def FireNet_generic(FireNet_module_func, choose_num_output_func, batch_size, pool_after, s):
print s
n = NetSpec()
FireNet_data_layer(n, batch_size) #add data layer to the net
layer_idx=1 #e.g. conv1, fire2, etc.
n.conv1 = L.Convolution(n.data, kernel_size=7, num_output=96, stride=2, weight_filler=dict(type='xavier'))
curr_bottom = 'conv1'
n.tops['relu_conv1'] = L.ReLU(n.tops[curr_bottom], in_place=True)
if curr_bottom in pool_after.keys():
curr_bottom = FireNet_pooling_layer(n, curr_bottom, pool_after[curr_bottom], layer_idx)
for layer_idx in xrange(2,10):
firenet_dict = choose_num_output_func(layer_idx-2, s)
print firenet_dict
curr_bottom = FireNet_module_func(n, curr_bottom, firenet_dict, layer_idx)
if curr_bottom in pool_after.keys():
curr_bottom = FireNet_pooling_layer(n, curr_bottom, pool_after[curr_bottom], layer_idx)
n.tops['drop'+str(layer_idx)] = L.Dropout(n.tops[curr_bottom], dropout_ratio=0.5, in_place=True)
n.tops['conv_final'] = L.Convolution(n.tops[curr_bottom], kernel_size=1, num_output=1000, weight_filler=dict(type='gaussian', std=0.01, mean=0.0))
n.tops['relu_conv_final'] = L.ReLU(n.tops['conv_final'], in_place=True)
n.tops['pool_final'] = L.Pooling(n.tops['conv_final'], global_pooling=1, pool=P.Pooling.AVE)
if phase == 'trainval':
n.loss = L.SoftmaxWithLoss(n.tops['pool_final'], n.label, include=dict(phase=caffe_pb2.TRAIN))
n.accuracy = L.Accuracy(n.tops['pool_final'], n.label, include=dict(phase=caffe_pb2.TEST))
n.accuracy_top5 = L.Accuracy(n.tops['pool_final'], n.label, include=dict(phase=caffe_pb2.TEST), top_k=5)
return n.to_proto()
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 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 NiN(opts):
n = NetSpec()
FireNet_data_layer(n, batch_size) #add data layer to the net
curr_bottom = 'data'
#TODO: possibly rename layers to conv1.1, 1.2, 1.3; 2.1, 2.2, etc.
curr_bottom = conv_relu_xavier(n, 11, 96, str(1), 4, 0, curr_bottom) #_, ksize, nfilt, layerIdx, stride, pad, _
if 'pool1' in opts:
curr_bottom = NiN_pool(n, str(3), curr_bottom)
curr_bottom = conv_relu_xavier(n, 1, 96, str(2), 1, 0, curr_bottom)
curr_bottom = conv_relu_xavier(n, 1, 96, str(3), 1, 0, curr_bottom)
curr_bottom = NiN_pool(n, str(3), curr_bottom)
curr_bottom = conv_relu_xavier(n, 5, 256, str(4), 1, 2, curr_bottom)
curr_bottom = conv_relu_xavier(n, 1, 256, str(5), 1, 0, curr_bottom)
curr_bottom = conv_relu_xavier(n, 1, 256, str(6), 1, 0, curr_bottom)
curr_bottom = NiN_pool(n, str(6), curr_bottom)
#conv8 and conv9 are the least computationally intensive layers
curr_bottom = conv_relu_xavier(n, 3, 384, str(7), 1, 1, curr_bottom)
conv8_nfilt = get_conv8_nfilt(opts)
curr_bottom = conv_relu_xavier(n, 1, conv8_nfilt, str(8), 1, 0, curr_bottom)
curr_bottom = conv_relu_xavier(n, 1, 384, str(9), 1, 0, curr_bottom)
curr_bottom = NiN_pool(n, str(9), curr_bottom)
n.tops['drop9'] = L.Dropout(n.tops[curr_bottom], dropout_ratio=0.5, in_place=True)
curr_bottom = conv_relu_xavier(n, 3, 1024, str(10), 1, 1, curr_bottom)
curr_bottom = conv_relu_xavier(n, 1, 1024, str(11), 1, 0, curr_bottom)
num_output=1000
if 'out10k' in opts:
num_output=10000
n.tops['conv_12'] = L.Convolution(n.tops[curr_bottom], kernel_size=1, num_output=num_output, weight_filler=dict(type='gaussian', std=0.01, mean=0.0))
n.tops['relu_conv_12'] = L.ReLU(n.tops['conv_12'], in_place=True)
n.tops['pool_12'] = L.Pooling(n.tops['conv_12'], global_pooling=1, pool=P.Pooling.AVE)
if phase == 'trainval':
n.loss = L.SoftmaxWithLoss(n.tops['pool_12'], n.label, include=dict(phase=caffe_pb2.TRAIN))
n.accuracy = L.Accuracy(n.tops['pool_12'], n.label, include=dict(phase=caffe_pb2.TEST))
n.accuracy_top5 = L.Accuracy(n.tops['pool_12'], n.label, include=dict(phase=caffe_pb2.TEST), top_k=5)
out_dir = 'nets/NiN_' + '_'.join(opts)
return [n.to_proto(), out_dir]
def FireNet(batch_size, pool_after, s, c1):
print s
n = NetSpec()
FireNet_data_layer(n, batch_size) #add data layer to the net
layer_idx=1 #e.g. conv1, fire2, etc.
