-
Notifications
You must be signed in to change notification settings - Fork 0
/
maze_dataset.py
206 lines (173 loc) · 7.81 KB
/
maze_dataset.py
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
194
195
196
197
198
199
200
201
202
203
204
205
206
"""
Create torch dataset for maze
"""
import torch
from torch.utils.data import Dataset
from maze_util import generate_maze_data_sample, CELL_SIZE, GO, MOVE_DOWN, MOVE_LEFT, MOVE_RIGHT, MOVE_STOP, MOVE_UP
class MazeDataset(Dataset):
r"""
Base Maze dataset.
Args:
`n_sample` (int): Number of samples
`grid_size` (int): Size of the maze grid
`max_path_length` (int): Maximum solution length
`shortest_path` (boolean): If `True`, the solution will be the shortest possible path
"""
def __init__(self, n_sample:int, grid_size:int, max_path_length:int, shortest_path:bool) -> None:
super().__init__()
self.n_sample = n_sample
self.grid_size = grid_size
self.max_path_length = max_path_length
self.samples = []
for i in range(self.n_sample):
grid, moves, route = generate_maze_data_sample(self.grid_size, self.max_path_length, shortest_path)
self.samples.append({'grid':torch.from_numpy(grid),'moves':moves,'route':route})
def __getitem__(self, index):
return self.samples[index]['grid'], self.samples[index]['moves'], self.samples[index]['route']
def __len__(self):
return len(self.samples)
def count_similar_samples(self, dataset):
cnt = 0
for i in range(len(dataset)):
grid, moves, route = dataset[i]
for j in range(len(self)):
self_grid, self_moves, self_route = self[j]
if len(moves) != len(self_moves):
continue
if torch.any(grid != self_grid):
continue
route_m = 1
for m,self_m in zip(moves, self_moves):
if m != self_m:
route_m = 0
break
cnt += route_m
return cnt
class MazeDatasetSnapshots(MazeDataset):
r"""
Converts maze samples into step-by-step snapshots (suitable for convnets)
"""
def __init__(self, n_sample: int, grid_size: int, max_path_length: int, shortest_path: bool) -> None:
super().__init__(n_sample, grid_size, max_path_length, shortest_path)
self.snapshots = []
self.labels = []
self.start_points = []
self.end_points = []
for i in range(super().__len__()):
grid, moves, route = super().__getitem__(i)
Xs = [grid.clone()]
self.start_points.append(route[0])
self.end_points.append(route[-1])
Ys = []
for m, (r,c) in zip(moves[1:-1],route[:-1]):
Ys.append(m)
self.add_move_to_grid(grid, r, c, m)
Xs.append(grid.clone())
self.start_points.append(route[0])
self.end_points.append(route[-1])
Ys.append(MOVE_STOP)
self.snapshots.extend(Xs)
self.labels.extend(Ys)
@staticmethod
def add_move_to_grid(grid, r, c, move):
cs = CELL_SIZE
dr,dc = GO[move]
nr,nc = r+dr, c+dc
center = grid[r*cs+cs//2,c*cs+cs//2].item()
ncenter = grid[nr*cs+cs//2,nc*cs+cs//2].item()
if move == MOVE_DOWN:
grid[r*cs+cs//2:r*cs+cs,c*cs+cs//2] = 4
grid[nr*cs:nr*cs+cs//2+1,nc*cs+cs//2] = 4
elif move == MOVE_LEFT:
grid[r*cs+cs//2,c*cs:c*cs+cs//2+1] = 4
grid[nr*cs+cs//2,nc*cs+cs//2:nc*cs+cs] = 4
elif move == MOVE_RIGHT:
grid[r*cs+cs//2,c*cs+cs//2:c*cs+cs] = 4
grid[nr*cs+cs//2,nc*cs:nc*cs+cs//2+1] = 4
elif move == MOVE_UP:
grid[r*cs:r*cs+cs//2+1,c*cs+cs//2] = 4
grid[nr*cs+cs//2:nr*cs+cs,nc*cs+cs//2] = 4
