FOLDER STRUCTURE
In this submission, you will be finding the main code-files organized into two folders. The folder libraries contain the main driver codes along with other essential helper functions. The folder models contain the model architecture codes along with their configuration codes written in tensorflow. Codes for both time-series classification as well as sequence prediction ae available in the models folder itself. Logs saved during training the models, including the configuration details are saved in the logs folder with the name of the model binded as prefix.
.
├── datasets
│ ├── figs
│ ├── corrupted_series.npz
│ ├── full.npz
│ └── data_reader.py # Code file to read clean and corrupted data
├── libraries
│ ├── losses.py # categorical_crossentorpy, cosine_distance, regression_error, hinge_loss (with basic tf ops)
│ ├── helpers.py # train_test_split, progressBar functions available </br>
│ ├── deep_learning.py # driver file for the three parts described below </br>
│ └── data_analysis.py # data analysis on clean data; pca, tSNE, correlation, fft, etc.
├── models
│ ├── ts_classification
│ ├── __init__
│ ├── ts_mlp # A dense layer neural network for time-series classification
│ ├── ts_rnn # A recurrent neural network for time-series classification
│ └── ts_cnn # A (1D) convolutional neural network for time-series classification
│ ├── ts_prediction
│ ├── __init__
│ └── ts_seq2seq # A recurrent sequence2sequence neural network for time-series prediction
├── logs
│ ├── _model_name_+_loss_function_ # configurations details, training logs, variable summaries, tensorboard, etc.
│ ├── tf_logs # tensorflow Writer logs for tensorboard, variable summaries, histograms and projections
│ │ └── ...
│ ├── infer_best # Model saving during validation/inference
│ ├── train_best # Model saving during training
│ └── model_configs # Configuration files with architectural parameter details
├── images
│ └── _model_name_+_loss_function_.png # Loss and Accuracy Images for different models with different loss functions
This is a three-fold project where-in deep learning models were developed to classify as well as to predict time-series data. Two sets of data were utilized: (i) a time-series data consisting of 30K samples each of length 457 time-units called as clean-data (cl-data) with class labels, (ii) a time-series data consisting of 30K samples each of variable length ranging from [150,457] time-units called as corrupt-data (cr-data) without class labels and drawn from same data distribution as cl-data.
Some custom loss functions have been implemented using basic tf ops.
Before developing deep learning models for time-series classification, a basic data analysis was performed on the (cl-data). The data consists of 30K samples each of length 457 time-units. Instances of null or invalid data has not been found and the data ranged between 0 and 1. There are 10 different labels for the entire dataset. The number of samples-per-class across all classes has a mean of 3000 and standard deviation of 178.2 samples, which implies that the data is not skewed (Had it been skewed, techniques such as weighing outputs wrt inverse-class frequency, drawing samples from skewed classes with higher probability in each mini-batch, etc can be used while training). Time-series plots of randomly picked samples from each class were drawn for visual understanding; looking for any visual patterns in common among same class samples. While it was observed that a visual similarity exists, a tSNE simulation was also done with perplexity 30 and an image of the same can be found in images folder. It was found that the data samples formed dense clusters in tSNE plot. A PCA analysis revealed that first 83, 143 and 239 principal componenets explained a variance of 90%, 95% and 98% respectively. While techniques such as fft, #mean-crossings, signal-energy were also computed, their analysis didn't prove to be of much importance.
Different models such as mlp, rnn, cnn, and fc_rnn were developed in tensorflow for time-series classification task. While mlp and cnn approaches had constrains using variable length inputs (based on the model architectures adopted), the rnn and fc_rnn models do not have that constraint. Individual configure class files were also developed for each model so as to facilitate quick hyper-parameter and architectural modifications during training. In cnn, rnn and fc_rnn models, the choice to provide more than one dimenstion per time-series unit has been added i.e. simple configure file changes will reflect in architectural modifications. These models were then trained on cl-data using different loss functions and the details of same can be found in respective logs folders. Regularization, to be wary of overfitting, at necessary layers in all the models was also incorporated.
