training data

#"""
#with open(config.DATA_FLIST[config.DATASET][0]) as f:
    #fnames = f.read().splitlines()
fnames=['img010.png','img000.png']

I change the code in train.py. Then it can run now. In the end it will give me information like this: 2018-10-15 15:52:31 @weights_viewer.py:60] Total size of trainable weights: 0G 10M 108K 392B (Assuming32-bit data type.) But it also produced the model. Then I run the test.

tensorflow.python.framework.errors_impl.DataLossError: Unable to open table file model_logs\20181015153040966844_imagenet_NORMAL_wgan_gp_LOG_DIR\events.out.tfevents.1539588647.NBA1903: Data loss: not an sstable (bad magic number): perhaps your file is in a different file format and you need to use a different restore operator?

Is this means my training fails? Should I increase my training dataset or doing something else to solve the problem? I am still trying it on Win10.

@aiueogawa Hi, I have opened a specific issue for you. You have asked five questions and I have answered all your questions. If you do not understand each one, please ask here and we can communicate here. Thanks.

The change of loss. the loss of D_net is: g_loss, d_loss = self.gan_hinge_loss(pos_global, neg_global) the complete loss of G loss is : losses['g_loss'] = a1 * g_loss +a2 * losses['l1_loss'] Q1.What happens to this g_loss of D_net , it can drop to -19. Of cource ,this result is awful. what's the range of g_loss, d_loss? Q2.the a1, a2 that weights of complete losses['g_loss'] is 1 ? I'm looking forward to your reply.

PS. I change loss below to make g_loss and l1 loss same order of magnitude. d_loss = tf.reduce_mean(tf.nn.softplus(-pos)) + tf.reduce_mean(tf.nn.softplus(neg)) g_loss = tf.reduce_mean(tf.nn.softplus(-neg)) Below is my result. Still has artifacts.

Originally posted by @huangqianfirst in https://github.com/JiahuiYu/generative_inpainting/issues/172#issuecomment-438930420

Hi, Thank for your work. I want to generate the fixed mask for my image with the mask's size = 1/2 image's size and the mask is on the right side of original image. so what i should change to get that because so far your code generate random mask in random_bbox and bbox2mask function. Original

mask Thanks

During handling of the above exception, another exception occurred:

Traceback (most recent call last): File "train.py", line 6, in import tensorflow as tf File "C:\Users\admin\AppData\Local\Programs\Python\Python37\lib\site-packages\tensorflow_init_.py", line 24, in from tensorflow.python import pywrap_tensorflow # pylint: disable=unused-import File "C:\Users\admin\AppData\Local\Programs\Python\Python37\lib\site-packages\tensorflow\python_init_.py", line 49, in from tensorflow.python import pywrap_tensorflow File "C:\Users\admin\AppData\Local\Programs\Python\Python37\lib\site-packages\tensorflow\python\pywrap_tensorflow.py", line 74, in raise ImportError(msg) ImportError: Traceback (most recent call last): File "C:\Users\admin\AppData\Local\Programs\Python\Python37\lib\site-packages\tensorflow\python\pywrap_tensorflow.py", line 58, in from tensorflow.python.pywrap_tensorflow_internal import * File "C:\Users\admin\AppData\Local\Programs\Python\Python37\lib\site-packages\tensorflow\python\pywrap_tensorflow_internal.py", line 28, in _pywrap_tensorflow_internal = swig_import_helper() File "C:\Users\admin\AppData\Local\Programs\Python\Python37\lib\site-packages\tensorflow\python\pywrap_tensorflow_internal.py", line 24, in swig_import_helper _mod = imp.load_module('_pywrap_tensorflow_internal', fp, pathname, description) File "C:\Users\admin\AppData\Local\Programs\Python\Python37\lib\imp.py", line 242, in load_module return load_dynamic(name, filename, file) File "C:\Users\admin\AppData\Local\Programs\Python\Python37\lib\imp.py", line 342, in load_dynamic return _load(spec) ImportError: DLL load failed: 找不到指定的模块。

Failed to load the native TensorFlow runtime.

See https://www.tensorflow.org/install/errors

for some common reasons and solutions. Include the entire stack trace above this error message when asking for help.

Really great works your deepfill! Is it possible to build graph once then feed it with multi images/masks with different resolutions?

Similar to https://github.com/JiahuiYu/generative_inpainting/issues/12#issuecomment-381798944 a loop was created to process multi images but he built graph for every input image.

In your code https://github.com/JiahuiYu/generative_inpainting/issues/12#issuecomment-381807753 graph only has to be built once but all image/mask should be resized to same size then fed to the graph.

Can we build graph once and feed images/masks with different resolutions?

Thanks.

Hi @JiahuiYu

I saw the numeric results report in your paper includes mean L1 loss, mean L2 loss and mean TV loss. I wonder why they all have percentage ? for example, I used to know L1_loss=abs(real-fake), does the percentage comes from dividing 255 like L1_loss=abs(real-fake)/255?

Could I know how many epoch you have used to train your model (e.g., release_places2_256 model)? Any suggest min epoch to re-produce your impainting result?

Also, I observe the loss kept increasing during my training (e.g., from -0.47 to -0.21 and -0.06 ...). Does that indicate that I am not on the right track? (I test the resulted model and they do have bad accuracy.) What parameter should I check if it is not normal? Thanks!

|####################| 100.00%, 74660/0 sec. train epoch 1, iter 10000/10000, loss -0.472547, 0.13 batches/sec. |####################| 100.00%, 74743/0 sec. train epoch 2, iter 10000/10000, loss -0.217382, 0.13 batches/sec. |####################| 100.00%, 74793/0 sec. train epoch 3, iter 10000/10000, loss -0.061427, 0.13 batches/sec. |#################---| 87.40%, 65406/9362 sec. train epoch 4, iter 8740/10000, loss -0.031020, 0.13 batches/sec.

