Spaces:
Running
on
Zero
Running
on
Zero
Create app.py
Browse files
app.py
ADDED
|
@@ -0,0 +1,261 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
import os
|
| 2 |
+
import spaces
|
| 3 |
+
import trimesh
|
| 4 |
+
import traceback
|
| 5 |
+
import numpy as np
|
| 6 |
+
import gradio as gr
|
| 7 |
+
from multiprocessing import Process, Queue
|
| 8 |
+
|
| 9 |
+
import torch
|
| 10 |
+
from torch import nn
|
| 11 |
+
from transformers import (
|
| 12 |
+
AutoTokenizer, Qwen2ForCausalLM, Qwen2Model, PreTrainedModel)
|
| 13 |
+
from transformers.modeling_outputs import CausalLMOutputWithPast
|
| 14 |
+
|
| 15 |
+
|
| 16 |
+
class FourierPointEncoder(nn.Module):
|
| 17 |
+
def __init__(self, hidden_size):
|
| 18 |
+
super().__init__()
|
| 19 |
+
frequencies = 2.0 ** torch.arange(8, dtype=torch.float32)
|
| 20 |
+
self.register_buffer('frequencies', frequencies, persistent=False)
|
| 21 |
+
self.projection = nn.Linear(54, hidden_size)
|
| 22 |
+
|
| 23 |
+
def forward(self, points):
|
| 24 |
+
x = points[..., :3]
|
| 25 |
+
x = (x.unsqueeze(-1) * self.frequencies).view(*x.shape[:-1], -1)
|
| 26 |
+
x = torch.cat((points[..., :3], x.sin(), x.cos()), dim=-1)
|
| 27 |
+
x = self.projection(torch.cat((x, points[..., 3:]), dim=-1))
|
| 28 |
+
return x
|
| 29 |
+
|
| 30 |
+
|
| 31 |
+
class CADRecode(Qwen2ForCausalLM):
|
| 32 |
+
def __init__(self, config):
|
| 33 |
+
PreTrainedModel.__init__(self, config)
|
| 34 |
+
self.model = Qwen2Model(config)
|
| 35 |
+
self.vocab_size = config.vocab_size
|
| 36 |
+
self.lm_head = nn.Linear(config.hidden_size, config.vocab_size, bias=False)
|
| 37 |
+
|
| 38 |
+
torch.set_default_dtype(torch.float32)
|
| 39 |
+
self.point_encoder = FourierPointEncoder(config.hidden_size)
|
| 40 |
+
torch.set_default_dtype(torch.bfloat16)
|
| 41 |
+
|
| 42 |
+
def forward(self,
|
| 43 |
+
input_ids=None,
|
| 44 |
+
attention_mask=None,
|
| 45 |
+
point_cloud=None,
|
| 46 |
+
position_ids=None,
|
| 47 |
+
past_key_values=None,
|
| 48 |
+
inputs_embeds=None,
|
| 49 |
+
labels=None,
|
| 50 |
+
use_cache=None,
|
| 51 |
+
output_attentions=None,
|
| 52 |
+
output_hidden_states=None,
|
| 53 |
+
return_dict=None,
|
| 54 |
+
cache_position=None):
|
| 55 |
+
output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
|
| 56 |
+
output_hidden_states = output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
|
| 57 |
+
return_dict = return_dict if return_dict is not None else self.config.use_return_dict
|
| 58 |
+
|
| 59 |
+
# concatenate point and text embeddings
|
| 60 |
+
if past_key_values is None or past_key_values.get_seq_length() == 0:
|
| 61 |
+
assert inputs_embeds is None
|
| 62 |
+
inputs_embeds = self.model.embed_tokens(input_ids)
|
| 63 |
+
point_embeds = self.point_encoder(point_cloud).bfloat16()
|
| 64 |
+
inputs_embeds[attention_mask == -1] = point_embeds.reshape(-1, point_embeds.shape[2])
|
| 65 |
+
attention_mask[attention_mask == -1] = 1
|
| 66 |
+
input_ids = None
