From 493d3610220b011349203ef04fe5d589c6c93689 Mon Sep 17 00:00:00 2001
From: Tongzheng Ren <rtz19970824@gmail.com>
Date: Sun, 5 Nov 2017 15:30:35 +0800
Subject: [PATCH] modified the network

---
 AlphaGo/Network.py | 90 ++++++++++++++++++++++++++--------------------
 1 file changed, 51 insertions(+), 39 deletions(-)

diff --git a/AlphaGo/Network.py b/AlphaGo/Network.py
index c98c6e2..f594be2 100644
--- a/AlphaGo/Network.py
+++ b/AlphaGo/Network.py
@@ -4,13 +4,6 @@ import time
 import multi_gpu
 import tensorflow.contrib.layers as layers
 
-data = np.load("data.npz")
-boards = data["boards"]
-wins = data["wins"]
-ps = data["ps"]
-print (boards.shape)
-print (wins.shape)
-print (ps.shape)
 def residual_block(input, is_training):
 	normalizer_params = {'is_training': is_training,
 						 'updates_collections': None}
@@ -44,7 +37,6 @@ def value_heads(input, is_training):
 	h = layers.fully_connected(h, 1, activation_fn=tf.nn.tanh, weights_regularizer=layers.l2_regularizer(1e-4))
 	return h
 
-
 x = tf.placeholder(tf.float32,shape=[None,19,19,17])
 is_training = tf.placeholder(tf.bool, shape=[])
 z = tf.placeholder(tf.float32, shape=[None, 1])
@@ -62,34 +54,54 @@ train_op = tf.train.RMSPropOptimizer(1e-2).minimize(total_loss)
 
 var_list = tf.get_collection(tf.GraphKeys.TRAINABLE_VARIABLES)
 saver = tf.train.Saver(max_to_keep=10, var_list=var_list)
-epochs = 100
-batch_size = 32
-batch_num = boards.shape[0] // batch_size
-result_path = "./results/"
-with multi_gpu.create_session() as sess:
-	sess.run(tf.global_variables_initializer())
-	ckpt_file = tf.train.latest_checkpoint(result_path)
-	if ckpt_file is not None:
-		print('Restoring model from {}...'.format(ckpt_file))
-		saver.restore(sess, ckpt_file)
-	for epoch in range(epochs):
-		time_train = -time.time()
-		index = np.arange(boards.shape[0])
-		np.random.shuffle(index)
-		losses = []
-		regs = []
-		for iter in range(batch_num):
-			_, l, r, value, prob = sess.run([train_op, loss, reg, v, p], feed_dict={x:boards[index[iter*batch_size:(iter+1)*batch_size]],
-																				z:wins[index[iter*batch_size:(iter+1)*batch_size]],
-																				pi:ps[index[iter*batch_size:(iter+1)*batch_size]],
-																				is_training:True})
-			losses.append(l)
-			regs.append(r)
-			if iter % 1 == 0:
-				print("Epoch: {}, Iteration: {}, Time: {}, Loss: {}, Reg: {}".format(epoch, iter, time.time()+time_train, np.mean(np.array(losses)), np.mean(np.array(regs))))
-				time_train=-time.time()
-				losses = []
-				regs = []
-			if iter % 20 == 0:
-				save_path = "Epoch{}.Iteration{}.ckpt".format(epoch, iter)
-				saver.save(sess, result_path + save_path)
+def train():
+	data = np.load("data.npz")
+	boards = data["boards"]
+	wins = data["wins"]
+	ps = data["ps"]
+	print (boards.shape)
+	print (wins.shape)
+	print (ps.shape)
+	epochs = 100
+	batch_size = 32
+	batch_num = boards.shape[0] // batch_size
+	result_path = "./results/"
+	with multi_gpu.create_session() as sess:
+		sess.run(tf.global_variables_initializer())
+		ckpt_file = tf.train.latest_checkpoint(result_path)
+		if ckpt_file is not None:
+			print('Restoring model from {}...'.format(ckpt_file))
+			saver.restore(sess, ckpt_file)
+		for epoch in range(epochs):
+			time_train = -time.time()
+			index = np.arange(boards.shape[0])
+			np.random.shuffle(index)
+			losses = []
+			regs = []
+			for iter in range(batch_num):
+				_, l, r, value, prob = sess.run([train_op, loss, reg, v, p], feed_dict={x:boards[index[iter*batch_size:(iter+1)*batch_size]],
+																					z:wins[index[iter*batch_size:(iter+1)*batch_size]],
+																					pi:ps[index[iter*batch_size:(iter+1)*batch_size]],
+																					is_training:True})
+				losses.append(l)
+				regs.append(r)
+				if iter % 1 == 0:
+					print("Epoch: {}, Iteration: {}, Time: {}, Loss: {}, Reg: {}".format(epoch, iter, time.time()+time_train, np.mean(np.array(losses)), np.mean(np.array(regs))))
+					time_train=-time.time()
+					losses = []
+					regs = []
+				if iter % 20 == 0:
+					save_path = "Epoch{}.Iteration{}.ckpt".format(epoch, iter)
+					saver.save(sess, result_path + save_path)
+
+def forward(board):
+	result_path = "./results/"
+	with multi_gpu.create_session() as sess:
+		sess.run(tf.global_variables_initializer())
+		ckpt_file = tf.train.latest_checkpoint(result_path)
+		if ckpt_file is not None:
+			print('Restoring model from {}...'.format(ckpt_file))
+			saver.restore(sess, ckpt_file)
+		else:
+			raise ValueError("No model loaded")
+		return sess.run([p,v], feed_dict={x:board})
\ No newline at end of file