def copy_file_from_remote(file_path, main_hostname, username):
"""
Copy file or folder from remote to local machine using scp.
This uses Paramiko's scp module and assumes that ssh keys between the two machines have been setup.
@file_path: This should lead with the generic '~/' user sign. This makes the home path compatible
between different machines
@main_hostname: The name or ... of the remote machine
@username: The username used for login to remote machine
"""
raw_file_path = file_path
local_file_path = os.path.expanduser(file_path)
# if not os.path.isfile(local_file_path+filename) and main_hostname != 'localhost':
local_parent_dir, filename = ntpath.split(local_file_path)
if not os.path.isdir(local_parent_dir):
print('Path to parent directory %s does not exist, creating dir.' %
local_parent_dir)
os.makedirs(local_parent_dir)
print('Copying data from %s via scp...' % main_hostname)
tic()
# copy the data folder (and contents) from the remote machine to the parent directory on the local machine
# parent_dir = os.path.dirname(os.path.dirname(self.local_file_path))
# local_parent_dir = os.path.dirname(local_file_path)
# ssh.scp_get(username, main_hostname, raw_file_path, os.path.join(file_path, os.pardir), recursive=True)
scp_get(username,
main_hostname,
raw_file_path,
local_parent_dir,
recursive=True)
print('Completed copying data from main host.')
toc()
return
def fit_model(phi, eps=0.):
assert phi.ndim == 2, 'data has to be two-dimensional'
assert phi.shape[1] > phi.shape[0], 'data samples have to be in columns'
d, nsamples = phi.shape
nij = d**2-d # number of coupling terms
adata, arow, acol, b = fill_model_matrix(phi)
a = sparse.coo_matrix((adata,(arow,acol)), (nij,nij))
tic('matrix inversion')
if eps > 0:
a2 = np.dot(a.T,a) + eps*nsamples*sparse.eye(nij,nij,format='coo')
b2 = np.dot(a.todense().T,np.atleast_2d(b).T)
# this sparse multiplication is buggy !!!!, I can't get the shape of b2 to be = (b.size,)
b3 = b2.copy().flatten().T
b3.shape = (b3.size,)
k_vec = dsolve.spsolve(a2.tocsr(),b3)
k_mat = np.zeros((d,d),complex)
k_mat.T[np.where(np.diag(np.ones(d))-1)] = k_vec.ravel()
else:
k_vec = dsolve.spsolve(a.tocsr(),b)
k_mat = np.zeros((d,d),complex)
k_mat.T[np.where(np.diag(np.ones(d))-1)] = k_vec
toc('matrix inversion')
return k_mat
def fit_model(phi, eps=0.):
assert phi.ndim == 2, 'data has to be two-dimensional'
assert phi.shape[1] > phi.shape[0], 'data samples have to be in columns'
d, nsamples = phi.shape
nij = d**2 - d # number of coupling terms
adata, arow, acol, b = fill_model_matrix(phi)
a = sparse.coo_matrix((adata, (arow, acol)), (nij, nij))
tic('matrix inversion')
if eps > 0:
a2 = np.dot(a.T,
a) + eps * nsamples * sparse.eye(nij, nij, format='coo')
b2 = np.dot(a.todense().T, np.atleast_2d(b).T)
# this sparse multiplication is buggy !!!!, I can't get the shape of b2 to be = (b.size,)
b3 = b2.copy().flatten().T
b3.shape = (b3.size, )
k_vec = dsolve.spsolve(a2.tocsr(), b3)
k_mat = np.zeros((d, d), complex)
k_mat.T[np.where(np.diag(np.ones(d)) - 1)] = k_vec.ravel()
else:
k_vec = dsolve.spsolve(a.tocsr(), b)
k_mat = np.zeros((d, d), complex)
k_mat.T[np.where(np.diag(np.ones(d)) - 1)] = k_vec
toc('matrix inversion')
return k_mat
def mainLoop(modelType, modelArgs, positives, trainingList, featuresDir,
featuresExt, modelOut, maxNegOverlap, iter):
pos, posIdx, featSize, fmSize = positives
featureSpace = pos.shape[1]
startTime = cu.tic()
if iter == 0:
## Random Negatives
print ' >>> RANDOM NEGATIVES'
N, negIdx = learn.getRandomNegs(featuresDir, trainingList, featuresExt,
featSize, maxVectorsCache,
maxNegativeImages)
cellsPerImage = featSize / featureSpace
N = N.reshape((N.shape[0], featureSpace, fmSize,
fmSize)) # Recover original feature layout
neg = np.zeros((cellsPerImage * N.shape[0], featureSpace))
for i in range(N.shape[0]):
neg[i * cellsPerImage:(i + 1) * cellsPerImage] = N[i].T.reshape(
(cellsPerImage, featureSpace)) # Unfold features
hards = {'features': np.zeros((0, neg.shape[1])), 'index': []}
lap = cu.toc(
'Random negatives matrix (' + str(neg.shape[0]) + ' instances)',
startTime)
else:
## Mine hard negatives
print ' >>> MINING HARD NEGATIVES'
model = det.createDetector(modelType, modelArgs)
model.load(modelOut + '.' + str(iter - 1))
detList, detMatrix = maskDetector.detectObjects(
model, trainingList, featuresDir, featuresExt, -10.0)
hdnList, detMatrix = maskDetector.selectHardNegatives(
detList, detMatrix, posIdx, maxNegativeVectors)
neg = maskDetector.loadHardNegativesFromMatrix(featuresDir, hdnList,
detMatrix, featuresExt,
featureSpace,
maxNegativeVectors)
hards = cu.loadMatrixNoCompression(modelOut + '.hards').item()
lap = cu.toc(
'Hard negatives (' + str(neg.shape[0]) + ' mined + ' +
str(hards['features'].shape[0]) + ' previous instances)',
startTime)
## Learn Detector
neg = np.concatenate((neg, hards['features']))
clf = det.createDetector(modelType, modelArgs)
clf.learn(pos, neg)
clf.save(modelOut + '.' + str(iter))
lap = cu.toc('Classifier learned:', lap)
## Keep hard negatives for next iterations
scores = clf.predict(neg)
hardNegsIdx = np.argsort(scores)
hardNeg = np.concatenate(
(hards['features'], neg[hardNegsIdx[-cu.topHards:]]))
cu.saveMatrixNoCompression({'features': hardNeg}, modelOut + '.hards')
print ' ** Iteration', iter, 'done'
def crank(V, L1, R1x, L2, R2x, dt, n, crumbs=[], callback=None):
V = V.copy()
dt *= 0.5
L1e = flatten_tensor(L1)
L1i = L1e.copy()
R1 = np.array(R1x).T
L2e = flatten_tensor(L2)
L2i = L2e.copy()
R2 = np.array(R2x)
m = 2
# L = (As + Ass - r*np.eye(nspots))*-dt + np.eye(nspots)
L1e.data *= dt
L1e.data[m, :] += 1
L1i.data *= -dt
L1i.data[m, :] += 1
R1 *= dt
L2e.data *= dt
L2e.data[m, :] += 1
L2i.data *= -dt
L2i.data[m, :] += 1
R2 *= dt
offsets1 = (abs(min(L1i.offsets)), abs(max(L1i.offsets)))
offsets2 = (abs(min(L2i.offsets)), abs(max(L2i.offsets)))
print_step = max(1, int(n / 10))
to_percent = 100.0 / n
utils.tic("Crank:")
R = R1 + R2
normal_shape = V.shape
transposed_shape = normal_shape[::-1]
for k in xrange(n):
if not k % print_step:
if isnan(V).any():
print "Crank fail @ t = %f (%i steps)" % (dt * k, k)
return crumbs
print int(k * to_percent),
if callback is not None:
callback(V, ((n - k) * dt))
V = (L2e.dot(V.flat).reshape(normal_shape) + R).T
V = spl.solve_banded(offsets1, L1i.data, V.flat, overwrite_b=True)
V = (L1e.dot(V).reshape(transposed_shape).T) + R
V = spl.solve_banded(offsets2, L2i.data, V.flat,
overwrite_b=True).reshape(normal_shape)
crumbs.append(V.copy())
utils.toc()
return crumbs
def crank(V, L1, R1x, L2, R2x, dt, n, crumbs=[], callback=None):
V = V.copy()
dt *= 0.5
L1e = flatten_tensor(L1)
L1i = L1e.copy()
R1 = np.array(R1x).T
L2e = flatten_tensor(L2)
L2i = L2e.copy()
R2 = np.array(R2x)
m = 2
# L = (As + Ass - r*np.eye(nspots))*-dt + np.eye(nspots)
L1e.data *= dt
L1e.data[m, :] += 1
L1i.data *= -dt
L1i.data[m, :] += 1
R1 *= dt
L2e.data *= dt
L2e.data[m, :] += 1
L2i.data *= -dt
L2i.data[m, :] += 1
R2 *= dt
offsets1 = (abs(min(L1i.offsets)), abs(max(L1i.offsets)))
offsets2 = (abs(min(L2i.offsets)), abs(max(L2i.offsets)))
print_step = max(1, int(n / 10))
to_percent = 100.0 / n
utils.tic("Crank:")
R = R1 + R2
normal_shape = V.shape
transposed_shape = normal_shape[::-1]
for k in xrange(n):
if not k % print_step:
if isnan(V).any():
print "Crank fail @ t = %f (%i steps)" % (dt * k, k)
return crumbs
print int(k * to_percent),
if callback is not None:
callback(V, ((n - k) * dt))
V = (L2e.dot(V.flat).reshape(normal_shape) + R).T
V = spl.solve_banded(offsets1, L1i.data, V.flat, overwrite_b=True)
V = (L1e.dot(V).reshape(transposed_shape).T) + R
V = spl.solve_banded(offsets2, L2i.data, V.flat, overwrite_b=True).reshape(normal_shape)
crumbs.append(V.copy())
utils.toc()
return crumbs
def impl(V, L1, R1x, L2, R2x, dt, n, crumbs=[], callback=None):
V = V.copy()
# L1i = flatten_tensor(L1)
L1i = L1.copy()
R1 = np.array(R1x).T
# L2i = flatten_tensor(L2)
L2i = L2.copy()
R2 = np.array(R2x)
m = 2
# L = (As + Ass - H.interest_rate*np.eye(nspots))*-dt + np.eye(nspots)
L1i.data *= -dt
L1i.data[m, :] += 1
R1 *= dt
L2i.data *= -dt
L2i.data[m, :] += 1
R2 *= dt
offsets1 = (abs(min(L1i.offsets)), abs(max(L1i.offsets)))
offsets2 = (abs(min(L2i.offsets)), abs(max(L2i.offsets)))
dx = np.gradient(spots)[:,np.newaxis]
dy = np.gradient(vars)
X, Y = [dim.T for dim in np.meshgrid(spots, vars)]
gradgrid = dt * coeffs[(0,1)](0, X, Y) / (dx * dy)
gradgrid[:,0] = 0; gradgrid[:,-1] = 0
gradgrid[0,:] = 0; gradgrid[-1,:] = 0
print_step = max(1, int(n / 10))
to_percent = 100.0 / n
utils.tic("Impl:")
for k in xrange(n):
if not k % print_step:
if np.isnan(V).any():
print "Impl fail @ t = %f (%i steps)" % (dt * k, k)
return crumbs
print int(k * to_percent),
if callback is not None:
callback(V, ((n - k) * dt))
Vsv = np.gradient(np.gradient(V)[0])[1] * gradgrid
V = spl.solve_banded(offsets2, L2i.data,
(V + Vsv + R2).flat, overwrite_b=True).reshape(V.shape)
V = spl.solve_banded(offsets1, L1i.data,
(V + R1).T.flat, overwrite_b=True).reshape(V.shape[::-1]).T
crumbs.append(V.copy())
utils.toc()
return crumbs
def train(self):
data = self.train_data
labels = self.train_labels
for epoch in range(self.epochs):
utils.tic()
total_loss = 0.0 # every epoch loss should start with zero
good = 0.0
total_size = 0.0
# TODO: shuffle?
data, labels = utils.shuffle(data, labels)
for d, l in zip(data, labels):
total_size += 1
pred, cache = self.fprop(d)
# check the prediction
y_hat = np.argmax(pred)
if y_hat == l:
good += 1
err_cost = float(pred[int(l)]) # loss = -1 * log(err_cost)
cross_entropy = utils.cross_entropy_loss(err_cost)
if self.L2:
cross_entropy += utils.L2_cost(self.parameters["W"],
self.L2)
total_loss += cross_entropy
grads = self.bprop(cache, d, l)
self.weights_updates(grads)
print('epoch {}:'.format(epoch + 1))
acc = good * 100 / total_size
train_acc.append(acc)
avg_loss = total_loss / total_size
train_loss.append(avg_loss)
print('train accuracy: {:2.2f}%'.format(acc))
print('train AVG loss: {:2.2f}'.format(avg_loss))
self.validation_acc()
print('time:')
utils.toc()
# end of epoch
# cache all about model
trained_model = {
"norm": self.norm,
"parameters": self.parameters,
"lr": self.lr
}
directory = str(len(self.hidden)) + 'Hidden/L2/'
np.save(directory + 'model_' + self.model_name, trained_model)
self.printGraph(directory)
def mainLoop(modelType,modelArgs,positives,trueObjectBoxes,trainingList,featuresDir,featuresExt,modelOut,maxNegOverlap,iter):
pos,posIdx,ari,osi = positives
startTime = cu.tic()
if iter == 0:
## Random Negatives
print ' >>> RANDOM NEGATIVES'
neg,negIdx = learn.getRandomNegs(featuresDir,trainingList,featuresExt,pos.shape[1],maxVectorsCache,maxNegativeImages)
detectionsList = [ [x[0],'0.0']+x[1:]+['1'] for x in negIdx]
hards = {'features':np.zeros((0,neg.shape[1])),'index':[]}
lap = cu.toc('Random negatives matrix ('+str(neg.shape[0])+' instances)',startTime)
else:
## Mine hard negatives
print ' >>> MINING HARD NEGATIVES'
model = det.createDetector(modelType,modelArgs)
model.load(modelOut+'.'+ str( iter-1 ))
detectionsList = detector.detectObjects(model,trainingList,featuresDir,featuresExt,0.3,-10.0)
hards = cu.loadMatrixNoCompression(modelOut+'.hards').item()
lap = cu.toc('Hard negatives matrix ('+str(hards['features'].shape[0])+' instances)',startTime)
## Rank and clean negative detections
detectionsData = evaluation.loadDetections(detectionsList)
groundTruth = evaluation.loadGroundTruthAnnotations(trueObjectBoxes)
detectionsLog = evaluation.evaluateDetections(groundTruth,detectionsData,0.5,allowDuplicates=True) # overlapMeasure=validRegion,
evaluation.computePrecisionRecall(len(posIdx),detectionsLog['tp'],detectionsLog['fp'],'tmp.txt')
evaluation.computePrecAt(detectionsLog['tp'],[20,50,100,200,300,400,500])
logData = learn.parseRankedDetectionsFile(detectionsLog['log'],maxNegOverlap,maxNegativeVectors)
print ' >>> LOADING HARD NEGATIVES'
neg,negIdx = learn.loadHardNegativesFromList(featuresDir,logData['negExamples'],featuresExt,pos.shape[1],logData['negTaken'])
del(detectionsList,detectionsData,detectionsLog,logData)
lap = cu.toc('Ranked negatives matrix ('+str(neg.shape[0])+' instances)',lap)
neg = np.concatenate( (neg,hards['features']) )
negIdx = negIdx + hards['index']
## Learn Detector
clf = det.createDetector(modelType,modelArgs)
clf.learn(pos,neg,posIdx,negIdx)
clf.save(modelOut+'.'+str(iter))
lap = cu.toc('Classifier learned:',lap)
## Keep hard negatives for next iterations
scores = clf.predict(neg,negIdx)
hardNegsIdx = np.argsort(scores)
hardNeg = np.concatenate( (hards['features'], neg[hardNegsIdx[-cu.topHards:]]) )
negIdx = hards['index'] + [negIdx[j] for j in hardNegsIdx[-cu.topHards:]]
print 'Hard negatives:',hardNeg.shape[0]
hards = {'features':hardNeg, 'index':negIdx}
cu.saveMatrixNoCompression({'features':hardNeg,'index':negIdx},modelOut+'.hards')
print ' ** Iteration',iter,'done'
return {'detector':clf,'pos':pos,'posIdx':posIdx,'neg':neg,'negIdx':negIdx}
def processImg(info, filename, idx, batchSize, layers, output):
startTime = tic()
allFeat = {}
n = len(info)
for l in layers.keys():
allFeat[l] = emptyMatrix([n,layers[l]['dim']])
numBatches = (n + batchSize - 1) / batchSize
# Write the index file
[idx.write(b[4]) for b in info]
# Prepare boxes, make sure that extra rows are added to fill the last batch
boxes = [x[:-1] for x in info] + [ [0,0,0,0] for x in range(numBatches * batchSize - n) ]
# Initialize the image
net.caffenet.InitializeImage(filename, ImageNetMean)
for k in range(numBatches):
s,f = k*batchSize,(k+1)*batchSize
e = batchSize if f <= n else n-s
# Forward this batch
net.caffenet.ForwardRegions(boxes[s:f])
#outputBlobs = [ np.empty((batch, 1000, 1, 1), dtype=np.float32) ]
#net.caffenet.ForwardRegions(boxes[s:f],filename, outputBlobs)
#print outputBlobs[0][0].shape, np.argmax(outputBlobs[0][0]), np.max(outputBlobs[0][0])
outputs = net.caffenet.blobs()
f = n if f > n else f
# Collect outputs
for l in layers.keys():
allFeat[l][s:f,:] = outputs[layers[l]['idx']].data[0:e,:,:,:].reshape([e,layers[l]['dim']])
# Release image data
net.caffenet.ReleaseImageData()
# Save files for this image
for l in layers.keys():
saveMatrix(allFeat[l][0:n,:],output+'.'+l)
lap = toc('GPU is done with '+str(len(info))+' boxes in:',startTime)
def processImg(info, filename, batchSize, layers):
startTime = tic()
allFeat = {}
n = len(info)
for l in layers.keys():
allFeat[l] = emptyMatrix([n,layers[l]['dim']])
numBatches = (n + batchSize - 1) / batchSize
# Prepare boxes, make sure that extra rows are added to fill the last batch
boxes = [x[:-1] for x in info] + [ [0,0,0,0] for x in range(numBatches * batchSize - n) ]
# Initialize the image
net.caffenet.InitializeImage(filename, IMG_DIM, ImageNetMean, CROP_SIZE)
for k in range(numBatches):
s,f = k*batchSize,(k+1)*batchSize
e = batchSize if f <= n else n-s
# Forward this batch
net.caffenet.ForwardRegions(boxes[s:f],CONTEXT_PAD)
outputs = net.caffenet.blobs
