shap-e/shap_e/examples/gc/ShapeGenerator.py

import torch
from shap_e.diffusion.sample import sample_latents
from shap_e.diffusion.gaussian_diffusion import diffusion_from_config
from shap_e.models.download import load_model, load_config
from shap_e.util.notebooks import decode_latent_mesh
from tqdm import tqdm
import pygltflib
from pygltflib import GLTF2
import trimesh
import open3d as o3d
import os
import datetime

class ShapeGenerator:
    def __init__(self, output_path, batch_size, step_size, guidance):
        self.device = None
        self.xm = None
        self.model = None
        self.diffusion = None
        self.iterations = 0
        self.latents = None
        self.output_path = output_path
        self.batch_size = batch_size
        self.step_size = step_size
        self.guidance = guidance

    def run(self):
        print("Loading Models..")
        self.load_models()
        print("Finished Loading Models!")


    def load_models(self):
        self.device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
        self.xm = load_model('transmitter', device=self.device)
        self.model = load_model('text300M', device=self.device)
        self.diffusion = diffusion_from_config(load_config('diffusion'))
       
    
    def generate_object(self, prompt):
        batch_size = 2
        # Create random latents
        latent_dim = self.model.d_latent
        random_latents = torch.randn(batch_size, latent_dim).to(self.model.device)
        print(random_latents.shape)
        model_kwargs = {}
        model_kwargs = dict(texts=[prompt] * self.batch_size)


        self.latents = sample_latents(
            batch_size=self.batch_size,
            model=self.model,
            diffusion=self.diffusion,
            guidance_scale=self.guidance,
            model_kwargs=model_kwargs,
            progress=True,  # This should already show progress
            clip_denoised=True,
            use_fp16=True,
            use_karras=True,
            karras_steps=self.step_size,
            sigma_min=1e-3,
            sigma_max=160,
            s_churn=0,
            device = self.model.device,
        )

        mesh = self.export_model(prompt)
        return mesh, prompt

    def export_model(self, prompt):
        timestamp = datetime.datetime.now().strftime("%Y%m%d-%H%M%S")
        obj_filepath = f'{prompt}-{self.iterations}.obj'
        output_filepath = f'{self.output_path}/{prompt}-{timestamp}.gltf'
        print(output_filepath)
        for i, latent in enumerate(self.latents):
            t = decode_latent_mesh(self.xm, latent).tri_mesh()
            with open(obj_filepath, 'w') as f:
                t.write_obj(f)
        final_mesh = self.construct_mesh(obj_filepath)
        o3d.io.write_triangle_mesh(output_filepath, final_mesh)
        self.iterations += 1
        return final_mesh

    def construct_mesh(self, obj_fp):
        mesh = o3d.io.read_triangle_mesh(obj_fp)

        if os.path.exists(obj_fp):
            os.remove(obj_fp)

        original_triangle_count = len(mesh.triangles)
        target_triangle_count = original_triangle_count // 3
        decimated_mesh = mesh.simplify_quadric_decimation(
            target_number_of_triangles=target_triangle_count)
        filtered_mesh = decimated_mesh.filter_smooth_simple(number_of_iterations=5)
        filtered_mesh.compute_vertex_normals()
        return filtered_mesh