这个地方将用户id和item id映射出索引不应该从0开始映射

不然下面用0padding 就会有问题

Hi, I used your bert4rec model to predict a sequence, but I don't know how to let the tensor readable. Can you give me some tips? thx!
Best regards

    # load one seq
    seq = dataset['test'][i]
    mask_token = len(dataset['smap']) + 1
    seq = seq + [mask_token]
    seq = seq[-args.bert_max_len:]
    padding_len = args.bert_max_len - len(seq)
    seq = [0] * padding_len + seq
    seq_ = torch.LongTensor([seq]).to(device)
    # recommend 50 item
    rec_list = []
    for _ in range(50):
        print(seq)
        #input 1...T time seq to predict
        logit = model(seq_)[0]
        #get the last tensor which is prediction? (T+1) time
        rec = torch.argmax(logit[-1], 0).it em()
        rec_list.append(datamap[rec])
        seq.insert(-1, rec)
        seq.pop(0)
        seq_ = torch.LongTensor([seq]).to(device)
        break

Why do I use mL-1m data set to run ndcg@10 result is only 0.26, in this paper 0.48. But the ML-20 dataset ,The results are similar to those in the paper.

Hi!!

Thanks for great code!
I have one question about padding at dataloaders/bert.py.

At this code, if item length is not enough, then pad using [0].
But I think, index 0 already have a Corresponding itemID, so that is padding using real itemID, right?

Thanks

Hi, Thank you so much for your code.
I got a question about AverageMeter updating way.

It seems that when updating recall or ndcg, in AverageMeterSet.update() and AverageMeter.update(),
that instance calculates average with batch average.

for example in Trainer > validate(), test(),
batch 1's ndcg = (0.1 + 0.2 + 0.3) / 3 =0.2 if batch size is 3,
batch 2's ndcg = (0.4 + 0.5) / 2 = 0.45 if last batch's size is 2
in codes, it seems calculating average like (0.2+0.45) / 2 = 0.325.

but isn't it right like (0.1+0.2+0.3+0.4+0.5) / 5 = 0.3 ?

Why your results are so much better than the original paper? is there anything different?

Dear @jaywonchung,

For the movielens 1m dataset the reported result is NDCG@10 = 0.4049. I am currently unable to reproduce this result with the default configuration template train_vae_search_beta used and 'vae_num_hidden': 0. The results I get are closer to 0.37, so I would like to know how did you achieve those results.

What configuration have you used to achieve the above-mentioned results and how did you evaluate?

Thank you.

Hi, I'm wondering if you also have datasets(Steam,Beauty) used in the original paper implementation of the BERT4REC available.

不好意思 请问怎么用训练出的模型对特定用户进行预测

I want to infer the model ，but the result is wrong

看起来测试集会比验证集上的效果低30%，这是正常的吗

I initially tested the functionality with one day’s worth of my custom data, and successfully trained and evaluated BERT4Rec without any issues.
However, when I tried to input a week’s worth of data for additional training, I encountered an out of memory CUDA error.

By reducing the 'bert_max_len' and 'batch size' in the 'templates.py' file, I managed to get it to work, but I do not understand why the increase in data size would cause an out of memory error.

Is there a part of the code where the increased data size might be impacting memory usage?
If so, could you point out where this might be occurring?

Thank you for your assistance.

Hi!

Thanks for your grate repository!
I'm trying to use this code and process step by step.
And I found strange point on my environment.

At recalls_and_ndcgs_for_ks def, I checked scores and ranks value.
I found difference between score order and rank order.

Below, ranks says index 13 is minimum, but scores index 13 is a '-2.9711'. It's not minimum at score.

Code is around here

# ranks[0] value
tensor([ 70,  86,   7,  97,  63,  95,  24,  58,  65,  56,  16,   5,  93,   0,
         18,  31,  68,  48,  59,  98,  29,  76,  54,  79,  23,  55,  39,  72,
         36,  89,  38,  44,  78,   4,  61,  53,   8,  74, 100,  22,   3,  91,
         80,  84,  51,  57,  82,  92,  83,  67,  34,  64,  50,  25,  85,  52,
         26,  43,  15,  12,  40,  27,   1,  88,  94,  71,   2,  66,  96,  28,
         19,  42,  20,  60,  49,  87,  77,  47,   9,  69,  41,  99,  17,  46,
         75,  73,  21,  14,  45,  32,  13,  11,  10,  33,  90,  35,   6,  37,
         30,  62,  81])

