def load_resnet152_notop_from_caffemodel(load_path='models/'):
prototxt = 'train_resnet152_places365.prototxt'
caffemodel = 'resnet152_places365.caffemodel'
debug = False # display input shape etc. and saved to 'debug.prototxt' if True
# Note this is old style ResNet 152 with ReLU on the main path such that gradient may explode/vanish
with Timer("Converting caffe model .."): # 141 secs
model_loaded_from_caffe = convert.caffe_to_keras(load_path + prototxt,
load_path +
caffemodel,
debug=debug)
# cut the top layer before average pooling
model_loaded_from_caffe.layers.pop() # drop 'prob' Activation layer
model_loaded_from_caffe.layers.pop() # drop 'fc365' Dense layer
model_loaded_from_caffe.layers.pop() # drop 'fc365_flatten' Flatten layer
model_loaded_from_caffe.layers.pop(
) # drop 'pool5' Average Pooling layer, sub with rr pooling
# model.output is still: Softmax.0, revise by
model_loaded_from_caffe.outputs = [
model_loaded_from_caffe.layers[-1].output
]
model_loaded_from_caffe.output_layers = [
model_loaded_from_caffe.layers[-1]
]
model_loaded_from_caffe.layers[-1].outbound_nodes = []
return model_loaded_from_caffe
def load_vgg16_notop_from_caffemodel(load_path='models/'):
prototxt = 'train_vgg16_places365.prototxt'
caffemodel = 'vgg16_places365.caffemodel'
debug = False # display input shape etc. and saved to 'debug.prototxt' if True
with Timer("Converting caffe model .."): # 151 secs
model_loaded_from_caffe = convert.caffe_to_keras(load_path + prototxt,
load_path +
caffemodel,
debug=debug)
# prepare a vgg model_loaded_from_caffe without top layers
img_width = 224
img_height = 224
img_size = (3, img_width, img_height)
input_tensor = Input(batch_shape=(None, ) + img_size)
model_vgg_places365_notop = vgg16.VGG16(input_tensor=input_tensor,
include_top=False)
model_vgg_places365_notop.get_layer('block1_conv1')\
.set_weights(model_loaded_from_caffe.get_layer('conv1_1').get_weights())
model_vgg_places365_notop.get_layer('block1_conv2') \
.set_weights(model_loaded_from_caffe.get_layer('conv1_2').get_weights())
model_vgg_places365_notop.get_layer('block2_conv1') \
.set_weights(model_loaded_from_caffe.get_layer('conv2_1').get_weights())
model_vgg_places365_notop.get_layer('block2_conv2') \
.set_weights(model_loaded_from_caffe.get_layer('conv2_2').get_weights())
model_vgg_places365_notop.get_layer('block3_conv1') \
.set_weights(model_loaded_from_caffe.get_layer('conv3_1').get_weights())
model_vgg_places365_notop.get_layer('block3_conv2') \
.set_weights(model_loaded_from_caffe.get_layer('conv3_2').get_weights())
model_vgg_places365_notop.get_layer('block3_conv3') \
.set_weights(model_loaded_from_caffe.get_layer('conv3_3').get_weights())
model_vgg_places365_notop.get_layer('block4_conv1') \
.set_weights(model_loaded_from_caffe.get_layer('conv4_1').get_weights())
model_vgg_places365_notop.get_layer('block4_conv2') \
.set_weights(model_loaded_from_caffe.get_layer('conv4_2').get_weights())
model_vgg_places365_notop.get_layer('block4_conv3') \
.set_weights(model_loaded_from_caffe.get_layer('conv4_3').get_weights())
model_vgg_places365_notop.get_layer('block5_conv1') \
.set_weights(model_loaded_from_caffe.get_layer('conv5_1').get_weights())
model_vgg_places365_notop.get_layer('block5_conv2') \
.set_weights(model_loaded_from_caffe.get_layer('conv5_2').get_weights())
model_vgg_places365_notop.get_layer('block5_conv3') \
.set_weights(model_loaded_from_caffe.get_layer('conv5_3').get_weights())
# TODO rename the model
# model_vgg_places365_notop.get_config().__setitem__('name', 'vgg_places365_model')
return model_vgg_places365_notop
def test_convertGoogleNet():
load_path = 'keras/caffe/models'
store_path = 'keras/caffe/models'
prototxt = 'train_val_for_keras.prototxt'
caffemodel = 'bvlc_googlenet.caffemodel'
# Convert model from caffe to keras
model = convert.caffe_to_keras(load_path+'/'+prototxt, load_path+'/'+caffemodel, debug=False)
assert(model.__class__.__name__ == 'Model')
# Save converted model structure
json_string = model.to_json()
open(store_path + '/Keras_model_structure.json', 'w').write(json_string)
# Save converted model weights
model.save_weights(store_path + '/Keras_model_weights.h5', overwrite=True)
def test_convertGoogleNet():
load_path = "keras/caffe/models"
store_path = "keras/caffe/models"
prototxt = "train_val_for_keras.prototxt"
caffemodel = "bvlc_googlenet.caffemodel"
# Convert model from caffe to keras
model = convert.caffe_to_keras(load_path + "/" + prototxt, load_path + "/" + caffemodel, debug=False)
assert model.__class__.__name__ == "Graph"
# Save converted model structure
json_string = model.to_json()
open(store_path + "/Keras_model_structure.json", "w").write(json_string)
# Save converted model weights
model.save_weights(store_path + "/Keras_model_weights.h5", overwrite=True)
def load_vgg16_notop_from_caffemodel():
prototxt = 'Caffe/prototxt/train_vgg16_places365_wtop.prototxt'
caffemodel = 'Caffe/weights/vgg16_places365_wtop.caffemodel'
