def get_phocnet(self, word_image_lmdb_path, phoc_lmdb_path,
phoc_size=604, generate_deploy=False):
'''
Returns a NetSpec definition of the PHOCNet. The definition can then be transformed
into a protobuffer message by casting it into a str.
'''
n = NetSpec()
# Data
self.set_phocnet_data(n=n, generate_deploy=generate_deploy,
word_image_lmdb_path=word_image_lmdb_path,
phoc_lmdb_path=phoc_lmdb_path)
# Conv Part
self.set_phocnet_conv_body(n=n, relu_in_place=True)
# FC Part
n.spp5 = L.SPP(n.relu4_3, spp_param=dict(pool=P.SPP.MAX, pyramid_height=3, engine=self.spp_engine))
n.fc6, n.relu6, n.drop6 = self.fc_relu(bottom=n.spp5, layer_size=4096,
dropout_ratio=0.5, relu_in_place=True)
n.fc7, n.relu7, n.drop7 = self.fc_relu(bottom=n.drop6, layer_size=4096,
dropout_ratio=0.5, relu_in_place=True)
n.fc8 = L.InnerProduct(n.drop7, num_output=phoc_size,
weight_filler=dict(type=self.initialization),
bias_filler=dict(type='constant'))
n.sigmoid = L.Sigmoid(n.fc8, include=dict(phase=self.phase_test))
# output part
if not generate_deploy:
n.silence = L.Silence(n.sigmoid, ntop=0, include=dict(phase=self.phase_test))
n.loss = L.SigmoidCrossEntropyLoss(n.fc8, n.phocs)
return n.to_proto()
def train_head(self, subset):
n = NetSpec()
# train
image_data_param = dict(source=subset.get_list_absolute_path(),
batch_size=self.batch_sizes[0],
new_width=self.infmt.new_width,
new_height=self.infmt.new_height,
rand_skip=self.batch_size,
shuffle=True)
if subset.root_folder is not None:
image_data_param['root_folder'] = subset.root_folder
transform_param = dict(
mirror=self.infmt.mirror,
crop_size=self.infmt.crop_size,
# mean_value = self.infmt.mean_pixel,
)
if self.infmt.scale is not None:
transform_param['scale'] = self.infmt.scale
if self.infmt.mean_file is not None:
transform_param['mean_file'] = self.infmt.mean_file
elif self.infmt.mean_pixel is not None:
transform_param['mean_value'] = self.infmt.mean_pixel
n.data, n.label = L.ImageData(ntop=2,
image_data_param=image_data_param,
transform_param=transform_param,
include=dict(phase=caffe.TRAIN))
net = n.to_proto()
net.name = self.name
return net
def __init__(self, number_of_neighbors=6, inner_product_output=100, weight_lr_mult=1, weight_decay_mult=1, b_lr_mult=2, b_decay_mult=0): self.number_of_neighbors = number_of_neighbors #self.netP = self.net = NetSpec() self.shared_weight_counter = 0 self.num_output = inner_product_output self.weight_lr_mult = weight_lr_mult self.weight_decay_mult = weight_decay_mult self.b_lr_mult = b_lr_mult self.b_decay_mult = b_decay_mult
def val_tail(self, last_top, stage=None):
n = NetSpec()
include_param = dict(phase=caffe.TEST)
if stage is not None:
include_param['stage'] = stage
if stage is None:
n.loss = L.SoftmaxWithLoss(bottom=[last_top, "label"])
n.accuracy = L.Accuracy(bottom=[last_top, "label"],
include=include_param)
return n.to_proto()
def val_head(self, subset, stage=None):
image_data_param = dict(
source=subset.get_list_absolute_path(),
batch_size=self.batch_sizes[1],
# root_folder=subset.root_folder,
rand_skip=self.batch_sizes[1],
shuffle=True,
# new_width,
# new_height
)
transform_param = dict(
mirror=False,
# crop_size = self.infmt.crop_size,
# mean_value = self.infmt.mean_pixel,
# mean_file,
# scale,
)
if subset.root_folder is not None:
image_data_param['root_folder'] = subset.root_folder
if self.crop_on_test:
image_data_param['new_width'] = self.infmt.new_width
image_data_param['new_height'] = self.infmt.new_height
transform_param['crop_size'] = self.infmt.crop_size
else:
image_data_param['new_width'] = self.infmt.crop_size
image_data_param['new_height'] = self.infmt.crop_size
if self.infmt.scale is not None:
transform_param['scale'] = self.infmt.scale
if self.infmt.mean_file is not None:
transform_param['mean_file'] = self.infmt.mean_file
elif self.infmt.mean_pixel is not None:
transform_param['mean_value'] = self.infmt.mean_pixel
