forked from waldeland/CNN-for-ASI
-
Notifications
You must be signed in to change notification settings - Fork 2
/
utils.py
244 lines (193 loc) · 7.99 KB
/
utils.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
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
# Compatability Imports
from __future__ import print_function
import torch
import numpy as np
from torch.autograd import Variable
from scipy.interpolate import interpn
def interpret( network, data, data_info, slice, slice_no, im_size, subsampl, return_full_size=True, use_gpu=True):
# Wrap np.linspace in compact function call
def ls(N): return np.linspace(0, N - 1, N, dtype='int')
#Size of cube
N0, N1, N2 = data.shape
#Coords for full cube
x0_range = ls(N0)
x1_range = ls(N1)
x2_range = ls(N2)
#Coords for subsampled cube
pred_points = (x0_range[::subsampl], x1_range[::subsampl], x2_range[::subsampl])
#Select slice
if slice == 'full':
class_cube = data[::subsampl, ::subsampl, ::subsampl] * 0
elif slice == 'inline':
slice_no = slice_no - data_info['inline_start']
class_cube = data[::subsampl, 0:1, ::subsampl] * 0
x1_range = np.array([slice_no])
pred_points = (pred_points[0],pred_points[2])
elif slice == 'crossline':
slice_no = slice_no - data_info['crossline_start']
class_cube = data[::subsampl, ::subsampl, 0:1,] * 0
x2_range = np.array([slice_no])
pred_points = (pred_points[0], pred_points[1])
elif slice == 'timeslice':
slice_no = slice_no - data_info['timeslice_start']
class_cube = data[0:1, ::subsampl, ::subsampl] * 0
x0_range = np.array([slice_no])
pred_points = (pred_points[1], pred_points[2])
#Grid for small class slice/cube
n0,n1,n2 = class_cube.shape
x0_grid, x1_grid, x2_grid = np.meshgrid(ls(n0,), ls(n1), ls(n2), indexing='ij')
#Grid for full slice/cube
X0_grid, X1_grid, X2_grid = np.meshgrid(x0_range, x1_range, x2_range, indexing='ij')
#Indexes for large cube at small cube pixels
X0_grid_sub = X0_grid[::subsampl, ::subsampl, ::subsampl]
X1_grid_sub = X1_grid[::subsampl, ::subsampl, ::subsampl]
X2_grid_sub = X2_grid[::subsampl, ::subsampl, ::subsampl]
#Get half window size
w = im_size//2
#Loop through center pixels in output cube
for i in range(X0_grid_sub.size):
#Get coordinates in small and large cube
x0 = x0_grid.ravel()[i]
x1 = x1_grid.ravel()[i]
x2 = x2_grid.ravel()[i]
X0 = X0_grid_sub.ravel()[i]
X1 = X1_grid_sub.ravel()[i]
X2 = X2_grid_sub.ravel()[i]
#Only compute when a full 65x65x65 cube can be extracted around center pixel
if X0>w and X1>w and X2>w and X0<N0-w+1 and X1<N1-w+1 and X2<N2-w+1:
#Get mini-cube around center pixel
mini_cube = data[X0-w:X0+w+ 1, X1-w:X1+w+ 1, X2-w:X2+w+ 1]
#Get predicted "probabilities"
mini_cube = Variable( torch.FloatTensor(mini_cube[np.newaxis,np.newaxis,:,:,:]) )
if use_gpu: mini_cube = mini_cube.cuda()
out = network(mini_cube)
out = out.data.cpu().numpy()
out = out[:,:, out.shape[2]//2, out.shape[3]//2, out.shape[4]//2]
out = np.squeeze(out)
# Make one output pr output channel
if type(class_cube) != type(list()):
class_cube = np.split( np.repeat(class_cube[:,:,:,np.newaxis],out.size,3),out.size, axis=3)
# Insert into output
if out.size == 1:
class_cube[0][x0, x1, x2] = out
else:
for i in range(out.size):
class_cube[i][x0,x1,x2] = out[i]
#Keep user informed about progress
if slice == 'full': printProgressBar(i,x0_grid.size)
#Resize to input size
if return_full_size:
if slice == 'full': print('Interpolating down sampled results to fit input cube')
N = X0_grid.size
#Output grid
if slice == 'full':
grid_output_cube = np.concatenate( [X0_grid.reshape([N, 1]), X1_grid.reshape([N, 1]), X2_grid.reshape([N, 1])], 1)
