Separate train and evaluation functions od ac3 module in two files

a3c/{discrete_A3C.py → eval.py}

@@ -1,46 +1,8 @@
-"""
-Reinforcement Learning (A3C) using Pytroch + multiprocessing.
-The most simple implementation for continuous action.
-
-View more on my Chinese tutorial page [莫烦Python](https://morvanzhou.github.io/).
-"""
 import os
 import torch
-import torch.multiprocessing as mp
-from .shared_adam import SharedAdam
-from .net import Net
-from .utils import v_wrap
-from .worker import Worker
-
-
-def train(env, max_ep, model_checkpoint_dir, pretrained_model_path=None):
-    os.environ["OMP_NUM_THREADS"] = "1"
-    if not os.path.exists(model_checkpoint_dir):
-        os.makedirs(model_checkpoint_dir)
-    n_s = env.observation_space.shape[0]
-    n_a = env.action_space.n
-    words_list = env.words
-    word_width = len(env.words[0])
-    gnet = Net(n_s, n_a, words_list, word_width)        # global network
-    if pretrained_model_path:
-        gnet.load_state_dict(torch.load(pretrained_model_path))
-    gnet.share_memory()         # share the global parameters in multiprocessing
-    opt = SharedAdam(gnet.parameters(), lr=1e-4, betas=(0.92, 0.999))      # global optimizer
-    global_ep, global_ep_r, res_queue, win_ep = mp.Value('i', 0), mp.Value('d', 0.), mp.Queue(), mp.Value('i', 0)
 
-    # parallel training
-    workers = [Worker(max_ep, gnet, opt, global_ep, global_ep_r, res_queue, i, env, n_s, n_a,
-                      words_list, word_width, win_ep, model_checkpoint_dir, pretrained_model_path) for i in range(mp.cpu_count())]
-    [w.start() for w in workers]
-    res = []                    # record episode reward to plot
-    while True:
-        r = res_queue.get()
-        if r is not None:
-            res.append(r)
-        else:
-            break
-    [w.join() for w in workers]
-    return global_ep, win_ep, gnet, res
+from .net import GreedyNet
+from .utils import v_wrap
 
 
 def evaluate_checkpoints(dir, env):
@@ -48,7 +10,7 @@ def evaluate_checkpoints(dir, env):
     n_a = env.action_space.n
     words_list = env.words
     word_width = len(env.words[0])
-    net = Net(n_s, n_a, words_list, word_width)
+    net = GreedyNet(n_s, n_a, words_list, word_width)
     results = {}
     for checkpoint in os.listdir(dir):
         checkpoint_path = os.path.join(dir, checkpoint)

a3c/net.py

@@ -54,3 +54,14 @@ class Net(nn.Module):
         a_loss = -exp_v
         total_loss = (c_loss + a_loss).mean()
         return total_loss
+
+
+class GreedyNet(Net):
+    def choose_action(self, s):
+        self.eval()
+        logits, _ = self.forward(s)
+        probabilities = logits.exp().squeeze(dim=-1)
+        prob_np = probabilities.data.cpu().numpy()
+
+        actions = np.argmax(prob_np, axis=1)
+        return actions[0]

a3c/train.py

@@ -0,0 +1,36 @@
+import os
+import torch
+import torch.multiprocessing as mp
+from .shared_adam import SharedAdam
+from .net import Net
+from .worker import Worker
+
+
+def train(env, max_ep, model_checkpoint_dir, pretrained_model_path=None):
+    os.environ["OMP_NUM_THREADS"] = "1"
+    if not os.path.exists(model_checkpoint_dir):
+        os.makedirs(model_checkpoint_dir)
+    n_s = env.observation_space.shape[0]
+    n_a = env.action_space.n
+    words_list = env.words
+    word_width = len(env.words[0])
+    gnet = Net(n_s, n_a, words_list, word_width)  # global network
+    if pretrained_model_path:
+        gnet.load_state_dict(torch.load(pretrained_model_path))
+    gnet.share_memory()  # share the global parameters in multiprocessing
+    opt = SharedAdam(gnet.parameters(), lr=1e-4, betas=(0.92, 0.999))  # global optimizer
+    global_ep, global_ep_r, res_queue, win_ep = mp.Value('i', 0), mp.Value('d', 0.), mp.Queue(), mp.Value('i', 0)
+
+    # parallel training
+    workers = [Worker(max_ep, gnet, opt, global_ep, global_ep_r, res_queue, i, env, n_s, n_a,
+                      words_list, word_width, win_ep, model_checkpoint_dir, pretrained_model_path) for i in range(mp.cpu_count())]
+    [w.start() for w in workers]
+    res = []  # record episode reward to plot
+    while True:
+        r = res_queue.get()
+        if r is not None:
+            res.append(r)
+        else:
+            break
+    [w.join() for w in workers]
+    return global_ep, win_ep, gnet, res

main.py

@@ -6,7 +6,8 @@ import os
 import sys
 import time
 import matplotlib.pyplot as plt
-from a3c.discrete_A3C import train, evaluate, evaluate_checkpoints
+from a3c.train import train
+from a3c.eval import evaluate, evaluate_checkpoints
 from wordle_env.wordle import WordleEnvBase