Dimension-Adaptive Neural Architecture (DANA)
(Accepted for ACM IMWUT Journal and will be presented in ACM UbiComp 2021 )
(The overall idea and part of the results are presented in MobiUK2021. Presentation video is here)
https://arxiv.org/pdf/2008.02397.pdf
Motion sensors embedded in wearable and mobile devices allow for dynamic selection of sensor streams and sampling rates, enabling useful applications, e.g. for power management or control of data sharing. While deep neural networks (DNNs) achieve competitive accuracy in sensor data classification, current DNN architectures only process data coming from a fixed set of sensors with a fixed sampling rate, and changes in the dimensions of their inputs cause considerable accuracy loss, unnecessary computations, or failure in operation. To address this problem, we introduce a dimension-adaptive pooling (DAP) layer that makes DNNs flexible and more robust to changes in sampling rate and in sensor availability. DAP operates on convolutional filter maps of variable dimensions and produces an input of fixed dimensions suitable for feedforward and recurrent layers. Further, we propose a dimension-adaptive training (DAT) procedure for enabling DNNs that use DAP to better generalize over the set of feasible data dimensions at inference time. DAT comprises the random selection of dimensions during the forward passes and optimization with accumulated gradients of several backward passes. Combining DAP and DAT, we show how to transform existing non-adaptive DNNs into a Dimension-Adaptive Neural Architecture (DANA), while keeping the same number of parameters. Compared to the existing approaches, our solution provides better average classification accuracy over the range of possible data dimensions at inference time, and it does not need up-sampling or imputation, thus reduces unnecessary computations at inference time. Experimental results, on four benchmark real-world datasets of human activity recognition as well as three synthetic datasets, show that DANA prevents significant losses in classification accuracy of the state-of-the-art DNNs and, compared to baselines, it better captures correlated patterns in sensor data under dynamic sensor availability and varying sampling rates.
Citation
Please use:
@article{malekzadeh2020dana,
title = {DANA: Dimension-Adaptive Neural Architecture for Multivariate Sensor Data},
author = {Malekzadeh, Mohammad and Clegg, Richard G and Cavallaro, Andrea and Haddadi, Hamed},
year = {2021},
issue_date = {September 2021},
publisher = {Association for Computing Machinery},
address = {New York, NY, USA},
volume = {5},
number = {3},
url = {https://doi.org/10.1145/3478074},
doi = {10.1145/3478074},
journal = {Proc. ACM Interact. Mob. Wearable Ubiquitous Technol.},
month = sep,
articleno = {120},
numpages = {27}
}To Beging With:
-
You can just click on this
-
For the Controlled Experiments see this
Another Example
import os
import sys
sys.path.append("../")
from dana import models as dana_models
from dana import trainers as dana_trainers
import numpy as np
import tensorflow as tf
print("TensorFlow Version: ", tf.__version__)TensorFlow Version: 2.3.0
## Download and UnZip UTwente Dataset into the Path: dana/datasets/Utwente
## https://www.utwente.nl/en/eemcs/ps/dataset-folder/ut-data-complex.rar
from dana.datasets import utwente
(X_train, Y_train, X_test, Y_test, data_class_names, data_class_weights) = utwente.load_data(shuffle=False)
rnd_seed = 0
tf.random.set_seed(rnd_seed)
np.random.RandomState(seed=rnd_seed).permutation(len(X_train))
indices = np.random.permutation(len(X_train))
