车道线任务简介

车道线检测是无人驾驶、地图导航场景中一个基础任务。比如，经典的(类似)行车记录仪视角：

https://ai.bdstatic.com/file/AA5AEBF1ED2C4867B5BFA3A699E96012

In [1]:
# 视频来源 https://github.com/udacity/CarND-Advanced-Lane-Lines/blob/master/project_video.mp4
import IPython
IPython.display.Video('project_video.mp4')
Out[1]:

当然，也有高空俯瞰视角：[第十六届全国大学生智能车竞赛线上资格赛：车道线检测](https://aistudio.baidu.com/aistudio/competition/detail/68) file

还有认为通过图像分割的方式，检测车道线速度太慢，对其进行改进的：[Ultra-Fast-Lane-Detection](https://arxiv.org/abs/2004.11757)

车道线检测相关数据集

BDD100K: A Large-scale Diverse Driving Video Database

2018年5月伯克利大学AI实验室（BAIR）发布了的大规模、内容多样的公开驾驶数据集，设计了一个图片标注系统。BDD100K 数据集包含10万段高清视频，每个视频约40秒\720p\30 fps 。每个视频的第10秒对关键帧进行采样，得到10万张图片（图片尺寸：1280*720），并进行标注。车道线检测的标注是其中一种。 http://bair.berkeley.edu/static/blog/bdd/lane_markings.png

ApolloScape

此大型数据集包含一组视频序列，记录在不同城市街道的自带驾驶场景，包括11万多帧的高质量像素级标注。 file

参考资料

车道线检测基础预处理技巧

图片裁剪

参考项目从头搭建无人车车道线检测挑战赛解决方案中提到：

通过仔细观察，我们发现这些数据有一个共同的特点，就是图片的上三分之一部分都是天空，是没有车道线存在的，知道了这点后，我们就可以进行一个裁剪的过程，一下子就可以节省下三分之一的显存，是不是很爽呢？这里我选择裁剪上方690个像素的高度。

在很多车道线检测的数据集中，这种现象普遍存在，所以截去天空再训练这类数据是一个普遍技巧，省下内存、显存空间。一般的做法是直接裁剪掉图片的上半部分或三分之一。

In [26]:
from PIL import Image, ImageFilter
import cv2
import numpy as np
def crop_data(input_path,output_path):
    img = Image.open(input_path)
    print(img.size)
    # 裁掉图片的上半部分
    cropped = img.crop((0, (img.height/2), img.width, img.height))  # (left, upper, right, lower)
    print(cropped.size)
    cropped.save(output_path)
In [28]:
crop_data('./crop_input.png','./crop_output.png')

图片锐化

In [ ]:
def sharp_data(input_path,output_path):
    img = Image.open(input_path)
    print(img.size)
    # 两次锐化
    img = img.filter(ImageFilter.SHARPEN)  
    img = img.filter(ImageFilter.SHARPEN)  
    img.save(output_path)
In [ ]:
sharp_data('./crop_input.png','./sharp_output.png')

(431, 241)

file file

Canny边缘提取

In [ ]:
img = cv2.imread('crop_input.png', cv2.COLOR_BGR2GRAY)
In [ ]:
low_threshold = 40
high_threshold = 150
canny_image = cv2.Canny(img, low_threshold, high_threshold)
In [ ]:
cv2.imwrite('canny.jpg', canny_image)
Out[ ]:
True

file

参考资料

使用PaddleSeg进行车道线分割

解压数据集

In [6]:
!unzip data/data68698/智能车数据集.zip
In [ ]:
# !git clone https://gitee.com/paddlepaddle/PaddleSeg.git

Cloning into 'PaddleSeg'...
remote: Enumerating objects: 10912, done.
remote: Counting objects: 100% (10912/10912), done.
remote: Compressing objects: 100% (5400/5400), done.
remote: Total 10912 (delta 7379), reused 8171 (delta 5353), pack-reused 0
Receiving objects: 100% (10912/10912), 156.94 MiB | 21.09 MiB/s, done.
Resolving deltas: 100% (7379/7379), done.
Checking connectivity... done.

准备数据集

PaddleSeg目前支持CityScapes、ADE20K、Pascal VOC等数据集的加载，在加载数据集时，如若本地不存在对应数据，则会自动触发下载(除Cityscapes数据集)。 这里可以直接使用比赛提供的脚本

In [7]:
%run make_list.py
# !python make_list.py

[('image_4000/3199.png', 'mask_4000/3199.png'), ('image_4000/3590.png', 'mask_4000/3590.png'), ('image_4000/3685.png', 'mask_4000/3685.png')]
4000
In [1]:
%set_env CUDA_VISIBLE_DEVICES=0

env: CUDA_VISIBLE_DEVICES=0

分析数据集

@BIT可达鸭这位大佬在【魔改版】全国大学生智能车竞赛-车道线检测Baseline项目中对数据集进行了分析，并尝试均衡化数据集。我们先直接看下数据集各类别的统计结果。

-[INFO] Label 1: 0.0170
-[INFO] Label 2: 0.4482
-[INFO] Label 3: 0.0843
-[INFO] Label 4: 0.0767
-[INFO] Label 5: 0.0334
-[INFO] Label 6: 0.2513
-[INFO] Label 7: 0.0070
-[INFO] Label 8: 0.0025
-[INFO] Label 9: 0.0158
-[INFO] Label 10: 0.0152
-[INFO] Label 11: 0.0292
-[INFO] Label 12: 0.0087
-[INFO] Label 13: 0.0061
-[INFO] Label 14: 0.0046
-[INFO] Label 15: 0.0000

也可以参考这篇文章[统计图像分割训练集中的类别分布 by @yhl_leo](http://blog.csdn.net/yhl_leo/article/details/52225600),由于计算时间过长，这里写了一个脚本任务，可以作为统计图像分割类别分布的一个模板：[图像分割训练集类别分布统计脚本：智能车竞赛-车道线检测](https://aistudio.baidu.com/aistudio/clusterprojectdetail/1759967)

In [ ]:
import cv2, os
import numpy as np

#amount of classer
CLASSES_NUM = 15

#find imagee in folder dir
def findImages(dir,topdown=True):
    im_list = []
    if not os.path.exists(dir):
        print("Path for {} not exist!".format(dir))
        raise
    else:
        for root, dirs, files in os.walk(dir, topdown):
            for fl in files:
                im_list.append(fl)
    return im_list

# amount of images corresponding to each classes
images_count = [0]*CLASSES_NUM
# amount of pixels corresponding to each class
class_pixels_count = [0]*CLASSES_NUM
# amount of pixels corresponding to the images of each class
image_pixels_count = [0]*CLASSES_NUM

image_folder = './mask_4000'
im_list = findImages(image_folder) 

for im in im_list:
    print(im)
    cv_img = cv2.imread(os.path.join(image_folder, im), cv2.IMREAD_UNCHANGED)
    size_img = cv_img.shape
    colors = set([])
    for i in range(size_img[0]):
        for j in range(size_img[1]):
            p_value = cv_img.item(i,j)
            if not p_value < CLASSES_NUM: # check
                print(p_value)
            else:
                class_pixels_count[p_value] = class_pixels_count[p_value] + 1
                colors.add(p_value)
    im_size = size_img[0]*size_img[1]
    for n in range(CLASSES_NUM):
        if n in colors:
            images_count[n] = images_count[n] + 1
            image_pixels_count[n] = image_pixels_count[n] + im_size

print(images_count)
print(class_pixels_count)
print(image_pixels_count)

计算每个类别的像素点数量占比

[4000, 457, 3437, 803, 481, 850, 2997, 118, 32, 182, 213, 205, 60, 24, 33]
[9284543967, 2595404, 68411608, 12860131, 11705492, 5098954, 38357780, 1066241, 383440, 2419091, 2312722, 4459757, 1332924, 927368, 709121]
[9437184000, 1078198272, 8108900352, 1894514688, 1134821376, 2005401600, 7070810112, 278396928, 75497472, 429391872, 502530048, 483655680, 141557760, 56623104, 77856768]
In [87]:
t = [9284543967, 2595404, 68411608, 12860131, 11705492, 5098954, 38357780, 1066241, 383440, 2419091, 2312722, 4459757, 1332924, 927368, 709121]
In [88]:
a = np.array(t).sum()
In [89]:
for i in range(len(t)):
    t[i] = round(t[i]/a, 4)
In [90]:
print(t)

[0.9838, 0.0003, 0.0072, 0.0014, 0.0012, 0.0005, 0.0041, 0.0001, 0.0, 0.0003, 0.0002, 0.0005, 0.0001, 0.0001, 0.0001]

类别不均衡的处理

在图像分割任务中，经常出现类别分布不均匀的情况，例如：工业产品的瑕疵检测、道路提取及病变区域提取等。

针对这个问题，PaddleSeg主要提供了带权的softmax loss和Lovasz loss两种解决方案。

Weighted softmax loss

Weighted softmax loss是按类别设置不同权重的softmax loss。

通过设置cfg.SOLVER.CROSS_ENTROPY_WEIGHT参数进行使用。
默认为None. 如果设置为'dynamic'，会根据每个batch中各个类别的数目，动态调整类别权重。 也可以设置一个静态权重(list的方式)，比如有3类，每个类别权重可以设置为[0.1, 2.0, 0.9]。

SOLVER:
    LR: 0.005
    LR_POLICY: "poly"
    OPTIMIZER: "sgd"
    NUM_EPOCHS: 40
    CROSS_ENTROPY_WEIGHT: [0.1, 2, 0.5, 2, 2, 2, 1, 2, 2, 2, 2, 2, 2 ,2, 2] #每一个类别loss权重

Lovasz loss

Lovasz loss基于子模损失(submodular losses)的凸Lovasz扩展，对神经网络的mean IoU损失进行优化。Lovasz loss根据分割目标的类别数量可分为两种：lovasz hinge loss和lovasz softmax loss. 其中lovasz hinge loss适用于二分类问题，lovasz softmax loss适用于多分类问题。该工作发表在CVPR 2018上，可点击参考文献查看具体原理。