include_param = dict(phase=caffe.TEST)
if stage is not None:
include_param['stage'] = stage
n = NetSpec()
n.data, n.label = L.ImageData(ntop=2,
image_data_param=image_data_param,
transform_param=transform_param,
include=include_param)
net = n.to_proto()
net.name = self.name
return net
def extract_head(self, subset):
image_data_param = dict(
source=subset.get_list_absolute_path(),
batch_size=self.batch_size,
root_folder=subset.root_folder,
# new_width,
# new_height
)
transform_param = dict(
mirror=False,
# crop_size = self.infmt.crop_size,
# mean_value = self.infmt.mean_pixel,
# mean_file,
# scale,
)
if self.crop_on_test:
image_data_param['new_width'] = self.infmt.new_width
image_data_param['new_height'] = self.infmt.new_height
transform_param['crop_size'] = self.infmt.crop_size
else:
image_data_param['new_width'] = self.infmt.crop_size
image_data_param['new_height'] = self.infmt.crop_size
if self.infmt.scale is not None:
transform_param['scale'] = self.infmt.scale
if self.infmt.mean_file is not None:
transform_param['mean_file'] = self.infmt.mean_file
elif self.infmt.mean_pixel is not None:
transform_param['mean_value'] = self.infmt.mean_pixel
n = NetSpec()
n.data, n.label = L.ImageData(ntop=2,
image_data_param=image_data_param,
transform_param=transform_param
) # , include=dict(phase=caffe.TEST))
net = n.to_proto()
net.name = self.name
return net
def main():
from argparse import ArgumentParser
from os import path
parser = ArgumentParser()
parser.add_argument('prototxt')
parser.add_argument('-l', '--load', help='Load a caffemodel')
parser.add_argument('-d',
'--data',
default=None,
help='Image list to use [default prototxt data]')
#parser.add_argument('-q', action='store_true', help='Quiet execution')
parser.add_argument('-sm', action='store_true', help='Summary only')
parser.add_argument('-q', action='store_true', help='Quiet execution')
parser.add_argument('-a',
'--all',
action='store_true',
help='Show the statistic for all layers')
parser.add_argument('-nc', action='store_true', help='Do not use color')
parser.add_argument('-s',
type=float,
default=1.0,
help='Scale the input [only custom data "-d"]')
parser.add_argument('-bs',
type=int,
default=16,
help='Batch size [only custom data "-d"]')
parser.add_argument('-nit',
type=int,
default=10,
help='Number of iterations')
parser.add_argument('--gpu',
type=int,
default=0,
help='What gpu to run it on?')
args = parser.parse_args()
if args.q:
from os import environ
environ['GLOG_minloglevel'] = '2'
import caffe, load
from caffe import NetSpec, layers as L
caffe.set_mode_gpu()
if args.gpu is not None:
caffe.set_device(args.gpu)
if args.data is not None:
model = load.ProtoDesc(args.prototxt)
net = NetSpec()
fl = getFileList(args.data)
if len(fl) == 0:
print("Unknown data type for '%s'" % args.data)
exit(1)
from tempfile import NamedTemporaryFile
f = NamedTemporaryFile('w')
f.write('\n'.join([path.abspath(i) + ' 0' for i in fl]))
f.flush()
net.data, net.label = L.ImageData(source=f.name,
batch_size=args.bs,
new_width=model.input_dim[-1],
new_height=model.input_dim[-1],
transform_param=dict(
mean_value=[104, 117, 123],
scale=args.s),
ntop=2)
net.out = model(data=net.data, label=net.label)
n = netFromString('force_backward:true\n' + str(net.to_proto()),
caffe.TRAIN)
else:
n = caffe.Net(args.prototxt, caffe.TRAIN)
if args.load is not None:
n.copy_from(args.load)
cvar = printMeanStddev(n,
NIT=args.nit,
show_all=args.all,
show_color=not args.nc,
quiet=args.sm)
cv, gr = computeGradientRatio(n, NIT=args.nit)
print()
print(' Summary ')
print('-----------')
print()
print(
'layer name out cvar rate cvar rate mean'
)
for l in n._layer_names:
if l in cvar and l in cv and l in gr:
print('%-30s %10.2f %10.2f %10.2f' %
(l, cvar[l], cv[l], gr[l]))
def test_tail(self, last_top):
n = NetSpec()
n.accuracy = L.Accuracy(bottom=[last_top, "label"],
include=dict(phase=caffe.TEST))
return n.to_proto()
