def train_zero_saia_longlist(word_list=noun_list):
with open('../indata/saia_zeroshot_nounslong_splits.json', 'r') as f:
ttsplit = json.load(f)
w2v = linwac.load_w2v()
X = np.load('../indata/saiapr.npz')
Xsaia = X['arr_0']
print "Xsaia", Xsaia.shape
W = np.load('../indata/saiapr_wmat.npz')
Wsaia = W['arr_0']
print "Wsaia", Wsaia.shape
long_noun_list = [
l.strip() for l in open('noun_list_long.txt').readlines()
]
long_noun_ind = [msim.word2ind[n] for n in long_noun_list]
for x, spl in enumerate(ttsplit):
print "SPLIT", x
if x > 0:
Xsaia_train = Xsaia[spl['train']]
Xsaia_test = Xsaia[spl['test']]
Wsaia_t = Wsaia.transpose()
Wsaia_t.shape
Wsaia_t_train = Wsaia_t[spl['train']]
Wsaia_train = Wsaia_t_train.transpose()
print "Train linwac"
linwac.train_all_nouns(Wsaia_train,
Xsaia_train,
w2v,
ssim="500n_zeroshot_split" + str(x),
word_list=long_noun_list)
print "Train transfer"
linmap.train_mappings(msim.w2v_vecs,
long_noun_ind,
Wsaia_train,
Xsaia_train,
split="500n_zeroshot_split" + str(x))
print "Train logwac"
logwac.train_saia_nosamp(Xsaia_train,
Wsaia_train,
word_list=noun_list,
ssim="nouns_zeroshot_split" + str(x))
def train_zero_mixed_plural_saia_models():
with open('../indata/saia_zeroshot_mixedpluralsplit.json', 'r') as f:
ttsplit = json.load(f)
w2v = linwac.load_w2v()
X = np.load('../indata/saiapr.npz')
Xsaia = X['arr_0']
print "Xsaia", Xsaia.shape
W = np.load('../indata/saiapr_wmat.npz')
Wsaia = W['arr_0']
print "Wsaia", Wsaia.shape
Xsaia_train = Xsaia[ttsplit['train']]
Xsaia_test = Xsaia[ttsplit['test']]
Wsaia_t = Wsaia.transpose()
Wsaia_t.shape
Wsaia_t_train = Wsaia_t[ttsplit['train']]
Wsaia_train = Wsaia_t_train.transpose()
print "Plurals", ttsplit['nouns']
print "Singulars", ttsplit['singulars']
print "Train linwac"
this_wordlist = ttsplit['nouns'] + ttsplit['singulars']
word_ind = [msim.word2ind[n] for n in this_wordlist]
print "Wordlist", len(this_wordlist)
linwac.train_all_nouns(Wsaia_train,
Xsaia_train,
w2v,
ssim="_zeroshot_mixedpluralsplit",
word_list=this_wordlist)
print "Train transfer"
linmap.train_mappings(msim.w2v_vecs,
word_ind,
Wsaia_train,
Xsaia_train,
split="_zeroshot_mixedpluralsplit")
print "Train logwac"
logwac.train_saia_nosamp(Xsaia_train,
Wsaia_train,
word_list=this_wordlist,
ssim="nouns_zeroshot_mixedpluralsplit")
def train_zero_hypern_saia_models():
with open('../indata/saia_zeroshot_hypernsplit.json', 'r') as f:
ttsplit = json.load(f)
w2v = linwac.load_w2v()
X = np.load('../indata/saiapr.npz')
Xsaia = X['arr_0']
print "Xsaia", Xsaia.shape
W = np.load('../indata/saiapr_wmat.npz')
Wsaia = W['arr_0']
print "Wsaia", Wsaia.shape
Xsaia_train = Xsaia[ttsplit['train']]
Xsaia_test = Xsaia[ttsplit['test']]
Wsaia_t = Wsaia.transpose()
Wsaia_t.shape
Wsaia_t_train = Wsaia_t[ttsplit['train']]
