3-unet-camvid

unet图像分割

导入包，数据集

%reload_ext autoreload
%autoreload 2
%matplotlib inline

from fastai.vision import *
from fastai.callbacks.hooks import *
from fastai.utils.mem import *

path = untar_data(URLs.CAMVID)
path.ls()

[WindowsPath('C:/Users/Wither8848/.fastai/data/camvid/codes.txt'),
 WindowsPath('C:/Users/Wither8848/.fastai/data/camvid/images'),
 WindowsPath('C:/Users/Wither8848/.fastai/data/camvid/labels'),
 WindowsPath('C:/Users/Wither8848/.fastai/data/camvid/valid.txt')]

path_lbl = path/'labels'
path_img = path/'images'

查看数据集

查看图像

fnames = get_image_files(path_img)
fnames[:3]

[WindowsPath('C:/Users/Wither8848/.fastai/data/camvid/images/0001TP_006690.png'),
 WindowsPath('C:/Users/Wither8848/.fastai/data/camvid/images/0001TP_006720.png'),
 WindowsPath('C:/Users/Wither8848/.fastai/data/camvid/images/0001TP_006750.png')]

img_f = fnames[0]
img = open_image(img_f)
img.show(figsize=(5,5))

查看标签

lbl_names = get_image_files(path_lbl)
lbl_names[:3]

[WindowsPath('C:/Users/Wither8848/.fastai/data/camvid/labels/0001TP_006690_P.png'),
 WindowsPath('C:/Users/Wither8848/.fastai/data/camvid/labels/0001TP_006720_P.png'),
 WindowsPath('C:/Users/Wither8848/.fastai/data/camvid/labels/0001TP_006750_P.png')]

print(fnames[0].suffix)
fnames[0].stem

.png





'0001TP_006690'

• 标签为数据集名字+‘_P’，生成这样的一个函数，自动寻找到便签

1. f’ 为字符串的代写格式，类似%f
2. path 对象的使用.stem得到文件名.suffix得到后缀

get_y_fn = lambda x: path_lbl/f'{x.stem}_P{x.suffix}'

• 这里使用了open_mask因为原来是掩码不太清楚

mask = open_mask(get_y_fn(img_f))
mask.show(figsize=(5,5), alpha=1)

src_size = np.array(mask.shape[1:])
src_size,mask.data

(array([720, 960]),
 tensor([[[ 4,  4,  4,  ...,  4,  4,  4],
          [ 4,  4,  4,  ...,  4,  4,  4],
          [ 4,  4,  4,  ...,  4,  4,  4],
          ...,
          [19, 19, 19,  ..., 30, 30, 30],
          [19, 19, 19,  ..., 30, 30, 30],
          [19, 19, 19,  ..., 30, 30, 30]]]))

• 映射储存在codes.txt里面，如4-对应建筑物

codes = np.loadtxt(path/'codes.txt', dtype=str); codes

array(['Animal', 'Archway', 'Bicyclist', 'Bridge', 'Building', 'Car', 'CartLuggagePram', 'Child', 'Column_Pole',
       'Fence', 'LaneMkgsDriv', 'LaneMkgsNonDriv', 'Misc_Text', 'MotorcycleScooter', 'OtherMoving', 'ParkingBlock',
       'Pedestrian', 'Road', 'RoadShoulder', 'Sidewalk', 'SignSymbol', 'Sky', 'SUVPickupTruck', 'TrafficCone',
       'TrafficLight', 'Train', 'Tree', 'Truck_Bus', 'Tunnel', 'VegetationMisc', 'Void', 'Wall'], dtype='<U17')

生成数据集

size = src_size//2

free = gpu_mem_get_free_no_cache()
# the max size of bs depends on the available GPU RAM
if free > 8200: bs=8
else:           bs=4
print(f"using bs={bs}, have {free}MB of GPU RAM free")

using bs=4, have 6813MB of GPU RAM free

• 这里使用图像分割数据集，验证集来自于给定得valid.txt因为这个数据集是来自于一个视频，为了保证验证集不连续

src = (SegmentationItemList.from_folder(path_img)
       .split_by_fname_file('../valid.txt')
       .label_from_func(get_y_fn, classes=codes))

data = (src.transform(get_transforms(), size=size, tfm_y=True)
        .databunch(bs=bs)
        .normalize(imagenet_stats))

• 查看数据集

data.show_batch(2, figsize=(10,7))