def train_tail(self, last_top):
n = NetSpec()
n.loss = L.SoftmaxWithLoss(bottom=[last_top, "label"])
return n.to_proto()
def deploy_tail(self, last_top):
n = NetSpec()
n.score = L.Softmax(bottom=last_top)
return n.to_proto()
def main():
from argparse import ArgumentParser
from os import path
import numpy as np
parser = ArgumentParser()
parser.add_argument('prototxt')
parser.add_argument('output_caffemodel')
parser.add_argument(
'-l',
'--load',
help=
'Load a pretrained model and rescale it [bias and type are not supported]'
)
parser.add_argument('-d',
'--data',
default=None,
help='Image list to use [default prototxt data]')
parser.add_argument('-b', '--bias', type=float, default=0.1, help='Bias')
parser.add_argument(
'-t',
'--type',
default='elwise',
help=
'Type: elwise, pca, zca, kmeans, rand (random input patches). Add fast_ to speed up the initialization, but you might lose in precision.'
)
parser.add_argument(
'--zero_from',
default=None,
help='Zero weights starting from this layer and reinitialize')
parser.add_argument('-z',
action='store_true',
help='Zero all weights and reinitialize')
parser.add_argument(
'--post_zero_from',
default=None,
help=
'AFTER everything else, zero weights starting from this layer (they will NOT be reinitialized)'
)
parser.add_argument('-cs', action='store_true', help='Correct for scaling')
parser.add_argument('-q', action='store_true', help='Quiet execution')
parser.add_argument('-s',
type=float,
default=1.0,
help='Scale the input [only custom data "-d"]')
parser.add_argument('-bs',
type=int,
default=16,
help='Batch size [only custom data "-d"]')
parser.add_argument('-nit',
type=int,
default=10,
help='Number of iterations')
parser.add_argument(
'--mem-limit',
type=int,
default=500,
help='How much memory should we use for the data buffer (MB)?')
parser.add_argument('--gpu',
type=int,
default=0,
help='What gpu to run it on?')
args = parser.parse_args()
if args.q:
from os import environ
environ['GLOG_minloglevel'] = '2'
import caffe, load
from caffe import NetSpec, layers as L
caffe.set_mode_gpu()
if args.gpu is not None:
caffe.set_device(args.gpu)
if args.data is not None:
model = load.ProtoDesc(args.prototxt)
net = NetSpec()
fl = getFileList(args.data)
if len(fl) == 0:
print("Unknown data type for '%s'" % args.data)
exit(1)
from tempfile import NamedTemporaryFile
f = NamedTemporaryFile('w')
f.write('\n'.join([path.abspath(i) + ' 0' for i in fl]))
f.flush()
net.data, net.label = L.ImageData(source=f.name,
batch_size=args.bs,
new_width=model.input_dim[-1],
new_height=model.input_dim[-1],
transform_param=dict(
mean_value=[104, 117, 123],
scale=args.s),
ntop=2)
net.out = model(data=net.data, label=net.label)
n = netFromString('force_backward:true\n' + str(net.to_proto()),
caffe.TRAIN)
else:
n = caffe.Net(args.prototxt, caffe.TRAIN)
if args.load is not None:
n.copy_from(args.load)
# Rescale existing layers?
#if args.fix:
#magicFix(n, args.nit)
if args.z or args.zero_from:
zeroLayers(n, start=args.zero_from)
if any([
np.abs(l.blobs[0].data).sum() < 1e-10 for l in n.layers
if len(l.blobs) > 0
]):
print([
m for l, m in zip(n.layers, n._layer_names)
if len(l.blobs) > 0 and np.abs(l.blobs[0].data).sum() < 1e-10
])
magicInitialize(n,
args.bias,
NIT=args.nit,
type=args.type,
max_data=args.mem_limit * 1024 * 1024 / 4)
else:
print("Network already initialized, skipping magic init")
if args.cs:
# A simply helper function that lets you figure out which layers are not
# homogeneous
#print( estimateHomogenety(n) )
print('Calibrating gradient ratio')
calibrateGradientRatio(n)
if args.post_zero_from:
zeroLayers(n, start=args.post_zero_from)
n.save(args.output_caffemodel)
def get_phocnet(self,
word_image_lmdb_path,
phoc_lmdb_path,
phoc_size=604,
generate_deploy=False):
'''
Returns a NetSpec definition of the PHOCNet. The definition can then be transformed
into a protobuffer message by casting it into a str.
