Models: [deeplab] the size of [513,513] in train.py and eval.py

Created on 22 May 2018 · 9Comments · Source: tensorflow/models

In deeplab, thetrain.py has the following setting

flags.DEFINE_multi_integer('train_crop_size', [513, 513], image crop size [height, width] during training.')

In eval.py, it has the following setting

flags.DEFINE_multi_integer('eval_crop_size', [513, 513],| 'Image crop size [height, width] for evaluation.')

Does this setting of [513,513] has to be modified for different data set? If the image set consists of images with different size, how to do that?

Source

liangjianfans

Most helpful comment

Hi, how can I modify the depth multiplier of MobileNet v2? I want to experiment with depth multiplier=0.75

kushagraagrawal on 22 May 2018

👍2

All 9 comments

See Q8 in FAQ.
During training, it is a trade-off to use large batch size and large train crop size, which requires some experiments to verify.

aquariusjay on 22 May 2018

Hi, how can I modify the depth multiplier of MobileNet v2? I want to experiment with depth multiplier=0.75

kushagraagrawal on 22 May 2018

👍2

Hi aquariusjay thank you for the reply. The following is the result of running eval.py, what does it mean, in specific, miou_1.0[1]?

INFO:tensorflow:Restoring parameters from /data/DL-Phase3/carvana/train_on_train_set/train/model.ckpt-50000
INFO:tensorflow:Starting evaluation at 2018-05-21-21:10:35
INFO:tensorflow:Evaluation [144/1449]
INFO:tensorflow:Evaluation [288/1449]
INFO:tensorflow:Evaluation [432/1449]
INFO:tensorflow:Evaluation [576/1449]
miou_1.0[1]

liangjianfans on 22 May 2018

1 is 100% here. I guess you can visualize the segmentation results to check if 100% results from some errors.

jytime on 23 May 2018

Hi， I have met same questions. I want to train the model with different size inputs, some are [<513, <513], some are [>513, >513], some are [<513, >513], each input has different size. Actually, I have trained the model using resize_range() in preprocess_utils.py to adjust the inputs, but the miou is lower than which is trained and tested using inputs are [<513,<513].

lumiaomiao on 28 May 2018

🚀1

@kushagraagrawal
Have you find the way to change the depth multiplier of Mobilenet V2?
I tried to add a flag --depth_multiplier=1.4 to train the deeplab with mobilenetv2_1.4 but failed.
Any idea?

dohai90 on 21 Sep 2018

For me it didn't work on any crop_size expect [513, 513]. Hope it helps!

SiddhantKapil on 18 Mar 2019

Hi aquariusjay thank you for the reply. The following is the result of running eval.py, what does it mean, in specific, miou_1.0[1]?
INFO:tensorflow:Restoring parameters from /data/DL-Phase3/carvana/train_on_train_set/train/model.ckpt-50000
INFO:tensorflow:Starting evaluation at 2018-05-21-21:10:35
INFO:tensorflow:Evaluation [144/1449]
INFO:tensorflow:Evaluation [288/1449]
INFO:tensorflow:Evaluation [432/1449]
INFO:tensorflow:Evaluation [576/1449]
miou_1.0[1]
Hi, When I use my data .I have the same question with you . Can you share your solution?
Thank you very much .I 'm looking forword to your reply

TulipDi on 19 Apr 2019

Hi There,
We are checking to see if you still need help on this, as this seems to be considerably old issue. Please update this issue with the latest information, code snippet to reproduce your issue and error you are seeing.
If we don't hear from you in the next 7 days, this issue will be closed automatically. If you don't need help on this issue any more, please consider closing this.