Spaces:
Runtime error
Runtime error
File size: 5,652 Bytes
7978a78 |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 |
import datetime
import json
import os
import time
import gradio as gr
import torch
import yaml
from huggingface_hub import hf_hub_download
from omegaconf import OmegaConf
from model.leo_agent import LeoAgentLLM
LOG_DIR = 'logs'
MESH_DIR = 'assets/scene_meshes'
MESH_NAMES = [os.path.splitext(fname)[0] for fname in os.listdir(MESH_DIR)]
ENABLE_BUTTON = gr.update(interactive=True)
DISABLE_BUTTON = gr.update(interactive=False)
DEVICE = 'cuda' if torch.cuda.is_available() else 'cpu'
ROLE_PROMPT = "You are an AI visual assistant situated in a 3D scene. "\
"You can perceive (1) an ego-view image (accessible when necessary) and (2) the objects (including yourself) in the scene (always accessible). "\
"You should properly respond to the USER's instruction according to the given visual information. "
EGOVIEW_PROMPT = "Ego-view image:"
OBJECTS_PROMPT = "Objects (including you) in the scene:"
TASK_PROMPT = "USER: {instruction} ASSISTANT:"
OBJ_FEATS_DIR = 'assets/obj_features'
def load_agent():
# build model
with open('model/cfg.yaml') as f:
cfg = yaml.safe_load(f)
cfg = OmegaConf.create(cfg)
agent = LeoAgentLLM(cfg)
# load checkpoint
if cfg.use_ckpt == 'hf':
ckpt_path = hf_hub_download(cfg.hf_ckpt_path[0], cfg.hf_ckpt_path[1])
else:
ckpt_path = cfg.local_ckpt_path
ckpt = torch.load(ckpt_path, map_location='cpu')
agent.load_state_dict(ckpt, strict=False)
agent.eval()
agent.to(DEVICE)
return agent
agent = load_agent()
def get_log_fname():
t = datetime.datetime.now()
fname = os.path.join(LOG_DIR, f'{t.year}-{t.month:02d}-{t.day:02d}.json')
return fname
def change_scene(dropdown_scene: str):
# reset 3D scene and chatbot history
return os.path.join(MESH_DIR, f'{dropdown_scene}.glb'), None
def receive_instruction(chatbot: gr.Chatbot, user_chat_input: gr.Textbox):
# display user input, after submitting user message, before inference
chatbot.append((user_chat_input, None))
return (chatbot, gr.update(value=""),) + (DISABLE_BUTTON,) * 5
def generate_response(
chatbot: gr.Chatbot,
dropdown_scene: gr.Dropdown,
dropdown_conversation_mode: gr.Dropdown,
repetition_penalty: float, length_penalty: float
):
# response starts
chatbot[-1] = (chatbot[-1][0], "▌")
yield (chatbot,) + (DISABLE_BUTTON,) * 5
# create data_dict, batch_size = 1
data_dict = {
'prompt_before_obj': [ROLE_PROMPT],
'prompt_middle_1': [EGOVIEW_PROMPT],
'prompt_middle_2': [OBJECTS_PROMPT],
'img_tokens': torch.zeros(1, 1, 4096).float(),
'img_masks': torch.zeros(1, 1).bool(),
'anchor_locs': torch.zeros(1, 3).float(),
}
# initialize prompt
prompt = ""
if 'Multi-round' in dropdown_conversation_mode:
# multi-round dialogue, with memory
for (q, a) in chatbot[:-1]:
prompt += f"USER: {q.strip()} ASSISTANT: {a.strip()}</s>"
prompt += f"USER: {chatbot[-1][0]} ASSISTANT:"
data_dict['prompt_after_obj'] = [prompt]
# anchor orientation
anchor_orient = torch.zeros(1, 4).float()
anchor_orient[:, -1] = 1
data_dict['anchor_orientation'] = anchor_orient
# load preprocessed scene features
data_dict.update(torch.load(os.path.join(OBJ_FEATS_DIR, f'{dropdown_scene}.pth'), map_location='cpu'))
# inference
for k, v in data_dict.items():
if isinstance(v, torch.Tensor):
data_dict[k] = v.to(DEVICE)
output = agent.generate(
data_dict,
repetition_penalty=float(repetition_penalty),
length_penalty=float(length_penalty),
)
output = output[0]
# display response
for out_len in range(1, len(output)-1):
chatbot[-1] = (chatbot[-1][0], output[:out_len] + '▌')
yield (chatbot,) + (DISABLE_BUTTON,) * 5
time.sleep(0.01)
chatbot[-1] = (chatbot[-1][0], output)
vote_response(chatbot, 'log', dropdown_scene, dropdown_conversation_mode)
yield (chatbot,) + (ENABLE_BUTTON,) * 5
def vote_response(
chatbot: gr.Chatbot, vote_type: str,
dropdown_scene: gr.Dropdown,
dropdown_conversation_mode: gr.Dropdown
):
t = datetime.datetime.now()
this_log = {
'time': f'{t.hour:02d}:{t.minute:02d}:{t.second:02d}',
'type': vote_type,
'scene': dropdown_scene,
'mode': dropdown_conversation_mode,
'dialogue': chatbot,
}
fname = get_log_fname()
if os.path.exists(fname):
with open(fname) as f:
logs = json.load(f)
logs.append(this_log)
else:
logs = [this_log]
with open(fname, 'w') as f:
json.dump(logs, f, indent=2)
def upvote_response(
chatbot: gr.Chatbot,
dropdown_scene: gr.Dropdown,
dropdown_conversation_mode: gr.Dropdown
):
vote_response(chatbot, 'upvote', dropdown_scene, dropdown_conversation_mode)
return ("",) + (DISABLE_BUTTON,) * 3
def downvote_response(
chatbot: gr.Chatbot,
dropdown_scene: gr.Dropdown,
dropdown_conversation_mode: gr.Dropdown
):
vote_response(chatbot, 'downvote', dropdown_scene, dropdown_conversation_mode)
return ("",) + (DISABLE_BUTTON,) * 3
def flag_response(
chatbot: gr.Chatbot,
dropdown_scene: gr.Dropdown,
dropdown_conversation_mode: gr.Dropdown
):
vote_response(chatbot, 'flag', dropdown_scene, dropdown_conversation_mode)
return ("",) + (DISABLE_BUTTON,) * 3
def clear_history():
# reset chatbot history
return (None, "",) + (DISABLE_BUTTON,) * 4
|