def rae_trainning_normal(wd_extra, w1, w2, neta, regu=None, wpresent=None):
err = 0.0
kerr = 0.0
samp_err = []
branch_samp_err = []
branchs = 0
t_dw1 = {}
t_dw2 = 0.0
t_epnum1 = {}
for i in wpresent:
t_epnum1[i] = 0
t_dw1[i] = 0.0
t_epnum2 = 0.0
for iword in wd_extra:
try:
vect = [preprocess(i) for i in iword['Word_vects']
] + [None for i in range(len(iword['h_vect']))]
vect_, dw1, dw2, epnum1, epnum2 = rae_operation(
vect1=vect,
w1=w1,
w2=w2,
w_size=iword['w_size'],
h_vect=iword['h_vect'],
Word_ids=iword['Word_ids'],
wp=iword['wp'],
h_index=iword['h_index'],
dep_order=iword['dep_order'],
hh_index=iword['hh_index'],
p=iword['p'],
regu=regu,
wpresent=wpresent)
terr = 0.0
bc = 0
for iv in range(iword['w_size']):
terr += np.sum(np.power((vect[iv] - vect_[iv]), 2))
bc += 1
branchs += bc
err += terr
if bc:
samp_err.append(terr)
branch_samp_err.append(terr / bc)
t_dw2 += dw2
t_epnum2 += epnum2
except KeyError:
kerr += 1
continue
for wi in wpresent:
t_dw1[wi] += dw1[wi]
t_epnum1[wi] += epnum1[wi]
w1, w2 = weight_update(w1=w1,
w2=w2,
dw1=t_dw1,
dw2=t_dw2,
neta=neta,
regu=regu,
epnum1=t_epnum1,
epnum2=t_epnum2,
wpresent=wpresent)
return w1, w2, None, None, kerr, err, samp_err, branch_samp_err, branchs
def rae_trainning_normal(wd_extra, neta=None, regu=0.9, w=None, wpresent=None):
err = 0.0
kerr = 0.0
samp_err = []
branch_samp_err = []
branchs = 0
t_dw = {}
t_epnum = {}
grads = {}
for i in wpresent:
t_dw[i] = 0.0
t_epnum[i] = 0
for iword in wd_extra:
try:
# vect = [i for i in iword['Word_vects']] + [None for i in range(len(iword['h_vect']))]
vect = [preprocess(i) for i in iword['Word_vects']
] + [None for i in range(len(iword['h_vect']))]
vect_, dw, epnum, grads = rae_operation(
vect=vect,
w=w,
w_size=iword['w_size'],
h_vect=iword['h_vect'],
Word_ids=iword['Word_ids'],
wp=iword['wp'],
h_index=iword['h_index'],
dep_order=iword['dep_order'],
hh_index=iword['hh_index'],
p=iword['p'],
neta=neta,
regu=regu,
wpresent=wpresent)
terr = 0.0
bc = 0
for iv in range(iword['w_size']):
terr += np.sum(np.power((vect[iv] - vect_[iv]), 2))
bc += 1
branchs += bc
err += terr
if bc:
samp_err.append(terr)
branch_samp_err.append(terr / bc)
except RuntimeError:
kerr += 1
continue
for wi in wpresent:
t_dw[wi] += dw[wi]
t_epnum[wi] += epnum[wi]
# weight_update(w, t_dw, neta=neta, regu=regu, epnum=t_epnum, wpresent=wpresent)
adam_weight_update(w,
grads,
neta=neta,
regu=regu,
epnum=t_epnum,
wpresent=wpresent)
return w, t_dw, t_epnum, kerr, err, samp_err, branch_samp_err, branchs
def get_vect_by_wd_dep(flag, words_data, mtype='normal'):
if flag == 't':
Word_ids = get_words_id(words_data)
Word_vects = []
for i in sorted(Word_ids):
Word_vects.append(
preprocess(
get_word_vect(words_data[i]['word'].lower(),
Global.v_size)))
w_size = len(Word_ids)
p = get_parents(words_data)
d = get_dep(words_data)
dep_order, d1 = pdep_2_deporder_dep(p, d)
h_index, h_vect, wp, _ = dep_2_hid_var(p, dep_order, d1, Word_ids)
# Word_vects = get_words_vect(words_data, Word_ids, Global.v_size)
vect = Word_vects + [None for i in range(len(h_vect))]
del Word_vects
w = pickle.load(open(Global.wfname, 'rb'))
if mtype == 'normal':
RAE_adam_herical.rae_encoding(vect=vect,
w=w,
w_size=w_size,
h_vect=h_vect,
wp=wp)
elif mtype == 'deep':
RAE_adam_herical_deep1.rae_encoding(vect=vect,
w=w,
w_size=w_size,
h_vect=h_vect,
wp=wp)
chunks = {}
chunks_vect = {}
