The project is based on this repository which is presented as a tutorial. It consists of Human Activity Recognition (HAR) using stacked residual bidirectional-LSTM cells (RNN) with TensorFlow.
It resembles to the architecture used in "Google’s Neural Machine Translation System: Bridging the Gap between Human and Machine Translation" without an attention mechanism and with just the encoder part. In fact, we started coding while thinking about applying residual connections to LSTMs - and it is only afterwards that we saw that such a deep LSTM architecture was already being used.
Here, we improve accuracy on the previously used dataset from 91% to 94% and we push the subject further by trying our architecture on another dataset.
Our neural network has been coded to be easy to adapt to new datasets (assuming it is given a fixed, non-dynamic, window of signal for every prediction) and to use different breadth, depth and length by using a new configuration file.
Here is a simplified overview of our architecture:
Simplified view of a "2x2" architecture. We obtain best results with a "3x3" architecture (details below figure).
Bear in mind that the time steps expands to the left for the whole sequence length and that this architecture example is what we call a "2x2" architecture: 2 residual cells as a block stacked 2 times for a total of 4 bidirectional cells, which is in reality 8 unidirectional LSTM cells. We obtain best results with a 3x3 architecture, consisting of 18 LSTM cells.
Neural network's architecture
Mainly, the number of stacked and residual layers can be parametrized easily as well as whether or not bidirectional LSTM cells are to be used. Input data needs to be windowed to an array with one more dimension: the training and testing is never done on full signal lengths and use shuffling with resets of the hidden cells' states.
We are using a deep neural network with stacked LSTM cells as well as residual (highway) LSTM cells for every stacked layer, a little bit like in ResNet, but for RNNs.
Our LSTM cells are also bidirectional in term of how they pass trough the time axis, but differ from classic bidirectional LSTMs by the fact we concatenate their output features rather than adding them in an element-wise fashion. A simple hidden ReLU layer then lowers the dimension of those concatenated features for sending them to the next stacked layer. Bidirectionality can be disabled easily.
TensorFlow 0.11 and
Sklearn is also used.
The two datasets can be loaded by running
python download_datasets.py in the
To preprocess the second dataset (opportunity challenge dataset), the
signal submodule of scipy is needed, as well as
Results using the previous public domain HAR dataset
This dataset named
A Public Domain Dataset for Human Activity Recognition Using Smartphones is about classifying the type of movement amongst six categories:
(WALKING, WALKING_UPSTAIRS, WALKING_DOWNSTAIRS, SITTING, STANDING, LAYING).
The bests results for a test accuracy of 94% are achieved with the 3x3 bidirectional architecture with a learning rate of
0.001 and an L2 regularization multiplier (weight decay) of
0.005, as seen in the
Training and testing can be launched by running the config:
Results from the Opportunity dataset
The neural network has also been tried on the Opportunity dataset to see if the architecture could be easily adapted to a similar task.
Don't miss out this nice video that offers a nice overview and understanding of the dataset.
We obtain a test F1-score of 0.893. Our results can be compared to the state of the art DeepConvLSTM that is used on the same dataset and achieving a test F1-score of 0.9157.
We only used a subset of the full dataset as done in other research in order to simulate the conditions of the competition, using 113 sensor channels and classifying on the 17 categories output (and with the NULL class for a total of 18 classes). The windowing of the series for feeding in our neural network is also the same 24 time steps per classification, on a 30 Hz signal. However, we observed that there was no significant difference between using 128 time steps or 24 time steps (0.891 vs 0.893 F1-score). Our LSTM cells' inner representation is always reset to 0 between series. We also used mean and standard deviation normalization rather than min to max rescaling to rescale features to a zero mean and a standard deviation of 0.5. More details about preprocessing are explained furthermore in their paper. Other details, such as the fact that the classification output is sampled only at the last timestep for the training of the neural network, can be found in their preprocessing script that we adapted in our repository.
The config file can be runned like this:
For best results, it is possible to readjust the learning rate such as in the