def run_numpy_pb():
args = parse_arg()
config = tf.ConfigProto(allow_soft_placement=True,
inter_op_parallelism_threads=1,
intra_op_parallelism_threads=1)
config.mlu_options.save_offline_model = True
model_name = os.path.basename(args.numpy_pb).split(".")[0]
image_name = os.path.basename(args.image).split(".")[0]
g = tf.Graph()
with g.as_default():
with tf.gfile.FastGFile(args.numpy_pb,'rb') as f:
graph_def = tf.GraphDef()
graph_def.ParseFromString(f.read())
tf.import_graph_def(graph_def, name='')
img = cv.imread(args.image)
X = cv.resize(img, (256, 256))
with tf.Session(config=config) as sess:
sess.graph.as_default()
sess.run(tf.global_variables_initializer())
input_tensor1 = sess.graph.get_tensor_by_name('X_content:0')
input_tensor2 = sess.graph.get_tensor_by_name('moments_15/PowerDifference:0')
output_tensor = sess.graph.get_tensor_by_name('add_37:0')
start_time = time.time()
input_2 = power_diff_numpy(sess.graph.get_tensor_by_name('Conv2D_13:0').eval(feed_dict={input_tensor1:[X]}),sess.graph.get_tensor_by_name('moments_15/StopGradient:0').eval(feed_dict={input_tensor1:[X]}),2)
ret = sess.run(output_tensor, feed_dict={input_tensor1:[X], input_tensor2:input_2})
end_time = time.time()
print("Numpy inference(MLU) time is: ",end_time-start_time)
img1 = tf.reshape(ret,[256,256,3])
img_numpy = img1.eval(session=sess)
cv.imwrite(image_name + '_' + model_name + '_mlu.jpg',img_numpy)
def main():
value = 256
start = time.time()
input_x = np.loadtxt("../data/in_x.txt")
input_y = np.loadtxt("../data/in_y.txt")
input_pow = np.loadtxt("../data/in_z.txt")
output = np.loadtxt("../data/out.txt")
end = time.time()
print("load data cost " + str((end - start)*1000) + "ms")
input_x = np.reshape(input_x,(1,value,value,3))
input_y = np.reshape(input_y,(1,1,1,3))
output = np.reshape(output, (-1))
start = time.time()
res = power_difference_op(input_x, input_y, input_pow)
end = time.time() - start
print("comput C++ op cost "+ str(end*1000) + "ms" )
res = np.reshape(res,(-1))
err = sum(abs(res - output))/sum(output)
print("C++ op err rate= " + str(err*100))
start = time.time()
res = power_diff_numpy(input_x, input_y,input_pow)
end = time.time()
print("comput numpy op cost "+ str((end-start)*1000) + "ms")
res = np.reshape(res,(-1))
err = sum(abs(res - output))/sum(output)
print("numpy op err rate= "+ str(err*100))
def main():
value = 256
input_x = np.random.randn(1, value, value, 3)
input_y = np.random.randn(1, 1, 1, 3)
input_pow = np.zeros((2))
start = time.time()
res = power_difference_op(input_x, input_y, input_pow)
end = time.time() - start
print("comput BCL op cost " + str(end * 1000) + "ms")
start = time.time()
output = power_diff_numpy(input_x, input_y, len(input_pow))
end = time.time()
res = np.reshape(res, (-1))
output = np.reshape(output, (-1))
print("comput op cost " + str((end - start) * 1000) + "ms")
err = sum(abs(res - output)) / sum(output)
print("err rate= " + str(err * 100))
def run_numpy_pb(numpy_pb, image):
config = tf.ConfigProto(allow_soft_placement=True,
inter_op_parallelism_threads=1,
intra_op_parallelism_threads=1)
model_name = os.path.basename(numpy_pb).split(".")[0]
image_name = os.path.basename(image).split(".")[0]
g = tf.Graph()
with g.as_default():
with tf.gfile.FastGFile(numpy_pb,'rb') as f:
graph_def = tf.GraphDef()
graph_def.ParseFromString(f.read())
tf.import_graph_def(graph_def, name='')
img = cv.imread(image)
X = cv.resize(img, (256, 256))
with tf.Session(config=config) as sess:
# TODO：完成Numpy版本 Pb模型的推理
sess.run(tf.global_variables_initializer())
# 根据输入名称获得输入的tensor
input_tensor = sess.graph.get_tensor_by_name('X_content:0')
# 获取两个输出节点， 作为numpy算子的输入
out_tmp_tensor_1 = sess.graph.get_tensor_by_name('Conv2D_13:0')
out_tmp_tensor_2 = sess.graph.get_tensor_by_name('moments_15/StopGradient:0')
start_time = time.time()
# 执行第一次sessionrun， 得到numpy算子的两个输入值,注意此时两个输入的shape不同
input_x, input_y = sess.run([out_tmp_tensor_1, out_tmp_tensor_2], feed_dict={input_tensor:[X]})
input_pow = 2
output = power_diff_numpy(input_x, input_y, input_pow).reshape(1, 256, 256, 3)
# 根据添加的输入节点名称获得输入tensor
input_tensor_new = sess.graph.get_tensor_by_name('moments_15/PowerDifference:0')
# 完整推理最终输出的tensor
output_tensor = sess.graph.get_tensor_by_name('add_37:0')
# 执行第二次session run，输入图片数据以及上一步骤numpy计算的数据
ret = sess.run(output_tensor, feed_dict={input_tensor:[X], input_tensor_new: output})
end_time = time.time()
print("Numpy inference(CPU) time is: ",end_time-start_time)
img1 = tf.reshape(ret,[256,256,3])
img_numpy = img1.eval(session=sess)
cv.imwrite(image_name + '_' + model_name + '_cpu.jpg',img_numpy)