alex04072000/sold

sold's Issues

Some questions about PWC-Net

• As stated in the README, the tfoptflow implementation of PWC-Net was used. However, tfoptflow has some alarming yet unresolved issues (e.g., regarding scaling and flow equation). What modification(s) were employed in this work?

• Also, the Requirements Setup says that:

  Please overwrite tfoptflow/model_pwcnet.py and tfoptflow/model_base.py using the ones in this repository.

  The "ones in this repository"—which I believe are to be used in place of model_pwcnet.py and model_base.py from the tfoptflow code base—seem to be missing.

overwrite files missing!

In your README, you said that Please overwrite tfoptflow/model_pwcnet.py and tfoptflow/model_base.py using the ones in this repository., but there are no files related to these two.
Thanks!

Code for Meta Learning with Reptile missing

Hi Alex,

Thanks for the inspiring work and the implementation.

Did you miss out checking in the code for Section 3.6 Meta Learning?

Thanks

Where are these checkpoints?

I tried the repo but encountering some problem, it seems missing checkpoints.

I found in the colab demo, the model.ckpt-239999 are successfully loaded.
INFO:tensorflow:Restoring parameters from train_dir_initFlow_Fence/model.ckpt-239999
INFO:tensorflow:Restoring parameters from train_dir_imgReconstruction_Fence/model.ckpt-239999

What are these checkpoints? I downloaded the ckpt.zip and unzip it, get some checkpoints with 'ckpt_' prefix.

Thanks.

Comparing to Alayrac et al.

Great work. Just wondering how did you use 'The Visual Centrifuge: Model-Free Layered Video Representations' paper in your paper comparison? Did you retrain that on your data? Or did you get their weights?

Supplementary Material missing in version 2 of the paper

Hi, I noticed that a version 2 of the arXiv paper is now available. However, this version has no Supplementary Material (which was present in version 1). Does this mean that the dataset generation and network architecture are different from that of v1? If so, will you be uploading a corresponding Supplementary Material for v2? Thanks.

How to use multiple image sequences for online training?

Say I have two five-frame image sequences, seq1_I{0,1,2,3,4}.png and seq2_I{0,1,2,3,4}.png. How can I run online training to use the said image sequences?

Looking at train_fence_online.py, I'm guessing it will have something to do with the --batch_size and the --training_scene flags:

SOLD/train_fence_online.py

Lines 22 to 23 in 519ee0d

	tf.app.flags.DEFINE_integer(
	'batch_size', 1, 'The number of samples in each batch.')

SOLD/train_fence_online.py

Lines 26 to 28 in 519ee0d

	tf.app.flags.DEFINE_string('training_scene', None,
	"""If specified, restore this pretrained model """
	"""before beginning any training.""")

However, I cannot tell how to specify the image sequence. Do I pass --batch_size 2 --training_scene seq when calling !python train_fence_online.py?