ZeroGPU (#5)

- Update for ZeroGPU (91189e1bee8b4418409045f231a721b869fd024a)


Co-authored-by: hysts <hysts@users.noreply.huggingface.co>

.gitattributes

@@ -33,3 +33,4 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
 *.zip filter=lfs diff=lfs merge=lfs -text
 *.zst filter=lfs diff=lfs merge=lfs -text
 *tfevents* filter=lfs diff=lfs merge=lfs -text
+*.whl filter=lfs diff=lfs merge=lfs -text

README.md

@@ -4,7 +4,8 @@ emoji: 🦀
 colorFrom: red
 colorTo: indigo
 sdk: gradio
-sdk_version: 4.17.0
+sdk_version: 4.20.1
+python_version: 3.10.13
 app_file: app.py
 pinned: false
 license: mit

app.py

@@ -1,4 +1,6 @@
 import os
+import shlex
+import subprocess
 import tyro
 import imageio
 import numpy as np
@@ -15,8 +17,7 @@ import gradio as gr
 from huggingface_hub import hf_hub_download
 ckpt_path = hf_hub_download(repo_id="ashawkey/LGM", filename="model_fp16.safetensors")
 
-# NOTE: no -e... else it's not working!
-os.system("pip install ./diff-gaussian-rasterization")
+subprocess.run(shlex.split("pip install wheel/diff_gaussian_rasterization-0.0.0-cp310-cp310-linux_x86_64.whl"))
 
 import kiui
 from kiui.op import recenter
@@ -66,7 +67,7 @@ model = model.half().to(device)
 model.eval()
 
 tan_half_fov = np.tan(0.5 * np.deg2rad(opt.fovy))
-proj_matrix = torch.zeros(4, 4, dtype=torch.float32, device=device)
+proj_matrix = torch.zeros(4, 4, dtype=torch.float32).to(device)
 proj_matrix[0, 0] = 1 / tan_half_fov
 proj_matrix[1, 1] = 1 / tan_half_fov
 proj_matrix[2, 2] = (opt.zfar + opt.znear) / (opt.zfar - opt.znear)
@@ -127,7 +128,7 @@ def process(input_image, prompt, prompt_neg='', input_elevation=0, input_num_ste
         image = image.astype(np.float32) / 255.0
         image = image[..., :3] * image[..., 3:4] + (1 - image[..., 3:4])
         mv_image = pipe_image(prompt, image, negative_prompt=prompt_neg, num_inference_steps=input_num_steps, guidance_scale=5.0,  elevation=input_elevation)
-        
+
     mv_image_grid = np.concatenate([
         np.concatenate([mv_image[1], mv_image[2]], axis=1),
         np.concatenate([mv_image[3], mv_image[0]], axis=1),
@@ -146,21 +147,21 @@ def process(input_image, prompt, prompt_neg='', input_elevation=0, input_num_ste
         with torch.autocast(device_type='cuda', dtype=torch.float16):
             # generate gaussians
             gaussians = model.forward_gaussians(input_image)
-        
+
         # save gaussians
         model.gs.save_ply(gaussians, output_ply_path)
-        
-        # render 360 video 
+
+        # render 360 video
         images = []
         elevation = 0
         if opt.fancy_video:
             azimuth = np.arange(0, 720, 4, dtype=np.int32)
             for azi in tqdm.tqdm(azimuth):
-                
+
                 cam_poses = torch.from_numpy(orbit_camera(elevation, azi, radius=opt.cam_radius, opengl=True)).unsqueeze(0).to(device)
 
                 cam_poses[:, :3, 1:3] *= -1 # invert up & forward direction
-                
+
                 # cameras needed by gaussian rasterizer
                 cam_view = torch.inverse(cam_poses).transpose(1, 2) # [V, 4, 4]
                 cam_view_proj = cam_view @ proj_matrix # [V, 4, 4]
@@ -173,11 +174,11 @@ def process(input_image, prompt, prompt_neg='', input_elevation=0, input_num_ste
         else:
             azimuth = np.arange(0, 360, 2, dtype=np.int32)
             for azi in tqdm.tqdm(azimuth):
-                
+
                 cam_poses = torch.from_numpy(orbit_camera(elevation, azi, radius=opt.cam_radius, opengl=True)).unsqueeze(0).to(device)
 
                 cam_poses[:, :3, 1:3] *= -1 # invert up & forward direction
-                
+
                 # cameras needed by gaussian rasterizer
                 cam_view = torch.inverse(cam_poses).transpose(1, 2) # [V, 4, 4]
                 cam_view_proj = cam_view @ proj_matrix # [V, 4, 4]
@@ -201,7 +202,7 @@ _DESCRIPTION = '''
 <a style="display:inline-block; margin-left: .5em" href="https://github.com/3DTopia/LGM"><img src='https://img.shields.io/github/stars/3DTopia/LGM?style=social'/></a>
 </div>
 
-* Input can be only text, only image, or both image and text. 
+* Input can be only text, only image, or both image and text.
 * Output is a `ply` file containing the 3D Gaussians, please check our [repo](https://github.com/3DTopia/LGM/blob/main/readme.md) for visualization and mesh conversion.
 * If you find the output unsatisfying, try using different seeds!
 '''
@@ -212,7 +213,7 @@ with block:
         with gr.Column(scale=1):
             gr.Markdown('# ' + _TITLE)
     gr.Markdown(_DESCRIPTION)
-    
+
     with gr.Row(variant='panel'):
         with gr.Column(scale=1):
             # input image
@@ -230,7 +231,7 @@ with block:
             # gen button
             button_gen = gr.Button("Generate")
 
-        
+
         with gr.Column(scale=1):
             with gr.Tab("Video"):
                 # final video results
@@ -242,7 +243,7 @@ with block:
                 output_image = gr.Image(interactive=False, show_label=False)
 
         button_gen.click(process, inputs=[input_image, input_text, input_neg_text, input_elevation, input_num_steps, input_seed], outputs=[output_image, output_video, output_file])
-    
+
     gr.Examples(
         examples=[
             "data_test/frog_sweater.jpg",
@@ -275,5 +276,5 @@ with block:
         cache_examples=True,
         label='Text-to-3D Examples'
     )
-    
-block.launch()
+
+block.launch()

requirements.txt

@@ -1,4 +1,3 @@
---extra-index-url https://download.pytorch.org/whl/cu118
 torch==2.0.0 
 xformers
 
@@ -28,4 +27,4 @@ trimesh
 kiui >= 0.2.3
 xatlas
 roma
-plyfile
+plyfile

wheel/diff_gaussian_rasterization-0.0.0-cp310-cp310-linux_x86_64.whl

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:42bf718442ba764469170abc09d99a70b7c1d891dc290f2e1247db09c95a0e88
+size 3021758