if center > 1:
grid[r*cs+cs//2,c*cs+cs//2] = center
if ncenter > 1:
grid[nr*cs+cs//2,nc*cs+cs//2] = ncenter
def __getitem__(self, index):
return self.snapshots[index], self.labels[index], self.start_points[index], self.end_points[index]
def __len__(self):
return len(self.snapshots)
class MazeDatasetSnapshotsTest(MazeDataset):
r"""
Converts maze samples into step-by-step snapshots (suitable for convnets during the text phase)
"""
def __init__(self, n_sample: int, grid_size: int, max_path_length: int, shortest_path: bool) -> None:
super().__init__(n_sample, grid_size, max_path_length, shortest_path)
self.snapshots = []
self.start_points = []
self.end_points = []
for i in range(super().__len__()):
grid, moves, route = super().__getitem__(i)
Xs = [grid.clone()]
self.start_points.append(route[0])
self.end_points.append(route[-1])
for m, (r,c) in zip(moves[1:-1],route[:-1]):
MazeDatasetSnapshots.add_move_to_grid(grid, r, c, m)
Xs.append(grid.clone())
self.snapshots.append(Xs)
def __getitem__(self, index):
return self.snapshots[index][0], self.snapshots[index][1], self.start_points[index], self.end_points[index]
def __len__(self):
return len(self.snapshots)
class MazeDatasetSequential(MazeDataset):
r"""
Converts maze samples into sequential format (suitable for transformers)
Args:
`d_embed` (int): Number of embedding dimensions
"""
def __init__(self, n_sample: int, grid_size: int, max_path_length: int, d_embed:int, shortest_path: bool) -> None:
super().__init__(n_sample, grid_size, max_path_length, shortest_path)
assert d_embed >= CELL_SIZE * CELL_SIZE + 2, f"d_embed should be >= {CELL_SIZE * CELL_SIZE + 2}"
self.d_embed = d_embed
self.in_sequences = []
self.out_sequences = []
self.start_points = []
self.end_points = []
for i in range(super().__len__()):
in_seq = []
out_seq = []
grid, moves, route = super().__getitem__(i)
cs = CELL_SIZE # cell size
for row in range(grid_size):
for col in range(grid_size):
x = torch.zeros((self.d_embed,))
x[:cs*cs] = grid[row*cs:row*cs+cs, col*cs:col*cs+cs].clone().reshape(-1) # grid cell info
x[cs*cs] = row # row info
x[cs*cs+1] = col # column info
in_seq.append(x)
self.in_sequences.append(torch.stack(in_seq))
self.start_points.append(route[0])
self.end_points.append(route[-1])
for m in range(self.max_path_length):
y = torch.zeros((self.d_embed,))
if m < len(moves):
y[0] = moves[m]
y[1] = m
if m == len(moves) - 1: # last move is always stop
y[2] = route[m-1][0]
y[3] = route[m-1][1]
else:
y[2] = route[m][0]
y[3] = route[m][1]
out_seq.append(y) # will be zeros after the last move
self.out_sequences.append(torch.stack(out_seq))
def sequence_to_grid(self, seq):
idx = 0
grid = torch.zeros(self.grid_size*CELL_SIZE,self.grid_size*CELL_SIZE)
for i in range(self.grid_size):
for j in range(self.grid_size):
grid[CELL_SIZE*i:CELL_SIZE*i+CELL_SIZE, CELL_SIZE*j:CELL_SIZE*j+CELL_SIZE] = seq[idx][:CELL_SIZE*CELL_SIZE].clone().reshape(CELL_SIZE,CELL_SIZE)
idx+=1
return grid
def __getitem__(self, index):
return self.in_sequences[index], self.out_sequences[index], self.start_points[index], self.end_points[index]
def __len__(self):
return len(self.in_sequences)