frankcan / gapointnet Goto Github PK

Sorry to bother you again. Recently I have to quote your article in some work, so I am writing the relevant pytorch code.
In the picture, spatial transform network of the paper illustration shows NxKx6. I understand that the 6D dimension feature is fed into the spatial transform network, but the [output_dim] defined by [gat_layers.py] in the code is 16-dimensional (16 channels), which means that in your paper the spatial transform network called [input_transform_net] has two inputs, one is 19-dimensional feature called [neighbors_features], and the other is 16-dimensional feature called [locals_max_transform]. What does this have to do with the NxKx6 displayed in the transformation matrix in the image? Did I understand wrong?, I am eager to get your answer, thank you very much.

Hello,When I run python.py , everything is fine. When I continue to run python evaluate.py --model=network --model_path=log/epoch_185_model.ckpt an error occurred
"raceback (most recent call last):
File "evaluate.py", line 169, in
evaluate(num_votes=12)
File "evaluate.py", line 78, in evaluate
saver.restore(sess, MODEL_PATH)
File "/home/student908/anaconda2/lib/python2.7/site-packages/tensorflow/python/training/saver.py", line 1538, in restore
+ compat.as_text(save_path))
ValueError: The passed save_path is not a valid checkpoint: log/epoch_185_model.ckpt",
I am using Xshell to connect to the server to run the program. I am not good at code, I want to know how to solve it, thank you

Getting following error while running part seg script

Traceback (most recent call last):
File "train_multi_gpu.py", line 395, in
train()
File "train_multi_gpu.py", line 377, in train
eval_one_epoch(epoch)
File "train_multi_gpu.py", line 332, in eval_one_epoch
pointclouds_phs[1]: cur_data[begidx: endidx, ...],
IndexError: list index out of range

Hello!

The more I work with your code, the more I like your approach!

Another thing. Could it be that the skip connections in the code are not properly reflected in the paper?

Here are the skip connections:
https://github.com/FrankCAN/GAPointNet/blob/master/models/network.py#L39
https://github.com/FrankCAN/GAPointNet/blob/master/models/network.py#L62

The paper mentions "skip-connection" exactly once. Would it make sense to show them in figure 3?

What do you think?

Hello!

I have decided to reimplement the paper. The approach of using attention to graph networks is excellent. My main motivation was putting a little bit more OOP on the table, plus Keras functionality.

Here it is: https://github.com/AI-Guru/pointcloud_experiments

It includes PointNet for comparisons.

Hello!

Sorry to bother you. I have worked with your code for a couple of days now. I am very excited and I found a couple of things that might be relevant.

Here is another one. In the paper you mention Multi-Layer Perceptrons as the basis for the transformer/attention-layers. So far so good. In the implementation, there are a lot of 2D convolutions with kernel-size (1, 1). I think this turns the convolutions effectively into plain dense-layers.
Example:
https://github.com/FrankCAN/GAPointNet/blob/master/models/network.py#L64

What do you think?

Hello!

First of all, I really like your approach. This is some very solid work! Congratulations.

I think I have found a slight discrepancy. Between lines
https://github.com/FrankCAN/GAPointNet/blob/master/models/network.py#L32
and
https://github.com/FrankCAN/GAPointNet/blob/master/models/gat_layers.py#L56

You see, attn_feature() returns "ret, coefs, edge_feature", while getmodel() unpacks the values as "edge_feature, coefs, locals". The order is slightly different. What do you think?

Namaste!

I think this is my final issue :D

In the implementation of the classifier, you start with a single-head-attention, then you do the spatial transformation followed by multi-head-attention. In the paper, on the other hand, you start with the transformation.

I have the feeling that I have missed something. What is your opinion?