app.py

import datetime
import os
import random
import time

import plotly.figure_factory as ff
import json

import pandas as pd

from compiled_jss.CPEnv import CompiledJssEnvCP

from stable_baselines3.common.vec_env import VecEnvWrapper
from torch.distributions import Categorical

import torch
import numpy as np

from MyDummyVecEnv import MyDummyVecEnv

import gradio as gr


class VecPyTorch(VecEnvWrapper):

    def __init__(self, venv, device):
        super(VecPyTorch, self).__init__(venv)
        self.device = device

    def reset(self):
        return self.venv.reset()

    def step_async(self, actions):
        self.venv.step_async(actions)

    def step_wait(self):
        return self.venv.step_wait()


def make_env(seed, instance):
    def thunk():
        _env = CompiledJssEnvCP(instance)
        return _env

    return thunk


def solve(file, num_workers, seed):
    seed = int(abs(seed))
    random.seed(seed)
    np.random.seed(seed)
    torch.manual_seed(seed)
    with torch.inference_mode():
        device = torch.device('cpu')
        actor = torch.jit.load('actor.pt', map_location=device)
        actor.eval()
        start_time = time.time()
        fn_env = [make_env(0, file.name)
                  for _ in range(num_workers)]
        async_envs = MyDummyVecEnv(fn_env, device)
        envs = VecPyTorch(async_envs, device)
        current_solution_cost = float('inf')
        current_solution = ''
        obs = envs.reset()
        total_episode = 0
        while total_episode < envs.num_envs:
            logits = actor(obs['interval_rep'], obs['attention_interval_mask'], obs['job_resource_mask'],
                           obs['action_mask'], obs['index_interval'], obs['start_end_tokens'])
            # temperature vector
            if num_workers >= 4:
                temperature = torch.arange(0.5, 2.0, step=(1.5 / num_workers), device=device)
            else:
                temperature = torch.ones(num_workers, device=device)
            logits = logits / temperature[:, None]
            probs = Categorical(logits=logits).probs
            # random sample based on logits
            actions = torch.multinomial(probs, probs.shape[1]).cpu().numpy()
            obs, reward, done, infos = envs.step(actions)
            total_episode += done.sum()
            # total_actions += 1
            # print(f'Episode {total_episode} / {envs.num_envs} - Actions {total_actions}', end='\r')
            for env_idx, info in enumerate(infos):
                if 'makespan' in info and int(info['makespan']) < current_solution_cost:
                    current_solution_cost = int(info['makespan'])
                    current_solution = json.loads(info['solution'])
        total_time = time.time() - start_time
        pretty_output = ""
        for job_id in range(len(current_solution)):
            pretty_output += f"Job {job_id}: {current_solution[job_id]}\n"

        jobs_data = []
        file.seek(0)
        line_str: str = file.readline()
        line_cnt: int = 1
        jobs_count: int = 0
        machines_count: int = 0
        while line_str:
            data = []
            split_data = line_str.split()
            if line_cnt == 1:
                jobs_count, machines_count = int(split_data[0]), int(
                    split_data[1]
                )
            else:
                i = 0
                this_job_op_count = 0
                while i < len(split_data):
                    machine, op_time = int(split_data[i]), int(split_data[i + 1])
                    data.append((machine, op_time))
                    i += 2
                    this_job_op_count += 1
                jobs_data.append(data)
            line_str = file.readline()
            line_cnt += 1
        # convert to integer the current_solution
        current_solution = [[int(x) for x in y] for y in current_solution]
        df = []
        for job_id in range(jobs_count):
            for task_id in range(len(current_solution[job_id])):
                dict_op = dict()
                dict_op["Task"] = "Job {}".format(job_id)
                start_sec = current_solution[job_id][task_id]
                finish_sec = start_sec + jobs_data[job_id][task_id][1]
                dict_op["Start"] = datetime.datetime.fromtimestamp(start_sec)
                dict_op["Finish"] = datetime.datetime.fromtimestamp(finish_sec)
                dict_op["Resource"] = "Machine {}".format(
                    jobs_data[job_id][task_id][0]
                )
                df.append(dict_op)
                i += 1
        fig = None
        colors = [
            tuple([random.random() for _ in range(3)]) for _ in range(machines_count)
        ]
        if len(df) > 0:
            df = pd.DataFrame(df)
            fig = ff.create_gantt(
                df,
                index_col="Resource",
                colors=colors,
                show_colorbar=True,
                group_tasks=True,
            )
            fig.update_yaxes(
                autorange=True
            )
        return current_solution_cost, str(total_time) + " seconds", pretty_output, fig


title = "Job-Shop Scheduling CP environment with RL dispatching"
description = """A Job-Shop Scheduling Reinforcement Learning based solver using an underlying CP model as an 
environment. <br> 
For fast inference, 
check out the cached examples below.<br> Any Job-Shop Scheduling instance following the standard specification is 
compatible. <a href='http://jobshop.jjvh.nl/index.php'>Check out this website for more instances</a>.<br>
Increasing the number of workers will provide better solutions, but will slow down the solving time.
This behavior is different than the one from the paper repository as here agents are run sequentially, 
whereas we run agents in parallel (technical limitation due to the platform here). <br>
<br>
For large instance, we recommend running the approach locally outside the interface, as it causes a lot 
of overhead and the resource available on this platform are low (1 vCPU and no GPU).<br> """

article = "<p style='text-align: center'>Article Under Review</p>"
# list all non-hidden files in the 'instances' directory
examples = [['instances/' + f, 16, 0] for f in os.listdir('instances') if not f.startswith('.')]
iface = gr.Interface(fn=solve,
                     inputs=[gr.File(label="Instance File"),
                             gr.Slider(8, 32, value=16, label="Number of Workers", step=1),
                             gr.Number(0, label="Random Seed", precision=0)],
                     outputs=[gr.Text(label="Makespan"), gr.Text(label="Elapsed Time"), gr.Text(label="Solution"),
                              gr.Plot(label="Solution's Gantt Chart")],
                     title=title,
                     description=description,
                     article=article,
                     examples=examples,
                     allow_flagging="never")
iface.launch(enable_queue=True)