Tested with TF 2.3.1

Tested with TF 2.3.1

10_Pong-v0_A3C/Pong-v0_A3C_TF2.py

@@ -0,0 +1,313 @@
+# Tutorial by www.pylessons.com
+# Tutorial written for - Tensorflow 2.3.1
+
+import os
+#os.environ['TF_CPP_MIN_LOG_LEVEL'] = '2'
+os.environ['CUDA_VISIBLE_DEVICES'] = '0'
+import random
+import gym
+import pylab
+import numpy as np
+import tensorflow as tf
+from tensorflow.keras.models import Model, load_model
+from tensorflow.keras.layers import Input, Dense, Lambda, Add, Conv2D, Flatten
+from tensorflow.keras.optimizers import Adam, RMSprop
+from tensorflow.keras import backend as K
+import cv2
+import threading
+from threading import Thread, Lock
+import time
+
+gpus = tf.config.experimental.list_physical_devices('GPU')
+if len(gpus) > 0:
+    print(f'GPUs {gpus}')
+    try: tf.config.experimental.set_memory_growth(gpus[0], True)
+    except RuntimeError: pass
+
+def OurModel(input_shape, action_space, lr):
+    X_input = Input(input_shape)
+
+    #X = Conv2D(32, 8, strides=(4, 4),padding="valid", activation="elu", data_format="channels_first", input_shape=input_shape)(X_input)
+    #X = Conv2D(64, 4, strides=(2, 2),padding="valid", activation="elu", data_format="channels_first")(X)
+    #X = Conv2D(64, 3, strides=(1, 1),padding="valid", activation="elu", data_format="channels_first")(X)
+    X = Flatten(input_shape=input_shape)(X_input)
+
+    X = Dense(512, activation="elu", kernel_initializer='he_uniform')(X)
+    #X = Dense(256, activation="elu", kernel_initializer='he_uniform')(X)
+    #X = Dense(64, activation="elu", kernel_initializer='he_uniform')(X)
+
+    action = Dense(action_space, activation="softmax", kernel_initializer='he_uniform')(X)
+    value = Dense(1, kernel_initializer='he_uniform')(X)
+
+    Actor = Model(inputs = X_input, outputs = action)
+    Actor.compile(loss='categorical_crossentropy', optimizer=RMSprop(lr=lr))
+
+    Critic = Model(inputs = X_input, outputs = value)
+    Critic.compile(loss='mse', optimizer=RMSprop(lr=lr))
+
+    return Actor, Critic
+
+class A3CAgent:
+    # Actor-Critic Main Optimization Algorithm
+    def __init__(self, env_name):
+        # Initialization
+        # Environment and PPO parameters
+        self.env_name = env_name       
+        self.env = gym.make(env_name)
+        self.action_size = self.env.action_space.n
+        self.EPISODES, self.episode, self.max_average = 20000, 0, -21.0 # specific for pong
+        self.lock = Lock()
+        self.lr = 0.000025
+
+        self.ROWS = 80
+        self.COLS = 80
+        self.REM_STEP = 4
+
+        # Instantiate plot memory
+        self.scores, self.episodes, self.average = [], [], []
+
+        self.Save_Path = 'Models'
+        self.state_size = (self.REM_STEP, self.ROWS, self.COLS)
+
+        if not os.path.exists(self.Save_Path): os.makedirs(self.Save_Path)
+        self.path = '{}_A3C_{}'.format(self.env_name, self.lr)
+        self.Model_name = os.path.join(self.Save_Path, self.path)
+
+        # Create Actor-Critic network model
+        self.Actor, self.Critic = OurModel(input_shape=self.state_size, action_space = self.action_size, lr=self.lr)
+
+    def act(self, state):
+        # Use the network to predict the next action to take, using the model
+        prediction = self.Actor.predict(state)[0]
+        action = np.random.choice(self.action_size, p=prediction)
+        return action
+
+    def discount_rewards(self, reward):
+        # Compute the gamma-discounted rewards over an episode
+        gamma = 0.99    # discount rate
+        running_add = 0
+        discounted_r = np.zeros_like(reward)
+        for i in reversed(range(0,len(reward))):
+            if reward[i] != 0: # reset the sum, since this was a game boundary (pong specific!)
