def extract(self,X):
assert self._models is not None, "Must be trained before calling extract"
if(len(X) == 2):
X, num_parts = X
else:
X, num_parts, orientation = X
assert X.ndim == 4, "Input X dimension is not correct"
assert num_parts == self._num_lower_parts
X = X.reshape((X.shape[0], X.shape[1], X.shape[2]))
secondLevelCurx = np.zeros((X.shape[0], X.shape[1] + self._lowerLayerShape[0] - self._part_shape[0], X.shape[2] + self._lowerLayerShape[1] - self._part_shape[1],1,1,self._num_lower_parts))
secondLevelCurxCenter = np.zeros((X.shape[0], X.shape[1] + self._lowerLayerShape[0] - self._part_shape[0], X.shape[2] + self._lowerLayerShape[1] - self._part_shape[1]))
extractedFeature = np.empty((X.shape[0] * (X.shape[1] + self._lowerLayerShape[0] - self._part_shape[0]) * (X.shape[2] + self._lowerLayerShape[1] - self._part_shape[1])))
totalRange = X.shape[1]
frame = (self._part_shape[0] - self._lowerLayerShape[0]) / 2
frame = int(frame)
for m in range(totalRange)[frame:totalRange-frame]:
for n in range(totalRange)[frame:totalRange-frame]:
secondLevelCurx[:,m-frame,n-frame] = index_map_pooling(X[:,m-frame:m+frame+1,n-frame:n+frame+1],self._num_lower_parts, (2 * frame + 1, 2 * frame + 1), (2 * frame + 1, 2 * frame + 1))
secondLevelCurxCenter[:,m-frame,n-frame] = X[:,m,n]
flattenSecondLevelCurx = secondLevelCurx.reshape((secondLevelCurx.shape[0] * secondLevelCurx.shape[1] * secondLevelCurx.shape[2], -1))
flattenSecondLevelCurxCenter = secondLevelCurxCenter.flatten()
#for m in range(X[0].shape[0] + self._lowerLayerShape[0]- self._part_shape[0]):
#for n in range(X[0].shape[1] + self._lowerLayerShape[1] - self._part_shape[1]):
#codedIndex = np.where(secondLevelCurxCenter[:,m,n]!=-1)[0]
codedIndex = np.where(flattenSecondLevelCurxCenter!=-1)[0]
#notCodedIndex = np.where(secondLevelCurxCenter[:,m,n] == 0)[0]
notCodedIndex = np.where(flattenSecondLevelCurxCenter == 0)[0]
print(codedIndex.shape)
#firstLevelPartIndex = secondLevelCurxCenter[codedIndex,m,n]
firstLevelPartIndex = flattenSecondLevelCurxCenter[codedIndex]
firstLevelPartIndex = np.array(firstLevelPartIndex,dtype = np.int)
firstLevelPartIndex = (firstLevelPartIndex - firstLevelPartIndex % self._rotation)//self._rotation
blockSize = 10000
for i in range(0, codedIndex.shape[0], blockSize):
blockend = min(codedIndex.shape[0], i + blockSize)
theta = self._models[firstLevelPartIndex[i:blockend]]
XX = flattenSecondLevelCurx[codedIndex[i:blockend]][:,np.newaxis]
print(XX.shape, theta.shape)
llh = XX * np.log(theta) + (1 - XX) * np.log(1 - theta)
bb = np.apply_over_axes
bb = np.apply_over_axes(np.sum, llh, [-3, -2, -1])[...,0,0,0]
maxIndex = np.argmax(bb,axis = -1)
print(maxIndex.shape)
extensionPart = maxIndex // self._rotation
rotatedAngle = maxIndex % self._rotation
extractedFeature[codedIndex[i:blockend]] = np.array(firstLevelPartIndex[i:blockend] * self._num_components * self._rotation + extensionPart * self._rotation + rotatedAngle, dtype = np.int)
extractedFeature[notCodedIndex] = -1
