Created
April 8, 2019 04:27
-
-
Save csvance/e5db387f464463783101093d00393f59 to your computer and use it in GitHub Desktop.
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
import keras | |
from keras.models import Model | |
from keras import backend as K | |
from keras.layers import Dense, Activation | |
import tensorflow as tf | |
from tensorflow.contrib import tensorrt as tftrt | |
import copy | |
import numpy as np | |
import sys | |
import time | |
class FrozenGraph(object): | |
def __init__(self, model, shape): | |
shape = (None, shape[0], shape[1], shape[2]) | |
x_name = 'image_tensor_x' | |
with K.get_session() as sess: | |
x_tensor = tf.placeholder(tf.float32, shape, x_name) | |
K.set_learning_phase(0) | |
y_tensor = model(x_tensor) | |
y_name = y_tensor.name[:-2] | |
graph = sess.graph.as_graph_def() | |
graph0 = tf.graph_util.convert_variables_to_constants(sess, graph, [y_name]) | |
graph1 = tf.graph_util.remove_training_nodes(graph0) | |
self.x_name = [x_name] | |
self.y_name = [y_name] | |
self.frozen = graph1 | |
class TfEngine(object): | |
def __init__(self, graph): | |
g = tf.Graph() | |
with g.as_default(): | |
x_op, y_op = tf.import_graph_def( | |
graph_def=graph.frozen, return_elements=graph.x_name + graph.y_name) | |
self.x_tensor = x_op.outputs[0] | |
self.y_tensor = y_op.outputs[0] | |
config = tf.ConfigProto(gpu_options= | |
tf.GPUOptions(per_process_gpu_memory_fraction=0.5, | |
allow_growth=True)) | |
self.sess = tf.Session(graph=g, config=config) | |
def infer(self, x): | |
y = self.sess.run(self.y_tensor, | |
feed_dict={self.x_tensor: x}) | |
return y | |
class TftrtEngine(TfEngine): | |
def __init__(self, graph, batch_size, precision): | |
tftrt_graph = tftrt.create_inference_graph( | |
graph.frozen, | |
outputs=graph.y_name, | |
max_batch_size=batch_size, | |
max_workspace_size_bytes=1 << 30, | |
precision_mode=precision, | |
minimum_segment_size=2) | |
self.tftrt_graph = tftrt_graph | |
opt_graph = copy.deepcopy(graph) | |
opt_graph.frozen = tftrt_graph | |
super(TftrtEngine, self).__init__(opt_graph) | |
self.batch_size = batch_size | |
def infer(self, x): | |
num_tests = x.shape[0] | |
y = np.empty((num_tests, 1), np.float32) | |
batch_size = self.batch_size | |
for i in range(0, num_tests, batch_size): | |
x_part = x[i: i + batch_size] | |
y_part = self.sess.run(self.y_tensor, | |
feed_dict={self.x_tensor: x_part}) | |
y[i: i + batch_size] = y_part | |
return y | |
mobilenet = keras.applications.mobilenet.MobileNet(input_shape=(224, 224, 3), weights='imagenet', pooling='max') | |
new_output = mobilenet.get_layer('global_average_pooling2d_1').output | |
new_output = Dense(1, activation='softmax')(new_output) | |
model = Model(inputs=mobilenet.input, outputs=new_output) | |
model.summary() | |
frozen_graph = FrozenGraph(model, (224, 224, 3)) | |
x_test = np.random.random((10, 224, 224, 3)) | |
tf_engine = TfEngine(frozen_graph) | |
t0 = time.time() | |
y_tf = tf_engine.infer(x_test) | |
print(y_tf) | |
t1 = time.time() | |
print('Tensorflow time', t1 - t0) | |
batch_size = 1 | |
tftrt_engine = TftrtEngine(frozen_graph, batch_size, 'FP32') | |
t0 = time.time() | |
y_tftrt = tftrt_engine.infer(x_test) | |
print(y_tftrt) | |
t1 = time.time() | |
print('TFTRT time', t1 - t0) | |
tftrt_engine = TftrtEngine(frozen_graph, batch_size, 'FP16') | |
t0 = time.time() | |
y_tftrt = tftrt_engine.infer(x_test) | |
print(y_tftrt) | |
t1 = time.time() | |
print('TFTRT_FP16 time', t1 - t0) | |
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment