When I run celebA_main_torch.py​​, I can not generate a checkpoint file.why?

The 'train_epoch' in OWM_CNN/owm.py always fills up the GPU's memory.
It seems that the memory space for each batch training has been accumulating.

Hi, I have two questions about your work:

1. What is the difference between OWM and GEM algorithm? GEM projects the gradient on the subspace of gradients from old samples. Is the subspace of inputs x better for solving continual learning problem?

2. Why should we update p before projecting the gradient?
   p.sub_(torch.mm(k, torch.t(k)) / (alpha + torch.mm(r, k)))
w.grad.data = torch.mm(w.grad.data, torch.t(p.data))
   This means the gradient is orthogonal to the subspace of new task and will affect the performance of on the new task.

Thank you.

[1] Lopez-Paz, David. "Gradient episodic memory for continual learning." NIPS. 2017.

RuntimeError: CUDA out of memory.

Hello, may I ask what is the memory size of your GPU?

I'm running the OWM code with 8G GPU and want to fix this(RuntimeError: CUDA out of memory.).

How can I modify this program to ensure that the experimental results are consistent with yours?

thank you

I have trained the datasets about CASIA_HWDB, but I can`t run the CHW_OWM.py. I found my "test_each chwdata_mat" and "train_each chwdata_mat" is wrong.
It seems the code about followed have the error.

ss = np.arange(train_length)
        np.random.shuffle(ss)
        trainimages = trainimages[ss, :]
        trainlabels = trainlabels[ss]

Traceback (most recent call last):
  File "/media/hdd/yike/OWM/CASIA_HWDB/CHW_OWM.py", line 60, in <module>
    trainlabels = trainlabels[ss]
IndexError: index 289 is out of bounds for axis 0 with size 1

So how I get the data?

The code of OWM_CNN can only use the pytorch0.3 version, and the 0.4 version and above will run out of memory. Would you like to ask if there is any code compatible with version 0.4 and above?

作者，您好，我想复现汉字识别的实现。但是我发现我的现存总是爆了，请问您做实验时候的显存是多大呀？self.w2.data -= lr_list[0] * dw2这里就爆了

Greetings,

I had read your paper but was unable to find any supplementary material along with it. I had used this link : https://arxiv.org/pdf/1810.01256.pdf

Thank you for your outstanding work！I have doubts about the CDP module.What is the context and how is it generated ? I found "wordvet.mat" in the source code and I think it is related to context but how is it generated ？I would appreciate it if you could answer.

Hi,
The website "http://python.jobbole.com/87509/" is unreachable. So I can`t handle CASIA_HWDB dataset. Can you give me the solution.

Running OWM/Shuffled MNIST/run_shuffled_100_mnist_3Layers_2000.py

It occurs the error.

force_layer1.force_learn(w1, input, learning_rate, alpha=alpha_array[0])
TypeError: force_learn() got an unexpected keyword argument 'alpha'

When running "OWM/CASIA_HWDB/CHW_ResNet18/CHW_3755_all.py", it occurs the following error:

Traceback (most recent call last):
  File "/media/hdd/yike/OWM/CASIA_HWDB/CHW_ResNet18/CHW_3755_all.py", line 17, in <module>
    from OWM.CHW.CHW_96_18_Norma.myresnet import *                                     ImportError: No module named 'OWM'

请问有解决方法吗？

Dear author, hello, I have read through your thesis, the idea is very good and the method is very effective. Because of my own limited level, I can't understand why the construction of the projection factor p is mathematically feasible. If you can, please Give a mathematically detailed proof, thank you!

