-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathhamming_dist_analysis.py
168 lines (140 loc) · 5.89 KB
/
hamming_dist_analysis.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
import torch
import pickle
import numpy as np
task = 'futoshiki'
ifile = '/home/yatin/nlm/futo_models/pretrain_chk/e-0_lrh-0.0005_min-1_d-30_nul-16_nmc-5_nm-14_s-3120_tsb-6_tse-6_ts-10000_tr-5_wds-ambiguous_we-200/train_amb_5_5_14_pred_dump.pkl'
#ifile = '/home/yatin/nlm/models/regime_seed_val/e-0_hds-rs_hmin-0_lrh-0.0005_min-0_d-30_nm-5_s-31_tesb-11_tese-11_trs-10_wds-unique_we-200/train10k_10_5_pred_dump.pkl'
#task = 'nqueens'
#ifile = '/home/yatin/nlm/sudoku_models/pretrain_chk_add_e/arb-0_e-0_clip-5.0_lr-0.001_min-1_s-3120_tesb-9_tese-9_trs-9_wds-three-one_we-200_wtd-0.0001/train_e_all_pred_dump.pkl'
#ifile = '/home/yatin/hpcscratch/nlm/sudoku_models/analysis_add/arb-0_e-0_clip-5.0_lr-0.001_min-1_s-3120_tesb-9_tese-9_trs-9_wds-three-one_we-200_wtd-0.0001/train_d_all_pred_dump.pkl'
def match_query(query, pred):
mask = (query>0)
return torch.equal(query[mask], pred[mask])
def check_validity(grid, constraints=None):
grid = grid.cpu().numpy()
constraints = constraints.cpu().numpy()
grid = grid.argmax(axis=2)
for x in range(len(grid)):
row = set(grid[x])
if len(row)!=len(grid):
return False
col = set(grid[:,x])
if len(col)!=len(grid):
return False
if constraints is None:
return True
gt = zip(*np.nonzero(constraints[0]))
for ind in gt:
next_ind = (ind[0],ind[1]+1)
if grid[next_ind]>grid[ind]:
return False
lt = zip(*np.nonzero(constraints[1]))
for ind in lt:
next_ind = (ind[0],ind[1]+1)
if grid[next_ind]<grid[ind]:
return False
return True
def is_safe_futoshiki(grid,constraints):
size = int(len(grid)**0.3334)
grid = grid.reshape(size,size,size).float()
gold = torch.ones(size,size)
if torch.sum(torch.abs(grid.sum(dim=0)-gold))>0:
return False
if torch.sum(torch.abs(grid.sum(dim=1)-gold))>0:
return False
if torch.sum(torch.abs(grid.sum(dim=2)-gold))>0:
return False
constraints = constraints.transpose(0,1)
constraints = constraints.reshape(2,size,size,size)
constraints = constraints[:,:,:,0]
return check_validity(grid,constraints)
def is_safe_sudoku(x,n=9):
grid = x.cpu().numpy().astype(int)
grid = grid.reshape(n,n)
b_size = int(np.sqrt(n))
for i in range(n):
if len(set(grid[i]))<n:
return False
if len(set(grid[:,i]))<n:
return False
#
b_row = i//b_size
b_col = i%b_size
if len(set(grid[b_size*b_row:b_size*(b_row+1),b_size*b_col:b_size*(b_col+1)].flatten()))<n:
return False
return True
def is_safe_nqueens(grid):
size = int(len(grid)**0.5)
grid = grid.reshape(size, size)
indices = torch.nonzero(grid)
if len(indices) != size:
return False
for x in range(size):
r1, c1 = indices[x]
for y in range(x+1, size):
r2, c2 = indices[y]
if (r1 == r2) or (c1 == c2) or (torch.abs(r1-r2) == torch.abs(c1-c2)):
return False
return True
train = pickle.load(open(ifile,'rb'))
compare_func_futo = lambda x,query: match_query(query[:,0].float(),x) and is_safe_futoshiki(x,query[:,1:])
compare_func_nqueens = lambda x,query: match_query(query[:,0].float(),x) and is_safe_nqueens(x)
compare_func_tower = lambda x,query: is_safe_towers(query[:,0].float(),x, args.test_number_begin if return_float else args.train_number)
# query doesn't have to be matched for towers
compare_func_sudoku = lambda x,query: match_query(query,x) and is_safe_sudoku(x, args.test_number_begin if return_float else args.train_number)
if task=='futoshiki':
compare_func = compare_func_futo
pred = [(x['output_dict']['pred'] > 0.5).int() for x in train]
elif task=='tower':
compare_func = compare_func_tower
elif task=='sudoku':
compare_func = compare_func_sudoku
pred = [x['output_dict']['pred'].argmax(dim=1) for x in train]
elif task == 'nqueens':
pred = [(x['output_dict']['pred'] > 0.5).int() for x in train]
compare_func = compare_func_nqueens
#
pred = torch.cat(pred,dim=0)
count = [x['feed_dict']['count'] for x in train]
count = torch.cat(count)
ts = [x['feed_dict']['target_set'] for x in train]
ts = torch.cat(ts,dim=0)
dif = (pred.int().unsqueeze(1).expand_as(ts) != ts.int())
hd = dif.sum(dim=-1)
mask = [x['feed_dict']['mask'] for x in train]
mask = torch.cat(mask,dim=0)
if task == 'sudoku':
k = mask.size(-1) - 1
else:
k = mask.size(-1)
hd = hd[:,:k]
mask = mask[:,:k]
hda = hd[count > 1]
maska = mask[count > 1]
hda[maska == 0] = 0
hdatk = hda.topk(k=k,dim=1)
hdatk= hdatk[0]
hdatk.sum(dim=0).float()/maska.sum(dim=0).float()
counta = count[count > 1]
for i in range(2,(k+1)):
print(i,(counta == i).sum(),hdatk[counta == i].sum(dim=0).float()/maska[counta == i].sum(dim=0).float())
query = [x['feed_dict']['query'] for x in train]
query = torch.cat(query,dim=0)
querya = query[count > 1]
preda = pred[count > 1]
correcta = [compare_func(x.float(),y.float()) for x,y in zip(preda,querya)]
correcta = torch.tensor(correcta)
hda[maska == 0] = -1
ints = (hda == 0).any(dim=1)
hda[maska == 0] = 0
for i in range(2,(k+1)):
ind = ((counta == i) & (correcta != 0))
print('{},{},{},{}'.format("correct",i,ind.sum(),list((hdatk[ind].sum(dim=0).float()/maska[ind].sum(dim=0).float()).numpy())))
if i == k:
ind = ((counta == i) & (correcta != 0) & (ints))
print('{},{},{},{}'.format("correct In TS",i,ind.sum(),list((hdatk[ind].sum(dim=0).float()/maska[ind].sum(dim=0).float()).numpy())))
ind = ((counta == i) & (correcta != 0) & (~ints))
print('{},{},{},{}'.format("correct Not in TS",i,ind.sum(),list((hdatk[ind].sum(dim=0).float()/maska[ind].sum(dim=0).float()).numpy())))
for i in range(2,(k+1)):
ind = ((counta == i) & (correcta == 0))
print('{},{},{},{}'.format("INcorrect",i,ind.sum(),list((hdatk[ind].sum(dim=0).float()/maska[ind].sum(dim=0).float()).numpy())))