def __init__(self,
name,
dim_in,
dim_c,
dim,
dim_class,
active=tensor.tanh,
maxout=2,
verbose=False):
'''
Initialize the parameters of the layer
'''
self.name = name
self.active = active
self.dim = dim
self.maxout = maxout
self.verbose = verbose
self.readout_emb = tools.init_weight((dim_in, dim),
name + '_readoutemb')
self.input_emb = tools.init_weight((dim_in, dim), name + '_inputemb')
self.gate_emb = tools.init_weight((dim_in, dim), name + '_gateemb')
self.reset_emb = tools.init_weight((dim_in, dim), name + '_resetemb')
self.readout_context = tools.init_weight((dim_c, dim),
name + '_readoutcontext')
self.input_context = tools.init_weight((dim_c, dim),
name + '_inputcontext')
self.gate_context = tools.init_weight((dim_c, dim),
name + '_gatecontext')
self.reset_context = tools.init_weight((dim_c, dim),
name + '_resetcontext')
self.readout_hidden = tools.init_weight((dim, dim),
name + '_readouthidden')
self.input_hidden = tools.init_weight((dim, dim),
name + '_inputhidden')
self.gate_hidden = tools.init_weight((dim, dim), name + '_gatehidden')
self.reset_hidden = tools.init_weight((dim, dim),
name + '_resethidden')
self.att_hidden = tools.init_weight((dim, dim), name + '_atthidden')
self.att_context = tools.init_weight((dim_c, dim),
name + '_attcontext')
self.att = tools.init_weight((dim, 1), name + '_att')
self.params = [
self.input_emb, self.gate_emb, self.reset_emb, self.input_context,
self.gate_context, self.reset_context, self.input_hidden,
self.gate_hidden, self.reset_hidden, self.readout_hidden,
self.readout_context, self.readout_emb, self.att_hidden,
self.att_context, self.att
]
self.probs_emb = tools.init_weight((dim / maxout, dim_in),
name + '_probsemb')
self.probs = tools.init_weight((dim_in, dim_class), name + '_probs')
self.params += [self.probs_emb, self.probs]
self.input_emb_offset = tools.init_bias((dim), name + '_inputoffset')
self.readout_offset = tools.init_bias((dim), name + '_readoutoffset')
self.probs_offset = tools.init_bias((dim_class), name + '_probsoffset')
self.params += [
self.input_emb_offset, self.readout_offset, self.probs_offset
]
def __init__(self, name, num, dim_embed, offset = True): ''' Initialize the parameters of the layer ''' self.name = name self.emb = tools.init_weight((num, dim_embed), name + '_emb') self.offset = offset self.params = [self.emb] if offset: self.b = tools.init_bias((dim_embed), name + '_b') self.params += [self.b]
def __init__(self, name, dim_in, dim_out, active = tensor.tanh, offset = False): ''' Initialize the parameters ''' self.name = name self.active = active self.offset = offset self.W = tools.init_weight((dim_in, dim_out), name + '_W') self.params = [self.W] if offset: self.b = tools.init_bias((dim_out), name + '_b') self.params += [self.b]
def __init__(self, name, dim_in, dim, active = tensor.tanh, verbose = False): self.name = name self.active = active self.dim = dim self.verbose = verbose self.input_emb = tools.init_weight((dim_in, dim), name + '_inputemb') self.gate_emb = tools.init_weight((dim_in, dim), name + '_gateemb') self.reset_emb = tools.init_weight((dim_in, dim), name + '_resetemb') self.input_hidden = tools.init_weight((dim, dim), name + '_inputhidden') self.gate_hidden = tools.init_weight((dim, dim), name + '_gatehidden') self.reset_hidden = tools.init_weight((dim, dim), name + '_resethidden') self.params = [self.input_emb, self.gate_emb, self.reset_emb, self.input_hidden, self.gate_hidden, self.reset_hidden] self.input_emb_offset = tools.init_bias((dim), name + '_inputoffset') self.params += [self.input_emb_offset]
def __init__(self, src_word_num, trg_word_num, config):
self.encoder_val = EncoderVal(src_word_num, config)
self.back_encoder_val = BackEncoderVal(src_word_num, config)
self.decoder_val = DecoderVal(src_word_num, trg_word_num, config)
self.name = 'GRU_enc'
self.src_word_num = src_word_num