extractedFeature = extractedFeature.reshape((X.shape[0], X.shape[1] + self._lowerLayerShape[0] - self._part_shape[0], X.shape[2] + self._lowerLayerShape[1] - self._part_shape[1]))
extractedFeature = np.array(extractedFeature[:,:,:,np.newaxis],dtype = np.int64)
return (extractedFeature,self._num_parts, self._rotation)
def extract(self,X):
assert self._models is not None, "Must be trained before calling extract"
if(len(X) == 2):
X, num_parts = X
else:
X, num_parts, orientation = X
assert X.ndim == 4, "Input X dimension is not correct"
assert num_parts == self._num_lower_parts
X = X.reshape((X.shape[0], X.shape[1], X.shape[2]))
secondLevelCurx = np.zeros((X.shape[0], X.shape[1] + self._lowerLayerShape[0] - self._part_shape[0], X.shape[2] + self._lowerLayerShape[1] - self._part_shape[1],1,1,self._num_lower_parts))
secondLevelCurxCenter = np.zeros((X.shape[0], X.shape[1] + self._lowerLayerShape[0] - self._part_shape[0], X.shape[2] + self._lowerLayerShape[1] - self._part_shape[1]))
extractedFeature = np.empty((X.shape[0],X.shape[1] + self._lowerLayerShape[0] - self._part_shape[0],X.shape[2] + self._lowerLayerShape[1] - self._part_shape[1]))
totalRange = X.shape[1]
frame = (self._part_shape[0] - self._lowerLayerShape[0]) / 2
frame = int(frame)
for m in range(totalRange)[frame:totalRange-frame]:
for n in range(totalRange)[frame:totalRange-frame]:
secondLevelCurx[:,m-frame,n-frame] = index_map_pooling(X[:,m-frame:m+frame+1,n-frame:n+frame+1],self._num_lower_parts, (2 * frame + 1, 2 * frame + 1), (2 * frame + 1, 2 * frame + 1))
secondLevelCurxCenter[:,m-frame,n-frame] = X[:,m,n]
for i in range(X.shape[0]):
patch = X[i]
for m in range(patch.shape[0] + self._lowerLayerShape[0]- self._part_shape[0]):
for n in range(patch.shape[1] + self._lowerLayerShape[1] - self._part_shape[1]):
if(secondLevelCurxCenter[i,m,n]!=-1):
firstLevelPartIndex = secondLevelCurxCenter[i,m,n]
firstLevelPartIndex = int(firstLevelPartIndex)
firstLevelPartIndex = (firstLevelPartIndex - firstLevelPartIndex % self._rotation)//self._rotation
theta = self._models[firstLevelPartIndex]
XX = secondLevelCurx[i, m, n][np.newaxis]
llh = XX * np.log(theta) + (1 - XX) * np.log(1 - theta)
bb = np.apply_over_axes(np.sum, llh, [-3, -2, -1])[...,0,0,0]
maxIndex = np.argmax(bb,axis = 0)
extensionPart = maxIndex // self._rotation
rotatedAngle = maxIndex % self._rotation
extractedFeature[i,m,n] = int( firstLevelPartIndex * self._num_components * self._rotation + extensionPart * self._rotation + rotatedAngle)
else:
extractedFeature[i,m,n] = -1
extractedFeature = np.array(extractedFeature[:,:,:,np.newaxis],dtype = np.int64)
return (extractedFeature,self._num_parts, self._rotation)
for a in range(2):
for b in range(2):
for m in range(totalRange)[a * (totalRange - subRegionSize) + frame: a * (totalRange - subRegionSize)+ subRegionSize - frame]:
for n in range(totalRange)[a * (totalRange - subRegionSize) + frame: a * (totalRange - subRegionSize) + subRegionSize - frame]:
if(X[i,m,n]!=-1):
partsGrid = self.partsPool(X[i,m-frame:m+frame+1, n-frame:n+frame+1], self._num_lower_parts)