需要注意的是，通常的直接训练方式不一定管用，PaddleSeg推荐的是另外2种训练方式：

  • （1）与cross entropy loss或bce loss(binary cross-entropy loss)加权结合使用。
  • （2）先使用cross entropy loss或bce loss进行训练，再使用lovasz softmax loss或lovasz hinge loss进行finetuning。 通过coef参数对不同loss进行权重配比，从而灵活地进行训练调参。
loss:
  types:
    - type: MixedLoss
      losses:
        - type: CrossEntropyLoss
        - type: LovaszSoftmaxLoss
      coef: [0.8, 0.2]
    - type: MixedLoss
      losses:
        - type: CrossEntropyLoss
        - type: LovaszSoftmaxLoss
      coef: [0.8, 0.2]
  coef: [1, 0.4]

SOLVER:
    LR: 0.005
    LR_POLICY: "poly"
    OPTIMIZER: "sgd"
    NUM_EPOCHS: 40
    CROSS_ENTROPY_WEIGHT: dynamic
In [8]:
%cd PaddleSeg

/home/aistudio/PaddleSeg

训练

In [29]:
!python train.py \
       --config configs/ocrnet/ocrnet_hrnetw18_cityscapes_1024x512_160k_lovasz_softmax.yml \
       --do_eval \
       --use_vdl \
       --save_interval 1000 \
       --save_dir output

In [18]:
!python train.py \
       --config configs/ocrnet/ocrnet_hrnetw18_cityscapes_1024x512_160k_lovasz_softmax.yml \
       --resume_model output/iter_6000 \
       --do_eval \
       --use_vdl \
       --save_interval 1000 \
       --save_dir output

/opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages/setuptools/depends.py:2: DeprecationWarning: the imp module is deprecated in favour of importlib; see the module's documentation for alternative uses
  import imp
2021-04-08 10:13:26 [INFO]	
------------Environment Information-------------
platform: Linux-4.4.0-150-generic-x86_64-with-debian-stretch-sid
Python: 3.7.4 (default, Aug 13 2019, 20:35:49) [GCC 7.3.0]
Paddle compiled with cuda: True
NVCC: Cuda compilation tools, release 10.1, V10.1.243
cudnn: 7.6
GPUs used: 1
CUDA_VISIBLE_DEVICES: 0
GPU: ['GPU 0: Tesla V100-SXM2-32GB']
GCC: gcc (Ubuntu 7.5.0-3ubuntu1~16.04) 7.5.0
PaddlePaddle: 2.0.1
OpenCV: 4.1.1
------------------------------------------------
2021-04-08 10:13:26 [INFO]	
---------------Config Information---------------
SOLVER:
  CROSS_ENTROPY_WEIGHT: dynamic
  LR: 0.005
  LR_POLICY: poly
  NUM_EPOCHS: 40
  OPTIMIZER: sgd
batch_size: 4
iters: 35000
learning_rate:
  decay:
    end_lr: 0.0
    power: 0.9
    type: poly
  value: 0.0025
loss:
  coef:
  - 1
  - 0.4
  types:
  - coef:
    - 0.8
    - 0.2
    losses:
    - type: CrossEntropyLoss
    - type: LovaszSoftmaxLoss
    type: MixedLoss
  - coef:
    - 0.8
    - 0.2
    losses:
    - type: CrossEntropyLoss
    - type: LovaszSoftmaxLoss
    type: MixedLoss
model:
  backbone:
    pretrained: https://bj.bcebos.com/paddleseg/dygraph/hrnet_w18_ssld.tar.gz
    type: HRNet_W18
  backbone_indices:
  - 0
  type: OCRNet
optimizer:
  momentum: 0.9
  type: sgd
  weight_decay: 4.0e-05
train_dataset:
  dataset_root: /home/aistudio/
  mode: train
  num_classes: 15
  train_path: /home/aistudio/train_list.txt
  transforms:
  - max_scale_factor: 2.0
    min_scale_factor: 0.5
    scale_step_size: 0.25
    type: ResizeStepScaling
  - max_rotation: 30
    type: RandomRotation
  - type: RandomHorizontalFlip
  - type: RandomVerticalFlip
  - crop_size:
    - 1024
    - 512
    type: RandomPaddingCrop
  - type: RandomBlur
  - brightness_range: 0.4
    contrast_range: 0.4
    saturation_range: 0.4
    type: RandomDistort
  - type: Normalize
  type: Dataset
val_dataset:
  dataset_root: /home/aistudio/
  mode: val
  num_classes: 15
  transforms:
  - type: Normalize
  type: Dataset
  val_path: /home/aistudio/val_list.txt
------------------------------------------------
W0408 10:13:26.278049 10398 device_context.cc:362] Please NOTE: device: 0, GPU Compute Capability: 7.0, Driver API Version: 10.1, Runtime API Version: 10.1
W0408 10:13:26.278107 10398 device_context.cc:372] device: 0, cuDNN Version: 7.6.
2021-04-08 10:13:31 [INFO]	Loading pretrained model from https://bj.bcebos.com/paddleseg/dygraph/hrnet_w18_ssld.tar.gz
2021-04-08 10:13:31,306 - INFO - Lock 140184385978832 acquired on /home/aistudio/.paddleseg/tmp/hrnet_w18_ssld
2021-04-08 10:13:31,307 - INFO - Lock 140184385978832 released on /home/aistudio/.paddleseg/tmp/hrnet_w18_ssld
2021-04-08 10:13:32 [INFO]	There are 1525/1525 variables loaded into HRNet.
2021-04-08 10:13:32 [INFO]	Resume model from output/iter_6000
2021-04-08 10:13:45 [INFO]	[TRAIN] epoch=7, iter=6010/35000, loss=0.2093, lr=0.002110, batch_cost=1.1784, reader_cost=0.10351, ips=3.3943 samples/sec | ETA 09:29:23
2021-04-08 10:13:55 [INFO]	[TRAIN] epoch=7, iter=6020/35000, loss=0.2493, lr=0.002110, batch_cost=0.9857, reader_cost=0.00008, ips=4.0580 samples/sec | ETA 07:56:05
2021-04-08 10:14:05 [INFO]	[TRAIN] epoch=7, iter=6030/35000, loss=0.2214, lr=0.002109, batch_cost=1.0455, reader_cost=0.00009, ips=3.8259 samples/sec | ETA 08:24:48
2021-04-08 10:14:16 [INFO]	[TRAIN] epoch=7, iter=6040/35000, loss=0.2016, lr=0.002108, batch_cost=1.0664, reader_cost=0.00008, ips=3.7509 samples/sec | ETA 08:34:43
2021-04-08 10:14:27 [INFO]	[TRAIN] epoch=7, iter=6050/35000, loss=0.2157, lr=0.002108, batch_cost=1.0611, reader_cost=0.00009, ips=3.7695 samples/sec | ETA 08:32:00
2021-04-08 10:14:38 [INFO]	[TRAIN] epoch=7, iter=6060/35000, loss=0.1924, lr=0.002107, batch_cost=1.1031, reader_cost=0.00012, ips=3.6261 samples/sec | ETA 08:52:03
2021-04-08 10:14:48 [INFO]	[TRAIN] epoch=7, iter=6070/35000, loss=0.2162, lr=0.002106, batch_cost=1.0151, reader_cost=0.00010, ips=3.9404 samples/sec | ETA 08:09:27
2021-04-08 10:14:58 [INFO]	[TRAIN] epoch=7, iter=6080/35000, loss=0.1959, lr=0.002106, batch_cost=1.0447, reader_cost=0.00010, ips=3.8288 samples/sec | ETA 08:23:32
2021-04-08 10:15:09 [INFO]	[TRAIN] epoch=7, iter=6090/35000, loss=0.2184, lr=0.002105, batch_cost=1.0520, reader_cost=0.00010, ips=3.8022 samples/sec | ETA 08:26:53
2021-04-08 10:15:20 [INFO]	[TRAIN] epoch=7, iter=6100/35000, loss=0.2238, lr=0.002104, batch_cost=1.1000, reader_cost=0.00008, ips=3.6363 samples/sec | ETA 08:49:50
2021-04-08 10:15:30 [INFO]	[TRAIN] epoch=7, iter=6110/35000, loss=0.2002, lr=0.002104, batch_cost=0.9978, reader_cost=0.00008, ips=4.0089 samples/sec | ETA 08:00:25
2021-04-08 10:15:41 [INFO]	[TRAIN] epoch=7, iter=6120/35000, loss=0.2047, lr=0.002103, batch_cost=1.0724, reader_cost=0.00008, ips=3.7300 samples/sec | ETA 08:36:10
2021-04-08 10:15:51 [INFO]	[TRAIN] epoch=8, iter=6130/35000, loss=0.2244, lr=0.002102, batch_cost=1.0544, reader_cost=0.00008, ips=3.7935 samples/sec | ETA 08:27:21
2021-04-08 10:16:02 [INFO]	[TRAIN] epoch=8, iter=6140/35000, loss=0.2038, lr=0.002102, batch_cost=1.0655, reader_cost=0.00011, ips=3.7541 samples/sec | ETA 08:32:30
2021-04-08 10:16:13 [INFO]	[TRAIN] epoch=8, iter=6150/35000, loss=0.2257, lr=0.002101, batch_cost=1.0989, reader_cost=0.00010, ips=3.6400 samples/sec | ETA 08:48:23
2021-04-08 10:16:23 [INFO]	[TRAIN] epoch=8, iter=6160/35000, loss=0.2056, lr=0.002100, batch_cost=0.9964, reader_cost=0.00008, ips=4.0143 samples/sec | ETA 07:58:57
2021-04-08 10:16:33 [INFO]	[TRAIN] epoch=8, iter=6170/35000, loss=0.2248, lr=0.002100, batch_cost=1.0123, reader_cost=0.00010, ips=3.9515 samples/sec | ETA 08:06:23
2021-04-08 10:16:43 [INFO]	[TRAIN] epoch=8, iter=6180/35000, loss=0.2562, lr=0.002099, batch_cost=1.0206, reader_cost=0.00008, ips=3.9192 samples/sec | ETA 08:10:14
2021-04-08 10:16:53 [INFO]	[TRAIN] epoch=8, iter=6190/35000, loss=0.1847, lr=0.002098, batch_cost=1.0048, reader_cost=0.00008, ips=3.9809 samples/sec | ETA 08:02:28
2021-04-08 10:17:03 [INFO]	[TRAIN] epoch=8, iter=6200/35000, loss=0.2059, lr=0.002098, batch_cost=1.0068, reader_cost=0.00008, ips=3.9731 samples/sec | ETA 08:03:14
2021-04-08 10:17:13 [INFO]	[TRAIN] epoch=8, iter=6210/35000, loss=0.1989, lr=0.002097, batch_cost=1.0177, reader_cost=0.00008, ips=3.9304 samples/sec | ETA 08:08:19
2021-04-08 10:17:24 [INFO]	[TRAIN] epoch=8, iter=6220/35000, loss=0.2231, lr=0.002096, batch_cost=1.0990, reader_cost=0.00014, ips=3.6396 samples/sec | ETA 08:47:09
2021-04-08 10:17:35 [INFO]	[TRAIN] epoch=8, iter=6230/35000, loss=0.2176, lr=0.002096, batch_cost=1.1169, reader_cost=0.00010, ips=3.5813 samples/sec | ETA 08:55:33
2021-04-08 10:17:46 [INFO]	[TRAIN] epoch=8, iter=6240/35000, loss=0.2173, lr=0.002095, batch_cost=1.0865, reader_cost=0.00008, ips=3.6816 samples/sec | ETA 08:40:47
2021-04-08 10:17:57 [INFO]	[TRAIN] epoch=8, iter=6250/35000, loss=0.1872, lr=0.002094, batch_cost=1.0429, reader_cost=0.00008, ips=3.8353 samples/sec | ETA 08:19:44
2021-04-08 10:18:08 [INFO]	[TRAIN] epoch=8, iter=6260/35000, loss=0.2105, lr=0.002094, batch_cost=1.1581, reader_cost=0.00010, ips=3.4539 samples/sec | ETA 09:14:44
2021-04-08 10:18:18 [INFO]	[TRAIN] epoch=8, iter=6270/35000, loss=0.1888, lr=0.002093, batch_cost=0.9761, reader_cost=0.00008, ips=4.0978 samples/sec | ETA 07:47:24
2021-04-08 10:18:28 [INFO]	[TRAIN] epoch=8, iter=6280/35000, loss=0.2201, lr=0.002092, batch_cost=0.9710, reader_cost=0.00008, ips=4.1197 samples/sec | ETA 07:44:45
2021-04-08 10:18:39 [INFO]	[TRAIN] epoch=8, iter=6290/35000, loss=0.2116, lr=0.002092, batch_cost=1.0923, reader_cost=0.00008, ips=3.6618 samples/sec | ETA 08:42:41
2021-04-08 10:18:49 [INFO]	[TRAIN] epoch=8, iter=6300/35000, loss=0.2101, lr=0.002091, batch_cost=1.0442, reader_cost=0.00008, ips=3.8306 samples/sec | ETA 08:19:29
2021-04-08 10:19:00 [INFO]	[TRAIN] epoch=8, iter=6310/35000, loss=0.2121, lr=0.002090, batch_cost=1.0931, reader_cost=0.00008, ips=3.6594 samples/sec | ETA 08:42:40