self.trg_word_num = trg_word_num
dim_in = config['dim_emb_src']
self.dim_in = dim_in
dim = config['dim_rec_enc']
self.dim = dim
dim_c = config['dim_rec_enc'] * 2
self.dim_c = dim_c
dim_class = config['num_vocab_trg']
src_vocab_size = config['num_vocab_src']
trg_vocab_size = config['num_vocab_trg']
self.weight = 0.5
self.src_emb = tools.init_weight((src_vocab_size, dim_in),
'emb_src_emb')
self.src_emb_offset = tools.init_bias((dim_in), 'emb_src_b')
self.trg_emb = tools.init_weight((trg_vocab_size, dim_in),
'emb_trg_emb')
self.trg_emb_offset = tools.init_bias((dim_in), 'emb_trg_b')
self.input_emb = tools.init_weight((dim_in, dim),
self.name + '_inputemb')
self.gate_emb = tools.init_weight((dim_in, dim),
self.name + '_gateemb')
self.reset_emb = tools.init_weight((dim_in, dim),
self.name + '_resetemb')
self.input_hidden = tools.init_weight((dim, dim),
self.name + '_inputhidden')
self.gate_hidden = tools.init_weight((dim, dim),
self.name + '_gatehidden')
self.reset_hidden = tools.init_weight((dim, dim),
self.name + '_resethidden')
self.params = [
self.input_emb, self.gate_emb, self.reset_emb, self.input_hidden,
self.gate_hidden, self.reset_hidden
]
self.input_emb_offset = tools.init_bias((dim),
self.name + '_inputoffset')
self.params += [
self.input_emb_offset, self.src_emb, self.src_emb_offset
]
self.name = 'GRU_enc_back'
self.input_emb_back = tools.init_weight((dim_in, dim),
self.name + '_inputemb')
self.gate_emb_back = tools.init_weight((dim_in, dim),
self.name + '_gateemb')
self.reset_emb_back = tools.init_weight((dim_in, dim),
self.name + '_resetemb')
self.input_hidden_back = tools.init_weight((dim, dim),
self.name + '_inputhidden')
self.gate_hidden_back = tools.init_weight((dim, dim),
self.name + '_gatehidden')
self.reset_hidden_back = tools.init_weight((dim, dim),
self.name + '_resethidden')
self.params += [
self.input_emb_back, self.gate_emb_back, self.reset_emb_back,
self.input_hidden_back, self.gate_hidden_back,
self.reset_hidden_back
]
self.input_emb_offset_back = tools.init_bias(
(dim), self.name + '_inputoffset')
self.params += [self.input_emb_offset_back]
self.name = 'GRU_dec'
self.dec_readout_emb = tools.init_weight((dim_in, dim),
self.name + '_readoutemb')
self.dec_input_emb = tools.init_weight((dim_in, dim),
self.name + '_inputemb')
self.dec_gate_emb = tools.init_weight((dim_in, dim),
self.name + '_gateemb')
self.dec_reset_emb = tools.init_weight((dim_in, dim),
self.name + '_resetemb')
self.dec_readout_context = tools.init_weight(
(dim_c, dim), self.name + '_readoutcontext')
self.dec_input_context = tools.init_weight((dim_c, dim),
self.name + '_inputcontext')
self.dec_gate_context = tools.init_weight((dim_c, dim),
self.name + '_gatecontext')
self.dec_reset_context = tools.init_weight((dim_c, dim),
self.name + '_resetcontext')
self.dec_readout_hidden = tools.init_weight(
(dim, dim), self.name + '_readouthidden')
self.dec_input_hidden = tools.init_weight((dim, dim),
self.name + '_inputhidden')
self.dec_gate_hidden = tools.init_weight((dim, dim),
self.name + '_gatehidden')
self.dec_reset_hidden = tools.init_weight((dim, dim),
self.name + '_resethidden')
self.dec_input_emb_offset = tools.init_bias((dim),
self.name + '_inputoffset')
self.dec_probs_emb = tools.init_weight((dim / 2, dim_in),
self.name + '_probsemb')
self.dec_probs = tools.init_weight((dim_in, dim_class),
self.name + '_probs')
self.params += [
self.dec_readout_emb, self.dec_input_emb, self.dec_gate_emb,
self.dec_reset_emb, self.dec_readout_context,
self.dec_input_context, self.dec_gate_context,
self.dec_reset_context, self.dec_readout_hidden,
self.dec_input_hidden, self.dec_gate_hidden, self.dec_reset_hidden,
self.dec_input_emb_offset, self.dec_probs_emb, self.dec_probs
]
self.ep = numpy.float32(0.4)
self.ep = theano.shared(self.ep, name='ep')
self.idx = TT.iscalar()
self.R_c_x = TT.tensor3()