def setup_prototypical_network(netName='vgg', layerName='pool4'): ''' Sets up a network in a configuration in which I commonly use it. ''' modelFile, meanFile = get_model_mean_file(netName) defFile = get_layer_def_files(netName, layerName=layerName) meanDat = mpio.read_mean(meanFile) net = MyNet(defFile, modelFile) net.set_preprocess(ipName='data', meanDat=meanDat, imageDims=(256,256,3)) return net
def setup_prototypical_network(netName='vgg', layerName='pool4'):
'''
Sets up a network in a configuration in which I commonly use it.
'''
modelFile, meanFile = get_model_mean_file(netName)
defFile = get_layer_def_files(netName, layerName=layerName)
meanDat = mpio.read_mean(meanFile)
net = MyNet(defFile, modelFile)
net.set_preprocess(ipName='data', meanDat=meanDat, imageDims=(256, 256, 3))
return net
def vis_generic_window_data(protoDef,
numLabels,
layerName='window_data',
phase='TEST',
maxVis=100):
'''
layerName: The name of the generic_window_data layer
numLabels: The number of labels.
'''
#Just write the data part of the file.
if not isinstance(protoDef, mpu.ProtoDef):
protoDef = mpu.ProtoDef(protoDef)
protoDef.del_all_layers_above(layerName)
randInt = np.random.randint(1e+10)
outProto = os.path.join(TMP_DATA_DIR, 'gn_window_%d.prototxt' % randInt)
protoDef.write(outProto)
#Extract the name of the data and the label blobs.
dataName = protoDef.get_layer_property(layerName, 'top', propNum=0)[1:-1]
labelName = protoDef.get_layer_property(layerName, 'top', propNum=1)[1:-1]
crpSize = int(protoDef.get_layer_property(layerName, ['crop_size']))
mnFile = protoDef.get_layer_property(layerName, ['mean_file'])[1:-1]
mnDat = mpio.read_mean(mnFile)
ch, nr, nc = mnDat.shape
xMn = int((nr - crpSize) / 2)
mnDat = mnDat[:, xMn:xMn + crpSize, xMn:xMn + crpSize]
print mnDat.shape
#Create a network
if phase == 'TRAIN':
net = caffe.Net(outProto, caffe.TRAIN)
else:
net = caffe.Net(outProto, caffe.TEST)
lblStr = ''.join('lb-%d: %s, ' % (i, '%.2f') for i in range(numLabels))
figDt = plt.figure()
plt.ion()
for i in range(maxVis):
allDat = net.forward([dataName, labelName])
imData = allDat[dataName] + mnDat
lblDat = allDat[labelName]
batchSz = imData.shape[0]
for b in range(batchSz):
#Plot network data.
im1 = imData[b, 0:3].transpose((1, 2, 0))
im2 = imData[b, 3:6].transpose((1, 2, 0))
im1 = im1[:, :, [2, 1, 0]]
im2 = im2[:, :, [2, 1, 0]]
lb = lblDat[b].squeeze()
lbStr = lblStr % tuple(lb)
plot_pairs(im1, im2, figDt, lbStr)
raw_input()
def set_preprocess(self, ipName='data',chSwap=(2,1,0), meanDat=None,
imageDims=None, isBlobFormat=False, rawScale=None, cropDims=None,
noTransform=False, numCh=3):
'''
isBlobFormat: if the images are already coming in blobFormat or not.
ipName : the blob for which the pre-processing parameters need to be set.
meanDat : the mean which needs to subtracted
imageDims : the size of the images as H * W * K where K is the number of channels
cropDims : the size to which the image needs to be cropped.
if None - then it is automatically determined
this behavior is undesirable for some deploy prototxts
noTransform: if no transform needs to be applied
numCh : number of channels
'''
if chSwap is not None:
assert len(chSwap) == numCh, 'Number of channels mismatch (%d, %d)'\
%(len(chSwap), numCh)
if noTransform:
self.transformer[ipName] = None
return
self.transformer[ipName] = caffe.io.Transformer({ipName: self.net.blobs[ipName].data.shape})
#Note blobFormat will be so used that finally the image will need to be flipped.
self.transformer[ipName].set_transpose(ipName, (2,0,1))
if isBlobFormat:
assert chSwap is None, 'With Blob format chSwap should be none'
if chSwap is not None:
#Required for eg RGB to BGR conversion.