# scores[0] value
tensor([ 0.5852, -3.4007, -3.5931, -1.8515, -1.1481,  0.5925, -7.0834,  2.5950,
        -1.6262, -4.7560, -6.4868, -6.4362, -2.9711, -6.3241, -5.8845, -2.9454,
         0.7134, -5.0486,  0.5127, -4.3835, -4.5180, -5.8608, -1.8072, -0.6608,
         1.9210, -2.5598, -2.8975, -3.0994, -4.1074, -0.2445, -7.5053,  0.1854,
        -6.3146, -6.4986, -2.3304, -6.8591, -0.8691, -7.4218, -1.0146, -0.7875,
        -3.0707, -4.9039, -4.3940, -2.9124, -1.0281, -6.2175, -5.1383, -4.7329,
        -0.0837, -4.6264, -2.4418, -2.1516, -2.5802, -1.6060, -0.3369, -0.7312,
         0.7645, -2.1919,  0.8558, -0.1193, -4.5226, -1.3819, -7.8501,  2.4229,
        -2.4081,  0.8114, -3.6097, -2.3246,  0.0658, -4.8747,  2.9387, -3.4794,
        -0.7969, -5.7531, -1.7263, -5.4659, -0.3193, -4.7225, -1.1454, -0.3836,
        -1.9473, -8.0547, -2.2138, -2.3139, -2.0216, -2.5738,  2.7595, -4.6759,
        -3.4287, -0.9745, -6.6980, -1.8827, -2.2390,  0.5889, -3.4509,  2.3932,
        -4.0403,  2.5251, -0.1454, -4.9200, -1.7323])

Do you have any idea? Or if I have some mistake, then please let me know.
Thank you.

Dependencies in requirements.txt all have specified versions. Some of them are quite out of date. I realized that with the most recent version of them, we can run the program all good. How about we remove those version constraints?

In some cases, such as the sequence length is short or the value of mask_prob is small, there will be a situation where the whole training sequence is not masked, and the loss at this time will be the value of nan, how to solve this situation? I don't want the loss to be a nan value, can I only adjust the value of prob?

Hi,
I am experimenting with the original implementation of BERT4rec, but I used the same sampling strategy as you do.

When I trained the original implementation for 1 hour I've got following results:
R@1: 0.341391
R@5: 0.661589
R@10: 0.765728
NDCG@5: 0.512567
NDCG@10: 0.546344

This is well aligned with what you reported in the table.

However, when I gave the model 16 hours to train, I got much better results:
R@1: 0.405960
R@5: 0.714570
R@10: 0.803974
NDCG@5: 0.571801
NDCG@10 0.600875

I think it worth for you to re-evaluate your model with more epochs on the ML-1M dataset

For the ML-20M dataset my results are well aligned with yours:
R@1 0.613104
R@5 0.887872
R@10 0.945339
NDCG@5 0.763871
NDCG@10 0.782702

Could you please add an explicit LICENSE file to the repo so that it's clear under what terms the content is provided.

Per GitHub docs on licensing:

Public repositories on GitHub are often used to share open source software. For your repository to truly be open source, you'll need to license it so that others are free to use, change, and distribute the software.

Thanks in advance!

can you help in solving the dependencies

I am just wondering like yours code, we can have our own implementation of BERT and that means to say we can implement of without going into the parent library https://github.com/google-research/bert

In the paper BERT4REC, the authors uses Hit Ratio as a metrics for evaluation, and in your code, you use the Recall. The question which I want to ask is "Two metrics are equivalent?"

Why does it include BERT4Rec and Netflix VAE? I only need the copy torch version of BERT4Rec, how can I getit

def densify_index(self, df):
    print('Densifying index')
    umap = {u: i for i, u in enumerate(set(df['uid']))}
    smap = {s: i for i, s in enumerate(set(df['sid']))}
    df['uid'] = df['uid'].map(umap)
    df['sid'] = df['sid'].map(smap)
    return df, umap, smap

Is the 'umap' and 'smap' both beginning with 0? If so, then it is the same index with the label when the token was masked.

I want to ask a question that why you give '0' label to the none masked token when generate a example here:

    for s in seq:
           prob = self.rng.random()
           if prob < self.mask_prob:
               prob /= self.mask_prob

               if prob < 0.8:
                   tokens.append(self.mask_token)
               elif prob < 0.9: 
                   tokens.append(self.rng.randint(1, self.num_items))
               else: 
                   tokens.append(s)

               labels.append(s) 
           else:
               tokens.append(s)
               labels.append(0) # **why give 0 label here and not the index of s ?**

Thanks !