2021-04-08 10:19:10 [INFO]	[TRAIN] epoch=8, iter=6320/35000, loss=0.2183, lr=0.002090, batch_cost=1.0179, reader_cost=0.00008, ips=3.9298 samples/sec | ETA 08:06:32
2021-04-08 10:19:21 [INFO]	[TRAIN] epoch=8, iter=6330/35000, loss=0.2113, lr=0.002089, batch_cost=1.0378, reader_cost=0.00008, ips=3.8544 samples/sec | ETA 08:15:52
2021-04-08 10:19:31 [INFO]	[TRAIN] epoch=8, iter=6340/35000, loss=0.2115, lr=0.002089, batch_cost=1.0446, reader_cost=0.00011, ips=3.8293 samples/sec | ETA 08:18:57
2021-04-08 10:19:42 [INFO]	[TRAIN] epoch=8, iter=6350/35000, loss=0.2100, lr=0.002088, batch_cost=1.1134, reader_cost=0.00012, ips=3.5925 samples/sec | ETA 08:51:39
2021-04-08 10:19:53 [INFO]	[TRAIN] epoch=8, iter=6360/35000, loss=0.1888, lr=0.002087, batch_cost=1.0259, reader_cost=0.00008, ips=3.8989 samples/sec | ETA 08:09:42
2021-04-08 10:20:03 [INFO]	[TRAIN] epoch=8, iter=6370/35000, loss=0.2396, lr=0.002087, batch_cost=1.0202, reader_cost=0.00008, ips=3.9208 samples/sec | ETA 08:06:48
2021-04-08 10:20:13 [INFO]	[TRAIN] epoch=8, iter=6380/35000, loss=0.2279, lr=0.002086, batch_cost=1.0538, reader_cost=0.00009, ips=3.7957 samples/sec | ETA 08:22:40
2021-04-08 10:20:24 [INFO]	[TRAIN] epoch=8, iter=6390/35000, loss=0.2020, lr=0.002085, batch_cost=1.0387, reader_cost=0.00009, ips=3.8509 samples/sec | ETA 08:15:17
2021-04-08 10:20:35 [INFO]	[TRAIN] epoch=8, iter=6400/35000, loss=0.1952, lr=0.002085, batch_cost=1.1040, reader_cost=0.00011, ips=3.6231 samples/sec | ETA 08:46:14
2021-04-08 10:20:45 [INFO]	[TRAIN] epoch=8, iter=6410/35000, loss=0.1740, lr=0.002084, batch_cost=1.0441, reader_cost=0.00014, ips=3.8310 samples/sec | ETA 08:17:31
2021-04-08 10:20:55 [INFO]	[TRAIN] epoch=8, iter=6420/35000, loss=0.2108, lr=0.002083, batch_cost=1.0202, reader_cost=0.00010, ips=3.9208 samples/sec | ETA 08:05:57
2021-04-08 10:21:06 [INFO]	[TRAIN] epoch=8, iter=6430/35000, loss=0.2581, lr=0.002083, batch_cost=1.0691, reader_cost=0.00010, ips=3.7415 samples/sec | ETA 08:29:03
2021-04-08 10:21:17 [INFO]	[TRAIN] epoch=8, iter=6440/35000, loss=0.1802, lr=0.002082, batch_cost=1.1262, reader_cost=0.00014, ips=3.5517 samples/sec | ETA 08:56:05
2021-04-08 10:21:29 [INFO]	[TRAIN] epoch=8, iter=6450/35000, loss=0.2011, lr=0.002081, batch_cost=1.1370, reader_cost=0.00015, ips=3.5181 samples/sec | ETA 09:01:00
2021-04-08 10:21:41 [INFO]	[TRAIN] epoch=8, iter=6460/35000, loss=0.1804, lr=0.002081, batch_cost=1.1968, reader_cost=0.00013, ips=3.3421 samples/sec | ETA 09:29:17
2021-04-08 10:21:51 [INFO]	[TRAIN] epoch=8, iter=6470/35000, loss=0.1894, lr=0.002080, batch_cost=1.0739, reader_cost=0.00011, ips=3.7249 samples/sec | ETA 08:30:37
2021-04-08 10:22:02 [INFO]	[TRAIN] epoch=8, iter=6480/35000, loss=0.2016, lr=0.002079, batch_cost=1.0634, reader_cost=0.00009, ips=3.7615 samples/sec | ETA 08:25:28
2021-04-08 10:22:12 [INFO]	[TRAIN] epoch=8, iter=6490/35000, loss=0.1693, lr=0.002079, batch_cost=0.9690, reader_cost=0.00009, ips=4.1280 samples/sec | ETA 07:40:25
2021-04-08 10:22:22 [INFO]	[TRAIN] epoch=8, iter=6500/35000, loss=0.2311, lr=0.002078, batch_cost=1.0540, reader_cost=0.00009, ips=3.7952 samples/sec | ETA 08:20:38
2021-04-08 10:22:33 [INFO]	[TRAIN] epoch=8, iter=6510/35000, loss=0.2029, lr=0.002077, batch_cost=1.0669, reader_cost=0.00009, ips=3.7490 samples/sec | ETA 08:26:37
2021-04-08 10:22:43 [INFO]	[TRAIN] epoch=8, iter=6520/35000, loss=0.2145, lr=0.002077, batch_cost=1.0271, reader_cost=0.00010, ips=3.8944 samples/sec | ETA 08:07:32
2021-04-08 10:22:54 [INFO]	[TRAIN] epoch=8, iter=6530/35000, loss=0.2249, lr=0.002076, batch_cost=1.0502, reader_cost=0.00009, ips=3.8088 samples/sec | ETA 08:18:19
2021-04-08 10:23:04 [INFO]	[TRAIN] epoch=8, iter=6540/35000, loss=0.1861, lr=0.002075, batch_cost=1.0206, reader_cost=0.00009, ips=3.9194 samples/sec | ETA 08:04:05
2021-04-08 10:23:14 [INFO]	[TRAIN] epoch=8, iter=6550/35000, loss=0.2141, lr=0.002075, batch_cost=1.0540, reader_cost=0.00009, ips=3.7951 samples/sec | ETA 08:19:45
2021-04-08 10:23:26 [INFO]	[TRAIN] epoch=8, iter=6560/35000, loss=0.1925, lr=0.002074, batch_cost=1.1108, reader_cost=0.00008, ips=3.6009 samples/sec | ETA 08:46:32
2021-04-08 10:23:36 [INFO]	[TRAIN] epoch=8, iter=6570/35000, loss=0.2098, lr=0.002073, batch_cost=1.0847, reader_cost=0.00010, ips=3.6878 samples/sec | ETA 08:33:57
2021-04-08 10:23:47 [INFO]	[TRAIN] epoch=8, iter=6580/35000, loss=0.1465, lr=0.002073, batch_cost=1.0294, reader_cost=0.00008, ips=3.8857 samples/sec | ETA 08:07:36
2021-04-08 10:23:57 [INFO]	[TRAIN] epoch=8, iter=6590/35000, loss=0.2230, lr=0.002072, batch_cost=1.0131, reader_cost=0.00008, ips=3.9484 samples/sec | ETA 07:59:41
2021-04-08 10:24:07 [INFO]	[TRAIN] epoch=8, iter=6600/35000, loss=0.1702, lr=0.002071, batch_cost=1.0057, reader_cost=0.00008, ips=3.9773 samples/sec | ETA 07:56:01
2021-04-08 10:24:18 [INFO]	[TRAIN] epoch=8, iter=6610/35000, loss=0.2132, lr=0.002071, batch_cost=1.0792, reader_cost=0.00009, ips=3.7065 samples/sec | ETA 08:30:37
2021-04-08 10:24:28 [INFO]	[TRAIN] epoch=8, iter=6620/35000, loss=0.1934, lr=0.002070, batch_cost=0.9973, reader_cost=0.00008, ips=4.0108 samples/sec | ETA 07:51:43
2021-04-08 10:24:38 [INFO]	[TRAIN] epoch=8, iter=6630/35000, loss=0.2026, lr=0.002070, batch_cost=1.0194, reader_cost=0.00008, ips=3.9241 samples/sec | ETA 08:01:58
2021-04-08 10:24:48 [INFO]	[TRAIN] epoch=8, iter=6640/35000, loss=0.1923, lr=0.002069, batch_cost=1.0225, reader_cost=0.00008, ips=3.9121 samples/sec | ETA 08:03:16
2021-04-08 10:24:59 [INFO]	[TRAIN] epoch=8, iter=6650/35000, loss=0.1979, lr=0.002068, batch_cost=1.0857, reader_cost=0.00010, ips=3.6843 samples/sec | ETA 08:32:59
2021-04-08 10:25:09 [INFO]	[TRAIN] epoch=8, iter=6660/35000, loss=0.2122, lr=0.002068, batch_cost=0.9929, reader_cost=0.00010, ips=4.0284 samples/sec | ETA 07:49:00
2021-04-08 10:25:20 [INFO]	[TRAIN] epoch=8, iter=6670/35000, loss=0.2092, lr=0.002067, batch_cost=1.0789, reader_cost=0.00010, ips=3.7073 samples/sec | ETA 08:29:26
2021-04-08 10:25:30 [INFO]	[TRAIN] epoch=8, iter=6680/35000, loss=0.2578, lr=0.002066, batch_cost=1.0312, reader_cost=0.00011, ips=3.8791 samples/sec | ETA 08:06:42
2021-04-08 10:25:40 [INFO]	[TRAIN] epoch=8, iter=6690/35000, loss=0.2238, lr=0.002066, batch_cost=1.0484, reader_cost=0.00010, ips=3.8154 samples/sec | ETA 08:14:39
2021-04-08 10:25:52 [INFO]	[TRAIN] epoch=8, iter=6700/35000, loss=0.1982, lr=0.002065, batch_cost=1.1295, reader_cost=0.00012, ips=3.5415 samples/sec | ETA 08:52:44
2021-04-08 10:26:02 [INFO]	[TRAIN] epoch=8, iter=6710/35000, loss=0.2036, lr=0.002064, batch_cost=1.0453, reader_cost=0.00010, ips=3.8267 samples/sec | ETA 08:12:51
2021-04-08 10:26:13 [INFO]	[TRAIN] epoch=8, iter=6720/35000, loss=0.2155, lr=0.002064, batch_cost=1.0411, reader_cost=0.00012, ips=3.8421 samples/sec | ETA 08:10:42
2021-04-08 10:26:23 [INFO]	[TRAIN] epoch=8, iter=6730/35000, loss=0.2365, lr=0.002063, batch_cost=1.0675, reader_cost=0.00010, ips=3.7470 samples/sec | ETA 08:22:58
2021-04-08 10:26:34 [INFO]	[TRAIN] epoch=8, iter=6740/35000, loss=0.2238, lr=0.002062, batch_cost=1.0999, reader_cost=0.00012, ips=3.6369 samples/sec | ETA 08:38:01
2021-04-08 10:26:45 [INFO]	[TRAIN] epoch=8, iter=6750/35000, loss=0.2034, lr=0.002062, batch_cost=1.0843, reader_cost=0.00013, ips=3.6891 samples/sec | ETA 08:30:30
2021-04-08 10:26:56 [INFO]	[TRAIN] epoch=8, iter=6760/35000, loss=0.1962, lr=0.002061, batch_cost=1.1255, reader_cost=0.00013, ips=3.5540 samples/sec | ETA 08:49:43
2021-04-08 10:27:07 [INFO]	[TRAIN] epoch=8, iter=6770/35000, loss=0.2059, lr=0.002060, batch_cost=1.0386, reader_cost=0.00009, ips=3.8514 samples/sec | ETA 08:08:39
2021-04-08 10:27:17 [INFO]	[TRAIN] epoch=8, iter=6780/35000, loss=0.3749, lr=0.002060, batch_cost=1.0281, reader_cost=0.00009, ips=3.8907 samples/sec | ETA 08:03:32
2021-04-08 10:27:27 [INFO]	[TRAIN] epoch=8, iter=6790/35000, loss=0.2470, lr=0.002059, batch_cost=1.0143, reader_cost=0.00009, ips=3.9438 samples/sec | ETA 07:56:52
2021-04-08 10:27:39 [INFO]	[TRAIN] epoch=8, iter=6800/35000, loss=0.1812, lr=0.002058, batch_cost=1.1326, reader_cost=0.00014, ips=3.5316 samples/sec | ETA 08:52:19
