def make_keras_model_org_netvlad(im_rows=240, im_cols=320, im_chnls=3):
# im_rows = 240
# im_cols = 320
# im_chnls = 3
input_img = keras.layers.Input(batch_shape=(1, im_rows, im_cols, im_chnls))
cnn = make_from_vgg16(input_img, weights=None, layer_name='block5_pool')
out, out_amap = NetVLADLayer(num_clusters=64)(cnn)
model = keras.models.Model(inputs=input_img, outputs=out)
return model
def make_keras_model_relja_netvlad(DATA_DIR,
im_rows=240,
im_cols=320,
im_chnls=3):
# im_rows = 240
# im_cols = 320
# im_chnls = 3
input_img = keras.layers.Input(batch_shape=(1, im_rows, im_cols, im_chnls))
cnn = make_from_vgg16(input_img, weights=None, layer_name='block5_pool')
out, out_amap = NetVLADLayer(num_clusters=64)(cnn)
model = keras.models.Model(inputs=input_img, outputs=out)
model.load_weights(DATA_DIR + '/matlab_model.keras')
WPCA_M = scipy.io.loadmat(DATA_DIR +
'/WPCA_1.mat')['the_mat'] # 1x1x32768x4096
WPCA_b = scipy.io.loadmat(DATA_DIR + '/WPCA_2.mat')['the_mat'] # 4096x1
WPCA_M = WPCA_M[0, 0] # 32768x4096
WPCA_b = np.transpose(WPCA_b) #1x4096
return model, WPCA_M, WPCA_b
def load_basic_model():
K.set_learning_phase(0)
from CustomNets import make_from_mobilenet, make_from_vgg16
from CustomNets import NetVLADLayer, GhostVLADLayer
# Please choose only one of these.
if False: # VGG
input_img = keras.layers.Input(shape=(240, 320, 3))
cnn = make_from_vgg16(input_img,
weights=None,
layer_name='block5_pool',
kernel_regularizer=keras.regularizers.l2(0.01))
model = keras.models.Model(inputs=input_img, outputs=cnn)
if True: #mobilenet
input_img = keras.layers.Input(shape=(240, 320, 3))
cnn = make_from_mobilenet(
input_img,
layer_name='conv_pw_5_relu',
weights=None,
kernel_regularizer=keras.regularizers.l2(0.01))
model = keras.models.Model(inputs=input_img, outputs=cnn)
if False: #mobilenet+netvlad
input_img = keras.layers.Input(shape=(240, 320, 3))
cnn = make_from_mobilenet(
input_img,
layer_name='conv_pw_5_relu',
weights=None,
kernel_regularizer=keras.regularizers.l2(0.01))
# cnn = make_from_vgg16( input_img, weights=None, layer_name='block5_pool', kernel_regularizer=keras.regularizers.l2(0.01) )
out = NetVLADLayer(num_clusters=16)(cnn)
model = keras.models.Model(inputs=input_img, outputs=out)
if False: #netvlad only
input_img = keras.layers.Input(shape=(60, 80, 256))
out = NetVLADLayer(num_clusters=16)(input_img)
model = keras.models.Model(inputs=input_img, outputs=out)
model.summary()
return model
image_nchnl))
if CNN_type == 'mobilenet' or CNN_type == 'vgg16' or CNN_type == 'mobilenetv2':
pass
else:
assert (False)
# weights=imagenet will only work with nchanls=3
cnn = None
if CNN_type == 'mobilenet':
cnn = make_from_mobilenet( input_img, layer_name=layer_name,\
weights=init_model_weights, kernel_regularizer=keras.regularizers.l2(0.01) )
if CNN_type == 'vgg16':
cnn = make_from_vgg16(input_img,
weights=init_model_weights,
layer_name=layer_name,
kernel_regularizer=keras.regularizers.l2(0.001))
if CNN_type == 'mobilenetv2':
cnn = make_from_mobilenetv2( input_img, layer_name=layer_name,\
weights=init_model_weights, kernel_regularizer=keras.regularizers.l2(0.01) )
# Reduce nChannels of the output.
# @ Downsample (Optional)
if False: #Downsample last layer (Reduce nChannels of the output.)
