| import tensorflow as tf |
| from tensorflow.keras import models |
| from tensorflow.keras import layers |
|
|
| class FlowNet: |
| """ |
| Uses the optical flow calculated by the Lucas-Kanade algorithm as input. |
| It splits the x and y shifts and runs them though lstms before combining them again. |
| The output are 2 tensors. One for the xyz position values and one for the 4 quaternions. |
| """ |
| def __init__(self, lstms=1, activ=None, bn=0): |
| self.lstms = lstms |
| self.activ = activ |
| self.bn = bn |
| print('build IMU ...') |
|
|
| def create_model(self): |
| inputs = tf.keras.Input(shape=(120, 160, 2)) |
| if self.bn == 1: |
| bnorm = layers.BatchNormalization()(inputs) |
| r = layers.Lambda(lambda x: x[:, :, :, 0], name="lambda_layer1")(bnorm) |
| g = layers.Lambda(lambda x: x[:, :, :, 1], name="lambda_layer2")(bnorm) |
| else: |
| r = layers.Lambda(lambda x: x[:, :, :, 0], name="lambda_layer1")(inputs) |
| g = layers.Lambda(lambda x: x[:, :, :, 1], name="lambda_layer2")(inputs) |
| if self.lstms == 2: |
| lstm1 = layers.LSTM(units=100, return_sequences=True)(r) |
| cls3_fc1 = layers.LSTM(units=50)(lstm1) |
| lstm2 = layers.LSTM(units=100, return_sequences=True)(g) |
| cls3_fc2 = layers.LSTM(units=50)(lstm2) |
| elif self.lstms == 1: |
| cls3_fc1 = layers.LSTM(units=50)(r) |
| cls3_fc2 = layers.LSTM(units=50)(g) |
| concat = layers.Concatenate(name='concat', axis=-1)([cls3_fc1, cls3_fc2]) |
| if self.activ == "relu": |
| concat = layers.Activation("relu")(concat) |
| elif self.activ == "softmax": |
| concat = layers.Activation("softmax")(concat) |
| pose_xyz = layers.Dense(3,name='pose_xyz')(concat) |
| pose_wpqr = layers.Dense(4,name='pose_wpqr')(concat) |
| return models.Model(inputs=inputs, outputs=[pose_xyz, pose_wpqr]) |
|
|
|
|
| if __name__ == "__main__": |
| |
| m = FlowNet(lstms=1) |
| Model = m.create_model() |
| Model.summary() |
|
|
