Today I am going to share an experiment for sentiment analysis on IMDB movie dataset from Kaggle. In short, it is sentiment analysis done using bi directional multi layer RNN on Pytorch, glove embedding is used beneath for vectors.

Exploring Glove embedding

Here we can see that, there are 400000 word embeddings in vocabulary and is trained using 6 billion words. Shape here represents that all of these 400000 words are represented by dimensionality of 100.

With this vector representation we can find nearest words in vocab for a specific word.

Exploring Data Set

Splitting the data set into train and validation.

There is some basic NLP pre processing involved that can be seen here. Let us get moving to loading data to Pytorch TabularDataset now.

Let us look at common frequent words from our train data.

And here is the integer string mapping of our keywords. You can see that there are 2 extra words ‘<unk>’,’<pad>’ for unknown words and padding respectively.

I am going to use BucketIterator from Pytorch’s torch.text, which clubs together similar length strings within a batch and avoid extra padding.

Modeling

Let us see how the data looks.

I have tried different epochs and parameters, this is smaller one for reference.

Here are some sample checks.

Conclusion

RNNs perform very well for text data and bring that extra edge performance. This is project report for zero to gans course by Free code Camp and Jovian. There are many other techniques like regularization, gradient clipping and hyper parameter tunning that could help improving the model. There are other embedding options and you can even train your own embedding to have better results. Do checkout full notebook here, the notebook is hosted on Jovian which is an awesome platform for organizing such notebooks.

References:

• Deep Learning with PyTorch: Zero to GANs
• torchtext.data - torchtext 0.4.0 documentation
• Natural Language Processing with PyTorch

PyTorch : Open source deep learning framework by Facebook.

pytorch.org says “PyTorch is defined as a Python-based scientific computing package that targets to replace NumPy, and use the power of GPUs. It is also a platform to perform deep learning algorithms”.

Tensors are usually defined as multi dimensional arrays.

Tensors are represented by multi dimensional arrays and have a geometric meaning. A scalar (torch.tensor(4.)) is zero-order tensor or rank zero tensor. A vector(torch.tensor([1., 2, 3, 4])) is a one-dimensional or first order tensor, a matrix(torch.tensor([[5., 6], [7, 8], [9, 10]])) is a two-dimensional or second order tensor, and so on.

Tensors: ref

1. Tensor as an object is invariant under change of coordinates, but it has components that change in a special predictable way under change of coordinates.

2. Tensor is a collection of vectors and co-vectors combined together using tensor product.

In this article, I will focus on 5 common PyTorch tensor functions:

• torch.where()
• torch.mean()
• torch.split()
• torch.t()
• torch.bitwise_and()

Bonus: I will create tensors in different ways while exploring these 5 functions :)

Let’s get started.

torch.where(condition, x, y) → Tensor

torch.where() does element-wise conditional assignment. In general, it returns a tensor of elements selected from x or y, depending on condition. This is an awesome equivalent of numpy.where().

Example 1: Normal case💡 Tensor created directly with matrix.

Example 2: Normal case💡 Tensor created using torch.randn().

Example 3 : break case💡 Tensor using zeros() and ones()

Note : Make sure that condition returns a BoolTensor and not bool.

torch.mean(input) → Tensor

torch.mean() returns mean value of all elements of input tensor.

Example 1: Normal case💡 Tensor using as_tensor()

Example 2: Normal case

Example 3 : break case

torch.split(tensor, split_size_or_sections, dim=0)

torch.split() splits the tensor into chunks of size specified.

Example 1: Normal case💡 Tensor using torch.linspace()

Example 2: Normal case

Example 3 : break case

Note: split_size_or_sections must be int or tuple of ints, not float. split_with_sizes expects split_sizes to sum exactly to size of input.

torch.t(input) → Tensor

Transpose a tensor.

Example 1: Normal case💡 Tensor using torch.as_strided()

Example 2: Normal case with BoolTensor💡 Tensor using numpy.array()

Example 3 : break case

Note: Function expects input to be <= 2-D tensor and transposes dimensions 0 and 1. It returns scalar and 1D tensors as is, for 2D tesnors it is equivalent to transpose(input, 0, 1).

torch.bitwise_and(input, other, out=None) → Tensor

Computes bitwise AND for tensors (logical AND for BoolTensors).

Example 1: Normal case💡 Tensor using torch.arange()

Example 2: Normal case with BoolTensor

Example 3 : break case

Note: The input tensor must be of integral or Boolean types.

Conclusion

This article is a peek to week 1 of course entitled Deep Learning with PyTorch — Zero to GANs. It’s a 6 weeks course, livestreamed in freeCodeCamp.org’s YouTube page. Hope you liked the brief introduction to PyTorch tesnor functions.

Refer the full code here. For more functions and details visit official PyTorch docs.

References

• Official documentation for torch.Tensor: https://pytorch.org/docs/stable/tensors.html
• Eignchris channel https://www.youtube.com/watch?v=TvxmkZmBa-k
• https://pytorch.org/tutorials/beginner/blitz/tensor_tutorial.html