def showrpca():
recall = np.load('rpca_rec.npy')
diffall = np.load('rpca_diff.npy')
from bigrec import bigrec
from layerbase import DrawPatch
drec = recall.reshape((-1,1,70-25,90-0))
drec2 = np.load('glassline.npy').astype('f').reshape((-1,1,70-25,90-0))
glassorig = np.load('glassorig.npy').reshape((-1,105,90))
glassall = np.copy(glassorig)[:drec.shape[0]]
drecall = np.zeros((drec.shape[0]+drec2.shape[0],1,70-25,90-0),'f')
drecall[::2]=drec
drecall[1::2]=drec2
for i in range(drec.shape[0]):
glassall[i] = bigrec(drec[i,0], glassall[i])
misc.toimage(DrawPatch(glassall.reshape((-1,1,105,90)))).save('glasses.jpg')
drecall = np.zeros((drec.shape[0]+drec.shape[0],105,90),'f')
drecall[::2]=glassall
drecall[1::2]=glassorig
misc.toimage(DrawPatch(drecall.reshape((-1,1,105,90)))).save('recoverypair.jpg')
#return
#misc.toimage(DrawPatch(drecall)).save('rpcapair.jpg')
drec = diffall.reshape((-1,1,70-25,90-0))
misc.toimage(DrawPatch(drec)).save('rpcadiff.jpg')
def cutratio_all(ratio):
import cPickle
sparsedirect = cPickle.load(file('sparsepcadirect','rb'))
sparsedirect.set_params(transform_algorithm='lasso_lars')
glassmodel = np.load('glassline.npy').astype('f')
glassmodel = glassmodel[:200]
recall = np.empty_like(glassmodel)
diffall = np.empty_like(glassmodel)
maskall = np.empty_like(glassmodel)
for idx in range(glassmodel.shape[0]):
print idx,glassmodel.shape[0]
recall[idx],diffall[idx],maskall[idx] = cutratio(ratio, idx, sparsedirect, glassmodel)
np.save('cut_rec.npy',recall)
np.save('cut_diff.npy',diffall)
np.save('cut_mask.npy',maskall)
from layerbase import DrawPatch
drec = recall.reshape((-1,1,70-25,90-0))
misc.toimage(DrawPatch(drec)).save('irec.jpg')
drec = diffall.reshape((-1,1,70-25,90-0))
misc.toimage(DrawPatch(drec)).save('idiff.jpg')
drec = maskall.reshape((-1,1,70-25,90-0))
misc.toimage(DrawPatch(drec)).save('imask.jpg')
def scvisdecomp():
"可视化结果"
from layerbase import DrawPatch
import cPickle
sparsedirect = cPickle.load(file('sparsepcadirect','rb'))
drec = sparsedirect.components_.reshape((-1,1,70-25,90-0))
misc.toimage(DrawPatch(drec)).save('dict.jpg')
drec = np.load('glassline.npy').astype('f').reshape((-1,1,70-25,90-0))
misc.toimage(DrawPatch(drec)).save('src.jpg')
drec = np.load('glassdecomppca.npy').reshape((-1,1,70-25,90-0))
misc.toimage(DrawPatch(drec)).save('decomppca.jpg')
drec = np.abs(np.load('glassdiffpca.npy')).reshape((-1,1,70-25,90-0))
misc.toimage(DrawPatch(drec)).save('diffpca.jpg')
def showcut():
recall = np.load('cut_rec.npy')
diffall = np.load('cut_diff.npy')
maskall = np.load('cut_mask.npy')
from layerbase import DrawPatch
