-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathproof_of_concept.py
309 lines (296 loc) · 12.7 KB
/
proof_of_concept.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
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
method=['opencv','pil','mpl','big_mpl'][-1]
import PIL
if method=='opencv':
import cv2
elif method=='mpl' or method=='big_mpl':
import matplotlib.pyplot as plt
import matplotlib.image as mim
import numpy as np
import utils as ARU
import tensorflow as tf
from edger import edger
import time
from utils import kerasdatasets
import scipy as sp
import imgaug as iag
class KvarLayer(): #K.layers.convolutional._Conv):#layers.Layer):
"""
:param num_c: number of cell in this layer
:filter size of the filter/kernel (int or container)
:bigB use a bias for the kernel(False) or for the whole output(True)
:padding type of padding to use, todo
"""
def __init__(self, weights,
sqrt=0,format='NHWC',
KCD=False,channal=False,t2=1,cast=None,
**kwargs):
self.window=weights[0].shape
self.channal=channal
self.W=np.array(weights)
if format=='NHWC':
if t2:
self.W=tf.expand_dims(self.W,len(self.W.shape))
else:
self.W=np.transpose(self.W, (1,2,0))
if self.channal:
self.W=tf.expand_dims(self.W, 1)
else:
if self.channal:
self.W=tf.expand_dims(self.W, 1)
self.sqrt=sqrt
self.KCD=KCD
assert (format=="NCHW")or(format=="NHWC")
self.format=format
self.cast=cast
self.zerr=tf.constant(0,dtype=cast)
self.one=tf.constant(1,dtype=cast)
if self.format=='NHWC':
self.sb=(1,1,self.W.shape[-1])
self.xi=(-3,-2)
self.x2=(-1,-3,-2)
elif self.format=='NCHW':
self.sb=(self.W.shape[1],1,1)
self.xi=(-2,-1)
self.x2=(-2,-1,-3)
def tfwindow(self,arr,nc_to_nh=False,pad="VALID",stride=1):
"""pad can be "SAME",will add 0s to get the same output shape as input
nc_to_nh: change from "NHWC" to "NCHW"
stride = stride for the 2 internal axis
"""
assert pad=="VALID" or pad=="SAME"
if self.format=="NCHW":
temp=tf.extract_image_patches(arr, ksizes=(1,1,*self.window), strides=(1,1,stride,stride),rates=(1,1,1,1),padding=pad)
elif self.format=="NHWC":
temp=tf.extract_image_patches(arr, ksizes=(1,*self.window,1), strides=(1,stride,stride,1),rates=(1,1,1,1),padding=pad)
if nc_to_nh and self.format=='NCHW':
return(tf.transpose(tf.reshape(temp,(-1,*self.convshape,1,self.ch,*self.window)),perm=[0,1,2,3,6,4,5]))
else:
if self.format=="NCHW":
#print((-1,*self.convshape,1,self.ch,*self.window))
return(tf.reshape(temp,(-1,*self.convshape,1,self.ch,*self.window)))
elif self.format=="NHWC":
#print((-1,*self.convshape,self.ch,*self.window,1))
return(tf.reshape(temp,(-1,*self.convshape,self.ch,*self.window,1)))
def call(self, array,outfrmt='NCHW'):#you can crunch by doing all channels
self.shape=array.shape
if self.format=="NCHW":
self.convshape=ARU.convshape(self.shape[-2:], self.window)
self.ashp=self.window[-1]
self.ch=self.shape[1]
elif self.format=="NHWC":
self.convshape=ARU.convshape(self.shape[1:-1], self.window)
self.ch=self.shape[-1]
if self.channal:
self.ashp=self.ch
else:
self.ashp=self.W.shape[-1]
reshaped=self.tfwindow(array)
if self.cast:
self.W=tf.cast(self.W,self.cast)
reshaped=tf.cast(reshaped,self.cast)
