def init_adadelta_extra_parameters(algo, state): algo.large_W_0_enc_approx_embdr_g2 = sample_zeros( algo.state['large_vocab_source'], algo.state['rank_n_approx'], -1, algo.state['weight_scale'], algo.rng) algo.large_W_0_enc_approx_embdr_d2 = sample_zeros( algo.state['large_vocab_source'], algo.state['rank_n_approx'], -1, algo.state['weight_scale'], algo.rng) algo.large_W_0_dec_approx_embdr_g2 = sample_zeros( algo.state['large_vocab_target'], algo.state['rank_n_approx'], -1, algo.state['weight_scale'], algo.rng) algo.large_W_0_dec_approx_embdr_d2 = sample_zeros( algo.state['large_vocab_target'], algo.state['rank_n_approx'], -1, algo.state['weight_scale'], algo.rng) algo.large_W2_dec_deep_softmax_g2 = sample_zeros( algo.state['rank_n_approx'], algo.state['large_vocab_target'], -1, algo.state['weight_scale'], algo.rng) algo.large_W2_dec_deep_softmax_d2 = sample_zeros( algo.state['rank_n_approx'], algo.state['large_vocab_target'], -1, algo.state['weight_scale'], algo.rng) algo.large_b_dec_deep_softmax_g2 = init_bias( algo.state['large_vocab_target'], 0., algo.rng) algo.large_b_dec_deep_softmax_d2 = init_bias( algo.state['large_vocab_target'], 0., algo.rng) if state['save_gs']: algo.large_W_0_enc_approx_embdr_gs = sample_zeros( algo.state['large_vocab_source'], algo.state['rank_n_approx'], -1, algo.state['weight_scale'], algo.rng) algo.large_W_0_dec_approx_embdr_gs = sample_zeros( algo.state['large_vocab_target'], algo.state['rank_n_approx'], -1, algo.state['weight_scale'], algo.rng) algo.large_W2_dec_deep_softmax_gs = sample_zeros( algo.state['rank_n_approx'], algo.state['large_vocab_target'], -1, algo.state['weight_scale'], algo.rng) algo.large_b_dec_deep_softmax_gs = init_bias( algo.state['large_vocab_target'], 0., algo.rng)
def init_adadelta_extra_parameters(algo, state): algo.large_W_0_enc_approx_embdr_g2 = sample_zeros(algo.state['large_vocab_source'], algo.state['rank_n_approx'], -1, algo.state['weight_scale'], algo.rng) algo.large_W_0_enc_approx_embdr_d2 = sample_zeros(algo.state['large_vocab_source'], algo.state['rank_n_approx'], -1, algo.state['weight_scale'], algo.rng) algo.large_W_0_dec_approx_embdr_g2 = sample_zeros(algo.state['large_vocab_target'], algo.state['rank_n_approx'], -1, algo.state['weight_scale'], algo.rng) algo.large_W_0_dec_approx_embdr_d2 = sample_zeros(algo.state['large_vocab_target'], algo.state['rank_n_approx'], -1, algo.state['weight_scale'], algo.rng) algo.large_W2_dec_deep_softmax_g2 = sample_zeros(algo.state['rank_n_approx'], algo.state['large_vocab_target'], -1, algo.state['weight_scale'], algo.rng) algo.large_W2_dec_deep_softmax_d2 = sample_zeros(algo.state['rank_n_approx'], algo.state['large_vocab_target'], -1, algo.state['weight_scale'], algo.rng) algo.large_b_dec_deep_softmax_g2 = init_bias(algo.state['large_vocab_target'], 0., algo.rng) algo.large_b_dec_deep_softmax_d2 = init_bias(algo.state['large_vocab_target'], 0., algo.rng) if state['save_gs']: algo.large_W_0_enc_approx_embdr_gs = sample_zeros(algo.state['large_vocab_source'], algo.state['rank_n_approx'], -1, algo.state['weight_scale'], algo.rng) algo.large_W_0_dec_approx_embdr_gs = sample_zeros(algo.state['large_vocab_target'], algo.state['rank_n_approx'], -1, algo.state['weight_scale'], algo.rng) algo.large_W2_dec_deep_softmax_gs = sample_zeros(algo.state['rank_n_approx'], algo.state['large_vocab_target'], -1, algo.state['weight_scale'], algo.rng) algo.large_b_dec_deep_softmax_gs = init_bias(algo.state['large_vocab_target'], 0., algo.rng)
def init_extra_parameters(model, state): # May want to add skip_init later model.large_W_0_enc_approx_embdr = eval(state['weight_init_fn'])( state['large_vocab_source'], state['rank_n_approx'], -1, state['weight_scale'], model.rng) model.large_W_0_dec_approx_embdr = eval(state['weight_init_fn'])( state['large_vocab_target'], state['rank_n_approx'], -1, state['weight_scale'], model.rng) model.large_W2_dec_deep_softmax = eval(state['weight_init_fn'])( state['rank_n_approx'], state['large_vocab_target'], -1, state['weight_scale'], model.rng) model.large_b_dec_deep_softmax = init_bias(state['large_vocab_target'], 0., model.rng)
def init_extra_parameters(model, state): # May want to add skip_init later model.large_W_0_enc_approx_embdr = eval(state['weight_init_fn'])(state['large_vocab_source'], state['rank_n_approx'], -1, state['weight_scale'], model.rng) model.large_W_0_dec_approx_embdr = eval(state['weight_init_fn'])(state['large_vocab_target'], state['rank_n_approx'], -1, state['weight_scale'], model.rng) model.large_W2_dec_deep_softmax = eval(state['weight_init_fn'])(state['rank_n_approx'], state['large_vocab_target'], -1, state['weight_scale'], model.rng) model.large_b_dec_deep_softmax = init_bias(state['large_vocab_target'], 0., model.rng)