+                running_add = 0
+            running_add = running_add * gamma + reward[i]
+            discounted_r[i] = running_add
+
+        discounted_r -= np.mean(discounted_r) # normalizing the result
+        discounted_r /= np.std(discounted_r) # divide by standard deviation
+        return discounted_r
+
+    def replay(self, states, actions, rewards):
+        # reshape memory to appropriate shape for training
+        states = np.vstack(states)
+        actions = np.vstack(actions)
+
+        # Compute discounted rewards
+        discounted_r = self.discount_rewards(rewards)
+
+        # Get Critic network predictions
+        value = self.Critic.predict(states)[:, 0]
+        # Compute advantages
+        advantages = discounted_r - value
+        # training Actor and Critic networks
+        self.Actor.fit(states, actions, sample_weight=advantages, epochs=1, verbose=0)
+        self.Critic.fit(states, discounted_r, epochs=1, verbose=0)
+
+    def load(self, Actor_name, Critic_name):
+        self.Actor = load_model(Actor_name, compile=False)
+        #self.Critic = load_model(Critic_name, compile=False)
+
+    def save(self):
+        self.Actor.save(self.Model_name + '_Actor.h5')
+        #self.Critic.save(self.Model_name + '_Critic.h5')
+
+    pylab.figure(figsize=(18, 9))
+    def PlotModel(self, score, episode):
+        self.scores.append(score)
+        self.episodes.append(episode)
+        self.average.append(sum(self.scores[-50:]) / len(self.scores[-50:]))
+        if str(episode)[-2:] == "00":# much faster than episode % 100
+            pylab.plot(self.episodes, self.scores, 'b')
+            pylab.plot(self.episodes, self.average, 'r')
+            pylab.ylabel('Score', fontsize=18)
+            pylab.xlabel('Steps', fontsize=18)
+            try:
+                pylab.savefig(self.path+".png")
+            except OSError:
+                pass
+
+        return self.average[-1]
+
+    def imshow(self, image, rem_step=0):
+        cv2.imshow(self.Model_name+str(rem_step), image[rem_step,...])
+        if cv2.waitKey(25) & 0xFF == ord("q"):
+            cv2.destroyAllWindows()
+            return
+
+    def GetImage(self, frame, image_memory):
+        if image_memory.shape == (1,*self.state_size):
+            image_memory = np.squeeze(image_memory)
+
+        # croping frame to 80x80 size
+        frame_cropped = frame[35:195:2, ::2,:]
+        if frame_cropped.shape[0] != self.COLS or frame_cropped.shape[1] != self.ROWS:
+            # OpenCV resize function 
+            frame_cropped = cv2.resize(frame, (self.COLS, self.ROWS), interpolation=cv2.INTER_CUBIC)
+
+        # converting to RGB (numpy way)
+        frame_rgb = 0.299*frame_cropped[:,:,0] + 0.587*frame_cropped[:,:,1] + 0.114*frame_cropped[:,:,2]
+
+        # convert everything to black and white (agent will train faster)
+        frame_rgb[frame_rgb < 100] = 0
+        frame_rgb[frame_rgb >= 100] = 255
+        # converting to RGB (OpenCV way)
+        #frame_rgb = cv2.cvtColor(frame_cropped, cv2.COLOR_RGB2GRAY)     
+
+        # dividing by 255 we expresses value to 0-1 representation
+        new_frame = np.array(frame_rgb).astype(np.float32) / 255.0
+
+        # push our data by 1 frame, similar as deq() function work
+        image_memory = np.roll(image_memory, 1, axis = 0)
+
+        # inserting new frame to free space
+        image_memory[0,:,:] = new_frame
+
+        # show image frame   
+        #self.imshow(image_memory,0)
+        #self.imshow(image_memory,1)
+        #self.imshow(image_memory,2)
+        #self.imshow(image_memory,3)
+
+        return np.expand_dims(image_memory, axis=0)
+
+    def reset(self, env):
+        image_memory = np.zeros(self.state_size)
+        frame = env.reset()
+        for i in range(self.REM_STEP):
+            state = self.GetImage(frame, image_memory)
+        return state
+
+    def step(self, action, env, image_memory):
+        next_state, reward, done, info = env.step(action)
+        next_state = self.GetImage(next_state, image_memory)
+        return next_state, reward, done, info
+
+    def run(self):
+        for e in range(self.EPISODES):
+            state = self.reset(self.env)
+            done, score, SAVING = False, 0, ''
+            # Instantiate or reset games memory
+            states, actions, rewards = [], [], []
