def simple_load_demo():
"""
The way to a)load json-formatted models b)loading weights c) sample prediction
"""
# LOG_DIR = 'models.keras/Apr2019/K16_gray_training/'
LOG_DIR = 'models.keras/Apr2019/gray_conv6_K16Ghost1__centeredinput/'
# Load JSON formatted model
json_string = open_json_file( LOG_DIR+'/model.json' )
print '======================='
pprint.pprint( json_string, indent=4 )
print '======================='
model = keras.models.model_from_json(str(json_string), custom_objects={'NetVLADLayer': NetVLADLayer, 'GhostVLADLayer': GhostVLADLayer} )
print 'OLD MODEL: '
model.summary()
quit()
# Load Weights from model-file
model_fname = LOG_DIR+'/core_model.%d.keras' %(100)
print 'Load model: ', model_fname
model.load_weights( model_fname )
# Replace Input Layer
new_model = change_model_inputshape( model, new_input_shape=(1,1500,500,1) )
# Sample Predict
# test new model on a random input image
X = np.random.rand(new_input_shape[0], new_input_shape[1], new_input_shape[2], new_input_shape[3] )
y_pred = new_model.predict(X)
print('try predict with a random input_img with shape='+str(X.shape)+ str(y_pred) )
def render_and_predict():
#---------- Setup model
LOG_DIR = 'models.keras/Apr2019/gray_conv6_K16__centeredinput/'
# Load json-model
json_model_fname = LOG_DIR + '/model.json'
print 'Load json-model: ', json_model_fname
json_string = open_json_file(json_model_fname)
model = keras.models.model_from_json(
str(json_string), custom_objects={'NetVLADLayer': NetVLADLayer})
# Load Weights
model_fname = LOG_DIR + '/core_model.%d.keras' % (400)
print 'Load model: ', model_fname
model.load_weights(model_fname)
# change_model_inputshape
new_model = change_model_inputshape(model,
new_input_shape=(11, 480, 640, 1))
#--------------- Setup renderer
PITS_VAL_PATH = '/Bulk_Data/data_Akihiko_Torii/Pitssburg_validation/'
pr = PittsburgRenderer(PITS_VAL_PATH)
for _ in range(100):
print '---'
a, b = pr.step(nP=5,
nN=5,
ENABLE_IMSHOW=False,
return_gray=True,
resize=(640, 480))
a = np.copy(a) #if you dont do a copy, keras.predict gives sigsegv
# from CustomNets import do_typical_data_aug
# a = do_typical_data_aug( a )
# Predict
print 'a.shape=', a.shape, '\ta.dtype=', a.dtype
a_pred = new_model.predict((a.astype('float32') - 128.) / 255.)
# a0_pred = new_model.predict( np.expand_dims(a[0],0).astype('float32') )
# a_p0_pred = new_model.predict( np.expand_dims(a[1],0).astype('float32'))
# a_n0_pred = new_model.predict(np.expand_dims(a[6],0).astype('float32') )
DOT = np.matmul(a_pred, a_pred[0])
print '<a, a0> ', DOT
# imshow
cv2.imshow('query', a[0].astype('uint8'))
imshow_set('sim', a[1:6], str(DOT[1:6]))
imshow_set('dif', a[6:], str(DOT[6:]))
key = cv2.waitKey(0)
if key == ord('q'):
print 'Quit...'
break
code.interact(local=locals())
def compare_jsonload_with_staticload():
# Load Image
# im = cv2.resize( cv2.imread( '/app/lena.jpg', 0 ), (320,240) )
im = cv2.imread('/app/lena.jpg', 0)
im_rows, im_cols = im.shape[0:2]
im_centered = (im - 128.) / 255.
