def __init__(self, input_shapes, input_type='yx', **optsettings):
lr = optsettings.get('learning_rate') or 0.01
optimizer = optsettings.get('optimizer') or 'adam'
loss_fn = optsettings.get('loss_fn') or 'categorical_crossentropy'
builder = builders.GraphBuilder()
shape = input_shapes[input_type]
in_name = builder.add_input_layer(shape, input_type, name='input')
builder.start_new_path()
builder.add_reshape_layer((*shape, 1))
builder.add_conv2d_layer(32, 3, activation='relu', regularizer="L2")
builder.add_maxpool2d_layer(2)
builder.add_lrn_layer()
builder.add_conv2d_layer(64, 3, activation='relu', regularizer="L2")
builder.add_maxpool2d_layer(2)
builder.add_lrn_layer()
builder.add_fc_layer(128, activation='tanh')
builder.add_dropout_layer(0.8)
builder.add_fc_layer(256, activation='tanh')
builder.add_dropout_layer(0.8)
out_name = builder.add_fc_layer(2, activation='softmax')
builder.end_current_path()
builder.finalize(out_name,
name='target',
learning_rate=lr,
optimizer=optimizer,
loss=loss_fn)
super(self.__class__, self).__init__(builder)
self.input_type = input_type
self.input_shape = input_shapes[input_type]
self.in_name, self.out_name = in_name, out_name
def __init__(self):
builder = builders.GraphBuilder()
input_layer = builder.add_input_layer((3, 3, 1), 'test')
conv1 = builder.add_conv2d_layer(10,
3,
filter_strides=1,
weights_init='zeros',
bias_init='zeros')
fc1 = builder.add_fc_layer(10, weights_init='zeros', bias_init='zeros')
fc2 = builder.add_fc_layer(3, weights_init='zeros', bias_init='zeros')
builder.finalize(fc2)
super(MockNeuralNetwork, self).__init__(builder)
all_layers = builder.layers_dict
self._in, self._out = input_layer, fc2
self.exp_inputs = {input_layer: all_layers[input_layer]}
self.exp_output = {fc2: all_layers[fc2]}
self.exp_trainable_layers = {
conv1: all_layers[conv1],
fc1: all_layers[fc1],
fc2: all_layers[fc2],
}
self.exp_hidden_layers = {
conv1: all_layers[conv1],
fc1: all_layers[fc1],
}
self.exp_paths = {input_layer: [conv1, fc1, fc2]}
example_input = nprand.randint(low=0, high=255, size=3 * 3)
self.example_input = {self._in: example_input.reshape((1, 3, 3, 1))}
def __init__(self, input_shapes, input_type='yx', **optsettings):
lr = optsettings.get('learning_rate') or 0.001
optimizer = optsettings.get('optimizer') or 'adam'
loss_fn = optsettings.get('loss_fn') or 'categorical_crossentropy'
builder = builders.GraphBuilder()
shape = input_shapes[input_type]
in_name = builder.add_input_layer(shape, input_type, name='input')
builder.start_new_path()
builder.add_reshape_layer((*shape, 1))
builder.add_conv2d_layer(10,
3,
filter_strides=1,
activation='relu',
regularizer='L2',
padding='same')
encoder = builder.add_conv2d_layer(10,
3,
filter_strides=1,
activation='relu',
regularizer='L2',
padding='same')
builder.add_conv2d_layer(1,
3,
filter_strides=1,
activation='relu',
regularizer='L2',
padding='same')
decoder = builder.add_reshape_layer(shape)
builder.end_current_path()
builder.finalize(decoder,
name='target',
learning_rate=lr,
optimizer=optimizer,
loss=loss_fn)
super(self.__class__, self).__init__(builder, encoder)
self.input_type = input_type
self.input_shape = input_shapes[input_type]
self.in_name, self.out_name = in_name, decoder
def __init__(self, input_shapes, **optsettings):
lr = optsettings.get('learning_rate') or 0.001
optimizer = optsettings.get('optimizer') or 'adam'
loss_fn = optsettings.get('loss_fn') or 'categorical_crossentropy'
builder = builders.GraphBuilder()
shape = input_shapes['yx']
in_yx = builder.add_input_layer(shape, name='input_yx')
builder.start_new_path()
builder.add_reshape_layer((*shape, 1))
builder.add_conv2d_layer(32, 3, filter_strides=1, activation='relu',
regularizer='L2', padding='same')
builder.add_maxpool2d_layer(2)
builder.add_lrn_layer()
builder.add_conv2d_layer(64, 3, filter_strides=1, activation='relu',
regularizer='L2', padding='same')
builder.add_maxpool2d_layer(2)
builder.add_lrn_layer()
yx_out = builder.add_flatten_layer()
builder.end_current_path()
shape = input_shapes['gtux']
in_gtux = builder.add_input_layer(shape, name='input_gtux')
builder.start_new_path()
builder.add_reshape_layer((*shape, 1))
builder.add_conv2d_layer(32, 3, filter_strides=1, activation='relu',
regularizer='L2', padding='same')
builder.add_maxpool2d_layer(2)
builder.add_lrn_layer()
builder.add_conv2d_layer(64, 3, filter_strides=1, activation='relu',
regularizer='L2', padding='same')
builder.add_maxpool2d_layer(2)
builder.add_lrn_layer()
gtux_out = builder.add_flatten_layer()
builder.end_current_path()
shape = input_shapes['gtuy']
in_gtuy = builder.add_input_layer(shape, name='input_gtuy')
builder.start_new_path()
builder.add_reshape_layer((*shape, 1))
builder.add_conv2d_layer(32, 3, filter_strides=1, activation='relu',
regularizer='L2', padding='same')
builder.add_maxpool2d_layer(2)
builder.add_lrn_layer()
builder.add_conv2d_layer(64, 3, filter_strides=1, activation='relu',
regularizer='L2', padding='same')
builder.add_maxpool2d_layer(2)
builder.add_lrn_layer()
gtuy_out = builder.add_flatten_layer()
builder.end_current_path()
builder.add_merge_layer((yx_out, gtux_out, gtuy_out),
'concat')
builder.start_new_path()
builder.add_fc_layer(128, activation='relu')
builder.add_dropout_layer(0.5)
builder.add_fc_layer(50, activation='relu')
builder.add_dropout_layer(0.5)
out_name = builder.add_fc_layer(2, activation='softmax')
builder.end_current_path()
builder.finalize(out_name, name='target', learning_rate=lr,
optimizer=optimizer, loss=loss_fn)
super(self.__class__, self).__init__(builder)
self.input_shapes = input_shapes.copy()
self.in_yx, self.in_gtux, self.in_gtuy = in_yx, in_gtux, in_gtuy
self.out_name = out_name
