I was not able to understand one thing , when it says "fine-tuning of BERT", what does it actually mean:
Are we retraining the entire model again with new data.
Or are we just training top few transformer layers with new data.
Or we are training the entire model but considering the pretrained weights as initial weight.
Or there is already few layers of ANN on top of transformer layers which is only getting trained keeping transformer weight freeze.
Tried Google but I am getting confused, if someone can help me on this.
Thanks in advance!
I remember reading about a Twitter poll with similar context, and it seems that most people tend to accept your suggestion 3. (or variants thereof) as the standard definition.
However, this obviously does not speak for every single work, but I think it's fairly safe to say that 1. is usually not included when talking about fine-tuning. Unless you have vast amounts of (labeled) task-specific data, this step would be referred to as pre-training a model.
2. and 4. could be considered fine-tuning as well, but from personal/anecdotal experience, allowing all parameters to change during fine-tuning has provided significantly better results. Depending on your use case, this is also fairly simple to experiment with, since freezing layers is trivial in libraries such as Huggingface transformers.
In either case, I would really consider them as variants of 3., since you're implicitly assuming that we start from pre-trained weights in these scenarios (correct me if I'm wrong).
Therefore, trying my best at a concise definition would be:
Fine-tuning refers to the step of training any number of parameters/layers with task-specific and labeled data, from a previous model checkpoint that has generally been trained on large amounts of text data with unsupervised MLM (masked language modeling).
Related
I would like to extend a zero-shot text classification (NLI) model's vocabulary, to include domain-specific vocabulary or just to keep it up-to-date. For example, I would like the model to know the names of the latest COVID-19 variants are related to the topic 'Healthcare'.
I've added the tokens to the tokenizer and resized the token embeddings. However, I don't know how to finetune the weights in the embedding layer, as suggested here.
To do the finetuning, can I use simply use texts containing a mixture of new vocabulary and existing vocabulary, and have the tokenizer recognise the relations between tokens through co-occurrences in an unsupervised fashion?
Any help is appreciated, thank you!
If you resized the corresponding embedding weights with resize_token_embeddings, they will be initialised randomly.
Technically, you can fine-tune the model on your target task (NLI, in your case), without touching the embedding weights. In practice, it will be harder for your model to learn anything meaningful about the newly added tokens, since their embeddings are randomly initialised.
To learn the embedding weights you can do further pre-training, before fine-tuning on the target task. This is done by training the model on the pre-training objective(s) (such as Masked Language Modelling). Pre-training is more expensive than fine-tuning of course, but remember that you aren't pre-training from scratch, since you start pre-training from the checkpoint of the already pre-trained model. Therefore, the number of epochs/steps will be significantly less than what was used in the original pre-training setup.
When doing pre-training it will be beneficial to include in-domain documents, so that it can learn the newly added tokens. Depending on whether you want the model to be more domain specific or remain varied so as to not "forget" any previous domains, you might also want to include documents from a variety of domains.
The Don't Stop Pretraining paper might also be an interesting reference, which delves into specifics regarding the type of data used as well as training steps.
I'm training a token classification (AKA named entity recognition) model with the HuggingFace Transformers library, with a customized data loader.
Like most NER datasets (I'd imagine?) there's a pretty significant class imbalance: A large majority of tokens are other - i.e. not an entity - and of course there's a little variation between the different entity classes themselves.
As we might expect, my "accuracy" metrics are getting distorted quite a lot by this: It's no great achievement to get 80% token classification accuracy if 90% of your tokens are other... A trivial model could have done better!
I can calculate some additional and more insightful evaluation metrics - but it got me wondering... Can/should we somehow incorporate these weights into the training loss? How would this be done using a typical *ForTokenClassification model e.g. BERTForTokenClassification?
This is actually a really interesting question, since it seems there is no intention (yet) to modify losses in the models yourself. Specifically for BertForTokenClassification, I found this code segment:
loss_fct = CrossEntropyLoss()
# ...
loss = loss_fct(logits.view(-1, self.num_labels), labels.view(-1))
To actually change the loss computation and add other parameters, e.g., the weights you mention, you can go about either one of two ways:
You can modify a copy of transformers locally, and install the library from there, which makes this only a small change in the code, but potentially quite a hassle to change parts during different experiments, or
You return your logits (which is the case by default), and calculate your own loss outside of the actual forward pass of the huggingface model. In this case, you need to be aware of any potential propagation from the loss calculated within the forward call, but this should be within your power to change.
