app.py

import openai
import numpy as np
from tempfile import NamedTemporaryFile
import copy
import shapely
from shapely.geometry import *
from shapely.affinity import *
from omegaconf import OmegaConf
from moviepy.editor import ImageSequenceClip
import gradio as gr


from lmp import LMP, LMPFGen
from sim import PickPlaceEnv, LMP_wrapper
from consts import ALL_BLOCKS, ALL_BOWLS


class DemoRunner:

    def __init__(self):
        self._cfg = OmegaConf.to_container(OmegaConf.load('cfg.yaml'), resolve=True)
        self._env = None
        self._model_name = ''

    def make_LMP(self, env):
        # LMP env wrapper
        cfg = copy.deepcopy(self._cfg)
        cfg['env'] = {
            'init_objs': list(env.obj_name_to_id.keys()),
            'coords': cfg['tabletop_coords']
        }
        for vs in cfg['lmps'].values():
            vs['engine'] = self._model_name

        LMP_env = LMP_wrapper(env, cfg)
        # creating APIs that the LMPs can interact with
        fixed_vars = {
            'np': np
        }
        fixed_vars.update({
            name: eval(name)
            for name in shapely.geometry.__all__ + shapely.affinity.__all__
        })
        variable_vars = {
            k: getattr(LMP_env, k)
            for k in [
                'get_bbox', 'get_obj_pos', 'get_color', 'is_obj_visible', 'denormalize_xy',
                'put_first_on_second', 'get_obj_names',
                'get_corner_name', 'get_side_name',
            ]
        }
        variable_vars['say'] = lambda msg: print(f'robot says: {msg}')

        # creating the function-generating LMP
        lmp_fgen = LMPFGen(cfg['lmps']['fgen'], fixed_vars, variable_vars)

        # creating other low-level LMPs
        variable_vars.update({
            k: LMP(k, cfg['lmps'][k], lmp_fgen, fixed_vars, variable_vars)
            for k in ['parse_obj_name', 'parse_position', 'parse_question', 'transform_shape_pts']
        })

        # creating the LMP that deals w/ high-level language commands
        lmp_tabletop_ui = LMP(
            'tabletop_ui', cfg['lmps']['tabletop_ui'], lmp_fgen, fixed_vars, variable_vars
        )

        return lmp_tabletop_ui

    def setup(self, api_key, model_name, n_blocks, n_bowls):
        openai.api_key = api_key
        self._model_name = model_name

        self._env = PickPlaceEnv(render=True, high_res=False, high_frame_rate=False)
        block_list = np.random.choice(ALL_BLOCKS, size=n_blocks, replace=False).tolist()
        bowl_list = np.random.choice(ALL_BOWLS, size=n_bowls, replace=False).tolist()
        obj_list = block_list + bowl_list
        self._env.reset(obj_list)

        self._lmp_tabletop_ui = self.make_LMP(self._env)

        info = '## Available objects: \n- ' + '\n- '.join(obj_list)
        img = self._env.get_camera_image()

        return info, img

    def run(self, instruction):
        if self._env is None:
            return 'Please run setup first'

        self._env.cache_video = []

        self._lmp_tabletop_ui(instruction, f'objects = {self._env.object_list}')

        video_file_name = ''
        if self._env.cache_video:
            rendered_clip = ImageSequenceClip(self._env.cache_video, fps=25)
            video_file_name = NamedTemporaryFile(suffix='.mp4', delete=False).name
            rendered_clip.write_videofile(video_file_name, fps=25)

        return 'Done', video_file_name


if __name__ == '__main__':
    demo_runner = DemoRunner()
    demo = gr.Blocks()

    with demo:
        with gr.Row():
            with gr.Column():
                with gr.Row():
                    inp_api_key = gr.Textbox(label='OpenAI API Key', lines=1, value='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')
                    inp_model_name = gr.Dropdown(label='Model Name', choices=['code-davinci-002', 'text-davinci-002'], value='code-davinci-002')
                with gr.Row():
                    inp_n_blocks = gr.Slider(label='Num Blocks', minimum=0, maximum=3, value=3, step=1)
                    inp_n_bowls = gr.Slider(label='Num Bowls', minimum=0, maximum=3, value=3, step=1)
            
                btn_setup = gr.Button("1) Setup/Reset Env")
                info_setup = gr.Markdown(label='Setup Info')
            with gr.Column():
                img_setup = gr.Image(label='Setup Image')

        with gr.Row():
            with gr.Column():
                
                inp_instruction = gr.Textbox(label='Instruction', lines=1) 
                btn_run = gr.Button("2) Run Instruction")
                info_run = gr.Label(label='Run Info')
            with gr.Column():
                video_run = gr.Video(label='Run Video')
        
        btn_setup.click(
            demo_runner.setup, 
            inputs=[inp_api_key, inp_model_name, inp_n_blocks, inp_n_bowls], 
            outputs=[info_setup, img_setup]
        )
        btn_run.click(
            demo_runner.run, 
            inputs=[inp_instruction], 
            outputs=[info_run, video_run]
        )
        
    demo.launch()