Feat: TNPG (but it has strange point. Not perfect)

Author: choru-k

PG/5-TNPG/config.py

@@ -0,0 +1,8 @@
+import torch
+
+env_name = 'CartPole-v1'
+gamma = 0.99
+lr = 0.01
+goal_score = 200
+log_interval = 10
+device = torch.device("cuda" if torch.cuda.is_available() else "cpu")

PG/5-TNPG/model.py

@@ -0,0 +1,130 @@
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+import numpy as np
+
+from config import gamma, lr
+
+def flat_grad(grads):
+    grad_flatten = []
+    for grad in grads:
+        grad_flatten.append(grad.view(-1))
+    grad_flatten = torch.cat(grad_flatten)
+    return grad_flatten
+
+def flat_hessian(hessians):
+    hessians_flatten = []
+    for hessian in hessians:
+        hessians_flatten.append(hessian.contiguous().view(-1))
+    hessians_flatten = torch.cat(hessians_flatten).data
+    return hessians_flatten
+
+def flat_params(model):
+    params = []
+    for param in model.parameters():
+        params.append(param.data.view(-1))
+    params_flatten = torch.cat(params)
+    return params_flatten
+
+def update_model(model, new_params):
+    index = 0
+    for params in model.parameters():
+        params_length = len(params.view(-1))
+        new_param = new_params[index: index + params_length]
+        new_param = new_param.view(params.size())
+        params.data.copy_(new_param)
+        index += params_length
+
+def kl_divergence(net, old_net, states):
+    policy = net(states)
+    old_policy = old_net(states).detach()
+
+    kl = old_policy * torch.log(old_policy / policy)
+
+    kl = kl.sum(1, keepdim=True)
+    return kl
+
+def fisher_vector_product(net, states, p, cg_damp=0.1):
+    kl = kl_divergence(net, net, states)
+    kl = kl.mean()
+    kl_grad = torch.autograd.grad(kl, net.parameters(), create_graph=True) # create_graph is True if we need higher order derivative products
+    kl_grad = flat_grad(kl_grad)
+
+    kl_grad_p = (kl_grad * p.detach()).sum()
+    kl_hessian_p = torch.autograd.grad(kl_grad_p, net.parameters())
+    kl_hessian_p = flat_hessian(kl_hessian_p)
+
+    return kl_hessian_p + cg_damp * p.detach()
+
+
+def conjugate_gradient(net, states, loss_grad, n_step=10, residual_tol=1e-10):
+    x = torch.zeros(loss_grad.size())
+    r = loss_grad.clone()
+    p = loss_grad.clone()
+    r_dot_r = torch.dot(r, r)
+
+    for i in range(n_step):
+        A_dot_p = fisher_vector_product(net, states, p)
+        alpha = r_dot_r / torch.dot(p, A_dot_p)
+        x += alpha * p
+        r -= alpha * A_dot_p
+        new_r_dot_r = torch.dot(r,r)
+        betta = new_r_dot_r / r_dot_r
+        p = r + betta * p
+        r_dot_r = new_r_dot_r
+        if r_dot_r < residual_tol:
+            break
+    return x
+
+class QNet(nn.Module):
+    def __init__(self, num_inputs, num_outputs):
+        super(QNet, self).__init__()
+        self.t = 0
+        self.num_inputs = num_inputs
+        self.num_outputs = num_outputs
+
+        self.fc_1 = nn.Linear(num_inputs, 128)
+        self.fc_2 = nn.Linear(128, num_outputs)
+
+        for m in self.modules():
+            if isinstance(m, nn.Linear):
+                nn.init.xavier_uniform(m.weight)
+
+    def forward(self, input):
+        x = torch.tanh(self.fc_1(input))
+        policy = F.softmax(self.fc_2(x))
+
+        return policy
+
+    @classmethod
+    def train_model(cls, net, transitions, k):
+        states, actions, rewards, masks = transitions
+        states = torch.stack(states)
+        actions = torch.stack(actions)
+        rewards = torch.Tensor(rewards)
+        masks = torch.Tensor(masks)
+
+        policy = net(states)
+        policy = policy.view(-1, net.num_outputs)
+        policy_action = (policy * actions.detach()).sum(dim=1)
+
+        loss = (policy_action * rewards).mean()
+
+        loss_grad = torch.autograd.grad(loss, net.parameters())
+        loss_grad = flat_grad(loss_grad)
+
+        step_dir = conjugate_gradient(net, states, loss_grad.data)
+
+        params = flat_params(net)
+        new_params = params + lr * step_dir
+
+        update_model(net, new_params)
+
+        return -loss
+
+    def get_action(self, input):
+        policy = self.forward(input)
+        policy = policy[0].data.numpy()
+
+        action = np.random.choice(self.num_outputs, 1, p=policy)[0]
+        return action

