def convert_caffemodel_to_prototxt(caffemodel_filename):
model = caffe_pb2.NetParameter()
# load model from binary file
f = open(caffemodel_filename, 'rb')
model.ParseFromString(f.read())
f.close()
layers = model.layer
res = list()
res.append('name: ' + model.name)
for layer in layers:
# name
res.append('layer {')
res.append(' name: "%s"' % layer.name)
# type
res.append(' type: "%s"' % layer.type)
# bottom
for bottom in layer.bottom:
res.append(' bottom: "%s"' % bottom)
# top
for top in layer.top:
res.append(' top: "%s"' % top)
# loss_weight
for loss_weight in layer.loss_weight:
res.append(' loss_weight: %0.3f' % loss_weight)
# param
for param in layer.param:
param_res = list()
if param.lr_mult is not None:
param_res.append(' lr_mult: %0.3f' % param.lr_mult)
if param.decay_mult != 1:
param_res.append(' decay_mult: %0.3f' % param.decay_mult)
if len(param_res) > 0:
res.append(' param{')
res.extend(param_res)
res.append(' }')
# lrn_param
if layer.lrn_param is not None:
lrn_res = list()
if layer.lrn_param.local_size != 5:
lrn_res.append(' local_size: %d' %
layer.lrn_param.local_size)
if layer.lrn_param.alpha != 1:
lrn_res.append(' alpha: %f' % layer.lrn_param.alpha)
if layer.lrn_param.beta != 0.75:
lrn_res.append(' beta: %f' % layer.lrn_param.beta)
NormRegionMapper = {'0': 'ACROSS_CHANNELS', '1': 'WITHIN_CHANNEL'}
if layer.lrn_param.norm_region != 0:
lrn_res.append(
' norm_region: %s' %
NormRegionMapper[str(layer.lrn_param.norm_region)])
EngineMapper = {'0': 'DEFAULT', '1': 'CAFFE', '2': 'CUDNN'}
if layer.lrn_param.engine != 0:
lrn_res.append(' engine: %s' %
EngineMapper[str(layer.lrn_param.engine)])
if len(lrn_res) > 0:
res.append(' lrn_param{')
res.extend(lrn_res)
res.append(' }')
# include
if len(layer.include) > 0:
include_res = list()
includes = layer.include
phase_mapper = {'0': 'TRAIN', '1': 'TEST'}
for include in includes:
if include.phase is not None:
include_res.append(' phase: %s' %
phase_mapper[str(include.phase)])
if len(include_res) > 0:
res.append(' include {')
res.extend(include_res)
res.append(' }')
# transform_param
if layer.transform_param is not None:
transform_param_res = list()
if layer.transform_param.scale != 1:
transform_param_res.append(' scale: %s' %
layer.transform_param.scale)
if layer.transform_param.mirror != False:
transform_param_res.append(' mirror: true')
if len(transform_param_res) > 0:
res.append(' transform_param {')
res.extend(transform_param_res)
res.append(' }')
# data_param
if layer.data_param is not None and (layer.data_param.source != "" or
layer.data_param.batch_size != 0
or layer.data_param.backend != 0):
data_param_res = list()
if layer.data_param.source is not None:
data_param_res.append(' source: "%s"' %
layer.data_param.source)
if layer.data_param.batch_size is not None:
data_param_res.append(' batch_size: %d' %
layer.data_param.batch_size)
if layer.data_param.backend is not None:
data_param_res.append(' backend: %s' %
layer.data_param.backend)
if len(data_param_res) > 0:
res.append(' data_param: {')
res.extend(data_param_res)
res.append(' }')
# convolution_param
if layer.convolution_param is not None:
convolution_param_res = list()
conv_param = layer.convolution_param
if conv_param.num_output != 0:
convolution_param_res.append(' num_output: %d' %
conv_param.num_output)
if len(conv_param.kernel_size) > 0:
for kernel_size in conv_param.kernel_size:
convolution_param_res.append(' kernel_size: %d' %
kernel_size)
if len(conv_param.pad) > 0:
for pad in conv_param.pad:
convolution_param_res.append(' pad: %d' % pad)
if len(conv_param.stride) > 0:
for stride in conv_param.stride:
convolution_param_res.append(' stride: %d' % stride)
if conv_param.weight_filler is not None and conv_param.weight_filler.type != 'constant':
convolution_param_res.append(' weight_filler {')
convolution_param_res.append(' type: "%s"' %
conv_param.weight_filler.type)
convolution_param_res.append(' }')
if conv_param.bias_filler is not None and conv_param.bias_filler.type != 'constant':
convolution_param_res.append(' bias_filler {')
convolution_param_res.append(' type: "%s"' %
conv_param.bias_filler.type)
convolution_param_res.append(' }')
if len(convolution_param_res) > 0:
res.append(' convolution_param {')
res.extend(convolution_param_res)
res.append(' }')
# pooling_param
if layer.pooling_param is not None:
pooling_param_res = list()
if layer.pooling_param.kernel_size > 0:
pooling_param_res.append(' kernel_size: %d' %
layer.pooling_param.kernel_size)
pooling_param_res.append(' stride: %d' %
layer.pooling_param.stride)
pooling_param_res.append(' pad: %d' %
layer.pooling_param.pad)
PoolMethodMapper = {'0': 'MAX', '1': 'AVE', '2': 'STOCHASTIC'}
pooling_param_res.append(
' pool: %s' %
PoolMethodMapper[str(layer.pooling_param.pool)])
if len(pooling_param_res) > 0:
res.append(' pooling_param {')
res.extend(pooling_param_res)
res.append(' }')
# inner_product_param
if layer.inner_product_param is not None:
inner_product_param_res = list()
if layer.inner_product_param.num_output != 0:
inner_product_param_res.append(
' num_output: %d' %
layer.inner_product_param.num_output)
if len(inner_product_param_res) > 0:
res.append(' inner_product_param {')
res.extend(inner_product_param_res)
res.append(' }')
# drop_param
if layer.dropout_param is not None:
dropout_param_res = list()
if layer.dropout_param.dropout_ratio != 0.5: #or layer.dropout_param.scale_train!=True:
dropout_param_res.append(' dropout_ratio: %f' %
layer.dropout_param.dropout_ratio)
#dropout_param_res.append(' scale_train: ' + str(layer.dropout_param.scale_train))
if len(dropout_param_res) > 0:
res.append(' dropout_param {')
res.extend(dropout_param_res)
res.append(' }')
res.append('}')
net_string = ""
for line in res:
net_string += line + '\n'
return net_string
def make_test_files(testable_net_path, train_weights_path, num_iterations,
in_h, in_w):
# load the train net prototxt as a protobuf message
with open(testable_net_path) as f:
testable_str = f.read()
testable_msg = caffe_pb2.NetParameter()
text_format.Merge(testable_str, testable_msg)
bn_layers = [l.name for l in testable_msg.layer if l.type == "BN"]
bn_blobs = [l.top[0] for l in testable_msg.layer if l.type == "BN"]
bn_means = [l.top[1] for l in testable_msg.layer if l.type == "BN"]
bn_vars = [l.top[2] for l in testable_msg.layer if l.type == "BN"]
net = caffe.Net(testable_net_path, train_weights_path, caffe.TEST)
# init our blob stores with the first forward pass
res = net.forward()
bn_avg_mean = {
bn_mean: np.squeeze(res[bn_mean]).copy()
for bn_mean in bn_means
}
bn_avg_var = {bn_var: np.squeeze(res[bn_var]).copy() for bn_var in bn_vars}
# iterate over the rest of the training set
for i in xrange(1, num_iterations):
res = net.forward()
for bn_mean in bn_means:
bn_avg_mean[bn_mean] += np.squeeze(res[bn_mean])
for bn_var in bn_vars:
bn_avg_var[bn_var] += np.squeeze(res[bn_var])
print 'progress: {}/{}'.format(i, num_iterations)
# compute average means and vars
for bn_mean in bn_means:
bn_avg_mean[bn_mean] /= num_iterations
for bn_var in bn_vars:
bn_avg_var[bn_var] /= num_iterations
for bn_blob, bn_var in zip(bn_blobs, bn_vars):
m = np.prod(net.blobs[bn_blob].data.shape) / np.prod(
bn_avg_var[bn_var].shape)
bn_avg_var[bn_var] *= (m / (m - 1))
# calculate the new scale and shift blobs for all the BN layers
scale_data = {
bn_layer: np.squeeze(net.params[bn_layer][0].data)
for bn_layer in bn_layers
}
shift_data = {
bn_layer: np.squeeze(net.params[bn_layer][1].data)
for bn_layer in bn_layers
}
var_eps = 1e-9
new_scale_data = {}
new_shift_data = {}
for bn_layer, bn_mean, bn_var in zip(bn_layers, bn_means, bn_vars):
gamma = scale_data[bn_layer]
beta = shift_data[bn_layer]
Ex = bn_avg_mean[bn_mean]
Varx = bn_avg_var[bn_var]
new_gamma = gamma / np.sqrt(Varx + var_eps)
new_beta = beta - (gamma * Ex / np.sqrt(Varx + var_eps))
new_scale_data[bn_layer] = new_gamma
new_shift_data[bn_layer] = new_beta
print "New data:"
print new_scale_data.keys()
print new_shift_data.keys()
# assign computed new scale and shift values to net.params
for bn_layer in bn_layers:
net.params[bn_layer][0].data[...] = new_scale_data[bn_layer].reshape(
net.params[bn_layer][0].data.shape)
net.params[bn_layer][1].data[...] = new_shift_data[bn_layer].reshape(
net.params[bn_layer][1].data.shape)
# build a test net prototxt
test_msg = testable_msg
# replace data layers with 'input' net param
data_layers = [l for l in test_msg.layer if l.type.endswith("Data")]
for data_layer in data_layers:
test_msg.layer.remove(data_layer)
test_msg.input.append("data")
test_msg.input_dim.append(1)
test_msg.input_dim.append(3)
test_msg.input_dim.append(in_h)
test_msg.input_dim.append(in_w)
# Set BN layers to INFERENCE so they use the new stat blobs
# and remove mean, var top blobs.
for l in test_msg.layer:
if l.type == "BN":
if len(l.top) > 1:
dead_tops = l.top[1:]
for dl in dead_tops:
l.top.remove(dl)
l.bn_param.bn_mode = caffe_pb2.BNParameter.INFERENCE
# replace output loss, accuracy layers with a softmax
dead_outputs = [
l for l in test_msg.layer if l.type in ["SoftmaxWithLoss", "Accuracy"]
]
out_bottom = dead_outputs[0].bottom[0]
for dead in dead_outputs:
test_msg.layer.remove(dead)
test_msg.layer.add(name="prob",
type="Softmax",
bottom=[out_bottom],
top=['prob'])
return net, test_msg
def modify_net_prototxt(net_prototxt,
new_net_prototxt,
moving_average_fraction=0.9999):
# Reads net parameters from prototxt
net_param = caffe_pb2.NetParameter()
with open(net_prototxt) as f:
txtf.Merge(str(f.read()), net_param)
flag = True
for l in net_param.layer:
# Comment / uncomment / adapt the following blocks as needed
# if l.type == "Convolution" and l.convolution_param.kernel_size[0] > 1:
# l.param[0].lr_mult = 0.0
# l.param[1].lr_mult = 0.0
# if l.type in ["Convolution", "InnerProduct"]: # and l.name != "mid_fc8":
# l.param[0].lr_mult = 1.0
# l.param[1].lr_mult = 1.0
# if l.type == "Convolution":
# if l.convolution_param.group !=1:
# l.type = "DepthwiseConvolution"
# if l.type in ["Convolution","DepthwiseConvolution"]:
# l.param[0].lr_mult = 1.0
# l.param[1].lr_mult = 1.0
# if l.name == "conv5_4/sep":
# flag = False
if l.type == "BatchReNorm":
# l.batch_renorm_param.step_to_init = 48
# l.batch_renorm_param.step_to_r_max = 96
# l.batch_renorm_param.step_to_d_max = 96
# l.batch_renorm_param.r_max = 3
# l.batch_renorm_param.d_max = 5
# l.batch_renorm_param.use_global_stats = flag
l.batch_renorm_param.step_to_init = 0
l.batch_renorm_param.step_to_r_max = 1
l.batch_renorm_param.step_to_d_max = 1
l.batch_renorm_param.moving_average_fraction = moving_average_fraction # 0.99995
l.batch_renorm_param.r_max = 1.25
l.batch_renorm_param.d_max = 0.5
# l.batch_renorm_param.r_max = 1.5
# l.batch_renorm_param.d_max = 2.5
# if l.type == "Scale":
# l.param[0].lr_mult = 0.0
# l.param[1].lr_mult = 0.0
# if l.type == "Convolution" and l.convolution_param.kernel_size[0] > 1:
# l.param[0].lr_mult = 0.0
# if l.type == "DepthwiseConvolution" and l.convolution_param.kernel_size[0] > 1:
# l.param[0].lr_mult = 0.0
# if l.type == "Convolution" and l.convolution_param.kernel_size[0] == 1:
# l.param[0].lr_mult = 0.0
with open(new_net_prototxt, 'w') as f:
f.write(str(net_param))
def __init__(self, name="network"):
self.net = caffe_pb2.NetParameter()
self.net.name = name
self.bottom = None
self.cur = None
self.this = None
def caffe_to_sas(network_file, model_name, network_param=None,
phase=caffe.TEST, verbose=False):
'''
Generate a SAS deep learning model from Caffe definition
Parameters
----------
network_file : string
Fully qualified file name of network definition file (*.prototxt).
sas_file : string
Fully qualified file name of SAS deep learning Python model definition.
model_name : string
Name for deep learning model.
network_param : string, optional
Fully qualified file name of network parameter file (*.caffemodel).
phase : int, optional
One of {caffe.TRAIN, caffe.TEST, None}.
verbose : bool, optional
To view all Caffe information messages, set to True.
