小 佬 鼠i · 7月15日

【周易AIPU 仿真】部署ResNet V1 152

1. 模型

拉取源码

git clone https://github.com/tensorflow/models.git
Screenshot from 2021-07-15 20-54-02.png
git clone -b v1.15.5 https://github.com/tensorflow/tensorflow.git
Screenshot from 2021-07-15 20-59-20.png
搭建环境:tensorflow==1.15.0 官方的预训练权重,导出pb模型
Screenshot from 2021-07-15 21-05-52.png

2.docker

sudo docker pull zepan/zhouyi
# 链接:https://pan.baidu.com/s/1yaKBPDxR_oakdTnqgyn5fg 
# 提取码:f8dr

3. 准备量化矫正数据集

下载ILSVRC2012\_img\_valScreenshot from 2021-07-15 21-13-43.png

import tensorflow as tf
import numpy as np
import sys
import os
import cv2

sys.path.append('/project/ai/scratch01/salyua01/sharing/guide_acc')
img_dir='./img/'
label_file='./label.txt'

#RESNET PARAM
input_height=224
input_width=224
input_channel = 3
mean = [123.68, 116.78, 103.94]
var = 1

tf.enable_eager_execution()

def smallest_size_at_least(height, width, resize_min):
    """Computes new shape with the smallest side equal to `smallest_side`.

    Computes new shape with the smallest side equal to `smallest_side` while
    preserving the original aspect ratio.

    Args:
      height: an int32 scalar tensor indicating the current height.
      width: an int32 scalar tensor indicating the current width.
      resize_min: A python integer or scalar `Tensor` indicating the size of
        the smallest side after resize.

    Returns:
      new_height: an int32 scalar tensor indicating the new height.
      new_width: an int32 scalar tensor indicating the new width.
    """
    resize_min = tf.cast(resize_min, tf.float32)

    # Convert to floats to make subsequent calculations go smoothly.
    height, width = tf.cast(height, tf.float32), tf.cast(width, tf.float32)

    smaller_dim = tf.minimum(height, width)
    scale_ratio = resize_min / smaller_dim

    # Convert back to ints to make heights and widths that TF ops will accept.
    new_height = tf.cast(tf.round(height * scale_ratio), tf.int32)
    new_width = tf.cast(tf.round(width * scale_ratio), tf.int32)

    return new_height, new_width

def resize_image(image, height, width, method='BILINEAR'):
    """Simple wrapper around tf.resize_images.

    This is primarily to make sure we use the same `ResizeMethod` and other
    details each time.

    Args:
      image: A 3-D image `Tensor`.
      height: The target height for the resized image.
      width: The target width for the resized image.

    Returns:
      resized_image: A 3-D tensor containing the resized image. The first two
        dimensions have the shape [height, width].
    """
    resize_func = tf.image.ResizeMethod.NEAREST_NEIGHBOR if method == 'NEAREST' else tf.image.ResizeMethod.BILINEAR
    return tf.image.resize_images(image, [height, width], method=resize_func, align_corners=False)


def aspect_preserving_resize(image, resize_min, channels=3, method='BILINEAR'):
    """Resize images preserving the original aspect ratio.

    Args:
      image: A 3-D image `Tensor`.
      resize_min: A python integer or scalar `Tensor` indicating the size of
        the smallest side after resize.

    Returns:
      resized_image: A 3-D tensor containing the resized image.
    """
    shape = tf.shape(image)
    height, width = shape[0], shape[1]
    new_height, new_width = smallest_size_at_least(height, width, resize_min)
    return resize_image(image, new_height, new_width, method)


def central_crop(image, crop_height, crop_width, channels=3):
    """Performs central crops of the given image list.

    Args:
      image: a 3-D image tensor
      crop_height: the height of the image following the crop.
      crop_width: the width of the image following the crop.