2021-04-08 10:27:50 [INFO]	[TRAIN] epoch=8, iter=6810/35000, loss=0.2375, lr=0.002058, batch_cost=1.1320, reader_cost=0.00010, ips=3.5337 samples/sec | ETA 08:51:50
2021-04-08 10:28:01 [INFO]	[TRAIN] epoch=8, iter=6820/35000, loss=0.2120, lr=0.002057, batch_cost=1.0803, reader_cost=0.00009, ips=3.7027 samples/sec | ETA 08:27:23
2021-04-08 10:28:11 [INFO]	[TRAIN] epoch=8, iter=6830/35000, loss=0.2117, lr=0.002056, batch_cost=1.0676, reader_cost=0.00009, ips=3.7465 samples/sec | ETA 08:21:15
2021-04-08 10:28:22 [INFO]	[TRAIN] epoch=8, iter=6840/35000, loss=0.2029, lr=0.002056, batch_cost=1.0229, reader_cost=0.00008, ips=3.9105 samples/sec | ETA 08:00:04
2021-04-08 10:28:32 [INFO]	[TRAIN] epoch=8, iter=6850/35000, loss=0.2184, lr=0.002055, batch_cost=1.0579, reader_cost=0.00008, ips=3.7812 samples/sec | ETA 08:16:18
2021-04-08 10:28:42 [INFO]	[TRAIN] epoch=8, iter=6860/35000, loss=0.1779, lr=0.002054, batch_cost=0.9809, reader_cost=0.00008, ips=4.0778 samples/sec | ETA 07:40:03
2021-04-08 10:28:52 [INFO]	[TRAIN] epoch=8, iter=6870/35000, loss=0.2097, lr=0.002054, batch_cost=1.0308, reader_cost=0.00009, ips=3.8807 samples/sec | ETA 08:03:15
2021-04-08 10:29:03 [INFO]	[TRAIN] epoch=8, iter=6880/35000, loss=0.2192, lr=0.002053, batch_cost=1.0967, reader_cost=0.06607, ips=3.6474 samples/sec | ETA 08:33:58
2021-04-08 10:29:14 [INFO]	[TRAIN] epoch=8, iter=6890/35000, loss=0.2067, lr=0.002052, batch_cost=1.1184, reader_cost=0.00012, ips=3.5764 samples/sec | ETA 08:43:59
2021-04-08 10:29:25 [INFO]	[TRAIN] epoch=8, iter=6900/35000, loss=0.2314, lr=0.002052, batch_cost=1.0439, reader_cost=0.00008, ips=3.8317 samples/sec | ETA 08:08:53
2021-04-08 10:29:35 [INFO]	[TRAIN] epoch=8, iter=6910/35000, loss=0.1971, lr=0.002051, batch_cost=1.0595, reader_cost=0.00011, ips=3.7755 samples/sec | ETA 08:16:00
2021-04-08 10:29:47 [INFO]	[TRAIN] epoch=8, iter=6920/35000, loss=0.2354, lr=0.002050, batch_cost=1.1289, reader_cost=0.00012, ips=3.5433 samples/sec | ETA 08:48:19
2021-04-08 10:29:57 [INFO]	[TRAIN] epoch=8, iter=6930/35000, loss=0.2110, lr=0.002050, batch_cost=1.0456, reader_cost=0.00009, ips=3.8257 samples/sec | ETA 08:09:08
2021-04-08 10:30:08 [INFO]	[TRAIN] epoch=8, iter=6940/35000, loss=0.2220, lr=0.002049, batch_cost=1.0347, reader_cost=0.00008, ips=3.8660 samples/sec | ETA 08:03:52
2021-04-08 10:30:18 [INFO]	[TRAIN] epoch=8, iter=6950/35000, loss=0.2090, lr=0.002048, batch_cost=1.0788, reader_cost=0.00011, ips=3.7078 samples/sec | ETA 08:24:20
2021-04-08 10:30:29 [INFO]	[TRAIN] epoch=8, iter=6960/35000, loss=0.2089, lr=0.002048, batch_cost=1.0996, reader_cost=0.00012, ips=3.6376 samples/sec | ETA 08:33:53
2021-04-08 10:30:40 [INFO]	[TRAIN] epoch=8, iter=6970/35000, loss=0.2111, lr=0.002047, batch_cost=1.0604, reader_cost=0.00011, ips=3.7721 samples/sec | ETA 08:15:23
2021-04-08 10:30:51 [INFO]	[TRAIN] epoch=8, iter=6980/35000, loss=0.1905, lr=0.002047, batch_cost=1.0727, reader_cost=0.00010, ips=3.7288 samples/sec | ETA 08:20:58
2021-04-08 10:31:00 [INFO]	[TRAIN] epoch=8, iter=6990/35000, loss=0.1796, lr=0.002046, batch_cost=0.9592, reader_cost=0.00008, ips=4.1703 samples/sec | ETA 07:27:45
2021-04-08 10:31:11 [INFO]	[TRAIN] epoch=8, iter=7000/35000, loss=0.2226, lr=0.002045, batch_cost=1.0320, reader_cost=0.00008, ips=3.8759 samples/sec | ETA 08:01:36
2021-04-08 10:31:11 [INFO]	Start evaluating (total_samples=500, total_iters=500)...
/opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages/paddle/fluid/dygraph/math_op_patch.py:238: UserWarning: The dtype of left and right variables are not the same, left dtype is VarType.INT32, but right dtype is VarType.BOOL, the right dtype will convert to VarType.INT32
  format(lhs_dtype, rhs_dtype, lhs_dtype))
/opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages/paddle/fluid/dygraph/math_op_patch.py:238: UserWarning: The dtype of left and right variables are not the same, left dtype is VarType.INT64, but right dtype is VarType.BOOL, the right dtype will convert to VarType.INT64
  format(lhs_dtype, rhs_dtype, lhs_dtype))
500/500 [==============================] - 103s 206ms/step - batch_cost: 0.2054 - reader cost: 0.001
2021-04-08 10:32:54 [INFO]	[EVAL] #Images=500 mIoU=0.1979 Acc=0.9857 Kappa=0.5681 
2021-04-08 10:32:54 [INFO]	[EVAL] Class IoU: 
[0.9872 0.2389 0.4774 0.472  0.154  0.2929 0.3465 0.     0.     0.
 0.     0.     0.     0.     0.    ]
2021-04-08 10:32:54 [INFO]	[EVAL] Class Acc: 
[0.9937 0.3976 0.6139 0.5784 0.2579 0.399  0.4729 0.     0.     0.
 0.     0.     0.     0.     0.    ]
2021-04-08 10:32:56 [INFO]	[EVAL] The model with the best validation mIoU (0.1979) was saved at iter 7000.
2021-04-08 10:33:07 [INFO]	[TRAIN] epoch=9, iter=7010/35000, loss=0.1783, lr=0.002045, batch_cost=1.0728, reader_cost=0.00012, ips=3.7286 samples/sec | ETA 08:20:27
2021-04-08 10:33:17 [INFO]	[TRAIN] epoch=9, iter=7020/35000, loss=0.2172, lr=0.002044, batch_cost=1.0500, reader_cost=0.00013, ips=3.8095 samples/sec | ETA 08:09:39
2021-04-08 10:33:28 [INFO]	[TRAIN] epoch=9, iter=7030/35000, loss=0.2276, lr=0.002043, batch_cost=1.1482, reader_cost=0.00011, ips=3.4838 samples/sec | ETA 08:55:14
2021-04-08 10:33:39 [INFO]	[TRAIN] epoch=9, iter=7040/35000, loss=0.2135, lr=0.002043, batch_cost=1.0231, reader_cost=0.00009, ips=3.9096 samples/sec | ETA 07:56:46
2021-04-08 10:33:50 [INFO]	[TRAIN] epoch=9, iter=7050/35000, loss=0.2092, lr=0.002042, batch_cost=1.1049, reader_cost=0.00014, ips=3.6201 samples/sec | ETA 08:34:42
2021-04-08 10:34:00 [INFO]	[TRAIN] epoch=9, iter=7060/35000, loss=0.1946, lr=0.002041, batch_cost=1.0493, reader_cost=0.00009, ips=3.8121 samples/sec | ETA 08:08:36
2021-04-08 10:34:11 [INFO]	[TRAIN] epoch=9, iter=7070/35000, loss=0.2025, lr=0.002041, batch_cost=1.0962, reader_cost=0.00013, ips=3.6489 samples/sec | ETA 08:30:17
2021-04-08 10:34:21 [INFO]	[TRAIN] epoch=9, iter=7080/35000, loss=0.1925, lr=0.002040, batch_cost=0.9794, reader_cost=0.00008, ips=4.0842 samples/sec | ETA 07:35:44
2021-04-08 10:34:32 [INFO]	[TRAIN] epoch=9, iter=7090/35000, loss=0.1994, lr=0.002039, batch_cost=1.0704, reader_cost=0.00008, ips=3.7370 samples/sec | ETA 08:17:54
2021-04-08 10:34:42 [INFO]	[TRAIN] epoch=9, iter=7100/35000, loss=0.2191, lr=0.002039, batch_cost=1.0315, reader_cost=0.00009, ips=3.8778 samples/sec | ETA 07:59:39
2021-04-08 10:34:52 [INFO]	[TRAIN] epoch=9, iter=7110/35000, loss=0.2162, lr=0.002038, batch_cost=0.9734, reader_cost=0.00008, ips=4.1094 samples/sec | ETA 07:32:27
2021-04-08 10:35:02 [INFO]	[TRAIN] epoch=9, iter=7120/35000, loss=0.1996, lr=0.002037, batch_cost=1.0414, reader_cost=0.00008, ips=3.8411 samples/sec | ETA 08:03:53
2021-04-08 10:35:13 [INFO]	[TRAIN] epoch=9, iter=7130/35000, loss=0.2196, lr=0.002037, batch_cost=1.0631, reader_cost=0.00008, ips=3.7626 samples/sec | ETA 08:13:48
2021-04-08 10:35:24 [INFO]	[TRAIN] epoch=9, iter=7140/35000, loss=0.2188, lr=0.002036, batch_cost=1.0844, reader_cost=0.00010, ips=3.6888 samples/sec | ETA 08:23:30
2021-04-08 10:35:34 [INFO]	[TRAIN] epoch=9, iter=7150/35000, loss=0.2410, lr=0.002035, batch_cost=1.0702, reader_cost=0.00010, ips=3.7375 samples/sec | ETA 08:16:46
2021-04-08 10:35:45 [INFO]	[TRAIN] epoch=9, iter=7160/35000, loss=0.2094, lr=0.002035, batch_cost=1.0379, reader_cost=0.00010, ips=3.8538 samples/sec | ETA 08:01:35
2021-04-08 10:35:55 [INFO]	[TRAIN] epoch=9, iter=7170/35000, loss=0.2016, lr=0.002034, batch_cost=1.0376, reader_cost=0.00009, ips=3.8552 samples/sec | ETA 08:01:15
2021-04-08 10:36:06 [INFO]	[TRAIN] epoch=9, iter=7180/35000, loss=0.2105, lr=0.002033, batch_cost=1.0601, reader_cost=0.00009, ips=3.7734 samples/sec | ETA 08:11:30
2021-04-08 10:36:17 [INFO]	[TRAIN] epoch=9, iter=7190/35000, loss=0.2069, lr=0.002033, batch_cost=1.0858, reader_cost=0.00009, ips=3.6839 samples/sec | ETA 08:23:16
2021-04-08 10:36:27 [INFO]	[TRAIN] epoch=9, iter=7200/35000, loss=0.2015, lr=0.002032, batch_cost=1.0550, reader_cost=0.00009, ips=3.7916 samples/sec | ETA 08:08:47
2021-04-08 10:36:37 [INFO]	[TRAIN] epoch=9, iter=7210/35000, loss=0.1748, lr=0.002031, batch_cost=0.9704, reader_cost=0.00009, ips=4.1220 samples/sec | ETA 07:29:27