'''
# open output file
try:
output_code = ''
# initialize Caffe logging facility
caffe.init_log(0, verbose)
# instantiate a model and read network parameters
if (network_param is None):
model = caffe.Net(network_file, phase)
else:
model = caffe.Net(network_file, phase, weights=network_param)
net = caffe_pb2.NetParameter()
text_format.Merge(open(network_file + '.tmp').read(), net)
# remove temporary file created
if os.path.isfile(network_file + '.tmp'):
os.remove(network_file + '.tmp')
# identify common Caffe/SAS computation layers
layer_list = []
for layer in net.layer:
include_layer = False
if (len(layer.include) == 0):
include_layer = True
else:
for layer_phase in layer.include:
if (caffe.TEST == layer_phase.phase):
include_layer = True
# exclude layers not implemented (or implemented in a different fashion)
if (layer.type.lower() not in common_layers):
include_layer = False
if include_layer:
layer_list.append(make_composite_layer(layer))
# associate activations with computation layers
for layer in net.layer:
layer_type = layer.type.lower()
if (layer_type in ['relu', 'prelu', 'elu', 'sigmoid', 'tanh']):
layer_index = None
for ii in range(len(layer_list)):
if (layer.top[0] == layer_list[ii].layer_parm.top[0]):
layer_index = ii
if layer_index is not None:
layer_list[layer_index].related_layers.append(layer)
else:
raise CaffeParseError(
'Activation layer ' + layer.name +
' is not associated with any computation layer.')
# associate dropout with computation layers
for layer in net.layer:
layer_type = layer.type.lower()
if (layer_type == 'dropout'):
layer_index = None
for ii in range(len(layer_list)):
if (layer.top[0] == layer_list[ii].layer_parm.top[0]):
layer_index = ii
if layer_index is not None:
layer_list[layer_index].related_layers.append(layer)
else:
raise CaffeParseError(
'Dropout layer ' + layer.name +
' is not associated with any computation layer.')
# associate softmax with a fully-connected layer
for layer in net.layer:
layer_type = layer.type.lower()
if (layer_type in ['softmax', 'softmaxwithloss']):
layer_index = None
for ii in range(len(layer_list)):
for jj in range(len(layer.bottom)):
if (layer.bottom[jj] == layer_list[ii].layer_parm.top[0]):
layer_index = ii
if layer_index is not None:
layer_list[layer_index].related_layers.append(layer)
else:
raise CaffeParseError(
'Softmax layer ' + layer.name +
' is not associated with any fully-connected layer.')
# determine source layer(s) for computation layers
for ii in range(len(layer_list)):
for kk in range(len(layer_list[ii].layer_parm.bottom)):
name = None
for jj in range(ii):
if (layer_list[ii].layer_parm.bottom[kk] ==
layer_list[jj].layer_parm.top[0]):
name = layer_list[jj].layer_parm.name
if name:
layer_list[ii].source_layer.append(name)
# associate scale layer with batchnorm layer
for layer in net.layer:
if (layer.type.lower() == 'scale'):
bn_found = False
for ii in range(len(layer_list)):
if ((layer_list[ii].layer_parm.type.lower() == 'batchnorm') and
(layer_list[ii].layer_parm.top[0] == layer.top[0])):
layer_list[ii].related_layers.append(layer)
bn_found = True
break
if not bn_found:
raise CaffeParseError(
'Scale layer ' + layer.name +
' is not associated with a batch normalization layer')
# loop over included layers
for clayer in layer_list:
layer_type = clayer.layer_parm.type.lower()
if (layer_type == 'pooling'): # average/max pooling
sas_code = caffe_pooling_layer(clayer, model_name)
elif (layer_type == 'convolution'): # 2D convolution
sas_code = caffe_convolution_layer(clayer, model_name)
elif (layer_type == 'batchnorm'): # batch normalization
sas_code = caffe_batch_normalization_layer(clayer, model_name)
elif (layer_type in ['data', 'memorydata']): # input layer
sas_code = caffe_input_layer(clayer, model_name)
elif (layer_type == 'eltwise'): # residual
sas_code = caffe_residual_layer(clayer, model_name)
elif (layer_type == 'innerproduct'): # fully connected
sas_code = caffe_full_connect_layer(clayer, model_name)
else:
raise CaffeParseError(layer_type +
' is an unsupported layer type')
# write SAS code associated with Caffe layer
if sas_code:
output_code = output_code + sas_code + '\n\n'
else:
raise CaffeParseError(
'Unable to generate SAS definition for layer ' +
clayer.layer_parm.name)
# convert from BINARYPROTO to HDF5
if network_param is not None:
sas_hdf5 = os.path.join(os.getcwd(), '{}_weights.h5'.format(model_name))
write_caffe_hdf5(model, layer_list, sas_hdf5)
return output_code
except CaffeParseError as err_msg:
print(err_msg)
def proto_decomposition(network_in, network_out, layers_ranks):
proto_in = caffe_pb2.NetParameter()
with open(network_in, 'r') as file:
text_format.Merge(str(file.read()), proto_in)
proto_out = caffe_pb2.NetParameter()
proto_out.CopyFrom(proto_in)
proto_out.ClearField('layer')
channel_buffer = {}
for layer in proto_in.layer:
if layer.type != 'Convolution' and layer.type != 'InnerProduct':
proto_out.layer.add()
proto_out.layer[-1].CopyFrom(layer)
if layer.type == 'Data':
channel_buffer[layer.top[0]] = 3
channel_buffer[layer.top[1]] = 1
else:
channel_buffer[layer.top[0]] = channel_buffer[layer.bottom[0]]
elif layer.type == 'Convolution':
channel_buffer[layer.top[0]] = layer.convolution_param.num_output
layer.convolution_param.ClearField('weight_filler')
layer.convolution_param.ClearField('bias_filler')
if layer.name not in layers_ranks:
proto_out.layer.add()
proto_out.layer[-1].CopyFrom(layer)
else:
if layers_ranks[layer.name][0] != channel_buffer[
layer.bottom[0]]:
proto_out.layer.add()
lra_a_layer = proto_out.layer[-1]
lra_a_layer.CopyFrom(layer)
lra_a_layer.name += '_lra_a'
lra_a_layer.convolution_param.kernel_size[0] = 1
lra_a_layer.convolution_param.num_output = layers_ranks[
layer.name][0]
lra_a_layer.convolution_param.ClearField('pad')
lra_a_layer.convolution_param.ClearField('stride')
lra_a_layer.top[0] = layer.name + '_lra_a'
channel_buffer[lra_a_layer.top[0]] = layers_ranks[
layer.name][0]
proto_out.layer.add()
lra_b_layer = proto_out.layer[-1]
lra_b_layer.CopyFrom(layer)
lra_b_layer.name += '_lra_b'
lra_b_layer.convolution_param.num_output = layers_ranks[
layer.name][1]
if layer.name + '_lra_a' in channel_buffer:
lra_b_layer.bottom[0] = layer.name + '_lra_a'
if layers_ranks[layer.name][1] != channel_buffer[layer.top[0]]:
lra_b_layer.top[0] = layer.name + '_lra_b'
proto_out.layer.add()
lra_c_layer = proto_out.layer[-1]
lra_c_layer.CopyFrom(layer)
lra_c_layer.name += '_lra_c'
lra_c_layer.convolution_param.kernel_size[0] = 1
lra_c_layer.convolution_param.ClearField('pad')
lra_c_layer.convolution_param.ClearField('stride')
lra_c_layer.bottom[0] = layer.name + '_lra_b'
channel_buffer[lra_c_layer.bottom[0]] = layers_ranks[
layer.name][1]
elif layer.type == 'InnerProduct':
channel_buffer[layer.top[0]] = layer.inner_product_param.num_output
layer.inner_product_param.ClearField('weight_filler')
layer.inner_product_param.ClearField('bias_filler')
if layer.name not in layers_ranks:
proto_out.layer.add()
proto_out.layer[-1].CopyFrom(layer)
else:
proto_out.layer.add()
svd_a_layer = proto_out.layer[-1]
svd_a_layer.CopyFrom(layer)
svd_a_layer.name += '_svd_a'
svd_a_layer.inner_product_param.num_output = layers_ranks[
layer.name]
if layers_ranks[layer.name] != channel_buffer[layer.top[0]]:
svd_a_layer.top[0] = layer.name + '_svd_a'
channel_buffer[svd_a_layer.top[0]] = layers_ranks[
layer.name]
proto_out.layer.add()
svd_b_layer = proto_out.layer[-1]
svd_b_layer.CopyFrom(layer)
svd_b_layer.name += '_svd_b'
svd_b_layer.bottom[0] = layer.name + '_svd_a'
with open(network_out, 'w') as file:
file.write(text_format.MessageToString(proto_out))
import caffe.proto.caffe_pb2 as caffe_pb2
#caffemodel_filename = 'D:/caffe-windows/examples/mnist/lenet.caffemodel'
caffemodel_filename = 'G:/py/tf2caffe/testBN/lenet.caffemodel'
model = caffe_pb2.NetParameter()
with open(caffemodel_filename, 'rb') as f:
model.ParseFromString(f.read())
layers = model.layer
print('name: "%s"' % model.name)
layer_id = -1
fname = 'testBN/lenetmodel.prototxt'
with open(fname, 'w') as f:
for layer in layers:
f.write('layer{\n')
f.write(str(layer))
f.write('\n}\n')
#f.write('test')
for layer in layers:
if layer.name in ['data_bn']:
print(layer)
for layer in layers:
layer_id = layer_id + 1
def Inpt_OPT_New_Bias(original_prototxt_path, original_model_path,
optimized_prototxt_path, new_model_path, mean_vector,
scale, H, W, input_channel):
net_param = caffe_pb2.NetParameter()
with open(original_prototxt_path, 'rt') as f:
Parse(f.read(), net_param)
layer_num = len(net_param.layer)
new_net_param = caffe_pb2.NetParameter()
new_net_param.name = 'calc_new_bias'
new_net_param.layer.add()
new_net_param.layer[-1].name = "data"
new_net_param.layer[-1].type = 'Input'
new_net_param.layer[-1].top.append('data')
new_net_param.layer[-1].input_param.shape.add()
new_net_param.layer[-1].input_param.shape[-1].dim.append(1)
new_net_param.layer[-1].input_param.shape[-1].dim.append(
int(input_channel))
new_net_param.layer[-1].input_param.shape[-1].dim.append(int(H))
new_net_param.layer[-1].input_param.shape[-1].dim.append(int(W))
target_blob_name = ''
target_layer_name = ''
input_layer_type = ['Data', 'Input', 'AnnotatedData']
for layer_idx in range(0, layer_num):
layer = net_param.layer[layer_idx]
if layer.type not in input_layer_type:
assert (layer.type == 'Convolution' or layer.type == 'InnerProduct'
), "## ERROR : First Layer MUST BE CONV or IP. ##"
new_net_param.layer.extend([layer])
if layer.type == 'Convolution':
try:
assert (
new_net_param.layer[-1].convolution_param.pad[0] == 0
), '## ERROR : MEAN cannot be mearged into CONV with padding > 0. ##'
except:
# padding not set
pass
target_blob_name = layer.top[0]
target_layer_name = layer.name
break
new_proto_name = './tmpfile.prototxt'
with open(new_proto_name, 'wt') as f:
f.write(MessageToString(new_net_param))
caffe.set_mode_cpu()
net = caffe.Net(new_proto_name, str(original_model_path), caffe.TEST)
mean_array = mean_vector * (-1.0) * scale
mean_array = mean_array.reshape(input_channel, 1)
mean_array = np.tile(mean_array,
(1, H * W)).reshape(1, input_channel, H, W)
os.remove(new_proto_name)
net.blobs['data'].data[...] = mean_array
net.forward()
mean_data = net.blobs[target_blob_name].data[...]
mean_data = mean_data.reshape(mean_data.shape[1],
mean_data.shape[2] * mean_data.shape[3])
new_bias = np.mean(mean_data, 1)
print "INPUT PREPROCESS (SUB MEAN) OPT : Calc New Bias Done."
caffe.set_mode_cpu()
net = caffe.Net(original_prototxt_path, str(original_model_path),
caffe.TEST)
if len(net.params[target_layer_name]) == 2:
# with bias
net.params[target_layer_name][1].data[...] += new_bias[...]
net.save(new_model_path)
try:
shutil.copyfile(original_prototxt_path, optimized_prototxt_path)
except:
# same file, not need to copy
pass
print "INPUT PREPROCESS (SUB MEAN) OPT : Merge Mean Done."
