improvements

MyDummyVecEnv.py

@@ -7,6 +7,8 @@ import numpy as np
 from stable_baselines3.common.vec_env.base_vec_env import VecEnv, VecEnvIndices, VecEnvObs, VecEnvStepReturn
 from stable_baselines3.common.vec_env.util import dict_to_obs, obs_space_info
 
+import torch
+
 
 class MyDummyVecEnv(VecEnv):
     """
@@ -20,14 +22,16 @@ class MyDummyVecEnv(VecEnv):
         that return environments to vectorize
     """
 
-    def __init__(self, env_fns: List[Callable[[], gym.Env]]):
+    def __init__(self, env_fns: List[Callable[[], gym.Env]], device):
         self.envs = [fn() for fn in env_fns]
         env = self.envs[0]
         VecEnv.__init__(self, len(env_fns), env.observation_space, env.action_space)
         obs_space = env.observation_space
         self.keys, shapes, dtypes = obs_space_info(obs_space)
+        self.device = device
 
-        self.buf_obs = OrderedDict([(k, np.zeros((self.num_envs,) + tuple(shapes[k]), dtype=dtypes[k])) for k in self.keys])
+        self.buf_obs = OrderedDict(
+            [(k, torch.zeros((self.num_envs,) + tuple(shapes[k]), dtype=torch.float, device=self.device)) for k in self.keys])
         self.buf_dones = np.zeros((self.num_envs,), dtype=bool)
         self.buf_rews = np.zeros((self.num_envs,), dtype=np.float32)
         self.buf_infos = [{} for _ in range(self.num_envs)]
@@ -86,10 +90,7 @@ class MyDummyVecEnv(VecEnv):
 
     def _save_obs(self, env_idx: int, obs: VecEnvObs) -> None:
         for key in self.keys:
-            if key is None:
-                self.buf_obs[key][env_idx] = obs
-            else:
-                self.buf_obs[key][env_idx] = obs[key]
+            self.buf_obs[key][env_idx] = torch.from_numpy(obs[key]).to(self.device, non_blocking=True)
 
     def _obs_from_buf(self) -> VecEnvObs:
         return dict_to_obs(self.observation_space, self.buf_obs)

app.py

@@ -3,6 +3,7 @@ import os
 import random
 import time
 
+import gym
 import plotly.figure_factory as ff
 import json
 
@@ -17,6 +18,7 @@ from torch.distributions import Categorical
 import torch
 import numpy as np
 
+from MyDummyVecEnv import MyDummyVecEnv
 from MyVecEnv import WrapperRay
 
 import gradio as gr
@@ -46,11 +48,11 @@ def make_env(seed, instance):
     return thunk
 
 
-def solve(file):
-    random.seed(0)
-    np.random.seed(0)
-    torch.manual_seed(0)
-    num_workers = 1 # only one CPU available
+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)
@@ -58,9 +60,8 @@ def solve(file):
         start_time = time.time()
         fn_env = [make_env(0, file.name)
                   for _ in range(num_workers)]
-        ray_wrapper_env = WrapperRay(lambda n: fn_env[n](),
-                                     num_workers, 1, device)
-        envs = VecPyTorch(ray_wrapper_env, device)
+        async_envs = MyDummyVecEnv(fn_env, device)
+        envs = VecPyTorch(async_envs, device)
         current_solution_cost = float('inf')
         current_solution = ''
         obs = envs.reset()
@@ -146,24 +147,31 @@ def solve(file):
             fig.update_yaxes(
                 autorange=True
             )
-        return current_solution_cost,  str(total_time) + " seconds", pretty_output, fig
-
-ray.init(log_to_driver=False,
-        ignore_reinit_error=True,
-        include_dashboard=False)
-title = "Job-Shop Scheduling CP RL"
-description = """A Job-Shop Scheduling Reinforcement Learning based solver using an underlying CP model as an environment. <br>
-                However, the results you obtain here don't represent the full potential of the approach due to resource limitations on the HuggingFace platform (a single vCPU available, no GPU).<br>
-                We recommend running this locally outside the interface for large instances, as it causes a lot of overhead.<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>."""
+        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 for f in os.listdir('instances') if not f.startswith('.')]
+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"),
-                     outputs=[gr.Text(label="Makespan"), gr.Text(label="Elapsed Time"), gr.Text(label="Solution"), gr.Plot(label="Solution's Gantt Chart")],
+                     inputs=[gr.File(label="Instance File"),
+                             gr.Slider(8, 32, value=16, label="Number of Workers", step=1),
+                             gr.Number(0, label="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,