def __init__(self, layer: Conv2D):
super().__init__(layer)
W, b = layer.get_weights()
i = Input(shape=layer.input_shape[1:])
u = Conv2D(
filters=layer.filters,
kernel_size=layer.kernel_size,
strides=layer.strides,
dilation_rate=layer.dilation_rate,
padding=layer.padding,
kernel_initializer=tf.constant_initializer(W),
bias_initializer=tf.constant_initializer(b),
)
self.up_func = K.function([i], [u(i)])
# Flip each filter vertically and horizontally
W_flipped = W[::-1, ::-1, :, :]
# Ensure the reconstructed image has correct number of channels
W_t = np.moveaxis(W_flipped, 2, 3)
b_t = np.zeros(W_t.shape[3])
o = Input(shape=layer.output_shape[1:])
d = Conv2D(
filters=W_t.shape[-1],
kernel_size=W_t.shape[:2],
strides=layer.strides,
dilation_rate=layer.dilation_rate,
padding=layer.padding,
kernel_initializer=tf.constant_initializer(W_t),
bias_initializer=tf.constant_initializer(b_t),
)
self.down_func = K.function([o], [d(o)])
def safe_init_conv2d(layer: layers.Conv2D):
weights = layer.get_weights()
w_kernel = scaled_constant(1, weights[0].shape)
_logger.info(f"Setting weights for layer `{layer.name}` :: "
f"{layer.__class__.__name__}{w_kernel.shape}")
w = [w_kernel]
if len(weights) > 1:
w_bias = np.zeros_like(weights[1])
w.append(w_bias)
layer.set_weights(w)
_check_activation(layer)
def get_deconv2d(conv2d_layer: Conv2D):
W = conv2d_layer.get_weights()[0]
# W: kernel_width, kernel_height, kernel_depth, n_filters
# Reverse the conv operation
W = np.transpose(W, (0, 1, 3, 2))
# Transpose the columns and rows
W = W[::-1, ::-1, :, :]
n_filters = W.shape[3]
kernel_size = W.shape[0]
strides = conv2d_layer.strides
padding = conv2d_layer.padding
b = np.zeros(n_filters)
return DeConv2D(n_filters,
kernel_size=kernel_size,
strides=strides,
padding=padding,
kernel_initializer=tf.constant_initializer(W),
bias_initializer=tf.constant_initializer(b),
trainable=False)
def get_deconv2d_reverse_bias(conv2d_layer: Conv2D):
return ReverseBiasLayer(conv2d_layer.get_weights()[1])
