- Notifications
You must be signed in to change notification settings - Fork749
Closed
Labels
Description
Environment
- Pythonnet version: master
- Python version: 3.5
- Operating System: Windows 7
Details
- I am calling python function from my c# code . Initially I import python module using Py.Import function . After importing , will pass a List data type of size 299x299x3 to python function as argument . This passing of argument taking 1.5 seconds where as a block of python code takes only 90 msec .
Here is c# code
using(Py.GIL()){dynamicPython_predict=Py.Import("Prediction");Python_predict.__init__(modelpath);List<float>imagelist=GetNormalizedFloatBGRImage(image).Data.Cast<float>().ToList();DateTimet1=System.DateTime.Now;Predicted_Type=Python_predict.Findtype(imagelist);DateTimet2=System.DateTime.Now;Console.WriteLine("Predict timing from c# {0} seconds",(t2.Second+((float)t2.Millisecond/1000))-(t1.Second+((float)t1.Millisecond/1000)));}
Here is python code
def__init__(model_path):kr.model=kr.load_model(model_path)#Model is loaded only oncedefFindtype(ImageData):t1=time.time()N_array=np.array(ImageData,dtype=np.float64)#creates Numpy array from .net List<float> arrayN_array_reshaped=np.reshape(N_array,(299,299,3))#convert BGR to RGBtemp_array[:,:,0]=N_array_reshaped[:,:,2]#bgR -> Rgbtemp_array[:,:,2]=N_array_reshaped[:,:,0]#Bgr -> rgBtemp_array[:,:,1]=N_array_reshaped[:,:,1]#bGr -> rGbimg_list[0]=temp_array#prediction = kr.model.predict(img_list, batch_size=1, verbose=0) # Predictions can be made throughout the lifecycle of MakePrediction class's instance#predicted_class = np.argmax(prediction, axis=1)t2=time.time()print('time on python ',t2-t1)returnpredicted_class[0]+1
So c# code send image array of size 2992993 of float time to python , by which it gets converted into numpy array and makes the prediction on gpu .
if i run the standalone python code , it takes 90msec but if i call from .net 1.5 sec for passing an argument .
Kindly help me to reduce the timing since it is for real time application .
here is a attached screenshots