project tutorial

[yzbx]

start

# torchseg
python test/pspnet_test.py --batch_size=4 --use_reg=True \
--backbone_pretrained=False --midnet_scale=15 \
--upsample_type=bilinear --dataset_use_part=32 \
--hyperkey=dataset.norm_ways --test=hyperopt \
--note=hyperopt010 --hyperopt=loop

# imagecls
for depth in 3 6 9
do 
    python test/train.py --net_name=fc_net --dataset_name=cifar100 \
--note=conv_number_wise_${depth} --fcc_block_number=${depth} \
--fcc_width_wise_ways=conv_number_wise 
done

# deeplab
python src/pspnet.py --logtostderr --model_variant=xception_65 \
--dataset=pascal_voc_seg --train_batch_size=2 --eval_batch_size=2 \
--train_crop_size=512  --train_crop_size=512 --eval_crop_size=512 \
--eval_crop_size=512 --note=bs2_512_name

[yzbx]

develop

• use_lr_mult
• config & hparm
• evaluation loop
  • wait for new checkpoint
  • evaluate repeatedly
• image augmentation
• high level api
  • tf.keras
  • tf.estimator
  • add metric
  • input fn

[yzbx]

dataset

voc2012

class number=21 class ratio in train dataset

background 0.7335528011120875
aeroplane 0.0071760763900985056
bicycle 0.0030561376986442553
bird 0.0089963803892935
boat 0.006102756087606602
bottle 0.006114548424663869
bus 0.017634589698972022
car 0.014084503806748567
cat 0.0272139210062373
chair 0.01153074142977194
cow 0.00830592011269573
diningtable 0.013628620761325854
dog 0.017510979729770577
horse 0.009203902952435257
motorbike 0.011641773174791661
person 0.048345571382578896
pottedplant 0.00673178820240435
sheep 0.00908591509881648
sofa 0.014557970572807251
train 0.01605723766656555
tvmonitor 0.009467864301684318

input image size	number
(375, 500)	598
(333, 500)	199
(500, 375)	114

cityscapes

input image size = (1024,2048) class number = 19 class ratio in train dataset

road 0.3276572161101393
sidewalk 0.054050920646456835
building 0.20282762061597998
wall 0.005829324406525249
fence 0.007801711883949818
pole 0.010908623685334776
traffic light 0.0018481932869411077
traffic sign 0.004899685641206073
vegetation 0.14146105146348614
terrain 0.010289616533938619
sky 0.03572185008073886
person 0.010841261042926983
rider 0.0011997494104227298
car 0.06210579377717644
truck 0.002368482198075143
bus 0.0020918691895904433
train 0.002071525048384442
motorcycle 0.000877411320154188