partsRegion[a * 2 + b, X[i,m,n]].append(partsGrid)
return partsRegion
for m in range(totalRange)[frame:totalRange-frame]:
for n in range(totalRange)[frame:totalRange-frame]:
secondLevelCurx[:,m-frame,n-frame] = index_map_pooling(X[:,m-frame:m+frame+1,n-frame:n+frame+1],self._num_lower_parts, (2 * frame + 1, 2 * frame + 1), (2 * frame + 1, 2 * frame + 1))
secondLevelCurxCenter[:,m-frame,n-frame] = X[:,m,n]
for a in range(2):
for b in range(2):
for m in range(totalRange)[a * (totalRange - subRegionSize) + frame: a * (totalRange - subRegionSize) + subRegionSize - frame]:
for n in range(totalRange)[b * (totalRange - subRegionSize) + frame: b * (totalRange - subRegionSize) + subRegionSize - frame]:
secondLevelindex_map_pooling(X[:,m-frame:m+frame+1,n-frame:n+frame+1],self._num_lower_parts, (2 * frame + 1, 2 * frame + 1), (2 * frame + 1, 2 * frame + 1))
sup_ims.append(ims0)
sup_labels.append(d * np.ones(len(ims0),dtype = np.int64))
sup_ims = np.concatenate(sup_ims, axis = 0)
sup_labels = np.concatenate(sup_labels,axis = 0)
curX = extract(sup_ims,allLayer[0:2])[0]
#print(curX.shape)
curX = curX.reshape(curX.shape[0:3])
secondLevelCurx = np.zeros((10 * classificationTrainingNum,29 - secondLayerShape,29 - secondLayerShape,1,1,numParts))
secondLevelCurxCenter = np.zeros((10 * classificationTrainingNum,29- secondLayerShape,29 - secondLayerShape))
#for i in range(10 * classificationTrainingNum):
# codeParts = curX[i]
for m in range(totalRange)[frame:totalRange-frame]:
for n in range(totalRange)[frame:totalRange-frame]:
secondLevelCurx[:,m-frame,n-frame] = index_map_pooling(curX[:,m-frame:m+frame+1,n-frame:n+frame+1],numParts,(2 * frame + 1,2 * frame + 1),(2 * frame + 1,2 * frame + 1))
secondLevelCurxCenter[:,m-frame,n-frame] = curX[:,m,n]
thirdLevelCurx = np.zeros((10 * classificationTrainingNum, 29 - secondLayerShape,29 - secondLayerShape))
for i in range(int(10 * classificationTrainingNum)):
for m in range(29 - secondLayerShape):
for n in range(29 - secondLayerShape):
if(secondLevelCurxCenter[i,m,n]!=-1):
firstLevelPartIndex = secondLevelCurxCenter[i,m,n]
#print(firstLevelPartIndex)
firstLevelPartIndex = int(firstLevelPartIndex)
extractedFeaturePart = extract(np.array(secondLevelCurx[i,m,n][np.newaxis,:],dtype = np.uint8),allPartsLayer[firstLevelPartIndex])[0]
#print("secondLayerExtraction")
#print(extractedFeaturePart.shape)
thirdLevelCurx[i,m,n] = int(numSecondLayerParts * firstLevelPartIndex + extractedFeaturePart)
#print(numSecondLayerParts,firstLevelPartIndex,extractedFeaturePart,thirdLevelCurx[i,m,n])
def extract(self, X):
assert self._classificationLayers is not None, "Must be trained before calling extract"
if len(X) == 2:
X, num_parts = X
else:
X, num_parts, orientation = X
assert X.ndim == 4, "Input X dimension is not correct"
assert num_parts == self._num_lower_parts
X = X.reshape((X.shape[0], X.shape[1], X.shape[2]))
secondLevelCurx = np.zeros(
(
X.shape[0],
X.shape[1] + self._lowerLayerShape[0] - self._part_shape[0],
X.shape[2] + self._lowerLayerShape[1] - self._part_shape[1],
