def __init__(self, config, is_training=True):
super(BiLSTM, self).__init__()
self.is_training = is_training
self.batch_size = config.batch_size * config.rnn_batch_size
print("batch size is {} ".format(self.batch_size))
self.input_size = config.input_size
self.hidden_size = config.hidden_size
self.num_step = config.num_step
self.reshape = P.Reshape()
self.cast = P.Cast()
k = (1 / self.hidden_size) ** 0.5
self.rnn1 = P.DynamicRNN(forget_bias=0.0)
self.rnn_bw = P.DynamicRNN(forget_bias=0.0)
self.w1 = Parameter(np.random.uniform(-k, k, \
(self.input_size + self.hidden_size, 4 * self.hidden_size)).astype(np.float32), name="w1")
self.w1_bw = Parameter(np.random.uniform(-k, k, \
(self.input_size + self.hidden_size, 4 * self.hidden_size)).astype(np.float32), name="w1_bw")
self.b1 = Parameter(np.random.uniform(-k, k, (4 * self.hidden_size)).astype(np.float32), name="b1")
self.b1_bw = Parameter(np.random.uniform(-k, k, (4 * self.hidden_size)).astype(np.float32), name="b1_bw")
self.h1 = Tensor(np.zeros(shape=(1, self.batch_size, self.hidden_size)).astype(np.float32))
self.h1_bw = Tensor(np.zeros(shape=(1, self.batch_size, self.hidden_size)).astype(np.float32))
self.c1 = Tensor(np.zeros(shape=(1, self.batch_size, self.hidden_size)).astype(np.float32))
self.c1_bw = Tensor(np.zeros(shape=(1, self.batch_size, self.hidden_size)).astype(np.float32))
self.reverse_seq = P.ReverseV2(axis=[0])
self.concat = P.Concat()
self.transpose = P.Transpose()
self.concat1 = P.Concat(axis=2)
self.dropout = nn.Dropout(0.7)
self.use_dropout = config.use_dropout
self.reshape = P.Reshape()
self.transpose = P.Transpose()
def __init__(self,
config: GNMTConfig,
is_training: bool,
compute_type=mstype.float32):
super(GNMTEncoder, self).__init__()
self.input_mask_from_dataset = config.input_mask_from_dataset
self.max_positions = config.seq_length
self.attn_embed_dim = config.hidden_size
self.num_layers = config.num_hidden_layers
self.hidden_dropout_prob = config.hidden_dropout_prob
self.vocab_size = config.vocab_size
self.seq_length = config.seq_length
self.batch_size = config.batch_size
self.word_embed_dim = config.hidden_size
self.transpose = P.Transpose()
self.transpose_orders = (1, 0, 2)
self.reshape = P.Reshape()
self.concat = P.Concat(axis=-1)
encoder_layers = []
for i in range(0, self.num_layers + 1):
if i == 2:
# the bidirectional layer's output is [T,D,2N]
scaler = 2
else:
# the rest layer's output is [T,D,N]
scaler = 1
layer = DynamicRNNNet(seq_length=self.seq_length,
batchsize=self.batch_size,
word_embed_dim=scaler * self.word_embed_dim,
hidden_size=self.word_embed_dim)
encoder_layers.append(layer)
self.encoder_layers = nn.CellList(encoder_layers)
self.reverse_v2 = P.ReverseV2(axis=[0])
self.dropout = nn.Dropout(keep_prob=1.0 - config.hidden_dropout_prob)
def __init__(self, kernel_size=1, stride=1, pad_mode="valid"):
super(MaxPool2dPytorch, self).__init__()
self.maxpool = nn.MaxPool2d(kernel_size=kernel_size,
stride=stride,
pad_mode=pad_mode)
self.reverse = F.ReverseV2(axis=[2, 3])
def __init__(self, bidirectional=False):
super(LSTM_Ascend, self).__init__()
self.bidirectional = bidirectional
self.dynamic_rnn = P.DynamicRNN(forget_bias=0.0)
self.reverseV2 = P.ReverseV2(axis=[0])
self.concat = P.Concat(2)