PG/5-TNPG/train.py

@@ -0,0 +1,90 @@
+import os
+import sys
+import gym
+import random
+import numpy as np
+
+import torch
+import torch.optim as optim
+import torch.nn.functional as F
+from model import QNet
+from tensorboardX import SummaryWriter
+
+from config import env_name, goal_score, log_interval, device, lr, gamma
+
+
+def main():
+    env = gym.make(env_name)
+    env.seed(500)
+    torch.manual_seed(500)
+
+    num_inputs = env.observation_space.shape[0]
+    num_actions = env.action_space.n
+    print('state size:', num_inputs)
+    print('action size:', num_actions)
+
+    net = QNet(num_inputs, num_actions)
+
+    writer = SummaryWriter('logs')
+
+    net.to(device)
+    net.train()
+    running_score = 0
+    steps = 0
+    loss = 0
+    k=0
+    for e in range(3000):
+        done = False
+        memory = []
+
+        score = 0
+        state = env.reset()
+        state = torch.Tensor(state).to(device)
+        state = state.unsqueeze(0)
+
+        while not done:
+            steps += 1
+
+            action = net.get_action(state)
+            next_state, reward, done, _ = env.step(action)
+
+            next_state = torch.Tensor(next_state)
+            next_state = next_state.unsqueeze(0)
+
+            mask = 0 if done else 1
+            reward = reward if not done or score == 499 else -1
+
+            action_one_hot = torch.zeros(2)
+            action_one_hot[action] = 1
+            memory.append([state, next_state, action_one_hot, reward, mask])
+
+            score += reward
+            state = next_state
+
+        sum_reward = 0
+        memory.reverse()
+        states, actions, rewards, masks = [], [], [], []
+        for t, transition in enumerate(memory):
+            state, next_state, action, reward, mask = transition
+            sum_reward = (reward + gamma * sum_reward)
+            states.append(state)
+            actions.append(action)
+            rewards.append(sum_reward)
+            masks.append(mask)
+        loss = QNet.train_model(net, (states, actions, rewards, masks), k)
+        k+=1
+
+        score = score if score == 500.0 else score + 1
+        running_score = 0.99 * running_score + 0.01 * score
+        if e % log_interval == 0:
+            print('{} episode | score: {:.2f}'.format(
+                e, running_score))
+            writer.add_scalar('log/score', float(running_score), e)
+            writer.add_scalar('log/loss', float(loss), e)
+
+        if running_score > goal_score:
+            break
+
+
+if __name__=="__main__":
+    main()

README.md

@@ -20,10 +20,10 @@ So you can run this example in your computer(maybe it take just only 1~2 minitue
 - [x] Actor Critic [[10]](#reference)
 - [x] Advantage Actor Critic
 - [x] GAE(Generalized Advantage Estimation) [[12]](#reference)
-- [ ] ACER [[21]](#reference)
-- [ ] NPG [[20]](#reference)
+- [ ] TNPG [[20]](#reference)
 - [ ] TRPO [[13]](#reference)
 - [ ] PPO [[14]](#reference)
+- [ ] ACER [[21]](#reference)
 
 ## Parallel
 - [x] Asynchronous Q-learning [[11]](#reference)