print bcolors.OKGREEN + "INPUT PREPROCESS (SUB MEAN) OPT : Model at " + new_model_path + "." + bcolors.ENDC
print bcolors.OKGREEN + "INPUT PREPROCESS (SUB MEAN) OPT : Prototxt at " + optimized_prototxt_path + "." + bcolors.ENDC
print bcolors.WARNING + "INPUT PREPROCESS (SUB MEAN) OPT : ** WARNING ** Remember to set mean values to zero before test !!!" + bcolors.ENDC
else:
net_param = caffe_pb2.NetParameter()
with open(original_prototxt_path, 'rt') as f:
Parse(f.read(), net_param)
layer_num = len(net_param.layer)
for layer_idx in range(0, layer_num):
layer = net_param.layer[layer_idx]
if layer.name == target_layer_name:
if layer.type == 'Convolution':
net_param.layer[
layer_idx].convolution_param.bias_term = True
else:
net_param.layer[
layer_idx].inner_product_param.bias_term = True
break
with open(optimized_prototxt_path, 'wt') as f:
f.write(MessageToString(net_param))
new_net = caffe.Net(optimized_prototxt_path, caffe.TEST)
for param_name in net.params.keys():
for i in range(0, len(net.params[param_name])):
new_net.params[param_name][i].data[
...] = net.params[param_name][i].data[...]
new_net.params[target_layer_name][1].data[...] = new_bias[...]
new_net.save(new_model_path)
print "INPUT PREPROCESS (SUB MEAN) OPT : Merge Mean Done."
print bcolors.OKGREEN + "INPUT PREPROCESS (SUB MEAN) OPT : Model at " + new_model_path + "." + bcolors.ENDC
print bcolors.OKGREEN + "INPUT PREPROCESS (SUB MEAN) OPT : Prototxt at " + optimized_prototxt_path + "." + bcolors.ENDC
print bcolors.WARNING + "INPUT PREPROCESS (SUB MEAN) OPT : ** WARNING ** Remember to set mean values to zero before test !!!" + bcolors.ENDC
def transform_convolutions(model_path,
compiled_model_path,
top_blobs_map,
bottom_blobs_map,
use_unsigned_range,
concat_use_fp32,
unify_concat_scales,
conv_algo,
enable_1st_conv=False):
net = caffe_pb2.NetParameter()
with open(model_path) as f:
s = f.read()
txtf.Merge(s, net)
compiled_net = caffe_pb2.NetParameter()
with open(compiled_model_path) as f:
s = f.read()
txtf.Merge(s, compiled_net)
convs_output_with_relu = analyze_conv_output_with_relu(compiled_net)
# extended convs output with relu is used for convs that cannot fuse with relu due to negative slope
# extended_convs_output_with_relu = analyze_conv_output_with_relu_from_net(convs_output_with_relu, compiled_net, net)
new_net = copy.deepcopy(net)
convolution_layers = [(value, index)
for index, value in enumerate(net.layer)
if value.type in quantize_layers]
compiled_convolution_layers = [
(value, index) for index, value in enumerate(compiled_net.layer)
if value.type in quantize_layers
]
u8_max = 255
s8_max = 127
first_conv = True if enable_1st_conv else False
for (l, index) in convolution_layers:
for si in range(
0, len(new_net.layer[index].quantization_param.scale_out)):
if l.name in convs_output_with_relu: # u8
new_net.layer[index].quantization_param.scale_out[
si] = u8_max / new_net.layer[
index].quantization_param.scale_out[si]
else: # s8
if use_unsigned_range:
new_net.layer[index].quantization_param.scale_out[
si] = u8_max / new_net.layer[
index].quantization_param.scale_out[si]
else:
new_net.layer[index].quantization_param.scale_out[
si] = s8_max / new_net.layer[
index].quantization_param.scale_out[si]
index_in_compiled_net = find_index_by_name(
l.name, compiled_convolution_layers)
assert (index_in_compiled_net >= 0)
#conv_inputs = get_input_convolutions(l, compiled_net, index_in_compiled_net, ["Convolution"])
#conv_input_u8 = analyze_conv_input_u8(conv_inputs, convs_output_with_relu)
conv_input_u8 = is_convolution_input_u8(
l, compiled_net, index_in_compiled_net, ["Convolution"],
convs_output_with_relu) # FIXME: extended_convs_output_with_relu
for si in range(0,
len(new_net.layer[index].quantization_param.scale_in)):
if conv_input_u8: # u8
if first_conv:
new_net.layer[index].quantization_param.scale_in[
si] = s8_max / new_net.layer[
index].quantization_param.scale_in[si]
new_net.layer[
index].quantization_param.is_negative_input = True
first_conv = False
else:
new_net.layer[index].quantization_param.scale_in[
si] = u8_max / new_net.layer[
index].quantization_param.scale_in[si]
else:
new_net.layer[index].quantization_param.scale_in[
si] = s8_max / new_net.layer[
index].quantization_param.scale_in[si]
new_net.layer[
index].quantization_param.is_negative_input = True
for si in range(
0, len(new_net.layer[index].quantization_param.scale_params)):
if not isclose(
new_net.layer[index].quantization_param.scale_params[si],
0.0):
new_scale_param = s8_max / new_net.layer[
index].quantization_param.scale_params[si]
if np.isinf(new_scale_param):
new_scale_param = 0.0
new_net.layer[index].quantization_param.scale_params[
si] = new_scale_param
else:
new_net.layer[index].quantization_param.scale_params[si] = 0.0
if conv_algo:
for conv_input in conv_inputs:
index_bottom_layer = find_index_by_name(
conv_input[1], convolution_layers)
for si in range(
0,
len(new_net.layer[index_bottom_layer].
quantization_param.scale_out)):
new_net.layer[
index_bottom_layer].quantization_param.scale_out[
si] = new_net.layer[
index].quantization_param.scale_in[si]
concat_layers = [(value, index) for index, value in enumerate(net.layer)
if value.type == 'Concat']
if len(concat_layers) > 0:
compiled_concat_layers = [
(value, index) for index, value in enumerate(compiled_net.layer)
if value.type == 'Concat'
]
concat_layers.reverse()
if unify_concat_scales:
for (l, index) in concat_layers:
index_in_compiled_net = find_index_by_name(
l.name, compiled_concat_layers)
assert (index_in_compiled_net >= 0)
conv_inputs = get_input_convolutions(l, compiled_net,
index_in_compiled_net,
["Convolution"],
["Concat"])
# TODO: support resonable cross-levels concat scale unify
min_concat_scale = sys.float_info.max
concat_input_indexes = []
for conv_input in conv_inputs:
index_in_net = find_index_by_name(conv_input[1],
convolution_layers)
assert (index_in_net >= 0)
concat_input_indexes.append(index_in_net)
if new_net.layer[index_in_net].quantization_param.scale_out[
0] < min_concat_scale:
min_concat_scale = new_net.layer[
index_in_net].quantization_param.scale_out[0]
for concat_input_index in concat_input_indexes:
new_net.layer[
concat_input_index].quantization_param.scale_out[
0] = min_concat_scale
else:
if concat_use_fp32:
for (l, index) in concat_layers:
index_in_compiled_net = find_index_by_name(
l.name, compiled_concat_layers)
assert (index_in_compiled_net >= 0)
conv_inputs = get_input_convolutions(
l, compiled_net, index_in_compiled_net,
["Convolution"])
for conv_input in conv_inputs:
index_in_net = find_index_by_name(
conv_input[1], convolution_layers)
assert (index_in_net >= 0)
new_net.layer[
index_in_net].quantization_param.bw_layer_out = 32
new_net.layer[
index_in_net].quantization_param.scale_out[:] = [
1.0
]
with open(model_path, 'w') as f:
f.write(str(new_net))
def run(self):
logger = logging.getLogger('dan.svdtool')
import caffe
import caffe.proto.caffe_pb2 as caffepb2
if self.ori_solver is None:
validate_status = self.validate_conf()
if not validate_status:
return False
new_solver = caffepb2.NetParameter()
new_solver.CopyFrom(self.ori_solver)
new_solver.ClearField('layer')
layer_index_dict = {}
layer_index = 0
# 构建第一个拆分的prototxt
for i in range(len(self.ori_solver.layer)):
layer = self.ori_solver.layer[i]
if layer.name in self.svd_spec_dict:
mid_layer_name = get_svd_layer_name(layer.name)
svd_bottom_layer = modify_message(
layer,
in_place=False,
**{
'top': [mid_layer_name],
'inner_product_param.bias_term': False,
'param': [layer.param[0]]
}
)
svd_mid_layer = modify_message(
layer,
in_place=False,
**{
'bottom': [mid_layer_name],
'name': mid_layer_name
}
)
new_solver.layer.extend([svd_bottom_layer, svd_mid_layer])
layer_index_dict[svd_bottom_layer.name] = layer_index
layer_index += 1
layer_index_dict[svd_mid_layer.name] = layer_index
else:
new_solver.layer.extend([layer])
layer_index_dict[layer.name] = layer_index
layer_index += 1
# 写入拆分后的proto
with open(self.output_proto, 'w') as output_proto_file:
logger.info('Writing temporary prototxt to "%s".', self._log_output_proto)
output_proto_file.write(text_format.MessageToString(new_solver))
# 构建新的net方便计算
new_net = caffe.Net(self.output_proto, caffe.TEST)
final_solver = caffepb2.NetParameter()
text_format.Merge(open(self.output_proto, 'r').read(), final_solver)
final_param_dict = {}
for layer_name, param in self.ori_net.params.iteritems():
if layer_name not in self.svd_spec_dict:
continue
svd_spec = self.svd_spec_dict[layer_name]
logger.info("Start calculating svd of layer < %s >. Strategy: %s. Argument < %s >: %s",
layer_name, svd_spec['method'].method_name,
svd_spec['method'].method_arg, str(svd_spec['argument']))
hide_layer_size, new_param_list = svd_spec['method'](svd_spec['argument'],
param[0].data, net=self.ori_net, new_net=new_net)
logger.info("Finish calculating svd of layer < %s >.", layer_name)
svd_hidelayer_name = get_svd_layer_name(layer_name)
# Store the final data
final_param_dict[layer_name] = (new_param_list[1],)
modify_message(
final_solver.layer[layer_index_dict[layer_name]],
in_place=True,
**{
'inner_product_param.num_output': hide_layer_size
}
)
# bias设置在后一层
final_param_dict[svd_hidelayer_name] = (new_param_list[0], param[1])
with open(self.output_proto, 'w') as output_proto_file:
logger.info('Writing proto to file "%s".', self._log_output_proto)
output_proto_file.write(text_format.MessageToString(final_solver))
new_net = caffe.Net(self.output_proto, caffe.TEST)
# USE THIS, as caffe will insert some layer such as split
# the `layer_index_dict` in the above code is of no use! TODO: remove those codes
layer_index_dict = {name: i for i, name in enumerate(new_net._layer_names)}
# 读入新的prototxt，然后对需要赋值blobs的layer都赋值，最后save
for layer_name, param in self.ori_net.params.iteritems():
if layer_name not in self.svd_spec_dict:
# 其它层的layer.blobs就等于原来的blobs
update_blob_vec(new_net.layers[layer_index_dict[layer_name]].blobs,
param)
else:
svd_hidelayer_name = get_svd_layer_name(layer_name)
update_blob_vec(new_net.layers[layer_index_dict[layer_name]].blobs,
final_param_dict[layer_name])
update_blob_vec(new_net.layers[layer_index_dict[svd_hidelayer_name]].blobs,
final_param_dict[svd_hidelayer_name])
logger.info('Writing caffe model to file "%s".', self._log_output_caffemodel)
new_net.save(self.output_caffemodel)
logger.info('Finish processing svd of fc layer! Prototxt in file "%s"'
'. Caffemodel in file "%s".\n', self._log_output_proto,
self._log_output_caffemodel,
extra={
'important': True
})
return True
def AFFine_OPT_Create_Caffemodel(original_prototxt_path, original_model_path,
optimized_prototxt_path, new_model_path):
net_param = caffe_pb2.NetParameter()
with open(original_prototxt_path, 'rt') as f:
Parse(f.read(), net_param)
param_layer_type_list = [layer.type for layer in net_param.layer]
param_layer_name_list = [layer.name for layer in net_param.layer]
target_layer_type = ['Convolution', 'InnerProduct']
merge_layer_type = ['Scale', 'BatchNorm']
caffe.set_mode_cpu()
net = caffe.Net(original_prototxt_path, original_model_path, caffe.TEST)
new_net = caffe.Net(optimized_prototxt_path, caffe.TEST)
for param_name in new_net.params.keys():
param_layer_idx = param_layer_name_list.index(param_name)
param_layer_type = param_layer_type_list[param_layer_idx]
if param_layer_type not in target_layer_type:
# OTHER LAYERS
for i in range(0, len(net.params[param_name])):
new_net.params[param_name][i].data[
...] = net.params[param_name][i].data[...]
else:
kernel_num = net.params[param_name][0].num
new_net.params[param_name][0].data[
...] = net.params[param_name][0].data[...]
if len(net.params[param_name]) == 2:
new_net.params[param_name][1].data[
...] = net.params[param_name][1].data[...]
