def initialise_one(name, verbose=True):
"""
Initialise data structures with no extra variables
"""
initialise(dims=1, param_name=name, sparse=False)
data, rel, known, A, B, T, embed = load(name)
W = (A,B,T)
options = [-1, -0.5, -0.2, -0.1, 0, 0.1, 0.2, 0.5, 1]
for _ in range(5):
for item, points in rel.items():
if not item in known:
old = embed[item][0]
minscore = float('inf')
best = None
for sent in options:
embed[item][0] = sent
new = objective(points, embed, W, reg_emb=0) + abs(sent)
if new < minscore:
minscore = new
best = sent
embed[item][0] = best
if verbose and best != old:
print(item, best)
print('\n...\n')
save(A,B,T,embed,name)
def initialise(dims=3,
lex_file = '../../data/sentimerge_nospin.pk',
morph_file = '../../data/morph/preproc_nopos.pk',
data_file = '../../data/morph/data.pk',
param_name = 'out',
sparse = True):
"""
Initialise data structures for learning
"""
with open(lex_file, 'rb') as f:
full_lex, full_lex_weight = pickle.load(f)
lex, weight = remove_pos(full_lex, full_lex_weight)
with open(morph_file, 'rb') as f:
full_morph = pickle.load(f)
morph = {x:y for x,y in full_morph.items() if x in lex and len(y)>1}
relevant = {}
for whole, parts in morph.items():
for atom in parts:
relevant.setdefault(atom, []).append((lex[whole], parts))
data = [(lex[whole], parts) for whole, parts in morph.items()]
items = sorted(relevant.keys()) #{z for y in morph.values() for z in y})
embed, known = featurise(items, lex, dims-1)
A = array([[random() for _ in range(dims)] for _ in range(dims)])
B = array([[random() for _ in range(dims)] for _ in range(dims)])
T = array([[[random() for _ in range(dims)] for _ in range(dims)] for _ in range(dims)])
A[0,0] = 0.4
B[0,0] = 0.7
T[0,0,0] = 0.4
# Make the initialisation sparse:
if sparse:
for arr in (A,B,T):
it = nditer(arr, flags=['multi_index'], op_flags=['readwrite'])
for value in it:
old_val = value.copy()
old_obj = objective(data, embed, (A,B,T))
value[...] = 0
new_obj = objective(data, embed, (A,B,T))
if old_obj < new_obj:
value[...] = old_val
print(it.multi_index)
for atom in items:
if atom in known:
r = range(1,dims)
else:
r = range(dims)
vec = embed[atom]
for i in r:
old_val = vec[i].copy()
old_obj = objective(relevant[atom], embed, (A,B,T), reg_emb=0)
old_obj += abs(old_val)
vec[i] = 0
new_obj = objective(relevant[atom], embed, (A,B,T), reg_emb=0)
if old_obj < new_obj:
vec[i] = old_val
print(atom, i)
print(A, end='\n\n\n')
print(B, end='\n\n\n')
print(T)
with open(data_file, 'wb') as f:
pickle.dump((data, relevant, known), f)
save(A, B, T, embed, param_name)