f = n if f > n else f
# Collect outputs
for l in layers.keys():
allFeat[l][s:f,:] = outputs[layers[l]['idx']].data[0:e,:,:,:].reshape([e,layers[l]['dim']])
# Release image data
net.caffenet.ReleaseImageData()
# Return features of boxes for this image
for l in layers.keys():
allFeat[l] = allFeat[l][0:n,:]
lap = toc('GPU is done with '+str(len(info))+' boxes in:',startTime)
return allFeat
def fit_gen_model(phi):
assert phi.ndim == 2, 'data has to be two-dimensional'
assert phi.shape[1] > phi.shape[0], 'data samples have to be in columns'
d, nsamples = phi.shape
nij = 4 * d**2 # number of coupling terms
adata, arow, acol, b = fill_gen_model_matrix(phi)
a = sparse.coo_matrix((adata, (arow, acol)), (nij, nij))
tic('matrix inversion')
m_vec, flag = isolve.cg(a.tocsr(), b)
# print 'exit flag = ', flag
assert flag == 0
m = m_vec2mat(m_vec)
toc('matrix inversion')
return m
def processImg(info, filename, idx, batchSize, layers):
startTime = tic()
allFeat = {}
n = len(info)
for l in layers.keys():
allFeat[l] = emptyMatrix([n, layers[l]['dim']])
numBatches = (n + batchSize - 1) / batchSize
# Write the index file
[idx.write(b[4]) for b in info]
# Prepare boxes, make sure that extra rows are added to fill the last batch
boxes = [x[:-1]
for x in info] + [[0, 0, 0, 0]
for x in range(numBatches * batchSize - n)]
# Initialize the image
net.caffenet.InitializeImage(filename, IMG_DIM, ImageNetMean, CROP_SIZE)
for k in range(numBatches):
s, f = k * batchSize, (k + 1) * batchSize
e = batchSize if f <= n else n - s
# Forward this batch
net.caffenet.ForwardRegions(boxes[s:f], CONTEXT_PAD) #,filename)
outputs = net.caffenet.blobs
f = n if f > n else f
# Collect outputs
for l in layers.keys():
allFeat[l][s:f, :] = outputs[layers[l]['idx']].data[
0:e, :, :, :].reshape([e, layers[l]['dim']])
# Release image data
net.caffenet.ReleaseImageData()
# Return features of boxes for this image
for l in layers.keys():
allFeat[l] = allFeat[l][0:n, :]
lap = toc('GPU is done with ' + str(len(info)) + ' boxes in:', startTime)
return allFeat
def processImg(imgName, filename, idx, batchSize, layers, output):
startTime = tic()
# Initialize image and boxes
dims = net.caffenet.InitializeImage(filename, IMG_DIM, ImageNetMean, CROP_SIZE)
boxes = multiScaleBoxes(dims, CROP_SIZE)
# Write index file
[idx.write(imgName + ' ' + ' '.join(map(str,b)) + '\n') for b in boxes]
#Prepare boxes, make sure that extra rows are added to fill the last batch
allFeat = {}
n = len(boxes)
for l in layers.keys():
allFeat[l] = emptyMatrix([n,layers[l]['dim']])
numBatches = (n + batchSize - 1) / batchSize
boxes += [ [0,0,0,0] for x in range(numBatches * batchSize - n) ]
for k in range(numBatches):
s,f = k*batchSize,(k+1)*batchSize
e = batchSize if f <= n else n-s
# Forward this batch
net.caffenet.ForwardRegions(boxes[s:f],CONTEXT_PAD) #,filename)
outputs = net.caffenet.blobs
f = n if f > n else f
# Collect outputs
for l in layers.keys():
allFeat[l][s:f,:] = outputs[layers[l]['idx']].data[0:e,:,:,:].reshape([e,layers[l]['dim']])
# Release image data
net.caffenet.ReleaseImageData()
# Save files for this image
for l in layers.keys():
saveMatrix(allFeat[l][0:n,:],output+'.'+l)
lap = toc('GPU is done with '+str(n)+' boxes in:',startTime)
def fit_gen_model(phi):
assert phi.ndim == 2, 'data has to be two-dimensional'
assert phi.shape[1] > phi.shape[0], 'data samples have to be in columns'
d, nsamples = phi.shape
nij = 4*d**2 # number of coupling terms
adata, arow, acol, b = fill_gen_model_matrix(phi)
a = sparse.coo_matrix((adata,(arow,acol)), (nij,nij))
tic('matrix inversion')
m_vec,flag = isolve.cg(a.tocsr(),b)
# print 'exit flag = ', flag
assert flag==0
m = m_vec2mat(m_vec)
toc('matrix inversion')
return m
def fill_model_matrix(phi):
z = np.concatenate((np.exp(1j*phi),np.exp(-1j*phi)))
d, nsamples = phi.shape
z.shape = (2,d,nsamples)
nij = d**2-d # number of coupling terms
na = 4*d**3-10*d**2+6*d # upper bound for number of elements in sparse matrix
adata = np.zeros(na,complex)
arow = np.zeros(na,int)
acol = np.zeros(na,int)
b = np.zeros(nij,complex)
tic('weave')
weave.inline(phasemodel_code_blitz, ['z','adata','arow','acol','b'],
type_converters=weave.converters.blitz)
toc('weave')
return adata, arow, acol, b
def processImg(imgName, filename, idx, batchSize, layers, output):
startTime = tic()
# Initialize image and boxes
dims = net.caffenet.InitializeImage(filename, IMG_DIM, ImageNetMean,
CROP_SIZE)
boxes = multiScaleBoxes(dims, CROP_SIZE)
# Write index file
[idx.write(imgName + ' ' + ' '.join(map(str, b)) + '\n') for b in boxes]
#Prepare boxes, make sure that extra rows are added to fill the last batch
allFeat = {}
n = len(boxes)
for l in layers.keys():
allFeat[l] = emptyMatrix([n, layers[l]['dim']])
numBatches = (n + batchSize - 1) / batchSize
boxes += [[0, 0, 0, 0] for x in range(numBatches * batchSize - n)]
for k in range(numBatches):
s, f = k * batchSize, (k + 1) * batchSize
e = batchSize if f <= n else n - s
# Forward this batch
net.caffenet.ForwardRegions(boxes[s:f], CONTEXT_PAD) #,filename)
outputs = net.caffenet.blobs
f = n if f > n else f
# Collect outputs
for l in layers.keys():
allFeat[l][s:f, :] = outputs[layers[l]['idx']].data[
0:e, :, :, :].reshape([e, layers[l]['dim']])
# Release image data
net.caffenet.ReleaseImageData()
# Save files for this image
for l in layers.keys():
saveMatrix(allFeat[l][0:n, :], output + '.' + l)
lap = toc('GPU is done with ' + str(n) + ' boxes in:', startTime)
def train(self):
networkFile = config.get('networkDir') + config.get(
'snapshotPrefix') + '_iter_' + config.get(
'trainingIterationsPerBatch') + '.caffemodel'
interactions = config.geti('trainInteractions')
minEpsilon = config.getf('minTrainingEpsilon')
epochSize = len(self.environment.imageList) / 1
epsilon = 1.0
self.controller.setEpsilonGreedy(epsilon,
self.environment.sampleAction)
epoch = 1
exEpochs = config.geti('explorationEpochs')
while epoch <= exEpochs:
s = cu.tic()
print 'Epoch', epoch, ': Exploration (epsilon=1.0)'
self.runEpoch(interactions, len(self.environment.imageList))
self.task.flushStats()
self.doValidation(epoch)
s = cu.toc('Epoch done in ', s)
epoch += 1
self.learner = QLearning()
self.agent.learner = self.learner
egEpochs = config.geti('epsilonGreedyEpochs')
while epoch <= egEpochs + exEpochs:
s = cu.tic()
epsilon = epsilon - (1.0 - minEpsilon) / float(egEpochs)
if epsilon < minEpsilon: epsilon = minEpsilon
self.controller.setEpsilonGreedy(epsilon,
self.environment.sampleAction)
print 'Epoch', epoch, '(epsilon-greedy:{:5.3f})'.format(epsilon)
self.runEpoch(interactions, epochSize)
self.task.flushStats()
self.doValidation(epoch)
s = cu.toc('Epoch done in ', s)
epoch += 1
maxEpochs = config.geti('exploitLearningEpochs') + exEpochs + egEpochs
while epoch <= maxEpochs:
s = cu.tic()
print 'Epoch', epoch, '(exploitation mode: epsilon={:5.3f})'.format(
epsilon)
self.runEpoch(interactions, epochSize)
self.task.flushStats()
self.doValidation(epoch)
s = cu.toc('Epoch done in ', s)
shutil.copy(networkFile, networkFile + '.' + str(epoch))
epoch += 1
def load_data(self, file_path, data_loading_fn, *args, **kwargs):
# print(os.path.join(data_path,filename))
local_file_path = os.path.expanduser(file_path)
self.file_path = local_file_path
# print(os.path.isfile(os.path.join(local_file_path,filename)))
print('loading file: %s' %file_path)
if not os.path.isfile(local_file_path) and self.main_hostname != 'localhost':
print('Copying data from remote machine...')