self.arrs=reshaped.shape
mul=(reshaped*self.W)
size=tf.reduce_sum(self.W,axis=self.xi,keepdims=True)#shape=(outputs, channel)
szm=tf.constant(self.window[0]*self.window[1],shape=self.sb,dtype=self.cast)
mean=tf.reduce_sum(mul,axis=self.xi,keepdims=True)/szm
#print(self.W.dtype,size.dtype,self.one.dtype,self.zerr.dtype)
size=tf.cond(tf.equal(tf.reshape(size,([1])),self.zerr)[0], lambda: tf.add(size,self.one),lambda: size)
i=(tf.square(mul-mean))/size
if self.KCD:
out=tf.reduce_sum(i,axis=self.xi)
else:
out=tf.reduce_sum(i,axis=self.x2)
if self.sqrt:
out=tf.sqrt(out)
if not(self.cast is None):
out=tf.cast(out, self.cast)
#print(out.shape,self.format,(self.arrs[0],self.arrs[1],self.arrs[2],self.ashp[0]))
#print('mul',mul.shape,'size',size.shape,'mean',mean.shape,'i',i.shape,'out',out.shape)
#print(out.shape,self.format,(self.arrs[1],self.arrs[2],self.ashp))
if self.format=="NCHW":
if self.KCD:
return(tf.transpose(tf.reshape(out,(self.arrs[0],self.arrs[1],self.arrs[2],self.ashp*self.arrs[-3])),(0,3,1,2)).eval())
else:
assert out.shape[1:]==(self.arrs[1],self.arrs[2],self.ashp)
#print(tf.transpose(out, (0,3,1,2)).shape,"outshapenchw")
return(tf.transpose(out, (0,3,1,2)).eval())#tf.reshape(out,(self.arrs[0],self.ashp[0],self.arrs[1],self.arrs[2]))
else:
if self.KCD:
return(tf.reshape(out,(self.arrs[0],self.arrs[1],self.arrs[2],self.ch)).eval())
else:
if outfrmt=='NCHW':
out=tf.transpose(out, (3,0,1,2))
assert out.shape[1:]==(self.ashp,self.arrs[1],self.arrs[2])
elif outfrmt=='NHWC':
out=tf.transpose(out, (3,1,2,0))
assert out.shape[1:]==(self.arrs[1],self.arrs[2],self.ashp)
return(out.eval())
def jpg2arr(image_path,imfrmt='CHW'):
"""
Loads JPEG image into 3D Numpy array of shape
(width, height, channels)
"""
with PIL.Image.open(image_path) as image:
im_arr = np.fromstring(image.tobytes(), dtype=np.uint8)
im_arr = im_arr.reshape((image.size[1], image.size[0], 3))
if imfrmt=='CHW':
im_arr = np.transpose(im_arr, (2,0,1))
return(im_arr)
FORMAT=['NCHW','NHWC'][1]
holder=[]
bp=".\\POC\\"
kerdataa=1
if kerdataa:
for K in (4,3):
tr,ts=kerasdatasets(K)
for I in range(9):
holder.append(tr[0][I])
for imgpth in ["guit.jpg","meme.jpg",
"3pan.jpg","croc.jpg","spell.jpg",
"cell.jpg",
"tiger.jpg",
"butterfly.jpg",
]:
holder.append(bp+imgpth)
kernel_shape=[3,3]
weights=edger(kernel_shape,kernel_shape[0]-1,num_edges=None,batch_edges=3,
maxpoint=kernel_shape[0]*2-1,mode=2,operator=1,seed=10102,round=4)
sess=tf.Session()
onerw=True
sample=4
norm=0
saveplots=1
datcast=[None,np.uint8,np.float16,np.float32,np.uint16,np.float64][-1]
recast=[None,np.uint8,np.float16,np.float32,np.uint16][1]
with sess.as_default():
for I in holder:
if isinstance(I, str):
if not(method=='opencv'):
I=jpg2arr(I,imfrmt=FORMAT[1:])
elif method=='opencv':
og=cv2.imread(I,1)
Itype='RGB'
if not(method=='opencv'):
og=PIL.Image.fromarray(I, Itype)
if FORMAT=='NCHW':
og=np.transpose(og, (2,0,1))
else:
Itype='L'
if FORMAT=='NCHW':
I=np.reshape(I,(1,*I.shape))
if method=='pil':
og=PIL.Image.fromarray(I[0], Itype)
elif FORMAT=='NHWC':
I=np.reshape(I,(*I.shape,1))
if method=='pil':
og=PIL.Image.fromarray(I[:,:,0], Itype)