    Returns:
      3-D tensor with cropped image.
    """
    shape = tf.shape(image)
    height, width = shape[0], shape[1]
    amount_to_be_cropped_h = height - crop_height
    crop_top = amount_to_be_cropped_h // 2
    amount_to_be_cropped_w = width - crop_width
    crop_left = amount_to_be_cropped_w // 2
    # return tf.image.crop_to_bounding_box(image, crop_top, crop_left, crop_height, crop_width)

    size_assertion = tf.Assert(
        tf.logical_and(
            tf.greater_equal(height, crop_height),
            tf.greater_equal(width, crop_width)),
        ['Crop size greater than the image size.']
    )
    with tf.control_dependencies([size_assertion]):
        if channels == 1:
            image = tf.squeeze(image)
            crop_start = [crop_top, crop_left, ]
            crop_shape = [crop_height, crop_width, ]
        elif channels >= 3:
            crop_start = [crop_top, crop_left, 0]
            crop_shape = [crop_height, crop_width, -1]

        image = tf.slice(image, crop_start, crop_shape)

    return tf.reshape(image, [crop_height, crop_width, -1])

label_data = open(label_file)
filename_list = []
label_list = []
for line in label_data:
    filename_list.append(line.rstrip('\n').split(' ')[0])
    label_list.append(int(line.rstrip('\n').split(' ')[1]))
label_data.close()
img_num = len(label_list)

images = np.zeros([img_num, input_height, input_width, input_channel], np.float32)
for file_name, img_idx in zip(filename_list, range(img_num)):
    image_file = os.path.join(img_dir, file_name)
    img_s = tf.gfile.GFile(image_file, 'rb').read()
    image = tf.image.decode_jpeg(img_s)
    image = tf.cast(image, tf.float32)
    image = tf.clip_by_value(image, 0., 255.)
    image = aspect_preserving_resize(image, min(input_height, input_width), input_channel) 
    image = central_crop(image, input_height, input_width)
    image = tf.image.resize_images(image, [input_height, input_width])
    image = (image - mean) / var
    image = image.numpy()
    _, _, ch = image.shape
    if ch == 1:
        image = tf.tile(image, multiples=[1,1,3])
        image = image.numpy()
    images[img_idx] = image

np.save('dataset.npy', images)

labels = np.array(label_list)
np.save('label.npy', labels)

preprocess\_for\_dataset.py生成 dataset.npy和label.npy

4. 编辑NN compiler配置文件

[Common]
mode=run

[Parser]
model_name = resnet_152
detection_postprocess = 
model_domain = image_classification
output = resnet_v1_152/predictions/Reshape_1
input_model = ./model/frozen_resnet_152.pb
input = input
input_shape = [1,224,224,3]
output_dir = ./

[AutoQuantizationTool]
model_name = resnet_152
quantize_method = SYMMETRIC
ops_per_channel = DepthwiseConv
calibration_data = ./dataset/dataset.npy
calibration_label = ./dataset/label.npy
preprocess_mode = normalize
quant_precision=int8
reverse_rgb = False
label_id_offset = 0

[GBuilder]
inputs=./model/input.bin
simulator=aipu_simulator_z1
outputs=output_resnet_152.bin
profile= True
target=Z1_0701

simulator仿真
aipubuild config/resnet_152_run.cfg

[I]     step1: get max/min statistic value DONE
[I]     step2: quantization each op DONE
[I]     step3: build quantization forward DONE
[I]     step4: show output scale of end node:
[I]         layer_id:214, layer_top:resnet_v1_152/predictions/Reshape_1_0, output_scale:[372.5597]
[I] ==== auto-quantization DONE =
[I] Quantize model complete
[I] Building ...
[I] [common_options.h: 276] BuildTool version: 4.0.175. Build for target Z1_0701 at frequency 800MHz
[I] [common_options.h: 297] using default profile events to profile AIFF

