Reproduction of the 'Deep Convolutional and LSTM Recurrent Neural Networks for Multimodal Wearable Activity Recognition' paper.

Paper link

Dataset download link

Blog Post

The zip downloaded from the link above should be placed in the data directory in the root of the project.

To pre-process the data run the following command:

python preprocess_data.py -i "data/OpportunityUCIDataset.zip" -o "data/pre-processed.pkl"

To load the pre-trained model, use the following lines of code:

from model import DeepConvLSTM
import torch

SAVE_MODEL_NAME = "DeepConvLSTM_Opportunity_Model.pt"

model = DeepConvLSTM()
model.load_state_dict(torch.load(SAVE_MODEL_NAME))

After this you can insert any code you want for evaluating the model.

tensorboard --logdir runs/

In the paper they mention 10e-3 lr, but this is too high and they actually use 10e-4. 10e-3 is too high and will cause overshooting? so there is never any improvement We found this out by checking their own training code.

Also we found that the paper uses l2 regularization, but they don't mention it in the paper. They use a value of 0.0001 for this. This is called the weight_decay in the pytorch code for RMSprop.

Lastly, in the paper the do mention weight_decay, but this is understood wrong by them. Instead the weight_decay is actually the rho of the RMSprop algorithm. In pytorch this is called alpha.