os.mkdir(model_path)
shutil.copyfile(conf_path, os.path.join(model_path, "model.conf"))
for pyfile in ["featurizer.py"]:
shutil.copyfile(pyfile, os.path.join(model_path, pyfile))
if config["model_type"] == "pytorch":
shutil.copyfile("modeling/torch_model.py",
os.path.join(model_path, "torch_model.py"))
elif config["model_type"] == "tf":
shutil.copyfile("modeling/tf_model.py",
os.path.join(model_path, "tf_model.py"))
else:
raise Exception("model_type is not supported")
featurizer = HydraFeaturizer(config)
model = create_model(config, is_train=True)
evaluator = HydraEvaluator(model_path, config, featurizer, model, note)
is_meta = "meta_train" in config.keys() and config["meta_train"] == "True"
if "use_content" in config.keys() and config["use_content"] == "True":
processed_data_path = config["train_data_path"] +\
"_{}_{}_{}".format(
config["base_class"],
config["base_name"],
"filtered" if "filter_content" in config.keys() and config["filter_content"] == "True" else "unfilt"
)
else:
processed_data_path = config["train_data_path"] +\
"_{}_{}".format(config["base_class"], config["base_name"])
if is_meta:
processed_data_path += "_meta_train.pickle"
print(eval_file + ": " + result_str)
if "DEBUG" in self.config:
for text in sq:
print(text[0] + ":" + text[1] + "\t" + text[2])
else:
with open(self.eval_history_file, "a+", encoding="utf8") as f:
f.write("[{0}, epoch {1}] ".format(eval_file, epochs) +
result_str + "\n")
bad_case_file = os.path.join(
self.bad_case_dir,
"{0}_epoch_{1}.log".format(eval_file, epochs))
with open(bad_case_file, "w", encoding="utf8") as f:
for text in sq:
f.write(text[0] + ":" + text[1] + "\t" + text[2] +
"\n")
if __name__ == "__main__":
os.environ["CUDA_VISIBLE_DEVICES"] = "3"
config = utils.read_conf(os.path.join("conf", "wikisql.conf"))
config["DEBUG"] = 1
config["num_train_steps"] = 1000
config["num_warmup_steps"] = 100
featurizer = HydraFeaturizer(config)
model = create_model(config, is_train=True, num_gpu=1)
evaluator = HydraEvaluator("output", config, featurizer, model,
"debug evaluator")
evaluator.eval(0)
###================================================================================================###
# All Best
model_path = "output/20210505_235209"
epoch = 4
engine = DBEngine(db_file)
config = utils.read_conf(os.path.join(model_path, "model.conf"))
# config["DEBUG"] = 1
featurizer = HydraFeaturizer(config)
pred_data = SQLDataset(in_file, config, featurizer, False)
print("num of samples: {0}".format(len(pred_data.input_features)))
##======================EG + TOP_k=============================##
model = create_model(config, is_train=False)
model.load(model_path, epoch)
if "DEBUG" in config:
model_out_file = model_out_file + ".partial"
if os.path.exists(model_out_file):
model_outputs = pickle.load(open(model_out_file, "rb"))
else:
model_outputs = model.dataset_inference(pred_data)
pickle.dump(model_outputs, open(model_out_file, "wb"))
beam_size = args.beam_size
top_k = args.topk
print("===HydraNet+EG===")
def execute_one_test(dataset, shot, model_moment, epoch):
os.environ["CUDA_VISIBLE_DEVICES"] = "0, 1, 2, 3"
model_path = "output/" + model_moment
in_file = "data/wiki{}_content.jsonl".format(
dataset) if shot == "orig" else "data/wiki{}_{}_content.jsonl".format(
dataset, shot)
db_file = "WikiSQL/data/{}.db".format(dataset)
label_file = "WikiSQL/data/{}.jsonl".format(
dataset) if shot == "orig" else "WikiSQL/data_{}/{}.jsonl".format(
shot, dataset)
out_path = "predictions/{}_{}_{}_{}".format(model_moment, epoch, dataset,
shot)
if not os.path.exists(out_path):
os.mkdir(out_path)
out_file = os.path.join(out_path, "out.jsonl")
eg_out_file = os.path.join(out_path, "out_eg.jsonl")
model_out_file = os.path.join(out_path, "model_out.pkl")
test_result_file = os.path.join(out_path, "result.txt")
engine = DBEngine(db_file)
config = utils.read_conf(os.path.join(model_path, "model.conf"))
# config["DEBUG"] = 1
featurizer = HydraFeaturizer(config)
pred_data = SQLDataset(in_file, config, featurizer, False)
print("num of samples: {0}".format(len(pred_data.input_features)))
model = create_model(config, is_train=False)
model.load(model_path, epoch)
if "DEBUG" in config:
model_out_file = model_out_file + ".partial"
model_outputs = model.dataset_inference(pred_data)
pickle.dump(model_outputs, open(model_out_file, "wb"))
# model_outputs = pickle.load(open(model_out_file, "rb"))
print("===HydraNet===")
pred_sqls = model.predict_SQL(pred_data, model_outputs=model_outputs)
with open(out_file, "w") as g:
for pred_sql in pred_sqls:
# print(pred_sql)
result = {"query": {}}
result["query"]["agg"] = int(pred_sql[0])
result["query"]["sel"] = int(pred_sql[1])
result["query"]["conds"] = [(int(cond[0]), int(cond[1]),
str(cond[2])) for cond in pred_sql[2]]
g.write(json.dumps(result) + "\n")
normal_res = print_metric(label_file, out_file, db_file)
print("===HydraNet+EG===")
pred_sqls = model.predict_SQL_with_EG(engine,
pred_data,
model_outputs=model_outputs)
with open(eg_out_file, "w") as g:
for pred_sql in pred_sqls:
# print(pred_sql)
result = {"query": {}}
result["query"]["agg"] = int(pred_sql[0])
result["query"]["sel"] = int(pred_sql[1])
result["query"]["conds"] = [(int(cond[0]), int(cond[1]),
str(cond[2])) for cond in pred_sql[2]]
g.write(json.dumps(result) + "\n")
eg_res = print_metric(label_file, eg_out_file, db_file)
with open(test_result_file, "w") as g:
g.write("normal results:\n" + normal_res + "eg results:\n" + eg_res)