Thesis/pytorch-dnc
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

FLANN read

Browse Source
This commit is contained in:
ixaxaar 2017-11-27 13:51:17 +05:30
parent c4ec88a58e
commit 8aee625101
2 changed files with 43 additions and 167 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

3
dnc/__init__.py
View File

@ -1,3 +1,6 @@
#!/usr/bin/env python3 #!/usr/bin/env python3
from .dnc import DNC from .dnc import DNC
from .sdnc import SDNC
from .memory import Memory
from .sparse_memory import SparseMemory

181
dnc/sparse_memory.py
View File

@ -10,6 +10,7 @@ from pyflann import FLANN
from .util import * from .util import *
class SparseMemory(nn.Module): class SparseMemory(nn.Module):
def __init__( def __init__(
@ -25,7 +26,7 @@ class SparseMemory(nn.Module):
index_checks=32, index_checks=32,
rebuild_indexes_after=10 rebuild_indexes_after=10
): ):
super(Memory, self).__init__() super(SparseMemory, self).__init__()
self.mem_size = mem_size self.mem_size = mem_size
self.cell_size = cell_size self.cell_size = cell_size
@ -33,7 +34,7 @@ class SparseMemory(nn.Module):
self.gpu_id = gpu_id self.gpu_id = gpu_id
self.input_size = input_size self.input_size = input_size
self.independent_linears = independent_linears self.independent_linears = independent_linears
self.K = sparse_reads self.K = sparse_reads if self.mem_size > sparse_reads else self.mem_size
self.num_kdtrees = num_kdtrees self.num_kdtrees = num_kdtrees
self.index_checks = index_checks self.index_checks = index_checks
self.rebuild_indexes_after = rebuild_indexes_after self.rebuild_indexes_after = rebuild_indexes_after
@ -46,17 +47,11 @@ class SparseMemory(nn.Module):
if self.independent_linears: if self.independent_linears:
self.read_keys_transform = nn.Linear(self.input_size, w * r) self.read_keys_transform = nn.Linear(self.input_size, w * r)
self.read_strengths_transform = nn.Linear(self.input_size, r)
self.write_key_transform = nn.Linear(self.input_size, w) self.write_key_transform = nn.Linear(self.input_size, w)
self.write_strength_transform = nn.Linear(self.input_size, 1)
self.erase_vector_transform = nn.Linear(self.input_size, w)
self.write_vector_transform = nn.Linear(self.input_size, w) self.write_vector_transform = nn.Linear(self.input_size, w)
self.free_gates_transform = nn.Linear(self.input_size, r)
self.allocation_gate_transform = nn.Linear(self.input_size, 1)
self.write_gate_transform = nn.Linear(self.input_size, 1) self.write_gate_transform = nn.Linear(self.input_size, 1)
self.read_modes_transform = nn.Linear(self.input_size, 3 * r)
else: else:
self.interface_size = (w * r) + (3 * w) + (5 * r) + 3 self.interface_size = (w * r) + (2 * w) + 1
self.interface_weights = nn.Linear(self.input_size, self.interface_size) self.interface_weights = nn.Linear(self.input_size, self.interface_size)
self.I = cuda(1 - T.eye(m).unsqueeze(0), gpu_id=self.gpu_id) # (1 * n * n) self.I = cuda(1 - T.eye(m).unsqueeze(0), gpu_id=self.gpu_id) # (1 * n * n)
@ -67,10 +62,10 @@ class SparseMemory(nn.Module):
