Month: October 2021

Human Activity Recognition using high-dimensional visual streams has been gaining popularity in recent times. Using a video input to categorize human activity is primarily applied in surveillance of different kinds. At hospitals and nursing homes, this can be used to immediately alert caretakers when any of the residents are displaying any sign of sickness – clutching their chest, falling, vomiting, etc. At public places like airports, railway stations, bus stations, malls or even in your neighborhoods, the activity recognition becomes the means to alert authorities on recognizing suspicious behavior. This AI solution is even useful for identifying flaws in and improving the form of athletes and professionals ensuring improved performance and better training.

Why Use Multimodal Learning for Activity Recognition?

Our experience of the world is multimodal in nature, as quoted by Baltrušaitis et al. There are multiple modalities a human is blessed with. We can touch, see, smell, hear, and taste, and understand the world around us in a better way. Most parents would remember their kid’s understanding of “what a dog is” keeps improving after seeing actual dogs, videos of dogs, photographs of dogs and cartoon dogs and being told that they are all dogs. Just seeing a single video of an actual dog does not help the kid identify the character “Goofy” as a dog. It is the same for machines. Multimodal machine learning models can process and relate the information from multiple modalities, learning in a more holistic way.

This blog serves as the captain’s log on how we combined the effectiveness of two modalities – Static Images and Videos – to improve the classification of human activities from videos. Algorithms for video activity recognition are based on dealing with only spatial information (images), or both spatial and temporal information (videos). Algorithms were used for both static images and videos for the activity recognition modeling. Fusing both models together made the resultant multimodal model far better than each of the individual unimodal models.

Multimodal Learning for Human Activity Recognition – Our Recipe

Our goal was to recognize 10 activities – basketball, biking, diving, golf swing, horse riding, soccer juggling, tennis swing, trampoline jumping, volleyball spiking, and walking. We created the multimodal models for activity recognition by fusing the two unimodal models – image-based and video-based – using the ensemble method, thus enhancing the effect of the classifier.

The Dataset Used: We have used modified UCF11 dataset (removed Swing class as it has many mis-labelled data). For the 10 activities we need to classify, the dataset has 120-200 videos of different lengths ranging from 1-21 seconds. The link to the dataset is (CRCV | Center for Research in Computer Vision at the University of Central Florida (ucf.edu)).

One Modality at a Time

There are different methodologies for Multimodal learning as is described by Song et al., 2016, Tzirakis et al., 2017, and Yoon et al., 2018. One of the techniques is ensemble learning, in which 2DCNN model and 3DCNN models are trained separately and the final softmax probabilities are combined to get predictions. Other techniques include joint representation, coordinated representation etc. A detailed overview is available in Baltrušaitis et al., 2018.

The First Modality – Images: We trained a 2DCNN model with VGG-16 architecture and equipped our model with batch normalization and other regularization methods, since the enormous number of frames

caused overfitting. We observed that deeper architectures were less accurate. We achieved 81% clip-wise accuracy. However, the accuracies are very poor on some classes such as basketball, soccer juggling, tennis, and walking as can be seen below.

Second Modality – Videos: The second model was trained with 3DCNN architecture, which took 16 frames as one chunk called as a block. We incorporated various augmentations, batch normalization, and regularization methods. During the exercise, we observed that 3D architecture is sensitive to the learning rate. In the end, we achieved a clip-wise accuracy of 73%. Looking at the accuracies for individual classes, this model is not the best, but it is much better than the 2DCNN for Soccer Juggling and Golf classes. Class-wise accuracies can be seen below.

Our next objective was to ensure that the learnings from both modalities are combined to create a more accurate and robust model.

Our Secret Sauce for Multimodal learning – The Ensemble Method

For fusing the two modalities, we resorted to ensemble methods (Zhao, 2019). We experimented with two ensemble methods for the Multimodal learning:

1. Maximum Vote – Mode from the predictions of both models is taken as the predicted label.

2. Averaging and Maximum Pooling – Weighted sum of probabilities of both models is calculated to decide the predicted label.

