def Alex_net():
conv_input = input_data(shape=[None , 227 , 227 , 3] , name='input')
# conv layer1
conv1 = conv_2d(conv_input, 96, filter_size=11, strides=4, activation='relu', bias=True, weights_init=xavier(), name='conv1') # padding !?
lrn1 = local_response_normalization(conv1, name='lrn1') # ***************************
max_pool1 = max_pool_2d(lrn1, kernel_size=3, strides=2, name='max_pool1') # padding !?
# conv layer 2
conv2 = conv_2d(max_pool1, 256, filter_size=5, activation='relu', bias=True, weights_init=xavier(), name='conv2') # ***
lrn2 = local_response_normalization(conv2, name='lrn2')
max_pool2 = max_pool_2d(lrn2, kernel_size=3, strides=2, name='max_pool2')
# conv layer 3
conv3 = conv_2d(max_pool2, 384, filter_size=3, activation='relu', bias='True', weights_init=xavier(), name='conv3') # *** stride !?
# conv layer 4
conv4 = conv_2d(conv3, 384, filter_size=3, activation='relu', bias=True, weights_init=xavier(), name='conv4')
# conv layer 5
conv5 = conv_2d(conv4, 256, filter_size=3, activation='relu', bias=True, weights_init=xavier(), name='conv5')
# dense layer
# fc1
fc1 = fully_connected(conv5, 4096, activation='relu', bias=True, weights_init=xavier(), name='fc1') # built in flatten
fcd1 = dropout(fc1, 0.5, name='fcd1') # how does it handle testing
# fc2
fc2 = fully_connected(fcd1, 4096, activation='relu', bias=True, weights_init=xavier(), name='fc2')
fcd2 = dropout(fc2, 0.5, name='fcd2')
# fc3 (output layer)
fc3 = fully_connected(fcd2, 10, activation='softmax', bias=True, weights_init=xavier(), name='output')
estimator = regression(fc3, optimizer='adam', learning_rate=0.01, loss='categorical_crossentropy', name='estimator')
model = tflearn.DNN(estimator, tensorboard_dir='alex_net_graph', tensorboard_verbose=3) # verbous !?
return model
def _conv(self, x, c):
if 'padding' not in c:
c['padding'] = 'same'
x = tflearn.layers.conv.conv_2d(x,
c['num_filters'],
c['filter_size'],
c['stride'],
activation='linear',
regularizer='L2',
weight_decay=c['l2'],
weights_init=tfi.xavier(),
padding=c['padding'],
name='conv%d' % (self.conv_count))
logger.info('conv%d output' % (self.conv_count),
x.get_shape().as_list())
self.conv_count += 1
return tf.nn.relu(x)