def train_test_repeat(load_id, exp_name, n_repeat): hp = hyperparams.HPBert() e_config = ExperimentConfig() e_config.name = "RTE_{}".format("A") e_config.num_epoch = 10 e_config.save_interval = 30 * 60 # 30 minutes e_config.load_names = ['bert'] vocab_filename = "bert_voca.txt" data_loader = rte.DataLoader(hp.seq_max, vocab_filename, True) print(load_id) scores = [] for i in range(n_repeat): e = Experiment(hp) print(exp_name) e_config.name = "rte_{}".format(exp_name) save_path = e.train_rte(e_config, data_loader, load_id) acc = e.eval_rte(e_config, data_loader, save_path) scores.append(acc) print(exp_name) for e in scores: print(e, end="\t") print() r = average(scores) print("Avg\n{0:.03f}".format(r)) return r
def train_snli_ex(): hp = hyperparams.HPBert() hp.compare_deletion_num = 20 e = Experiment(hp) nli_setting = NLI() nli_setting.vocab_size = 30522 nli_setting.vocab_filename = "bert_voca.txt" e_config = ExperimentConfig() e_config.name = "SNLIEx_B" e_config.ex_val = False e_config.num_epoch = 1 e_config.save_interval = 30 * 60 # 30 minutes e_config.load_names = ['bert', 'cls_dense'] #, 'aux_conflict'] #explain_tag = 'match' # 'dontcare' 'match' 'mismatch' #explain_tag = 'mismatch' #explain_tag = 'conflict' data_loader = nli.SNLIDataLoader(hp.seq_max, nli_setting.vocab_filename, True) #load_id = ("NLI_run_nli_warm", "model-97332") #load_id = ("NLIEx_A", "model-16910") #load_id = ("uncased_L-12_H-768_A-12", 'bert_model.ckpt') #load_id = ("NLIEx_D", "model-1964") #load_id = ("NLIEx_D", "model-1317") load_id = ("SNLI_Only_A", 'model-0') e.train_nli_any_way(nli_setting, e_config, data_loader, load_id)
def analyze_nli_ex(): hp = hyperparams.HPBert() e = Experiment(hp) nli_setting = NLI() nli_setting.vocab_size = 30522 nli_setting.vocab_filename = "bert_voca.txt" explain_tag = 'match' e_config = ExperimentConfig() #e_config.name = "NLIEx_{}_premade_analyze".format(explain_tag) e_config.name = "NLIEx_{}_analyze".format(explain_tag) e_config.num_epoch = 4 e_config.save_interval = 30 * 60 # 30 minutes e_config.load_names = ['bert', 'cls_dense', 'aux_conflict'] data_loader = nli.DataLoader(hp.seq_max, nli_setting.vocab_filename, True) #load_id = ("NLIEx_E_align", "model-23621") #load_id = ("NLIEx_I_match", "model-1238") if explain_tag == 'conflict': load_id = ("NLIEx_Y_conflict", "model-12039") #load_id = ("NLIEx_HB", "model-2684") elif explain_tag == 'match': load_id = ("NLIEx_P_match", "model-1636") load_id = ("NLIEx_X_match", "model-12238") elif explain_tag == 'mismatch': load_id = ("NLIEx_U_mismatch", "model-10265") e.nli_visualization(nli_setting, e_config, data_loader, load_id, explain_tag)
def ukp_train_test_repeat(load_id, exp_name, topic, n_repeat): hp = hyperparams.HPBert() e_config = ExperimentConfig() e_config.num_epoch = 2 e_config.save_interval = 100 * 60 # 30 minutes e_config.voca_size = 30522 e_config.load_names = ['bert'] encode_opt = "is_good" print(load_id) scores = [] for i in range(n_repeat): e = Experiment(hp) print(exp_name) e_config.name = "arg_{}_{}_{}".format(exp_name, topic, encode_opt) data_loader = BertDataLoader(topic, True, hp.seq_max, "bert_voca.txt", option=encode_opt) save_path = e.train_ukp(e_config, data_loader, load_id) f1_last = e.eval_ukp(e_config, data_loader, save_path) scores.append(f1_last) print(exp_name) print(encode_opt) for e in scores: print(e, end="\t") print() print("Avg\n{0:.03f}".format(average(scores)))