2021-04-08 10:36:48 [INFO]	[TRAIN] epoch=9, iter=7220/35000, loss=0.1989, lr=0.002031, batch_cost=1.0699, reader_cost=0.00010, ips=3.7386 samples/sec | ETA 08:15:22
2021-04-08 10:36:58 [INFO]	[TRAIN] epoch=9, iter=7230/35000, loss=0.2276, lr=0.002030, batch_cost=1.0851, reader_cost=0.00014, ips=3.6862 samples/sec | ETA 08:22:14
2021-04-08 10:37:09 [INFO]	[TRAIN] epoch=9, iter=7240/35000, loss=0.2133, lr=0.002029, batch_cost=1.0206, reader_cost=0.00010, ips=3.9192 samples/sec | ETA 07:52:12
2021-04-08 10:37:19 [INFO]	[TRAIN] epoch=9, iter=7250/35000, loss=0.2121, lr=0.002029, batch_cost=1.0623, reader_cost=0.00008, ips=3.7655 samples/sec | ETA 08:11:17
2021-04-08 10:37:29 [INFO]	[TRAIN] epoch=9, iter=7260/35000, loss=0.2147, lr=0.002028, batch_cost=1.0064, reader_cost=0.00008, ips=3.9747 samples/sec | ETA 07:45:16
2021-04-08 10:37:39 [INFO]	[TRAIN] epoch=9, iter=7270/35000, loss=0.2258, lr=0.002027, batch_cost=0.9983, reader_cost=0.00008, ips=4.0069 samples/sec | ETA 07:41:22
2021-04-08 10:37:50 [INFO]	[TRAIN] epoch=9, iter=7280/35000, loss=0.1994, lr=0.002027, batch_cost=1.0797, reader_cost=0.00011, ips=3.7048 samples/sec | ETA 08:18:48
2021-04-08 10:38:01 [INFO]	[TRAIN] epoch=9, iter=7290/35000, loss=0.1831, lr=0.002026, batch_cost=1.1136, reader_cost=0.00013, ips=3.5918 samples/sec | ETA 08:34:18
2021-04-08 10:38:12 [INFO]	[TRAIN] epoch=9, iter=7300/35000, loss=0.2439, lr=0.002025, batch_cost=1.1283, reader_cost=0.00009, ips=3.5452 samples/sec | ETA 08:40:53
2021-04-08 10:38:24 [INFO]	[TRAIN] epoch=9, iter=7310/35000, loss=0.2297, lr=0.002025, batch_cost=1.1026, reader_cost=0.00013, ips=3.6277 samples/sec | ETA 08:28:51
2021-04-08 10:38:34 [INFO]	[TRAIN] epoch=9, iter=7320/35000, loss=0.2291, lr=0.002024, batch_cost=1.0667, reader_cost=0.00011, ips=3.7500 samples/sec | ETA 08:12:05
2021-04-08 10:38:45 [INFO]	[TRAIN] epoch=9, iter=7330/35000, loss=0.2180, lr=0.002023, batch_cost=1.0527, reader_cost=0.00012, ips=3.7996 samples/sec | ETA 08:05:29
2021-04-08 10:38:55 [INFO]	[TRAIN] epoch=9, iter=7340/35000, loss=0.1709, lr=0.002023, batch_cost=1.0047, reader_cost=0.00009, ips=3.9813 samples/sec | ETA 07:43:10
2021-04-08 10:39:05 [INFO]	[TRAIN] epoch=9, iter=7350/35000, loss=0.2211, lr=0.002022, batch_cost=1.0589, reader_cost=0.00008, ips=3.7777 samples/sec | ETA 08:07:57
2021-04-08 10:39:17 [INFO]	[TRAIN] epoch=9, iter=7360/35000, loss=0.2350, lr=0.002022, batch_cost=1.1764, reader_cost=0.00014, ips=3.4001 samples/sec | ETA 09:01:56
2021-04-08 10:39:28 [INFO]	[TRAIN] epoch=9, iter=7370/35000, loss=0.2058, lr=0.002021, batch_cost=1.1304, reader_cost=0.00013, ips=3.5385 samples/sec | ETA 08:40:33
2021-04-08 10:39:38 [INFO]	[TRAIN] epoch=9, iter=7380/35000, loss=0.2030, lr=0.002020, batch_cost=1.0050, reader_cost=0.00009, ips=3.9799 samples/sec | ETA 07:42:39
2021-04-08 10:39:50 [INFO]	[TRAIN] epoch=9, iter=7390/35000, loss=0.1859, lr=0.002020, batch_cost=1.1548, reader_cost=0.00014, ips=3.4639 samples/sec | ETA 08:51:23
2021-04-08 10:40:02 [INFO]	[TRAIN] epoch=9, iter=7400/35000, loss=0.2126, lr=0.002019, batch_cost=1.1655, reader_cost=0.00014, ips=3.4321 samples/sec | ETA 08:56:07
2021-04-08 10:40:12 [INFO]	[TRAIN] epoch=9, iter=7410/35000, loss=0.2128, lr=0.002018, batch_cost=1.0643, reader_cost=0.00009, ips=3.7585 samples/sec | ETA 08:09:23
2021-04-08 10:40:22 [INFO]	[TRAIN] epoch=9, iter=7420/35000, loss=0.2045, lr=0.002018, batch_cost=1.0039, reader_cost=0.00009, ips=3.9843 samples/sec | ETA 07:41:28
2021-04-08 10:40:33 [INFO]	[TRAIN] epoch=9, iter=7430/35000, loss=0.1948, lr=0.002017, batch_cost=1.0320, reader_cost=0.00009, ips=3.8758 samples/sec | ETA 07:54:13
2021-04-08 10:40:43 [INFO]	[TRAIN] epoch=9, iter=7440/35000, loss=0.2034, lr=0.002016, batch_cost=1.0771, reader_cost=0.00009, ips=3.7138 samples/sec | ETA 08:14:43
2021-04-08 10:40:54 [INFO]	[TRAIN] epoch=9, iter=7450/35000, loss=0.2285, lr=0.002016, batch_cost=1.1053, reader_cost=0.00009, ips=3.6189 samples/sec | ETA 08:27:30
2021-04-08 10:41:05 [INFO]	[TRAIN] epoch=9, iter=7460/35000, loss=0.1733, lr=0.002015, batch_cost=1.0539, reader_cost=0.00012, ips=3.7953 samples/sec | ETA 08:03:45
2021-04-08 10:41:16 [INFO]	[TRAIN] epoch=9, iter=7470/35000, loss=0.2410, lr=0.002014, batch_cost=1.0781, reader_cost=0.00008, ips=3.7103 samples/sec | ETA 08:14:39
2021-04-08 10:41:26 [INFO]	[TRAIN] epoch=9, iter=7480/35000, loss=0.2051, lr=0.002014, batch_cost=1.0267, reader_cost=0.00009, ips=3.8960 samples/sec | ETA 07:50:54
2021-04-08 10:41:37 [INFO]	[TRAIN] epoch=9, iter=7490/35000, loss=0.2213, lr=0.002013, batch_cost=1.0964, reader_cost=0.00014, ips=3.6482 samples/sec | ETA 08:22:42
2021-04-08 10:41:48 [INFO]	[TRAIN] epoch=9, iter=7500/35000, loss=0.2178, lr=0.002012, batch_cost=1.1163, reader_cost=0.00012, ips=3.5834 samples/sec | ETA 08:31:36
2021-04-08 10:41:59 [INFO]	[TRAIN] epoch=9, iter=7510/35000, loss=0.1996, lr=0.002012, batch_cost=1.0849, reader_cost=0.00012, ips=3.6869 samples/sec | ETA 08:17:04
2021-04-08 10:42:10 [INFO]	[TRAIN] epoch=9, iter=7520/35000, loss=0.2209, lr=0.002011, batch_cost=1.0595, reader_cost=0.00010, ips=3.7752 samples/sec | ETA 08:05:16
2021-04-08 10:42:20 [INFO]	[TRAIN] epoch=9, iter=7530/35000, loss=0.1771, lr=0.002010, batch_cost=1.0364, reader_cost=0.00008, ips=3.8595 samples/sec | ETA 07:54:30
2021-04-08 10:42:31 [INFO]	[TRAIN] epoch=9, iter=7540/35000, loss=0.1948, lr=0.002010, batch_cost=1.0658, reader_cost=0.00010, ips=3.7531 samples/sec | ETA 08:07:46
2021-04-08 10:42:41 [INFO]	[TRAIN] epoch=9, iter=7550/35000, loss=0.2044, lr=0.002009, batch_cost=1.0817, reader_cost=0.00008, ips=3.6979 samples/sec | ETA 08:14:52
2021-04-08 10:42:53 [INFO]	[TRAIN] epoch=9, iter=7560/35000, loss=0.2094, lr=0.002008, batch_cost=1.1292, reader_cost=0.00011, ips=3.5424 samples/sec | ETA 08:36:24
2021-04-08 10:43:03 [INFO]	[TRAIN] epoch=9, iter=7570/35000, loss=0.1898, lr=0.002008, batch_cost=1.0158, reader_cost=0.00012, ips=3.9377 samples/sec | ETA 07:44:24
2021-04-08 10:43:13 [INFO]	[TRAIN] epoch=9, iter=7580/35000, loss=0.2550, lr=0.002007, batch_cost=1.0497, reader_cost=0.00015, ips=3.8105 samples/sec | ETA 07:59:43
2021-04-08 10:43:23 [INFO]	[TRAIN] epoch=9, iter=7590/35000, loss=0.1830, lr=0.002006, batch_cost=0.9935, reader_cost=0.00008, ips=4.0261 samples/sec | ETA 07:33:52
2021-04-08 10:43:34 [INFO]	[TRAIN] epoch=9, iter=7600/35000, loss=0.1888, lr=0.002006, batch_cost=1.0526, reader_cost=0.00010, ips=3.8000 samples/sec | ETA 08:00:41
2021-04-08 10:43:45 [INFO]	[TRAIN] epoch=9, iter=7610/35000, loss=0.2366, lr=0.002005, batch_cost=1.0825, reader_cost=0.00011, ips=3.6952 samples/sec | ETA 08:14:09
2021-04-08 10:43:56 [INFO]	[TRAIN] epoch=9, iter=7620/35000, loss=0.2138, lr=0.002004, batch_cost=1.0789, reader_cost=0.00008, ips=3.7076 samples/sec | ETA 08:12:18
2021-04-08 10:44:06 [INFO]	[TRAIN] epoch=9, iter=7630/35000, loss=0.2137, lr=0.002004, batch_cost=1.0301, reader_cost=0.00010, ips=3.8831 samples/sec | ETA 07:49:53
2021-04-08 10:44:16 [INFO]	[TRAIN] epoch=9, iter=7640/35000, loss=0.2155, lr=0.002003, batch_cost=1.0623, reader_cost=0.00008, ips=3.7655 samples/sec | ETA 08:04:24
2021-04-08 10:44:28 [INFO]	[TRAIN] epoch=9, iter=7650/35000, loss=0.1630, lr=0.002002, batch_cost=1.1322, reader_cost=0.00012, ips=3.5331 samples/sec | ETA 08:36:04
2021-04-08 10:44:39 [INFO]	[TRAIN] epoch=9, iter=7660/35000, loss=0.1974, lr=0.002002, batch_cost=1.1523, reader_cost=0.00014, ips=3.4712 samples/sec | ETA 08:45:04
2021-04-08 10:44:50 [INFO]	[TRAIN] epoch=9, iter=7670/35000, loss=0.2032, lr=0.002001, batch_cost=1.0271, reader_cost=0.00011, ips=3.8944 samples/sec | ETA 07:47:50
2021-04-08 10:45:01 [INFO]	[TRAIN] epoch=9, iter=7680/35000, loss=0.2560, lr=0.002000, batch_cost=1.1139, reader_cost=0.00012, ips=3.5911 samples/sec | ETA 08:27:11
2021-04-08 10:45:11 [INFO]	[TRAIN] epoch=9, iter=7690/35000, loss=0.1690, lr=0.002000, batch_cost=1.0349, reader_cost=0.00013, ips=3.8651 samples/sec | ETA 07:51:03
2021-04-08 10:45:22 [INFO]	[TRAIN] epoch=9, iter=7700/35000, loss=0.1883, lr=0.001999, batch_cost=1.0980, reader_cost=0.00011, ips=3.6431 samples/sec | ETA 08:19:34
