lynetcha / completion3d Goto Github PK

the code for the tree structure is just residual connection? The novel decoder is just mutiple use of folding net?

Change pytorch split to validation.

Hi, what's your test data "*.h5" order? Are you following the test.list? Why did you rename them from "0000.h5" to "1199.h5"?

The page remains in 'Please be patient while we verify that the file contents match the expected format.' for quite a long time. I wonder whether the server is broken.

In the TopNet paper, there is a version to generate N=16384 points.

For N=2048, it is clearly demonstrated in the Sec. 4.2:

When generating N = 2048 points, the root node has 4 children and all other internal nodes in subsequent level generate 8 children. yielding a total of N = 4 × (8^3) = 2048 points generated by the decoder. Each MLP in the decoder is a has 3 stages with 256, 64, and C channels respectively, where C = 8 for inner nodes and C = 3 for leaf nodes.

Could you provide me more details of the network generating 16,384 points?
Thank you!

503:service unavailable

Just followed the steps in tensorflow-setup.md. And when I building Chamfer and EMD functions met this error.

I'm running in windows environment tho.

Hi, I get 503 service unavailable error for the given link for downloading the dataset.

Any suggestion to get the dataset?

THX!

Is this dataset still available for downloading? The link in the project webpage has no response...

Hi,

I looked into the create_decoder function of TopNet. It looks more like a sequential structure to me, rather than the proposed tree structure in the paper. the inp variable at each level is fed directly into next level without tree expansion. May I know how to interpret it as a tree?
Thanks!

    def create_level(self, level, input_channels, output_channels, inputs, bn):
        with tf.variable_scope('level_%d' % (level), reuse=tf.AUTO_REUSE):
            features = mlp_conv(inputs, [input_channels, int(input_channels/2),
                                         int(input_channels/4), int(input_channels/8),
                                         output_channels*int(self.args.tarch[level])],
                                        self.args.phase, bn)
            features = tf.reshape(features, [tf.shape(features)[0], -1, output_channels])
        return features

    def create_decoder(self, code):
        Nin = self.args.NFEAT + self.args.code_nfts
        Nout = self.args.NFEAT
        bn = True
        N0 = int(self.args.tarch[0])
        nlevels = len(self.args.tarch)
        with tf.variable_scope('decoder', reuse=tf.AUTO_REUSE):
            level0 = mlp(code, [256, 64, self.args.NFEAT * N0], self.args.phase, bn=True)
            level0 = tf.tanh(level0, name='tanh_0')
            level0 = tf.reshape(level0, [-1, N0, self.args.NFEAT])
            outs = [level0, ]
            for i in range(1, nlevels):
                if i == nlevels - 1:
                    Nout = 3
                    bn = False
                inp = outs[-1]
                y = tf.expand_dims(code, 1)
                y = tf.tile(y, [1, tf.shape(inp)[1], 1])
                y = tf.concat([inp, y], 2)
                outs.append(tf.tanh(self.create_level(i, Nin, Nout, y, bn), name='tanh_%d' % (i)))
        return outs[-1]

Hi

the provided link is not working. I can not download the dataset.

2048K: http://download.cs.stanford.edu/downloads/completion3d/dataset2019.zip
16384K: http://download.cs.stanford.edu/downloads/completion3d/shapenet16K2019.zip

Hi~ I have trouble submitting my results. I'm sure that the format of my submission is matched with yours but I still can't submit results and get 'Please be patient while we verify that the file contents match the expected format.' The size of my submission.zip is about 27MB. The format is like 'all/0000.h5'.

Hello,
thanks a lot for your great work! Currently, I am trying to integrate the "completion3d" dataset into torch-point3d. To load the point cloud "h5" files I just used the same code as presented in this repo.

def load_h5(path, key='data', dtype=None, device=None):
    f = h5py.File(path, 'r')
    data = torch.tensor(f[key], dtype=dtype, device=device)
    f.close()
    return data

The only thing is that it takes about 200 milliseconds to load one h5-file, which is really a lot of time for ~50KB data!
My env:
OS: Ubunut 20.04
RAM: 32GB
Drive: SSD

Are you also observing the same effect?
Do you plan to provide the dataset in other formats like plain text-file!

Thanks

Hi, all,

I followed the pytorch-setup.md. However, when I build the EMD module, I got this building error. My version of pytorch is python3-pytorch-0.4.1.

According to the building step on EMD module, only python-27-pytorch-0.11.1 is suggested. But pytorch-0.11.1 seems quite old.

Any suggestion to fix this issue?

THX!

Hi~ thanks for providing the dataset!

I have download it and use the 'Car' category(02958343) specially, and I found that the model seems to be 'slanting' : this two points looks like in the same height(z-dir) but they have completely different z-dir coordinates.
Could you please tell me what should I do if I want to normalize the model to [0,1] ?

Hi, When I follow the instructions to train a TopNet model with 16384 output points, I found the batchsize can be only set to 1 for the high memory occupation to compute a chamfer distance, can you give me some suggestions? Besides, as mentioned in the paper, the ground truth points number keeps 2048 while the output points number is 16384, and it seems that somewhere should be moified in the code, because the emd can't be computed for the different number of ground truth points and output points?

Hi, @lynetcha ,

In addition to tensorflow implementation, is there a pytorch implementation on topnet model?

THX!

I followed the instruction to train the topnet for 300 epochs using the data provided.
However, the results don't seem as good as the paper presents.
Could you please provide your pre-trained model?

Is it plausible to train the model on a given class, if so, how to do it?

python setup.py install

Hi, thanks for your great work! I am running your code rencently, however, i met a problem that descrobed as the title:

ValueError: too many values to unpack (expected 2)

Then i output the shape of "outputs", its shape was (2,2116,3), while the shape of "targets" is (2,16384,3). There must be something wrong. Is there a problem in your AtlasNet model? Cause the input (2,3,2048) was feed into your model.forward(), and get the outputs was (2,2116,3), isn't the shape should be (2,16384,3)??

Thanks in advance!

Hi,

I am unable to download the dataset. I am getting the 'Not Found' webpage error. Kindly help.