|
| 67 |
+
position_ids = None
|
| 68 |
+
|
| 69 |
+
# decoder outputs consists of (dec_features, layer_state, dec_hidden, dec_attn)
|
| 70 |
+
outputs = self.model(
|
| 71 |
+
input_ids=input_ids,
|
| 72 |
+
attention_mask=attention_mask,
|
| 73 |
+
position_ids=position_ids,
|
| 74 |
+
past_key_values=past_key_values,
|
| 75 |
+
inputs_embeds=inputs_embeds,
|
| 76 |
+
use_cache=use_cache,
|
| 77 |
+
output_attentions=output_attentions,
|
| 78 |
+
output_hidden_states=output_hidden_states,
|
| 79 |
+
return_dict=return_dict,
|
| 80 |
+
cache_position=cache_position)
|
| 81 |
+
|
| 82 |
+
hidden_states = outputs[0]
|
| 83 |
+
logits = self.lm_head(hidden_states)
|
| 84 |
+
logits = logits.float()
|
| 85 |
+
|
| 86 |
+
loss = None
|
| 87 |
+
if labels is not None:
|
| 88 |
+
# Shift so that tokens < n predict n
|
| 89 |
+
shift_logits = logits[..., :-1, :].contiguous()
|
| 90 |
+
shift_labels = labels[..., 1:].contiguous()
|
| 91 |
+
# Flatten the tokens
|
| 92 |
+
loss_fct = nn.CrossEntropyLoss()
|
| 93 |
+
shift_logits = shift_logits.view(-1, self.config.vocab_size)
|
| 94 |
+
shift_labels = shift_labels.view(-1)
|
| 95 |
+
# Enable model parallelism
|
| 96 |
+
shift_labels = shift_labels.to(shift_logits.device)
|
| 97 |
+
loss = loss_fct(shift_logits, shift_labels)
|
| 98 |
+
|
| 99 |
+
if not return_dict:
|
| 100 |
+
output = (logits,) + outputs[1:]
|
| 101 |
+
return (loss,) + output if loss is not None else output
|
| 102 |
+
|
| 103 |
+
return CausalLMOutputWithPast(
|
| 104 |
+
loss=loss,
|
| 105 |
+
logits=logits,
|
| 106 |
+
past_key_values=outputs.past_key_values,
|
| 107 |
+
hidden_states=outputs.hidden_states,
|
| 108 |
+
attentions=outputs.attentions)
|
| 109 |
+
|
| 110 |
+
def prepare_inputs_for_generation(self, *args, **kwargs):
|
| 111 |
+
model_inputs = super().prepare_inputs_for_generation(*args, **kwargs)
|
| 112 |
+
model_inputs['point_cloud'] = kwargs['point_cloud']
|
| 113 |
+
return model_inputs
|
| 114 |
+
|
| 115 |
+
|
| 116 |
+
def mesh_to_point_cloud(mesh, n_points=256):
|
| 117 |
+
vertices, faces = trimesh.sample.sample_surface(mesh, n_points)
|
| 118 |
+
point_cloud = np.concatenate((
|
| 119 |
+
np.asarray(vertices),
|
| 120 |
+
mesh.face_normals[faces]
|
| 121 |
+
), axis=1)
|
| 122 |
+
ids = np.lexsort((point_cloud[:, 0], point_cloud[:, 1], point_cloud[:, 2]))
|
| 123 |
+
point_cloud = point_cloud[ids]
|
| 124 |
+
return point_cloud
|
| 125 |
+
|
| 126 |
+
|
| 127 |
+
def py_string_to_mesh_file(py_string, mesh_path, queue):
|
| 128 |
+
try:
|
| 129 |
+
exec(py_string, globals())
|
| 130 |
+
compound = globals()['r'].val()
|
| 131 |
+
vertices, faces = compound.tessellate(0.001, 0.1)
|
| 132 |
+
mesh = trimesh.Trimesh([(v.x, v.y, v.z) for v in vertices], faces)
|
| 133 |
+
mesh.export(mesh_path)
|
| 134 |
+
except:
|
| 135 |
+
queue.put(traceback.format_exc())
|
| 136 |
+
|
| 137 |
+
|
| 138 |
+
def py_string_to_mesh_file_safe(py_string, mesh_path):
