def build(
self,
rng,
n_fms_in,
n_fms_out,
conv_kernel_dims,
pool_prms, # Can be []
conv_w_init_method,
conv_pad_mode,
use_bn,
moving_avg_length, #If this is <= 0, we are not using BatchNormalization, even if above is True.
activ_func="relu",
dropout_rate=0.0):
"""
param rng: numpy.random.RandomState used to initialize weights
param inputToLayer: tensor5 of shape inputToLayerShape
param filterShape: (number of filters, num input feature maps,
filter height, filter width, filter depth)
param inputToLayerShape: (batch size, num input feature maps,
image height, image width, filter depth)
use_bn: True of False. Used to not allow doing BN on first layers straight on image, even if rollingAvForBnOverThayManyBatches > 0.
"""
self._n_fms_in = n_fms_in
self._n_fms_out = n_fms_out
# Order of what is applied, ala He et al "Identity Mappings in Deep Residual Networks" 2016
# Input -> [ BatchNorm OR biases applied] -> NonLinearity -> DropOut -> Pooling --> Conv ]
#------------------ Batch Normalization ------------------
if use_bn and moving_avg_length > 0:
self._bn_l = dm_layers.BatchNormLayer(moving_avg_length,
n_channels=n_fms_in)
self._layers.append(self._bn_l)
else: #Not using batch normalization
#make the bias terms and apply them. Like the old days before BN's own learnt bias terms.
bias_l = dm_layers.BiasLayer(n_fms_in)
self._layers.append(bias_l)
#------------ Apply Activation/ non-linearity -----------
act_l = dm_layers.get_act_layer(activ_func, n_fms_in)
self._layers.append(act_l)
#------------- Dropout --------------
dropout_l = dm_layers.DropoutLayer(dropout_rate, rng)
self._layers.append(dropout_l)
#----------- Pooling ----------------------------------
if len(pool_prms) > 0: #Max pooling is actually happening here...
# if len == 0, pool_prms == [], no max pooling before this conv.
pooling_l = PoolingLayer(pool_prms[0], pool_prms[1], pool_prms[2],
pool_prms[3])
self._layers.append(pooling_l)
# --------- Convolution ---------------------------------
conv_l = self._create_conv_layer(n_fms_in, n_fms_out, conv_kernel_dims,
conv_w_init_method, conv_pad_mode,
rng)
self._layers.append(conv_l)
def build(self, rng, n_fms, t=1):
# t: temperature. Scalar
self._n_fms_in = n_fms
self._n_fms_out = n_fms
self._temperature = t
self._bias_l = dm_layers.BiasLayer(n_fms)
self._layers.append(self._bias_l)