[I] [IRChecker] Start to check IR: /tmp/AIPUBuilder_1626355647.7874727/resnet_152_int8.txt
[I] [IRChecker] model_name: resnet_152
[I] [IRChecker] IRChecker: All IR pass
[I] [graph.cpp : 846] loading graph weight: /tmp/AIPUBuilder_1626355647.7874727/resnet_152_int8.bin size: 0x398d9a4
[I] [builder.cpp:1059] Total memory for this graph: 0x4055790 Bytes
[I] [builder.cpp:1060] Text   section:    0x0005ca90 Bytes
[I] [builder.cpp:1061] RO     section:    0x0000a400 Bytes
[I] [builder.cpp:1062] Desc   section:    0x00013800 Bytes
[I] [builder.cpp:1063] Data   section:    0x0396d100 Bytes
[I] [builder.cpp:1064] BSS    section:    0x0062dc00 Bytes
[I] [builder.cpp:1065] Stack         :    0x00040400 Bytes
[I] [builder.cpp:1066] Workspace(BSS):    0x000c4000 Bytes
[I] [main.cpp  : 467] # autogenrated by aipurun, do NOT modify!
LOG_FILE=log_default
FAST_FWD_INST=0
INPUT_INST_CNT=1
INPUT_DATA_CNT=2
CONFIG=Z1-0701
LOG_LEVEL=0
INPUT_INST_FILE0=/tmp/temp_7f213b561036e3c5c72971300849.text
INPUT_INST_BASE0=0x0
INPUT_INST_STARTPC0=0x0
INPUT_DATA_FILE0=/tmp/temp_7f213b561036e3c5c72971300849.ro
INPUT_DATA_BASE0=0x10000000
INPUT_DATA_FILE1=/tmp/temp_7f213b561036e3c5c72971300849.data
INPUT_DATA_BASE1=0x20000000
OUTPUT_DATA_CNT=2
OUTPUT_DATA_FILE0=output_resnet_152.bin
OUTPUT_DATA_BASE0=0x24100000
OUTPUT_DATA_SIZE0=0x3e8
OUTPUT_DATA_FILE1=profile_data.bin
OUTPUT_DATA_BASE1=0x23a71500
OUTPUT_DATA_SIZE1=0x1500
RUN_DESCRIPTOR=BIN[0]

[I] [main.cpp  : 118] run simulator:
aipu_simulator_z1 /tmp/temp_7f213b561036e3c5c72971300849.cfg
[INFO]:SIMULATOR START!
[INFO]:========================================================================
[INFO]:                             STATIC CHECK
[INFO]:========================================================================
[INFO]:  INST START ADDR : 0x0(0)
[INFO]:  INST END ADDR   : 0x5ca8f(379535)
[INFO]:  INST SIZE       : 0x5ca90(379536)
[INFO]:  PACKET CNT      : 0x5ca9(23721)
[INFO]:  INST CNT        : 0x172a4(94884)
[INFO]:------------------------------------------------------------------------
[WARN]:[0803] INST WR/RD REG CONFLICT! PACKET 0x44e: 0x472021b(POP R27,Rc7) vs 0x5f00000(MVI R0,0x0,Rc7), PACKET:0x44e(1102) SLOT:0 vs 3
[WARN]:[0803] INST WR/RD REG CONFLICT! PACKET 0x45b: 0x472021b(POP R27,Rc7) vs 0x5f00000(MVI R0,0x0,Rc7), PACKET:0x45b(1115) SLOT:0 vs 3
[WARN]:[0803] INST WR/RD REG CONFLICT! PACKET 0x5c0: 0x472021b(POP R27,Rc7) vs 0x9f80020(ADD.S R0,R0,0x1,Rc7), PACKET:0x5c0(1472) SLOT:0 vs 3
[WARN]:[0803] INST WR/RD REG CONFLICT! PACKET 0x7fd: 0x4520180(BRL R0) vs 0x47a03e4(ADD R4,R0,R31,Rc7), PACKET:0x7fd(2045) SLOT:0 vs 3
[WARN]:[0803] INST WR/RD REG CONFLICT! PACKET 0x996: 0x4720204(POP R4,Rc7) vs 0x9f80020(ADD.S R0,R0,0x1,Rc7), PACKET:0x996(2454) SLOT:0 vs 3
[WARN]:[0803] INST WR/RD REG CONFLICT! PACKET 0xe40: 0x4720204(POP R4,Rc7) vs 0x47a1be0(ADD R0,R6,R31,Rc7), PACKET:0xe40(3648) SLOT:0 vs 3
[WARN]:[0803] INST WR/RD REG CONFLICT! PACKET 0x1266: 0x472021b(POP R27,Rc7) vs 0x5f00000(MVI R0,0x0,Rc7), PACKET:0x1266(4710) SLOT:0 vs 3
[WARN]:[0803] INST WR/RD REG CONFLICT! PACKET 0x1273: 0x472021b(POP R27,Rc7) vs 0x5f00000(MVI R0,0x0,Rc7), PACKET:0x1273(4723) SLOT:0 vs 3
[WARN]:[0803] INST WR/RD REG CONFLICT! PACKET 0x13d8: 0x472021b(POP R27,Rc7) vs 0x9f80020(ADD.S R0,R0,0x1,Rc7), PACKET:0x13d8(5080) SLOT:0 vs 3
[WARN]:[0803] INST WR/RD REG CONFLICT! PACKET 0x1575: 0x4520180(BRL R0) vs 0x47a03e5(ADD R5,R0,R31,Rc7), PACKET:0x1575(5493) SLOT:0 vs 3
[INFO]:========================================================================
[INFO]:                             STATIC CHECK END
[INFO]:========================================================================