2021-04-08 10:45:33 [INFO]	[TRAIN] epoch=9, iter=7710/35000, loss=0.2164, lr=0.001998, batch_cost=1.0833, reader_cost=0.00009, ips=3.6923 samples/sec | ETA 08:12:44
2021-04-08 10:45:44 [INFO]	[TRAIN] epoch=9, iter=7720/35000, loss=0.2087, lr=0.001998, batch_cost=1.1057, reader_cost=0.00012, ips=3.6175 samples/sec | ETA 08:22:44
2021-04-08 10:45:55 [INFO]	[TRAIN] epoch=9, iter=7730/35000, loss=0.2024, lr=0.001997, batch_cost=1.1481, reader_cost=0.00014, ips=3.4839 samples/sec | ETA 08:41:49
2021-04-08 10:46:06 [INFO]	[TRAIN] epoch=9, iter=7740/35000, loss=0.2127, lr=0.001996, batch_cost=1.0996, reader_cost=0.00013, ips=3.6378 samples/sec | ETA 08:19:34
2021-04-08 10:46:16 [INFO]	[TRAIN] epoch=9, iter=7750/35000, loss=0.1886, lr=0.001996, batch_cost=0.9753, reader_cost=0.00014, ips=4.1013 samples/sec | ETA 07:22:56
2021-04-08 10:46:28 [INFO]	[TRAIN] epoch=9, iter=7760/35000, loss=0.1873, lr=0.001995, batch_cost=1.2101, reader_cost=0.07465, ips=3.3056 samples/sec | ETA 09:09:22
2021-04-08 10:46:38 [INFO]	[TRAIN] epoch=9, iter=7770/35000, loss=0.1910, lr=0.001995, batch_cost=1.0024, reader_cost=0.00008, ips=3.9906 samples/sec | ETA 07:34:54
2021-04-08 10:46:49 [INFO]	[TRAIN] epoch=9, iter=7780/35000, loss=0.2153, lr=0.001994, batch_cost=1.0494, reader_cost=0.00008, ips=3.8118 samples/sec | ETA 07:56:04
2021-04-08 10:47:00 [INFO]	[TRAIN] epoch=9, iter=7790/35000, loss=0.2261, lr=0.001993, batch_cost=1.0863, reader_cost=0.00008, ips=3.6823 samples/sec | ETA 08:12:37
2021-04-08 10:47:11 [INFO]	[TRAIN] epoch=9, iter=7800/35000, loss=0.2153, lr=0.001993, batch_cost=1.1479, reader_cost=0.00011, ips=3.4847 samples/sec | ETA 08:40:22
2021-04-08 10:47:22 [INFO]	[TRAIN] epoch=9, iter=7810/35000, loss=0.1830, lr=0.001992, batch_cost=1.0728, reader_cost=0.00011, ips=3.7287 samples/sec | ETA 08:06:08
2021-04-08 10:47:33 [INFO]	[TRAIN] epoch=9, iter=7820/35000, loss=0.1824, lr=0.001991, batch_cost=1.1096, reader_cost=0.00014, ips=3.6048 samples/sec | ETA 08:22:39
2021-04-08 10:47:45 [INFO]	[TRAIN] epoch=9, iter=7830/35000, loss=0.1868, lr=0.001991, batch_cost=1.1620, reader_cost=0.00013, ips=3.4424 samples/sec | ETA 08:46:10
2021-04-08 10:47:56 [INFO]	[TRAIN] epoch=9, iter=7840/35000, loss=0.2101, lr=0.001990, batch_cost=1.1051, reader_cost=0.00012, ips=3.6195 samples/sec | ETA 08:20:14
2021-04-08 10:48:06 [INFO]	[TRAIN] epoch=9, iter=7850/35000, loss=0.1703, lr=0.001989, batch_cost=1.0187, reader_cost=0.00010, ips=3.9266 samples/sec | ETA 07:40:57
2021-04-08 10:48:17 [INFO]	[TRAIN] epoch=9, iter=7860/35000, loss=0.2123, lr=0.001989, batch_cost=1.0715, reader_cost=0.00008, ips=3.7329 samples/sec | ETA 08:04:41
2021-04-08 10:48:27 [INFO]	[TRAIN] epoch=9, iter=7870/35000, loss=0.2546, lr=0.001988, batch_cost=1.0805, reader_cost=0.00009, ips=3.7021 samples/sec | ETA 08:08:33
2021-04-08 10:48:38 [INFO]	[TRAIN] epoch=10, iter=7880/35000, loss=0.2201, lr=0.001987, batch_cost=1.0298, reader_cost=0.00009, ips=3.8841 samples/sec | ETA 07:45:29
2021-04-08 10:48:48 [INFO]	[TRAIN] epoch=10, iter=7890/35000, loss=0.1705, lr=0.001987, batch_cost=1.0292, reader_cost=0.00009, ips=3.8866 samples/sec | ETA 07:45:00
2021-04-08 10:48:58 [INFO]	[TRAIN] epoch=10, iter=7900/35000, loss=0.1844, lr=0.001986, batch_cost=1.0092, reader_cost=0.00009, ips=3.9636 samples/sec | ETA 07:35:48
2021-04-08 10:49:08 [INFO]	[TRAIN] epoch=10, iter=7910/35000, loss=0.1921, lr=0.001985, batch_cost=1.0208, reader_cost=0.00009, ips=3.9185 samples/sec | ETA 07:40:53
2021-04-08 10:49:19 [INFO]	[TRAIN] epoch=10, iter=7920/35000, loss=0.1991, lr=0.001985, batch_cost=1.1182, reader_cost=0.00011, ips=3.5773 samples/sec | ETA 08:24:39
2021-04-08 10:49:30 [INFO]	[TRAIN] epoch=10, iter=7930/35000, loss=0.2251, lr=0.001984, batch_cost=1.0473, reader_cost=0.00008, ips=3.8194 samples/sec | ETA 07:52:29
2021-04-08 10:49:41 [INFO]	[TRAIN] epoch=10, iter=7940/35000, loss=0.1971, lr=0.001983, batch_cost=1.0838, reader_cost=0.00009, ips=3.6908 samples/sec | ETA 08:08:46
2021-04-08 10:49:51 [INFO]	[TRAIN] epoch=10, iter=7950/35000, loss=0.1989, lr=0.001983, batch_cost=1.0728, reader_cost=0.00013, ips=3.7285 samples/sec | ETA 08:03:40
2021-04-08 10:50:03 [INFO]	[TRAIN] epoch=10, iter=7960/35000, loss=0.2182, lr=0.001982, batch_cost=1.1091, reader_cost=0.00014, ips=3.6064 samples/sec | ETA 08:19:50
2021-04-08 10:50:13 [INFO]	[TRAIN] epoch=10, iter=7970/35000, loss=0.2061, lr=0.001981, batch_cost=1.0185, reader_cost=0.00010, ips=3.9275 samples/sec | ETA 07:38:48
2021-04-08 10:50:24 [INFO]	[TRAIN] epoch=10, iter=7980/35000, loss=0.2233, lr=0.001981, batch_cost=1.0997, reader_cost=0.00008, ips=3.6374 samples/sec | ETA 08:15:13
2021-04-08 10:50:35 [INFO]	[TRAIN] epoch=10, iter=7990/35000, loss=0.3235, lr=0.001980, batch_cost=1.1299, reader_cost=0.00011, ips=3.5400 samples/sec | ETA 08:28:39
2021-04-08 10:50:47 [INFO]	[TRAIN] epoch=10, iter=8000/35000, loss=0.2507, lr=0.001979, batch_cost=1.1588, reader_cost=0.00011, ips=3.4518 samples/sec | ETA 08:41:27
2021-04-08 10:50:47 [INFO]	Start evaluating (total_samples=500, total_iters=500)...
500/500 [==============================] - 100s 200ms/step - batch_cost: 0.1996 - reader cost: 7.0843e-0
2021-04-08 10:52:27 [INFO]	[EVAL] #Images=500 mIoU=0.1877 Acc=0.9878 Kappa=0.5409 
2021-04-08 10:52:27 [INFO]	[EVAL] Class IoU: 
[0.9889 0.238  0.4705 0.4509 0.117  0.2911 0.2587 0.     0.     0.
 0.     0.     0.     0.     0.    ]
2021-04-08 10:52:27 [INFO]	[EVAL] Class Acc: 
[0.9912 0.3593 0.721  0.5205 0.4217 0.4955 0.7894 0.     0.     0.
 0.     0.     0.     0.     0.    ]
2021-04-08 10:52:28 [INFO]	[EVAL] The model with the best validation mIoU (0.1979) was saved at iter 7000.
2021-04-08 10:52:39 [INFO]	[TRAIN] epoch=10, iter=8010/35000, loss=0.1795, lr=0.001979, batch_cost=1.0620, reader_cost=0.00013, ips=3.7664 samples/sec | ETA 07:57:43
2021-04-08 10:52:49 [INFO]	[TRAIN] epoch=10, iter=8020/35000, loss=0.1845, lr=0.001978, batch_cost=1.0571, reader_cost=0.00011, ips=3.7841 samples/sec | ETA 07:55:19
2021-04-08 10:53:00 [INFO]	[TRAIN] epoch=10, iter=8030/35000, loss=0.2287, lr=0.001977, batch_cost=1.1007, reader_cost=0.00010, ips=3.6342 samples/sec | ETA 08:14:44
2021-04-08 10:53:11 [INFO]	[TRAIN] epoch=10, iter=8040/35000, loss=0.2048, lr=0.001977, batch_cost=1.0854, reader_cost=0.00014, ips=3.6851 samples/sec | ETA 08:07:43
2021-04-08 10:53:23 [INFO]	[TRAIN] epoch=10, iter=8050/35000, loss=0.1982, lr=0.001976, batch_cost=1.1484, reader_cost=0.00013, ips=3.4832 samples/sec | ETA 08:35:48
2021-04-08 10:53:34 [INFO]	[TRAIN] epoch=10, iter=8060/35000, loss=0.2150, lr=0.001975, batch_cost=1.1500, reader_cost=0.00015, ips=3.4782 samples/sec | ETA 08:36:21
^C
Traceback (most recent call last):
  File "train.py", line 154, in <module>
    main(args)
  File "train.py", line 149, in main
    keep_checkpoint_max=args.keep_checkpoint_max)
  File "/home/aistudio/PaddleSeg/paddleseg/core/train.py", line 151, in train
    edges=edges)
  File "/home/aistudio/PaddleSeg/paddleseg/core/train.py", line 46, in loss_computation
    loss_list.append(losses['coef'][i] * loss_i(logits, labels))
  File "/opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages/paddle/fluid/dygraph/layers.py", line 902, in __call__
    outputs = self.forward(*inputs, **kwargs)
  File "/home/aistudio/PaddleSeg/paddleseg/models/losses/mixed_loss.py", line 56, in forward
    output = loss(logits, labels)
  File "/opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages/paddle/fluid/dygraph/layers.py", line 902, in __call__
    outputs = self.forward(*inputs, **kwargs)
  File "/home/aistudio/PaddleSeg/paddleseg/models/losses/lovasz_loss.py", line 53, in forward
    loss = lovasz_softmax_flat(vprobas, vlabels, classes=self.classes)
  File "/home/aistudio/PaddleSeg/paddleseg/models/losses/lovasz_loss.py", line 191, in lovasz_softmax_flat
    grad = lovasz_grad(fg_sorted)
  File "/home/aistudio/PaddleSeg/paddleseg/models/losses/lovasz_loss.py", line 98, in lovasz_grad
    jaccard[1:p] = jaccard[1:p] - jaccard[0:-1]
KeyboardInterrupt

