Skip to content

Commit

Permalink
Merge pull request #38 from akshaychawla/master
Browse files Browse the repository at this point in the history 
Support for Deconv and Crop layers in caffe2keras converter
  • Loading branch information
@MarcBS
MarcBS authored Apr 21, 2017
2 parents 9fe9ab3 + bb9b159 commit 51cb559
Show file tree
Hide file tree
Showing 9 changed files with 1,659 additions and 7 deletions.
3 changes: 2 additions & 1 deletion keras/caffe/caffe2keras.py
View file
Original file line number Diff line number Diff line change
Expand Up @@ -38,9 +38,10 @@ def main(args):
print("Storing model...")
json_string = model.to_json()
open(store_path + '/Keras_model_structure.json', 'w').write(json_string)
# Save converted model weights
# # Save converted model weights
model.save_weights(store_path + '/Keras_model_weights.h5', overwrite=True)
print("Finished storing the converted model to "+ store_path)


main(args)

104 changes: 98 additions & 6 deletions keras/caffe/convert.py
View file
Original file line number Diff line number Diff line change
Expand Up @@ -8,6 +8,7 @@

from . import caffe_pb2 as caffe
import google.protobuf
from google.protobuf import text_format
from .caffe_utils import *
from .extra_layers import *

Expand All @@ -26,7 +27,7 @@ def caffe_to_keras(prototext, caffemodel, phase='train', debug=False):
'''
config = caffe.NetParameter()
prototext = preprocessPrototxt(prototext, debug)
google.protobuf.text_format.Merge(prototext, config)
text_format.Merge(prototext, config)

if len(config.layers) != 0:
raise Exception("Prototxt files V1 are not supported.")
Expand All @@ -36,10 +37,19 @@ def caffe_to_keras(prototext, caffemodel, phase='train', debug=False):
else:
raise Exception('could not load any layers from prototext')

input_dim = []
if len(config.input_dim[:])==0:
input_dim.append(int(layers[0].input_param.shape[0].dim[0]))
input_dim.append(int(layers[0].input_param.shape[0].dim[1]))
input_dim.append(int(layers[0].input_param.shape[0].dim[2]))
input_dim.append(int(layers[0].input_param.shape[0].dim[3]))
else:
input_dim = tuple(config.input_dim[:])

print("CREATING MODEL")
model = create_model(layers,
0 if phase == 'train' else 1,
tuple(config.input_dim[1:]),
tuple(input_dim[1:]),
debug)

params = caffe.NetParameter()
Expand Down Expand Up @@ -95,7 +105,6 @@ def preprocessPrototxt(prototxt, debug=False):
f.close()
return p


def create_model(layers, phase, input_dim, debug=False):
'''
layers:
Expand Down Expand Up @@ -186,8 +195,81 @@ def create_model(layers, phase, input_dim, debug=False):

if pad_h + pad_w > 0:
input_layers = ZeroPadding2D(padding=(pad_h, pad_w), name=name + '_zeropadding')(input_layers)
net_node[layer_nb] = Convolution2D(nb_filter, nb_row, nb_col, bias=has_bias, subsample=(stride_h, stride_w), name=name)(input_layers)
net_node[layer_nb] = Convolution2D(nb_filter, nb_row, nb_col, bias=has_bias, subsample=(stride_h, stride_w), name=name, border_mode='valid')(input_layers)

elif type_of_layer == 'deconvolution':
has_bias = layer.convolution_param.bias_term
nb_filter = int(layer.convolution_param.num_output)
nb_col = int(layer.convolution_param.kernel_size[0])
nb_row = int(layer.convolution_param.kernel_size[0])
stride_h = int(layer.convolution_param.stride[0])
stride_w = int(layer.convolution_param.stride[0])
pad_h = 0
pad_w = 0
try:
pad_h = int(layer.convolution_param.pad[0])
pad_w = int(layer.convolution_param.pad[0])
except Exception as e:
pass

if(debug):
print("Deconv kernel")
print(str(nb_filter)+'x'+str(nb_col)+'x'+str(nb_row))
print("stride")
print(stride_h)
print("pad")
print(pad_h)