Maximum Vote Ensemble Approach: Let us consider a video consisting of 64 frames. We fed these frames to both 2D and 3D models. The 2D model produced an output for each frame resulting in 64 tensors with a predicted label. Similarly, the 3D model produced a label for 16 frames, thus finally generating 4 tensors for 64 frames video. To balance the weightage of the 3D model, we augment the output tensors of the 3D model by repeating the tensor multiple times. Finally, we concatenated the output tensors from both models, and selected the modal class (the class label with the highest occurrence) as the prediction label.

We iteratively experimented with different weightages for both models and checked the corresponding overall accuracy as shown below. Maximum overall accuracy of 84% was achieved at 60% weightage to the 2D model and 40% weightage to the 3D model.

At 60:40 weightage for 2D:3D, the Maximum Vote ensemble method has performed better than either of the unimodal models except for Biking and Soccer Juggling.

Averaging and Maximum Pooling Ensemble Method: Consider that we are feeding a video with 64 frames to both models, the 2D model which generates 64 tensors with probabilities for every class and the 3D model will generate 4 tensors with probabilities for each of 10 classes. We calculate the average of all 64 tensors from the 2D model to get the average probability for every class. Similarly, we calculate the average of the 4 tensors from the 3D model. Finally, we take the weighted sum of resulting tensors from both models. We select the class with maximum probability as the predicted class.

The equation for the weight in Averaging and Maximum Pooling method is:

Multimodal = α*(2D Model) + (1-α)*(3D Model)

We experimented with different values of α and compared the resulting overall accuracies as shown below. With equal weightage to both models, we achieved 77% accuracy. With the increase in α, i.e., weightage to 2D model, multimodal accuracy increased. We achieved the best accuracy of 87% at 90% weightage to the 2D model and 10% weightage of the 3D model.

The class-wise accuracy for the Averaging and Maximum Pooling method is low for Basketball and Walking classes. However, the class-wise accuracies are better than or at least the same as the unimodal models. The accuracies for Soccer Juggling and Tennis Swing improved the most from 71% and 67% for the 2D model to 90% and 85% respectively.

Conclusion

Beyond doubt, our Multimodal models performed better than the Unimodal ones. Comparing the multimodal engines, Averaging and Maximum Pooling performed better than the Maximum Vote method, as is evident from the overall accuracies of 87% and 84% respectively. The reason is that the Averaging and Maximum Pooling method considers the confidence of the predicted label whereas, the Maximum Vote method considers only the label with maximum probability.

In Human Activity Recognition, we believe the multimodal learning approach can be improved further by incorporating other modalities as well. Such as Facebook’s Detectron model or pose estimation method.

Our next plan of action is to explore more forms of multimodal learning for activity recognition. Using features addition/ layers are fusing features can be effective in learning features better. Another way of proceeding would be to add different modalities like pose detection feed, motion detection feed and object detection feed to provide better results. No matter the approach, fusing modalities has a corresponding cost factor associated with it. While we have 40.4 million trainable parameters in the 2DCNN model and 78 million parameters in the 3DCNN models, the multimodal model has 118.5 million parameters to train on. But this is a small amount to pay considering the limitless applications that can be made viable because of the performance improvement provided by the multimodal models.

Natural Language Inferencing (NLI) task is one of the most important subsets of Natural Language Processing (NLP) which has seen a series of development in recent years. There are standard benchmark publicly available datasets like Stanford Natural Language Inference (SNLI) Corpus, Multi-Genre NLI (MultiNLI) Corpus, etc. which are dedicated to NLI tasks. Few state-of-the-art models trained on these datasets possess decent accuracy. In this blog I will start with briefing the reader about NLI terminologies, applications of NLI, NLI state-of-the-art model architectures and eventually demonstrate the NLI task using Kaggle Contradictory My Dear Watson Challenge Dataset by the end.