#else:
# print new_net.params[param_name][1].data[...]
if param_layer_idx + 1 < len(param_layer_type_list):
for i in range(param_layer_idx + 1,
len(param_layer_type_list)):
# CHECK : CONV + BN +SCALE / CONV + BN / IP + ...
affine_layer_type = param_layer_type_list[i]
affine_layer_name = param_layer_name_list[i]
if affine_layer_type in merge_layer_type:
# MERGE BN/SCALE
if affine_layer_type == "Scale":
if len(net_param.layer[i].bottom) >= 2:
# NOT In-place Scale
try:
for j in range(
0,
len(net.params[affine_layer_name])
):
new_net.params[affine_layer_name][
j].data[...] = net.params[
affine_layer_name][j].data[...]
except:
# no parameter
break
else:
# In-place Scale
scale = net.params[affine_layer_name][0].data
if len(net.params[affine_layer_name]) == 2:
bias = net.params[affine_layer_name][
1].data
else:
bias = 0.0 * scale
for k in range(0, kernel_num):
new_net.params[param_name][0].data[
k] = new_net.params[param_name][
0].data[k] * scale[k]
new_net.params[param_name][1].data[
k] = new_net.params[param_name][
1].data[k] * scale[k] + bias[k]
elif affine_layer_type == "BatchNorm":
scale = net.params[affine_layer_name][2].data[0]
if scale != 0:
mean = net.params[affine_layer_name][
0].data / scale
std = np.sqrt(
net.params[affine_layer_name][1].data /
scale)
else:
mean = net.params[affine_layer_name][0].data
std = np.sqrt(
net.params[affine_layer_name][1].data)
for k in range(0, kernel_num):
new_net.params[param_name][0].data[
k] = new_net.params[param_name][0].data[
k] / std[k]
new_net.params[param_name][1].data[k] = (
new_net.params[param_name][1].data[k] -
mean[k]) / std[k]
else:
# TODO
assert (
1 > 2
), "## TODO ## : Other layers haven't been supported yet. ##"
else:
# NOT BN or SCALE, then BREAK
break
else:
# LAST LAYER, then BREAK
break
new_net.save(new_model_path)
print bcolors.OKGREEN + "BN SCALE OPT : Model at " + new_model_path + "." + bcolors.ENDC
def extract_cnn_features(img_filelist, img_root, prototxt, caffemodel, feat_name, output_path=None, output_type=None, caffe_path=None, mean_file=None, gpu_index=0, pool_size=8):
if caffe_path:
sys.path.append(os.path.join(caffe_path, 'python'))
import caffe
from caffe.proto import caffe_pb2
imagenet_mean = np.array([104, 117, 123])
if mean_file:
imagenet_mean = np.load(mean_file)
net_parameter = caffe_pb2.NetParameter()
text_format.Merge(open(prototxt, 'r').read(), net_parameter)
print("input dim",net_parameter.input_dim, len(net_parameter.input_dim),'input_shape',net_parameter.input_shape,len(net_parameter.input_shape))
if len(net_parameter.input_dim) != 0:
input_shape = net_parameter.input_dim
else:
input_shape = net_parameter.input_shape[0].dim
imagenet_mean = caffe.io.resize_image(imagenet_mean.transpose((1, 2, 0)), input_shape[2:]).transpose((2, 0, 1))
# INIT NETWORK
caffe.set_mode_gpu()
caffe.set_device(gpu_index)
net = caffe.Classifier(prototxt,caffemodel,
mean=imagenet_mean,
raw_scale=255,
channel_swap=(2, 1, 0))
img_filenames = [os.path.abspath(img_filelist)]
if img_filelist.endswith('.txt'):
img_filenames = [img_root+'/'+x.rstrip() for x in open(img_filelist, 'r')]
N = len(img_filenames)
if output_path != None and os.path.exists(output_path):
if os.path.isdir(output_path):
shutil.rmtree(output_path)
else:
os.remove(output_path)
## BATCH FORWARD
batch_size = int(net.blobs['data'].data.shape[0])
batch_num = int(math.ceil(N/float(batch_size)))
print('batch_num:', batch_num)
def compute_feat_array(batch_idx):
start_idx = batch_size * batch_idx
end_idx = min(batch_size * (batch_idx+1), N)
print(datetime.datetime.now().time(), '- batch: ', batch_idx, 'of', batch_num, 'idx range:[', start_idx, end_idx, ']')
input_data = []
for img_idx in range(start_idx, end_idx):
im = caffe.io.load_image(img_filenames[img_idx])
input_data.append(im)
while len(input_data) < batch_size:
input_data.append(input_data[0])
net.predict(input_data, oversample=False)
feat_array = net.blobs[feat_name].data
return feat_array
if output_type == None:
feat_list = []
for batch_idx in range(batch_num):
start_idx = batch_size * batch_idx
end_idx = min(batch_size * (batch_idx+1), N)
batch_count = end_idx - start_idx
feat_array = compute_feat_array(batch_idx)
for n in range(batch_count):
feat_list.append(feat_array[n, ...].copy())
return feat_list
elif output_type == 'txt':
with open(output_path, 'w') as output_file:
for batch_idx in range(batch_num):
start_idx = batch_size * batch_idx
end_idx = min(batch_size * (batch_idx+1), N)
batch_count = end_idx - start_idx
feat_array = compute_feat_array(batch_idx)
for n in range(batch_count):
output_file.write(' '.join([str(x) for x in feat_array[n, ...].flat])+'\n')
elif output_type == 'lmdb':
import lmdb
env = lmdb.open(output_path, map_size=int(1e12))
pool = Pool(pool_size)
for batch_idx in range(batch_num):
feat_array = compute_feat_array(batch_idx)
start_idx = batch_size * batch_idx
end_idx = min(batch_size * (batch_idx+1), N)
array4d_idx = [(feat_array, idx, idx+start_idx) for idx in range(end_idx-start_idx)]
datum_strings = pool.map(_array4d_idx_to_datum_string, array4d_idx)
with env.begin(write=True) as txn:
for idx in range(end_idx-start_idx):
txn.put('{:0>10d}'.format(start_idx+idx).encode('utf-8'), datum_strings[idx])
env.close();
def __init__(self):
self.net = pb.NetParameter()
def image_classification_model_create():
form = ImageClassificationModelForm()
form.dataset.choices = get_datasets()
form.standard_networks.choices = get_standard_networks()
form.standard_networks.default = get_default_standard_network()
form.previous_networks.choices = get_previous_networks()
prev_network_snapshots = get_previous_network_snapshots()
if not form.validate_on_submit():
return render_template(
'models/images/classification/new.html',
form=form,
previous_network_snapshots=prev_network_snapshots), 400
datasetJob = scheduler.get_job(form.dataset.data)
if not datasetJob:
return 'Unknown dataset job_id "%s"' % form.dataset.data, 500
job = None
try:
job = ImageClassificationModelJob(
name=form.model_name.data,
dataset_id=datasetJob.id(),
)
network = caffe_pb2.NetParameter()
pretrained_model = None
if form.method.data == 'standard':
found = False
networks_dir = os.path.join(os.path.dirname(digits.__file__),
'standard-networks')
for filename in os.listdir(networks_dir):
path = os.path.join(networks_dir, filename)
if os.path.isfile(path):
match = re.match(
r'%s.prototxt' % form.standard_networks.data, filename)
if match:
with open(path) as infile:
text_format.Merge(infile.read(), network)
found = True
break
if not found:
raise Exception('Unknown standard model "%s"' %
form.standard_networks.data)
elif form.method.data == 'previous':
old_job = scheduler.get_job(form.previous_networks.data)
if not old_job:
raise Exception('Job not found: %s' %
form.previous_networks.data)
network.CopyFrom(old_job.train_task().network)
for i, choice in enumerate(form.previous_networks.choices):
if choice[0] == form.previous_networks.data:
epoch = int(request.form['%s-snapshot' %
form.previous_networks.data])
if epoch != 0:
for filename, e in old_job.train_task().snapshots:
if e == epoch:
pretrained_model = filename
break
if pretrained_model is None:
raise Exception(
"For the job %s, selected pretrained_model for epoch %d is invalid!"
% (form.previous_networks.data, epoch))
if not (os.path.exists(pretrained_model)):
raise Exception(
"Pretrained_model for the selected epoch doesn't exists. May be deleted by another user/process. Please restart the server to load the correct pretrained_model details"
)
break
elif form.method.data == 'custom':
text_format.Merge(form.custom_network.data, network)
pretrained_model = form.custom_network_snapshot.data.strip()
else:
raise Exception('Unrecognized method: "%s"' % form.method.data)
policy = {'policy': form.lr_policy.data}
if form.lr_policy.data == 'fixed':
pass
elif form.lr_policy.data == 'step':
policy['stepsize'] = form.lr_step_size.data
policy['gamma'] = form.lr_step_gamma.data
elif form.lr_policy.data == 'multistep':
policy['stepvalue'] = form.lr_multistep_values.data
policy['gamma'] = form.lr_multistep_gamma.data
elif form.lr_policy.data == 'exp':
policy['gamma'] = form.lr_exp_gamma.data
elif form.lr_policy.data == 'inv':
policy['gamma'] = form.lr_inv_gamma.data
policy['power'] = form.lr_inv_power.data
elif form.lr_policy.data == 'poly':
policy['power'] = form.lr_poly_power.data
elif form.lr_policy.data == 'sigmoid':
policy['stepsize'] = form.lr_sigmoid_step.data
policy['gamma'] = form.lr_sigmoid_gamma.data
else:
return 'Invalid policy', 404
job.tasks.append(
tasks.CaffeTrainTask(
job_dir=job.dir(),
dataset=datasetJob,
train_epochs=form.train_epochs.data,
learning_rate=form.learning_rate.data,
lr_policy=policy,
batch_size=form.batch_size.data,
val_interval=form.val_interval.data,
pretrained_model=pretrained_model,
crop_size=form.crop_size.data,
use_mean=form.use_mean.data,
network=network,
))
scheduler.add_job(job)
return redirect(url_for('models_show', job_id=job.id()))
except:
if job:
scheduler.delete_job(job)
raise
def importPrototxt(request):
if request.method == 'POST':
try:
prototxt = request.FILES['file']
except Exception:
return JsonResponse({
'result': 'error',
'error': 'No Prototxt model file found'
})
caffe_net = caffe_pb2.NetParameter()
try:
text_format.Merge(prototxt.read(), caffe_net)
except Exception:
return JsonResponse({
'result': 'error',
'error': 'Invalid Prototxt'
})
net = {}
i = 0
blobMap = {}
net_name = caffe_net.name
for layer in caffe_net.layer:
id = "l" + str(i)
input = []
# this logic for phase has to be improved
if len(layer.include):
phase = layer.include[0].phase
else:
phase = None
params = {}
if (layer.type == 'Data'):
params['source'] = layer.data_param.source
params['batch_size'] = layer.data_param.batch_size
params['backend'] = layer.data_param.backend
params['scale'] = layer.transform_param.scale
elif (layer.type == 'Convolution'):
if len(layer.convolution_param.kernel_size):
params[
'kernel_h'] = layer.convolution_param.kernel_h or layer.convolution_param.kernel_size[
0]
params[
'kernel_w'] = layer.convolution_param.kernel_w or layer.convolution_param.kernel_size[
0]
if len(layer.convolution_param.pad):
params[
'pad_h'] = layer.convolution_param.pad_h or layer.convolution_param.pad[
0]
params[
'pad_w'] = layer.convolution_param.pad_w or layer.convolution_param.pad[
0]
if len(layer.convolution_param.stride):
params[
'stride_h'] = layer.convolution_param.stride_h or layer.convolution_param.stride[
0]
params[
'stride_w'] = layer.convolution_param.stride_w or layer.convolution_param.stride[
0]
params[
'weight_filler'] = layer.convolution_param.weight_filler.type
params[
'bias_filler'] = layer.convolution_param.bias_filler.type
params['num_output'] = layer.convolution_param.num_output
elif (layer.type == 'ReLU'):
if (layer.top == layer.bottom):
params['inplace'] = True
elif (layer.type == 'Pooling'):
params[
'pad_h'] = layer.pooling_param.pad_h or layer.pooling_param.pad
params[
'pad_w'] = layer.pooling_param.pad_w or layer.pooling_param.pad
params[
'stride_h'] = layer.pooling_param.stride_h or layer.pooling_param.stride
params[
'stride_w'] = layer.pooling_param.stride_w or layer.pooling_param.stride
params[
'kernel_h'] = layer.pooling_param.kernel_h or layer.pooling_param.kernel_size
params[
'kernel_w'] = layer.pooling_param.kernel_w or layer.pooling_param.kernel_size
params['pool'] = layer.pooling_param.pool
elif (layer.type == 'InnerProduct'):
params['num_output'] = layer.inner_product_param.num_output
params[
'weight_filler'] = layer.inner_product_param.weight_filler.type
params[
'bias_filler'] = layer.inner_product_param.bias_filler.type
elif (layer.type == 'SoftmaxWithLoss'):
pass
elif (layer.type == 'Accuracy'):
pass
elif (layer.type == 'Input'):
params['dim'] = str(map(int,
layer.input_param.shape[0].dim))[1:-1]
# string '64,1,28,28'
jsonLayer = {
'info': {
'type': layer.type,
'phase': phase
},
'connection': {
'input': [],
'output': []
},
'params': params
}
# this logic was written for a general scenerio(where train and test layers are mixed up)
# But as we know, the only differences between the train and test phase are:
# 1) input layer with different source in test
# 2) some accuracy layers in test
# If we consider these constraint, the below logic can be vastly reduced
for bottom_blob in layer.bottom:
if (bottom_blob != 'label'):
# if the current layer has a phase
# then only connect with layers of same phase
# if it has no phase then connect with all layers
if jsonLayer['info']['phase'] is not None:
phase = jsonLayer['info']['phase']
for bottomLayerId in blobMap[bottom_blob]:
if (net[bottomLayerId]['info']['phase'] == phase
) or (net[bottomLayerId]['info']['phase'] is
None):
input.append(bottomLayerId)
net[bottomLayerId]['connection'][
'output'].append(id)
else:
for bottomLayerId in blobMap[bottom_blob]:
input.append(bottomLayerId)
net[bottomLayerId]['connection']['output'].append(
id)
for top_blob in layer.top:
if (top_blob != 'label'):
if top_blob in blobMap:
if top_blob in layer.bottom:
# check for inplace operations
# layer has no phase
# then remove all layer history
# and add this one to the top
# layer has phase then remove all layers with same phase and append this
if jsonLayer['info']['phase'] is not None:
phase = jsonLayer['info']['phase']
for layerId in blobMap[bottom_blob]:
if net[layerId]['info']['phase'] == phase:
blobMap[bottom_blob].remove(layerId)
blobMap[top_blob].append(id)
else:
blobMap[top_blob] = [id]
else:
blobMap[top_blob].append(id)
else:
blobMap[top_blob] = [id]
jsonLayer['connection']['input'] = input
net[id] = jsonLayer
i = i + 1
return JsonResponse({
'result': 'success',
'net': net,
'net_name': net_name
})
action='store_true',
help='replace ReLU layers by ReLU6'
)
parser.add_argument(
'--tfpad',
action='store_true',
help='use tensorflow pad=SAME'
)
parser.add_argument(
'--nobn',
action='store_true',
help='do not use batch_norm, defualt is false'
)
parser.add_argument(
'-v','--version',
type=str,
default="2",
help='MobileNet version, 1|2'
)
parser.add_argument(
'-t','--type',
type=str,
default="ssdlite",
help='ssd type, ssd|ssdlite'
)
FLAGS, unparsed = parser.parse_known_args()
net_specs = caffe_pb2.NetParameter()
net = CaffeNetGenerator(net_specs)
net.generate(FLAGS)
print(text_format.MessageToString(net_specs, float_format=".5g"))
import sys
from sys import argv
from utils import *
from config import *
sys.path.append(pycaffe_path)
import caffe
from caffe.proto import caffe_pb2
from google.protobuf.text_format import Parse, MessageToString
print ">> start activation quantization ..."