生成模型

name2id = {v:k for k,v in enumerate(codes)}
void_code = name2id['Void']

def acc_camvid(input, target):
    target = target.squeeze(1)
    mask = target != void_code
    return (input.argmax(dim=1)[mask]==target[mask]).float().mean()

metrics=acc_camvid
wd=1e-2
learn = unet_learner(data, models.resnet34, metrics=metrics, wd=wd)

lr_find(learn)
learn.recorder.plot()

lr=3e-3
learn.fit_one_cycle(10, slice(lr), pct_start=0.9)

epoch	train_loss	valid_loss	acc_camvid	time
0	1.069481	0.792951	0.815857	01:54
1	0.744278	0.600508	0.847082	01:51
2	0.649365	0.559110	0.848990	01:49
3	0.598212	0.453965	0.875559	01:51
4	0.617887	0.492955	0.866971	01:49
5	0.573497	0.563069	0.854965	01:51
6	0.552920	0.494836	0.874791	01:48
7	0.519904	0.403810	0.891915	01:53
8	0.554230	0.464974	0.883040	01:52
9	0.427734	0.338918	0.903289	01:53

learn.save('stage-1')

learn.load('stage-1');

learn.show_results(rows=3, figsize=(8,9))

lr=3e-3
learn.unfreeze()
lrs = slice(lr/400,lr/4)
learn.fit_one_cycle(12, lrs, pct_start=0.8)

epoch	train_loss	valid_loss	acc_camvid	time
0	0.378398	0.326795	0.906832	02:02
1	0.378235	0.313857	0.910155	01:54
2	0.375283	0.307506	0.911930	01:56
3	0.364699	0.321208	0.909612	01:56
4	0.353212	0.285177	0.920163	01:56
5	0.347013	0.309324	0.912606	01:53
6	0.333858	0.323652	0.905763	01:55
7	0.332646	0.302886	0.914157	01:57
8	0.328257	0.286069	0.922094	01:56
9	0.320706	0.287675	0.922618	01:55
10	0.294463	0.279678	0.922595	01:53
11	0.266878	0.272097	0.925978	01:54

• 学习率曲线应该是先高后低，如果一直降低，尝试加大一点点学习率

learn.recorder.plot_losses()
learn.recorder.plot_lr()

learn.save('stage-2');

使用原始分辨率进一步训练

learn.destroy()

size = src_size

free = gpu_mem_get_free_no_cache()
# the max size of bs depends on the available GPU RAM
if free > 8200: bs=3
else:           bs=1
print(f"using bs={bs}, have {free}MB of GPU RAM free")

this Learner object self-destroyed - it still exists, but no longer usable
using bs=1, have 6027MB of GPU RAM free

data = (src.transform(get_transforms(), size=size, tfm_y=True)
        .databunch(bs=bs)
        .normalize(imagenet_stats))

• 使用16位浮点数加快运算，甚至结果可能会更好

learn = unet_learner(data, models.resnet34, metrics=metrics, wd=wd).to_fp16()

learn.load('stage-2');
lr_find(learn)
learn.recorder.plot()

ame}.recorder.plot() to see the graph.

[外链图片转存失败,源站可能有防盗链机制,建议将图片保存下来直接上传(img-GesvvZoT-1583647675339)(output_45_2.png)]

lr=1e-3

learn.fit_one_cycle(10, slice(lr), pct_start=0.8)

epoch	train_loss	valid_loss	acc_camvid	time
0	0.423464	0.330359	0.911409	04:13
1	0.373553	0.334004	0.907827	04:08
2	0.354978	0.316964	0.912270	04:08
3	0.393460	0.316810	0.917438	04:08
4	0.342301	0.332319	0.917796	04:10
5	0.343624	0.313800	0.919534	04:07
6	0.328565	0.312378	0.909933	04:14
7	0.336849	0.317563	0.913902	04:20
8	0.269040	0.300068	0.918740	04:22
9	0.238536	0.281892	0.923772	04:10

learn.save('stage-1-big')

learn.load('stage-1-big');
learn.unfreeze()
lrs = slice(1e-6,lr/10)

learn.fit_one_cycle(10, lrs)
learn.save('stage-2-big')

epoch	train_loss	valid_loss	acc_camvid	time
0	0.256014	0.284205	0.923452	04:41
1	0.232119	0.283108	0.923677	04:29
2	0.235852	0.288952	0.920897	04:26
3	0.219901	0.261550	0.928312	04:24
4	0.227803	0.264939	0.927502	04:23
5	0.207069	0.275547	0.927656	04:25
6	0.199358	0.262479	0.928366	04:28
7	0.200243	0.259015	0.930449	04:31
8	0.197228	0.250665	0.932849	04:32
9	0.194318	0.265213	0.928627	04:31

learn.load('stage-2-big');
learn.show_results(rows=3, figsize=(10,10))