# shape inference
semi_model = Model(input=net_node[0], output=input_layers[0])
ip_shape = semi_model.layers[-1].output_shape
del semi_model

##### FORMULA FOR O/P SHAPE OF DECONV ########
# o = s (i - 1) + a + k - 2p
# where:
# i - input size (rows or cols),
# k - kernel size (nb_filter),
# s - stride (subsample for rows or cols respectively),
# p - padding size
# a - (not used)
##############################################

i_h, i_w = ip_shape[2], ip_shape[3]
output_shape = [ None,
nb_filter,
stride_h*(i_h - 1) + nb_row - 2*pad_h,
stride_w*(i_w - 1) + nb_col - 2*pad_w
]
if pad_h + pad_w > 0:
input_layers = ZeroPadding2D(padding=(pad_h, pad_w), name=name + '_zeropadding')(input_layers)

net_node[layer_nb] = Deconvolution2D(nb_filter, nb_row, nb_col, output_shape, name=name, subsample=(stride_h, stride_w))(input_layers)

elif type_of_layer == "crop":
assert (len(input_layers)==2), "Caffe crop layer must have \
only 2 Bottom blobs"

# shape inference - Input Layer [1]
semi_model = Model(input=net_node[0], output=input_layers[0])
shape1 = semi_model.layers[-1].output_shape
del semi_model

# shape inference - Input Layer [2]
semi_model = Model(input=net_node[0], output=input_layers[1])
shape2 = semi_model.layers[-1].output_shape
del semi_model

# offset parameter
offset = int(layer.crop_param.offset[0])

crop_param = (
(offset, shape2[2] - (offset + shape1[2])),
(offset, shape2[3] - (offset + shape1[3]))
)

net_node[layer_nb] = Cropping2D(cropping=crop_param, name=name)(input_layers[1])

elif type_of_layer == 'dropout':
prob = layer.dropout_param.dropout_ratio
net_node[layer_nb] = Dropout(prob, name=name)(input_layers)
Expand Down Expand Up @@ -256,6 +338,15 @@ def create_model(layers, phase, input_dim, debug=False):
net_node[layer_nb] = Activation('sigmoid', name=name)(input_layers)

elif type_of_layer == 'softmax' or type_of_layer == 'softmaxwithloss':
# check output shape
semi_model = Model(input=net_node[0], output=input_layers[0])
op_shape = semi_model.layers[-1].output_shape
del semi_model

if len(op_shape)==4: # for img segmentation - i/p to softmax is (None, num_classes, height, width)
interm_layer = Reshape((op_shape[1], op_shape[2]*op_shape[3]))(input_layers)
input_layers = Permute((2,1))(interm_layer) # reshaped to (None, height*width, num_classes)

net_node[layer_nb] = Activation('softmax', name=name)(input_layers)

elif type_of_layer == 'split':
Expand Down Expand Up @@ -383,7 +474,7 @@ def convert_weights(param_layers, v='V1', debug=False):

weights[layer.name] = [weights_gamma.astype(dtype=np.float32), weights_beta.astype(dtype=np.float32)]

elif typ == 'convolution':
elif typ == 'convolution' or typ == 'deconvolution':
blobs = layer.blobs

if(v == 'V1'):
Expand Down Expand Up @@ -415,7 +506,7 @@ def convert_weights(param_layers, v='V1', debug=False):
if layer.convolution_param.bias_term:
weights_b = np.array(blobs[1].data)
else:
weights_b = None
weights_b = np.zeros((nb_filter,))

group_data_size = len(blobs[0].data) // group
stacks_size_per_group = stack_size // group
Expand Down Expand Up @@ -453,5 +544,6 @@ def load_weights(model, weights):
for layer in model.layers:
if weights.has_key(layer.name):
model.get_layer(layer.name).set_weights(weights[layer.name])
print "Copied wts for layer:",layer.name


Loading

0 comments on commit 51cb559

Please sign in to comment.