def gen_DNN():
net = RandomDNN()
data_layer_str = random_net_defs.dataLayerStr_deploy(256, 3, 227, 227)
#data_layer_str = random_net_defs.dataLayerStr_trainval('LMDB', '/path/to/train_lmdb', 'LMDB', '/path/to/test_lmdb', 256, 50, 227, 'examples/imagenet/ilsvrc12_train_lmdb')
print data_layer_str
prev_layer_type = 'data'
prev_layer_name = 'data_layer'
for i in xrange(0, 10):
curr_layer_type = net.chooseNextLayer(prev_layer_type)
if curr_layer_type == 'conv':
curr_layer_dict = net.convLayer(i, prev_layer_name)
if curr_layer_type == 'pool':
curr_layer_dict = net.poolLayer(i, prev_layer_name)
if curr_layer_type == 'relu':
curr_layer_dict = net.reluLayer(i, prev_layer_name)
if curr_layer_type == 'lrn':
curr_layer_dict = net.lrnLayer(i, prev_layer_name)
print curr_layer_dict['prototxt']
prev_layer_name = curr_layer_dict[
'name'] #TODO: only update this for layers that can't be computed in place?
prev_layer_type = curr_layer_type
#boilerplate fully-connected (fast to compute, can't hurt.)
print random_net_defs.fcLayerStr('fc6', prev_layer_name, 'fc6', 4096)
print random_net_defs.reluLayerStr('relu6', 'fc6', 'fc6') #in-place ReLU
print random_net_defs.dropoutLayerStr('drop6', 'fc6', 'fc6')
print random_net_defs.fcLayerStr('fc7', 'fc6', 'fc7', 4096)
print random_net_defs.reluLayerStr('relu7', 'fc7', 'fc7')
print random_net_defs.dropoutLayerStr('drop7', 'fc7', 'fc7')
print random_net_defs.fcLayerStr('fc8', 'fc7', 'fc8', 1000)
def gen_DNN(phase):
net = RandomDNN()
CAFFE_ROOT='/lustre/atlas/scratch/forresti/csc103/dnn_exploration/caffe-bvlc-master'
DATA_PATH='/lustre/atlas/scratch/forresti/csc103/dnn_exploration'
if phase == 'deploy':
data_layer_str = random_net_defs.dataLayerStr_deploy(256, 3, 227, 227)
elif phase == 'trainval':
data_layer_str = random_net_defs.dataLayerStr_trainval('LMDB', DATA_PATH+'/ilsvrc2012_train_256x256_lmdb', 'LMDB', DATA_PATH+'/ilsvrc2012_val_256x256_lmdb', 256, 50, 227, CAFFE_ROOT+'/data/ilsvrc12/imagenet_mean.binaryproto')
else:
print "Warning: didn't generate data_layer. phase must be 'deploy' or 'trainval'"
print data_layer_str
downsampleTo = 227 #keep this below 227 (input img height = width = 227)
prev_layer_type='data'
prev_layer_name='data_layer'
for i in xrange(0, 10):
curr_layer_type = net.chooseNextLayer(prev_layer_type)
if curr_layer_type == 'conv':
tmp_layer_dict = net.convLayer(i, prev_layer_name, downsampleTo)
if tmp_layer_dict['downsampleTo'] > 0:
curr_layer_dict = tmp_layer_dict
downsampleTo = curr_layer_dict['downsampleTo']
#print 'downsampleTo: ', curr_layer_dict['downsampleTo']
else: #if the new layer would downsample below 1x1, ignore the layer.
continue
if curr_layer_type == 'pool':
tmp_layer_dict = net.poolLayer(i, prev_layer_name, downsampleTo)
if tmp_layer_dict['downsampleTo'] > 0:
curr_layer_dict = tmp_layer_dict
downsampleTo = curr_layer_dict['downsampleTo']
#print 'downsampleTo: ', curr_layer_dict['downsampleTo']
else:
continue
if curr_layer_type == 'relu':
curr_layer_dict = net.reluLayer(i, prev_layer_name)
if curr_layer_type == 'lrn':
curr_layer_dict = net.lrnLayer(i, prev_layer_name)
print curr_layer_dict['prototxt']
prev_layer_name = curr_layer_dict['name'] #TODO: only update this for layers that can't be computed in place?
prev_layer_type = curr_layer_type
#boilerplate fully-connected (fast to compute, can't hurt.)