Prerequisites:

• Basics of NLP
• Moderate Python coding

What is NLI?

Natural Language Inference which is also known as Recognizing Textual Entailment (RTE) is a task of determining whether the given “hypothesis” and “premise” logically follow (entailment) or unfollow (contradiction) or are undetermined (neutral) to each other. For example, let us consider hypothesis as “The game is played by only males” and premise as “Female players are playing the game”. The task of NLI model is to predict whether the two sentences are either entailment, contradiction, or neutral. In this case, it is a contradiction.

How NLI is different from NLP?

The main difference between NLP and NLI is that NLP is a broader set that contains two subsets Natural Language Understanding (NLU) and Natural Language Generation (NLG). We are more concerned about NLU as NLI comes under this. NLU is basically making the computer capable of comprehending what the given text block represents. NLI, which comes under NLU is the task of understanding the given two statements and categorizing them either as entailment, contradiction, or neutral sentences. When dealing with data most of the NLP tasks include pre-processing steps like removing stop words, special characters, etc. But in case of NLI, one has to just provide the model with two sentences. The model then processes the data itself and outputs the relationship between the two sentences.

Applications of NLI

NLI is been used in many domains like banking, retail, finance, etc. It is widely used in cases where there is a requirement to check if generated or obtained result from the end-user follows the hypothesis. One of the use cases includes automatic auditing tasks. NLI can replace human auditing to some extent by comparing if sentences in generated document entail with the reference documents.

Models used to Demonstrate NLI Task

In this blog, I have demonstrated the NLI task using two models: RoBERTa and XLM-RoBERTa. Let us understand these models in this section.

In order to understand RoBERTa model, one should have a brief knowledge about BERT model.

BERT

Bidirectional Encoder Representation Transformers (BERT) was published by Google AI researchers in 2018. It has shown state-of-the-art results in many NLP tasks like question and answering, NLI task etc. It is basically an encoder stack of transformer architecture. It has two versions BERT base and BERT large. BERT base has 12 layers in its encoder stack and 110M total parameters whereas BERT large has 24 layers and 340M total parameters.

BERT pre-training consists of two tasks:

1. Masked Language Model (MLM)
2. Next Sentence Prediction (NSP)

Masked LM

In the input sequence sent to the model as input, randomly 15% of the words are masked and the model is tasked to predict these masks by understanding the context from unmasked words at the end of training. This helps model in understanding the context of the sentence.

Next Sentence Prediction

Model is fed with two-sentence pairs as input. In this task, a model must predict at the end of training whether the sentences follow or unfollow each other. This helps in understanding the relationship between two sentences which is the major objective for tasks like question and answering, NLI, etc.

Both the tasks are executed simultaneously while training.

Model Input

Input to the BERT model is a sequence of tokens that are converted to embeddings. Each token embedding is a combination of 3 embeddings.

1. Token Embeddings – These are word embeddings from WordPiece token vocabulary.
2. Segment Embeddings – As BERT model takes pair of sentences as input, in order to help model distinguish the embeddings from different sentences these embeddings are used. In the above picture, EA represents embeddings of sentence A while EB represents embeddings from sentence B.
3. Position Embeddings – In order to capture “sequence” or “order” information these embeddings are used to express the position of words in a sentence.

Model Output

The output of BERT model has the same no. of tokens as input with additional classification token which gives the classification results ie. whether sentence B follows sentence A or not.

Fine-Tuning BERT

A pre-trained BERT model can be fine-tuned to achieve a specific task on specific data. Fine-tuning uses same architecture as the pre-trained model only an additional output layer is added depending on the task. In case of NLI task classification token is fed into the output classification layer which determines the probabilities of entailment, contradiction, and neutral classes.

BERT GLUE Task Results

As you can see in figure 4, BERT outperforms all the previous models on GLUE tests.