Wsaia_train = Wsaia_t_train.transpose()
print ttsplit['nouns']
print "Train linwac"
this_wordlist = noun_list + [
n for n in ttsplit['nouns'] if not n in noun_list
]
print "Wordlist", len(this_wordlist)
linwac.train_all_nouns(Wsaia_train,
Xsaia_train,
w2v,
ssim="_zeroshot_hypernsplit",
word_list=this_wordlist)
print "Train transfer"
linmap.train_mappings(msim.w2v_vecs,
noun_ind,
Wsaia_train,
Xsaia_train,
split="_zeroshot_hypernsplit")
print "Train logwac"
logwac.train_saia_nosamp(Xsaia_train,
Wsaia_train,
word_list=noun_list,
ssim="nouns_zeroshot_hypernsplit")
def train_zero_saia_models(word_list=noun_list):
with open('../indata/saia_zeroshot_nounsplits.json', 'r') as f:
ttsplit = json.load(f)
w2v = linwac.load_w2v()
X = np.load('../indata/saiapr.npz')
Xsaia = X['arr_0']
print "Xsaia", Xsaia.shape
W = np.load('../indata/saiapr_wmat.npz')
Wsaia = W['arr_0']
print "Wsaia", Wsaia.shape
for x, spl in enumerate(ttsplit):
print "SPLIT", x
if x > 0:
Xsaia_train = Xsaia[spl['train']]
Xsaia_test = Xsaia[spl['test']]
Wsaia_t = Wsaia.transpose()
Wsaia_t.shape
Wsaia_t_train = Wsaia_t[spl['train']]
Wsaia_train = Wsaia_t_train.transpose()
print "Train linwac"
linwac.train_all_nouns(Wsaia_train,
Xsaia_train,
w2v,
ssim="_zeroshot_split" + str(x))
print "Train transfer"
linmap.train_mappings(msim.w2v_vecs,
noun_ind,
Wsaia_train,
Xsaia_train,
split="_zeroshot_split" + str(x))
print "Train logwac"
logwac.train_saia_nosamp(Xsaia_train,
Wsaia_train,
word_list=noun_list,
ssim="nouns_zeroshot_split" + str(x))
def make_results_randomsplits():
with gzip.open('../indata/saia_zeroshot_nounsplits_testsets.pklz',
'r') as f:
testsets = pickle.load(f)
with open('../indata/saia_zeroshot_nounsplits.json', 'r') as f:
ttsplit = json.load(f)
print "testsets", len(testsets)
w2v = linwac.load_w2v()
nouns_w2v_mat = np.array([msim.w2v_vecs[msim.word2ind[n]] for n in NOUNS])
print "W2v matrix", nouns_w2v_mat.shape
mapmodels = [
'../linmodels/linmap_nouns_zeroshot_split' + str(x) + '.pklz'
for x in range(10)
]
linmodels = [
'../linmodels/linwac_nouns_w2v_zeroshot_split' + str(x) + '.pklz'
for x in range(10)
]
logmodels = [
'../logmodels/logwac_saia_nouns_zeroshot_split' + str(x) +
'_nosamp.pklz' for x in range(10)
]
print "*****Eval mapmodels"
res1 = eval_map_models(testsets[:10], ttsplit[:10], mapmodels,
nouns_w2v_mat, NOUNS)
print "*****Eval logmodels"
res2 = eval_log_models(testsets[:10], ttsplit[:10], logmodels,
nouns_w2v_mat, NOUNS)
print "*****Eval linmodels"
res3 = eval_lin_models(testsets[:10], ttsplit[:10], linmodels,
nouns_w2v_mat, NOUNS)
results = []
for (x, y, a, b, c, d) in res1 + res2 + res3:
results.append((x, y, "%.2f" % (a * 100), "%.2f" % (b * 100),
"%.2f" % (c * 100), "%.2f" % (d * 100)))