X_train = X_train[indices]
Y_train = Y_train[indices]
X_train = np.expand_dims(X_train,3)
X_test = np.expand_dims(X_test,3)
print(X_train.shape, Y_train.shape, X_test.shape, Y_test.shape)
print("Activity Weights",dict(zip(data_class_names,data_class_weights.values())))(14611, 128, 9, 1) (14611,) (3634, 128, 9, 1) (3634,)
Activity Weights {'walk': 0.9999, 'stand': 0.9999, 'jog': 0.9999, 'sit': 0.9999, 'bike': 0.9999, 'ups': 0.9999, 'downs': 0.9999, 'type': 0.9999, 'write': 0.9999, 'coffee': 0.9999, 'talk': 0.9999, 'smoke': 0.9999, 'eat': 1.0008}
def test_model(model, X_val, Y_val, batch_size, W_comb=None, H_comb=None):
if not W_comb:
W_comb = X_val.shape[1]
if not H_comb:
H_comb = list(range(X_val.shape[2]))
val_accuracy = tf.keras.metrics.Accuracy()
val_dataset = tf.data.Dataset.from_tensor_slices((X_val, Y_val))
val_dataset = val_dataset.batch(batch_size)
for (X, Y) in val_dataset:
X = X.numpy()
X = X[:,:,H_comb,:]
X = tf.image.resize(X, (W_comb, len(H_comb)))
logits = model(X)
prediction = tf.argmax(logits, axis=1, output_type=tf.int32)
val_accuracy(prediction, Y)
return val_accuracy.result()W_combinations = list(np.arange(16,129,16))
H_combinations = [[0,1,2], [3,4,5], [6,7,8],
[0,1,2,3,4,5], [0,1,2,6,7,8],
[3,4,5,0,1,2], [3,4,5,6,7,8],
[6,7,8,0,1,2], [6,7,8,3,4,5],
[0,1,2,3,4,5,6,7,8],[0,1,2,6,7,8,3,4,5],
[3,4,5,0,1,2,6,7,8],[3,4,5,6,7,8,0,1,2],
[6,7,8,3,4,5,0,1,2],[6,7,8,0,1,2,3,4,5]] ## Acc & Gyr & Mgn
W_combinations_validation = W_combinations
H_combinations_validation = [[0,1,2], [3,4,5], [6,7,8],
[0,1,2,3,4,5], [0,1,2,6,7,8],[3,4,5,6,7,8],
[0,1,2,3,4,5,6,7,8]]
n_batch_per_train_setp = 5runs = np.arange(1,2)
for run in runs:
## ["standard", "reptile", "weight_avg", "dat"]
for method in ["dat"]:
history_accuracy = []
history_time = []
model = dana_models.Ordonez2016DeepWithDAP((None, None, 1), len(np.unique(Y_train)), (8,9))
print(model.summary())
loss_fn = tf.keras.losses.SparseCategoricalCrossentropy(from_logits=False)
optimizer = tf.keras.optimizers.Adam()
# print("**__** Run {} -- Method {}".format(run, method))
trainer = dana_trainers.ModelTrainer(model, loss_fn, optimizer, data_class_weights,
num_epochs=1000, batch_size=128, patience=100,
W_combinations=W_combinations,
H_combinations=H_combinations,
W_combinations_validation=W_combinations_validation,
H_combinations_validation=H_combinations_validation,
n_batch_per_train_setp=n_batch_per_train_setp)
save_dir_model="saved_models/exp_ut_dana/Ordonez2016DeepWithDAP_2_"+method+"_"+\
str((len(W_combinations),len(H_combinations),n_batch_per_train_setp))+"/best_"+str(run)
##### Uncomment this if you want to train the model
# _, best_accuracy_record, avg_training_time = trainer.train_model(X_train[len(X_train)//10:], Y_train[len(X_train)//10:],
# X_train[:len(X_train)//10], Y_train[:len(X_train)//10],
# method,
# save_dir=save_dir_model,
# verbose=10)
# history_accuracy.append(best_accuracy_record)
# history_time.append(avg_training_time)
# save_dir_results = "saved_models/exp_ut_dana/Ordonez2016DeepWithDAP_2_"+method+"_"+\
# str((len(W_combinations),len(H_combinations),n_batch_per_train_setp))
# if not os.path.exists(save_dir_results):
# os.makedirs(save_dir_results)
# np.save(save_dir_results+"/accuracy_results_"+str(run)+".npy", history_accuracy)
# np.save(save_dir_results+"/timing_results_"+str(run)+".npy", history_time)
#####
model.load_weights(save_dir_model)
best_accuracy_record = np.zeros((len(W_combinations_validation),len(H_combinations_validation)))
for w in range(len(W_combinations_validation)):
for h in range(len(H_combinations_validation)):
best_accuracy_record[w,h] = test_model(model, X_test, Y_test, 128,
