str(gen_subsample_manifold_iter), '--model.init_args.gen_refine_iter', str(gen_refine_iter), ] sys.argv = args subprocess.run(['python', 'ppsurf/pps.py'] + args[1:]) # cli_main() print('Finished inference at {}'.format(datetime.datetime.now())) result_3d_model = out_file output_file = out_file progress_text = 'done' return result_3d_model, output_file, progress_text def main(): description = '''# [PPSurf](https://github.com/cg-tuwien/ppsurf) Supported file formats: - PLY, STL, OBJ and other mesh files, - XYZ as whitespace-separated text file, - NPY and NPZ (key='arr_0'), - LAS and LAZ (version 1.0-1.4), COPC and CRS. Best results for 50k-250k points. This method is meant for scans of single and few objects. Quality for scenes and landscapes will be lower. Inference takes about 2 minutes. ''' # can't render many input types directly in Gradio Model3D # so we need to convert to supported format # Gradio can't draw point clouds anyway, so we skip this for now # def convert_to_ply(input_point_cloud_upload: gr.utils.NamedString): # # # add absolute path to import dirs # import sys # import os # sys.path.append(os.path.abspath('ppsurf')) # # # import os # # os.chdir('ppsurf') # # print('Inputs:', input_point_cloud_upload, type(input_point_cloud_upload)) # input_shape: str = input_point_cloud_upload.name # if not input_shape.endswith('.ply'): # # load file # from ppsurf.source.occupancy_data_module import OccupancyDataModule # pts_np = OccupancyDataModule.load_pts(input_shape) # # # convert to ply # import trimesh # mesh = trimesh.Trimesh(vertices=pts_np[:, :3]) # input_shape = input_shape + '.ply' # mesh.export(input_shape) # # print('ls:\n', subprocess.run(['ls', os.path.dirname(input_shape)])) # # # show in viewer # print(type(input_tabs)) # # print(type(input_point_cloud_viewer)) # # input_tabs.selected = 'pc_viewer' # # input_point_cloud_viewer.value = input_shape if (SPACE_ID := os.getenv('SPACE_ID')) is not None: description += (f'\n

For faster inference without waiting in queue, ' f'you may duplicate the space and upgrade to GPU in settings. ' f'' f'Duplicate Space

') with gr.Blocks(css='style.css') as demo: gr.Markdown(description) with gr.Row(): with gr.Column(): # with gr.Tabs() as input_tabs: # with gr.TabItem(label='Input Point Cloud Upload', id='pc_upload'): input_point_cloud_upload = gr.File(show_label=False, file_count='single') # input_point_cloud_upload.upload( # fn=convert_to_ply, # inputs=[ # input_point_cloud_upload, # ], # outputs=[ # # input_point_cloud_viewer, # not available here # ]) # with gr.TabItem(label='Input Point Cloud Viewer', id='pc_viewer'): # input_point_cloud_viewer = gr.Model3D(show_label=False) gen_resolution_global = gr.Slider( label='Grid Resolution (larger for more details)', minimum=17, maximum=513, value=129, step=2) padding_factor = gr.Slider( label='Padding Factor (larger if object is cut off at boundaries)', minimum=0, maximum=1.0, value=0.05, step=0.05) gen_subsample_manifold_iter = gr.Slider( label='Subsample Manifold Iterations (larger for larger point clouds)', minimum=3, maximum=30, value=10, step=1) gen_refine_iter = gr.Slider( label='Edge Refinement Iterations (larger for more details)', minimum=3, maximum=30, value=10, step=1) with gr.Column(): progress_text = gr.Text(label='Progress') with gr.Tabs(): with gr.TabItem(label='Reconstructed 3D model'): result_3d_model = gr.Model3D(show_label=False) with gr.TabItem(label='Output mesh file'): output_file = gr.File(show_label=False) # with gr.Row(): # examples = [ # ['shapes/dragon1.obj', 'a photo of a dragon', 0, 7.5], # ['shapes/dragon2.obj', 'a photo of a dragon', 0, 7.5], # ['shapes/eagle.obj', 'a photo of an eagle', 0, 7.5], # ['shapes/napoleon.obj', 'a photo of Napoleon Bonaparte', 3, 7.5], # ['shapes/nascar.obj', 'A next gen nascar', 2, 10], # ] # gr.Examples(examples=examples, # inputs=[ # input_point_cloud_viewer, # text, # seed, # guidance_scale, # ], # outputs=[ # result_3d_model, # output_file, # ], # cache_examples=False) with gr.Row(): run_button = gr.Button('Reconstruct with PPSurf') run_button.click(fn=run_on_gpu, inputs=[ input_point_cloud_upload, gen_resolution_global, padding_factor, gen_subsample_manifold_iter, gen_refine_iter, ], outputs=[ result_3d_model, output_file, progress_text, ]) demo.queue(max_size=5) demo.launch(debug=True) if __name__ == '__main__': main()