def show_simple(run_name, data_id, tex_visulizer):
num_tags = 3
num_select = 20
pickle_name = "save_view_{}_{}".format(run_name, data_id)
tokenizer = get_tokenizer()
data_loader = get_modified_data_loader2(HPSENLI3(), NLIExTrainConfig())
explain_entries = load_from_pickle(pickle_name)
explain_entries = explain_entries
selected_instances = [[], [], []]
for idx, entry in enumerate(explain_entries):
x0, logits, scores = entry
pred = np.argmax(logits)
input_ids = x0
p, h = data_loader.split_p_h_with_input_ids(input_ids, input_ids)
p_tokens = tokenizer.convert_ids_to_tokens(p)
h_tokens = tokenizer.convert_ids_to_tokens(h)
p_tokens = restore_capital_letter(p_tokens)
h_tokens = restore_capital_letter(h_tokens)
target_tag = ["match", "mismatch", "conflict"][pred]
tag_idx = data_generator.NLI.nli_info.tags.index(target_tag)
tag_name = data_generator.NLI.nli_info.tags[tag_idx]
p_score, h_score = data_loader.split_p_h_with_input_ids(scores[tag_idx], input_ids)
normalize_fn = normalize
p_score = normalize_fn(p_score)
h_score = normalize_fn(h_score)
p_row = [Cell("\\textbf{P:}")] + cells_from_tokens(p_tokens, p_score)
h_row = [Cell("\\textbf{H:}")] + cells_from_tokens(h_tokens, h_score)
pred_str = ["entailment", "neutral" , "contradiction"][pred]
apply_color(p_row, tag_name)
apply_color(h_row, tag_name)
#tex_visulizer.write_paragraph(pred_str)
if len(selected_instances[pred]) < num_select :
e = pred_str, [p_row, h_row]
#tex_visulizer.write_instance(pred_str, gold_label, [p_row, h_row])
selected_instances[pred].append(e)
if all([len(s) == num_select for s in selected_instances]):
break
for insts in selected_instances:
for e in insts:
pred_str, rows = e
tex_visulizer.write_instance(pred_str, rows)
return selected_instances
def run(args):
hp = hyperparams.HPSENLI3()
nli_setting = BertNLI()
data_loader = get_modified_data_loader2(hp, nli_setting)
predict_nli_ex(
hp,
nli_setting,
data_loader,
args.tag,
args.data_id,
args.model_path,
args.run_name,
args.modeling_option,
)
def run(args):
hp = hyperparams.HPSENLI3()
nli_setting = BertNLI()
data_loader = get_modified_data_loader2(hp, nli_setting)
if args.method_name in [
'deletion_seq', "random", 'idf', 'deletion', 'LIME'
]:
predictor = nli_baseline_predict
elif args.method_name in [
"elrp",
"deeplift",
"saliency",
"grad*input",
"intgrad",
]:
predictor = nli_attribution_predict
else:
raise Exception(
"method_name={} is not in the known method list.".format(
args.method_name))
predictor(hp, nli_setting, data_loader, args.tag, args.method_name,
args.data_id, args.sub_range, args.model_path)
def show_all(run_name, data_id):
num_tags = 3
num_select = 1
pickle_name = "save_view_{}_{}".format(run_name, data_id)
tokenizer = get_tokenizer()
data_loader = get_modified_data_loader2(HPSENLI3(), NLIExTrainConfig())
explain_entries = load_from_pickle(pickle_name)
explain_entries = explain_entries
visualizer = HtmlVisualizer(pickle_name + ".html")
tex_visulizer = TexTableNLIVisualizer(pickle_name + ".tex")
tex_visulizer.begin_table()
selected_instances = [[], [], []]
for idx, entry in enumerate(explain_entries):
x0, logits, scores = entry
pred = np.argmax(logits)
input_ids = x0
p, h = data_loader.split_p_h_with_input_ids(input_ids, input_ids)
p_tokens = tokenizer.convert_ids_to_tokens(p)
h_tokens = tokenizer.convert_ids_to_tokens(h)
p_rows = []
h_rows = []
p_rows.append(cells_from_tokens(p_tokens))
h_rows.append(cells_from_tokens(h_tokens))
p_score_list = []
h_score_list = []
for j in range(num_tags):
tag_name = data_generator.NLI.nli_info.tags[j]
p_score, h_score = data_loader.split_p_h_with_input_ids(scores[j], input_ids)
normalize_fn = normalize
add = True
if pred == "0":
add = tag_name == "match"
if pred == "1":
add = tag_name == "mismatch"
if pred == "2":
add = tag_name == "conflict"
def format_scores(raw_scores):
def format_float(s):
return "{0:.2f}".format(s)
norm_scores = normalize_fn(raw_scores)
cells = [Cell(format_float(s1), s2, False, False) for s1, s2 in zip(raw_scores, norm_scores)]
if tag_name == "mismatch":
set_cells_color(cells, "G")
elif tag_name == "conflict":
set_cells_color(cells, "R")
return cells
if add:
p_rows.append(format_scores(p_score))
h_rows.append(format_scores(h_score))
p_score_list.append(p_score)
h_score_list.append(h_score)
pred_str = ["Entailment", "Neutral" , "Contradiction"][pred]
out_entry = pred_str, p_tokens, h_tokens, p_score_list, h_score_list
if len(selected_instances[pred]) < num_select :
selected_instances[pred].append(out_entry)
visualizer.write_headline(pred_str)
visualizer.multirow_print_from_cells_list(p_rows)
visualizer.multirow_print_from_cells_list(h_rows)
visualizer.write_instance(pred_str, p_rows, h_rows)
tex_visulizer.write_paragraph(str(pred))
tex_visulizer.multirow_print_from_cells_list(p_rows, width=13)
tex_visulizer.multirow_print_from_cells_list(h_rows, width=13)
if all([len(s) == num_select for s in selected_instances]):
break
tex_visulizer.close_table()
return selected_instances