BattleSnakeArena/tensor_observer.py at master · conlan/BattleSnakeArena · GitHub

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
import numpy as np
import constants

from observer import Observer

class TensorObserver(Observer):
    def __init__(self):
        super().__init__()

    # translates our XY (x is left to right, y is top to bottom) to the tensor's XY
    def assign_tensor_val(self, tensor, layer, x, y, val):
        tensor[layer][y][x] = val

    def convert_data_to_tensor_no_mirror(self, data) -> np.ndarray:
        return self.convert_data_to_tensor(data, False)

    def convert_data_to_tensor(self, data, can_mirror) -> np.ndarray:
        # initialize a 13 x 11 x 11 ndarray

        # layer0: snake health on heads {0,...,100}
        # layer1: snake bodies {0,1}
        # layer2: body segment numbers {0,...,255}
        # layer3: snake length >= player {0,1}
        # layer4: food {0,1}
        # layer5: head_mask {0,1}
        # layer6: double_tail_mask {0,1}
        # layer7: snake bodies >= us {0,1}
        # layer8: snake bodies < us {0,1}
        # layer9-12: alive count mask

        snakes = data["board"]["snakes"]

        pov_snake = data["you"]
        pov_snake_id = pov_snake["id"]

        food = data["board"]["food"]

        player_size = len(pov_snake["body"])

        # print(snakes)
        # print(pov_snake)
        # print(data)

        board_tensor = np.zeros((13, 11, 11), dtype=float)

        # which direction the snake is pointing in
        pov_snake_orientation = None

        # check if there's a snake in the snakes list with pov snake's id
        pov_snake_is_alive = (pov_snake_id in [s["id"] for s in snakes])

        # if the pov snake is dead, we don't need to do anything else
        # Terminal state is an all-zero tensor
        if not pov_snake_is_alive:
            return {
                "tensor" : board_tensor,
                "is_mirror" : False
            }

        for s in snakes:
            snake_id = s["id"]
            health = s ["health"]
            body = s["body"]
            head_x, head_y = body[0]['x'], body[0]['y']

            snake_size = len(body)

            # layer0: snake health on heads {0,...,100}
            self.assign_tensor_val(board_tensor, 0, head_x, head_y, health * 1.0 / constants.MAX_SNAKE_HEALTH)

            if (snake_id == pov_snake_id):
                # layer5: head_mask {0,1}
                self.assign_tensor_val(board_tensor, 5, head_x, head_y, 1)

                neck_x, neck_y = body[1]['x'], body[1]['y']
                if (head_x == neck_x):
                    if (head_y < neck_y):
                        pov_snake_orientation = 'up'
                    elif (head_y > neck_y):
                        pov_snake_orientation = 'down'
                elif (head_y == neck_y):
                    if (head_x < neck_x):
                        pov_snake_orientation = 'left'
                    elif (head_x > neck_x):
                        pov_snake_orientation = 'right'

            # snake body
            tail_1 = None
            tail_2 = None
            for i in reversed(range(1, len(body))):
                # print(i)
                segment = body[i]
                segment_x, segment_y = segment['x'], segment['y']

                if (tail_1 == None):
                    tail_1 = segment
                elif (tail_2 == None):
                    tail_2 = segment

                    if (tail_1 == tail_2):
                        # layer6: double_tail_mask {0,1}
                        self.assign_tensor_val(board_tensor, 6, segment_x, segment_y, 1)

                # layer1: snake bodies {0,1}
                self.assign_tensor_val(board_tensor, 1, segment_x, segment_y, 1)

                # layer2: body segment numbers {0,...,255}
                self.assign_tensor_val(board_tensor, 2, segment_x, segment_y, i * 0.001)

                if (snake_id != pov_snake_id):
                    if (snake_size >= player_size):
                        # layer7: snake bodies >= us {0,1}
                        self.assign_tensor_val(board_tensor, 7, segment_x, segment_y, 1)# + snake_size - player_size)
                    else:
                        # layer8: snake bodies < us {0,1}
                        self.assign_tensor_val(board_tensor, 8, segment_x, segment_y, 1)#player_size - snake_size)

            # layer3: snake length >= player {0,1}
            if (snake_id != pov_snake_id):
                if (snake_size >= player_size):
                    self.assign_tensor_val(board_tensor, 3, head_x, head_y, 1)


        for f in food:
            food_x, food_y = f['x'], f['y']

            # layer4: food {0,1}
            self.assign_tensor_val(board_tensor, 4, food_x, food_y, 1)

        board_width = data["board"]["width"]
        board_height = data["board"]["height"]

        num_snakes_alive = len(snakes)

        if (num_snakes_alive > 0):
            for x in range(board_width):
                for y in range(board_height):
                    # layer9-12: alive count mask
                    self.assign_tensor_val(board_tensor, 8 + num_snakes_alive, x, y, 1)

        # rotate the board so that the POV snake is always facing UP
        if (pov_snake_orientation == 'right'):
            # rotate once to the left
            board_tensor = np.rot90(board_tensor, 1, axes=(1, 2)).copy()
        elif (pov_snake_orientation == 'left'):
            # rotate thrice to the left
            board_tensor = np.rot90(board_tensor, 3, axes=(1, 2)).copy()
        elif (pov_snake_orientation == 'down'):
            # rotate twice so it's flipped upside down
            board_tensor = np.rot90(board_tensor, 2, axes=(1, 2)).copy()

        # print('-------------------')
        # np.set_printoptions(threshold=np.inf)
        # print(board_tensor)

        is_mirror = False

        # get column from head mask where equals 1
        if (can_mirror):
            head_mask_column = np.where(board_tensor[5] == 1)[1]


            # if snake head is on the right side of the board
            if (head_mask_column > int(board_width / 2)):
                # flip the board horizontally (that way our training is equally applicable to both sides of the board)
                board_tensor = np.flip(board_tensor, 2).copy()

                is_mirror = True

        return {
                "tensor" : board_tensor,
                "is_mirror" : is_mirror
            }

    def observe(self, data : dict, should_store_observation : bool, can_mirror : bool) -> dict:
        converted_obj = self.convert_data_to_tensor(data, can_mirror)

        game_id = data["game"]["id"]

        if (game_id not in self.observations):
            self.observations[game_id] = []

        observation = {
            "tensor" : converted_obj["tensor"],
            "is_mirror" : converted_obj["is_mirror"]
        }

        if (should_store_observation):
            observation["image"] = self.convert_data_to_image(data)

            self.observations[game_id].append(observation)

        return observation