Don't know how to inference

hello @snosov1 @alexey-trushkov @alexey-sidnev AlexanderDokuchaev and everyone, Thank for your code. I have already trained and exported to frozen model (.pb) without optimize. Then i use frozen model for tensorrt inference server. but I got the wrong result compared to using infer.py file (checkpoints)

I don’t know if my export .pb file is incorrect or my clients code are wrong

#### MY CLIENTS CODE #####
import argparse
import os
import random
import cv2
import numpy as np
import tensorflow as tf
import tensorflow.contrib.slim as slim
from builtins import range
from tensorrtserver.api import *
import tensorrtserver.api.model_config_pb2 as model_config


if __name__ == '__main__':
    input_name = 'input'
    output_name = 'd_predictions'
    model_name = 'lp-recognitor'
    protocol = ProtocolType.from_str('http')
    ctx = InferContext('localhost:8000', protocol, model_name, -1, False)

    img = cv2.imread("/data/tmp/plate/000111.png")
    img = np.float32(img)
    img = cv2.resize(img, (24, 94))

    in_frame = img.reshape((24, 94, 3)) 

    
    input_data = []
    input_data.append(in_frame) 

    results = [] 
    results.append(ctx.run(
            { input_name : input_data },
            { output_name : (InferContext.ResultFormat.RAW) }))
    print("****************results*********************", results)

[AlexanderDokuchaev]

Hi @tienduchoang, You should add preprocessing for input data in your code

    img = cv2.resize(image, (94, 24))
    img = cv2.cvtColor(img, cv2.COLOR_BGR2RGB)
    img = np.float32(img)
    img = np.multiply(img, 1.0/255.0)

For IE model it done by arguments for mo.py.

[snosov1]

I’ve heard from @AlexanderDokuchaev that he sees the same (strange) behavior. We’ll take a closer look and let you know when we have something.