tic()
ssh.copy_file_from_remote(file_path, self.main_hostname, self.username)
print('Completed copying data from main host.')
toc()
print('Loading data...')
tic()
data = data_loading_fn(local_file_path, *args, **kwargs)
toc()
return data
def fill_gen_model_matrix(phi):
d, nsamples = phi.shape
x = p2torus(phi)
q = p2dtorus(phi)
x.shape = (d,2,nsamples)
q.shape = (d,2,nsamples)
nij = 4*d**2 # number of coupling terms
na = 32*d**3 - 16*d**2 # number of elements in large matrix multiplying mij
adata = np.zeros(na,float)
arow = np.zeros(na,int)
acol = np.zeros(na,int)
b = np.zeros(nij,float)
tic('weave')
weave.inline(gen_phasemodel_code, ['x','q','adata','arow','acol','b'])
toc('weave')
return adata, arow, acol, b
def run(self, image, features, boxes):
s = cu.tic()
result = {}
boxSet = [ map(float, b[1:]) for b in boxes ]
for i in self.catIndex:
scores = features[:,i]
fb,fs = det.nonMaximumSuppression(boxSet, scores, self.maxOverlap)
result[i] = (image, fb, fs)
s = cu.toc(image, s)
return result
def runEpoch(self, interactions, maxImgs):
img = 0
s = cu.tic()
while img < maxImgs:
self.experiment.doInteractions(interactions)
self.agent.learn()
self.agent.reset()
self.environment.loadNextEpisode()
img += 1
s = cu.toc('Run epoch with ' + str(maxImgs) + ' episodes', s)
def runEpoch(self, interactions, maxImgs):
img = 0
s = cu.tic()
while img < maxImgs:
self.experiment.doInteractions(interactions)
self.agent.learn()
self.agent.reset()
self.environment.loadNextEpisode()
img += 1
s = cu.toc('Run epoch with ' + str(maxImgs) + ' episodes', s)
def run(self, image, features, boxes):
s = cu.tic()
result = {}
boxSet = [map(float, b[1:]) for b in boxes]
for i in self.catIndex:
scores = features[:, i]
fb, fs = det.nonMaximumSuppression(boxSet, scores, self.maxOverlap)
result[i] = (image, fb, fs)
s = cu.toc(image, s)
return result
def impl(V, L1, R1x, L2, R2x, dt, n, crumbs=[], callback=None):
V = V.copy()
L1i = flatten_tensor(L1)
R1 = np.array(R1x).T
L2i = flatten_tensor(L2)
R2 = np.array(R2x)
m = 2
# L = (As + Ass - r*np.eye(nspots))*-dt + np.eye(nspots)
L1i.data *= -dt
L1i.data[m, :] += 1
R1 *= dt
L2i.data *= -dt
L2i.data[m, :] += 1
R2 *= dt
offsets1 = (abs(min(L1i.offsets)), abs(max(L1i.offsets)))
offsets2 = (abs(min(L2i.offsets)), abs(max(L2i.offsets)))
print_step = max(1, int(n / 10))
to_percent = 100.0 / n
utils.tic("Impl:")
for k in xrange(n):
if not k % print_step:
if isnan(V).any():
print "Impl fail @ t = %f (%i steps)" % (dt * k, k)
return crumbs
print int(k * to_percent),
if callback is not None:
callback(V, ((n - k) * dt))
V = spl.solve_banded(offsets2,
L2i.data, (V + R2).flat,
overwrite_b=True).reshape(V.shape)
V = spl.solve_banded(offsets1,
L1i.data, (V + R1).T.flat,
overwrite_b=True).reshape(V.shape[::-1]).T
crumbs.append(V.copy())
utils.toc()
return crumbs
def train(self):
networkFile = config.get('networkDir') + config.get('snapshotPrefix') + '_iter_' + config.get('trainingIterationsPerBatch') + '.caffemodel'
interactions = config.geti('trainInteractions')
minEpsilon = config.getf('minTrainingEpsilon')
epochSize = len(self.environment.imageList)/1
epsilon = 1.0
self.controller.setEpsilonGreedy(epsilon, self.environment.sampleAction)
epoch = 1
exEpochs = config.geti('explorationEpochs')
while epoch <= exEpochs:
s = cu.tic()
print 'Epoch',epoch,': Exploration (epsilon=1.0)'
self.runEpoch(interactions, len(self.environment.imageList))
self.task.flushStats()
self.doValidation(epoch)
s = cu.toc('Epoch done in ',s)
epoch += 1
self.learner = QLearning()
self.agent.learner = self.learner
egEpochs = config.geti('epsilonGreedyEpochs')
while epoch <= egEpochs + exEpochs:
s = cu.tic()
epsilon = epsilon - (1.0-minEpsilon)/float(egEpochs)
if epsilon < minEpsilon: epsilon = minEpsilon
self.controller.setEpsilonGreedy(epsilon, self.environment.sampleAction)
print 'Epoch',epoch ,'(epsilon-greedy:{:5.3f})'.format(epsilon)
self.runEpoch(interactions, epochSize)
self.task.flushStats()
self.doValidation(epoch)
s = cu.toc('Epoch done in ',s)
epoch += 1
maxEpochs = config.geti('exploitLearningEpochs') + exEpochs + egEpochs
while epoch <= maxEpochs:
s = cu.tic()
print 'Epoch',epoch,'(exploitation mode: epsilon={:5.3f})'.format(epsilon)
self.runEpoch(interactions, epochSize)
self.task.flushStats()
self.doValidation(epoch)
s = cu.toc('Epoch done in ',s)
shutil.copy(networkFile, networkFile + '.' + str(epoch))
epoch += 1
def impl(V, L1, R1x, L2, R2x, dt, n, crumbs=[], callback=None):
V = V.copy()
L1i = flatten_tensor(L1)
R1 = np.array(R1x).T
L2i = flatten_tensor(L2)
R2 = np.array(R2x)
m = 2
# L = (As + Ass - r*np.eye(nspots))*-dt + np.eye(nspots)
L1i.data *= -dt
L1i.data[m, :] += 1
R1 *= dt
L2i.data *= -dt
L2i.data[m, :] += 1
R2 *= dt
offsets1 = (abs(min(L1i.offsets)), abs(max(L1i.offsets)))
offsets2 = (abs(min(L2i.offsets)), abs(max(L2i.offsets)))
print_step = max(1, int(n / 10))
to_percent = 100.0 / n
utils.tic("Impl:")
for k in xrange(n):
if not k % print_step:
if isnan(V).any():
print "Impl fail @ t = %f (%i steps)" % (dt * k, k)
return crumbs
print int(k * to_percent),
if callback is not None:
callback(V, ((n - k) * dt))
V = spl.solve_banded(offsets2, L2i.data,
(V + R2).flat, overwrite_b=True).reshape(V.shape)
V = spl.solve_banded(offsets1, L1i.data,
(V + R1).T.flat, overwrite_b=True).reshape(V.shape[::-1]).T
crumbs.append(V.copy())
utils.toc()
return crumbs
def runEpoch(self, interactions, maxImgs):
img = 0
s = cu.tic()
while img < maxImgs:
k = 0
while not self.environment.episodeDone and k < interactions:
self.experiment._oneInteraction()
k += 1
self.agent.learn()
self.agent.reset()
self.environment.loadNextEpisode()
img += 1
s = cu.toc('Run epoch with ' + str(maxImgs) + ' episodes', s)
def runEpoch(self, interactions, maxImgs):
img = 0
s = cu.tic()
while img < maxImgs:
k = 0
while not self.environment.episodeDone and k < interactions:
self.experiment._oneInteraction()
k += 1
self.agent.learn()
self.agent.reset()
self.environment.loadNextEpisode()
img += 1
s = cu.toc('Run epoch with ' + str(maxImgs) + ' episodes', s)
def do_next_group(self, tiffs_to_run):
"""
Replaces the OnlineAnalysis.do_next_group() so we can fake experiments and test downstream
analysis.