if self.rebuild_indexes_after == self.index_reset_ctr or 'dict' not in hidden: if self.rebuild_indexes_after == self.index_reset_ctr or 'dict' not in hidden:
self.index_reset_ctr = 0 self.index_reset_ctr = 0
hidden['dict'] = [ FLANN() for x in range(b) ] hidden['dict'] = [FLANN() for x in range(b)]
hidden['dict'] = [ \ [
x.build_index(hidden['sparse'][n], algorithm='kdtree', trees=self.num_kdtrees, checks=self.index_checks) x.build_index(hidden['sparse'][n], algorithm='kdtree', trees=self.num_kdtrees, checks=self.index_checks)
for n,x in enumerate(hidden['dict']) for n, x in enumerate(hidden['dict'])
] ]
self.index_reset_ctr += 1 self.index_reset_ctr += 1
return hidden return hidden
@ -82,156 +77,50 @@ class SparseMemory(nn.Module):
b = batch_size b = batch_size
if hidden is None: if hidden is None:
hx = { hidden = {
# warning can be a huge chunk of contiguous memory # warning can be a huge chunk of contiguous memory
'sparse': np.zeros((b, m, w)), 'sparse': np.zeros((b, m, w), dtype=np.float32),
# 'memory': cuda(T.zeros(b, m, w).fill_(δ), gpu_id=self.gpu_id),
'link_matrix': cuda(T.zeros(b, 1, m, m), gpu_id=self.gpu_id),
'precedence': cuda(T.zeros(b, 1, m), gpu_id=self.gpu_id),
'read_weights': cuda(T.zeros(b, r, m).fill_(δ), gpu_id=self.gpu_id), 'read_weights': cuda(T.zeros(b, r, m).fill_(δ), gpu_id=self.gpu_id),
'write_weights': cuda(T.zeros(b, 1, m).fill_(δ), gpu_id=self.gpu_id), 'write_weights': cuda(T.zeros(b, 1, m).fill_(δ), gpu_id=self.gpu_id)
'usage_vector': cuda(T.zeros(b, m), gpu_id=self.gpu_id)
} }
# Build FLANN randomized k-d tree indexes for each batch # Build FLANN randomized k-d tree indexes for each batch
hx = rebuild_indexes(hx) hidden = self.rebuild_indexes(hidden)
else: else:
# hidden['memory'] = hidden['memory'].clone() # hidden['memory'] = hidden['memory'].clone()
hidden['link_matrix'] = hidden['link_matrix'].clone()
hidden['precedence'] = hidden['precedence'].clone()
hidden['read_weights'] = hidden['read_weights'].clone() hidden['read_weights'] = hidden['read_weights'].clone()
hidden['write_weights'] = hidden['write_weights'].clone() hidden['write_weights'] = hidden['write_weights'].clone()
hidden['usage_vector'] = hidden['usage_vector'].clone()
if erase: if erase:
hidden = self.rebuild_indexes(hidden) hidden = self.rebuild_indexes(hidden)
hidden['sparse'].fill(0) hidden['sparse'].fill(0)
# hidden['memory'].data.fill_(δ) # hidden['memory'].data.fill_(δ)
hidden['link_matrix'].data.zero_()
hidden['precedence'].data.zero_()
hidden['read_weights'].data.fill_(δ) hidden['read_weights'].data.fill_(δ)
hidden['write_weights'].data.fill_(δ) hidden['write_weights'].data.fill_(δ)
hidden['usage_vector'].data.zero_()
return hidden return hidden
def get_usage_vector(self, usage, free_gates, read_weights, write_weights):
# write_weights = write_weights.detach() # detach from the computation graph
usage = usage + (1 - usage) * (1 - T.prod(1 - write_weights, 1))
ψ = T.prod(1 - free_gates.unsqueeze(2) * read_weights, 1)
return usage * ψ
def allocate(self, usage, write_gate):