|
| 139 |
+
# CadQuery code predicted by LLM may be unsafe and cause memory leaks.
|
| 140 |
+
# That's why we execute it in a separace Process with timeout.
|
| 141 |
+
queue = Queue()
|
| 142 |
+
process = Process(
|
| 143 |
+
target=py_string_to_mesh_file,
|
| 144 |
+
args=(py_string, mesh_path, queue))
|
| 145 |
+
process.start()
|
| 146 |
+
process.join(3)
|
| 147 |
+
|
| 148 |
+
if process.is_alive():
|
| 149 |
+
process.terminate()
|
| 150 |
+
process.join()
|
| 151 |
+
raise RuntimeError('Process is alive after 3 seconds')
|
| 152 |
+
|
| 153 |
+
if not queue.empty():
|
| 154 |
+
raise RuntimeError(queue.get())
|
| 155 |
+
|
| 156 |
+
|
| 157 |
+
@spaces.GPU(duration=20)
|
| 158 |
+
def run_gpu(model, input_ids, attention_mask, point_cloud, pad_token_id):
|
| 159 |
+
if torch.cuda.is_available():
|
| 160 |
+
model = model.cuda()
|
| 161 |
+
with torch.no_grad():
|
| 162 |
+
batch_ids = model.generate(
|
| 163 |
+
input_ids=torch.tensor(input_ids).unsqueeze(0).to(model.device),
|
| 164 |
+
attention_mask=torch.tensor(attention_mask).unsqueeze(0).to(model.device),
|
| 165 |
+
point_cloud=torch.tensor(point_cloud.astype(np.float32)).unsqueeze(0).to(model.device),
|
| 166 |
+
max_new_tokens=768,
|
| 167 |
+
pad_token_id=pad_token_id).cpu()
|
| 168 |
+
return batch_ids
|
| 169 |
+
|
| 170 |
+
|
| 171 |
+
def run_test(in_mesh_path, seed, results):
|
| 172 |
+
mesh = trimesh.load(in_mesh_path)
|
| 173 |
+
mesh.apply_translation(-(mesh.bounds[0] + mesh.bounds[1]) / 2.0)
|
| 174 |
+
mesh.apply_scale(2.0 / max(mesh.extents))
|
| 175 |
+
np.random.seed(seed)
|
| 176 |
+
point_cloud = mesh_to_point_cloud(mesh)
|
| 177 |
+
|
| 178 |
+
pcd_path = '/tmp/pcd.obj'
|
| 179 |
+
trimesh.points.PointCloud(point_cloud[:, :3]).export(pcd_path)
|
| 180 |
+
results.append(pcd_path)
|
| 181 |
+
|
| 182 |
+
tokenizer = AutoTokenizer.from_pretrained(
|
| 183 |
+
'Qwen/Qwen2-1.5B',
|
| 184 |
+
pad_token='<|im_end|>',
|
| 185 |
+
padding_side='left')
|
| 186 |
+
model = CADRecode.from_pretrained(
|
| 187 |
+
'filapro/cad-recode',
|
| 188 |
+
torch_dtype='auto').eval()
|
| 189 |
+
|
| 190 |
+
input_ids = [tokenizer.pad_token_id] * len(point_cloud) + [tokenizer('<|im_start|>')['input_ids'][0]]
|
| 191 |
+
attention_mask = [-1] * len(point_cloud) + [1]
|
| 192 |
+
batch_ids = run_gpu(model, input_ids, attention_mask, point_cloud, tokenizer.pad_token_id)
|
| 193 |
+
py_string = tokenizer.batch_decode(batch_ids)[0]
|
| 194 |
+
begin = py_string.find('<|im_start|>') + 12
|
| 195 |
+
end = py_string.find('<|endoftext|>')
|
| 196 |
+
py_string = py_string[begin: end]
|
| 197 |
+
results.append(py_string)
|
| 198 |
+
|