"""
t = tic()
self.validate_tiffs()
self.opts.change_params(dict(fnames=tiffs_to_run))
self.make_mmap(tiffs_to_run)
self.make_movie()
self.C = self.do_fit()
self.trial_lengths.append(self.splits)
self.group_lenths.append(toc(t))
#self.save_json()
self.advance(by=1)
def timing_benchmark(eval_dim=None,dims=[2, 4, 6, 8, 10],nsamps=10**4):
ind = 0
t = np.zeros(len(dims),float)
for d in dims:
print 'Benchmarking dim: ', d
phi = 2*np.pi*np.random.rand(d,nsamps)
tic('fit parameters')
c_inv = fit_model(phi)
t[ind] = toc('fit parameters')
ind += 1
pol = np.polyfit(dims[1:],t[1:],3)
if eval_dim:
print np.polyval(pol,eval_dim)
else:
return pol
def timing_benchmark(eval_dim=None, dims=[2, 4, 6, 8, 10], nsamps=10**4):
ind = 0
t = np.zeros(len(dims), float)
for d in dims:
print 'Benchmarking dim: ', d
phi = 2 * np.pi * np.random.rand(d, nsamps)
tic('fit parameters')
c_inv = fit_model(phi)
t[ind] = toc('fit parameters')
ind += 1
pol = np.polyfit(dims[1:], t[1:], 3)
if eval_dim:
print np.polyval(pol, eval_dim)
else:
return pol
def computeFeatures(batch, n, data, net, layers, output):
startTime = tic()
allFeat = {}
for l in layers.keys():
allFeat[l] = emptyMatrix([n,layers[l]['dim']])
# Extract and store CNN Features
outputBlobs = [np.empty((batch, 1000, 1, 1), dtype=np.float32)]
for i in range(batch,n+batch,batch):
inputBlobs = np.empty((batch, 3, 227, 227), dtype=np.float32)
start = i-batch
finish = min(i,n)
elems = finish-start
inputBlobs[0:elems,:,:,:] = data[start:finish,:,:,:]
net.caffenet.Forward([inputBlobs], outputBlobs)
outputs = net.caffenet.blobs()
for l in layers.keys():
allFeat[l][start:finish,:] = outputs[layers[l]['idx']].data[0:elems,:,:,:].reshape([elems,layers[l]['dim']])
# Save files for this image
for l in layers.keys():
saveMatrix(allFeat[l][0:n,:],output+'.'+l)
lap = toc('Image ready with '+str(n)+' boxes in:',startTime)
def computeFeatures(batch, n, data, net, layers, output):
startTime = tic()
allFeat = {}
for l in layers.keys():
allFeat[l] = emptyMatrix([n, layers[l]['dim']])
# Extract and store CNN Features
outputBlobs = [np.empty((batch, 1000, 1, 1), dtype=np.float32)]
for i in range(batch, n + batch, batch):
inputBlobs = np.empty((batch, 3, 227, 227), dtype=np.float32)
start = i - batch
finish = min(i, n)
elems = finish - start
inputBlobs[0:elems, :, :, :] = data[start:finish, :, :, :]
net.caffenet.Forward([inputBlobs], outputBlobs)
outputs = net.caffenet.blobs()
for l in layers.keys():
allFeat[l][start:finish, :] = outputs[layers[l]['idx']].data[
0:elems, :, :, :].reshape([elems, layers[l]['dim']])
# Save files for this image
for l in layers.keys():
saveMatrix(allFeat[l][0:n, :], output + '.' + l)
lap = toc('Image ready with ' + str(n) + ' boxes in:', startTime)
import utils as cu
import libDetection as det
from dataProcessor import processData
class Checker():
def __init__(self):
print 'Starting checker'
def run(self,img,features,bboxes):
return img,features.shape[0] == len(bboxes)
## Main Program Parameters
params = cu.loadParams("testImageList featuresDir featuresExt")
imageList = [x.replace('\n','') for x in open(params['testImageList'])]
## Run Detector
task = Checker()
start = cu.tic()
result = processData(imageList,params['featuresDir'],params['featuresExt'],task)
cu.toc('All images checked',start)
totalP = 0
for data in result:
img,r = data
if not r:
print 'Problems with',img
totalP += 1
print 'Total problems:',totalP
def main(argv):
global STATES
year, state, ext, shift = (None, None, None, None)
usage = '05-cand-info-vices.py -y 2016 -s SC -t 1 or 2 -e data.csv'
try:
opts, _args = getopt.getopt(
argv, 'hy:s:t:e:', ['year=', 'state=', 'shift=', 'ext='])
except getopt.GetoptError:
print(usage)
sys.exit()
for opt, arg in opts:
if opt == '-h':
print(usage)
sys.exit()
elif opt in ('-y', '--year'):
year = arg
elif opt in ('-s', '--state'):
state = arg
elif opt in ('-t', '--shift'):
shift = arg
elif opt in ('-e', '--ext='):
ext = arg
if year == '2020' or year == '2016' or year == '2012':
STATES.remove('DF')
else:
print('Year is invalid!')
print(usage)
sys.exit()
if state:
STATES = list(filter(lambda x: x == str(state), STATES))
if not shift:
print('The shift number is required! 1 or 2 shift.')
