| from core.leras import nn | |
| tf = nn.tf | |
| class TLU(nn.LayerBase): | |
| """ | |
| Tensorflow implementation of | |
| Filter Response Normalization Layer: Eliminating Batch Dependence in theTraining of Deep Neural Networks | |
| https://arxiv.org/pdf/1911.09737.pdf | |
| """ | |
| def __init__(self, in_ch, dtype=None, **kwargs): | |
| self.in_ch = in_ch | |
| if dtype is None: | |
| dtype = nn.floatx | |
| self.dtype = dtype | |
| super().__init__(**kwargs) | |
| def build_weights(self): | |
| self.tau = tf.get_variable("tau", (self.in_ch,), dtype=self.dtype, initializer=tf.initializers.zeros() ) | |
| def get_weights(self): | |
| return [self.tau] | |
| def forward(self, x): | |
| if nn.data_format == "NHWC": | |
| shape = (1,1,1,self.in_ch) | |
| else: | |
| shape = (1,self.in_ch,1,1) | |
| tau = tf.reshape ( self.tau, shape ) | |
| return tf.math.maximum(x, tau) | |
| nn.TLU = TLU |