datas = list(restaurants_train())
for g, att_func in attentions:
if att_func == rbf_attention:
r.vectors[r.items["<UNK>"]] += 10e5
else:
r.vectors[r.items["<UNK>"]] *= 0
for n_noun in noun_cands:
a = aspects[:n_noun]
for idx, (inst, y, label_set) in enumerate(datas):
s = get_scores(inst,
a,
r,
label_set,
gamma=g,
remove_oov=False,
attention_func=att_func)
y_pred = s.argmax(1)
f1_macro = precision_recall_fscore_support(
y, y_pred, average="weighted")[:-1] # noqa
row = (g, fun2name[att_func], n_noun, idx, *f1_macro)
df.append(row)
pbar.update(1)
df = pd.DataFrame(df,
columns=("gamma", "function", "n_noun", "dataset", "p",
"r", "f1"))
df.to_csv("results_grid_search.csv")
for k, v in d.items():
if k.lower() in r.items:
nouns[k.lower()] += v
if att == rbf_attention:
r.vectors[r.items["<UNK>"]] = r.vectors.max()
if att == rbf_attention:
candidates, _ = zip(*nouns.most_common(BEST_RBF["n_noun"]))
else:
candidates, _ = zip(*nouns.most_common(BEST_ATT["n_noun"]))
aspects = [[x] for x in candidates]
for idx, (instances, y, label_set) in enumerate(datums):
s = get_scores(instances,
aspects,
r,
label_set,
gamma=GAMMA,
remove_oov=False,
attention_func=att)
y_pred = s.argmax(1)
f1_score = precision_recall_fscore_support(y, y_pred)
f1_macro = precision_recall_fscore_support(y,
y_pred,
average="weighted")
scores[idx] = (f1_score, f1_macro)
if __name__ == "__main__":
scores = defaultdict(dict)
r = Reach.load("embeddings/restaurant_vecs_w2v.vec",
unk_word="<UNK>")
datums = list(restaurants_test())
for path in sorted(glob("embeddings/restaurant_vecs_w2v.vec")):
r = Reach.load(path, unk_word="<UNK>")
for idx, (instances, y, label_set) in enumerate(datums):
s = get_scores(instances,
[["food"]],
r,
label_set,
remove_oov=False,
attention_func=mean)
y_pred = s.argmax(1)
f1_score = precision_recall_fscore_support(y, y_pred)
f1_macro = precision_recall_fscore_support(y,
y_pred,
average="weighted")
scores[path][idx] = (f1_score, f1_macro)
score_per_class = [[z[x][0][:-1] for x in range(3)]
for z in scores.values()]
score_per_class = np.stack(score_per_class).mean(0)
macro_score = [v[2][1][:-1] for v in scores.values()]