sys.exit()
engine = create_engine(DATABASE, echo=False)
tic()
for st in STATES:
print(
'Reading candidates (%s) for vice-mayor of all cities in the state of: %s in shift: %s' %
(year, st, shift))
df = pd.read_sql("""SELECT nm_city FROM cand_info
WHERE election_year = '{}' AND sg_uf = '{}' GROUP BY 1 ORDER BY 1""".format(year, st), engine)
# df = pd.read_sql("""SELECT nm_city FROM cand_info
# WHERE election_year = '{}' AND sg_uf = '{}' AND nm_city = 'OSASCO'
# GROUP BY 1 ORDER BY 1""".format(year, st), engine)
dfcount = df['nm_city'].count()
bar = Bar('Progress', max=dfcount)
# for ct in ['RIO BRANCO']:
for ct in df['nm_city'].tolist():
if year == '2020' or year == '2016' or year == '2012':
df0 = pd.read_sql("""
SELECT
t2.election_year,
t2.sg_uf,
t1.sq_candidate,
t1.nr_cpf_candidate,
t1.nm_candidate,
t1.sg_party,
t1.nr_party,
t1.nm_ballot_candidate,
t1.ds_position,
t1.ds_situ_tot_shift,
t2.nm_city,
t1.ds_situ_cand,
t1.nm_email,
t1.ds_genre,
t1.ds_degree_instruction,
t1.ds_race_color,
t1.ds_occupation,
t1.nr_campaign_max_expenditure,
t1.st_reelection,
t1.dt_birth,
t1.nr_shift,
t1.ds_election,
t1.sq_alliance
FROM raw_tse_consult_candidates t1
INNER JOIN cand_info AS t2 ON (t1.sq_alliance = t2.sq_alliance)
WHERE t1.election_year = '{}' AND t1.sg_uf = '{}'
AND t1.cd_position = 12 AND t2.nm_city = "{}" AND t1.nr_shift = '{}'
GROUP BY 1,2,3,4,5,6,7,8,9,10,11,12,13,14,15,16,17,18,19,20,21,22,23""".format(year, st, ct, shift), engine)
df0 = df0.applymap(
lambda s: s.upper() if isinstance(
s, str) else s)
df1 = pd.read_sql("""SELECT * FROM cand_info
WHERE sg_uf = '{}' AND nm_city = "{}"
AND ds_position = 'PREFEITO' ORDER BY qt_votes_nominal_int DESC""".format(st, ct), engine)
df2 = pd.merge(df1, df0, on='sq_alliance', how='inner')
df3 = df2[COLS_VICES_XY]
df4 = df3.rename(columns=COLS_VICES_XY_NEW, inplace=False)
df4 = df4.applymap(
lambda s: s.upper() if isinstance(
s, str) else s)
if int(shift) == 1:
if any(df4['ds_situ_tot_shift'] == '2º TURNO'):
df4 = df4.where(df4['ds_situ_tot_shift'] != '2º TURNO')
if any(df4['ds_situ_tot_shift'] == '#NULO#'):
df4.loc[df4['ds_situ_tot_shift'] == '#NULO#',
['ds_situ_tot_shift']] = 'NÃO ELEITO'
elif int(shift) == 2:
if not df4.empty:
df4.sort_values(
by=['ds_situ_tot_shift'],
inplace=False,
ascending=False)
rank = df4['qt_votes_nominal_int'].nlargest(4).tolist()
for i in range(len(rank)):
if i == 0:
df4.loc[df4['qt_votes_nominal_int'] == rank[0], [
'ds_situ_tot_shift']] = 'ELEITO'
if i == 1:
df4.loc[df4['qt_votes_nominal_int'] == rank[1], [
'ds_situ_tot_shift']] = '2º TURNO'
if i == 2:
df4.loc[df4['qt_votes_nominal_int'] == rank[2], [
'ds_situ_tot_shift']] = 'NÃO ELEITO'
if i == 3:
df4.loc[df4['qt_votes_nominal_int'] == rank[3], [
'ds_situ_tot_shift']] = '2º TURNO'
else:
pass
df5 = pd.read_sql("""SELECT * FROM cand_info
WHERE sg_uf = '{}' AND nm_city = "{}"
AND ds_position = 'VICE-PREFEITO'""".format(st, ct), engine)
for i in df4['sq_candidate'].tolist():
if any(df5['sq_candidate'] == i):
df4 = df4[df4['sq_candidate'] != i]
df6 = pd.read_sql("""
SELECT
sq_candidate,
format(sum(amount_goods_declared), 0, 'de_DE') as amount_goods_declared,
sum(amount_goods_declared) as amount_goods_declared_float
FROM raw_tse_cand_goods_declared
WHERE election_year = '{}' AND sg_uf = '{}' GROUP BY 1
ORDER BY 3 DESC""".format(year, st), engine)
df7 = pd.merge(df4, df6, on='sq_candidate', how='inner')
if df7.empty:
df4['amount_goods_declared'] = ''
df4['amount_goods_declared_float'] = 0
df7 = df4
if not df7.empty:
final = df7.sort_values(
by=['qt_votes_nominal'],
inplace=False,
ascending=False)
if ext:
write_to_csv(final)
final.to_sql(
con=engine,
name=CAND_TABLE_NAME,
if_exists='append',
index=False,
index_label=CAND_TABLE_NAME_ID)
else:
raise ValueError('Invalid year')
bar.next()
bar.finish()
toc()
def saveMatrix(matrix,outFile):
outf = open(outFile,'w')
np.savez_compressed(outf,matrix)
outf.close()
#################################
# Extract Features
#################################
startTime = tic()
totalItems = len(images)
layers = {'fc6_neuron_cudanet_out': {'dim':4096,'idx':'fc6'}, 'fc7_neuron_cudanet_out': {'dim':4096,'idx':'fc7'}}
batch = 50
print 'Extracting features for',totalItems,'total images'
for name in images:
# Check if files already exist
processed = 0
for l in layers.keys():
if os.path.isfile(outDir+'/'+name+'.'+l):
processed += 1
if processed == len(layers):
continue
# Get features for patches
indexFile = open(outDir+'/'+name+'.idx','w')
processImg(name, imgsDir+'/'+name+'.jpg', indexFile, batch, layers, outDir+'/'+name)
indexFile.close()
toc('Total processing time:',startTime)
plt.title("Rejection Method (Constant Function)")
plt.legend(["rho(x)", "f(x)=3/10", "Histogram data"])
plt.show()
plt.figure(4)
plt.hist(y_inverse2, bins, normed=1)
plt.plot(x, const_function)
plt.show()
area_cauchy_exact = 3 * 2 * math.atan(xmax) / pi
area_rho_exact = 0.921348
area_constant_exact = 0.3 * 20
# Acceptance Rates
print("Acceptance rate cauchy exact: {:.2f}%".format(area_rho_exact /
area_cauchy_exact * 100))
print("Acceptance rate rejection method (cauchy): {:.2f}%\n".format(
acceptance_rate))
print("Acceptance rate constant function exact: {:.2f}%".format(
area_rho_exact / area_constant_exact * 100))
print("Acceptance rate rejection method (constant): {:.2f}%\n".format(
acceptance_rate2))
# Save plots to pdf file
utils.save_fig_to_pdf("ComSim1.pdf", fig_functions, fig_inverse, fig_rejection,
fig_rejection2)
# Measure total program runtime
utils.toc(start_time)
print('\n\n************ Computing Distance Matrix ************\n\n')
for i in range(total_shapes):
f = open(save_path + '/' + PDs[i][0], 'rb')
pts1 = pickle.load(f)
f.close()
for j in range(i, total_shapes):
f = open(save_path + '/' + PDs[j][0], 'rb')
pts2 = pickle.load(f)
f.close()
for k in range(len(pts1)):
utils.tic()
distmat1[i, j] = distmat1[i, j] + utils.subspace_angle(
pts1[k], pts2[k])
time_taken_SubspaceAngle.append(utils.toc())
utils.tic()
distmat2[i, j] = distmat2[i, j] + utils.distChordalGrass(
pts1[k], pts2[k])
time_taken_Chordal.append(utils.toc())
if not (i + 1) % 8:
print('#', end=" ")
distmat = distmat1 + distmat1.T
f = open(code_path + '/distmat_SubspaceAngle_' + descriptor + '.pckl',
'wb')
pickle.dump(distmat, f)
f.close()
time_taken_SubspaceAngle = time_taken_SubspaceAngle[0:3000]
f = open(code_path + '/time_taken_SubspaceAngle_' + descriptor + '.pckl',
'wb')
dim = 3
M = np.random.randn(2*dim,2*dim)
M += M.T.copy()
for i in np.arange(M.shape[0]/2):
s = M[2*i,2*i] + M[2*i+1,2*i+1]
M[2*i,2*i] -= s/2
M[2*i+1,2*i+1] -= s/2
tic('sampling')
nsamples = 10**4
burnin = 10**3
lf_steps = 50
step_sz = .15
phi,E,diagn = sample_hmc(M,nsamples,burnin,lf_steps,step_sz,diagnostics=True)
toc('sampling')
tic('fiting')
M_hat = fit_model(phi)
Mneg_hat,Mpos_hat= m2kappa(M_hat)
# anti-symmetrize diagonal elements for estimation matrix
for i in np.arange(M_hat.shape[0]/2):
s = M_hat[2*i,2*i] + M_hat[2*i+1,2*i+1]
M_hat[2*i,2*i] -= s/2
M_hat[2*i+1,2*i+1] -= s/2
toc('fiting')
M_error = M - M_hat
M_max = max(abs(M).max(),abs(M_hat).max())
print 'M_error norm = ', (M_error**2).sum()
scipy.io.savemat(output, mat, do_compression=True)
##################################
# Organize boxes by source image
#################################
startTime = tic()
images = {}
for s,box in bboxes:
# Subtract 1 because RCNN proposals have 1-based indexes for Matlab
b = map(lambda x: int(x)-1,box[1:]) + [s.replace('\n','')]
try:
images[ box[0] ].append(b)
except:
images[ box[0] ] = [b]
lap = toc('Reading boxes file:',startTime)
groundTruth, categories = loadBoxAnnotationsFile(groundTruthFile)
print 'Found categories:',categories
lap = toc('Reading ground truth file:',lap)
#################################
# Extract Features
#################################