I want to train a Doc2Vec model with a generic corpus and, then, continue training with a domain-specific corpus (I have read that is a common strategy and I want to test results).
I have all the documents, so I can build and tag the vocab at the beginning.
As I understand, I should train initially all the epochs with the generic docs, and then repeat the epochs with the ad hoc docs. But, this way, I cannot place all the docs in a corpus iterator and call train() once (as it is recommended everywhere).
So, after building the global vocab, I have created two iterators, the first one for the generic docs and the second one for the ad hoc docs, and called train() twice.
Is it the best way or it is a more appropriate way?
If the best, how I should manage alpha and min_alpha? Is it a good decision not to mention them in the train() calls and let the train() manage them?
Best
Alberto
This is probably not a wise strategy, because:
the Python Gensim Doc2Vec class hasn't ever properly supported expanding its known vocabulary after a 1st single build_vocab() call. (Up through at least 3.8.3, such attempts typically cause a Segmentation Fault process crash.) Thus if there are words that are only in your domain-corpus, an initial typical initialization/training on the generic-corpus would leave them out of the model entirely. (You could work around this, with some atypical extra steps, but the other concerns below would remain.)
if there is truly an important contrast between the words/word-senses used in your generic and the different words/word-senses used in your domain corpus, influence of the words from the generic corpus may not be beneficial, diluting domain-relevant meanings
further, any followup training that just uses a subset of all documents (the domain corpus) will only be updating the vectors for that subset of words/word-senses, and the model's internal weights used for further unseen-document inference, in directions that make sense for the domain-corpus alone. Such later-trained vectors may be nudged arbitrarily far out of comparable alignment with other words not appearing in the domain-corpus, and earlier-trained vectors will find themselves no longer tuned in relation to the model's later-updated internal-weights. (Exactly how far will depend on the learning-rate alpha & epochs choices in the followup training, and how well that followup training optimizes model loss.)
If your domain dataset is sufficient, or can be grown with more domain data, it may not be necessary to mix in other training steps/data. But if you think you must try that, the best-grounded approach would be to shuffle all training data together, and train in one session where all words are known from the beginning, and all training examples are presented in balanced, interleaved fashion. (Or possibly, where some training texts considered extra-important are oversampled, but still mixed in with the variety of all available documents, in all epochs.)
If you see an authoritative source suggesting such a "train with one dataset, then another disjoint dataset" approach with the Doc2Vec algorithms, you should press them for more details on what they did to make that work: exact code steps, and the evaluations which showed an improvement. (It's not impossible that there's some way to manage all the issues! But I've seen many vague impressions that this separate-pretraining is straightforward or beneficial, and zero actual working writeups with code and evaluation metrics showing that it's working.)
Update with respect to the additional clarifications you provided at https://stackoverflow.com/a/64865886/130288:
Even with that context, my recommendation remains: don't do this segmenting of training into two batches. It's almost certain to degrade the model compared to a combined training.
I would be interested to see links to the "references in the literature" you allude to. They may be confused or talking about algorithms other than the Doc2Vec ("Paragraph Vectors") algorithm.
If there is any reason to give your domain docs more weight, a better-grounded way would be to oversample them in the combined corpus.
Bu by all means, test all these variants & publish the relative results. If you're exploring shaky hypotheses, I would ignore any advice from StackOverflow-like sources & just run all the variants that your reading of the literature suggest, to see which, if any actually help.
You're right to recognized that the choice of alpha parameters is a murky area that could majorly influence what impact such add-on training has. There's no right answer, so you'll have to search-for and reason-out what might make sense. The inherent issues I've mentioned with such subset-followup-training could make it so that even if you find benefits in some combos, they may be more a product of a lucky combination of data & arbitrary parameters than a generalizable practice.
And: your specific question "if it is better to set such values or not provide them at all" reduces to: "do you want to use the default values, or values set when the model was created, or not?"
Which values might be workable, if at all, for this unproven technique is something that'd need to be experimentally discovered. That is, if you wanted to have comparable (or publishable) results here, I think you'd have to justify from your own novel work some specific strategy for choosing good alpha/epochs and other parameters, rather than adopt any practice merely recommended in a StackOverflow answer.
I want to fine tune BERT on a specific domain. I have texts of that domain in text files. How can I use these to fine tune BERT?
I am looking here currently.
My main objective is to get sentence embeddings using BERT.
The important distinction to make here is whether you want to fine-tune your model, or whether you want to expose it to additional pretraining.
The former is simply a way to train BERT to adapt to a specific supervised task, for which you generally need in the order of 1000 or more samples including labels.