Running the "/media/hdd/yike/OWM/Disjoint MNIST/OWM/run_dis_mnist_2Layers.py".
It occurs the following errors:

Traceback (most recent call last):
  File "/media/hdd/yike/OWM/Disjoint MNIST/OWM/run_dis_mnist_2Layers.py", line 27, in <module>
    mnist = input_data.read_data_sets("./data/MNIST_data/", one_hot=True)
  File "/home/gzz/.local/lib/python3.5/site-packages/tensorflow/contrib/learn/python/learn/datasets/mnist.py", line 240, in read_data_sets
    source_url + TRAIN_IMAGES)
  File "/home/gzz/.local/lib/python3.5/site-packages/tensorflow/contrib/learn/python/learn/datasets/base.py", line 208, in maybe_download
    temp_file_name, _ = urlretrieve_with_retry(source_url)
  File "/home/gzz/.local/lib/python3.5/site-packages/tensorflow/contrib/learn/python/learn/datasets/base.py", line 165, in wrapped_fn
    return fn(*args, **kwargs)
  File "/home/gzz/.local/lib/python3.5/site-packages/tensorflow/contrib/learn/python/learn/datasets/base.py", line 190, in urlretrieve_with_retry
    return urllib.request.urlretrieve(url, filename)
  File "/media/hdd/yike/anaconda3/envs/own/lib/python3.5/urllib/request.py", line 188, in urlretrieve
    with contextlib.closing(urlopen(url, data)) as fp:
  File "/media/hdd/yike/anaconda3/envs/own/lib/python3.5/urllib/request.py", line 163, in urlopen
    return opener.open(url, data, timeout)
  File "/media/hdd/yike/anaconda3/envs/own/lib/python3.5/urllib/request.py", line 466, in open
    response = self._open(req, data)
  File "/media/hdd/yike/anaconda3/envs/own/lib/python3.5/urllib/request.py", line 484, in _open
    '_open', req)
  File "/media/hdd/yike/anaconda3/envs/own/lib/python3.5/urllib/request.py", line 444, in _call_chain
    result = func(*args)
  File "/media/hdd/yike/anaconda3/envs/own/lib/python3.5/urllib/request.py", line 1297, in https_open
    context=self._context, check_hostname=self._check_hostname)
  File "/media/hdd/yike/anaconda3/envs/own/lib/python3.5/urllib/request.py", line 1257, in do_open
    r = h.getresponse()
  File "/media/hdd/yike/anaconda3/envs/own/lib/python3.5/http/client.py", line 1198, in getresponse
    response.begin()
  File "/media/hdd/yike/anaconda3/envs/own/lib/python3.5/http/client.py", line 297, in begin
    version, status, reason = self._read_status()
  File "/media/hdd/yike/anaconda3/envs/own/lib/python3.5/http/client.py", line 258, in _read_status
    line = str(self.fp.readline(_MAXLINE + 1), "iso-8859-1")
  File "/media/hdd/yike/anaconda3/envs/own/lib/python3.5/socket.py", line 576, in readinto
    return self._sock.recv_into(b)
  File "/media/hdd/yike/anaconda3/envs/own/lib/python3.5/ssl.py", line 937, in recv_into
    return self.read(nbytes, buffer)
  File "/media/hdd/yike/anaconda3/envs/own/lib/python3.5/ssl.py", line 799, in read
    return self._sslobj.read(len, buffer)
  File "/media/hdd/yike/anaconda3/envs/own/lib/python3.5/ssl.py", line 583, in read
    v = self._sslobj.read(len, buffer)
ConnectionResetError: [Errno 104] Connection reset by peer

I ran the OWM_ CNN code, but the running results is not similar with res.txt . After each task training, its test accuracy in other tasks is zero. The parameters are all default parameters. Is it that I did not run correctly?

In the file of OWM_CNN, your convolution layer uses 2*2, stride=1. but line 149, 152 and 155 in owm.py uses the stride=2. Is there anything wrong here? How should I think about that? Thank you!

You give the dataset download url with HWDB1 in Baidu cloud disk. But it only has the images, I want to get the mat files so that it can directly run on your code with CASIA_HWDB. Can you give me the downloading URL.

Is there a plan to release the code for reproducing the imagenet experiments?

Thanks for your great work. I have some doubts about owm.
1.In the setting of jointly training the feature extractor and the classifier, the architecture of the feature extractor is very simple. If we replace the simple feature extractor with resnet18 or vgg, will OWM still work?
2. Why the bias of the conv weight is all set to zero? If we use resnet18 as the feature extractor, should we set bias of each conv to zero?