# ensure values are not too small prior to cumprod.
usage = δ + (1 - δ) * usage
batch_size = usage.size(0)
# free list
sorted_usage, φ = T.topk(usage, self.mem_size, dim=1, largest=False)
# cumprod with exclusive=True, TODO: unstable territory, revisit this shit
# essential for correct scaling of allocation_weights to prevent memory pollution
# during write operations
# https://discuss.pytorch.org/t/cumprod-exclusive-true-equivalences/2614/8
v = var(T.ones(batch_size, 1))
if self.gpu_id != -1:
v = v.cuda(self.gpu_id)
cat_sorted_usage = T.cat((v, sorted_usage), 1)[:, :-1]
prod_sorted_usage = fake_cumprod(cat_sorted_usage, self.gpu_id)
sorted_allocation_weights = (1 - sorted_usage) * prod_sorted_usage.squeeze()
# construct the reverse sorting index https://stackoverflow.com/questions/2483696/undo-or-reverse-argsort-python
_, φ_rev = T.topk(φ, k=self.mem_size, dim=1, largest=False)
allocation_weights = sorted_allocation_weights.gather(1, φ.long())
# update usage after allocating
# usage += ((1 - usage) * write_gate * allocation_weights)
return allocation_weights.unsqueeze(1), usage
def write_weighting(self, write_content_weights, allocation_weights, write_gate, allocation_gate):
ag = allocation_gate.unsqueeze(-1)
wg = write_gate.unsqueeze(-1)
return wg * (ag * allocation_weights + (1 - ag) * write_content_weights)
def get_link_matrix(self, link_matrix, write_weights, precedence):
precedence = precedence.unsqueeze(2)
write_weights_i = write_weights.unsqueeze(3)
write_weights_j = write_weights.unsqueeze(2)
prev_scale = 1 - write_weights_i - write_weights_j
new_link_matrix = write_weights_i * precedence
link_matrix = prev_scale * link_matrix + new_link_matrix
# elaborate trick to delete diag elems
return self.I.expand_as(link_matrix) * link_matrix
def update_precedence(self, precedence, write_weights):
return (1 - T.sum(write_weights, 2, keepdim=True)) * precedence + write_weights
def write(self, write_key, write_vector, write_gate, hidden): def write(self, write_key, write_vector, write_gate, hidden):
write_weights = write_gate * ( \ # write_weights = write_gate * ( \
interpolation_gate * hidden['read_weights'] + \ # interpolation_gate * hidden['read_weights'] + \
(1 - interpolation_gate)*cuda(T.ones(hidden['read_weights'].size()), gpu_id=self.gpu_id) ) # (1 - interpolation_gate)*cuda(T.ones(hidden['read_weights'].size()), gpu_id=self.gpu_id) )
# write_weights * write_vector
# get current usage
hidden['usage_vector'] = self.get_usage_vector(
hidden['usage_vector'],
free_gates,
hidden['read_weights'],
hidden['write_weights']
)
# lookup memory with write_key and write_strength
write_content_weights = self.content_weightings(hidden['memory'], write_key, write_strength)
# get memory allocation
alloc, _ = self.allocate(
hidden['usage_vector'],
allocation_gate * write_gate
)
# get write weightings
hidden['write_weights'] = self.write_weighting(
write_content_weights,
alloc,
write_gate,
allocation_gate
)
weighted_resets = hidden['write_weights'].unsqueeze(3) * erase_vector.unsqueeze(2)
reset_gate = T.prod(1 - weighted_resets, 1)
# Update memory
hidden['memory'] = hidden['memory'] * reset_gate
hidden['memory'] = hidden['memory'] + \
T.bmm(hidden['write_weights'].transpose(1, 2), write_vector)
# update link_matrix
hidden['link_matrix'] = self.get_link_matrix(
hidden['link_matrix'],
hidden['write_weights'],
hidden['precedence']
)
hidden['precedence'] = self.update_precedence(hidden['precedence'], hidden['write_weights'])
return hidden return hidden
def read_from_sparse_memory(self, sparse, dict, keys): def read_from_sparse_memory(self, sparse, dict, keys):
ks = keys.data.cpu().numpy() keys = keys.data.cpu().numpy()
read_vectors = [] read_vectors = []
positions = [] positions = []
read_weights = [] read_weights = []
# search nearest neighbor for each key # search nearest neighbor for each key
for k in range(ks.shape[1]): for key in range(keys.shape[1]):