Pretraining, on the other hand, is basically trying to help BERT better "understand" data from a certain domain, by basically continuing its unsupervised training objective ([MASK]ing specific words and trying to predict what word should be there), for which you do not need labeled data.
If your ultimate objective is sentence embeddings, however, I would strongly suggest you to have a look at Sentence Transformers, which is based on a slightly outdated version of Huggingface's transformers library, but primarily tries to generate high-quality embeddings. Note that there are ways to train with surrogate losses, where you try to emulate some form ofloss that is relevant for embeddings.
Edit: The author of Sentence-Transformers recently joined Huggingface, so I expect support to greatly improve over the upcoming months!
#dennlinger gave an exhaustive answer. Additional pretraining is also referred as "post-training", "domain adaptation" and "language modeling fine-tuning". here you will find an example how to do it.
But, since you want to have good sentence embeddings, you better use Sentence Transformers. Moreover, they provide fine-tuned models, which already capable of understanding semantic similarity between sentences. "Continue Training on Other Data" section is what you want to further fine-tune the model on your domain. You do have to prepare training dataset, according to one of available loss functions. E.g. ContrastLoss requires a pair of texts and a label, whether this pair is similar.
I believe transfer learning is useful to train the model on a specific domain. First you load the pretrained base model and freeze its weights, then you add another layer on top of the base model and train that layer based on your own training data. However, the data would need to be labelled.
Tensorflow has some useful guide on transfer learning.
You are talking about pre-training. Fine-tuning on unlabeled data is called pre-training and for getting started, you can take a look over here.
BERT pre-training of the base-model is done by a language modeling approach, where we mask certain percent of tokens in a sentence, and we make the model learn those missing mask. Then, I think in order to do downstream tasks, we add a newly initialized layer and we fine-tune the model.
However, suppose we have a gigantic dataset for sentence classification. Theoretically, can we initialize the BERT base architecture from scratch, train both the additional downstream task specific layer + the base model weights form scratch with this sentence classification dataset only, and still achieve a good result?
Thanks.
BERT can be viewed as a language encoder, which is trained on a humongous amount of data to learn the language well. As we know, the original BERT model was trained on the entire English Wikipedia and Book corpus, which sums to 3,300M words. BERT-base has 109M model parameters. So, if you think you have large enough data to train BERT, then the answer to your question is yes.
However, when you said "still achieve a good result", I assume you are comparing against the original BERT model. In that case, the answer lies in the size of the training data.
I am wondering why do you prefer to train BERT from scratch instead of fine-tuning it? Is it because you are afraid of the domain adaptation issue? If not, pre-trained BERT is perhaps a better starting point.
Please note, if you want to train BERT from scratch, you may consider a smaller architecture. You may find the following papers useful.
Well-Read Students Learn Better: On the Importance of Pre-training Compact Models
ALBERT: A Lite BERT for Self-supervised Learning of Language Representations
I can give help.
First of all, MLM and NSP (which are the original pre-training objectives from NAACL 2019) are meant to train language encoders with prior language knowledge. Like a primary school student who read many books in the general domain. Before BERT, many neural networks would be trained from scratch, from a clean slate where the model doesn't know anything. This is like a newborn baby.
So my question is, "is it a good idea to start teaching a newborn baby when you can begin with a primary school student?" My answer is no. This is supported by numerous State-of-The-Arts achieved by the pre-trained models, compared to the old methods of training a neural network from scratch.
As someone who works in the field, I can assure you that it is a much better idea to fine-tune a pre-trained model. It doesn't matter if you have a 200k dataset or a 1mil datapoints. In fact, more fine-tuning data will only make the downstream results better if you use the right hyperparameters.
Though I recommend the learning rate between 2e-6 ~ 5e-5 for sentence classification tasks, you can explore. If your dataset is very, very domain-specific, it's up to you to fine-tune with a higher learning rate, which will deviate the model further away from its "pre-trained" knowledge.
And also, regarding your question on
can we initialize the BERT base architecture from scratch, train both the additional downstream task specific layer + the base model weights form scratch with this sentence classification dataset only, and still achieve a good result?
I'm negative about this idea. Even though you have a dataset with 200k instances, BERT is pre-trained on 3300mil words. BERT is too inefficient to be trained with 200k instances (both size-wise and architecture-wise). If you want to train a neural network from scratch, I'd recommend you look into LSTMs or RNNs.
I'm not saying I recommend LSTMs. Just fine-tune BERT. 200k is not even too big anyways.
All the best luck with your NLP studies :)