n.conv1 = L.Convolution(n.data, kernel_size=c1['dim'], num_output=c1['nfilt'], stride=2, weight_filler=dict(type='xavier'))
curr_bottom = 'conv1'
n.tops['relu_conv1'] = L.ReLU(n.tops[curr_bottom], in_place=True)
#if curr_bottom in pool_after.keys():
# curr_bottom = FireNet_pooling_layer(n, curr_bottom, pool_after[curr_bottom], layer_idx)
if layer_idx in pool_after:
n.tops['pool1'] = L.Pooling(n.tops[curr_bottom], kernel_size=3, stride=2, pool=P.Pooling.MAX)
curr_bottom = 'pool1'
for layer_idx in xrange(2, s['n_layers']+2):
firenet_dict = choose_num_output(layer_idx-2, s)
print firenet_dict
curr_bottom = FireNet_module(n, curr_bottom, firenet_dict, layer_idx)
if layer_idx in pool_after:
next_bottom = 'pool%d' %layer_idx
n.tops[next_bottom] = L.Pooling(n.tops[curr_bottom], kernel_size=3, stride=2, pool=P.Pooling.MAX)
curr_bottom = next_bottom
n.tops['drop'+str(layer_idx)] = L.Dropout(n.tops[curr_bottom], dropout_ratio=0.5, in_place=True)
#optional pre_conv_final (w/ appropriate CEratio)
#n.pre_conv_final = L.Convolution(n.tops[curr_bottom], kernel_size=1, num_output=int(1000*s['CEratio']), stride=1, weight_filler=dict(type='xavier'))
#n.tops['relu_pre_conv_final'] = L.ReLU(n.tops['pre_conv_final'], in_place=True)
#curr_bottom='pre_conv_final'
n.tops['conv_final'] = L.Convolution(n.tops[curr_bottom], kernel_size=1, num_output=1000, weight_filler=dict(type='gaussian', std=0.01, mean=0.0))
n.tops['relu_conv_final'] = L.ReLU(n.tops['conv_final'], in_place=True)
n.tops['pool_final'] = L.Pooling(n.tops['conv_final'], global_pooling=1, pool=P.Pooling.AVE)
if phase == 'trainval':
n.loss = L.SoftmaxWithLoss(n.tops['pool_final'], n.label, include=dict(phase=caffe_pb2.TRAIN))
n.accuracy = L.Accuracy(n.tops['pool_final'], n.label, include=dict(phase=caffe_pb2.TEST))
n.accuracy_top5 = L.Accuracy(n.tops['pool_final'], n.label, include=dict(phase=caffe_pb2.TEST), top_k=5)
return n.to_proto()
def lenet(lmdbData, lmdbLabel, batch_size):
n = NetSpec()
n.data = L.Data(batch_size=batch_size, backend=P.Data.LMDB, source=lmdbData,
transform_param=dict(scale=1./255), ntop=1)
n.label = L.Data(batch_size=batch_size, backend=P.Data.LMDB, source=lmdbLabel,
transform_param=dict(scale=1./255), ntop=1)
n.conv1 = L.Convolution(n.data, kernel_size=4, num_output=200, weight_filler=dict(type='xavier'))
n.pool1 = L.Pooling(n.conv1, kernel_size=2, stride=2, pool=P.Pooling.MAX)
n.conv2 = L.Convolution(n.pool1, kernel_size=3, num_output=50, weight_filler=dict(type='xavier'))
n.pool2 = L.Pooling(n.conv2, kernel_size=2, stride=1, pool=P.Pooling.MAX)
n.fc1 = L.InnerProduct(n.pool2, num_output=200, weight_filler=dict(type='xavier'))
n.relu1 = L.ReLU(n.fc1, in_place=True)
n.score = L.InnerProduct(n.relu1, num_output=1200, weight_filler=dict(type='xavier'))
n.loss = L.Python(n.score, n.label, module='pyloss', layer='EuclideanLossLayer')
return n.to_proto()
def gen_net(batch_size=512):
n=NetSpec();
n.data = L.DummyData(shape={"dim":[batch_size,3,96,96]})
n.select1 = L.DummyData(shape={"dim":[2]})
n.select2 = L.DummyData(shape={"dim":[2]})
n.label = L.DummyData(shape={"dim":[2]})
caffenet_stack(n.data, n)
n.first = L.BatchReindex(n.fc6, n.select1)
n.second = L.BatchReindex(n.fc6, n.select2)
n.fc6_concat=L.Concat(n.first, n.second);
n.fc7, n.bn7, n.relu7 = fc_relu(n.fc6_concat, 4096, batchnorm=True);
n.fc8, n.relu8 = fc_relu(n.relu7, 4096);
n.fc9 = L.InnerProduct(n.relu8, num_output=8,
weight_filler=dict(type='xavier'));
n.loss = L.SoftmaxWithLoss(n.fc9, n.label, loss_param=dict(normalization=P.Loss.NONE));
prot=n.to_proto()
prot.debug_info=True
return prot;
def train_tail(self, last_top):
n = NetSpec()
n.loss = L.SoftmaxWithLoss(bottom=[last_top, "label"])
return n.to_proto()
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 StickNet(batch_size, s):
inImgH = 224 #TODO: put inImg{H,W} into 's' if necessary.