print (ipName, chSwap)
self.transformer[ipName].set_channel_swap(ipName, chSwap)
if rawScale is not None:
self.transformer[ipName].set_raw_scale(ipName, rawScale)
#Crop Dimensions
ipDims = np.array(self.net.blobs[ipName].data.shape)
if cropDims is not None:
assert len(cropDims)==2, 'Length of cropDims needs to be corrected'
self.cropDims = np.array(cropDims)
else:
self.cropDims = ipDims[2:]
self.isBlobFormat = isBlobFormat
if imageDims is None:
imageDims = np.array([ipDims[2], ipDims[3], ipDims[1]])
else:
assert len(imageDims)==3
imageDims = np.array([imageDims[0], imageDims[1], imageDims[2]])
self.imageDims = imageDims
self.get_crop_dims()
#Mean Subtraction
if not meanDat is None:
isTuple = False
if isinstance(meanDat, string_types):
meanDat = mpio.read_mean(meanDat)
elif type(meanDat) == tuple:
meanDat = np.array(meanDat).reshape(numCh,1,1)
meanDat = meanDat * (np.ones((numCh, self.crop[2] - self.crop[0],\
self.crop[3]-self.crop[1])).astype(np.float32))
isTuple = True
_,h,w = meanDat.shape
assert self.imageDims[0]<=h and self.imageDims[1]<=w,\
'imageDims must match mean Image size, (h,w), (imH, imW): (%d, %d), (%d,%d)'\
% (h,w,self.imageDims[0],self.imageDims[1])
if not isTuple:
meanDat = meanDat[:, self.crop[0]:self.crop[2], self.crop[1]:self.crop[3]]
self.transformer[ipName].set_mean(ipName, meanDat)
def set_preprocess(self,
ipName='data',
chSwap=(2, 1, 0),
meanDat=None,
imageDims=None,
isBlobFormat=False,
rawScale=None,
cropDims=None,
noTransform=False,
numCh=3):
'''
isBlobFormat: if the images are already coming in blobFormat or not.
ipName : the blob for which the pre-processing parameters need to be set.
meanDat : the mean which needs to subtracted
imageDims : the size of the images as H * W * K where K is the number of channels
cropDims : the size to which the image needs to be cropped.
if None - then it is automatically determined
this behavior is undesirable for some deploy prototxts
noTransform: if no transform needs to be applied
numCh : number of channels
'''
if chSwap is not None:
assert len(chSwap) == numCh, 'Number of channels mismatch (%d, %d)'\
%(len(chSwap), numCh)
if noTransform:
self.transformer[ipName] = None
return
self.transformer[ipName] = caffe.io.Transformer(
{ipName: self.net.blobs[ipName].data.shape})
#Note blobFormat will be so used that finally the image will need to be flipped.
self.transformer[ipName].set_transpose(ipName, (2, 0, 1))
if isBlobFormat:
assert chSwap is None, 'With Blob format chSwap should be none'
if chSwap is not None:
#Required for eg RGB to BGR conversion.
print(ipName, chSwap)
self.transformer[ipName].set_channel_swap(ipName, chSwap)
if rawScale is not None:
self.transformer[ipName].set_raw_scale(ipName, rawScale)
#Crop Dimensions
ipDims = np.array(self.net.blobs[ipName].data.shape)
if cropDims is not None:
assert len(
cropDims) == 2, 'Length of cropDims needs to be corrected'
self.cropDims = np.array(cropDims)
else:
self.cropDims = ipDims[2:]
self.isBlobFormat = isBlobFormat
if imageDims is None:
imageDims = np.array([ipDims[2], ipDims[3], ipDims[1]])
else:
assert len(imageDims) == 3
imageDims = np.array([imageDims[0], imageDims[1], imageDims[2]])
self.imageDims = imageDims
self.get_crop_dims()
#Mean Subtraction
if not meanDat is None:
isTuple = False
if isinstance(meanDat, string_types):
meanDat = mpio.read_mean(meanDat)
elif type(meanDat) == tuple:
meanDat = np.array(meanDat).reshape(numCh, 1, 1)
meanDat = meanDat * (np.ones((numCh, self.crop[2] - self.crop[0],\
self.crop[3]-self.crop[1])).astype(np.float32))
isTuple = True
_, h, w = meanDat.shape
assert self.imageDims[0]<=h and self.imageDims[1]<=w,\
'imageDims must match mean Image size, (h,w), (imH, imW): (%d, %d), (%d,%d)'\
% (h,w,self.imageDims[0],self.imageDims[1])
if not isTuple:
meanDat = meanDat[:, self.crop[0]:self.crop[2],
self.crop[1]:self.crop[3]]
self.transformer[ipName].set_mean(ipName, meanDat)
