def one_match(asts, ast2, rewrite=True, ignore_arg_value=False):
if rewrite:
with er.DBConnection() as db:
ast_rewrites = get_rewrites(asts, db)
else:
ast_rewrites = asts
cmd2 = ignore_differences(data_tools.ast2template(
ast2, loose_constraints=True, arg_type_only=ignore_arg_value))
for ast1 in ast_rewrites:
cmd1 = data_tools.ast2template(
ast1, loose_constraints=True, arg_type_only=ignore_arg_value)
cmd1 = ignore_differences(cmd1)
if cmd1 == cmd2:
return True
return False
def manual_eval(model, dataset, rev_nl_vocab, FLAGS, output_dir, num_eval=30):
num_top1_correct_temp = 0.0
num_top3_correct_temp = 0.0
num_top5_correct_temp = 0.0
num_top10_correct_temp = 0.0
num_top1_correct = 0.0
num_top3_correct = 0.0
num_top5_correct = 0.0
num_top10_correct = 0.0
num_evaled = 0
grouped_dataset = data_utils.group_data_by_nl(dataset, use_bucket=False) \
.values()
random.shuffle(grouped_dataset, lambda: 0.5208484091114275)
cmd_parser = data_tools.bash_parser if FLAGS.dataset == "bash" \
else data_utils.parse_brackets
o_f = open(os.path.join(output_dir, "manual.eval.results"), 'w')
rejudge = False
with DBConnection() as db:
db.create_schema()
while num_evaled < len(grouped_dataset):
nl_strs, cm_strs, nls, search_historys = grouped_dataset[num_evaled]
nl_str = nl_strs[0].decode('utf-8')
if num_evaled == num_eval:
break
gt_trees = [cmd_parser(cmd) for cmd in cm_strs]
predictions = db.get_top_k_predictions(model, nl_str, k=10)
top1_correct_temp, top3_correct_temp, top5_correct_temp, top10_correct_temp = \
False, False, False, False
top1_correct, top3_correct, top5_correct, top10_correct = \
False, False, False, False
# evaluation ignoring ordering of flags
print("Example %d (%d)" % (num_evaled+1, len(cm_strs)))
o_f.write("Example %d (%d)" % (num_evaled+1, len(cm_strs)) + "\n")
print("English: " + nl_str.strip())
o_f.write("English: " + nl_str.encode('utf-8'))
for j in xrange(len(cm_strs)):
print("GT Command %d: " % (j+1) + cm_strs[j].strip())
o_f.write("GT Command %d: " % (j+1) + cm_strs[j].strip() + "\n")
pred_id = 0
while pred_id < min(1, len(predictions)):
pred_cmd, score = predictions[pred_id]
tree = cmd_parser(pred_cmd)
print("Prediction {}: {} ({})".format(pred_id+1, pred_cmd, score))
o_f.write("Prediction {}: {} ({})\n".format(pred_id+1, pred_cmd, score))
print()
pred_temp = data_tools.ast2template(tree, loose_constraints=True)
str_judge = db.get_str_judgement((nl_str, pred_cmd))
temp_judge = db.get_temp_judgement((nl_str, pred_temp))
if temp_judge is not None and not rejudge:
judgement_str = "y" if temp_judge == 1 \
else "n ({})".format(error_types[temp_judge])
print("Correct template [y/n]: %s" % judgement_str)
else:
temp_judge = ast_based.one_match(gt_trees, tree, rewrite=False,
ignore_arg_value=True)
if not temp_judge:
inp = raw_input("Correct template [y/n]: ")
if inp == "REVERSE":
rejudge = True
else:
if inp == "y":
temp_judge = True
db.add_temp_judgement((nl_str, pred_temp, 1))
else:
temp_judge = False
error_type = raw_input(
"Error type: \n"
"(2) extra utility \n"
"(3) missing utility \n"
"(4) confused utility \n"
"(5) extra flag \n"
"(6) missing flag \n"
"(7) confused flag \n"
"(8) logic error\n"
"(9) count error\n"
)
db.add_temp_judgement((nl_str, pred_temp, int(error_type)))
rejudge = False
else:
print("Correct template [y/n]: y")
if temp_judge == 1:
if pred_id < 1:
top1_correct_temp = True