+            while not done:
+                #self.env.render()
+                # Actor picks an action
+                action = self.act(state)
+                # Retrieve new state, reward, and whether the state is terminal
+                next_state, reward, done, _ = self.step(action, self.env, state)
+                # Memorize (state, action, reward) for training
+                states.append(state)
+                action_onehot = np.zeros([self.action_size])
+                action_onehot[action] = 1
+                actions.append(action_onehot)
+                rewards.append(reward)
+                # Update current state
+                state = next_state
+                score += reward
+                if done:
+                    average = self.PlotModel(score, e)
+                    # saving best models
+                    if average >= self.max_average:
+                        self.max_average = average
+                        self.save()
+                        SAVING = "SAVING"
+                    else:
+                        SAVING = ""
+                    print("episode: {}/{}, score: {}, average: {:.2f} {}".format(e, self.EPISODES, score, average, SAVING))
+
+                    self.replay(states, actions, rewards)
+         # close environemnt when finish training  
+        self.env.close()
+
+    def train(self, n_threads):
+        self.env.close()
+        # Instantiate one environment per thread
+        envs = [gym.make(self.env_name) for i in range(n_threads)]
+
+        # Create threads
+        threads = [threading.Thread(
+                target=self.train_threading,
+                daemon=True,
+                args=(self,
+                    envs[i],
+                    i)) for i in range(n_threads)]
+
+        for t in threads:
+            time.sleep(2)
+            t.start()
+
+        for t in threads:
+            time.sleep(10)
+            t.join()
+
+    def train_threading(self, agent, env, thread):
+        while self.episode < self.EPISODES:
+            # Reset episode
+            score, done, SAVING = 0, False, ''
+            state = self.reset(env)
+            # Instantiate or reset games memory
+            states, actions, rewards = [], [], []
+            while not done:
+                action = agent.act(state)
+                next_state, reward, done, _ = self.step(action, env, state)
+
+                states.append(state)
+                action_onehot = np.zeros([self.action_size])
+                action_onehot[action] = 1
+                actions.append(action_onehot)
+                rewards.append(reward)
+
+                score += reward
+                state = next_state
+
+            self.lock.acquire()
+            self.replay(states, actions, rewards)
+            self.lock.release()
+
+            # Update episode count
+            with self.lock:
+                average = self.PlotModel(score, self.episode)
+                # saving best models
+                if average >= self.max_average:
+                    self.max_average = average
+                    self.save()
+                    SAVING = "SAVING"
+                else:
+                    SAVING = ""
+                print("episode: {}/{}, thread: {}, score: {}, average: {:.2f} {}".format(self.episode, self.EPISODES, thread, score, average, SAVING))
+                if(self.episode < self.EPISODES):
+                    self.episode += 1
+        env.close()            
+
+    def test(self, Actor_name, Critic_name):
+        self.load(Actor_name, Critic_name)
+        for e in range(100):
+            state = self.reset(self.env)
+            done = False
+            score = 0
+            while not done:
+                self.env.render()
+                action = np.argmax(self.Actor.predict(state))
+                state, reward, done, _ = self.step(action, self.env, state)
+                score += reward
+                if done:
+                    print("episode: {}/{}, score: {}".format(e, self.EPISODES, score))
+                    break
+        self.env.close()
+
+if __name__ == "__main__":
+    env_name = 'PongDeterministic-v4'
+    #env_name = 'Pong-v0'
+    agent = A3CAgent(env_name)
+    #agent.run() # use as A2C
+    agent.train(n_threads=5) # use as A3C
+    #agent.test('Models/Pong-v0_A3C_2.5e-05_Actor.h5', '')