RoBERTa

Robustly Optimised BERT Pre-training Approach (RoBERTa) was proposed by Facebook researchers. They found with a much more robustly pre-training BERT model it can still perform better on GLUE tasks. RoBERTa model is a BERT model with modified pre-training approach.

Below are the few changes incorporated in RoBERTa model when compared to BERT model.

1. Data – RoBERTa model is trained using much more data when compared to BERT. It is trained on 160GB uncompressed data.
2. Static vs Dynamic Masking – In BERT model, data was masked only once during pre-processing which results in single static masks. These masks are used for all the iterations while training. In contrast, data used for RoBERTa training was duplicated 10 times with 10 different mask patterns and was trained over 40 epochs. This means a single mask pattern is used only in 4 epochs. This is static masking. While in dynamic masking different mask pattern is generated for every epoch during training.

3. Removal of Next Sentence Prediction (NSP) objective – Researches have found that removing NSP loss significantly improved the model performance on GLUE tasks.

4. Trained on Large Batch Sizes – Training model on large batch sizes improved the model accuracy.

5. Tokenization – RoBERTa uses a byte-level Byte-Pair Encoding (BPE) encoding scheme with a containing 50K vocabulary in contrast to BERT’s character-level BPE with a 30K vocabulary.

RoBERTa Results on GLUE Tasks

RoBERTa clearly outperforms when compared to previous models.

XLM-RoBERTa

XLM-R is a large multilingual model trained on 100 different languages. It is basically an update to Facebook XLM-100 model which is also trained in 100 different languages. It uses the same training procedure as RoBERTa model which used only Masked Language Model (MLM) technique without using Next Sentence Prediction (NSP) technique.

Noticeable changes in XLM-R model are:

1. Data – XLM-R model is trained on large cleaned CommonCrawl data scaled up to 2.5TB which is a way larger than Wiki-100 corpus which was used in training other multilingual models.
2. Vocabulary – XLM-R vocabulary contains 250k tokens in contrast to RoBERTa which has 50k tokens in its vocabulary. It uses one large shared Sentence Piece Model (SPM) to tokenize words of all languages instead of XLM-100 model which uses different tokenizers for different languages. XLM-R authors assume that similar words across all the languages have similar representation in space.
3. XLM-R is self-supervised, whereas XLM-100 is supervised model. XLM-R samples stream of text from each language and trains the model to predict masked tokens. XLM-100 model required parallel sentences (sentences that have same meaning) in two different languages as input which is a supervised method.

XLM-R Results on Cross-Lingual-Classification on XNLI dataset

XLM-R is now the state-of-the-art multilingual model which outperforms all the previous multi-language models.

Demonstration of NLI Task Using Kaggle Dataset

In this section, we will implement the NLI task using Kaggle dataset.

Kaggle has launched Contradictory My Dear Watson challenge to detect contradiction and entailment in multilingual text. It has shared a training and validation dataset that contains 12120 and 5195 text pairs respectively. This dataset contains textual pairs from 15 different languages – Arabic, Bulgarian, Chinese, German, Greek, English, Spanish, French, Hindi, Russian, Swahili, Thai, Turkish, Urdu, and Vietnamese. Sentence pairs are classified into three classes entailment (0), neutral (1), and contradiction (2).

We will be using the training dataset of this challenge to demonstrate the NLI task. One can run the following code blocks using Google Colab and can download a dataset from this link.

Code Flow

1. Install transformers library

!pip install transformers

2. Load XLM-RoBERTa model –

Since our dataset contains multilingual text, we will be using XLM-R model for checking the accuracy on training dataset.

from transformers import AutoModelForSequenceClassification, AutoTokenizer xlmr= AutoModelForSequenceClassification.from_pretrained(‘joeddav/xlm-roberta-large-xnli’) tokenizer = AutoTokenizer.from_pretrained(‘joeddav/xlm-roberta-large-xnli’)

3. Load training dataset –

import pandas as pd train_data = pd.read_csv(<dataset path>) train_data.head(3)

4. XLM-R model classes –

Before going further do a sanity check to confirm if the model classes notation and the dataset classes notation is same

xlmr.config.label2id

{‘contradiction’: 0, ‘entailment’: 2, ‘neutral’: 1}

We can see that the model classes notation and Kaggle dataset classes notation (entailment (0), neutral (1), and contradiction (2)) is different. Therefore, change the training dataset classes notation to match with model.