model_loaded_from_caffe = convert.caffe_to_keras(prototxt,
caffemodel,
debug=False)
# prepare a vgg model_loaded_from_caffe without top layers
img_width = 224
img_height = 224
img_size = (3, img_width, img_height)
input_tensor = Input(batch_shape=(None, ) + img_size)
model_vgg16_places365_notop = vgg16.VGG16(input_tensor=input_tensor,
include_top=False)
model_vgg16_places365_notop.get_layer('block1_conv1')\
.set_weights(model_loaded_from_caffe.get_layer('conv1_1').get_weights())
model_vgg16_places365_notop.get_layer('block1_conv2') \
.set_weights(model_loaded_from_caffe.get_layer('conv1_2').get_weights())
model_vgg16_places365_notop.get_layer('block2_conv1') \
.set_weights(model_loaded_from_caffe.get_layer('conv2_1').get_weights())
model_vgg16_places365_notop.get_layer('block2_conv2') \
.set_weights(model_loaded_from_caffe.get_layer('conv2_2').get_weights())
model_vgg16_places365_notop.get_layer('block3_conv1') \
.set_weights(model_loaded_from_caffe.get_layer('conv3_1').get_weights())
model_vgg16_places365_notop.get_layer('block3_conv2') \
.set_weights(model_loaded_from_caffe.get_layer('conv3_2').get_weights())
model_vgg16_places365_notop.get_layer('block3_conv3') \
.set_weights(model_loaded_from_caffe.get_layer('conv3_3').get_weights())
model_vgg16_places365_notop.get_layer('block4_conv1') \
.set_weights(model_loaded_from_caffe.get_layer('conv4_1').get_weights())
model_vgg16_places365_notop.get_layer('block4_conv2') \
.set_weights(model_loaded_from_caffe.get_layer('conv4_2').get_weights())
model_vgg16_places365_notop.get_layer('block4_conv3') \
.set_weights(model_loaded_from_caffe.get_layer('conv4_3').get_weights())
model_vgg16_places365_notop.get_layer('block5_conv1') \
.set_weights(model_loaded_from_caffe.get_layer('conv5_1').get_weights())
model_vgg16_places365_notop.get_layer('block5_conv2') \
.set_weights(model_loaded_from_caffe.get_layer('conv5_2').get_weights())
model_vgg16_places365_notop.get_layer('block5_conv3') \
.set_weights(model_loaded_from_caffe.get_layer('conv5_3').get_weights())
# rename the model
model_vgg16_places365_notop.name = 'vgg16_notop_places365_model'
return model_vgg16_places365_notop
def test_convertGoogleNet():
load_path = 'keras/caffe/models'
store_path = 'keras/caffe/models'
prototxt = 'train_val_for_keras.prototxt'
caffemodel = 'bvlc_googlenet.caffemodel'
# Convert model from caffe to keras
model = convert.caffe_to_keras(load_path + '/' + prototxt,
load_path + '/' + caffemodel,
debug=False)
assert (model.__class__.__name__ == 'Model')
# Save converted model structure
json_string = model.to_json()
open(store_path + '/Keras_model_structure.json', 'w').write(json_string)
# Save converted model weights
model.save_weights(store_path + '/Keras_model_weights.h5', overwrite=True)
def main(args):
if(not args.store_path):
store_path = args.load_path
else:
store_path = args.store_path
print("Converting model...")
model = convert.caffe_to_keras(args.load_path+'/'+args.prototxt, args.load_path+'/'+args.caffemodel, debug=args.debug)
print("Finished converting model.")
# Save converted model structure
print("Storing model...")
json_string = model.to_json()
open(store_path + '/Keras_model_structure.json', 'w').write(json_string)
# Save converted model weights
model.save_weights(store_path + '/Keras_model_weights.h5', overwrite=True)
print("Finished storing the converted model to "+ store_path)
def main(args):
if (not args.store_path):
store_path = args.load_path
else:
store_path = args.store_path
print("Converting model...")
model = convert.caffe_to_keras(args.load_path + '/' + args.prototxt,
args.load_path + '/' + args.caffemodel,
debug=args.debug)
print("Finished converting model.")
# Save converted model structure
print("Storing model...")
json_string = model.to_json()
open(store_path + '/Keras_model_structure.json', 'w').write(json_string)
# Save converted model weights
model.save_weights(store_path + '/Keras_model_weights.h5', overwrite=True)
print("Finished storing the converted model to " + store_path)
path + "/deploy.prototxt"), "Err. Couldn't find the debug.prototxt file"
assert os.path.exists(
path + "/horse.png"), "Err. Couldn't find the horse.png image file"
if not os.path.exists(path + "/fcn8s-heavy-pascal.caffemodel"):
call([
"wget http://dl.caffe.berkeleyvision.org/fcn8s-heavy-pascal.caffemodel -O "
+ "./" + path + "/fcn8s-heavy-pascal.caffemodel"
],
shell=True)
assert os.path.exists(
path +
"/fcn8s-heavy-pascal.caffemodel"), "Err. Cannot find .caffemodel file. \
please download file using command : wget http://dl.caffe.berkeleyvision.org/fcn8s-heavy-pascal.caffemodel "
model = convert.caffe_to_keras(path + "/deploy.prototxt",
path + "/fcn8s-heavy-pascal.caffemodel",
debug=1)
print "Yay!"
# 1. load image
img = misc.imread(path + "./horse.png")
# modify it
img = np.rollaxis(img, 2)
img = np.expand_dims(img, 0)
# 2. run forward pass
op = model.predict(img)
# 3. reshape output