totalItems = len(bboxes)
del(bboxes)
layers = {'pool5': {'dim':9216,'idx':'pool5'}}
print 'Extracting features for',totalItems,'total regions'
for name in images.keys():
# Get window proposals
def __init__(self, boxes):
t = cu.tic()
self.auxBoxes = []
frame = [999,999,0,0]
id = 0
for box in boxes.tolist():
frame = min(frame[0:2], box[0:2]) + max(frame[2:],box[2:])
self.auxBoxes.append(Box(box, id))
id += 1
self.frame = map(int,frame)
self.auxBoxes.sort(key=lambda x:x.area, reverse=True)
self.adjacency = np.zeros( (len(self.auxBoxes),len(self.auxBoxes)) )
for i in range(len(self.auxBoxes)):
self.adjacency[i,i] = 1.0
for j in range(i+1,len(self.auxBoxes)):
iou = self.auxBoxes[i].IoU( self.auxBoxes[j] )
self.adjacency[i,j] = iou
self.adjacency[j,i] = iou
knn = range(-2,-GRAPH_NEIGHBORS-2,-1) # Avoid last element (same box)
self.graph = {'nodes':[], 'edges':[]}
for i in range(len(self.auxBoxes)):
center = self.auxBoxes[i].center()
node = {'data':{'id':str(i), 'box':self.auxBoxes[i].box},
'position':{'x':int(center[0]),'y':int(center[1])}}
self.graph['nodes'].append(node)
edges = {}
neighbors = np.argsort( self.adjacency[i,:] )
knn = [ (j,self.adjacency[i,j]) for j in neighbors[-NEAREST_NEIGHBORS-2:-2] ]
for j,iou in knn:
edge = { 'data': { 'id': str(i)+'_'+str(j), 'weight': iou, 'source': str(i), 'target': str(j) } }
self.graph['edges'].append(edge)
t = cu.toc('Graph construction:',t)
return
self.layout = []
scaleRange = len(auxBoxes)/SCALES
for s in range(SCALES):
scaleElems = []
self.layout.append([])
for i in range(scaleRange):
scaleElems.append(auxBoxes[scaleRange*s + i])
scaleElems.sort(key=lambda x:x.box[0], reverse=True)
horizontalRange = len(scaleElems)/HORIZONTAL_BINS
for h in range(HORIZONTAL_BINS):
horizontalRangeElems = []
self.layout[s].append([])
for j in range(horizontalRange):
horizontalRangeElems.append(scaleElems[horizontalRange*h + j])
horizontalRangeElems.sort(key=lambda x:x.box[1], reverse=True)
verticalRange = len(horizontalRangeElems)/VERTICAL_BINS
for v in range(VERTICAL_BINS):
self.layout[s][h].append([])
for k in range(verticalRange):
self.layout[s][h][v].append(horizontalRangeElems[verticalRange*v + k])
self.numBoxes = len(auxBoxes)
self.boxesPerBin = float(self.numBoxes)/WORLD_SIZE
self.actionCounter = 0
self.scale = SCALES/2 # (Greedy: Set to zero)
self.horizontal = 0
self.vertical = 0
self.percentExplored = 0
self.selectedIds = []
self.status = np.zeros( (SCALES, HORIZONTAL_BINS, VERTICAL_BINS), dtype=np.int32 )
self.currentPosition = np.zeros( (SCALES, HORIZONTAL_BINS, VERTICAL_BINS), dtype=np.int32 )
fetched = sess.run(fetches, feed_dict)
loss, p, gnorm = fetch(fetched, ('loss', 'p', 'gnorm'))
tensors, tensor_shapes = fetch(fetched,
('tensors', 'tensor_shapes'))
#####################################
losses.append(loss)
gnorms.append(gnorm)
TRAIN_PRED.extend(np.squeeze(p))
TRAIN_Y.extend(np.squeeze(b.y))
word_count = batcher.word_count(reset=False)
sec = U.toc(reset=False)
wps = int(word_count / sec)
if batch % FLAGS.print_every == 0:
if ACC:
acc = U.nacc(TRAIN_Y[-k:], TRAIN_PRED[-k:])
sys.stdout.write('\tacc={0:0.3f}'.format(acc))
else:
kappa = U.nkappa(TRAIN_Y[-k:], TRAIN_PRED[-k:])
sys.stdout.write('\tqwk={0:0.3f}'.format(kappa))
sys.stdout.write('|loss={0:0.3f}'.format(loss))
#sys.stdout.write('|ploss={0:0.3g}'.format(p_loss))
sys.stdout.write('|wps={0}'.format(wps))
sys.stdout.write('|bs={0}'.format(b.n))
#sys.stdout.write('|gnm={0:0.2f}'.format(gnorm))
sys.stdout.flush()
def crank(V, L1, R1x, L2, R2x, dt, n, crumbs=[], callback=None):
V = V.copy()
theta = 0.5
# dt *= 0.5
# L1e = flatten_tensor(L1)
L1e = L1.copy()
L1i = L1e.copy()
R1 = np.array(R1x).T
# L2e = flatten_tensor(L2)
L2e = L2.copy()
L2i = L2e.copy()
R2 = np.array(R2x)
# print "L var"
# fp(L2e.data, 2)
# print "FD op var"
# fp(F.operators[1].data, 2)
# print "diff"
# fp(F.operators[1].data - L2e.data, 2, 'f')
# assert np.allclose(F.operators[1].data, L2e.data)
m = 2
# L = (As + Ass - H.interest_rate*np.eye(nspots))*-dt + np.eye(nspots)
L1i.data *= -theta*dt
L1i.data[m, :] += 1
# R1 *= dt
L2i.data *= -theta*dt
L2i.data[m, :] += 1
# R2 *= dt
offsets1 = (abs(min(L1i.offsets)), abs(max(L1i.offsets)))
offsets2 = (abs(min(L2i.offsets)), abs(max(L2i.offsets)))
dx = np.gradient(spots)[:,np.newaxis]
dy = np.gradient(vars)
X, Y = [dim.T for dim in np.meshgrid(spots, vars)]
gradgrid = dt * coeffs[(0,1)](0, X, Y) / (dx*dy)
gradgrid[:,0] = 0; gradgrid[:,-1] = 0
gradgrid[0,:] = 0; gradgrid[-1,:] = 0
print_step = max(1, int(n / 10))
to_percent = 100.0 / n
utils.tic("Crank:")
R = R1 + R2
normal_shape = V.shape
transposed_shape = normal_shape[::-1]
for k in xrange(n):
if not k % print_step:
if np.isnan(V).any():
print "Crank fail @ t = %f (%i steps)" % (dt * k, k)
return crumbs
print int(k * to_percent),
if callback is not None:
callback(V, ((n - k) * dt))
Vsv = np.gradient(np.gradient(V)[0])[1] * gradgrid
# V12 = (V
# + Vsv
# + (1-theta)*dt*L1e.dot(V.T.flat).reshape(transposed_shape).T
# + (1-theta)*dt*L2e.dot(V.flat).reshape(normal_shape)
# + dt * R)
# V1 = spl.solve_banded(offsets2, L2i.data, V12.flat, overwrite_b=True).reshape(normal_shape)
# V = spl.solve_banded(offsets1, L1i.data, V1.T.flat, overwrite_b=True).reshape(transposed_shape).T
V1 = (L1e.dot(V.T.flat).reshape(transposed_shape)).T
V2 = (L2e.dot(V.flat).reshape(normal_shape))
Y0 = V + Vsv + dt*(V1 + V2 + R)
V1 = Y0 - theta * dt * L1e.dot(V.T.flat).reshape(transposed_shape).T
Y1 = spl.solve_banded(offsets1, L1i.data, V1.T.flat, overwrite_b=True).reshape(transposed_shape).T
V2 = Y1 - theta * dt * L2e.dot(V.flat).reshape(normal_shape)
Y2 = spl.solve_banded(offsets2, L2i.data, V2.flat, overwrite_b=True).reshape(normal_shape)
V = Y2
crumbs.append(V.copy())
utils.toc()
return crumbs
inQueue.task_done()
return True
##################################
# Organize boxes by source image
#################################
startTime = tic()
images = {}
for s,box in bboxes:
b = map(int,box[1:]) + [s]
try:
images[ box[0] ].append(b)
except:
images[ box[0] ] = [b]
lap = toc('Reading boxes file:',startTime)
#################################
# Extract Features
#################################
totalItems = len(bboxes)
del(bboxes)
layers = {'fc6_neuron_cudanet_out': {'dim':4096,'idx':15} }
batch = 200
taskQueue = Queue.Queue()
p = threading.Thread(target=worker, args=(taskQueue, net, layers, batch))
p.daemon = True
p.start()
print 'Extracting features for',totalItems,'total images'
for name in images.keys():
s = M[2 * i, 2 * i] + M[2 * i + 1, 2 * i + 1]
M[2 * i, 2 * i] -= s / 2
M[2 * i + 1, 2 * i + 1] -= s / 2
tic('sampling')
nsamples = 10**4
burnin = 10**3
lf_steps = 50
step_sz = .15
phi, E, diagn = sample_hmc(M,
nsamples,
burnin,
lf_steps,
step_sz,
diagnostics=True)
toc('sampling')
tic('fiting')
M_hat = fit_model(phi)
Mneg_hat, Mpos_hat = m2kappa(M_hat)
# anti-symmetrize diagonal elements for estimation matrix
for i in np.arange(M_hat.shape[0] / 2):
s = M_hat[2 * i, 2 * i] + M_hat[2 * i + 1, 2 * i + 1]
M_hat[2 * i, 2 * i] -= s / 2
M_hat[2 * i + 1, 2 * i + 1] -= s / 2
toc('fiting')
M_error = M - M_hat
M_max = max(abs(M).max(), abs(M_hat).max())
print 'M_error norm = ', (M_error**2).sum()
V = np.copy(V_init)
# bs, delta = [x for x in bs_call_delta(spots[:, newaxis], k, r,
# np.sqrt(vars)[newaxis, :], t)]
bs = BlackScholesOption(spot=spots[:, np.newaxis],
strike=k,
interest_rate=r,
variance=vars[np.newaxis, :],
tenor=t).analytical
utils.tic("Heston Analytical:")
# hss = array([hs_call(spots, k, r, np.sqrt(vars),
# dt*i, kappa, theta, sigma, rho) for i in range(int(t/dt)+1)])
# hs = hss[-1]
hs = hs_call_vector(spots, k, r, np.sqrt(vars),
t, kappa, theta, sigma, rho)
utils.toc()
hs[isnan(hs)] = 0.0
if max(hs.flat) > spots[-1] * 2:
BADANALYTICAL = True
print "Warning: Analytical solution looks like trash."
