eval_onnx.py

#!/usr/bin/env python

import argparse
import math
import numpy as np
import os
import time

import tensorflow.compat.v1 as tf
tf.disable_v2_behavior()

import onnxruntime as ort

parser = argparse.ArgumentParser()
parser.add_argument(
    "--onnx_model", default="inceptionv4_int8.onnx", help="Input onnx model")
parser.add_argument(
    "--val_data_dir",
    default="/workspace/dataset/imagenet/val",
    help="Data directory of validation set")
parser.add_argument(
    "--val_image_list",
    default="/workspace/dataset/imagenet/val.txt",
    help="Validation images list")
parser.add_argument(
    "--subset_len",
    default=50000,
    type=int,
    help="Subset length of validation set to use")
parser.add_argument(
    "--batch_size", default=1, type=int, help="Validation batch size")
parser.add_argument(
    "--ipu",
    action="store_true",
    help="Use IPU for inference.",
)
parser.add_argument(
    "--provider_config",
    type=str,
    default="vaip_config.json",
    help="Path of the config file for seting provider_options.",
)
args = parser.parse_args()

class DataLoader(object):

  def __init__(self, height=224, width=224):
    self.output_height = height
    self.output_width = width

  def _inception_preprocess(self,
                            image,
                            central_fraction=0.875,
                            central_crop=True):
    image = tf.image.convert_image_dtype(image, dtype=tf.float32)
    # Crop the central region of the image with an area containing 87.5% of
    # the original image.
    if central_crop and central_fraction:
      image = tf.image.central_crop(image, central_fraction=central_fraction)

    if self.output_height and self.output_width:
      image = tf.expand_dims(image, 0)
      image = tf.image.resize_bilinear(
          image, [self.output_height, self.output_width], align_corners=False)
    image = tf.subtract(image, 0.5)
    image = tf.multiply(image, 2.0)
    return image

  def _build_placeholder(self):
    input_image_path = tf.placeholder(
        tf.string, shape=(None), name="input_image_path")
    image = tf.io.read_file(input_image_path)
    image = tf.image.decode_jpeg(image, channels=3)
    return image, input_image_path

  def build_preprocess(self):
    """Returns image tensor used to read image."""
    image, input_image_path = self._build_placeholder()
    image = self._inception_preprocess(image)
    return image, input_image_path

def main():
  input_shape = (299, 299, 3)
  label_offset = 0

  with tf.Session() as tf_session:
    loader = DataLoader(input_shape[0], input_shape[1])
    image, image_path = loader.build_preprocess()
    in_image = tf.placeholder(
        tf.float32, shape=(None,) + input_shape, name='in_image')
    in_label = tf.placeholder(tf.int64, shape=(None, 1), name='in_label')
    num_classes = 1001 - label_offset
    logits = tf.placeholder(
        tf.float32, shape=(None, num_classes), name='logits')
    top1, top1_update = tf.metrics.recall_at_k(
        in_label, logits, 1, name="precision_top1")
    top5, top5_update = tf.metrics.recall_at_k(
        in_label, logits, 5, name="precision_top5")

    var_list = tf.get_collection(
        tf.GraphKeys.LOCAL_VARIABLES, scope="precision")
    vars_initializer = tf.variables_initializer(var_list=var_list)
    tf_session.run(vars_initializer)

    with open(args.val_image_list, 'r') as fr:
      lines = fr.readlines()
    if args.subset_len > len(lines):
      raise ValueError(
          "subset_len(%d) should be less or equal than total images(%d)." %
          (args.subset_len, len(lines)))
    eval_steps = math.ceil(args.subset_len / args.batch_size)
    start_t = time.time()

    if args.ipu:
      providers = ["VitisAIExecutionProvider"]
      provider_options = [{"config_file": args.provider_config}]
    else:
      providers = ['CUDAExecutionProvider', 'CPUExecutionProvider']
      provider_options = None
    ort_session = ort.InferenceSession(
        args.onnx_model, providers=providers, provider_options=provider_options)

    for step in range(eval_steps):
      print(f'Eval step {step} / {eval_steps}')
      batch_images = []
      batch_labels = []
      for i in range(args.batch_size):
        index = step * args.batch_size + i
        if index >= args.subset_len:
          break

        img_path, label = lines[index].strip().split(" ")
        img_path = os.path.join(args.val_data_dir, img_path)
        # Run session to get image from feeded image path.
        image_val = tf_session.run(image, feed_dict={image_path: img_path})
        batch_images.append(image_val)

        label = int(label) + 1 - label_offset
        label = np.array([label], dtype=np.int64)
        batch_labels.append(label)

      batch_images = batch_images[0] if args.batch_size == 1 else np.squeeze(
          batch_images)
      ort_inputs = {ort_session.get_inputs()[0].name: batch_images}
      outputs = ort_session.run(None, ort_inputs)

      # Update top1/5 metric.
      tf_session.run([top1_update, top5_update],
                     feed_dict={
                         in_image: batch_images,
                         in_label: batch_labels,
                         logits: outputs[0]
                     })
    end_t = time.time()
    top1_val, top5_val = tf_session.run([top1, top5])
    print('Recall_1 = [%s]' % str(top1_val))
    print('Recall_5 = [%s]' % str(top5_val))
    print('Use_time = [%s]' % str(end_t - start_t))

if __name__ == "__main__":
  main()