cnn_dwn = keras.layers.Conv2D(256, (1, 1),
padding='same',
activation='relu')(cnn)
cnn_dwn = keras.layers.normalization.BatchNormalization()(cnn_dwn)
cnn_dwn = keras.layers.Conv2D(64, (1, 1),
padding='same',
# int_logr = InteractiveLogger( './models.keras/mobilenet_conv7_quash_chnls_allpairloss/' )
# int_logr = InteractiveLogger( './models.keras/mobilenet_conv7_quash_chnls_tripletloss2_K64/' )
# int_logr = InteractiveLogger( './models.keras/vgg16/block5_pool_k48_tripletloss2' )
int_logr = InteractiveLogger(
'./models.keras/vgg16_new/block5_pool_k16_allpairloss')
#--------------------------------------------------------------------------
# Core Model Setup
#--------------------------------------------------------------------------
# Build
input_img = keras.layers.Input(shape=(image_nrows, image_ncols,
image_nchnl))
# cnn = make_from_mobilenet( input_img )
cnn = make_from_vgg16(input_img, layer_name='block5_pool')
# Reduce nChannels of the output.
# @ Downsample (Optional)
if False: #Downsample last layer (Reduce nChannels of the output.)
cnn_dwn = keras.layers.Conv2D(256, (1, 1),
padding='same',
activation='relu')(cnn)
cnn_dwn = keras.layers.normalization.BatchNormalization()(cnn_dwn)
cnn_dwn = keras.layers.Conv2D(32, (1, 1),
padding='same',
activation='relu')(cnn_dwn)
cnn_dwn = keras.layers.normalization.BatchNormalization()(cnn_dwn)
cnn = cnn_dwn
out, out_amap = NetVLADLayer(num_clusters=16)(cnn)
model = keras.models.Model(inputs=input_img, outputs=out)
def do_demo():
im_rows = 240
im_cols = 320
im_chnls = 3
im_rows = 480
im_cols = 640
im_chnls = 3
##------ Create Model (from json file)
# [[[[[Option-A]]]]] (load from json) this doesn't work. seem like i need to implement get_config in my custom layer which I am not able to do correctly, TODO fix in the future
# json_fname = base_path+'/model.json'
# print 'Read model json: ', json_fname
# with open(json_fname, 'r') as myfile:
# model_json_string=myfile.read()
# model = keras.models.model_from_json( model_json_string, custom_objects={'NetVLADLayer': NetVLADLayer} )
# [[[[[Option-B]]]]] (from h5 files). Apparently h5 stores the model config and the weights.
# h5_fname = base_path+'/model.h5'
# model = keras.models.load_model( h5_fname, custom_objects={'NetVLADLayer': NetVLADLayer} )
# [[[[[Option-C]]]]] : Construct manually
# @ Input
# input_img = keras.layers.Input( shape=(480, 640, 3 ) )
# input_img = keras.layers.Input( shape=(240, 320, 3 ) )
# input_img = keras.layers.Input( batch_shape=(1,480, 640, 3 ) )
input_img = keras.layers.Input(batch_shape=(1, im_rows, im_cols, im_chnls))
# @ CNN
cnn = make_from_vgg16(input_img, weights=None, layer_name='block5_pool')
# cnn = make_from_mobilenet( input_img, weights=None, layer_name='conv_pw_13_relu' )
# @ Downsample (Optional)
if False: #Downsample last layer (Reduce nChannels of the output.)
cnn_dwn = keras.layers.Conv2D(256, (1, 1),
padding='same',
activation='relu')(cnn)
cnn_dwn = keras.layers.normalization.BatchNormalization()(cnn_dwn)
cnn_dwn = keras.layers.Conv2D(32, (1, 1),
padding='same',
activation='relu')(cnn_dwn)
cnn_dwn = keras.layers.normalization.BatchNormalization()(cnn_dwn)
cnn = cnn_dwn
# @ NetVLADLayer
out, out_amap = NetVLADLayer(num_clusters=16)(cnn)
model = keras.models.Model(inputs=input_img, outputs=out)
##---------- Print Model Info, Memory Usage, FLOPS
if True: # Set this to `True` to display FLOPs, memory etc .
model.summary()
# print_flops_report( model )
# print_model_memory_usage( 1, model )
print 'input_shape=%s\toutput_shape=%s' % (model.input_shape,
model.output_shape)
##------------ Load Weights
model_file = './models.keras/vgg16_new/block5_pool_k16_allpairloss/core_model.1000.keras'
print 'load_weights=', model_file
model.load_weights(model_file)
code.interact(local=locals())
##------------model.predict
# Load your image here.
tmp_zer = np.random.rand(1, im_rows, im_cols, im_chnls).astype('float32')
start = time.time()
tmp_zer_out = model.predict(tmp_zer)
print 'Exec in %4.2fms' % (1000. * (time.time() - start))
tmp_zer = np.random.rand(1, im_rows, im_cols, im_chnls).astype('float32')
start = time.time()
tmp_zer_out = model.predict(tmp_zer)
print 'Exec in %4.2fms' % (1000. * (time.time() - start))
