Updated server code.

Experimentation

See merge request Pownie/backgammon!8

app.py

@@ -2,6 +2,7 @@ from flask import Flask, request, jsonify
 from flask_json import FlaskJSON, as_json_p
 from flask_cors import CORS
 from board import Board
+from eval import Eval
 import main
 import random
 from network import Network
@@ -17,8 +18,8 @@ CORS(app)
 
 config = main.config.copy()
 config['model'] = "player_testings"
-config['ply'] = "1"
+config['ply'] = "0"
-config['board_representation'] = 'quack-fat'
+config['board_representation'] = 'tesauro'
 network = Network(config, config['model'])
 
 network.restore_model()
@@ -90,11 +91,16 @@ def bot_move():
     data = request.get_json(force=True)
 
     board = [int(x) for x in data['board'].split(',')]
+    use_pubeval = bool(data['pubeval'])
 
     roll = (random.randrange(1, 7), random.randrange(1, 7))
-    # print(roll)
+
+    if use_pubeval:
+        board, value = Eval.make_pubeval_move(tuple(board), 1, roll)
+    else:
         board, _ = network.make_move(tuple(board), roll, 1)
-    # print("Boards!:",board)
+
+    # print("Board!:",board)
 
     return ",".join([str(x) for x in list(board)])
 

bin/0-ply-tests.rb

@@ -50,7 +50,7 @@ run_stuff(board_rep, model_name, ply)
 
 board_rep = "quack-norm"
 model_name = "quack-norm_test_0_ply"
-ply = 1
+ply = 0
 
 run_stuff(board_rep, model_name, ply)
 
@@ -67,7 +67,7 @@ run_stuff(board_rep, model_name, ply)
 
 board_rep = "tesauro"
 model_name = "tesauro_test_0_ply"
-ply = 1
+ply = 0
 
 run_stuff(board_rep, model_name, ply)
 

bin/train-evaluate-save

@@ -1,30 +1,30 @@
 #!/usr/bin/env ruby
+MODELS_DIR = 'models'
+
 def save(model_name)
   require 'date'
 
-  models_dir = 'models'
+  model_path = File.join(MODELS_DIR, model_name)
-  model_path = File.join(models_dir, model_name)
-  if not File.exists? model_path then
-    return false
-  end
 
   episode_count = (File.read File.join(model_path, 'episodes_trained')).to_i
 
   puts "Found model #{model_name} with episodes #{episode_count} trained!"
 
   file_name = "model-#{model_name}-#{episode_count}-#{Time.now.strftime('%Y%m%d-%H%M%S')}.tar.gz"
-  save_path = File.join(models_dir, 'saves', file_name)
+  save_path = File.join(MODELS_DIR, 'saves', file_name)
   puts "Saving to #{save_path}"
   
-  system("tar", "-cvzf", save_path, "-C", models_dir, model_name)
+  system("tar", "-cvzf", save_path, "-C", MODELS_DIR, model_name)
-
-  return true
 end
 
 def train(model, episodes)
   system("python3", "main.py", "--train", "--model", model, "--episodes", episodes.to_s)
 end
 
+def force_train(model, episodes)
+  system("python3", "main.py", "--train", "--force-creation", "--model", model, "--episodes", episodes.to_s)
+end
+
 def evaluate(model, episodes, method)
   system("python3", "main.py", "--eval" , "--model", model, "--episodes", episodes.to_s, "--eval-methods", method)
 end
@@ -33,11 +33,9 @@ model = ARGV[0]
 
 if model.nil? then raise "no model specified" end
 
-while true do
+if not File.exists? File.join(MODELS_DIR, model) then
+  force_train model, 10
   save model
-  train model, 1000
-  save model
-  train model, 1000
   3.times do
     evaluate model, 250, "pubeval"
   end
@@ -45,3 +43,27 @@ while true do
     evaluate model, 250, "dumbeval"
   end
 end
+
+# while true do
+#   save model
+#   train model, 1000
+#   save model
+#   train model, 1000
+#   3.times do
+#     evaluate model, 250, "pubeval"
+#   end
+#   3.times do
+#     evaluate model, 250, "dumbeval"
+#   end
+# end
+
+while true do
+  save model
+  train model, 500
+  5.times do
+    evaluate model, 250, "pubeval"
+  end
+  5.times do
+    evaluate model, 250, "dumbeval"
+  end
+end

board.py

@@ -16,8 +16,6 @@ class Board:
         return quack.idxs_with_checkers_of_player(board, player)
 
