-
Notifications
You must be signed in to change notification settings - Fork 0
/
see_tunning.py
136 lines (110 loc) · 3.99 KB
/
see_tunning.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
"""This is the main file to run Weakly supervised supervised dictionary learning
The default data type is torch.tensor with precision float32
"""
#%%
from utils2 import *
os.environ["CUDA_VISIBLE_DEVICES"] = "0"
opts = OPT(C=10, K0=1, K=2)
opts.init, opts.shuffle, opts.show_details = 'rand', False, True
opts.Dh, opts.Dw, opts.batch_size = 100, 29, -1
opts.lamb, opts.lamb0, opts.eta, opts.mu = 0.1, 0.1, 0.01, 0.1 #sparsity, label, low rank
# %% analysis result
record = torch.load('tunning_rough.pt')
n = len(record)
res = torch.rand(n, 3)
param = torch.rand(n,5)
for i,v in enumerate(record):
res[i] = torch.tensor(v[0]) # [n, acc, recall, f1]
param[i] = torch.tensor(v[1]) # [n, Dw, lamb, lamb0, eta, mu]
value, index = res.max(0)
print('max acc, recall, f1, vlaues :', value, '\n')
for i, v in enumerate(index):
print(f"max {['acc', 'recall', 'f1'][i]} index and vlaues :", res[v])
print([v], param[v], '\n')
# a function of given parameters to return the result tensors
def get_result(res, param, Dw=0, lamb=0, lamb0=0, eta=0, mu=0):
""" if Dw, lamb etc. is 0, that means coresponding column are all selected
otherwise Dw etc. is a value from its pool,e.g.
pool_Dw = [ 7., 15., 21., 25., 29., 35., 45.]
Dw = 7. or Dw=15.
param has the shape of [n_record, 5]
each of the 5 columns means [Dw, lamb, lamb0, eta, mu]
res has the shape of [[n_record, 5]]
each of the 3 columns means [acc, recall, F1]
"""
# find the indecies of given param
n = param.shape[0]
idx = torch.arange(n)
if Dw!=0 :
res_ind = idx[param[:,0] == Dw]
else:
res_ind = idx.clone()
if lamb!=0: res_ind = np.intersect1d(idx[param[:,1] == lamb], res_ind)
if lamb0 !=0: res_ind = np.intersect1d(idx[param[:,2] == lamb0], res_ind)
if eta !=0: res_ind = np.intersect1d(idx[param[:,3] == eta], res_ind)
if mu !=0: res_ind = np.intersect1d(idx[param[:,4] == mu], res_ind)
return res[res_ind], res_ind
"Dw=29, lamb=0.1, lamb0=0.1, eta=0.01, mu=0.1"
"Dw=21, lamb=0.1, lamb0=1, eta=0.001, mu=1"
"Dw=21, lamb=0.1, lamb0=0.1, eta=0.001, mu=1"
r, idx = get_result(res, param, Dw=21, lamb=0.1, lamb0=0.1, eta=0.001, mu=1)
print(r)
print(param[idx])
for i in range(5):
if param[idx][:, i].unique().shape[0] >1 :
which_is_0 = i
break
fig = plt.figure()
fig.set_size_inches(w=6, h=4)
index = param[idx][:, which_is_0].sort()[1]
plt.plot(param[idx][:, which_is_0].sort()[0], r[:, -1][index], '-x')
#%%
record = torch.load('tunning.pt')
res = torch.zeros(80, 3)
para = torch.zeros(80, 5)
for i,r in enumerate(record):
res[i] = torch.tensor(r[0])
para[i] = torch.tensor(r[1])
res = res.reshape(5, 16, -1).mean(0)
para = para.reshape(5, 16, -1).mean(0)
v, i = res.sort()
"Dw=29, lamb=0.1, lamb0=0.1, eta=0.001, mu=0.1 is the best"
# %% compare with others' result
route = '/home/chenhao1/Matlab/WSCDL/'
# res = sio.loadmat(route+'res_knn.mat')
# res = res['Pre_Labels']
# res[res==-1]=0
# res = res.T
# metrics.f1_score(Y_test.cpu().flatten(), res.flatten())
with open(route+'you_raich_0.txt') as f:
data =f.readlines()
rec = []
prec = []
count = 0
for i, d in enumerate(data):
if d == 'rec =\n':
rec.append(float(data[i+2][4:10]))
count += 1
if d == 'prec =\n':
prec.append(float(data[i+2][4:10]))
if count == 10:
print(rec[-1])
print(prec[-1])
rec, prec = torch.tensor(rec), torch.tensor(prec)
f1 = 2/(1/rec+1/prec)
v, i = f1.sort()
# best lamb, winzize, N
# 10, 30, 200
# 10, 100, 10
# 10, 50, 50
#%% visualize learned atoms
param = str([opts.K, opts.K0, opts.Dw, opts.lamb, opts.lamb0, opts.eta , opts.mu])
D, D0, S, S0, W, opts, loss = \
torch.load('../saved_dicts/'+param+'DD0SS0Woptsloss.pt', map_location='cpu')
for i in range(10):
fig= plt.figure()
fig.set_size_inches(w=4, h=6)
d = D[i].permute(0,2,1).reshape(opts.Dh, opts.K*opts.Dw).cpu()
plt.imshow(d, aspect='auto', interpolation='None')
plt.title(f'Class {i} atoms')
# %%