top3_correct_temp = True
top5_correct_temp = True
top10_correct_temp = True
elif pred_id < 3:
top3_correct_temp = True
top5_correct_temp = True
top10_correct_temp = True
elif pred_id < 5:
top5_correct_temp = True
top10_correct_temp = True
elif pred_id < 10:
top10_correct_temp = True
o_f.write("C")
if str_judge is not None and not rejudge:
judgement_str = "y" if str_judge == 1 \
else "n ({})".format(error_types[str_judge])
print("Correct command [y/n]: %s" % judgement_str)
else:
str_judge = ast_based.one_match(gt_trees, tree, rewrite=False,
ignore_arg_value=False)
if not str_judge:
inp = raw_input("Correct command [y/n]: ")
if inp == "REVERSE":
rejudge = True
continue
elif inp == "y":
str_judge = True
o_f.write("C")
db.add_str_judgement((nl_str, pred_cmd, 1))
else:
str_judge = False
o_f.write("W")
db.add_str_judgement((nl_str, pred_cmd, 0))
else:
print("Correct command [y/n]: y")
if str_judge == 1:
if pred_id < 1:
top1_correct = True
top3_correct = True
top5_correct = True
top10_correct = True
elif pred_id < 3:
top3_correct = True
top5_correct = True
top10_correct = True
elif pred_id < 5:
top5_correct = True
top10_correct = True
elif pred_id < 10:
top10_correct = True
o_f.write("C")
else:
o_f.write("W")
else:
o_f.write("WW")
o_f.write("\n")
o_f.write("\n")
pred_id += 1
if rejudge:
num_evaled -= 1
else:
num_evaled += 1
if top1_correct_temp:
num_top1_correct_temp += 1
if top3_correct_temp:
num_top3_correct_temp += 1
if top5_correct_temp:
num_top5_correct_temp += 1
if top10_correct_temp:
num_top10_correct_temp += 1
if top1_correct:
num_top1_correct += 1
if top3_correct:
num_top3_correct += 1
if top5_correct:
num_top5_correct += 1
if top10_correct:
num_top10_correct += 1
rejudge = False
print()
print("%d examples evaluated" % num_eval)
print("Top 1 Template Match Score = %.2f" % (num_top1_correct_temp/num_eval))
print("Top 1 String Match Score = %.2f" % (num_top1_correct/num_eval))
if len(predictions) > 3:
print("Top 5 Template Match Score = %.2f" % (num_top5_correct_temp/num_eval))
print("Top 5 String Match Score = %.2f" % (num_top5_correct/num_eval))
print("Top 10 Template Match Score = %.2f" % (num_top10_correct_temp/num_eval))
print("Top 10 String Match Score = %.2f" % (num_top10_correct/num_eval))
print()
o_f.write("%d examples evaluated" % num_eval + "\n")
o_f.write("Top 1 Template MatchScore = %.2f" % (num_top1_correct_temp/num_eval) + "\n")
o_f.write("Top 1 String Match Score = %.2f" % (num_top1_correct/num_eval) + "\n")
if len(predictions) > 1:
o_f.write("Top 5 Template Match Score = %.2f" % (num_top5_correct_temp/num_eval) + "\n")
o_f.write("Top 5 String Match Score = %.2f" % (num_top5_correct/num_eval) + "\n")
o_f.write("Top 10 Template Match Score = %.2f" % (num_top10_correct_temp/num_eval) + "\n")
o_f.write("Top 10 String Match Score = %.2f" % (num_top10_correct/num_eval) + "\n")
o_f.write("\n")
def string_match(ast1, ast2):
str1 = ignore_differences(
data_tools.ast2template(ast1, loose_constraints=True, arg_type_only=False))
str2 = ignore_differences(
data_tools.ast2template(ast2, loose_constraints=True, arg_type_only=False))
return str1 == str2
def template_match(ast1, ast2):
temp1 = ignore_differences(
data_tools.ast2template(ast1, loose_constraints=True))
temp2 = ignore_differences(
data_tools.ast2template(ast2, loose_constraints=True))
return temp1 == temp2