im_centered = np.expand_dims(np.expand_dims(im_centered, 0), -1)
LOG_DIR = './models.keras/Apr2019/centeredinput-gray__mobilenet-conv6__K16__allpairloss/'
# LOG_DIR = './models.keras/Apr2019/default_model/'
# LOG_DIR = 'models.keras/Apr2019/gray_conv6_K16Ghost1__centeredinput/'
model_fname = LOG_DIR + '/core_model.%d.keras' % (1000)
# model_fname = LOG_DIR+'/mobilenet_conv7_allpairloss.keras'
# Load JSON
if True:
json_string = open_json_file(LOG_DIR + '/model.json')
model_json = keras.models.model_from_json(str(json_string),
custom_objects={
'NetVLADLayer':
NetVLADLayer,
'GhostVLADLayer':
GhostVLADLayer
})
print 'Load model into `model_json`: ', model_fname
model_json.load_weights(model_fname)
model_json1 = change_model_inputshape(model_json,
new_input_shape=(1, im_rows,
im_cols, 1),
verbose=True)
# Load Static
if True:
image_nrows = im_rows #240
image_ncols = im_cols #320
image_nchnl = 1
netvlad_num_clusters = 16
input_img = keras.layers.Input(shape=(image_nrows, image_ncols,
image_nchnl))
cnn = make_from_mobilenet(input_img,
layer_name='conv_pw_7_relu',
weights=None,
kernel_regularizer=None,
trainable=False)
out = NetVLADLayer(num_clusters=netvlad_num_clusters)(cnn)
model_static = keras.models.Model(inputs=input_img, outputs=out)
print 'Load model into `model_json`: ', model_fname
model_static.load_weights(model_fname)
# a = model_json.predict( im_centered )
# b = model_static.predict( im_centered )
code.interact(local=locals())
def model_config(model, im_w, im_h): # Replace Input Layer
new_input_shape = (1, im_w, im_h, 3) #(1,480,640,1)
new_model = change_model_inputshape(model, new_input_shape=new_input_shape)
return new_model
# model = load_basic_model()
# quit()
#-----
# Replace Input Layer's Dimensions
im_rows = None #480
im_cols = 752
im_chnls = 3
if im_rows == None or im_cols == None or im_chnls == None:
print tcol.WARNING, 'NOT doing `change_model_inputshape`', tcol.ENDC
new_model = model
else:
# change_model_inputshape uses model_from_json internally, I feel a bit uncomfortable about this.
new_model = change_model_inputshape(model,
new_input_shape=(1, im_rows,
im_cols,
im_chnls),
verbose=True)
print 'OLD MODEL: ', 'input_shape=', str(model.inputs)
print 'NEW MODEL: input_shape=', str(new_model.inputs)
#-----
# Write Tensorflow (atleast 1.12) proto-binary (.pb)
write_kerasmodel_as_tensorflow_pb(new_model,
LOG_DIR=LOG_DIR,
output_model_name='output_model.pb')
# write_kerasmodel_as_tensorflow_pb( new_model, LOG_DIR=LOG_DIR, output_model_name=kerasmodel_h5file.split('/')[-1]+'.pb' )
#-----
# Clean up graph with Nvidia's graphsurgeon
# currently not in use but might come in handly later...maybe
print '## kerasmodel_h5file = ', kerasmodel_h5file
print '## LOG_DIR = ', LOG_DIR
print '##------------------------------------------------------------##'
print tcol.ENDC
#---
# Load HDF5 Keras model
model = load_keras_hdf5_model(kerasmodel_h5file, verbose=True) #this
model.summary()
#---
# FLOPS
from CustomNets import print_model_memory_usage, print_flops_report
print tcol.OKGREEN + '====\n==== print_model_memory_usage\n====', tcol.ENDC
print_model_memory_usage(1, model)
print_flops_report(model)
#---
# Change Model Shape and do flops computation
print tcol.OKGREEN + '====\n==== Memory and FLOPS for various Input Shapes\n====', tcol.ENDC
for m in [0.5, 1.0, 2.0, 4.0]:
from predict_utils import change_model_inputshape
new_input_shape = (None, int(model.input.shape[1].value * m),
int(model.input.shape[2].value * m),
model.input.shape[3].value)
new_model = change_model_inputshape(model,
new_input_shape=new_input_shape)
print_model_memory_usage(1, new_model)
do_inference_on_random_input(new_model)
def __init__(self, kerasmodel_file, im_rows=600, im_cols=960, im_chnls=3):
## Build net
from keras.backend.tensorflow_backend import set_session
import tensorflow as tf
config = tf.ConfigProto()