inImgW = 224
round_to_nearest = 4
n = NetSpec()
FireNet_data_layer(n, batch_size) #add data layer to the net
curr_bottom='data'
#layer-to-layer counters
_totalStride = 1 #note that, using 1x1 conv, our (stride>1) is only in pooling layers.
_numPoolings = 1 #for indexing 'conv2_1', etc.
_ch=3
[activH, activW] = est_activ_size(inImgH, inImgW, _totalStride)
n_filt = choose_num_output(1, 1, _ch, activH, activW, s['mflop_per_img_target'], s['n_layers']) #only using this for conv1 to avoid oscillations.
n_filt = round_to(n_filt, round_to_nearest) #make divisible by 8
#FIXME: somehow account for num_output produced by conv1 when selecting number of filters for conv2. (else, conv2 goes way over budget on flops.)
# perhaps we need to find the number N such that N^2*activations = mflop_per_img_target?
idx_minor = 1
idx_major = 1
#this goes to (n_layers-1) ... then we do conv_final separately because it has a different weight init.
for layer_idx in xrange(1, s['n_layers']):
layer_str = '%d.%d' %(idx_major, idx_minor)
#select number of filters in this layer:
#[activH, activW] = est_activ_size(inImgH, inImgW, _totalStride)
#n_filt = choose_num_output(1, 1, _ch, activH, activW, s['mflop_per_img_target'], s['n_layers'])
#TODO: to avoid oscillations, perhaps just use choose_num_output for conv1,
# and then just double n_filt whenever we do stride=2.
#generate layer
ksize=1
stride=1
pad=0
curr_bottom = conv_relu_xavier(n, ksize, n_filt, layer_str, stride, pad, curr_bottom)
_ch = n_filt #for next layer
if layer_idx in s['pool_after'].keys():
pinfo = s['pool_after'][layer_idx]
#next_bottom = 'pool%d' %layer_idx
next_bottom = 'pool_' + layer_str
n.tops[next_bottom] = L.Pooling(n.tops[curr_bottom], kernel_size=pinfo['kernel_size'], stride=pinfo['stride'], pool=P.Pooling.MAX)
curr_bottom = next_bottom
_totalStride = _totalStride * pinfo['stride']
_numPoolings = _numPoolings + 1
n_filt = n_filt * pinfo['stride'] #to keep (most) layers at roughly the same complexity-per-layer
idx_major = idx_major + 1
idx_minor = 1
else:
idx_minor = idx_minor + 1
n.tops['drop'+str(layer_idx)] = L.Dropout(n.tops[curr_bottom], dropout_ratio=0.5, in_place=True)
n.tops['conv_final'] = L.Convolution(n.tops[curr_bottom], kernel_size=1, num_output=1000, weight_filler=dict(type='gaussian', std=0.01, mean=0.0))
n.tops['relu_conv_final'] = L.ReLU(n.tops['conv_final'], in_place=True)
n.tops['pool_final'] = L.Pooling(n.tops['conv_final'], global_pooling=1, pool=P.Pooling.AVE)
if phase == 'trainval':
n.loss = L.SoftmaxWithLoss(n.tops['pool_final'], n.label, include=dict(phase=caffe_pb2.TRAIN))
n.accuracy = L.Accuracy(n.tops['pool_final'], n.label, include=dict(phase=caffe_pb2.TEST))
n.accuracy_top5 = L.Accuracy(n.tops['pool_final'], n.label, include=dict(phase=caffe_pb2.TEST), top_k=5)
return n.to_proto()