print(key, keys.shape)
# search for K nearest neighbours given key for each batch # search for K nearest neighbours given key for each batch
search = [ h.nn_index(k[n], num_neighbours=self.K) for n,h in enumerate(dict) ] search = [h.nn_index(keys[b, key, :], num_neighbors=self.K) for b, h in enumerate(dict)]
distances = [ m[1] for m in search ] distances = [m[1] for m in search]
v = [ cudavec(sparse[m[0]], gpu_id=self.gpu_id) for m in search ] v = [cudavec(sparse[m[0]], gpu_id=self.gpu_id) for m in search]
v = v v = v
p = [ m[0] for m in search ] p = [m[0] for m in search]
read_vectors.append(T.stack(v, 0).contiguous()) read_vectors.append(T.stack(v, 0).contiguous())
positions.append(p) positions.append(p)
@ -244,7 +133,8 @@ class SparseMemory(nn.Module):
def read(self, read_keys, hidden): def read(self, read_keys, hidden):
# sparse read # sparse read
read_vectors, positions, read_weights = self.read_from_sparse_memory(hidden['sparse'], hidden['dict'], read_keys) read_vectors, positions, read_weights = \
self.read_from_sparse_memory(hidden['sparse'], hidden['dict'], read_keys)
hidden['read_positions'] = positions hidden['read_positions'] = positions
hidden['read_weights'] = read_weights hidden['read_weights'] = read_weights
@ -263,16 +153,8 @@ class SparseMemory(nn.Module):
read_keys = self.read_keys_transform(ξ).view(b, r, w) read_keys = self.read_keys_transform(ξ).view(b, r, w)
# write key (b * 1 * w) # write key (b * 1 * w)
write_key = self.write_key_transform(ξ).view(b, 1, w) write_key = self.write_key_transform(ξ).view(b, 1, w)
# write strength (b * 1)
write_strength = self.write_strength_transform(ξ).view(b, 1)
# erase vector (b * 1 * w)
erase_vector = F.sigmoid(self.erase_vector_transform(ξ).view(b, 1, w))
# write vector (b * 1 * w) # write vector (b * 1 * w)
write_vector = self.write_vector_transform(ξ).view(b, 1, w) write_vector = self.write_vector_transform(ξ).view(b, 1, w)
# r free gates (b * r)
free_gates = F.sigmoid(self.free_gates_transform(ξ).view(b, r))
# allocation gate (b * 1)
allocation_gate = F.sigmoid(self.allocation_gate_transform(ξ).view(b, 1))
# write gate (b * 1) # write gate (b * 1)
write_gate = F.sigmoid(self.write_gate_transform(ξ).view(b, 1)) write_gate = F.sigmoid(self.write_gate_transform(ξ).view(b, 1))
else: else:
@ -280,20 +162,11 @@ class SparseMemory(nn.Module):
# r read keys (b * w * r) # r read keys (b * w * r)
read_keys = ξ[:, :r * w].contiguous().view(b, r, w) read_keys = ξ[:, :r * w].contiguous().view(b, r, w)
# write key (b * w * 1) # write key (b * w * 1)
write_key = ξ[:, r * w + r:r * w + r + w].contiguous().view(b, 1, w) write_key = ξ[:, r * w:r * w + w].contiguous().view(b, 1, w)
# write strength (b * 1)
write_strength = 1 + F.relu(ξ[:, r * w + r + w].contiguous()).view(b, 1)
# erase vector (b * w)
erase_vector = F.sigmoid(ξ[:, r * w + r + w + 1: r * w + r + 2 * w + 1].contiguous().view(b, 1, w))
# write vector (b * w) # write vector (b * w)
write_vector = ξ[:, r * w + r + 2 * w + 1: r * w + r + 3 * w + 1].contiguous().view(b, 1, w) write_vector = ξ[:, r * w + w: r * w + 2 * w].contiguous().view(b, 1, w)
# r free gates (b * r)
free_gates = F.sigmoid(ξ[:, r * w + r + 3 * w + 1: r * w + 2 * r + 3 * w + 1].contiguous().view(b, r))
# allocation gate (b * 1)
allocation_gate = F.sigmoid(ξ[:, r * w + 2 * r + 3 * w + 1].contiguous().unsqueeze(1).view(b, 1))
# write gate (b * 1) # write gate (b * 1)
write_gate = F.sigmoid(ξ[:, r * w + 2 * r + 3 * w + 2].contiguous()).unsqueeze(1).view(b, 1) write_gate = F.sigmoid(ξ[:, -1].contiguous()).unsqueeze(1).view(b, 1)
hidden = self.write(write_key, write_vector, erase_vector, free_gates, hidden = self.write(write_key, write_vector, write_gate, hidden)
read_strengths, write_strength, write_gate, allocation_gate, hidden)
return self.read(read_keys, hidden) return self.read(read_keys, hidden)