def __init__(self,
num_lstm_layer,
input_size,
num_hidden,
num_embed,
num_label,
arg_params,
ctx=mx.cpu(),
dropout=0.):
self.sym = lstm_inference_symbol(num_lstm_layer,
input_size,
num_hidden,
num_embed,
num_label,
dropout)
batch_size = 1
init_c = [('l%d_init_c'%l, (batch_size, num_hidden)) for l in range(num_lstm_layer)]
init_h = [('l%d_init_h'%l, (batch_size, num_hidden)) for l in range(num_lstm_layer)]
data_shape = [("data", (batch_size,))]
input_shapes = dict(init_c + init_h + data_shape)
self.executor = self.sym.simple_bind(ctx=mx.cpu(), **input_shapes)
for key in self.executor.arg_dict.keys():
if key in arg_params:
arg_params[key].copyto(self.executor.arg_dict[key])
state_name = []
for i in range(num_lstm_layer):
state_name.append("l%d_init_c" % i)
state_name.append("l%d_init_h" % i)
self.states_dict = dict(zip(state_name, self.executor.outputs[1:]))
self.input_arr = mx.nd.zeros(data_shape[0][1])
def __init__(self,
num_lstm_layer,
seq_len,
num_hidden,
num_label,
arg_params,
data_size,
ctx=mx.cpu()):
self.sym = lstm_inference_symbol(num_lstm_layer, seq_len, num_hidden,
num_label)
batch_size = 1
init_c = [('l%d_init_c' % l, (batch_size, num_hidden))
for l in range(num_lstm_layer)]
init_h = [('l%d_init_h' % l, (batch_size, num_hidden))
for l in range(num_lstm_layer)]
data_shape = [("data", (batch_size, data_size))]
input_shapes = dict(init_c + init_h + data_shape)
self.executor = self.sym.simple_bind(ctx=ctx, **input_shapes)
for key in self.executor.arg_dict.keys():
if key in arg_params:
arg_params[key].copyto(self.executor.arg_dict[key])
state_name = []
for i in range(num_lstm_layer):
state_name.append("l%d_init_c" % i)
state_name.append("l%d_init_h" % i)
self.states_dict = dict(zip(state_name, self.executor.outputs[1:]))
self.input_arr = mx.nd.zeros(data_shape[0][1])
def __init__(self,
num_lstm_layer,
input_size,
num_hidden,
num_embed,
num_label,
arg_params,
ctx=mx.cpu(),
dropout=0.):
self.sym = lstm_inference_symbol(num_lstm_layer, input_size,
num_hidden, num_embed, num_label,
dropout)
batch_size = 1
init_c = [('l%d_init_c_beta' % l, (batch_size, num_hidden))
for l in range(num_lstm_layer)]
init_h = [('l%d_init_h_beta' % l, (batch_size, num_hidden))
for l in range(num_lstm_layer)]
data_shape = [('data', (batch_size, ))]
print(init_c)
input_shapes = dict(init_c + init_h + data_shape)
print input_shapes
self.executor = self.sym.simple_bind(ctx=ctx, **input_shapes)
for key in self.executor.arg_dict.keys():
if key not in input_shapes and key in arg_params and key != "softmax_label":
arg_params[key].copyto(self.executor.arg_dict[key])
state_name = []
for i in range(num_lstm_layer):
state_name.append("l%d_init_c_beta" % i)
state_name.append("l%d_init_h_beta" % i)
self.states_dict = dict(zip(state_name, self.executor.outputs[1:]))
self.input_arr = mx.nd.zeros(data_shape[0][1])