1,
1,
self._num_lower_parts,
)
)
secondLevelCurxCenter = np.zeros(
(
X.shape[0],
X.shape[1] + self._lowerLayerShape[0] - self._part_shape[0],
X.shape[2] + self._lowerLayerShape[1] - self._part_shape[1],
)
)
extractedFeature = np.empty(
(
X.shape[0],
X.shape[1] + self._lowerLayerShape[0] - self._part_shape[0],
X.shape[2] + self._lowerLayerShape[1] - self._part_shape[1],
)
)
totalRange = X.shape[1]
frame = (self._part_shape[0] - self._lowerLayerShape[0]) / 2
frame = int(frame)
for m in range(totalRange)[frame : totalRange - frame]:
for n in range(totalRange)[frame : totalRange - frame]:
secondLevelCurx[:, m - frame, n - frame] = index_map_pooling(
X[:, m - frame : m + frame + 1, n - frame : n + frame + 1],
self._num_lower_parts,
(2 * frame + 1, 2 * frame + 1),
(2 * frame + 1, 2 * frame + 1),
)
secondLevelCurxCenter[:, m - frame, n - frame] = X[:, m, n]
for i in range(X.shape[0]):
patch = X[i]
for m in range(patch.shape[0] + self._lowerLayerShape[0] - self._part_shape[0]):
for n in range(patch.shape[1] + self._lowerLayerShape[1] - self._part_shape[1]):
if secondLevelCurxCenter[i, m, n] != -1:
firstLevelPartIndex = secondLevelCurxCenter[i, m, n]
firstLevelPartIndex = int(firstLevelPartIndex)
extractedFeaturePart = self.codeParts(
np.array(secondLevelCurx[i, m, n][np.newaxis, :], dtype=np.uint8),
self._classificationLayers[firstLevelPartIndex],
)[0]
extractedFeature[i, m, n] = int(
self._num_components * firstLevelPartIndex + extractedFeaturePart
)
else:
extractedFeature[i, m, n] = -1
extractedFeature = np.array(extractedFeature[:, :, :, np.newaxis], dtype=np.int64)
return (extractedFeature, self._num_parts)
sup_ims.append(ims0)
sup_labels.append(d * np.ones(len(ims0),dtype = np.int64))
sup_ims = np.concatenate(sup_ims, axis = 0)
sup_labels = np.concatenate(sup_labels,axis = 0)
curX = extract(sup_ims,allLayer[0:2])[0]
#print(curX.shape)
curX = curX.reshape(curX.shape[0:3])
secondLevelCurx = np.zeros((10 * classificationTrainingNum,17,17,1,1,numParts))
secondLevelCurxCenter = np.zeros((10 * classificationTrainingNum,17,17))
#for i in range(10 * classificationTrainingNum):
# codeParts = curX[i]
for m in range(23)[3:20]:
for n in range(23)[3:20]:
secondLevelCurx[:,m-3,n-3] = index_map_pooling(curX[:,m-3:m+4,n-3:n+4],numParts,(7,7),(7,7))
secondLevelCurxCenter[:,m-3,n-3] = curX[:,m,n]
thirdLevelCurx = np.zeros((10 * classificationTrainingNum, 17,17))
for i in range(int(10 * classificationTrainingNum)):
for m in range(17):
for n in range(17):
if(secondLevelCurxCenter[i,m,n]!=-1):
firstLevelPartIndex = secondLevelCurxCenter[i,m,n]
#print(firstLevelPartIndex)
firstLevelPartIndex = int(firstLevelPartIndex)
extractedFeaturePart = extract(np.array(secondLevelCurx[i,m,n][np.newaxis,:],dtype = np.uint8),allPartsLayer[firstLevelPartIndex])[0]
#print("secondLayerExtraction")
#print(extractedFeaturePart.shape)
thirdLevelCurx[i,m,n] = int(numSecondLayerParts * firstLevelPartIndex + extractedFeaturePart)
#print(numSecondLayerParts,firstLevelPartIndex,extractedFeaturePart,thirdLevelCurx[i,m,n])