if method=='pil':
og.show()
I2=None
weigtz=[]
if onerw:
acm={}
for num in range(sample+1):
if num==0:
rw= [np.array([[0, -1, 0],[-1, 4, -1],[0, -1, 0]],dtype=datcast),]
elif num==1:
rw = [np.array([[0, 0, 0],[0, -4, 1],[0, 2, 1]],dtype=datcast),]
else:
rw=[weights[np.random.randint(0,high=len(weights))],]
weigtz.append(rw[0])
print('weights',rw,'weights')
lay=KvarLayer(rw,format=FORMAT,KCD=True,sqrt=1,cast=datcast)
outp=lay.call(np.reshape(I,(1,*I.shape,)),outfrmt='NHWC')[0]
conv1=np.transpose(np.array([sp.signal.correlate2d(I[:,:,ix],rw[0], mode='valid') for ix in range(I.shape[-1])]),(1,2,0))
convp=iag.augmenters.Sequential([iag.augmenters.Convolve(matrix=rw[0])]).augment_image(I)
rwex=np.expand_dims(np.expand_dims(rw[0],-1),-1).astype(datcast)
if not(Itype=='L'):
rwex=np.transpose(np.array([np.transpose(np.array([rw[0] for _ in range(I.shape[-1])]),(1,2,0)) for _ in range(I.shape[-1])]),(1,2,3,0))
else:
rwex=np.expand_dims(np.expand_dims(rw[0],-1),-1).astype(datcast)
tfconv=tf.nn.conv2d(np.expand_dims(I.astype(np.float16),0), rwex, [1,1,1,1], padding='VALID').eval().astype(recast)
if not(lay.KCD):
if norm:
outp *= 255.0/outp.max()
else:
outp=outp//3
suum=rw[0].sum(axis=(1,0)).astype(datcast)
if 0:
if suum==0:
suum+=1
convp = convp//suum
conv1 = conv1.astype(np.uint8)//suum
tfconv=tfconv//suum
if Itype=='L':
if I2 is None:
I2=I[:,:,0]
outp=outp[:,:,0]
convp=convp[:,:,0]
conv1=conv1[:,:,0]
tfconv=tfconv[0,:,:,0]
cmapv='gray'
else:
if I2 is None:
I2=I
cmapv=None
tfconv=tfconv[0]
if method=='pil':
com=PIL.Image.fromarray(convp, Itype)
co1=PIL.Image.fromarray(conv1, Itype)
img = PIL.Image.fromarray(outp, Itype)
com.show('conv')
img.show('var')
elif method=='opencv':
cv2.imshow('output',outp)
elif method=="big_mpl":
acm[num]={0:I2,1:convp,2:outp,3:conv1,4:tfconv,5:rw[0]}
if method=="big_mpl":
acmarg=['Original','imgaug conv','weighted variance','conv with scipy.signal','tf.conv2d','kernel/weight']
no_original=0
many_plots=1
if no_original:
start=1
col=2
row=2
subx=0.45
suby=0.5
else:
start=0
col=3
row=2
subx=0.48
suby=0.55
i=len(acm)
il=len(acm[0])
j=i*il
if many_plots:
for ij in range(i):
weight=weigtz[ij]
plt.subplots(nrows=row, ncols=col)#, sharex, sharey, squeeze, subplot_kw, gridspec_kw)
for column in range(start,il):
data=acm[ij][column]
#print(data)
#print(data.min(),data.max(),column,data.dtype,recast,recast(255.0).dtype)
if column==(5+start):
cmapv='gray'
if 1:
data=recast((data-data.min()) * (data.dtype.type(255.0)/(data.max()-data.min())))
plt.subplot(row,col,column+1-start)
plt.imshow(data,cmap=cmapv)
plt.title(acmarg[column],fontdict={'family':'serif','weight':'black','style':'oblique'})
plt.colorbar()#mappable, cax, ax)
#plt.figlegend(handles=,str(weight), loc='center')
plt.suptitle('kernel'+str(weight),x=subx,y=suby)
plt.tight_layout(pad=0.55, h_pad=2.71, w_pad=1.51)#, rect)
plt.show()
else:
plt.subplots(nrows=len(acm), ncols=4)#, sharex, sharey, squeeze, subplot_kw, gridspec_kw)
for jj in range(j):
plt.subplot(i,il,jj+1)
xj=jj%il
vj=jj//il
weight=weigtz[vj]
plt.imshow(acm[vj][xj],cmap=cmapv)
if vj==0:
plt.title(acmarg[xj])
plt.show()