df = pd.DataFrame(results,
columns=['testset', 'model', '@1', '@2', '@5', '@10'])
print df.to_latex(index=False)
def train_standard_plural_saia_models():
with open('../indata/saia_standard_pluralsplit.json', 'r') as f:
ttsplit = json.load(f)
print "Nouns", len(ttsplit['nouns'])
w2v = linwac.load_w2v()
Xsaia_t, Wsaia_t = linwac.load_saia_train()
Xsaia_train = Xsaia_t[ttsplit['train']]
Wsaia_tt = Wsaia_t.transpose()
print Wsaia_tt.shape
Wsaia_t_train = Wsaia_tt[ttsplit['train']]
Wsaia_train = Wsaia_t_train.transpose()
print "Train linwac"
this_wordlist = ttsplit['nouns']
word_ind = [msim.word2ind[n] for n in this_wordlist]
print "Wordlist", len(this_wordlist)
linwac.train_all_nouns(Wsaia_train,
Xsaia_train,
w2v,
ssim="_standard_pluralsplit",
word_list=this_wordlist)
print "Train transfer"
linmap.train_mappings(msim.w2v_vecs,
word_ind,
Wsaia_train,
Xsaia_train,
split="_standard_pluralsplit")
print "Train logwac"
logwac.train_saia_nosamp(Xsaia_train,
Wsaia_train,
word_list=this_wordlist,
ssim="nouns_standard_pluralsplit")
def train_zero_refcoco_models():
long_noun_list = [
l.strip() for l in open('noun_list_long.txt').readlines()
]
long_noun_ind = [msim.word2ind[n] for n in long_noun_list]
with open('../indata/refcoco_zeroshot_nounsplits.json', 'r') as f:
ttsplit = json.load(f)
w2v = linwac.load_w2v()
X = np.load('../indata/mscoco.npz')
Xcoco = X['arr_0']
print "Xcoco", Xcoco.shape
W = np.load('../indata/mscoco_refcoco_wmat.npz')
Wcoco = W['arr_0']
print "Wcoco", Wcoco.shape
total = 0
for rx in range(Wcoco.shape[1]):
if len(np.nonzero(Wcoco[:, rx])[0]):
total += 1
print "Total", total
with gzip.open('../indata/refcoco_refdf.pklz', 'r') as f:
refdf = pickle.load(f)
print "Regionids", len(set(refdf['region_id']))
for x, spl in enumerate(ttsplit):
print "SPLIT", x
if x == 2:
Xcoco_train = Xcoco[spl['train']]
Xcoco_test = Xcoco[spl['test']]
print "Xcoco", Xcoco_train.shape
Wcoco_t = Wcoco.transpose()
Wcoco_t.shape
Wcoco_t_train = Wcoco_t[spl['train']]
Wcoco_train = Wcoco_t_train.transpose()
print "Wcoco", Wcoco_train.shape
total = 0
for rx in range(Wcoco_train.shape[1]):
if len(np.nonzero(Wcoco_train[:, rx])[0]):
total += 1
print "Total", total
print "Train linwac"
linwac.train_vocab_abssample(Wcoco_train,Xcoco_train,w2v,\
nsamp=Xcoco_train.shape[0],scorp="rcoco",ssim="w2v_zeroshot_split"+str(x),word_list=long_noun_list)
#linwac.train_all_nouns(Wcoco_train,Xcoco_train,w2v,ssim="refcoco_w2v_zeroshot_split"+str(x),word_list=long_noun_list)
print "Train transfer"
linmap.train_mappings(msim.w2v_vecs,
long_noun_ind,
Wcoco_train,
Xcoco_train,
split="_refcoco_zeroshot_split" + str(x))
print "Train logwac"
logwac.train_saia_nosamp(Xsaia_train,
Wsaia_train,
word_list=noun_list,
ssim="nouns_zeroshot_split" + str(x))