| 199 |
+
out_mesh_path = '/tmp/mesh.stl'
|
| 200 |
+
py_string_to_mesh_file_safe(py_string, out_mesh_path)
|
| 201 |
+
results.append(out_mesh_path)
|
| 202 |
+
|
| 203 |
+
|
| 204 |
+
def run_test_safe(in_mesh_path, seed):
|
| 205 |
+
results, log = list(), str()
|
| 206 |
+
try:
|
| 207 |
+
run_test(in_mesh_path, seed, results)
|
| 208 |
+
except:
|
| 209 |
+
log += 'Status: FAILED\n' + traceback.format_exc()
|
| 210 |
+
return results + [None] * (3 - len(results)) + [log]
|
| 211 |
+
|
| 212 |
+
|
| 213 |
+
def run():
|
| 214 |
+
with gr.Blocks() as demo:
|
| 215 |
+
with gr.Row():
|
| 216 |
+
gr.Markdown('## CAD-Recode Demo\n'
|
| 217 |
+
'Upload mesh or select from examples and press Run! Mesh ⇾ 256 points ⇾ Python code by CAD-Recode ⇾ CAD model.')
|
| 218 |
+
|
| 219 |
+
with gr.Row(equal_height=True):
|
| 220 |
+
in_model = gr.Model3D(label='1. Input Mesh', interactive=True)
|
| 221 |
+
point_model = gr.Model3D(label='2. Sampled Point Cloud', display_mode='point_cloud', interactive=False)
|
| 222 |
+
out_model = gr.Model3D(
|
| 223 |
+
label='4. Result CAD Model', interactive=False
|
| 224 |
+
)
|
| 225 |
+
|
| 226 |
+
with gr.Row():
|
| 227 |
+
with gr.Column():
|
| 228 |
+
with gr.Row():
|
| 229 |
+
seed_slider = gr.Slider(label='Random Seed', value=42, interactive=True)
|
| 230 |
+
with gr.Row():
|
| 231 |
+
_ = gr.Examples(
|
| 232 |
+
examples=[
|
| 233 |
+
['./data/49215_5368e45e_0000.stl', 42],
|
| 234 |
+
['./data/00882236.stl', 6],
|
| 235 |
+
['./data/User Library-engrenage.stl', 18],
|
| 236 |
+
['./data/00010900.stl', 42],
|
| 237 |
+
['./data/21492_8bd34fc1_0008.stl', 42],
|
| 238 |
+
['./data/00375556.stl', 53],
|
| 239 |
+
['./data/49121_adb01620_0000.stl', 42]],
|
| 240 |
+
example_labels=[
|
| 241 |
+
'fusion360_table1', 'deepcad_star', 'cc3d_gear', 'deepcad_barrels',
|
| 242 |
+
'fusion360_gear', 'deepcad_house', 'fusion360_table2'],
|
| 243 |
+
inputs=[in_model, seed_slider],
|
| 244 |
+
cache_examples=False)
|
| 245 |
+
with gr.Row():
|
| 246 |
+
run_button = gr.Button('Run')
|
| 247 |
+
|
| 248 |
+
with gr.Column():
|
| 249 |
+
out_code = gr.Code(language='python', label='3. Generated Python Code', wrap_lines=True, interactive=False)
|
| 250 |
+
|
| 251 |
+
with gr.Column():
|
| 252 |
+
log_textbox = gr.Textbox(label='Log', placeholder='Status: OK', interactive=False)
|
| 253 |
+
|
| 254 |
+
run_button.click(
|
| 255 |
+
run_test_safe, inputs=[in_model, seed_slider], outputs=[point_model, out_code, out_model, log_textbox])
|
| 256 |
+
|
| 257 |
+
demo.launch()
|
| 258 |
+
|
| 259 |
+
|
| 260 |
+
os.environ['TOKENIZERS_PARALLELISM'] = 'False'
|
| 261 |
+
run()
|