caffe.set_mode_cpu()
net_param = caffe_pb2.NetParameter()
with open(act_int8_prototxt, 'rt') as f:
Parse(f.read(), net_param)
layer_num = len(net_param.layer)
relu_top = []
relu_layer_idx = []
for layer_idx in range(layer_num):
layer = net_param.layer[layer_idx]
if layer.type == 'ReLU':
relu_top.append(layer.top[0])
relu_layer_idx.append(layer_idx)
del net_param
def main(args):
# Set default output file names
if args.output_model is None:
file_name = osp.splitext(args.model)[0]
args.output_model = file_name + '_inference.prototxt'
if args.output_weights is None:
file_name = osp.splitext(args.weights)[0]
args.output_weights = file_name + '_inference.caffemodel'
with open(args.model) as f:
model = caffe_pb2.NetParameter()
pb.text_format.Parse(f.read(), model)
# Determince the BN layers to be absorbed or replaced
# Create the new layers
new_layers = []
absorbed, replaced = {}, {}
for i, layer in enumerate(model.layer):
if layer.type != 'BN':
new_layers.append(layer)
continue
assert len(layer.bottom) == 1
assert len(layer.top) == 1
bottom_blob = layer.bottom[0]
top_blob = layer.top[0]
# Check if can be absorbed. As there could be some inplace layers,
# for example, conv -> relu -> bn. In such case, the BN cannot be
# absorbed.
can_be_absorbed = False
for j in xrange(i - 1, -1, -1):
if bottom_blob in model.layer[j].top:
if model.layer[j].type not in ['Convolution', 'InnerProduct']:
can_be_absorbed = False
break
else:
can_be_absorbed = True
bottom_layer = model.layer[j]
if can_be_absorbed:
# Rename the blob in the top layers
for j in xrange(i + 1, len(model.layer)):
update_blob_name(model.layer[j].bottom, top_blob, bottom_blob)
update_blob_name(model.layer[j].top, top_blob, bottom_blob)
if bottom_layer.type == 'Convolution':
bottom_layer.convolution_param.bias_term = True
elif bottom_layer.type == 'InnerProduct':
bottom_layer.inner_product_param.bias_term = True
absorbed[layer.name] = bottom_layer.name
elif args.replace_by == 'affine':
# Replace by an scale bias layer
new_layer = caffe_pb2.LayerParameter()
new_layer.name = layer.name + '_affine'
new_layer.type = 'Scale'
new_layer.bottom.extend([bottom_blob])
new_layer.top.extend([top_blob])
new_layer.scale_param.bias_term = True
replaced[layer.name] = new_layer.name
new_layers.append(new_layer)
elif args.replace_by == 'frozen':
# Freeze the BN layer
layer.bn_param.frozen = True
del (layer.param[:])
param = caffe_pb2.ParamSpec()
param.lr_mult = 0
param.decay_mult = 0
layer.param.extend([param] * 2)
new_layers.append(layer)
# Save the prototxt
output_model = caffe_pb2.NetParameter()
output_model.CopyFrom(model)
del (output_model.layer[:])
output_model.layer.extend(new_layers)
with open(args.output_model, 'w') as f:
f.write(pb.text_format.MessageToString(output_model))
# Copy the parameters
weights = caffe.Net(args.model, args.weights, caffe.TEST)
output_weights = caffe.Net(args.output_model, caffe.TEST)
for name in np.intersect1d(weights.params.keys(),
output_weights.params.keys()):
# Some original conv / inner product layers do not have bias_term
for i in xrange(
min(len(weights.params[name]),
len(output_weights.params[name]))):
output_weights.params[name][i].data[...] = \
weights.params[name][i].data.copy()
# Absorb the BN parameters
for old, new in absorbed.iteritems():
scale, bias, mean, tmp = [p.data.ravel() for p in weights.params[old]]
invstd = tmp if args.bn_style == 'invstd' else \
np.power(tmp + args.epsilon, -0.5)
W, b = output_weights.params[new]
assert W.data.ndim == 4 or W.data.ndim == 2
assert b.data.ndim == 1
if W.data.ndim == 4:
W.data[...] = (W.data * scale[:, None, None, None] *
invstd[:, None, None, None])
elif W.data.ndim == 2:
W.data[...] = W.data * scale[:, None] * invstd[:, None]
b.data[...] = (b.data[...] - mean) * scale * invstd + bias
# Fill up the affine layers
for old, new in replaced.iteritems():
scale, bias, mean, tmp = [p.data.ravel() for p in weights.params[old]]
invstd = tmp if args.bn_style == 'invstd' else \
np.power(tmp + args.epsilon, -0.5)
W, b = output_weights.params[new]
assert W.data.ndim == 1
assert b.data.ndim == 1
W.data[...] = scale * invstd
b.data[...] = bias - scale * mean * invstd
# Check if the conversion is correct
check(weights, output_weights)
# Save the caffemodel
output_weights.save(args.output_weights)
def main():
args = parse_args()
net = caffe_pb2.NetParameter()
text_format.Merge(open(args.input_net_proto_file).read(), net)
print('Drawing net to %s' % args.output_image_file)
caffe.draw.draw_net_to_file(net, args.output_image_file, args.rankdir)
def convert(prototxt: str, caffemodel: str, dest: str = 'nnmodel.daq') -> None:
assert isinstance(prototxt, str) and isinstance(
caffemodel, str), 'prototxt and caffemodel shoule be filename'
skipped_layers = []
params = caffe_pb2.NetParameter()
with open(prototxt) as f:
text_format.Merge(f.read(), params)
net = caffe.Net(prototxt, caffemodel, caffe.TEST)
out_filename = dest
with open(out_filename, 'wb') as f:
model_writer = ModelWriter(f)
for i, layer in enumerate(params.layer):
if layer.type not in SUPPORTED_LAYERS:
raise ValueError("Not supported layer " + layer.type)
if layer.name in skipped_layers:
continue
top_name = layer.top[0]
if i == 0:
if layer.type != "Input":
raise ValueError("First layer should be input")
param = layer.input_param
model_writer.add_input(top_name, param.shape[0].dim)
elif layer.type == 'Convolution':
bottom_name = layer.bottom[0]
param = layer.convolution_param
if len(param.pad) == 0:
pad_left = pad_right = pad_top = pad_bottom = 0
elif len(param.pad) == 1:
pad_left = pad_right = pad_top = pad_bottom = param.pad[0]
elif len(param.pad) == 2:
pad_top, pad_bottom, pad_left, pad_right = param.pad[
0], param.pad[0], param.pad[1], param.pad[1]
else:
raise ValueError("Only conv 2d is supported.")
if param.pad_h != 0:
pad_top = pad_bottom = param.pad_h
if param.pad_w != 0:
pad_left = pad_right = param.pad_w
if len(param.stride) == 0:
stride_x = stride_y = 1
elif len(param.stride) == 1:
stride_x = stride_y = param.stride[0]
elif len(param.stride) == 2:
stride_y, stride_x = param.stride[0], param.stride[1]
else:
raise ValueError("Only conv 2d is supported.")
if param.stride_h != 0:
stride_y = param.stride_h
if param.stride_w != 0:
stride_x = param.stride_w
if len(param.kernel_size) == 1:
filter_height = filter_width = param.kernel_size[0]
elif len(param.kernel_size) == 2:
filter_height, filter_width = param.kernel_size[
0], param.kernel_size[1]
else:
raise ValueError("Only conv 2d is supported.")
if param.kernel_h != 0:
filter_height = param.kernel_h
if param.kernel_w != 0:
filter_width = param.kernel_w
if len(param.dilation) == 0:
dilation = 1
elif len(param.dilation) == 1:
dilation = param.dilation[0]
else:
raise ValueError("Only dilation == 1 is supported.")
if dilation != 1:
raise ValueError(
"Dilation == {}, only dilation == 1 is supported.".
format(dilation))
axis = param.axis
if axis != 1:
raise ValueError("Only axis == 1 is supported.")
num_output = param.num_output
weights = net.params[layer.name][
0].data # shape: [depth_out, depth_in, filter_height, filter_width]
bias = net.params[layer.name][
1].data if param.bias_term else None # np.zeros(swapped_weights.shape[0])
activation = find_inplace_activation(params, top_name,
skipped_layers)
group = param.group
model_writer.add_conv(bottom_name, top_name, pad_left,
pad_right, pad_top, pad_bottom, stride_x,
stride_y, dilation, group, filter_height,
filter_width, num_output, activation,
weights, bias)
elif layer.type == 'Pooling':
param = layer.pooling_param
bottom_name = layer.bottom[0]
pad = param.pad
pad_left = pad_right = pad_top = pad_bottom = pad
if param.pad_h != 0:
pad_top = pad_bottom = param.pad_h
if param.pad_w != 0:
pad_left = pad_right = param.pad_w
stride = param.stride
stride_x = stride_y = stride
if param.stride_h != 0:
stride_y = param.stride_h
if param.stride_w != 0:
stride_x = param.stride_w
kernel_size = param.kernel_size
filter_height = filter_width = kernel_size
if param.kernel_h != 0:
filter_height = param.kernel_h
if param.kernel_w != 0:
filter_width = param.kernel_w
if param.global_pooling:
filter_height, filter_width = -1, -1
activation = find_inplace_activation(params, top_name,
skipped_layers)
if param.pool == CAFFE_POOL_MAX:
model_writer.add_max_pool(bottom_name, top_name, pad_left,
pad_right, pad_top, pad_bottom,
stride_x, stride_y,
filter_height, filter_width,
activation)
elif param.pool == CAFFE_POOL_AVE:
model_writer.add_ave_pool(bottom_name, top_name, pad_left,
pad_right, pad_top, pad_bottom,
stride_x, stride_y,
filter_height, filter_width,
activation)
else:
raise ValueError("Not supported pool type")
elif layer.type == 'InnerProduct':
bottom_name = layer.bottom[0]
param = layer.inner_product_param
input_dim = list(net.blobs[bottom_name].data.shape)
weights = net.params[layer.name][0].data
num_output = param.num_output
if param.axis != 1:
raise ValueError(
"Only inner_product.axis == 1 is supported.")
if param.transpose:
raise ValueError(
"Only inner_product.transpose == True is supported")
if len(input_dim) == 4:
input_dim[0] = param.num_output
weights = weights.reshape(input_dim)
weights = np.moveaxis(weights, 1, 3)
bias = net.params[layer.name][
1].data if param.bias_term else None # np.zeros(num_output)
activation = find_inplace_activation(params, top_name,
skipped_layers)
model_writer.add_FC(bottom_name, top_name, num_output,
activation, weights, bias)
elif layer.type == 'ReLU':
bottom_name = layer.bottom[0]
param = layer.relu_param
model_writer.add_ReLU(bottom_name, top_name,
param.negative_slope)
elif layer.type == 'Softmax':
bottom_name = layer.bottom[0]
param = layer.softmax_param
if param.axis != 1:
raise ValueError("Only softmax.axis == 1 is supported.")