评估

In [20]:
!python val.py \
       --config configs/ocrnet/ocrnet_hrnetw18_cityscapes_1024x512_160k_lovasz_softmax.yml \
       --model_path output/iter_7000/model.pdparams

/opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages/setuptools/depends.py:2: DeprecationWarning: the imp module is deprecated in favour of importlib; see the module's documentation for alternative uses
  import imp
2021-04-08 10:54:19 [INFO]	
---------------Config Information---------------
SOLVER:
  CROSS_ENTROPY_WEIGHT: dynamic
  LR: 0.005
  LR_POLICY: poly
  NUM_EPOCHS: 40
  OPTIMIZER: sgd
batch_size: 4
iters: 35000
learning_rate:
  decay:
    end_lr: 0.0
    power: 0.9
    type: poly
  value: 0.0025
loss:
  coef:
  - 1
  - 0.4
  types:
  - coef:
    - 0.8
    - 0.2
    losses:
    - type: CrossEntropyLoss
    - type: LovaszSoftmaxLoss
    type: MixedLoss
  - coef:
    - 0.8
    - 0.2
    losses:
    - type: CrossEntropyLoss
    - type: LovaszSoftmaxLoss
    type: MixedLoss
model:
  backbone:
    pretrained: https://bj.bcebos.com/paddleseg/dygraph/hrnet_w18_ssld.tar.gz
    type: HRNet_W18
  backbone_indices:
  - 0
  type: OCRNet
optimizer:
  momentum: 0.9
  type: sgd
  weight_decay: 4.0e-05
train_dataset:
  dataset_root: /home/aistudio/
  mode: train
  num_classes: 15
  train_path: /home/aistudio/train_list.txt
  transforms:
  - max_scale_factor: 2.0
    min_scale_factor: 0.5
    scale_step_size: 0.25
    type: ResizeStepScaling
  - max_rotation: 30
    type: RandomRotation
  - type: RandomHorizontalFlip
  - type: RandomVerticalFlip
  - crop_size:
    - 1024
    - 512
    type: RandomPaddingCrop
  - type: RandomBlur
  - brightness_range: 0.4
    contrast_range: 0.4
    saturation_range: 0.4
    type: RandomDistort
  - type: Normalize
  type: Dataset
val_dataset:
  dataset_root: /home/aistudio/
  mode: val
  num_classes: 15
  transforms:
  - type: Normalize
  type: Dataset
  val_path: /home/aistudio/val_list.txt
------------------------------------------------
W0408 10:54:19.346495 13276 device_context.cc:362] Please NOTE: device: 0, GPU Compute Capability: 7.0, Driver API Version: 10.1, Runtime API Version: 10.1
W0408 10:54:19.346547 13276 device_context.cc:372] device: 0, cuDNN Version: 7.6.
2021-04-08 10:54:24 [INFO]	Loading pretrained model from https://bj.bcebos.com/paddleseg/dygraph/hrnet_w18_ssld.tar.gz
2021-04-08 10:54:24,349 - INFO - Lock 140006487673616 acquired on /home/aistudio/.paddleseg/tmp/hrnet_w18_ssld
2021-04-08 10:54:24,349 - INFO - Lock 140006487673616 released on /home/aistudio/.paddleseg/tmp/hrnet_w18_ssld
2021-04-08 10:54:25 [INFO]	There are 1525/1525 variables loaded into HRNet.
2021-04-08 10:54:25 [INFO]	Loading pretrained model from output/iter_7000/model.pdparams
2021-04-08 10:54:26 [INFO]	There are 1583/1583 variables loaded into OCRNet.
2021-04-08 10:54:26 [INFO]	Loaded trained params of model successfully
2021-04-08 10:54:26 [INFO]	Start evaluating (total_samples=500, total_iters=500)...
/opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages/paddle/fluid/dygraph/math_op_patch.py:238: UserWarning: The dtype of left and right variables are not the same, left dtype is VarType.INT32, but right dtype is VarType.BOOL, the right dtype will convert to VarType.INT32
  format(lhs_dtype, rhs_dtype, lhs_dtype))
/opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages/paddle/fluid/dygraph/math_op_patch.py:238: UserWarning: The dtype of left and right variables are not the same, left dtype is VarType.INT64, but right dtype is VarType.BOOL, the right dtype will convert to VarType.INT64
  format(lhs_dtype, rhs_dtype, lhs_dtype))
500/500 [==============================] - 93s 185ms/step - batch_cost: 0.1851 - reader cost: 8.3591e-
2021-04-08 10:55:59 [INFO]	[EVAL] #Images=500 mIoU=0.1979 Acc=0.9857 Kappa=0.5681 
2021-04-08 10:55:59 [INFO]	[EVAL] Class IoU: 
[0.9872 0.2389 0.4774 0.472  0.154  0.2929 0.3465 0.     0.     0.
 0.     0.     0.     0.     0.    ]
2021-04-08 10:55:59 [INFO]	[EVAL] Class Acc: 
[0.9937 0.3976 0.6139 0.5784 0.2579 0.399  0.4729 0.     0.     0.
 0.     0.     0.     0.     0.    ]

预测

In [21]:
!python predict.py \
       --config configs/ocrnet/ocrnet_hrnetw18_cityscapes_1024x512_160k_lovasz_softmax.yml \
       --model_path output/iter_7000/model.pdparams \
       --image_path ../infer/4346.png \
       --save_dir output/result

/opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages/setuptools/depends.py:2: DeprecationWarning: the imp module is deprecated in favour of importlib; see the module's documentation for alternative uses
  import imp
2021-04-08 10:59:44 [INFO]	
---------------Config Information---------------
SOLVER:
  CROSS_ENTROPY_WEIGHT: dynamic
  LR: 0.005
  LR_POLICY: poly
  NUM_EPOCHS: 40
  OPTIMIZER: sgd
batch_size: 4
iters: 35000
learning_rate:
  decay:
    end_lr: 0.0
    power: 0.9
    type: poly
  value: 0.0025
loss:
  coef:
  - 1
  - 0.4
  types:
  - coef:
    - 0.8
    - 0.2
    losses:
    - type: CrossEntropyLoss
    - type: LovaszSoftmaxLoss
    type: MixedLoss
  - coef:
    - 0.8
    - 0.2
    losses:
    - type: CrossEntropyLoss
    - type: LovaszSoftmaxLoss
    type: MixedLoss
model:
  backbone:
    pretrained: https://bj.bcebos.com/paddleseg/dygraph/hrnet_w18_ssld.tar.gz
    type: HRNet_W18
  backbone_indices:
  - 0
  type: OCRNet
optimizer:
  momentum: 0.9
  type: sgd
  weight_decay: 4.0e-05
train_dataset:
  dataset_root: /home/aistudio/
  mode: train
  num_classes: 15
  train_path: /home/aistudio/train_list.txt
  transforms:
  - max_scale_factor: 2.0
    min_scale_factor: 0.5
    scale_step_size: 0.25
    type: ResizeStepScaling
  - max_rotation: 30
    type: RandomRotation
  - type: RandomHorizontalFlip
  - type: RandomVerticalFlip
  - crop_size:
    - 1024
    - 512
    type: RandomPaddingCrop
  - type: RandomBlur
  - brightness_range: 0.4
    contrast_range: 0.4
    saturation_range: 0.4
    type: RandomDistort
  - type: Normalize
  type: Dataset
val_dataset:
  dataset_root: /home/aistudio/
  mode: val
  num_classes: 15
  transforms:
  - type: Normalize
  type: Dataset
  val_path: /home/aistudio/val_list.txt
------------------------------------------------
W0408 10:59:44.776221 13753 device_context.cc:362] Please NOTE: device: 0, GPU Compute Capability: 7.0, Driver API Version: 10.1, Runtime API Version: 10.1
W0408 10:59:44.776278 13753 device_context.cc:372] device: 0, cuDNN Version: 7.6.
2021-04-08 10:59:49 [INFO]	Loading pretrained model from https://bj.bcebos.com/paddleseg/dygraph/hrnet_w18_ssld.tar.gz
2021-04-08 10:59:49,681 - INFO - Lock 140715049213264 acquired on /home/aistudio/.paddleseg/tmp/hrnet_w18_ssld
2021-04-08 10:59:49,681 - INFO - Lock 140715049213264 released on /home/aistudio/.paddleseg/tmp/hrnet_w18_ssld
2021-04-08 10:59:50 [INFO]	There are 1525/1525 variables loaded into HRNet.
2021-04-08 10:59:50 [INFO]	Number of predict images = 1
2021-04-08 10:59:50 [INFO]	Loading pretrained model from output/iter_7000/model.pdparams
2021-04-08 10:59:51 [INFO]	There are 1583/1583 variables loaded into OCRNet.
2021-04-08 10:59:51 [INFO]	Start to predict...
1/1 [==============================] - 1s 527ms/step
In [42]:
%matplotlib inline
import matplotlib.pyplot as plt
img = Image.open('../infer/4346.png')
In [43]:
# 原始图片
img
Out[43]:
In [ ]:
img = Image.open('output/result/added_prediction/4346.png')
In [41]:
# 预测结果
img
Out[41]:
In [ ]:
img = Image.open('output/result/pseudo_color_prediction/4346.png')
In [39]:
# 伪彩色标注
img
Out[39]:

导出静态图模型

In [9]:
!python export.py \
       --config configs/ocrnet/ocrnet_hrnetw18_cityscapes_1024x512_160k_lovasz_softmax.yml \
       --model_path output/iter_8000/model.pdparams

Python预测部署

In [11]:
!ls

benchmark  deploy     legacy   paddleseg     README.md	       slim	 val.py
configs    docs       LICENSE  predict.py    requirements.txt  tools
contrib    export.py  output   README_CN.md  setup.py	       train.py
In [12]:
# 把infer.py复制到PaddleSeg主目录下
!cp deploy/python/infer.py infer.py
In [13]:
# 预测的伪彩色图片默认放在./output目录下
!python infer.py --config output/deploy.yaml --image_path ../infer/4346.png

/opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages/setuptools/depends.py:2: DeprecationWarning: the imp module is deprecated in favour of importlib; see the module's documentation for alternative uses
  import imp

番外篇：用脚本任务处理数据分析和训练任务

Python的OS命令与常用的终端命令行

  • 显示当前路径，对应终端bash中pwd
    import os
print(os.getcwd())
  • git获取PaddleSeg套件，对应终端bash中git clone https://gitee.com/paddlepaddle/PaddleSeg.git
    import os
os.system("git clone https://gitee.com/paddlepaddle/PaddleSeg.git")
  • 安装依赖库，并指定源
    import os
os.system("cd PaddleSeg && pip install -r requirements.txt -i https://mirror.baidu.com/pypi/simple")
  • 执行python文件 
    import os
os.system("python make_list.py")
    ## 脚本任务的基本操作 [脚本任务示例合集](https://aistudio.baidu.com/aistudio/projectdetail/913043) https://ai-studio-static-online.cdn.bcebos.com/c93f730754234f0a8b5e68ddf669817393d1722a4c524d77b1f8234e44282f16