     
-    # TODO: Write a test for this
-    # TODO: Make sure that the bars fit, 0 represents the -1 player and 25 represents the 1 player
     # index 26 is player 1 home, index 27 is player -1 home
     @staticmethod
     def board_features_to_pubeval(board, player):
@@ -249,9 +247,6 @@ class Board:
         # Find all points with checkers on them belonging to the player
         # Iterate through each index and check if it's a possible move given the roll
 
-        # TODO: make sure that it is not possible to do nothing on first part of
-        #       turn and then do something with the second die
-        
         def calc_moves(board, face_value):
             if face_value == 0:
                 return [board]
@@ -273,23 +268,13 @@ class Board:
         # print("Dice permuts:",dice_permutations)
         for roll in dice_permutations:
             # Calculate boards resulting from first move
-            #print("initial board: ", board)
-            #print("roll:", roll)
-            #print("Rest of roll:",roll[1:])
             boards = calc_moves(board, roll[0])
-            #print("Boards:",boards)
-            #print("Roll:",roll[0])
-            #print("boards after first die: ", boards)
 
             for die in roll[1:]:
                 # Calculate boards resulting from second move
                 nested_boards = [calc_moves(board, die) for board in boards]
-                #print("nested boards: ", nested_boards)
                 boards = [board for boards in nested_boards for board in boards]
-                # What the fuck
+
-                #for board in boards:
-                #    print(board)
-                #    print("type__:",type(board))
                 # Add resulting unique boards to set of legal boards resulting from roll
 
                 #print("printing boards from calculate_legal_states: ", boards)

main.py

@@ -209,6 +209,8 @@ if __name__ == "__main__":
 
     elif args.eval:
         network = Network(config, config['model'])
+        network.restore_model()
+
         for i in range(int(config['repeat_eval'])):
             start_episode = network.episodes_trained
             # Evaluation measures are described in `config`

network.py

@@ -114,12 +114,11 @@ class Network:
 
         with tf.GradientTape() as tape:
             value = self.model(prev_state.reshape(1,-1))
+
         grads = tape.gradient(value, self.model.variables)
 
         difference_in_values = tf.reshape(tf.subtract(value_next, value, name='difference_in_values'), [])
-        tf.summary.scalar("difference_in_values", tf.abs(difference_in_values))
 
-        with tf.variable_scope('apply_gradients'):
         for grad, train_var in zip(grads, self.model.variables):
             backprop_calc = self.learning_rate * difference_in_values * grad
             train_var.assign_add(backprop_calc)
@@ -299,7 +298,7 @@ class Network:
         length_list = []
         test_list = []
         # Prepping of data
-        start = time.time()
+        # start = time.time()
         for board in boards:
             length = 0
             for roll in all_rolls:
@@ -478,7 +477,6 @@ class Network:
         for episode in range(1, episodes + 1):
 
             sys.stderr.write("[TRAIN] Episode {}".format(episode + trained_eps))
-            # TODO decide which player should be here
 
             # player = 1
             player = random.choice([-1,1])

test.py

@@ -737,6 +737,23 @@ class TestBoardFlip(unittest.TestCase):
         self.assertTrue((Board.board_features_tesauro(board, 1) ==
                          np.array(expected).reshape(1, 198)).all())
 
+    def test_pubeval_features(self):
+        board = Board.initial_state
+
+        expected = (0,
+                    2, 0, 0, 0, 0, -5,
+                    0, -3, 0, 0, 0, 5,
+                    -5, 0, 0, 0, 3, 0,
+                    5, 0, 0, 0, 0, -2,
+                    0,
+                    0, 0)
+
+        import numpy as np
+        self.assertTrue((Board.board_features_to_pubeval(board, 1) ==
+                         np.array(expected).reshape(1, 28)).all())
+        self.assertTrue((Board.board_features_to_pubeval(board, -1) ==
+                         np.array(expected).reshape(1, 28)).all())
+
     def test_tesauro_bars(self):
         board = list(Board.initial_state)
         board[1] = 0