During training, it was trivial that the rnn model consumed a lot more time compared to other networks (457 time-units is a big number). In rnn network architecture, dense layers on top of recurrent bi-directional LSTM cells were used. In cnn network, 3, 5 and 7 sized 1D conv filters were used along with dense layers towards the end. In mlp network, series of dense layers were only used. For model training, 24k samples randomly drawn from the 30k sample set was used for training and the rest 6K was kept aside for testing/inference purpose. All logs during training and the best models obtained were saved in corresponding log folders.
Analysis such as what the conv layers have learnt, or analysis on latent embeddings, etc were left for future work. The architectural parameters like number of filters, cells, units, etc. were not fully explored during this experimentation. Also, some modifications that couldn't be incorporated in early stages of the work are:
- Attention mechanism (Input/Temporal-Local/Temporal-Global) in RNN. Although LSTMs are designed to remember information for over long time-units, too long time-units makes remeberance of very long term memory challenging. Input-Attention mechanism can come handy in such scenarios. Truncated BPTT technique can also be used along with Attention mechanism in RNNs.
- A fully-convolutional RNN (FC-RNN) overcomes the long training time of RNNs in time-series data, data such as the one in this experiment. While the fully-convolutional part of the architecture captures short-term patterns in the data, RNN on top of it can better model these recurring short-term patterns in lengthy time-series data.
- An approach similar to (2) wherein we employ auto-encoders to capture short-term time series pattern for further modelling. Also, fully-convolutional architecture that acts as an auto-encoder can be developed to encode each input time-series into latent vector. A fully-convolutional because it relaxes the need for fixed length time-series data.
- While using Deep Learning was the main objective in this experiment, other ML techniques such as k-NN can also be used for classification of the cl-data. But it again imposes a constraint on flexibilty wrt length of training and testing time-series data.
- Short-term feature extraction; Breaking the lengthy sequence into consecutive overlapping short-segments of data and extracting suitable features for each of the short-segments, which then act as a summarized time-series information. MFCC feature extraction is a suitable example here.
It was mentioned that the cr-data was sampled from same distribution as the cl-data. Hence, prior to inference on cr-data, data similar to cr-data was generated from the 6K sample set set aside for inference purpose. Analysis of the different models on this generated data gave some results to expect on cr-data as well as helped to choose the importance of each model in ensemble result. Also, data analysis, as performed on (cl-data) was also done on the (cr-data).
During inference, inputs to different models were given in different manner. While mlp and cnn models required a fixed length input (same as during training), rnn was fed with the effective length input. The effective length for cr-data samples is the legth of each corrupted sample without the trailing zeros.
Finally, an ensemble result collated from best models in mlp, rnn and cnn architectures was used in classification of corrupted data (cr-data). While a high weightage is alloted for rnn model, equal weightage is alloted to the rest of the models. The results are saved in corrupt_labels.npz file.
A simple recurrent seq2seq network was developed for the task of time-series prediction. At encoder, a bi-directional LSTM multi-cell unit was utilized. The output of encoder was then passed through a dense network and then utilized as input for uni-directional multi-cell decoder. Schedule Training technique was incorporated during training phase.
Due to unavailability of good GPU resources, not many experiments could be performed. For training the seq2seq model, (cl-data) was utilized. The encoder side was treated with 100 time-units (picked randomly) from each sample in cl-data; 100 is choosen because each sample in (cr-data) had at-least 100 time-units. A more effective approach would be to train the encoder with lengthier time-units as well so as to capture long-term data patterns. Several un-explored ideas described in Part 1 can also be adopted here, such as attention-mechanism, fully-convolutional layers, etc. to make the architecture more robust and effective to any length test-data. The results with this simple architecture are saved in corrupt_prediction.npz.
Requirements
To set up the environment, please do:
git clone https://github.com/murali1996/ts_classification_prediction
pip install -r requirements.txt