def eval_ukp_with_nli(exp_name): step_per_epoch = 24544 + 970 hp = hyperparams.HPBert() e_config = ExperimentConfig() e_config.num_steps = step_per_epoch e_config.voca_size = 30522 e_config.num_dev_batches = 30 e_config.load_names = ['bert'] encode_opt = "is_good" num_class_list = [3, 3] f1_list = [] save_path = "/mnt/scratch/youngwookim/Chair/output/model/runs/argmix_AN_B_40000_abortion_is_good/model-21306" for topic in data_generator.argmining.ukp_header.all_topics[:1]: e = Experiment(hp) print(exp_name) e_config.name = "argmix_{}_{}_{}".format(exp_name, topic, encode_opt) arg_data_loader = BertDataLoader(topic, True, hp.seq_max, "bert_voca.txt", option=encode_opt) f1_last = e.eval_ukp_on_shared(e_config, arg_data_loader, num_class_list, save_path) f1_list.append((topic, f1_last)) print(exp_name) print(encode_opt) print(f1_list) for key, score in f1_list: print("{0}\t{1:.03f}".format(key, score))
def test_fidelity(): hp = hyperparams.HPBert() e = Experiment(hp) nli_setting = NLI() nli_setting.vocab_size = 30522 nli_setting.vocab_filename = "bert_voca.txt" is_senn = False e_config = ExperimentConfig() e_config.name = "NLIEx_{}".format("Fidelity") e_config.num_epoch = 4 e_config.save_interval = 30 * 60 # 30 minutes if is_senn: e_config.load_names = ['bert', 'cls_dense', 'aux_conflict'] else: e_config.load_names = ['bert', 'cls_dense'] explain_tag = 'conflict' data_loader = nli.DataLoader(hp.seq_max, nli_setting.vocab_filename, True) load_id = ("NLIEx_Y_conflict", 'model-12039') #load_id = ("NLI_Only_C", 'model-0') #e.eval_fidelity(nli_setting, e_config, data_loader, load_id, explain_tag) e.eval_fidelity_gradient(nli_setting, e_config, data_loader, load_id, explain_tag)
def wikicont_cnn(): e = Experiment(hyperparams.HPCNN()) e_config = ExperimentConfig() e_config.num_epoch = 0 e_config.input_name = "WikiContrvCNN" e_config.name = "WikiContrvCNN_sigmoid" e.train_cnn_wiki_contrv(e_config)
def ukp_train_test(load_id, exp_name): hp = hyperparams.HPBert() e_config = ExperimentConfig() e_config.num_epoch = 2 e_config.save_interval = 100 * 60 # 30 minutes e_config.voca_size = 30522 e_config.load_names = ['bert'] encode_opt = "is_good" print(load_id) f1_list = [] for topic in data_generator.argmining.ukp_header.all_topics: e = Experiment(hp) print(exp_name) e_config.name = "arg_{}_{}_{}".format(exp_name, topic, encode_opt) data_loader = BertDataLoader(topic, True, hp.seq_max, "bert_voca.txt", option=encode_opt) save_path = e.train_ukp(e_config, data_loader, load_id) print(topic) f1_last = e.eval_ukp(e_config, data_loader, save_path) f1_list.append((topic, f1_last)) print(exp_name) print(encode_opt) print(f1_list) for key, score in f1_list: print("{0}\t{1:.03f}".format(key, score))
