How to train the attribute classifier to test custom attributes?

[DateBro]

Hi, csmliu, thank you so much for your brilliant work! I want to use STGAN as a baseline in my paper and want to get the attribute classification accuracy of STGAN on some attributes instead of your predefined attributes. But I find it confusing about the required tfrecord data, which is a little different from LynnHo/TfrecordCreator. Just want to avoid potential error, could you give a more detailed tutorial for the training of attribute classifier?

[csmliu]

Hi DateBro, Thanks for your attention.

1. The provided test model is trained with all 40 attributes provided by CelebA, and you only need to add other attributes into the att_id (https://github.com/csmliu/STGAN/blob/ce592029501724fddc47019cd2b25624ad166c1f/att_classification/test.py#L34-L46), the overall accuracy is similar to (should be slightly higher than) the value reported in the AttGAN paper.
2. I've found the code to generate the tfrecord file, and I did modify the code for easier use. If you need to train your own attribute classifier, you can use Create_TFRecord.zip to generate a training tfrecord file.

If there are any other problems, please feel free to reopen this repo.

[DateBro]

I tested the commands in https://github.com/csmliu/STGAN/blob/master/att_classification/README.md, but got some trouble in tackling it. When using tensorflow-gpu1.15, I got errors like

(0) Failed precondition: Attempting to use uninitialized value classifier/Conv_2/weights
     [[node classifier/Conv_2/weights/read (defined at /home/zhiyong/.local/lib/python3.6/site-packages/tensorflow_core/python/framework/ops.py:1748) ]]
     [[Cast/_3]]
  (1) Failed precondition: Attempting to use uninitialized value classifier/Conv_2/weights
     [[node classifier/Conv_2/weights/read (defined at /home/zhiyong/.local/lib/python3.6/site-packages/tensorflow_core/python/framework/ops.py:1748) ]]
0 successful operations.
0 derived errors ignored.

When using a new virtual environment of tf-gpu1.12 or 1.4, I always got

WARNING:tensorflow:From /home/zhiyong/RemoteServer/pycharm_projects/STGAN/att_classification/tflib/collection.py:62: The name tf.GraphKeys is deprecated. Please use tf.compat.v1.GraphKeys instead.

WARNING:tensorflow:
The TensorFlow contrib module will not be included in TensorFlow 2.0.
For more information, please see:
  * https://github.com/tensorflow/community/blob/master/rfcs/20180907-contrib-sunset.md
  * https://github.com/tensorflow/addons
  * https://github.com/tensorflow/io (for I/O related ops)
If you depend on functionality not listed there, please file an issue.

段错误 (核心已转储

I have no idea what to do, can you give me some advice?

[csmliu]

TF 1.12 should work according to my experience, and should not raise such warnings. Please check that whether you are using the right version by import tensorflow as tf; print(tf.__version__)

[csmliu]

I remember that, the warning about TF 2.0 occurs in TF 1.13 or 1.14, so maybe you are using the wrong version.

[DateBro]

Sorry, I forgot some features of Anaconda and use TF 1.15 as TF 1.12. But when I used TF 12 correctly, there is still the same error as TF 1.15.

tensorflow.python.framework.errors_impl.FailedPreconditionError: Attempting to use uninitialized value classifier/Conv/weights
     [[Node: classifier/Conv/weights/read = Identity[T=DT_FLOAT, _device="/job:localhost/replica:0/task:0/device:GPU:0"](classifier/Conv/weights)]]
     [[Node: Cast/_3 = _Recv[client_terminated=false, recv_device="/job:localhost/replica:0/task:0/device:CPU:0", send_device="/job:localhost/replica:0/task:0/device:GPU:0", send_device_incarnation=1, tensor_name="edge_228_Cast", tensor_type=DT_INT64, _device="/job:localhost/replica:0/task:0/device:CPU:0"]()]]

[csmliu]

Pretty weird error. Have you modified the network architecture?

BTW, to make sure that the environment is correctly set, it's better to create the virtual environment by

conda create -n NAME tensorflow-gpu=1.12

That is, assigning the tf version when creating the environment (in this way, anaconda will automatically install packages which will not cause conflicts, e.g., an older version of python.)

[csmliu]

Please refer to https://stackoverflow.com/questions/44624648/tensorflow-attempting-to-use-uninitialized-value-in-variable-initialization

[DateBro]

I only modified basic.py

return imageio.imread(path, pilmode=mode) / 127.5 - 1

and set the test_tfrecord_path = './tfrecords/test' to absolute path, because when I use relative path I got no such file or directory. I run the test.py on Ubuntu18.04 with RTX2070, should I switch to windows? It seems that your train.py was run in Windows. I am still confused about the advice in stackoverflow, shouldn't test.py just read in the checkpoint file and predict?

[DateBro]

Thanks for your detailed help, I can run the test after add the code in StackOverflow. 👍

[DateBro]

I got different accuracy by the command

python test.py --experiment_name 128 --test_int 2 --dataroot mydataroot
python att_classification/test.py --img_dir ./output/128/sample_testing

First test:

Acc.
[0.64472498 0.84395351 0.64632802 0.13325318 0.17969141 0.12954614
 0.93497646 0.39469993 0.50490933 0.90997896 0.16245867 0.93377417
 0.35196874]

Second test:

Acc.
[0.02609959 0.40992886 0.27326921 0.13325318 0.18154494 0.20203386
 0.92084961 0.48176535 0.49509067 0.39339746 0.76605551 0.04207995
 0.75288047]

Is there something I forget to do? The results are so weird.

[csmliu]

Did you run the two commands twice or only run the second command twice?

[DateBro]

I only repeat the second command several times and find the results are different from each other and your quantitative.results in #18 .

[csmliu]

Well, maybe I have found the problem. Please use the original code, and test via

cd att_classification
python test.py --img_dir ../output/128/sample_testing

[DateBro]

emmm The results still have the same problem.

[csmliu]

Could you download the repo and try again using the original att_classification folder?

[csmliu]

I've just tested the code with Ubuntu 18.04 and TensorFlow 1.12.

[DateBro]

After cloning the repo and try again, I got the correct accuracy as quantitative.results. In the last repo, I only add the init code following StackOverflow to make it run. Anyway, thanks for your help and I'll figure out what's the problem in my machine or the modified code.