drec = recall.reshape((-1,1,70-25,90-0))
drec2 = np.load('glassline.npy').astype('f').reshape((-1,1,70-25,90-0))[0:4]
drecall = np.zeros((drec.shape[0]+drec2.shape[0],1,70-25,90-0),'f')
drecall[::2]=drec
drecall[1::2]=drec2
misc.toimage(DrawPatch(drecall)).save('ipair.jpg')
drec = diffall.reshape((-1,1,70-25,90-0))
misc.toimage(DrawPatch(drec)).save('idiff.jpg')
drec = maskall.reshape((-1,1,70-25,90-0))
misc.toimage(DrawPatch(drec)).save('imask.jpg')
def domedmask():
diff = np.load('rpca_diff.npy').reshape((-1,70-25,90-0))
diffmask = np.where(diff>50,255,0)
from bigrec import bigrec
maskall = np.zeros((diff.shape[0],105,90),np.int0)
glassorig = np.load('glassorig.npy').reshape((-1,105,90))
noglass = np.zeros_like(glassorig)
for i in range(diff.shape[0]):
print i
maskall[i] = bigrec(diffmask[i], maskall[i])
medresult = maskedmedfilt(glassorig[i], maskall[i], 6,4)
noglass[i] = np.where(maskall[i], medresult, glassorig[i])
np.save('medfilt_noglass.npy',noglass)
from layerbase import DrawPatch
misc.toimage(DrawPatch(noglass.reshape((-1,1,105,90)))).save('mednoglass.png')
misc.toimage(DrawPatch(maskall.reshape((-1,1,105,90)))).save('medmask.png')
def pairdump():
"可视化模型列"
from layerbase import DrawPatch
import cPickle
sparsedirect = cPickle.load(file('sparselinked','rb'))
drec = sparsedirect.components_.reshape((-1,1,70-25,90-0))
misc.toimage(DrawPatch(drec)).save('dictpair.jpg')
def showtransform():
"显示神经网络转换结果"
trans = np.load('transform.npy')[:,0]
glassmodel = np.load('glassline.npy').astype('f').reshape((-1,45,90))
glassmodel /= np.max(glassmodel)
trans = np.insert(trans,2,glassmodel,axis=1)
p = trans[:,0].reshape((trans.shape[0],-1))
trans[:,0] -= ((p.max(axis=1)+p.mean(axis=1))*0.5).reshape((-1,1,1))
trans[:,0] = np.where(trans[:,0]>0, trans[:,0],0)
from layerbase import DrawPatch
misc.toimage(DrawPatch(trans,False,'bgy')).save('nntrans.png')
def rpca_noglass():
"用RPCA学习不戴眼镜的字典"
dic = np.load('nglassline.npy')
dicflat = dic.reshape((dic.shape[0],-1))
import os
if os.path.exists('eyemodel.npz'):
partialmodel = np.load('eyemodel.npz')
else:
partialmodel = {'components':None, 'nglasscodes':None}
from layerbase import DrawPatch
idx=0
for sdict, scode in rpca_learn(dicflat,stepdict=partialmodel['components'],stepcode=partialmodel['nglasscodes']):
np.savez('eyemodel.npz', components=sdict, nglasscodes=scode)
misc.toimage(DrawPatch(sdict.reshape((-1,1,45,90)),False,'bgy')).save('dictrpca_%s.png'%idx)
print "SAVED"
idx += 1
def getF(ab):
pr = []
#valid = inp[230:330]
#voutp = oup[230:330]
a, b = ab.astype('f')
for j in range(100 / BS):
#sys.stdout.write('.')
