Using WordPiece tokenization with RoBERTa - deep-learning

As far as I understood, the RoBERTa model implemented by the huggingface library, uses BPE tokenizer. Here is the link for the documentation:
RoBERTa has the same architecture as BERT, but uses a byte-level BPE as a tokenizer (same as GPT-2) and uses a different pretraining scheme.
However, I have a custom tokenizer based on WordPiece tokenization and I used the BertTokenizer.
Because my customized tokenizer is much more relevant for my task, I prefer not to use BPE.
When I pre-trained the RoBERTa from scratch (RobertaForMaskedLM) with my custom tokenizer the loss for the MLM task was much better than the loss with BPE. However, when it comes to fine-tuning, the model (RobertaForSequenceClassification) perform poorly. I am almost sure the problem is not about the tokenizer. I wonder if the huggingface library for the RobertaForSequenceClassification is not compatible with my tokenizer.
Details about the fine-tuning:
task: multilabel classification with imbalanced labels.
epochs: 20
loss: BCEWithLogitsLoss()
optimizer: Adam, weight_decay_rate:0.01, lr: 2e-5, correct_bias: True
The F1 and AUC was very low because the output probabilities for the labels was not in accordance with the actual labels (even with a very low threshold) which means the model couldn't learn anything.
*
Note: The pre-trained and fine-tuned RoBERTa with BPE tokenizer
performs better than the pre-trained and fine-tuned with custom
tokenizer although the loss for MLM with custom tokenizer was better
than BPE.

Related

Backbone network in Object detection

I am trying to understand the training process of a object deetaction deeplearng algorithm and I am having some problems understanding how the backbone network (the network that performs feature extraction) is trained.
I understand that it is common to use CNNs like AlexNet, VGGNet, and ResNet but I don't understand if these networks are pre-trained or not. If they are not trained what does the training consist of?
We directly use a pre-trained VGGNet or ResNet backbone. Although the backbone is pre-trained for classification task, the hidden layers learn features which can be used for object detection also. Initial layers will learn low level features such as lines, dots, curves etc. Next layer will learn learn high-level features that are built on top of low-level features to detect objects and larger shapes in the image.
Then the last layers are modified to output the object detection coordinates rather than class.
There are object detection specific backbones too. Check these papers:
DetNet: A Backbone network for Object Detection
CBNet: A Novel Composite Backbone Network Architecture for Object Detection
DetNAS: Backbone Search for Object Detection
High-Resolution Network: A universal neural architecture for visual recognition
Lastly, the pretrained weights will be useful only if you are using them for similar images. E.g.: weights trained on Image-net will be useless on ultrasound medical image data. In this case we would rather train from scratch.

How to choose which pre-trained weights to use for my model?

