yl-1993/hfsoftmax

Accelerated Training for Massive Classification via Dynamic Class Selection (AAAI 2018)

hfsoftmax's Issues

Training loss abnormal

Hello, I am using a non-sampling training model, and the loss always changes within a small range[-0.5,0.5]. When using the hf_sample method, the network loss rapidly increases to nan. Does the sample method require a pre-trained model?

pytorch version?

which version of pytorch is used in this project?

When the number of classes is nearly 1 million, how to implement the hash forest?

Hi, when the number of classes is nearly 100k(MS-Celeb-1M), i think annoy could build hash forest in an acceptable time.
But as the number of classes becomes larger, considered as a "massive classification", nearly 1 million, the HF building time seems unacceptable(I tried annoy building a 1 million features and set n_trees = 100, feat_dim = 2048, it costs ~22 minutes).
So is there a solution, or do you have any advice?
Thank you!

about Feature Extraction

Hello, I am a beginner, trying to extract features, can you tell me what the three parameters prefix, filelist, load_path mean?

'torch.dtype' object has no attribute 'type'

@yl-1993 Hi, i try to train the model with webface, then this problem appears。

How should I solve this problem

how to get the bin_file used in extract_feat.py？

I have thousands of pictures,i want to use extract_feat.py to extract features,I want to know how to get the bin file?
THX.

Problems about the training

Hi, I have two problems about the hfsoftmax

1. The sampler works from the beginning of training process, while the CPk (for example, top 1000) is very low at the beginning, how to ensure it works?
2. The adaptive allocation scheme not release the code ? only see sample_num=1000,

‘get_value_by_rows’ Fall into an infinite loop

Hi， It's a great work. When I am using the sampling layer（hnsw_sampler.py）, the program runs to 'update_hf ' function, falls into an infinite loop in 'get_value_by_rows', and cannot enter the conditional judgment. Part of the code is as follows：
#####################################################
while True:
sockets = dict(self._poll.poll(1000))
if self._socket in sockets:
msg = self._socket.recv()
if not meta:
meta = json.loads(msg)
else:
data = np.frombuffer(msg, dtype=meta['dtype'])
return data.reshape(meta['shape'])
#####################################################

text classification

your idea is very good! I am curious whether this method can be used for text classification？

When do you publish the source code?

I greatly admire your paper, can I ask for source code to have a look.

Thank you very much.

[email protected]

TypeError: conv2d(): argument 'input' (position 1) must be Tensor, not list

File "/home/adam/anaconda3/envs/torch04/lib/python3.6/site-packages/torch/nn/modules/conv.py", line 303, in forward
self.padding, self.dilation, self.groups)
TypeError: conv2d(): argument 'input' (position 1) must be Tensor, not list

Here, the convolution operation will report an error. You should enter a tensor. Actually enter a list composed of two tensors. I want to ask the author how to solve it.

How to generate training/validation list.

Hi， It's a great work， Could you tell me how to generate training/validation list since the rec2ims.py didn't generate these list files.

'Unsupported msg type'

def _recv(self):
    # TODO: 
    #   1. use logging to better log server info
    #   2. better exception handling
    print('waiting for message ... ')
    packet = self._socket.recv_multipart()
    print(packet)
    print(len(packet))
    if len(packet) == 2:
        ident, msg = packet
        msg = self._parse_json(msg)
    elif len(packet) == 4:
        ident, msg, meta, data = packet
        msg, meta = map(self._parse_json, [msg, meta])
        if msg['op'] == 'set_matrix':
            msg['data'] = self._buf_to_ndarray(data, meta)
        else:
            msg['rows'] = self._buf_to_ndarray(data, meta)
    elif len(packet) == 6:
        ident, msg, rows_meta, rows_data, val_meta, val_data = packet
        msg, rows_meta, val_meta = map(self._parse_json, [msg, rows_meta, val_meta])
        msg['rows'] = self._buf_to_ndarray(rows_data, rows_meta)
        msg['data'] = self._buf_to_ndarray(val_data, val_meta)
    else:
        raise RuntimeError('Unsupported msg type')
    self.handle(ident, msg)

always raise an error: RuntimeError('Unsupported msg type')
I find where the error shows up, but don't why?

两个face recognition模型，推荐使用哪个，hfsoftmax还要用到python 2 ？

https://github.com/yl-1993/hfsoftmax
https://github.com/XiaohangZhan/face_recognition_framework