W_combinations_validation[w],
H_combinations_validation[h])
print("\n Run {} *** Mean of Test Accuraccies: {:.3f}".format(run, np.mean(best_accuracy_record)))
print("\n Run {} *** Test Accuraccies: {} ".format(run, (best_accuracy_record)))Model: "functional_1"
_________________________________________________________________
Layer (type) Output Shape Param #
=================================================================
input_1 (InputLayer) [(None, None, None, 1)] 0
_________________________________________________________________
conv2d (Conv2D) (None, None, None, 64) 384
_________________________________________________________________
conv2d_1 (Conv2D) (None, None, None, 64) 20544
_________________________________________________________________
conv2d_2 (Conv2D) (None, None, None, 64) 20544
_________________________________________________________________
conv2d_3 (Conv2D) (None, None, None, 64) 20544
_________________________________________________________________
DAP (DimensionAdaptivePoolin (None, 8, 576) 0
_________________________________________________________________
lstm_1 (LSTM) (None, 8, 128) 360960
_________________________________________________________________
dot_1 (Dropout) (None, 8, 128) 0
_________________________________________________________________
lstm_2 (LSTM) (None, 128) 131584
_________________________________________________________________
dot_2 (Dropout) (None, 128) 0
_________________________________________________________________
act_smx (Dense) (None, 13) 1677
=================================================================
Total params: 556,237
Trainable params: 556,237
Non-trainable params: 0
_________________________________________________________________
None
Run 1 *** Mean of Test Accuraccies: 0.871
Run 1 *** Test Accuraccies: [[0.8503027 0.7264722 0.70886075 0.88855255 0.89323062 0.85690701
0.92460102]
[0.88442487 0.81783158 0.72922397 0.91001654 0.91496974 0.88112271
0.93698406]
[0.89460647 0.8450743 0.72812331 0.91964775 0.92432582 0.89378095
0.94221246]
[0.89955974 0.84617501 0.72812331 0.92129886 0.9235003 0.89378095
0.93588334]
[0.89955974 0.85250413 0.72564667 0.92212439 0.92377543 0.89790863
0.93753439]
[0.89570719 0.85112822 0.72592187 0.92102367 0.9232251 0.89900935
0.93450743]
[0.89020365 0.84892678 0.72289491 0.9179967 0.92267472 0.89488167
0.93120527]
[0.88992846 0.84727573 0.71711612 0.91909742 0.92515135 0.8970831
0.93175566]]
import matplotlib.pyplot as plt
labels = ["Only Accl is Available", "Only Gyro is Available", "Only Mgnt is Available",
"Accl and Gyro are Available", "Accl and Mgnt are Available", "Gyro and Mgnt are Available",
"All Sensors are Available "]
markers = ["-vr","-<r", "-^r",
"--pb", "--ob", "--sb",
"-.*g"]
fig, ax = plt.subplots(nrows=1, ncols=1, figsize=(20,12))
fig.tight_layout()
for i in range(len(H_combinations_validation)):
y_data = best_accuracy_record[:,i]*100
x_data = np.floor(np.array(W_combinations)/(2.56))
if i == len(H_combinations_validation)-1:
ax.plot(x_data, y_data, markers[i], label=labels[i], markersize= 50, linewidth=7,
markerfacecolor="None", markeredgewidth=3)
else:
ax.plot(x_data, y_data, markers[i], label=labels[i], markersize= 40, linewidth=5,
markerfacecolor="None", markeredgewidth=3)
ax.set_xticks(x_data)
plt.setp(ax.get_xticklabels(), fontsize=35)
ax.set_yticks(np.arange(50,101,10))
plt.setp(ax.get_yticklabels(), fontsize=35)
fig.text(0.4, -0.05, "Sampling Rate (Hz)", size = 40)
fig.text(-0.05, 0.5, "Classification Accuracy (%)", va='center', rotation='vertical', size = 40)
ax.legend(loc= "lower center",prop={'size': 30}, ncol=1)
plt.show()