model_writer.add_softmax(bottom_name, top_name, 1.)
elif layer.type == 'Dropout':
pass
elif layer.type == 'LRN':
bottom_name = layer.bottom[0]
param = layer.lrn_param
local_size = param.local_size
alpha = param.alpha
beta = param.beta
model_writer.add_LRN(bottom_name, top_name, local_size, alpha,
beta)
elif layer.type == 'Eltwise':
bottom0 = layer.bottom[0]
bottom1 = layer.bottom[1]
param = layer.eltwise_param
if param.operation == CAFFE_ELTWISE_SUM:
if np.count_nonzero(np.array(param.coeff) != 1) > 0:
raise ValueError(
"Only all coefficients in sum == 1 is supported.")
model_writer.add_add(bottom0, mw.TENSOR_OP, bottom1,
top_name)
elif param.operation == CAFFE_ELTWISE_PROD:
model_writer.add_mul(bottom0, mw.TENSOR_OP, bottom1,
top_name)
else:
raise ValueError("Unsupported EltwiseOp " +
str(param.operation))
elif layer.type == 'BatchNorm':
bottom_name = layer.bottom[0]
param = layer.batch_norm_param
if not param.use_global_stat:
raise ValueError(
"Only batch_norm.use_global_stat is true is supported. "
"(Did you load model in train phase?)")
scale_factor = net.params[layer.name][2].data[0]
mean = net.params[layer.name][0].data / scale_factor
var = net.params[layer.name][1].data / scale_factor + param.eps
model_writer.add_add(bottom_name, mw.ARRAY_OP, -mean, top_name)
# Append top into blobs so that the mul will use a new index as input
# It will be the index of output blob of add
model_writer.add_mul(top_name, mw.ARRAY_OP, 1 / np.sqrt(var),
top_name)
elif layer.type == 'Scale':
if len(layer.bottom) != 1:
raise ValueError(
"Only a learnable Scale layer is supported.")
bottom_name = layer.bottom[0]
param = layer.scale_param
if param.num_axes != 1:
raise ValueError("Only scale.num_axes == 2 is supported.")
multiplier = net.params[layer.name][0].data
model_writer.add_mul(bottom_name, mw.ARRAY_OP, multiplier,
top_name)
if param.bias_term:
bias = net.params[layer.name][1].data
model_writer.add_add(top_name, mw.ARRAY_OP, bias, top_name)
elif layer.type == 'Concat':
param = layer.concat_param
model_writer.add_concat(layer.bottom, top_name, param.axis)
elif layer.type == 'Power':
bottom_name = layer.bottom[0]
param = layer.power_param
power, scale, shift = param.power, param.scale, param.shift
internal_bottom_name = bottom_name
if scale != 1:
model_writer.add_mul(internal_bottom_name, mw.SCALAR_OP,
scale, top_name)
internal_bottom_name = top_name
if shift != 0:
model_writer.add_add(internal_bottom_name, mw.SCALAR_OP,
shift, top_name)
internal_bottom_name = top_name
if power != 1:
raise ValueError('Only power == 1 is supported')
model_writer.save()
# In[ ]:
# reading weights by using net.params
nice_edge_detectors = net.params['conv1_1']
higher_level_filter = net.params['fc7']
# In[ ]:
from google.protobuf import text_format
from caffe.draw import get_pydot_graph
from caffe.proto import caffe_pb2
from IPython.display import display, Image
_net = caffe_pb2.NetParameter()
f = open("gender_siamese_train_2.prototxt")
text_format.Merge(f.read(), _net)
display(Image(get_pydot_graph(_net,"TB").create_png()))
# In[ ]:
# not works
# http://research.beenfrog.com/code/2015/03/28/read-leveldb-lmdb-for-caffe-with-python.html
def get_data_for_case_from_lmdb(lmdb_name, id):
lmdb_env = lmdb.open(lmdb_name, readonly=True)
lmdb_txn = lmdb_env.begin()
raw_datum = lmdb_txn.get(id)
datum = caffe.proto.caffe_pb2.Datum()
def main():
args = parse_args()
if not args.verbose:
os.environ['GLOG_minloglevel'] = '2'
import caffe
from caffe.proto import caffe_pb2
if args.cpu:
caffe.set_mode_cpu()
else:
caffe.set_mode_gpu()
if args.verbose:
if args.cpu:
print("CPU mode set.")
else:
print("GPU mode set.")
errorSum = np.zeros(2)
for i in range(args.tests_number):
tmpNetProto = tempfile.NamedTemporaryFile()
tmpNetProto.write(createPoolingNet(args.params))
tmpNetProto.flush()
net = caffe.Net(tmpNetProto.name, caffe.TEST)
deploy = caffe_pb2.NetParameter()
Merge((open(tmpNetProto.name, 'r').read()), deploy)
tmpNetProto.close()
sys.stdout.write("{}. ".format(i + 1))
if not args.verbose:
sys.stdout.write("Input shape: {}, {},{},{} ".format(
net.blobs['data'].data.shape[0],
net.blobs['data'].data.shape[2],
net.blobs['data'].data.shape[3],
net.blobs['data'].data.shape[1]))
poolingParams = deploy.layer[1].pooling_param
sys.stdout.write("Pooling params: {},{} ".format(
poolingParams.kernel_size, poolingParams.stride))
sys.stdout.write("Output shape: {},{},{}".format(
net.blobs['pooling'].data.shape[2],
net.blobs['pooling'].data.shape[3],
net.blobs['pooling'].data.shape[1]))
net.blobs['data'].data[...] = np.random.random_sample(
net.blobs['data'].data.shape) - 0.5
net.blobs['pooling'].diff[...] = np.random.random_sample(
net.blobs['pooling'].diff.shape) - 0.5
errorSum += comparePooling(net,
deploy,
args.forward,
args.backward,
args.convbase_executable,
outputPreffix="pooling_",
verbose=args.verbose)
meanError = errorSum / args.tests_number
print("\n#############################################################")
print("Number of tests: {}\n".format(args.tests_number))
if args.forward:
print("Mean forward error: {}".format(meanError[0]))
if args.backward:
print("Mean backward error: {}".format(meanError[1]))
print("#############################################################")
parser.add_argument("--predict_net",
help="Caffe2 prediction net.",
default="predict_net.pb")
parser.add_argument("--remove_legacy_pad",
help="Remove legacy pad \
(Only works for nets with one input blob)",
action="store_true",
default=False)
parser.add_argument("--input_dims",
help="Dimension of input blob",
nargs='+',
type=int,
default=[])
args = parser.parse_args()
caffenet = caffe_pb2.NetParameter()
caffenet_pretrained = caffe_pb2.NetParameter()
input_proto = args.prototext
input_caffemodel = args.caffemodel
output_init_net = args.init_net
output_predict_net = args.predict_net
text_format.Merge(open(input_proto, 'r').read(), caffenet)
caffenet_pretrained.ParseFromString(open(input_caffemodel, 'rb').read())
#for l in caffenet_pretrained.layer:
# log.info(" the pretrained layer is {}".format(l))
net, pretrained_params = TranslateModel(
caffenet,
caffenet_pretrained,
is_test=True,
remove_legacy_pad=args.remove_legacy_pad,
metavar='image-params',
help='The parameters for image preprocess')
parser.add_argument('--ignore_layer_list',
metavar='ignore-layer-list',
help='Ignore these layers because they already tuned')
args = parser.parse_args()
caffe.set_mode_gpu()
caffe.set_device(0)
# Call it when you are using yolo
# caffe.yolo()
if not os.path.isdir(args.output_path):
os.mkdir(args.output_path)
param = caffe_pb2.NetParameter()
text_format.Merge(open(args.proto).read(), param)
first = True
for layer in param.layer:
if 'data' == layer.name:
continue
#if 'BatchNorm' == layer.type:
if layer.type in [
'BatchNorm', 'Data', 'Input', 'Python', 'Softmax', 'PriorBox',
'Reshape', 'Flatten', 'Reshape', 'DetectionOutput'
]:
continue
if layer.name in args.ignore_layer_list:
continue
best_model, best_proto = run_tune(args, layer.name, sys.float_info.max)
def import_prototxt(request):
prototxtIsText = False
if request.method == 'POST':
if ('file' in request.FILES) and \
(request.FILES['file'].content_type == 'application/octet-stream' or
request.FILES['file'].content_type == 'text/plain'):
try:
prototxt = request.FILES['file']
except Exception:
return JsonResponse({
'result': 'error',
'error': 'No Prototxt model file found'
})
elif 'sample_id' in request.POST:
try:
prototxt = open(
os.path.join(settings.BASE_DIR, 'example', 'caffe',
request.POST['sample_id'] + '.prototxt'), 'r')
except Exception:
return JsonResponse({
'result': 'error',
'error': 'No Prototxt model file found'
})
elif 'config' in request.POST:
prototxt = request.POST['config']
prototxtIsText = True
elif 'url' in request.POST:
try:
url = urlparse(request.POST['url'])
if url.netloc == 'github.com':
url = url._replace(netloc='raw.githubusercontent.com')
url = url._replace(path=url.path.replace('blob/', ''))
prototxt = urllib2.urlopen(url.geturl())
except Exception as ex:
return JsonResponse({
'result': 'error',
'error': 'Invalid URL\n' + str(ex)
})
caffe_net = caffe_pb2.NetParameter()
# try to convert to new prototxt
try:
if prototxtIsText is True:
content = prototxt
else:
content = prototxt.read()
tempFile = tempfile.NamedTemporaryFile()
tempFile.write(content)
tempFile.seek(0)
subprocess.call(
"~/caffe/caffe/build/tools/upgrade_net_proto_text " +
tempFile.name + " " + tempFile.name,
shell=True)
tempFile.seek(0)
content = tempFile.read()
tempFile.close()
except Exception as ex:
return JsonResponse({
'result': 'error',
'error': 'Invalid Prototxt\n' + str(ex)
})
try:
text_format.Merge(content, caffe_net)
except Exception as ex:
return JsonResponse({
'result': 'error',
'error': 'Invalid Prototxt\n' + str(ex)
})
net = {}
i = 0
blobMap = {}
net_name = caffe_net.name
hasTransformParam = ['ImageData', 'Data', 'WindowData']
for layer in caffe_net.layer:
id = "l" + str(i)
input = []
# this logic for phase has to be improved
if len(layer.include):
if (layer.include[0].HasField('phase')):
phase = layer.include[0].phase
else:
phase = None
else:
phase = None
params = {}
if (layer.type in hasTransformParam):
params['scale'] = layer.transform_param.scale
params['mirror'] = layer.transform_param.mirror
params['crop_size'] = layer.transform_param.crop_size
if (layer.transform_param.mean_file != ''):
params['mean_file'] = layer.transform_param.mean_file
elif (layer.transform_param.mean_value):
params['mean_value'] = str(
map(int, layer.transform_param.mean_value))[1:-1]
params['force_color'] = layer.transform_param.force_color
params['force_gray'] = layer.transform_param.force_gray
if layer.type in layer_dict:
layer_params = layer_dict[layer.type](layer)
params.update(layer_params)
jsonLayer = {
'info': {
'type': layer.type,
'phase': phase
},
'connection': {
'input': [],
'output': []
},
'params': params
}
# this logic was written for a scenario where train and test layers are mixed up
# But as we know, the only differences between the train and test phase are:
# 1) input layer with different source in test
# 2) some accuracy layers in test
# If we consider these constraint, the below logic can be vastly reduced
for bottom_blob in layer.bottom:
if (bottom_blob != 'label'):
# if the current layer has a phase
# then only connect with layers of same phase
# if it has no phase then connect with all layers
if jsonLayer['info']['phase'] is not None:
phase = jsonLayer['info']['phase']
for bottomLayerId in blobMap[bottom_blob]:
if (net[bottomLayerId]['info']['phase'] == phase) or\
(net[bottomLayerId]['info']['phase'] is None):
input.append(bottomLayerId)
net[bottomLayerId]['connection'][
'output'].append(id)
else:
for bottomLayerId in blobMap[bottom_blob]:
input.append(bottomLayerId)
net[bottomLayerId]['connection']['output'].append(
id)
for top_blob in layer.top:
if (top_blob != 'label'):
if top_blob in blobMap:
if top_blob in layer.bottom:
# check for in-place operations
# layer has no phase
# then remove all layer history
# and add this one to the top
# layer has phase then remove all layers with same phase and append this
if jsonLayer['info']['phase'] is not None:
phase = jsonLayer['info']['phase']
for layerId in blobMap[bottom_blob]:
if net[layerId]['info']['phase'] == phase:
blobMap[bottom_blob].remove(layerId)
blobMap[top_blob].append(id)
else:
blobMap[top_blob] = [id]
else:
blobMap[top_blob].append(id)
else:
blobMap[top_blob] = [id]
jsonLayer['connection']['input'] = input
net[id] = jsonLayer
i = i + 1
return JsonResponse({
'result': 'success',
'net': net,
'net_name': net_name
})
def read_network_dag(processed_deploy_prototxt):
"""
Reads from the caffe prototxt the network structure
:param processed_deploy_prototxt: name of prototxt to load, preferably the prototxt should
be processed before using a call to process_network_proto()
:return: network_def, layer_name_to_record, top_to_layers
network_def: caffe network structure, gives access to *all* the network information
layer_name_to_record: *ordered* dictionary which maps between layer name and a structure which
describes in a simple form the layer parameters
top_to_layers: dictionary which maps a blob name to an ordered list of layers which output it