print random_net_defs.fcLayerStr('fc6', prev_layer_name, 'fc6', 4096)
print random_net_defs.reluLayerStr('relu6', 'fc6', 'fc6') #in-place ReLU
print random_net_defs.dropoutLayerStr('drop6', 'fc6', 'fc6')
print random_net_defs.fcLayerStr('fc7', 'fc6', 'fc7', 4096)
print random_net_defs.reluLayerStr('relu7', 'fc7', 'fc7')
print random_net_defs.dropoutLayerStr('drop7', 'fc7', 'fc7')
print random_net_defs.fcLayerStr('fc8', 'fc7', 'fc8', 1000)
if phase == 'trainval':
#boilerplate scoring (use only if trainval)
print random_net_defs.scoringTrainvalStr('fc8')
def reluLayer(self, layerIdx, prevLayerStr):
myName = "layer" + str(layerIdx) + "_relu"
bottom = prevLayerStr
top = myName #not bothering with in-place
#print myName
retStr = random_net_defs.reluLayerStr(myName, bottom, top)
return {'name':myName, 'prototxt':retStr}
def reluLayer(self, layerIdx, prevLayerStr):
myName = "layer" + str(layerIdx) + "_relu"
bottom = prevLayerStr
top = myName #not bothering with in-place
#print myName
retStr = random_net_defs.reluLayerStr(myName, bottom, top)
return {'name': myName, 'prototxt': retStr}
def gen_DNN(phase):
net = RandomDNN()
CAFFE_ROOT = '/lustre/atlas/scratch/forresti/csc103/dnn_exploration/caffe-bvlc-master'
DATA_PATH = '/lustre/atlas/scratch/forresti/csc103/dnn_exploration'
if phase == 'deploy':
data_layer_str = random_net_defs.dataLayerStr_deploy(256, 3, 227, 227)
elif phase == 'trainval':
data_layer_str = random_net_defs.dataLayerStr_trainval(
'LMDB', DATA_PATH + '/ilsvrc2012_train_256x256_lmdb', 'LMDB',
DATA_PATH + '/ilsvrc2012_val_256x256_lmdb', 256, 50, 227,
CAFFE_ROOT + '/data/ilsvrc12/imagenet_mean.binaryproto')
else:
print "Warning: didn't generate data_layer. phase must be 'deploy' or 'trainval'"
print data_layer_str
downsampleTo = 227 #keep this below 227 (input img height = width = 227)
prev_layer_type = 'data'
prev_layer_name = 'data_layer'
for i in xrange(0, 10):
curr_layer_type = net.chooseNextLayer(prev_layer_type)
if curr_layer_type == 'conv':
tmp_layer_dict = net.convLayer(i, prev_layer_name, downsampleTo)
if tmp_layer_dict['downsampleTo'] > 0:
curr_layer_dict = tmp_layer_dict
downsampleTo = curr_layer_dict['downsampleTo']
#print 'downsampleTo: ', curr_layer_dict['downsampleTo']
else: #if the new layer would downsample below 1x1, ignore the layer.
continue
if curr_layer_type == 'pool':
tmp_layer_dict = net.poolLayer(i, prev_layer_name, downsampleTo)
if tmp_layer_dict['downsampleTo'] > 0:
curr_layer_dict = tmp_layer_dict
downsampleTo = curr_layer_dict['downsampleTo']
#print 'downsampleTo: ', curr_layer_dict['downsampleTo']
else:
continue
if curr_layer_type == 'relu':
curr_layer_dict = net.reluLayer(i, prev_layer_name)
if curr_layer_type == 'lrn':
curr_layer_dict = net.lrnLayer(i, prev_layer_name)
print curr_layer_dict['prototxt']
prev_layer_name = curr_layer_dict[
'name'] #TODO: only update this for layers that can't be computed in place?
prev_layer_type = curr_layer_type
#boilerplate fully-connected (fast to compute, can't hurt.)
print random_net_defs.fcLayerStr('fc6', prev_layer_name, 'fc6', 4096)
print random_net_defs.reluLayerStr('relu6', 'fc6', 'fc6') #in-place ReLU
print random_net_defs.dropoutLayerStr('drop6', 'fc6', 'fc6')
print random_net_defs.fcLayerStr('fc7', 'fc6', 'fc7', 4096)
print random_net_defs.reluLayerStr('relu7', 'fc7', 'fc7')
print random_net_defs.dropoutLayerStr('drop7', 'fc7', 'fc7')
print random_net_defs.fcLayerStr('fc8', 'fc7', 'fc8', 1000)
if phase == 'trainval':
#boilerplate scoring (use only if trainval)
print random_net_defs.scoringTrainvalStr('fc8')