def predict_rf(): hp = hyperparams.HPBert() hp.batch_size = 256 e = Experiment(hp) nli_setting = NLI() nli_setting.vocab_size = 30522 nli_setting.vocab_filename = "bert_voca.txt" target_label = 'mismatch' #data_id = 'test_conflict' data_id = "{}_1000".format(target_label) e_config = ExperimentConfig() #del_g = 0.7 #e_config.name = "X_match_del_{}".format(del_g) e_config.name = "NLIEx_AnyA_{}".format(target_label) e_config.load_names = ['bert', 'cls_dense', 'aux_conflict'] data_loader = nli.DataLoader(hp.seq_max, nli_setting.vocab_filename, True) #load_id = ("NLI_bare_A", 'model-195608') #load_id = ("NLIEx_O", 'model-10278') load_id = ("NLIEx_W_mismatch", "model-12030") load_id = ("NLIEx_Y_conflict", "model-12039") load_id = ("NLIEx_X_match", "model-12238") #load_id = ("NLIEx_match_del_{}".format(del_g), "model-4390") load_id = ("NLIEx_CE_{}".format(target_label), "model-12199") load_id = ("NLIEx_AnyA", "model-7255") e.predict_rf(nli_setting, e_config, data_loader, load_id, data_id, 5)
def train_nli_smart_rf(): hp = hyperparams.HPSENLI() hp.compare_deletion_num = 20 e = Experiment(hp) nli_setting = NLI() nli_setting.vocab_size = 30522 nli_setting.vocab_filename = "bert_voca.txt" e_config = ExperimentConfig() #explain_tag = 'mismatch' explain_tag = 'match' #explain_tag = 'mismatch' loss_type = 2 e_config.name = "NLIEx_Hinge_{}".format(explain_tag) e_config.num_epoch = 1 e_config.ex_val = True e_config.save_interval = 30 * 60 # 30 minutes e_config.load_names = ['bert', 'cls_dense'] #, 'aux_conflict'] e_config.save_eval = True e_config.save_name = "LossFn_{}_{}".format(loss_type, explain_tag) data_loader = nli.DataLoader(hp.seq_max, nli_setting.vocab_filename, True) load_id = ("NLI_run_A", 'model-0') print("Loss : ", loss_type) e.train_nli_smart(nli_setting, e_config, data_loader, load_id, explain_tag, loss_type)
def train_score_merger_on_vector(): hp = hyperparams.HPMerger() e = Experiment(hp) e_config = ExperimentConfig() e_config.name = "MergerE_{}".format("E") data_loader = score_loader.NetOutputLoader(hp.seq_max, hp.hidden_units, hp.batch_size) e.train_score_merger(e_config, data_loader)
def train_score_merger(): hp = hyperparams.HPMerger_BM25() e = Experiment(hp) e_config = ExperimentConfig() e_config.name = "Merger_{}".format("A") e_config.num_epoch = 4 data_loader = score_loader.DataLoader(hp.seq_max, hp.hidden_units) e.train_score_merger(e_config, data_loader)
def lm_tweets_train(): hp = HPTweets() data = tweets.TweetLoader("atheism", hp.seq_max, shared_setting.Tweets2Stance) e_config = ExperimentConfig() e_config.name = "LM_tweets" e_config.num_epoch = 30 e_config.save_interval = 30 * 60 # 30 minutes e = Experiment(hp) e.train_lm_batch(e_config, data)
def lm_guardian_train(): hp = Hyperparams() guardian_data = guardian.GuardianLoader("atheism", hp.seq_max, shared_setting.Guardian2Stance) e_config = ExperimentConfig() e_config.name = "LM_guardian" e_config.num_epoch = 30 e = Experiment(hp) e.train_lm_batch(e_config, guardian_data)
def gradient_rte_visulize(): hp = hyperparams.HPBert() e = Experiment(hp) vocab_filename = "bert_voca.txt" load_id = loader.find_model_name("RTE_A") e_config = ExperimentConfig() e_config.name = "RTE_{}".format("visual") e_config.save_interval = 30 * 60 # 30 minutes e_config.load_names = ['bert', 'cls_dense'] data_loader = rte.DataLoader(hp.seq_max, vocab_filename, True) e.rte_visualize(e_config, data_loader, load_id)
def pair_lm(): hp = HPPairTweet() topic = "atheism" setting = shared_setting.TopicTweets2Stance(topic) tweet_group = tweet_reader.load_per_user(topic) data = loader.PairDataLoader(hp.sent_max, setting, tweet_group) e_config = ExperimentConfig() e_config.name = "LM_pair_tweets_{}".format(topic) e_config.num_epoch = 1 e_config.save_interval = 30 * 60 # 30 minutes e = Experiment(hp) e.train_pair_lm(e_config, data)
