This repository is a PyTorch implementation of Human Activity Recognition using convolutional neural networks.
Notes:
Table of Contents:
# clone repo
pip install -r requirements.txtUse python main.py to run the preset configuration to train and evaluate the model. The preset
configuration can be found in hyperparams.ini.
To run a custom experiment use python main.py <experiment name> <params>. For example:
python main.py -n har_1 -d har -b 64 --lr 0.0001 You can evaluate a pre-trained model with the following:
python main.py -n har_1 --is-eval-onlyRunning will create a directory results/<saving-name>/ which contains:
To get the help menu run python main.py -h which yields the following output:
usage: main.py [-h] [-d {har,newdataset}] [-b BATCH_SIZE] [--lr LR]
[-e EPOCHS] [-s IS_STANDARDIZED] [-m {Cnn1,Cnn2}] [-n NAME]
[--is-eval-only] [--no-test]
PyTorch implementation of CNN's for Human Activity Recognition
optional arguments:
-h, --help show this help message and exit
Training specific options:
-d, --dataset {har,newdataset}
Path to training data. (default: har)
-b, --batch-size BATCH_SIZE
Batch size for training. (default: 32)
--lr LR Learning rate. (default: 0.0005)
-e, --epochs EPOCHS Maximum number of epochs to run for. (default: 20)
-s, --is_standardized IS_STANDARDIZED
Whether to standardize the data. (default: True)
Model specific options:
-m, --model-type {Cnn1,Cnn2}
Type of encoder to use. (default: Cnn2)
General options:
-n, --name NAME Name of the model for storing and loading purposes.
(default: HAR_1)
Evaluation specific options:
--is-eval-only Whether to only evaluate using precomputed model
`name`. (default: False)
--no-test Whether or not to compute the test losses.` (default:
False)
The repository uses the Human Activity Recognition Using Smartphones Data Set available at:
Two similar models have been implemented: Cnn1 & Cnn2. The only difference is that Cnn1 has a filter size of 3 and Cnn2 has a filter of size 5.
Cnn2 architecture:
CNN(
(encoder): Cnn2(
(conv1): Conv1d(9, 64, kernel_size=(5,), stride=(1,))
(relu1): Relu()
(conv2): Conv1d(64, 64, kernel_size=(5,), stride=(1,))
(relu2): Relu()
(drop): Dropout(p=0.6, inplace=False)
(pool): MaxPool1d(kernel_size=2, stride=2, padding=0, dilation=1, ceil_mode=False)
(lin3): Linear(in_features=3840, out_features=100, bias=True)
(relu3): Relu()
(lin4): Linear(in_features=100, out_features=6, bias=True)
(relu4): Relu()
)
)Pre-trained models for HAR_Cnn1_StdFalse, HAR_Cnn1_StdTrue, HAR_Cnn2_StdFalse and HAR_Cnn2_StdTrue
can be found in the results folder. These can be loaded and evaluated with the command:
python main.py -n HAR_Cnn2_StdTrue --is-eval-only
The best results were achieved with the HAR_Cnn2_StdTrue run:
***************************************************
* Evaluating Train Accuracy *
***************************************************
Train accuracy of the network on the 7352 train sequences: 98.86 %
Accuracy of class 0: walking : 100.00 %
Accuracy of class 1: walking upstairs : 100.00 %
Accuracy of class 2: walking downstairs : 100.00 %
Accuracy of class 3: sitting : 98.91 %
Accuracy of class 4: standing : 94.91 %
Accuracy of class 5: laying : 100.00 %
Confusion matrix:
--------------------------------
0 1 2 3 4 5
0 1226 0 0 0 0 0
1 0 1073 0 0 0 0
2 0 0 986 0 0 0
3 0 0 0 1272 14 0
4 0 0 0 70 1304 0
5 0 0 0 0 0 1407
***************************************************
* Evaluating Test Accuracy *
***************************************************
Test accuracy of the network on the 2947 test sequences: 92.57 %
Accuracy of class 0: walking : 98.39 %
Accuracy of class 1: walking upstairs : 95.54 %
Accuracy of class 2: walking downstairs : 98.81 %
Accuracy of class 3: sitting : 83.30 %
Accuracy of class 4: standing : 80.64 %
Accuracy of class 5: laying : 100.00 %
Confusion matrix:
--------------------------------
0 1 2 3 4 5
0 488 0 8 0 0 0
1 3 450 18 0 0 0
2 1 4 415 0 0 0
3 1 1 0 409 74 6
4 0 0 0 103 429 0
5 0 0 0 0 0 537