I am a beginner, and I am very confused about how we can choose a pre-trained model that will improve my model.
I am trying to create a cat breed classifier using pre-trained weights of a model, lets say VGG16 trained on digits dataset, will that improve the performance of the model? or if I train my model just on the database without using any other weights will be better, or will both be the same as those pre-trained weights will be just a starting point.
Also if I use weights of the VGG16 trained for cat vs dog data as a starting point of my cat breed classification model will that help me in improving the model?
Since you've mentioned that you are a beginner I'll try to be a bit more verbose than normal so please bear with me.
How neural models recognise images
The layers in a pre-trained model store multiple aspects of the images they were trained on like patterns(lines, curves), colours within the image which it uses to decide if an image is of a specific class or not
With each layer the complexity of what it can store increases initially it captures lines or dots or simple curves but with each layer, the representation power increases and it starts capturing features like cat ears, dog face, curves in a number etc.
The image below from Keras blog shows how initial layers learn to represent simple things like dots and lines and as we go deeper they start to learn to represent more complex patterns.
Read more about Conv net Filters at keras's blog here
How does using a pretrained model give better results ?
When we train a model we waste a lot of compute and time initially creating these representations and in order to get to those representations we need quite a lot of data too else we might not be able to capture all relevant features and our model might not be as accurate.
So when we say we want to use a pre-trained model we want to use these representations so if we use a model trained on imagenet which has lots of cat pics we can be sure that the model already has representations to identify important features required to identify a cat and will converge to a better point than if we used random weights.
How to use pre-trained weights
So when we say to use pre-trained weights we mean use the layers which hold the representations to identify cats but discard the last layer (dense and output) and instead add fresh dense and output layers with random weights. So our predictions can make use of the representations already learned.
In real life we freeze our pretrained weights during the initial training as we do not want our random weights at the bottom to ruin the learned representations. we only unfreeze the representations in the end after we have a good classification accuracy to fine-tune them, and that too with a very small learning rate.
Which kind of pre-trained model to use
Always choose those pretrained weights that you know has the most amount of representations which can help you in identifying the class you are interested in.
So will using a mnist digits trained weights give relatively bad results when compared with one trained on image net?
Yes, but given that the initial layers have already learned simple patterns like lines and curves for digits using these weights will still put you at an advantage when compared to starting from scratch in most of the cases.
Sane weight initialization
The pre-trained weights to choose depends upon the type of classes you wish to classify. Since, you wish to classify Cat Breeds, use pre-trained weights from a classifier that is trained on similar task. As mentioned by the above answers the initial layers learn things like edges, horizontal or vertical lines, blobs, etc. As you go deeper, the model starts learning problem specific features. So for generic tasks you can use say imagenet & then fine-tune it for the problem at hand.
However, having a pre-trained model which closely resembles your training data helps immensely. A while ago, I had participated in Scene Classification Challenge where we initialized our model with the ResNet50 weights trained on Places365 dataset. Since, the classes in the above challenge were all present in the Places365 dataset, we used the weights available here and fine-tuned our model. This gave us a great boost in our accuracy & we ended up at top positions on the leaderboard.
You can find some more details about it in this blog
Also, understand that the one of the advantages of transfer learning is saving computations. Using a model with randomly initialized weights is like training a neural net from scratch. If you use VGG16 weights trained on digits dataset, then it might have already learned something, so it will definitely save some training time. If you train a model from scratch then it will eventually learn all the patterns which using a pre-trained digits classifier weights would have learnt.
On the other hand using weights from a Dog-vs-Cat classifier should give you better performance as it already has learned features to detect say paws, ears, nose or whiskers.
Could you provide more information, what do you want to classify exactly? I see you wish to classify images, which type of images (containing what?) and in which classes?
As a general remark : If you use a trained model, it must fit your need, of course. Keep in mind that a model which was trained on a given dataset, learned only the information contained in that dataset and can classify / indentify information analogous to the one in the training dataset.
If you want to classify an image containing an animal with a Y/N (binary) classifier, (cat or not cat) you should use a model trained on different animals, cats among them.
If you want to classify an image of a cat into classes corresponding to cat races, let's say, you should use a model trained only on cats images.
I should say you should use a pipeline, containing steps 1. followed by 2.
it really depends on the size of the dataset you have at hand and how related the task and data that the model was pretrained on to your task and data. Read more about Transfer Learning http://cs231n.github.io/transfer-learning/ or Domain Adaptation if your task is the same.
I am trying to create a cat breed classifier using pre-trained weights of a model, lets say VGG16 trained on digits dataset, will that improve the performance of the model?
There are general characteristics that are still learned from digits like edge detection that could be useful for your target task, so the answer here is maybe. You can here try just training the top layers which is common in computer vision applications.
Also if I use weights of the VGG16 trained for cat vs dog data as a starting point of my cat breed classification model will that help me in improving the model?
Your chances should be better if the task and data are more related and similar

Training model in eval() mode gives better result in PyTorch?

I have a model with Dropout layers (with p=0.6). I ended up training the model in .eval() mode and again trained the model in .train() mode, I find that the training .eval() mode gave me better accuracy and quicker loss reduction on training data,
train(): Train loss : 0.832, Validation Loss : 0.821
eval(): Train loss : 0.323, Validation Loss : 0.251
Why is this so?
This seems like the model architecture is simple and when in train mode, is not able to capture the features in the data and hence undergoes underfitting.
eval() disables dropouts and Batch normalization, among other modules.
This means that the model trains better without dropout helping the model the learn better with more neurons, also increasing the layer size, increasing the number of layers, decreasing the dropout probability, helps.

Is the xception model in keras the best model was describe in the paper?

I read the Xception paper and in this paper it was mentioned in part 4.7 that best results are achivable without any activation. Now I want to use this network on videos using keras toolbox but the model in keras uses 'ReLU' activation function. Does the model in keras returns best model or it is better to omit the relu layers?
You are confusing normal activations used for convolutional and dense layers, with the ones mentioned in the paper. Section 4.7 only deals with varying the activation between depth-wise and point-wise convolutions, the rest of the activations in the architecture are kept unchanged.

Caffe Autoencoder

I wanna compare the performance of CNN and autoencoder in caffe. I'm completely familiar with cnn in caffe but I wanna is the autoencoder also has deploy.prototxt file ? is there any differences in using this two models rather than the architecture?
Yes it also has a deploy.prototxt.
both train_val.prototxt and 'deploy.prototxt' are cnn architecture description files. The sole difference between them is, train_val.prototxt takes training data and loss as input/output, but 'deploy.prototxt' takes testing image as input, and predicted value as out put.
Here is an example of a cnn and autoencoder for MINST: Caffe Examples. (I have not tried the examples.) Using the models is generally the same. Learning rates etc. depend on the model.
You need to implement an auto-encoder example using python or matlab. The example in Caffe is not true auto-encoder because it doesn't set layer-wise training stage and during training stage, it doesn't fix W{L->L+1} = W{L+1->L+2}^T. It is easily to find a 1D auto-encoder in github, but 2D auto-encoder may be hard to find.
The main difference between the Auto encoders and conventional network is
In Auto encoder your input is your label image for training.
Auto encoder tries to approximate the output similar as input.
Auto encoders does not have softmax layer while training.
It can be used as a pre-trained model for your network which converge faster comparing to other pre-trained models. It is because your network has already extracted the features for your data.
The Conventional training and testing you can perform on pre trained auto encoder network for faster convergence and accuracy.