init commit

README.md

@@ -0,0 +1,28 @@
+# Gait Based User Recognition from Channel State Information(CSI) collected from WiFi Data
+Deep learning methods for user recognition from Channel State Information(CSI) collected from WiFi Data. We used [Intel CSI Tool](https://dhalperi.github.io/linux-80211n-csitool/) and [Atheros CSI Tool](https://github.com/kdkalvik/Atheros-CSI-Tool) to collect our gait dataset(private). The methods in this repo convert CSI data to spectrograms and use deep learning methods for user recognition. Refer to the publications listed below for further details of the method.
+
+## Directory Structure
+- `models/`: Deep learning models/experiments
+- `preprocessing/`: Scripts used to preprocess the CSI data
+- `theoretical_derivation/`: Theoretical derivation of the expected spectrograms from human gait
+- `tools/`: Arbitrary tools to plot results/spectrograms, debug CSI tool
+
+## Citation
+Please cite the following paper if you use this repository in your research.
+```
+@inproceedings{pokkunuru2018neuralwave,
+  title={NeuralWave: Gait-based user identification through commodity WiFi and deep learning},
+  author={Pokkunuru, Akarsh and Jakkala, Kalvik and Bhuyan, Arupjyoti and Wang, Pu and Sun, Zhi},
+  booktitle={IECON 2018-44th Annual Conference of the IEEE Industrial Electronics Society},
+  pages={758--765},
+  year={2018},
+  organization={IEEE}
+}
+
+@article{jakkala2019deep,
+  title={Deep CSI learning for gait biometric sensing and recognition},
+  author={Jakkala, Kalvik and Bhuya, Arupjyoti and Sun, Zhi and Wang, Pu and Cheng, Zhuo},
+  journal={arXiv preprint arXiv:1902.02300},
+  year={2019}
+}
+```

models/resnet_500-270.py

@@ -0,0 +1,126 @@
+from tensorflow.keras.utils import multi_gpu_model
+from tensorflow.keras.models import Model
+from tensorflow.keras import optimizers
+from skimage.transform import resize
+from tensorflow.keras import layers
+import numpy as np
+import argparse
+import pickle
+import h5py
+
+def get_args():
+    parser = argparse.ArgumentParser(description="train resent on CSI data",
+                                     formatter_class=argparse.ArgumentDefaultsHelpFormatter)
+    parser.add_argument("--data_dir", type=str, required=True, help="h5py data file")
+    parser.add_argument("--save_path", type=str, required=True, help="h5py data file for weight")
+    parser.add_argument("--history_path", type=str, required=True, help="pkl data file for history")
+    parser.add_argument("--ngpu", type=int, required=True, help="number of gpus")
+    args = parser.parse_args()
+
+    return args
+
+def identity_block(input_tensor, filters, activation_func):
+    filters1, filters2 = filters
+
+    x = layers.Conv2D(filters1, (1, 1), kernel_initializer='he_normal')(input_tensor)
+    x = layers.BatchNormalization(axis=-1)(x)
+    x = layers.Activation('relu')(x)
+
+    x = layers.Conv2D(filters1, (3, 3), padding='same', kernel_initializer='he_normal')(x)
+    x = layers.BatchNormalization(axis=-1)(x)
+    x = layers.Activation('relu')(x)
+
+    x = layers.Conv2D(filters2, (1, 1), kernel_initializer='he_normal')(x)
+    x = layers.BatchNormalization(axis=-1)(x)
+
+    x = layers.add([x, input_tensor])
+    x = layers.Activation('relu')(x)    
+    return x
+
+def conv_block(input_tensor, filters, activation_func, strides=(2, 2)):
+    filters1, filters2 = filters
+
+    x = layers.Conv2D(filters1, (1, 1), strides=strides, kernel_initializer='he_normal')(input_tensor)
+    x = layers.BatchNormalization(axis=-1)(x)
+    x = layers.Activation('relu')(x)
+
+    x = layers.Conv2D(filters1, (3, 3), padding='same', kernel_initializer='he_normal')(x)
+    x = layers.BatchNormalization(axis=-1)(x)
+    x = layers.Activation('relu')(x)
+
+    x = layers.Conv2D(filters2, (1, 1), kernel_initializer='he_normal')(x)
+    x = layers.BatchNormalization(axis=-1)(x)
+
+    shortcut = layers.Conv2D(filters2, (1, 1), strides=strides, kernel_initializer='he_normal')(input_tensor)
+    shortcut = layers.BatchNormalization(axis=-1)(shortcut)
+
+
+    x = layers.add([x, shortcut])
+    x = layers.Activation('relu')(x)
+    return x
+
+def main():
+    args=get_args()
+    data_dir=args.data_dir
+    save_path=args.save_path
+    history_path=args.history_path
+    ngpu=args.ngpu
+
+    lr=1e-3
+    epochs=1000
+    decay=1e-3
+    height = 256
+    width = 2048
+
+    hf = h5py.File(data_dir, 'r')
+    train_classes = np.array(hf.get('labels')).astype(str)
+    num_classes = len(train_classes)
+    X_train = np.expand_dims(hf.get('X_train'), axis=-1)
+    X_test = np.expand_dims(hf.get('X_test'), axis=-1)
+    y_train = np.eye(num_classes)[hf.get('y_train')]
+    y_test = np.eye(num_classes)[hf.get('y_test')]
+    hf.close()
+
+    X_train = np.array([resize(X_train[i], (height, width), mode='reflect', anti_aliasing=True) for i in range(X_train.shape[0])])
+    X_test = np.array([resize(X_test[i], (height, width), mode='reflect', anti_aliasing=True) for i in range(X_test.shape[0])])
+
+    print(X_train.shape, X_test.shape, y_train.shape, y_test.shape, num_classes)
+
+    input_layer = layers.Input(shape=(X_train.shape[1:]))
+
+	x = layers.Conv2D(16, (7, 7),
+	                  activation=None,
+	                  padding='same',
+	                  kernel_initializer='he_normal')(input_layer) 
+	x = layers.BatchNormalization(axis=-1)(x)
+	x = layers.Activation('relu')(x)
+	x = layers.MaxPooling2D((3, 3), strides=(2, 2), padding="same")(x)
+
+	x = conv_block(x, [16, 64], "relu")
+	x = identity_block(x, [16, 64], "relu")
+
+	x = conv_block(x, [32, 128], "relu")
+	x = identity_block(x, [32, 128], "relu")
+
+	x = conv_block(x, [64, 128], "relu")
+	x = identity_block(x, [64, 128], "relu")
+
+	x = layers.GlobalAveragePooling2D()(x)
+	x = layers.Dense(num_classes, activation='softmax')(x)
+
+    model_base = Model(inputs=input_layer, outputs=x)
+    model_base.summary()
+    model = multi_gpu_model(model_base, gpus=ngpu)
+
+    model.compile(loss='categorical_crossentropy', optimizer=optimizers.Adam(lr=lr, decay=decay), metrics=['acc'])
+    history = model.fit(x=X_train, y=y_train, epochs=epochs, validation_data=(X_test, y_test), verbose=2, batch_size=16)
+
+    model_base.save(save_path)
+
+    outfile = open(history_path,'wb')
+    pickle.dump(history.history, outfile)
+    outfile.close()
+
+if __name__ == '__main__':
+    main()
+