'''
n = NetSpec()
relu_in_place = True
# Data
if generate_deploy:
n.word_images = L.Input(shape=dict(dim=[1, 1, 100, 250]))
relu_in_place = False
else:
n.word_images, n.label = L.Data(batch_size=1,
backend=P.Data.LMDB,
source=word_image_lmdb_path,
prefetch=20,
transform_param=dict(
mean_value=255,
scale=-1. / 255,
),
ntop=2)
n.phocs, n.label_phocs = L.Data(batch_size=1,
backend=P.Data.LMDB,
source=phoc_lmdb_path,
prefetch=20,
ntop=2)
# Conv Part
n.conv1_1, n.relu1_1 = self.conv_relu(n.word_images,
nout=64,
relu_in_place=relu_in_place)
n.conv1_2, n.relu1_2 = self.conv_relu(n.relu1_1,
nout=64,
relu_in_place=relu_in_place)
n.pool1 = L.Pooling(n.relu1_2,
pooling_param=dict(pool=P.Pooling.MAX,
kernel_size=2,
stride=2))
n.conv2_1, n.relu2_1 = self.conv_relu(n.pool1,
nout=128,
relu_in_place=relu_in_place)
n.conv2_2, n.relu2_2 = self.conv_relu(n.relu2_1,
nout=128,
relu_in_place=relu_in_place)
n.pool2 = L.Pooling(n.relu2_2,
pooling_param=dict(pool=P.Pooling.MAX,
kernel_size=2,
stride=2))
n.conv3_1, n.relu3_1 = self.conv_relu(n.pool2,
nout=256,
relu_in_place=relu_in_place)
n.conv3_2, n.relu3_2 = self.conv_relu(n.relu3_1,
nout=256,
relu_in_place=relu_in_place)
n.conv3_3, n.relu3_3 = self.conv_relu(n.relu3_2,
nout=256,
relu_in_place=relu_in_place)
n.conv3_4, n.relu3_4 = self.conv_relu(n.relu3_3,
nout=256,
relu_in_place=relu_in_place)
n.conv3_5, n.relu3_5 = self.conv_relu(n.relu3_4,
nout=256,
relu_in_place=relu_in_place)
n.conv3_6, n.relu3_6 = self.conv_relu(n.relu3_5,
nout=256,
relu_in_place=relu_in_place)
n.conv4_1, n.relu4_1 = self.conv_relu(n.relu3_6,
nout=512,
relu_in_place=relu_in_place)
n.conv4_2, n.relu4_2 = self.conv_relu(n.relu4_1,
nout=512,
relu_in_place=relu_in_place)
n.conv4_3, n.relu4_3 = self.conv_relu(n.relu4_2,
nout=512,
relu_in_place=relu_in_place)
# FC Part
n.spp5 = L.SPP(n.relu4_3,
spp_param=dict(pool=P.SPP.MAX,
pyramid_height=3,
engine=self.spp_engine))
n.fc6, n.relu6, n.drop6 = self.fc_relu(bottom=n.spp5,
layer_size=4096,
dropout_ratio=0.5,
relu_in_place=relu_in_place)
n.fc7, n.relu7, n.drop7 = self.fc_relu(bottom=n.drop6,
layer_size=4096,
dropout_ratio=0.5,
relu_in_place=relu_in_place)
n.fc8 = L.InnerProduct(n.drop7,
num_output=phoc_size,
weight_filler=dict(type=self.initialization),
bias_filler=dict(type='constant'))
n.sigmoid = L.Sigmoid(n.fc8, include=dict(phase=self.phase_test))
# output part
if not generate_deploy:
n.silence = L.Silence(n.sigmoid,
ntop=0,
include=dict(phase=self.phase_test))
n.loss = L.SigmoidCrossEntropyLoss(n.fc8, n.phocs)
return n.to_proto()
def __init__(self, number_of_neighbors, num_output=100, lr_mult):
self.number_of_neighbors = number_of_neighbors
#self.netP =
self.net = NetSpec()
self.shared_weight_counter = 0
self.num_output = num_output
