missing files

[rashaalshehhi]

batch_transform is missing in utils file?

[robinkraft]

You need to build the utils to get access to that function. C++ code lives here:

https://github.com/mitmul/ssai-cnn/blob/master/utils/src/transform.cpp#L100

[rashaalshehhi]

Thank you for your reply.
I did and it works with me.

Sent from my iPhone

On Jul 21, 2016, at 1:29 AM, Robin Kraft <[email protected]mailto:[email protected]> wrote:

You need to build the utilshttps://github.com/mitmul/ssai-cnn#build-utils to get access to that function. C++ code lives here:

https://github.com/mitmul/ssai-cnn/blob/master/utils/src/transform.cpp#L100

—
You are receiving this because you authored the thread.
Reply to this email directly, view it on GitHubhttps://github.com//issues/1#issuecomment-234089099, or mute the threadhttps://github.com/notifications/unsubscribe-auth/AIE9theuU0mxXA33WLiGDLHI3B93pux7ks5qXpMvgaJpZM4JOvP1.

[rashaalshehhi]

Two more questions?

1. If I want to enable GPU, what I should do?
   Should I enable one or more GPU(s) as following: python train.py --gpu 1 or python train.py --gpu 2 ??
2. I would like to try with 5 classes. As long as I know, imread doesn't work with image of 5 channels, so, how to run your method with five classes?
   Do you advise me to use CIS? each class in each image (5 images).

Rgs,
Rasha

---

From: Robin Kraft [email protected]
Sent: Thursday, July 21, 2016 1:29 AM
To: mitmul/ssai-cnn
Cc: Rasha Alshehhi; Author
Subject: Re: [mitmul/ssai-cnn] missing files (#1)

You need to build the utilshttps://github.com/mitmul/ssai-cnn#build-utils to get access to that function. C++ code lives here:

https://github.com/mitmul/ssai-cnn/blob/master/utils/src/transform.cpp#L100

You are receiving this because you authored the thread.
Reply to this email directly, view it on GitHubhttps://github.com//issues/1#issuecomment-234089099, or mute the threadhttps://github.com/notifications/unsubscribe-auth/AIE9theuU0mxXA33WLiGDLHI3B93pux7ks5qXpMvgaJpZM4JOvP1.

[zahidmadeel]

Hi guys,
I want to use this repo and your pre-trained models to check and see if it works well for segmentation of my data set. If not, I would do training on my dataset. So far, I cannot make out how to use pretrained models to obtain segmented images from my dataset. Is there a tutorial or something that can help me get going?
Regards
Adeel

[SarahwXU]

Also couldn't use the pre-trained model to predict.
In the "Prediction" part, I can't find where "results/MnihCNN_multi_2016-02-03_03-34-58/MnihCNN_multi.py" is.
Thank you

[mitmul]

Thank you for reporting it, you guys. I'll investigate it later. Sorry for inconvenience.

[mitmul]

@SarahwXU Could you explain the problem more in detail?

[polars05]

@mitmul, I'm having the same issue as @SarahwXU; for the model (MnihCNN_multi.py), using the pre-trained weights (MnihCNN_multi_2016-01-20_06-22-02/epoch-400.model), the output is merely a 1433x1433x3 array of zeroes.

I've put a single image file (22828930_15.tiff) into test_sat_dir, and the other arguments in predict.py are as follows:

• sat_size = 64
• map_size = 16
• channels = 3
• offset = 8
• batchsize = 128

Are you able to advise why the model is not giving any meaningful results? Thanks!

[mitmul]

All the values in the model output are exactly zeros? Could you give me some statistics of the output values?

[polars05]

@mitmul, yup, all the values in the output are exactly zeroes (displayed as 0.); tried to multiply by 255 as well but to no avail. In addition, ran max() on the output and it confirms that all the values in the output are zeroes.

Is it that another set of weights should be used? You seem to have provided other weight files (e.g. MnihCNN_multi_2016-01-20_06-22-02_1.372101/epoch-400.model) besides MnihCNN_multi_2016-01-20_06-22-02/epoch-400.model; do they mean something?

[mitmul]

OK, I actually almost forgot what my scripts are doing cause like so long time has passed after I ran my scripts last time. Well, could you provide the command you ran? I'd like to reproduce the problem first. Thanks.

[polars05]

@mitmul, no worries, haha; the command I used for prediction was as per what you stated in readme.md:

python scripts/predict.py
--model models/MnihCNN_multi.py
--param epoch-400.model
--test_sat_dir data/
--channels 3
--offset 8
--gpu -1 &

For the model argument, I used "models/MnihCNN_multi.py" instead of "results/MnihCNN_multi_2016-02-03_03-34-58/MnihCNN_multi.py" as I couldn't find the "results" folder you were referring to.

Similarly, for the param argument, I used "epoch-400.model" (downloaded from MnihCNN_multi_2016-01-20_06-22-0) as I couldn't find the "results/MnihCNN_multi_2016-02-03_03-34-58" folder you stated in readme.md.

I also set the gpu argument to -1 as I currently don't have a GPU.

[polars05]

Have found the issue; the canvas which is worked on by tile_patches() is not being output into get_predict() and hence, the canvas created in get_predict() remains blank - have modified the way the multiprocessing is done to output the canvas; pls see the fork here: (https://github.com/polars05/ssai-cnn/commit/faaa66fc9f5730dddcb1f7613b837143fa250de0)