config.gpu_options.per_process_gpu_memory_fraction = 0.1
# config.gpu_options.visible_device_list = "0"
set_session(tf.Session(config=config))
# Blackbox 4
# self.im_rows = 512
# self.im_cols = 640
# self.im_chnls = 3
# point grey
# self.im_rows = 600
# self.im_cols = 960
# self.im_chnls = 3
# EuroC
# self.im_rows = 480
# self.im_cols = 752
# self.im_chnls = 3
self.im_rows = int(im_rows)
self.im_cols = int(im_cols)
self.im_chnls = int(im_chnls)
LOG_DIR = '/'.join(kerasmodel_file.split('/')[0:-1])
model_type = LOG_DIR.split('/')[-1]
# code.interact( local=locals() )
assert os.path.isdir(
LOG_DIR
), "The LOG_DIR doesnot exist of there is a permission issue. LOG_DIR=" + LOG_DIR
assert os.path.isfile(
LOG_DIR + '/model.json'
), "model.json does not exist in LOG_DIR=" + LOG_DIR + '. This file is needed to load the network architecture from json format. This file should have been created when you learned the model using script in github.com/mpkuse/cartwheel_train'
#----- @ Load Model Structure from JSON
# Load JSON formatted model
json_string = open_json_file(LOG_DIR + '/model.json')
# print '======================='
# pprint.pprint( json_string, indent=4 )
# print '======================='
model = keras.models.model_from_json(str(json_string),
custom_objects={
'NetVLADLayer': NetVLADLayer,
'GhostVLADLayer':
GhostVLADLayer
})
old_input_shape = model._layers[0].input_shape
model._layers[0]
print 'OLD MODEL: ', 'input_shape=', str(old_input_shape)
model.summary()
model_visual_fname = '/app/core.png'
print 'Writing Model Visual to: ', model_visual_fname
keras.utils.plot_model(model,
to_file=model_visual_fname,
show_shapes=True)
#----- @ Load Weights
assert os.path.isfile(
kerasmodel_file
), 'The model weights file doesnot exists or there is a permission issue.' + "kerasmodel_file=" + kerasmodel_file
model_fname = kerasmodel_file
print tcol.OKGREEN, 'Load model: ', model_fname, tcol.ENDC
model.load_weights(model_fname)
# Replace Input Layer
new_model = change_model_inputshape(model,
new_input_shape=(1, self.im_rows,
self.im_cols,
self.im_chnls))
new_input_shape = new_model._layers[0].input_shape
print 'OLD MODEL: ', 'input_shape=', str(old_input_shape)
print 'NEW MODEL: input_shape=', str(new_input_shape)
self.model = new_model
self.model_type = model_type
# Doing this is a hack to force keras to allocate GPU memory. Don't comment this,
tmp_zer = np.zeros((1, self.im_rows, self.im_cols, self.im_chnls),
dtype='float32')
tmp_zer_out = self.model.predict(tmp_zer)
print 'model input.shape=', tmp_zer.shape, '\toutput.shape=', tmp_zer_out.shape
print 'model_type=', self.model_type
print '-----'
print '\tinput_image.shape=', tmp_zer.shape
print '\toutput.shape=', tmp_zer_out.shape
print '\tminmax=', np.min(tmp_zer_out), np.max(tmp_zer_out)
print '\tnorm=', np.linalg.norm(tmp_zer_out)
print '\tdtype=', tmp_zer_out.dtype
print '-----'
LOG_DIR = '/'.join( kerasmodel_h5file.split('/')[0:-1] )
print tcol.HEADER
print '##------------------------------------------------------------##'
print '## kerasmodel_h5file = ', kerasmodel_h5file
print '## LOG_DIR = ', LOG_DIR
print '## nrows = ', nrows, '\tncols = ', ncols
print '##------------------------------------------------------------##'
print tcol.ENDC
#---
# Load HDF5 Keras model
model = load_keras_hdf5_model( kerasmodel_h5file, verbose=True ) #this
#---
# Change Shape
new_input_shape= ( None, nrows, ncols, model.input.shape[3].value )
new_model = change_model_inputshape( model, new_input_shape=new_input_shape, verbose=True )
new_model.summary()
#---
# Model Save
new_model_fname = '.'.join( kerasmodel_h5file.split( '.' )[0:-1] )+'.%dx%dx%d.h5' %(new_input_shape[1], new_input_shape[2], new_input_shape[3])
print '====\n====Save new_model to:', tcol.OKBLUE, new_model_fname, tcol.ENDC, '\n===='
# import code
# code.interact( local=locals() )
new_model.save( new_model_fname )
print tcol.OKGREEN+'====DONE===='+tcol.ENDC