init_h3.tag.test_value = np_zeros((minibatch_size, n_hid))
init_kappa = tensor.matrix("init_kappa")
init_kappa.tag.test_value = np_zeros((minibatch_size, att_size))
init_w = tensor.matrix("init_w")
init_w.tag.test_value = np_zeros((minibatch_size, n_chars))
params = []
biases = []
cell1 = GRU(input_dim, n_hid, random_state)
cell2 = GRU(n_hid, n_hid, random_state)
cell3 = GRU(n_hid, n_hid, random_state)
params += cell1.get_params()
params += cell2.get_params()
params += cell3.get_params()
inp_to_h1 = GRUFork(input_dim, n_hid, random_state)
inp_to_h2 = GRUFork(input_dim, n_hid, random_state)
inp_to_h3 = GRUFork(input_dim, n_hid, random_state)
att_to_h1 = GRUFork(n_chars, n_hid, random_state)
att_to_h2 = GRUFork(n_chars, n_hid, random_state)
att_to_h3 = GRUFork(n_chars, n_hid, random_state)
h1_to_h2 = GRUFork(n_hid, n_hid, random_state)
h1_to_h3 = GRUFork(n_hid, n_hid, random_state)
h2_to_h3 = GRUFork(n_hid, n_hid, random_state)
params += inp_to_h1.get_params()
params += inp_to_h2.get_params()
init_h3.tag.test_value = np_zeros((minibatch_size, n_hid))
init_kappa = tensor.matrix("init_kappa")
init_kappa.tag.test_value = np_zeros((minibatch_size, att_size))
init_w = tensor.matrix("init_w")
init_w.tag.test_value = np_zeros((minibatch_size, n_chars))
params = []
biases = []
cell1 = GRU(input_dim, n_hid, random_state)
cell2 = GRU(n_hid, n_hid, random_state)
cell3 = GRU(n_hid, n_hid, random_state)
params += cell1.get_params()
params += cell2.get_params()
params += cell3.get_params()
inp_to_h1 = GRUFork(input_dim, n_hid, random_state)
inp_to_h2 = GRUFork(input_dim, n_hid, random_state)
inp_to_h3 = GRUFork(input_dim, n_hid, random_state)
att_to_h1 = GRUFork(n_chars, n_hid, random_state)
att_to_h2 = GRUFork(n_chars, n_hid, random_state)
att_to_h3 = GRUFork(n_chars, n_hid, random_state)
h1_to_h2 = GRUFork(n_hid, n_hid, random_state)
h1_to_h3 = GRUFork(n_hid, n_hid, random_state)
h2_to_h3 = GRUFork(n_hid, n_hid, random_state)
params += inp_to_h1.get_params()
params += inp_to_h2.get_params()
random_state)
b_conv1, = make_biases((n_kernels,))
w_conv2, = make_conv_weights(n_kernels, (n_kernels,),
(conv_size2, 1), random_state)
b_conv2, = make_biases((n_kernels,))
params += [w_conv1, b_conv1, w_conv2, b_conv2]
# Use GRU classes only to fork 1 inp to 2 inp:gate pairs
conv_to_h1 = GRUFork(n_kernels, n_hid, random_state)
conv_to_h2 = GRUFork(n_kernels, n_hid, random_state)
params += conv_to_h1.get_params()
params += conv_to_h2.get_params()
cell1 = GRU(n_kernels, n_hid, random_state)
cell2 = GRU(n_hid, n_hid, random_state)
params += cell1.get_params()
params += cell2.get_params()
# Use GRU classes only to fork 1 inp to 2 inp:gate pairs
att_to_h1 = GRUFork(n_chars, n_hid, random_state)
att_to_h2 = GRUFork(n_chars, n_hid, random_state)
h1_to_h2 = GRUFork(n_hid, n_hid, random_state)
params += att_to_h1.get_params()
params += att_to_h2.get_params()
params += h1_to_h2.get_params()
h1_to_att_a, h1_to_att_b, h1_to_att_k = make_weights(n_hid, 3 * [att_size],
random_state)
params += [h1_to_att_a, h1_to_att_b, h1_to_att_k]
random_state)
b_conv1, = make_biases((n_kernels, ))
w_conv2, = make_conv_weights(n_kernels, (n_kernels, ), (conv_size2, 1),
random_state)
b_conv2, = make_biases((n_kernels, ))
params += [w_conv1, b_conv1, w_conv2, b_conv2]
# Use GRU classes only to fork 1 inp to 2 inp:gate pairs
conv_to_h1 = GRUFork(n_kernels, n_hid, random_state)
conv_to_h2 = GRUFork(n_kernels, n_hid, random_state)
params += conv_to_h1.get_params()
params += conv_to_h2.get_params()
cell1 = GRU(n_kernels, n_hid, random_state)
cell2 = GRU(n_hid, n_hid, random_state)
params += cell1.get_params()
params += cell2.get_params()
# Use GRU classes only to fork 1 inp to 2 inp:gate pairs
att_to_h1 = GRUFork(n_chars, n_hid, random_state)
att_to_h2 = GRUFork(n_chars, n_hid, random_state)
h1_to_h2 = GRUFork(n_hid, n_hid, random_state)
params += att_to_h1.get_params()
params += att_to_h2.get_params()
params += h1_to_h2.get_params()
h1_to_att_a, h1_to_att_b, h1_to_att_k = make_weights(
n_hid, 3 * [att_size], random_state)
params += [h1_to_att_a, h1_to_att_b, h1_to_att_k]