def selectNegativeWindowsFromPositiveImages(groundTruths,featuresDir,featuresExt,maxVectors,overlap,model=False):
gtb = dict()
for x in groundTruths:
im,bx = x[0],map(float,x[1:])
try:
gtb[im].append(bx)
except:
gtb[im] = [bx]
task = NWFPIFilter(gtb,featuresDir,featuresExt,maxVectors/len(gtb.keys()),overlap,model)
result = dp.processData(gtb.keys(),featuresDir,featuresExt,task)
posIdx,posFeat,negIdx,negFeat = [],[],[],[]
for r in result:
posIdx += r[0]
posFeat += r[1]
negIdx += r[2]
negFeat += r[3]
Xp = emptyMatrix( (len(posIdx),posFeat[0].shape[1]) )
Xn = emptyMatrix( (len(negIdx),negFeat[0].shape[1]) )
k = 0
for i in range(len(posFeat)):
Xp[k:k+posFeat[i].shape[0],:] = posFeat[i]
k = k + posFeat[i].shape[0]
k = 0
for i in range(len(negFeat)):
Xn[k:k+negFeat[i].shape[0],:] = negFeat[i]
k + k + negFeat[i].shape[0]
print 'NegFromPos ready:',len(negIdx)
return {'posIdx':posIdx, 'posFeat':Xp, 'negIdx':negIdx, 'negFeat':Xn}
def getHardNegatives(negativesDir,negativesList,featuresExt,numFeatures,maxVectors,currentModel):
maxVectorsPerImage = maxVectors/len(negativesList)
i = 0
task = HardNegativeMining(currentModel,maxVectorsPerImage)
result = dp.processData(negativesList,negativesDir,featuresExt,task)
hardng = emptyMatrix([2*maxVectors,numFeatures])
boxes = []
while len(result) > 0:
data = result.pop(0)
if data[0].shape[0]+i > hardng.shape[0]:
print 'Not enough matrix space'
hardng = np.concatenate( (hardng,emptyMatrix([maxVectors,numFeatures])) )
hardng[i:i+data[0].shape[0],:] = data[0]
boxes += data[2]
i = i + data[0].shape[0]
return hardng[0:i,:],boxes[0:i]
def getActivations(self, imagePath, boxes, state):
n = len(boxes)
activations = cu.emptyMatrix([n, config.geti('outputActions')])
numBatches = (n + config.geti('deployBatchSize') -
1) / config.geti('deployBatchSize')
boxes += [[0, 0, 0, 0]
for x in range(numBatches * config.geti('deployBatchSize') -
n)]
stateFeatures = np.zeros((len(boxes), 20, 1, 1), dtype=np.float32)
stateFeatures[0:n, :, :, :] = state
dims = self.net.caffenet.InitializeImage(imagePath,
config.geti('imageSize'),
self.meanImage,
config.geti('cropSize'))
for k in range(numBatches):
s, f = k * config.geti('deployBatchSize'), (
k + 1) * config.geti('deployBatchSize')
e = config.geti('deployBatchSize') if f <= n else n - s
# Forward this batch
#self.net.caffenet.ForwardRegions(boxes[s:f], config.geti('contextPad'))
self.net.caffenet.ForwardRegionsAndState(
boxes[s:f], config.geti('contextPad'),
[stateFeatures[s:f, :, :, :]])
outputs = self.net.caffenet.blobs
f = n if f > n else f
# Collect outputs
activations[s:f, :] = outputs['prob'].data[0:e, :, :, :].reshape(
[e, config.geti('outputActions')])
# Release image data
self.net.caffenet.ReleaseImageData()
return activations
def loadHardNegativesFromMatrix(featuresDir,imagesIdx,detMatrix,featuresExt,numFeatures,totalNegatives):
i = 0
task = LoadHardNegatives(imagesIdx,detMatrix,numFeatures)
result = dp.processData(imagesIdx.keys(),featuresDir,featuresExt,task)
hardng = cu.emptyMatrix([totalNegatives,numFeatures])
while len(result) > 0:
data = result.pop(0)
hardng[i:i+data.shape[0],:] = data
i = i + data.shape[0]
return hardng[0:i,:]
def loadHardNegativesFromMatrix(featuresDir, imagesIdx, detMatrix, featuresExt,
numFeatures, totalNegatives):
i = 0
task = LoadHardNegatives(imagesIdx, detMatrix, numFeatures)
result = dp.processData(imagesIdx.keys(), featuresDir, featuresExt, task)
hardng = cu.emptyMatrix([totalNegatives, numFeatures])
while len(result) > 0:
data = result.pop(0)
hardng[i:i + data.shape[0], :] = data
i = i + data.shape[0]
return hardng[0:i, :]
def loadHardNegativesFromList(featuresDir,negativesInfo,featuresExt,numFeatures,totalNegatives,idx=False):
i = 0
task = LoadHardNegatives(negativesInfo)
result = dp.processData(negativesInfo.keys(),featuresDir,featuresExt,task)
hardng = emptyMatrix([totalNegatives,numFeatures])
hardNames = []
boxes = []
while len(result) > 0:
data,imgs,box = result.pop(0)
hardng[i:i+data.shape[0],:] = data
hardNames += imgs
boxes += box
i = i + data.shape[0]
return (hardng[0:i,:],boxes)
def getRandomNegs(featuresDir,negativeList,featuresExt,numFeatures,maxVectors,maxNegativeImages):
randomBoxes = maxVectors/maxNegativeImages
cu.rnd.shuffle(negativeList)
task = RandomNegativesFilter(numFeatures,randomBoxes)
negatives = [negativeList.pop(0) for i in range(maxNegativeImages)]
result = dp.processData(negatives,featuresDir,featuresExt,task)
neg = emptyMatrix([maxVectors,numFeatures])
boxes = []
n = 0
while len(result) > 0:
mat,box = result.pop()
neg[n:n+mat.shape[0]] = mat
n = n + mat.shape[0]
boxes += box
return (neg[0:n],boxes[0:n])
def getActivations(self, imagePath, boxes, state):
n = len(boxes)
activations = cu.emptyMatrix( [n, config.geti('outputActions')] )
numBatches = (n + config.geti('deployBatchSize') - 1) / config.geti('deployBatchSize')
boxes += [ [0,0,0,0] for x in range(numBatches * config.geti('deployBatchSize') - n) ]
stateFeatures = np.zeros( (len(boxes), 20, 1, 1), dtype=np.float32)
stateFeatures[0:n,:,:,:] = state
dims = self.net.caffenet.InitializeImage(imagePath, config.geti('imageSize'), self.meanImage, config.geti('cropSize'))
for k in range(numBatches):
s, f = k * config.geti('deployBatchSize'), (k + 1) * config.geti('deployBatchSize')
e = config.geti('deployBatchSize') if f <= n else n - s
# Forward this batch
#self.net.caffenet.ForwardRegions(boxes[s:f], config.geti('contextPad'))
self.net.caffenet.ForwardRegionsAndState(boxes[s:f], config.geti('contextPad'), [stateFeatures[s:f,:,:,:]])
outputs = self.net.caffenet.blobs
f = n if f > n else f
# Collect outputs
activations[s:f,:] = outputs['prob'].data[0:e,:,:,:].reshape([e,config.geti('outputActions')])
# Release image data
self.net.caffenet.ReleaseImageData()
return activations