Hello Jiahui, thanks for sharing such wonderful work. I read through the paper and have some questions regards to both structure and implementation. Sorry I am new to machine learning and my questions may seems a bit naive.

Q1. why did you use ELU for activation of the x and sigmoid activation for the mask? are these just on experimental trial? I searched around and most of the articles suggest to avoid use the sigmoid activation it kills the gradient in deep neural network. the ELU does not have the gradient kill issue but it is very computational heavy compare to ReLU or leaky ReLU.

Q2. the cnum is number of filters used right? is the number already reduced by 25% as indicated in the paper? like 48 = 64*0.75? i saw some post uses less filter numbers. is it going to affect the training result significantly? and what would be the suggested range to pick that number from?

Q3. what is are those parameters under #train? train_spe: 4000 max_iters: 100000000 viz_max_out: 10 val_psteps: 2000 those are the default value i got and i am not sure what effect they have on the training process.

Q4, I am trying to train the ImageNet and for my local PC, i have Nvidia GTX 1060 6GB but it cannot go up to 16 batch_size, and i tried to train them with around 8 batch size but the result shows blurry inpainted region. should I adjust the l1_loss_alpha to the 0.5? or 2? I saw two issue post suggesting different values. and I only have these three parameters under #loss section. ae_loss: True l1_loss: True l1_loss_alpha: 1.

Q5. Since I don't have enough graphic memories, I applied for google cloud compute engine and have VM created using their deep learning VM from the market place. with same version of the tensorflow and tensorflow-gpu from my local computer, the VM was not able to recognize the GPU installed. I tried change the trainer to multiGPU trainer and setup the parameters in inpaint.yml. I am only using one GPU for the VM and I was using the tesla V100 16gb and the network gpu utilization just blows up and the kernel got stuck and died. in the live nvidia-smi gpu monitoring I saw error like "cannot allocate the memory". after it blows up.

Thank you very much.

Q1: According to the paper discriminator's input is image, mask and optional user input. Just like for generator net. And according to this comment https://github.com/JiahuiYu/generative_inpainting/issues/158#issuecomment-433722306 you do NOT use ones in discriminator's input. So it's looks like that?

ones_x = tf.ones_like(x)[:, :, :, 0:1]
x = tf.concat([x, ones_x*mask], axis=3)

Q2: Do you use gated convolution for ALL layers? Even for last one of coarse and refine nets?

While running the following: python inpaint_ops.py --imageA examples/style_transfer/bnw_butterfly.png --imageB examples/style_transfer/bike.jpg --imageOut examples/style_transfer/bike_style_out.png

I get the following error: Traceback (most recent call last): File "/opt/virtual_tensorflow/g_tf_neuralgym/lib/python3.5/site-packages/tensorflow/python/framework/ops.py", line 558, in set_shape unknown_shape) tensorflow.python.framework.errors_impl.InvalidArgumentError: Dimension 1 in both shapes must be equal, but are 166 and 64. Shapes are [1,166,250,2] and [1,64,64,2].

During handling of the above exception, another exception occurred:

Traceback (most recent call last): File "inpaint_ops.py", line 513, in test_contextual_attention(args) File "inpaint_ops.py", line 360, in test_contextual_attention training=False, fuse=False) File "inpaint_ops.py", line 313, in contextual_attention offsets.set_shape(int_bs[:3] + [2]) File "/opt/virtual_tensorflow/g_tf_neuralgym/lib/python3.5/site-packages/tensorflow/python/framework/ops.py", line 561, in set_shape raise ValueError(str(e)) ValueError: Dimension 1 in both shapes must be equal, but are 166 and 64. Shapes are [1,166,250,2] and [1,64,64,2].

I can see this has been logged before here: https://github.com/JiahuiYu/generative_inpainting/issues/67 But cant find any fix. Can you please let me know how I can fix this?

Hello, I have a question. I would like to know if I can use this model to

1. whether it is possible to learn grayscale images (one dimensional).
2. whether it is possible to learn RGB plus depth information for a total of 4 dimensions.

Is it possible to achieve the above in this project? If so, I would like to know what method should be used.

Thank you!

I would like to change some layers of a trained model for transition learning, but I don't know how to change only a part of the model. Can someone please tell me how to do this?

when meet yi = yi*mm in contextual_attention,the error occurs. tensorflow.python.framework.errors_impl.InvalidArgumentError: required broadcastable shapes [Op:Mul] can anyone help me?

Hi,

Congratulations for your outstanding papers, they contain several innovations and contributions.

I am conducting research on face de-occlusion and inpainting, and came across your papers. The research requires the exact replication of your DeepFill V2 experiments, since I'll use it as a baseline in my research. I couldn't find some key information neither in the paper nor in the code and YAML file.

I need the following information for CelebA and CelebA-HQ datasets:

1. Number of images used in the train and validation sets.
2. Image preprocessing, such as data augmentation, resizing, alignment, etc.
3. The CelebA-HQ paper says there are 30k images, but the dataset comes with 100k images. Where did you get this dataset? The link in inpaint.yml file is broken.
4. The CelebA link in inpaint.yml file is also broken.
5. Just to confirm, the image size during training is 256x256 and testing is 512x512?

I searched the closed issues, but couldn't find the flist files; I found just the code to create them.

Thanks, Victor.