def __init__(self, N, M):
"""Initialize the NTM Memory matrix.
The memory's dimensions are (batch_size x N x M).
Each batch has it's own memory matrix.
:param N: Number of rows in the memory.
:param M: Number of columns/features in the memory.
"""
super(NTMMemory, self).__init__()
self.N = N
self.M = M
# The memory bias allows the heads to learn how to initially address
# memory locations by content
self.register_buffer('mem_bias', torch.Tensor(N, M))
self.memory = None
self.init_mem = None
# Initialize memory bias
stdev = 1 / (np.sqrt(N + M))
nn.init.uniform_(self.mem_bias, -stdev, stdev)
conf = Config()
lower_bound, upper_bound = conf.output_range()
threshold = conf.get_threshold()
length = conf.get_encoding_length()
mem = []
for num in range(lower_bound, upper_bound + 1):
mem.append(enc(num, threshold, length))
self.memory = torch.stack(mem, dim=0).float()
self.init_mem = torch.stack(mem, dim=0).float()
def encode(inarr):
conf = Config()
result = []
for element in inarr:
result.append(
encoder(element, conf.get_threshold(), conf.get_encoding_length()))
return torch.stack(result, dim=0)