if len(sys.argv) > 1:
if sys.argv[1] == '0':
print "Bail out with arg 0."
sys.exit()
L1_ = []
R1_ = []
utils.tic("Building As(s):")
print "(Up/Down)wind from:", flip_idx_spot
As_ = utils.nonuniform_complete_coefficients(dss, up_or_down=up_or_down_spot,
for i in range(len(self.imgBoxes[img])):
box = map( int, self.imgBoxes[img][i,:].tolist() )
key = img + ' ' + ' '.join( map(str, box) )
try: score = records[key]
except: score = -10.0
self.scores[img][i,fileIdx] = score
def saveDB(self, outputDir):
for img in self.imgBoxes.keys():
data = {'boxes':self.imgBoxes[img], 'scores':self.scores[img]}
scipy.io.savemat(outputDir+'/'+img+'.mat', data, do_compression=True)
out = open(outputDir+'/categories.txt','w')
for c in self.categories:
out.write(c + '\n')
out.close()
if __name__ == "__main__":
params = cu.loadParams('scoresDirectory proposalsFile outputDir')
cu.mem('Program started')
lap = tic()
builder = DBBuilder(params['scoresDirectory'], params['proposalsFile'])
lap = toc('Proposals loaded', lap)
cu.mem('DB initialized')
builder.parseDir()
lap = toc('Directory parsed', lap)
cu.mem('All files read')
builder.saveDB(params['outputDir'])
lap = toc('Database saved', lap)
cu.mem('Program ends')
return True
##################################
# Organize boxes by source image
#################################
startTime = tic()
images = {}
for s, box in bboxes:
b = map(int, box[1:]) + [s]
try:
images[box[0]].append(b)
except:
images[box[0]] = [b]
lap = toc('Reading boxes file:', startTime)
#################################
# Extract Features
#################################
totalItems = len(bboxes)
del (bboxes)
layers = {'fc6_neuron_cudanet_out': {'dim': 4096, 'idx': 15}}
batch = 200
taskQueue = Queue.Queue()
p = threading.Thread(target=worker, args=(taskQueue, net, layers, batch))
p.daemon = True
p.start()
print 'Extracting features for', totalItems, 'total images'
for name in images.keys():
np.savez_compressed(outf,matrix)
outf.close()
##################################
# Organize boxes by source image
#################################
startTime = tic()
images = {}
for s,box in bboxes:
b = map(int,box[1:]) + [s]
try:
images[ box[0] ].append(b)
except:
images[ box[0] ] = [b]
lap = toc('Reading boxes file:',startTime)
#################################
# Extract Features
#################################
totalItems = len(bboxes)
del(bboxes)
layers = {'fc6_neuron_cudanet_out': {'dim':4096,'idx':15}}
#layers = {'fc6_neuron_cudanet_out': {'dim':4096,'idx':15},'conv3_cudanet_out': {'dim':64896,'idx':9}}
batch = 200
print 'Extracting features for',totalItems,'total images'
for name in images.keys():
# Check if files already exist
processed = 0
dim = 3
M = np.random.randn(2 * dim, 2 * dim)
M += M.T.copy()
for i in np.arange(M.shape[0] / 2):
s = M[2 * i, 2 * i] + M[2 * i + 1, 2 * i + 1]
M[2 * i, 2 * i] -= s / 2
M[2 * i + 1, 2 * i + 1] -= s / 2
tic("sampling")
nsamples = 10 ** 4
burnin = 10 ** 3
lf_steps = 50
step_sz = 0.15
phi, E, diagn = sample_hmc(M, nsamples, burnin, lf_steps, step_sz, diagnostics=True)
toc("sampling")
tic("fiting")
M_hat = fit_model(phi)
Mneg_hat, Mpos_hat = m2kappa(M_hat)
# anti-symmetrize diagonal elements for estimation matrix
for i in np.arange(M_hat.shape[0] / 2):
s = M_hat[2 * i, 2 * i] + M_hat[2 * i + 1, 2 * i + 1]
M_hat[2 * i, 2 * i] -= s / 2
M_hat[2 * i + 1, 2 * i + 1] -= s / 2
toc("fiting")
M_error = M - M_hat
M_max = max(abs(M).max(), abs(M_hat).max())
print "M_error norm = ", (M_error ** 2).sum()
def mainLoop(modelType, modelArgs, positives, trainingList, featuresDir,
featuresExt, modelOut, maxNegOverlap, iter):
pos, posIdx, ari, osi = positives
startTime = cu.tic()
if iter == 0:
## Random Negatives
print ' >>> RANDOM NEGATIVES'
neg, negIdx = learn.getRandomNegs(featuresDir, trainingList,
featuresExt, pos.shape[1],
maxVectorsCache, maxNegativeImages)
detectionsList = [[x[0], '0.0'] + x[1:] + ['1'] for x in negIdx]
hards = {'features': np.zeros((0, neg.shape[1])), 'index': []}
lap = cu.toc(
'Random negatives matrix (' + str(neg.shape[0]) + ' instances)',
startTime)
else:
## Mine hard negatives
print ' >>> MINING HARD NEGATIVES'
model = det.createDetector(modelType, modelArgs)
model.load(modelOut + '.' + str(iter - 1))
detectionsList = detector.detectObjects(
model, trainingList, featuresDir, featuresExt, 1.0, -10.0
) # For RCNN the overlap parameter is 0.3 not 1.0(no suppression)
hards = cu.loadMatrixNoCompression(modelOut + '.hards').item()
lap = cu.toc(
'Hard negatives matrix (' + str(hards['features'].shape[0]) +
' instances)', startTime)
## Rank and clean negative detections
detectionsData = evaluation.loadDetections(detectionsList)
groundTruth = evaluation.loadGroundTruthAnnotations(posIdx)
detectionsLog = evaluation.evaluateDetections(
groundTruth, detectionsData, 0.5,
allowDuplicates=False) #,overlapMeasure=det.overlap
evaluation.computePrecisionRecall(len(posIdx), detectionsLog['tp'],
detectionsLog['fp'], 'tmp.txt')
evaluation.computePrecAt(detectionsLog['tp'],
[20, 50, 100, 200, 300, 400, 500])
logData = learn.parseRankedDetectionsFile(detectionsLog['log'],
maxNegOverlap,
maxNegativeVectors)
print ' >>> LOADING HARD NEGATIVES'
neg, negIdx = learn.loadHardNegativesFromList(featuresDir,
logData['negExamples'],
featuresExt, pos.shape[1],
logData['negTaken'])
del (detectionsList, detectionsData, detectionsLog, logData)
lap = cu.toc(
'Ranked negatives matrix (' + str(neg.shape[0]) + ' instances)', lap)
neg = np.concatenate((neg, hards['features']))
negIdx = negIdx + hards['index']
## Learn Detector
clf = det.createDetector(modelType, modelArgs)
clf.learn(pos, neg, posIdx, negIdx)
clf.save(modelOut + '.' + str(iter))
lap = cu.toc('Classifier learned:', lap)
## Keep hard negatives for next iterations
scores = clf.predict(neg, negIdx)
hardNegsIdx = np.argsort(scores)
hardNeg = np.concatenate(
(hards['features'], neg[hardNegsIdx[-cu.topHards:]]))
negIdx = hards['index'] + [negIdx[j] for j in hardNegsIdx[-cu.topHards:]]
print 'Hard negatives:', hardNeg.shape[0]
hards = {'features': hardNeg, 'index': negIdx}
cu.saveMatrixNoCompression({
'features': hardNeg,
'index': negIdx
}, modelOut + '.hards')
print ' ** Iteration', iter, 'done'
return {
'detector': clf,
'pos': pos,
'posIdx': posIdx,
'neg': neg,
'negIdx': negIdx
}