elif slice == 'inline':
grid_output_cube = np.concatenate( [X0_grid.reshape([N, 1]), X2_grid.reshape([N, 1])], 1)
elif slice == 'crossline':
grid_output_cube = np.concatenate( [X0_grid.reshape([N, 1]), X1_grid.reshape([N, 1])], 1)
elif slice == 'timeslice':
grid_output_cube = np.concatenate( [X1_grid.reshape([N, 1]), X2_grid.reshape([N, 1])], 1)
#Interpolation
for i in range(len(class_cube)):
is_int = np.sum(np.unique(class_cube[i]).astype('float') - np.unique(class_cube[i]).astype('int32').astype('float') ) == 0
class_cube[i] = interpn(pred_points, class_cube[i].astype('float').squeeze(), grid_output_cube, method='linear', fill_value=0, bounds_error=False)
class_cube[i] = class_cube[i].reshape([x0_range.size, x1_range.size, x2_range.size])
#If ouput is class labels we convert the interpolated array to ints
if is_int:
class_cube[i] = class_cube[i].astype('int32')
if slice == 'full': print('Finished interpolating')
#Squeeze outputs
for i in range(len(class_cube)):
class_cube[i]= class_cube[i].squeeze()
return class_cube
# Print progress information
import sys
import time
st = 0
last_update = 0
#Adapted from https://stackoverflow.com/questions/3173320/text-progress-bar-in-the-console/14879561#14879561
def printProgressBar(iteration, total, prefix = '', suffix = '', decimals = 1, length = 100, fill = '='):
global st, last_update
#Expect itteration to go from 0 to N-1
iteration = iteration + 1 ;
#Only update every 5 second
if time.time() - last_update < 5:
if iteration == total:
time.sleep(1)
else:
return
if iteration <= 1:
st=time.time()
exp_h = ''
exp_m = ''
exp_s = ''
elif iteration == total:
exp_time = (time.time() - st)
exp_h = int(exp_time / 3600 )
exp_m = int(exp_time / 60 - exp_h*60.)
exp_s = int(exp_time - exp_m * 60. - exp_h*3600.)
else:
exp_time = (time.time()-st)/(iteration-1)*total - (time.time()-st)
exp_h = int(exp_time / 3600)
exp_m = int(exp_time / 60 - exp_h * 60.)
exp_s = int(exp_time - exp_m * 60. - exp_h * 3600.)
percent = ("{0:." + str(decimals) + "f}").format(100 * (iteration / float(total)))
filledLength = int(length * iteration // total)
bar = fill * filledLength + '-' * (length - filledLength)
if iteration != total:
print('\r%s |%s| %s%% %s - %sh %smin %ss left' % (prefix, bar, percent, suffix, exp_h, exp_m, exp_s),)
else:
print('\r%s |%s| %s%% %s - %sh %smin %ss ' % (prefix, bar, percent, suffix, exp_h, exp_m, exp_s),)
sys.stdout.write("\033[F")
# Print New Line on Complete
if iteration == total:
print
last_update = time.time()
#Function that returns the GPU number of a variable/module (or False if on CPU)
def gpu_no_of_var(var):
try:
is_cuda = next(var.parameters()).is_cuda
except:
is_cuda = var.is_cuda
if is_cuda:
try:
return next(var.parameters()).get_device()
except:
return var.get_device()
else:
return False
#Take a pytorch variable and make numpy
def var_to_np(var):
if type(var) in [np.array, np.ndarray]:
return var
#If input is list we do this for all elements
if type(var) == type([]):
out = []
for v in var:
out.append(var_to_np(v))
return out
try:
var = var.cpu()
except:
None
try:
var = var.data
except:
None
try:
var = var.numpy()
except:
None
if type(var) == tuple:
var = var[0]
return var
def computeAccuracy(predicted_class, labels):
labels = var_to_np(labels)
predicted_class = var_to_np(predicted_class)
accuracies = {}
for cls in np.unique(labels):
if cls>=0:
accuracies['accuracy_class_' + str(cls)] = int(np.mean(predicted_class[labels==cls]==cls)*100)
accuracies['average_class_accuracy'] = np.mean([acc for acc in accuracies.values()])
return accuracies