#sys.stdout.flush()
v = valid_binary_model(valid[j * BS:(j + 1) * BS],
voutp[j * BS:(j + 1) * BS], a, b)
resp = vis_binary_model(valid[j * BS:(j + 1) * BS],
voutp[j * BS:(j + 1) * BS], a, b)
layer = 0
for i in resp:
layer += 1
if len(i.shape) != 4:
PIL.Image.fromarray(
np.array((i - np.min(i)) /
(np.max(i) - np.min(i)) * 255,
np.uint8)).save('stepf789/%s_resp%s.png' %
(j, layer))
continue
PIL.Image.fromarray(DrawPatch(i[0:1])).save(
'stepf789/%s_resp%s.jpg' % (j, layer), quality=100)
PIL.Image.fromarray(
DrawMaskedPatch(resp[0][0:1], resp[-1][0:1, 0:1])).save(
'stepf789/maskresult_%s.jpg' % (j), quality=100)
PIL.Image.fromarray(
DrawMaskedPatch(resp[0][0:1], resp[-2][0:1, 0:1])).save(
'stepf789/masktruth_%s.jpg' % (j), quality=100)
#v=v[:-1]
print v
pr.append(v)
pr = np.array(pr).sum(axis=0)
tp = pr[0]
fp = pr[1]
tn = pr[2]
fn = pr[3]
prec = tp / (tp + fp)
reca = tp / (tp + fn)
F = -2 * prec * reca / (prec + reca)
print ab, F
return F[:-1].sum()
inpt.std(axis=(2, 3), keepdims=True) + 1e-10)
d += [float(t) for t in train_model(inpt, oupt)][1]
sys.stdout.write('.')
sys.stdout.flush()
if i % BATCHSTEP == BATCHSTEP - 1:
print d, "DRAW"
d = 0
#Draw model
layer = 0
for i in model.paramlayers():
layer += 1
param = i.params[0].get_value()
if len(param.shape) != 4:
if hasattr(i, 'reshape') and i.reshape != None:
PIL.Image.fromarray(
DrawPatch(
param.reshape((-1, ) + i.reshape[1:]))).save(
'stepf789/layer_%s.jpg' % layer,
quality=100)
else:
PIL.Image.fromarray(
np.array((param - np.min(param)) /
(np.max(param) - np.min(param)) * 255,
np.uint8)).save('stepf789/layer_%s.png' %
layer)
continue
PIL.Image.fromarray(DrawPatch(param)).save(
'stepf789/layer_%s.jpg' % layer, quality=100)
#Draw response
resp = vis_model(inpt, oupt)
layer = 0
for i in resp:
print "Get precision & recall"
pr = []
valid = inp[230:330]
voutp = oup[230:330]
for j in range(100/BS):
#sys.stdout.write('.')
#sys.stdout.flush()
v=valid_binary_model(valid[j*BS:(j+1)*BS],voutp[j*BS:(j+1)*BS])
resp = vis_binary_model(valid[j*BS:(j+1)*BS],voutp[j*BS:(j+1)*BS])
layer = 0
for i in resp:
layer += 1
if len(i.shape)!=4:
PIL.Image.fromarray(np.array((i-np.min(i))/(np.max(i)-np.min(i))*255,np.uint8)).save('stepallface/%s_resp%s.png'%(j,layer))
continue
PIL.Image.fromarray(DrawPatch(i[0:1])).save('stepallface/%s_resp%s.jpg'%(j,layer), quality=100)
PIL.Image.fromarray(DrawMaskedPatch(resp[0][0:1], resp[-1][0:1,0:1])).save('stepallface/%s_maskresult_%s.jpg'%(v[-1],j), quality=100)
PIL.Image.fromarray(DrawMaskedPatch(resp[0][0:1], resp[-2][0:1,0:1])).save('stepallface/%s_masktruth_%s.jpg'%(v[-1],j), quality=100)
v=v[:-1]
print v
pr.append(v)
pr = np.array(pr).sum(axis=0)
tp=pr[0]
fp=pr[1]
tn=pr[2]
fn=pr[3]
prec = tp/(tp+fp)
reca = tp/(tp+fn)
F = 2*prec*reca/(prec+reca)
print F
sys.exit(0)
def showdata():
"绘出数据"
data = np.load('glassdata2.npz')
from layerbase import DrawPatch
misc.toimage(DrawPatch(data['input'])).save('datainput.png')
misc.toimage(DrawPatch(data['output'])).save('dataoutput.png')