when a top is used several times, like in inplace layhers, the list will contain all the layers
by order of appearance
"""
from caffe.proto import caffe_pb2
from google.protobuf import text_format # pylint: disable=relative-import
from collections import OrderedDict
# load prototxt file
network_def = caffe_pb2.NetParameter()
with open(processed_deploy_prototxt, 'r') as proto_file:
text_format.Merge(str(proto_file.read()), network_def)
# map layer name to layer record
layer_name_to_record = OrderedDict()
for layer_def in network_def.layer:
if (len(layer_def.include) == 0) or \
(caffe_pb2.TEST in [item.phase for item in layer_def.include]):
layer_name_to_record[layer_def.name] = LayerRecord(layer_def)
top_to_layers = dict()
for layer in network_def.layer:
# no specific phase, or TEST phase is specifically asked for
if (len(layer.include) == 0) or (caffe_pb2.TEST in [
item.phase for item in layer.include
]):
for top in layer.top:
if top not in top_to_layers:
top_to_layers[top] = list()
top_to_layers[top].append(layer.name)
# find parents and children of all layers
for child_layer_name in layer_name_to_record.keys(): # pylint: disable=too-many-nested-blocks
child_layer_def = layer_name_to_record[child_layer_name]
for bottom in child_layer_def.bottoms:
if bottom in top_to_layers:
for parent_layer_name in top_to_layers[bottom]:
if parent_layer_name in layer_name_to_record:
parent_layer_def = layer_name_to_record[
parent_layer_name]
if parent_layer_def not in child_layer_def.parents:
child_layer_def.parents.append(parent_layer_def)
if child_layer_def not in parent_layer_def.children:
parent_layer_def.children.append(child_layer_def)
# update filter, strid, pad for maxout "structures"
for layer_name in layer_name_to_record.keys():
layer_def = layer_name_to_record[layer_name]
if layer_def.type == 'Eltwise' and \
len(layer_def.parents) == 1 and \
layer_def.parents[0].type == 'Slice' and \
len(layer_def.parents[0].parents) == 1 and \
layer_def.parents[0].parents[0].type in ['Convolution', 'InnerProduct']:
layer_def.filter = layer_def.parents[0].parents[0].filter
layer_def.stride = layer_def.parents[0].parents[0].stride
layer_def.pad = layer_def.parents[0].parents[0].pad
return network_def, layer_name_to_record, top_to_layers
def save_prototxt_files(self):
"""
Save solver, train_val and deploy files to disk
"""
has_val_set = self.dataset.val_db_task() is not None
### Check what has been specified in self.network
tops = []
bottoms = {}
train_data_layer = None
val_data_layer = None
hidden_layers = caffe_pb2.NetParameter()
loss_layer = None
accuracy_layer = None
for layer in self.network.layer:
assert layer.type not in ['MemoryData', 'HDF5Data', 'ImageData'], 'unsupported data layer type'
if layer.type == 'Data':
for rule in layer.include:
if rule.phase == caffe_pb2.TRAIN:
assert train_data_layer is None, 'cannot specify two train data layers'
train_data_layer = layer
elif rule.phase == caffe_pb2.TEST:
assert val_data_layer is None, 'cannot specify two test data layers'
val_data_layer = layer
elif layer.type == 'SoftmaxWithLoss':
assert loss_layer is None, 'cannot specify two loss layers'
loss_layer = layer
elif layer.type == 'Accuracy':
assert accuracy_layer is None, 'cannot specify two accuracy layers'
accuracy_layer = layer
else:
hidden_layers.layer.add().CopyFrom(layer)
if len(layer.bottom) == 1 and len(layer.top) == 1 and layer.bottom[0] == layer.top[0]:
pass
else:
for top in layer.top:
tops.append(top)
for bottom in layer.bottom:
bottoms[bottom] = True
assert loss_layer is not None, 'must specify a SoftmaxWithLoss layer'
assert accuracy_layer is not None, 'must specify an Accuracy layer'
if not has_val_set:
self.logger.warning('Discarding Data layer for validation')
val_data_layer = None
output_name = None
for name in tops:
if name not in bottoms:
assert output_name is None, 'network cannot have more than one output'
output_name = name
assert output_name is not None, 'network must have one output'
for layer in hidden_layers.layer:
if output_name in layer.top and layer.type == 'InnerProduct':
layer.inner_product_param.num_output = len(self.labels)
break
if train_data_layer is None:
assert val_data_layer is None, 'cannot specify a test data layer without a train data layer'
### Write train_val file
train_val_network = caffe_pb2.NetParameter()
default_batch_size = 16 #XXX Reasonable default?
# data layers
if train_data_layer is not None:
if train_data_layer.HasField('data_param'):
assert not train_data_layer.data_param.HasField('source'), "don't set the data_param.source"
assert not train_data_layer.data_param.HasField('backend'), "don't set the data_param.backend"
max_crop_size = min(self.dataset.image_dims[0], self.dataset.image_dims[1])
if self.crop_size:
assert self.crop_size <= max_crop_size, 'crop_size is larger than the image size'
train_data_layer.transform_param.crop_size = self.crop_size
elif train_data_layer.transform_param.HasField('crop_size'):
cs = train_data_layer.transform_param.crop_size
if cs > max_crop_size:
# don't throw an error here
cs = max_crop_size
train_data_layer.transform_param.crop_size = cs
self.crop_size = cs
train_val_network.layer.add().CopyFrom(train_data_layer)
train_data_layer = train_val_network.layer[-1]
if val_data_layer is not None and has_val_set:
if val_data_layer.HasField('data_param'):
assert not val_data_layer.data_param.HasField('source'), "don't set the data_param.source"
assert not val_data_layer.data_param.HasField('backend'), "don't set the data_param.backend"
if self.crop_size:
# use our error checking from the train layer
val_data_layer.transform_param.crop_size = self.crop_size
train_val_network.layer.add().CopyFrom(val_data_layer)
val_data_layer = train_val_network.layer[-1]
else:
train_data_layer = train_val_network.layer.add(type = 'Data', name = 'data')
train_data_layer.top.append('data')
train_data_layer.top.append('label')
train_data_layer.include.add(phase = caffe_pb2.TRAIN)
train_data_layer.data_param.batch_size = default_batch_size
if self.crop_size:
train_data_layer.transform_param.crop_size = self.crop_size
if has_val_set:
val_data_layer = train_val_network.layer.add(type = 'Data', name = 'data')
val_data_layer.top.append('data')
val_data_layer.top.append('label')
val_data_layer.include.add(phase = caffe_pb2.TEST)
val_data_layer.data_param.batch_size = default_batch_size
if self.crop_size:
val_data_layer.transform_param.crop_size = self.crop_size
train_data_layer.data_param.source = self.dataset.path(self.dataset.train_db_task().db_name)
train_data_layer.data_param.backend = caffe_pb2.DataParameter.LMDB
if val_data_layer is not None:
val_data_layer.data_param.source = self.dataset.path(self.dataset.val_db_task().db_name)
val_data_layer.data_param.backend = caffe_pb2.DataParameter.LMDB
if self.use_mean:
train_data_layer.transform_param.mean_file = self.dataset.path(self.dataset.train_db_task().mean_file)
if val_data_layer is not None:
val_data_layer.transform_param.mean_file = self.dataset.path(self.dataset.train_db_task().mean_file)
if self.batch_size:
train_data_layer.data_param.batch_size = self.batch_size
if val_data_layer is not None:
val_data_layer.data_param.batch_size = self.batch_size
else:
if not train_data_layer.data_param.HasField('batch_size'):
train_data_layer.data_param.batch_size = default_batch_size
if val_data_layer is not None and not val_data_layer.data_param.HasField('batch_size'):
val_data_layer.data_param.batch_size = default_batch_size
# hidden layers
train_val_network.MergeFrom(hidden_layers)
# output layers
if loss_layer is not None:
train_val_network.layer.add().CopyFrom(loss_layer)
loss_layer = train_val_network.layer[-1]
else:
loss_layer = train_val_network.layer.add(
type = 'SoftmaxWithLoss',
name = 'loss')
loss_layer.bottom.append(output_name)
loss_layer.bottom.append('label')
loss_layer.top.append('loss')
if accuracy_layer is not None:
train_val_network.layer.add().CopyFrom(accuracy_layer)
accuracy_layer = train_val_network.layer[-1]
elif self.dataset.val_db_task():
accuracy_layer = train_val_network.layer.add(
type = 'Accuracy',
name = 'accuracy')
accuracy_layer.bottom.append(output_name)
accuracy_layer.bottom.append('label')
accuracy_layer.top.append('accuracy')
accuracy_layer.include.add(phase = caffe_pb2.TEST)
with open(self.path(self.train_val_file), 'w') as outfile:
text_format.PrintMessage(train_val_network, outfile)
### Write deploy file
deploy_network = caffe_pb2.NetParameter()
# input
deploy_network.input.append('data')
deploy_network.input_dim.append(1)
deploy_network.input_dim.append(self.dataset.image_dims[2])
if self.crop_size:
deploy_network.input_dim.append(self.crop_size)
deploy_network.input_dim.append(self.crop_size)
else:
deploy_network.input_dim.append(self.dataset.image_dims[0])
deploy_network.input_dim.append(self.dataset.image_dims[1])
# hidden layers
deploy_network.MergeFrom(hidden_layers)
# output layers
prob_layer = deploy_network.layer.add(
type = 'Softmax',
name = 'prob')
prob_layer.bottom.append(output_name)
prob_layer.top.append('prob')
with open(self.path(self.deploy_file), 'w') as outfile:
text_format.PrintMessage(deploy_network, outfile)
### Write solver file
solver = caffe_pb2.SolverParameter()
solver.net = self.train_val_file
if config_option('gpu_list'):
solver.solver_mode = caffe_pb2.SolverParameter.GPU
else:
solver.solver_mode = caffe_pb2.SolverParameter.CPU
solver.snapshot_prefix = self.snapshot_prefix
# Epochs -> Iterations
train_iter = int(math.ceil(float(self.dataset.train_db_task().entries_count) / train_data_layer.data_param.batch_size))
solver.max_iter = train_iter * self.train_epochs
solver.snapshot = train_iter
if self.dataset.val_db_task() and self.val_interval:
solver.test_iter.append(int(math.ceil(float(self.dataset.val_db_task().entries_count) / val_data_layer.data_param.batch_size)))
solver.test_interval = train_iter * self.val_interval
# Learning rate
solver.base_lr = self.learning_rate
solver.lr_policy = self.lr_policy['policy']
scale = float(solver.max_iter)/100.0
if solver.lr_policy == 'fixed':
pass
elif solver.lr_policy == 'step':
# stepsize = stepsize * scale
solver.stepsize = int(math.ceil(float(self.lr_policy['stepsize']) * scale))
solver.gamma = self.lr_policy['gamma']
elif solver.lr_policy == 'multistep':
for value in self.lr_policy['stepvalue']:
# stepvalue = stepvalue * scale
solver.stepvalue.append(int(math.ceil(float(value) * scale)))
solver.gamma = self.lr_policy['gamma']
elif solver.lr_policy == 'exp':
# gamma = gamma^(1/scale)
solver.gamma = math.pow(self.lr_policy['gamma'], 1.0/scale)
elif solver.lr_policy == 'inv':
# gamma = gamma / scale
solver.gamma = self.lr_policy['gamma'] / scale
solver.power = self.lr_policy['power']
elif solver.lr_policy == 'poly':
solver.power = self.lr_policy['power']
elif solver.lr_policy == 'sigmoid':
# gamma = -gamma / scale
solver.gamma = -1.0 * self.lr_policy['gamma'] / scale
# stepsize = stepsize * scale
solver.stepsize = int(math.ceil(float(self.lr_policy['stepsize']) * scale))
else:
raise Exception('Unknown lr_policy: "%s"' % solver.lr_policy)
# go with the suggested defaults
solver.momentum = 0.9
solver.weight_decay = 0.0005
# Display 8x per epoch, or once per 5000 images, whichever is more frequent
solver.display = min(
int(math.floor(float(solver.max_iter) / (self.train_epochs * 8))),
int(math.ceil(5000.0 / train_data_layer.data_param.batch_size))
)
with open(self.path(self.solver_file), 'w') as outfile:
text_format.PrintMessage(solver, outfile)
self.solver = solver # save for later
return True
args = parser.parse_args()
return args
if __name__ == '__main__':
args = parse_args()
print('Called with args:')
print(args)
# load files
print 'Loading caffemodel: {}'.format(args.caffemodel)
with open(args.caffemodel, 'rb') as f:
binary_content = f.read()
protobuf = caffe_pb2.NetParameter()
protobuf.ParseFromString(binary_content)
layers = protobuf.layer
params = {}
for layer in layers:
if layer.type in args.layer_types:
print (layer.name, layer.type)
params[layer.name+'_w'] = np.reshape(np.array(layer.blobs[0].data), layer.blobs[0].shape.dim)
params[layer.name+'_b'] = np.reshape(np.array(layer.blobs[1].data), layer.blobs[1].shape.dim)