[INFO]:AIPU START RUNNING: BIN[0]
[INFO]:TOTAL TIME: 16.073494s. 
[INFO]:SIMULATOR EXIT!
[I] [main.cpp  : 135] Simulator finished.
Total errors: 0,  warnings: 0

quant\_predict.py

from matplotlib import pyplot as plt
import matplotlib.patches as patches
import numpy as np
import os
import imagenet_classes as class_name

current_dir = os.getcwd()
label_offset = 1
outputfile = current_dir + '/output_resnet_152.bin'
npyoutput = np.fromfile(outputfile, dtype=np.uint8)
outputclass = npyoutput.argmax()
head5p = npyoutput.argsort()[-5:][::-1]

labelfile = current_dir + '/output_ref.bin'
npylabel = np.fromfile(labelfile, dtype=np.int8)
labelclass = npylabel.argmax()
head5t = npylabel.argsort()[-5:][::-1]

print("predict first 5 label:")
for i in head5p:
    print("    index %4d, prob %3d, name: %s"%(i, npyoutput[i], class_name.class_names[i-label_offset]))

print("true first 5 label:")
for i in head5t:
    print("    index %4d, prob %3d, name: %s"%(i, npylabel[i], class_name.class_names[i-label_offset]))

# Show input picture
print('Detect picture save to result.jpeg')

input_path = './model/input.bin'
npyinput = np.fromfile(input_path, dtype=np.int8)
image = np.clip(np.round(npyinput)+128, 0, 255).astype(np.uint8)
image = np.reshape(image, (224, 224, 3))
im = Image.fromarray(image)
im.save('result.jpeg')

quant\_predict.py解析

predict first 5 label:
    index  230, prob 216, name: Old English sheepdog, bobtail
    index  231, prob  39, name: Shetland sheepdog, Shetland sheep dog, Shetland
    index  999, prob   0, name: ear, spike, capitulum
    index  342, prob   0, name: hog, pig, grunter, squealer, Sus scrofa
    index  340, prob   0, name: sorrel
true first 5 label:
    index  230, prob 109, name: Old English sheepdog, bobtail
    index  231, prob  96, name: Shetland sheepdog, Shetland sheep dog, Shetland
    index  232, prob  57, name: collie
    index  226, prob  54, name: malinois
    index  263, prob  53, name: Brabancon griffon
Detect picture save to result.jpeg

result.jpeg
矫正集的data.npy和label.npy $ NN compiler的cfg文件
新人发帖 如有错误 欢迎指正

2 阅读 508
推荐阅读
0 条评论
关注数
1510
内容数
76
人工智能边缘计算软硬件解决方案,提供高性能、低成本、低功耗、易使用的硬件选型方案.
目录
极术微信服务号
关注极术微信号
实时接收点赞提醒和评论通知
Arm中国学堂公众号
关注Arm中国学堂
实时获取免费 Arm 教学资源信息
Arm中国招聘公众号
关注Arm中国招聘
实时获取 Arm 中国职位信息