5. Change training dataset classes notation –

train_data[‘label’] = train_data[‘label’].replace([0, 2], [2, 0]) train_data.head(3)

6. EDA on dataset –

Check the distribution of training data based on language

train_data_lang = train_data.groupby(‘language’).count().reset_index()[[‘language’,’id’]] # plot pie chart import matplotlib.pyplot as plt import numpy as np plt.figure(figsize=(10,10)) plt.pie(train_data_lang[‘id’], labels = train_data_lang[‘language’], autopct=’%1.1f%%‘) plt.title(‘Distribution of Train data based on Language’) plt.show()

We can see that English constitutes to more than 50% of the training data.

7. Sample data creation –

Since training data has 12120 textual pairs, evaluating all the pairs would be time-consuming. Therefore, we will create a sample data out of training data which will be a representative sampling ie. sample data created will have the same distribution of text pairs based on language as of the training data.

# create a column which tells how many random rows should be extracted for each language train_data_lang[‘sample_count’] = train_data_lang[‘id’]/10 # sample data sample_train_data = pd.DataFrame(columns = train_data.columns) for i in range(len(train_data_lang)): df = train_data[train_data[‘language’] == train_data_lang[‘language’][i]] n = int(train_data_lang[‘sample_count’][i]) df = df.sample(n).reset_index(drop=True) sample_train_data = sample_train_data.append(df) sample_train_data = sample_train_data.reset_index(drop=True) # plot distribution of sample data based on language sample_train_data_lang = sample_train_data.groupby(‘language’).count().reset_index()[[‘language’,’id’]] plt.figure(figsize=(10,10)) plt.pie(sample_train_data_lang[‘id’], labels = sample_train_data_lang[‘language’], autopct=’%1.1f%%‘) plt.title(‘Distribution of Sample Train data based on Language’) plt.show()

We can see that sample data created and the training data have nearly same distribution of text pairs based on language.

8. Functions to get predictions from XLM-R model –

def get_tokens_xlmr_model(data): ”’ Function which creats tokens for the passed data using xlmr model input – Dataframe Output – list of tokens ”’ batch_tokens = [] for i in range(len(data)): tokens = tokenizer.encode(data[‘premise’][i], data[‘hypothesis’][i], return_tensors=’pt’, truncation_strategy=’only_first’) batch_tokens.append(tokens) return batch_tokens def get_predicts_xlmr_model(tokens): ”’ Function which creats predictions for the passed tokens using xlmr model input – list of tokens Output – list of predictions ”’ batch_predicts = [] for i in tokens: predict = xlmr(i)[0][0] predict = int(predict.argmax()) batch_predicts.append(predict) return batch_predicts

9. Predictions on sample data –

sample_train_data_tokens = get_tokens_xlmr_model(sample_train_data) sample_train_data_predictions = get_predicts_xlmr_model(sample_train_data_tokens)

10. Find model accuracy on the predictions –

# plot the confusion matrix and classification report for original labels to the predicted labels import numpy as np import seaborn as sns from sklearn.metrics import classification_report sample_train_data[‘label’] = sample_train_data[‘label’].astype(str).astype(int) x = np.array(sample_train_data[‘label’]) y = np.array(sample_train_data_predictions) cm = np.zeros((3, 3), dtype=int) np.add.at(cm, [x, y], 1) sns.heatmap(cm,cmap=”YlGnBu”, annot=True, annot_kws={‘size’:16}, fmt=’g’, xticklabels=[‘contradiction’,’neutral’,’entailment’],

yticklabels=[‘contradiction’,’neutral’,’entailment’]) matrix = classification_report(x,y,labels=[0,1,2], target_names=[‘contradiction’,’neutral’,’entailment’]) print(‘Classification report : \n‘,matrix)

The model is able to give 93% accuracy on the sample data without any finetuning.