print params[layer.name+'_w'].shape, params[layer.name+'_b'].shape
# save the layers into a file
# if the file name is .pkl, save to pickle file.
# if the file name is .mat, save to mat file.
# otherwise, report file type not recognized.
def get_netparameter(self, model):
with open(model) as f:
net = cp.NetParameter()
pb.text_format.Parse(f.read(), net)
return net
def _build_rnn_network(self,
wtype='xavier',
std=0.01,
batchsize=100,
numstep=24):
network = caffe_pb2.NetParameter()
network.force_backward = True
network.name = 'rotation_rnn'
network.input.append('images')
network.input_dim.append(batchsize)
network.input_dim.append(3)
network.input_dim.append(64)
network.input_dim.append(64)
for t in range(numstep):
network.input.append('rotations%d' % t)
network.input_dim.append(batchsize)
network.input_dim.append(3)
network.input_dim.append(1)
network.input_dim.append(1)
layers = []
name_generator = self._connection_name_generator()
tensor_view = []
relu2_view = []
relu2_view_split = []
concat = []
dec_fc1 = []
dec_relu1 = []
dec_fc2 = []
dec_relu2 = []
dec_relu2_split = []
dec_img_fc1 = []
dec_img_relu1 = []
dec_img_fold = []
dec_img_up1 = []
dec_img_conv1 = []
dec_img_relu2 = []
dec_img_up2 = []
dec_img_conv2 = []
dec_img_relu3 = []
dec_img_up3 = []
dec_img_conv3 = []
dec_mask_fc1 = []
dec_mask_relu1 = []
dec_mask_fold = []
dec_mask_up1 = []
dec_mask_conv1 = []
dec_mask_relu2 = []
dec_mask_up2 = []
dec_mask_conv2 = []
dec_mask_relu3 = []
dec_mask_up3 = []
dec_mask_conv3 = []
conv1 = self._make_conv_layer(network,
kernel_size=5,
stride=2,
pad=2,
num_output=64,
shared_name='conv1')
conv1.bottom.append('images')
relu1 = self._make_relu_layer(network)
conv2 = self._make_conv_layer(network,
kernel_size=5,
stride=2,
pad=2,
num_output=128,
shared_name='conv2')
relu2 = self._make_relu_layer(network)
conv3 = self._make_conv_layer(network,
kernel_size=5,
stride=2,
pad=2,
num_output=256,
shared_name='conv3')
relu3 = self._make_relu_layer(network)
fc1 = self._make_inner_product_layer(network,
num_output=1024,
shared_name='fc1')
relu4 = self._make_relu_layer(network)
fc2 = self._make_inner_product_layer(network,
num_output=1024,
shared_name='fc2')
relu5 = self._make_relu_layer(network)
enc_split = self._make_split_layer(network)
fc1_id = self._make_inner_product_layer(network,
num_output=512,
shared_name='fc1_id')
relu1_id = self._make_relu_layer(network)
id_split = self._make_split_layer(network)
fc1_view = self._make_inner_product_layer(network,
num_output=512,
shared_name='fc1_view')
relu1_view = self._make_relu_layer(network)
tensor_view.append(
self._make_tensor_layer(network,
num_output=512,
shared_name='tensor_view'))
tensor_view[-1].bottom.append('rotations0')
relu2_view.append(self._make_relu_layer(network))
relu2_view_split.append(self._make_split_layer(network))
connections = []
connections.append(
(conv1.name, (conv1.top, relu1.bottom, relu1.top, conv2.bottom)))
connections.append(
(conv2.name, (conv2.top, relu2.bottom, relu2.top, conv3.bottom)))
connections.append(
(conv3.name, (conv3.top, relu3.bottom, relu3.top, fc1.bottom)))
connections.append(
(fc1.name, (fc1.top, relu4.bottom, relu4.top, fc2.bottom)))
connections.append((fc2.name, (fc2.top, relu5.bottom)))
connections.append((relu5.name, (relu5.top, enc_split.bottom)))
connections.append((enc_split.name, (enc_split.top, fc1_id.bottom)))
connections.append((fc1_id.name, (fc1_id.top, relu1_id.bottom,
relu1_id.top, id_split.bottom)))
connections.append((enc_split.name, (enc_split.top, fc1_view.bottom)))
connections.append(
(fc1_view.name, (fc1_view.top, relu1_view.bottom, relu1_view.top,
tensor_view[-1].bottom)))
for t in range(numstep):
# Action.
if t > 0:
tensor_view.append(
self._make_tensor_layer(network,
num_output=512,
shared_name='tensor_view'))
tensor_view[-1].bottom.append('rotations%d' % t)
relu2_view.append(self._make_relu_layer(network))
relu2_view_split.append(self._make_split_layer(network))
# Decoder.
concat.append(self._make_concat_layer(network))
dec_fc1.append(
self._make_inner_product_layer(network,
num_output=1024,
shared_name='dec_fc1'))
dec_relu1.append(self._make_relu_layer(network))
dec_fc2.append(
self._make_inner_product_layer(network,
num_output=1024,
shared_name='dec_fc2'))
dec_relu2.append(self._make_relu_layer(network))
dec_relu2_split.append(self._make_split_layer(network))
# Dec img path.
dec_img_fc1.append(
self._make_inner_product_layer(network,
num_output=16384,
shared_name='dec_img_fc1'))
dec_img_relu1.append(self._make_relu_layer(network))
dec_img_fold.append(self._make_folding_layer(network, 256, 8, 8))
dec_img_up1.append(self._make_upsampling_layer(network, stride=2))
dec_img_conv1.append(
self._make_conv_layer(network,
kernel_size=5,
stride=1,
pad=2,
num_output=128,
shared_name='dec_img_conv1'))
dec_img_relu2.append(self._make_relu_layer(network))
dec_img_up2.append(self._make_upsampling_layer(network, stride=2))
dec_img_conv2.append(
self._make_conv_layer(network,
kernel_size=5,
stride=1,
pad=2,
num_output=64,
shared_name='dec_img_conv2'))
dec_img_relu3.append(self._make_relu_layer(network))
dec_img_up3.append(self._make_upsampling_layer(network, stride=2))
dec_img_conv3.append(
self._make_conv_layer(network,
kernel_size=5,
stride=1,
pad=2,
num_output=3,
shared_name='dec_img_conv3'))
# Dec mask path.
dec_mask_fc1.append(
self._make_inner_product_layer(network,
num_output=8192,
shared_name='dec_mask_fc1'))
dec_mask_relu1.append(self._make_relu_layer(network))
dec_mask_fold.append(self._make_folding_layer(network, 128, 8, 8))
dec_mask_up1.append(self._make_upsampling_layer(network, stride=2))
dec_mask_conv1.append(
self._make_conv_layer(network,
kernel_size=5,
stride=1,
pad=2,
num_output=64,
shared_name='dec_mask_conv1'))
dec_mask_relu2.append(self._make_relu_layer(network))
dec_mask_up2.append(self._make_upsampling_layer(network, stride=2))
dec_mask_conv2.append(
self._make_conv_layer(network,
kernel_size=5,
stride=1,
pad=2,
num_output=32,
shared_name='dec_mask_conv2'))
dec_mask_relu3.append(self._make_relu_layer(network))
dec_mask_up3.append(self._make_upsampling_layer(network, stride=2))
dec_mask_conv3.append(
self._make_conv_layer(network,
kernel_size=5,
stride=1,
pad=2,
num_output=1,
shared_name='dec_mask_conv3'))
# dec connections.
if t > 0:
connections.append(
(relu2_view_split[-2].name, (relu2_view_split[-2].top,
tensor_view[-1].bottom)))
connections.append((tensor_view[-1].name, (tensor_view[-1].top,
relu2_view[-1].bottom)))
connections.append(
(relu2_view[-1].name, (relu2_view[-1].top,
relu2_view_split[-1].bottom)))
connections.append(
(id_split.name, (id_split.top, concat[-1].bottom)))
connections.append((relu2_view_split[-1].name,
(relu2_view_split[-1].top, concat[-1].bottom)))
connections.append(
(concat[-1].name, (concat[-1].top, dec_fc1[-1].bottom)))
connections.append(
(dec_fc1[-1].name, (dec_fc1[-1].top, dec_relu1[-1].bottom,
dec_relu1[-1].top, dec_fc2[-1].bottom)))
connections.append(
(dec_fc2[-1].name, (dec_fc2[-1].top, dec_relu2[-1].bottom)))
connections.append(
(dec_relu2[-1].name, (dec_relu2[-1].top,
dec_relu2_split[-1].bottom)))
# dec image connections.
connections.append(
(dec_relu2_split[-1].name, (dec_relu2_split[-1].top,
dec_img_fc1[-1].bottom)))
connections.append(
(dec_img_fc1[-1].name,
(dec_img_fc1[-1].top, dec_img_relu1[-1].bottom,
dec_img_relu1[-1].top, dec_img_fold[-1].bottom)))
connections.append(
(dec_img_fold[-1].name, (dec_img_fold[-1].top,
dec_img_up1[-1].bottom)))
connections.append(
(dec_img_up1[-1].name, (dec_img_up1[-1].top,
dec_img_conv1[-1].bottom)))
connections.append(
(dec_img_conv1[-1].name,
(dec_img_conv1[-1].top, dec_img_relu2[-1].bottom,
dec_img_relu2[-1].top, dec_img_up2[-1].bottom)))
connections.append(
(dec_img_up2[-1].name, (dec_img_up2[-1].top,
dec_img_conv2[-1].bottom)))
connections.append(
(dec_img_conv2[-1].name,
(dec_img_conv2[-1].top, dec_img_relu3[-1].bottom,
dec_img_relu3[-1].top, dec_img_up3[-1].bottom)))
connections.append(
(dec_img_up3[-1].name, (dec_img_up3[-1].top,
dec_img_conv3[-1].bottom)))
# dec mask connections.
connections.append(
(dec_relu2_split[-1].name, (dec_relu2_split[-1].top,
dec_mask_fc1[-1].bottom)))
connections.append(
(dec_mask_fc1[-1].name,
(dec_mask_fc1[-1].top, dec_mask_relu1[-1].bottom,
dec_mask_relu1[-1].top, dec_mask_fold[-1].bottom)))
connections.append(
(dec_mask_fold[-1].name, (dec_mask_fold[-1].top,
dec_mask_up1[-1].bottom)))
connections.append(
(dec_mask_up1[-1].name, (dec_mask_up1[-1].top,
dec_mask_conv1[-1].bottom)))
connections.append(
(dec_mask_conv1[-1].name,
(dec_mask_conv1[-1].top, dec_mask_relu2[-1].bottom,
dec_mask_relu2[-1].top, dec_mask_up2[-1].bottom)))
connections.append(
(dec_mask_up2[-1].name, (dec_mask_up2[-1].top,
dec_mask_conv2[-1].bottom)))
connections.append(
(dec_mask_conv2[-1].name,
(dec_mask_conv2[-1].top, dec_mask_relu3[-1].bottom,
dec_mask_relu3[-1].top, dec_mask_up3[-1].bottom)))
connections.append(
(dec_mask_up3[-1].name, (dec_mask_up3[-1].top,
dec_mask_conv3[-1].bottom)))
layers = [
conv1, relu1, conv2, relu2, conv3, relu3, fc1, relu4, fc2, relu5,
enc_split, fc1_id, relu1_id, id_split
]
layers += tensor_view
layers += relu2_view
layers += relu2_view_split
layers += concat
layers += dec_fc1
layers += dec_relu1
layers += dec_fc2
layers += dec_relu2
layers += dec_relu2_split
layers += dec_img_fc1
layers += dec_img_relu1
layers += dec_img_fold
layers += dec_img_up1
layers += dec_img_conv1
layers += dec_img_relu2
layers += dec_img_up2
layers += dec_img_conv2
layers += dec_img_relu3
layers += dec_img_up3
layers += dec_img_conv3
layers += dec_mask_fc1
layers += dec_mask_relu1
layers += dec_mask_fold
layers += dec_mask_up1
layers += dec_mask_conv1
layers += dec_mask_relu2
layers += dec_mask_up2
layers += dec_mask_conv2
layers += dec_mask_relu3
layers += dec_mask_up3
layers += dec_mask_conv3
final_img_concat = self._make_concat_layer(network)
for idx, l in enumerate(dec_img_conv3):
l.name = 't%d_%s' % (idx, l.name)
connections.append((l.name, (l.top, final_img_concat.bottom)))
final_img_concat.top.append('images_concat')
final_img_concat.loss_weight.append(10.0)
final_mask_concat = self._make_concat_layer(network)
for idx, l in enumerate(dec_mask_conv3):
l.name = 't%d_%s' % (idx, l.name)
connections.append((l.name, (l.top, final_mask_concat.bottom)))
final_mask_concat.top.append('masks_concat')
final_mask_concat.loss_weight.append(1.0)
layers += [final_img_concat, final_mask_concat]
# make connections.
for connection in connections:
self._tie(connection, name_generator)
for l in tensor_view[0:]:
tmp = reversed(l.bottom)
l.ClearField('bottom')
l.bottom.extend(tmp)
# Fix up the names based on the connections that were generated.
for pos, layer in enumerate(layers):
layer.name += '_%d' % pos
return network