for i in range(len(Word_ids)):
chunks[i] = words_data[Word_ids[i]]['word']
chunks_vect[i] = vect[i]
rev_h_index = {v: k for k, v in h_index.items()}
count = i + 1
for i in h_vect:
if len(i) > 1:
chunks_vect[count] = vect[h_index[Word_ids[min(i)]]]
chunks[count] = ' '.join([
words_data[rev_h_index[j]]['word']
if j >= len(Word_ids) else words_data[Word_ids[j]]['word']
for j in i
])
count += 1
return chunks, chunks_vect
def get_chk_vect_by_wd(flag, words_data, mtype='normal'):
if flag == 't':
Word_ids = get_words_id(words_data)
Word_vects = []
for i in sorted(Word_ids):
Word_vects.append(
preprocess(
get_word_vect(words_data[i]['word'].lower(),
Global.v_size)))
w_size = len(Word_ids)
p = get_parents(words_data)
d = get_dep(words_data)
dep_order, d1 = pdep_2_deporder_dep(p, d)
h_index, h_vect, wp, _ = dep_2_hid_var(p, dep_order, d1, Word_ids)
# Word_vects = get_words_vect(words_data, Word_ids, Global.v_size)
vect = Word_vects + [None for i in range(len(h_vect))]
del Word_vects
w = pickle.load(open(Global.wfname, 'rb'))
if mtype == 'normal':
RAE_adam_herical.rae_encoding(vect=vect,
w=w,
w_size=w_size,
h_vect=h_vect,
wp=wp)
elif mtype == 'deep':
RAE_adam_herical_deep1.rae_encoding(vect=vect,
w=w,
w_size=w_size,
h_vect=h_vect,
wp=wp)
chks = get_chunks(words_data)
chks_main = get_chunk_main(chks, dep_order)
chunks = {}
chunks_vect = {}
for c in range(len(chks)):
chunks[c] = ' '.join([words_data[i]['word'] for i in chks[c]])
del words_data
for c in range(len(chks_main)):
ind = h_index[chks_main[c]]
chunks_vect[c] = vect[ind]
return chunks, chunks_vect
def get_chk_vect(flag, line, mtype='normal'):
if flag == 't':
line = line_processing(line)
# print line
words_data = extract_feature_using_senna(line)
p = get_parents(words_data)
d = get_dep(words_data)
Word_ids = get_words_id(words_data)
for i in Word_ids:
words_data[i]['vect'] = preprocess(
get_word_vect(words_data[i]['word'].lower(), Global.v_size))
w_size = len(Word_ids)
dep_order, d1 = pdep_2_deporder_dep(p, d)
h_index, h_vect, wp, _ = dep_2_hid_var(p, dep_order, d1, Word_ids)
Word_vects = get_words_vect(words_data, Word_ids, Global.v_size)
vect = Word_vects + [None for i in h_vect]
del Word_vects
w = pickle.load(open(Global.wfname, 'rb'))
if mtype == 'normal':
RAE_adam_herical.rae_encoding(vect=vect,
w=w,
w_size=w_size,
h_vect=h_vect,
wp=wp)
elif mtype == 'deep':
RAE_adam_herical_deep1.rae_encoding(vect=vect,
w=w,
w_size=w_size,
h_vect=h_vect,
wp=wp)
chks = get_chunks_by_dep(Word_ids, h_index, h_vect)
chunks = {}
chunks_vect = {}
count = 0
order = get_order(d1, w_size)
for m in order:
chunks[count] = ' '.join([words_data[i]['word'] for i in chks[m]])
chunks_vect[count] = vect[h_index[m]]
count += 1
return chunks, chunks_vect
'Word_vects': Word_vects,
'wp': wp,
"dep_order": dep_order,
"hh_index": hh_index,
'Word_ids': Word_ids,
'p': p
})
for i in w:
if type(w[i]) == np.ndarray:
w1[i] = w[i].copy() + e
w2[i] = w[i].copy() - e
epnum = {}
for i in w:
epnum[i] = 0
for iword in wd_extra:
vect1 = [preprocess(i) for i in iword['Word_vects']
] + [None for i in range(len(iword['h_vect']))]
vect2 = [preprocess(i) for i in iword['Word_vects']
] + [None for i in range(len(iword['h_vect']))]
o1 = {}
o2 = {}
vect_1 = {}
vect_2 = {}
rae_encoding(vect=vect1,
w=w1,
w_size=iword['w_size'],
h_vect=iword['h_vect'],
wp=iword['wp'])
vect_1 = rae_decoding(vect=vect1,
o=o1,