11. Find model accuracy at language level –

sample_train_data[‘prediction’] = sample_train_data_predictions sample_train_data[‘true_prediction’] = np.where(sample_train_data[‘label’]==sample_train_data[‘prediction’], 1, 0) sample_train_data_predicted_lang = sample_train_data.groupby(‘language’).agg({‘id’:’count’, ‘true_prediction’:’sum’}).reset_index()[[‘language’,’id’,’true_prediction’]] sample_train_data_predicted_lang[‘accuracy’] = round(sample_train_data_predicted_lang[‘true_prediction’]/sample_train_data_predicted_lang[‘id’], 2) sample_train_data_predicted_lang = sample_train_data_predicted_lang.sort_values(by=[‘accuracy’],ascending=False) sample_train_data_predicted_lang

Except for English, rest of the languages are having accuracy greater than 94%. Therefore, use a different model for English pairs prediction to further improve accuracy.

12. RoBERTa model for English text pairs prediction –

!pip install regex requests !pip install omegaconf !pip install hydra-core import torch roberta = torch.hub.load(‘pytorch/fairseq’, ‘roberta.large.mnli’) roberta.eval()

RoBERTa model classes notation is same as XLM-R model notations. Therefore, we can directly use the sample data without any class notation changes.

13. Extract only English pairs from sample data –

sample_train_data_en = sample_train_data[sample_train_data[‘language’]==’English’].reset_index(drop=True) sample_train_data_en.shape

(687, 8)

14. Functions to get predictions from RoBERTa model –

def get_tokens_roberta(data): ”’ Function which generates tokens for the passed data using roberta model input – Dataframe Output – list of tokens ”’ batch_tokens = [] for i in range(len(data)): tokens = roberta.encode(data[‘premise’][i],data[‘hypothesis’][i]) batch_tokens.append(tokens) return batch_tokens def get_predictions_roberta(tokens): ”’ Function which generates predictions for the passed tokens using roberta model input – list of tokens Output – list of predictions ”’ batch_predictions = [] for i in range(len(tokens)): prediction = roberta.predict(‘mnli’, tokens[i]).argmax().item() batch_predictions.append(prediction) return batch_predictions

15. Predictions with RoBERTa model –

sample_train_data_tokens = get_tokens_xlmr_model(sample_train_data) sample_train_data_predictions = get_predicts_xlmr_model(sample_train_data_tokens)

16. Accuracy of RoBERTa model –

sample_train_data_en[‘prediction’] = sample_train_data_en_predictions # roberta model accuracy sample_train_data_en[‘true_prediction’] = np.where(sample_train_data_en[‘label’]==sample_train_data_en[‘prediction’], 1, 0) roberta_accuracy = round(sum(sample_train_data_en[‘true_prediction’])/len(sample_train_data_en), 2) print(“Accuracy of RoBERTa model {}“.format(roberta_accuracy))

Accuracy of RoBERTa model 0.92

Accuracy of English text pairs increased from 89% to 92%. Therefore, for predictions on test dataset of Kaggle challenge use RoBERTa for English pairs prediction and XLM-R for predictions of other language pairs.

By this approach, I was able to score 94.167% accuracy on the test dataset.

Conclusion

In this blog ,we have learned what NLI task is, how to achieve this using two state-of-the-art models. There are many more pre-trained models for achieving NLI tasks other than the models discussed in this blog. Few of them are language-specific like German BERT, French BERT, Finnish BERT, etc. multilingual models like Multilingual BERT, XLM-100, etc.

As future steps, one can further achieve task-specific accuracy by finetuning these models with specific data.

References

[1] BERT official paper

[2] RoBERTa official paper

[3] XLM-RoBERTa official paper