def train_nli_with_reinforce_old(): hp = hyperparams.HPNLI2() e = Experiment(hp) nli_setting = NLI() e_config = ExperimentConfig() e_config.name = "NLI_run_{}".format("retest") e_config.num_epoch = 4 e_config.save_interval = 30 * 60 # 30 minutes e_config.load_names = ['bert', 'dense_cls'] #, 'aux_conflict'] data_loader = nli.DataLoader(hp.seq_max, nli_setting.vocab_filename) load_id = ("interval", "model-48040") e.train_nli_ex_0(nli_setting, e_config, data_loader, load_id, True)
def train_nil(): hp = hyperparams.HPBert() e = Experiment(hp) nli_setting = NLI() nli_setting.vocab_size = 30522 nli_setting.vocab_filename = "bert_voca.txt" e_config = ExperimentConfig() e_config.name = "NLI_only_{}".format("B") e_config.num_epoch = 2 e_config.save_interval = 30 * 60 # 30 minutes data_loader = nli.DataLoader(hp.seq_max, nli_setting.vocab_filename, True) e.train_nli(nli_setting, e_config, data_loader)
def stance_with_consistency(): hp = HPStanceConsistency() topic = "atheism" e = Experiment(hp) e_config = ExperimentConfig() e_config.name = "stance_consistency_{}".format(topic) setting = shared_setting.TopicTweets2Stance(topic) stance_data = stance_detection.DataLoader(topic, hp.seq_max, setting.vocab_filename) tweet_group = tweet_reader.load_per_user(topic) aux_data = AuxPairLoader(hp.seq_max, setting, tweet_group) voca_size = setting.vocab_size e.train_stance_consistency(voca_size, stance_data, aux_data)
def protest_bert(): hp = hyperparams.HPBert() e = Experiment(hp) e_config = ExperimentConfig() e_config.name = "protest" e_config.num_epoch = 1 e_config.save_interval = 1 * 60 # 1 minutes e_config.load_names = ['bert'] vocab_size = 30522 vocab_filename = "bert_voca.txt" data_loader = protest.DataLoader(hp.seq_max, vocab_filename, vocab_size) load_id = ("uncased_L-12_H-768_A-12", 'bert_model.ckpt') e.train_protest(e_config, data_loader, load_id)
def pred_mnli_anyway(): hp = hyperparams.HPBert() e = Experiment(hp) nli_setting = NLI() nli_setting.vocab_size = 30522 nli_setting.vocab_filename = "bert_voca.txt" e_config = ExperimentConfig() e_config.name = "NLIEx_AnyA" e_config.load_names = ['bert', 'cls_dense', 'aux_conflict'] data_loader = nli.DataLoader(hp.seq_max, nli_setting.vocab_filename, True) target_label = 'mismatch' data_id = "{}_1000".format(target_label) load_id = ("NLIEx_AnyA", 'model-2785') e.predict_rf(nli_setting, e_config, data_loader, load_id, data_id)
def pred_snli_ex(): hp = hyperparams.HPBert() e = Experiment(hp) nli_setting = NLI() nli_setting.vocab_size = 30522 nli_setting.vocab_filename = "bert_voca.txt" e_config = ExperimentConfig() e_config.name = "SNLIEx_B" e_config.load_names = ['bert', 'cls_dense', 'aux_conflict'] data_loader = nli.SNLIDataLoader(hp.seq_max, nli_setting.vocab_filename, True) load_id = ("SNLIEx_B", 'model-10275') e.predict_rf(nli_setting, e_config, data_loader, load_id, "test")
def crs_stance_baseline(): hp = hyperparams.HPCRS() hp.batch_size = 16 e = Experiment(hp) e_config = ExperimentConfig() e_config.name = "CRS_{}".format("baseline") e_config.num_epoch = 4 e_config.save_interval = 10 * 60 # 60 minutes e_config.load_names = ['bert'] #, 'reg_dense'] e_config.voca_size = 30522 data_loader = DataGenerator() load_id = ("uncased_L-12_H-768_A-12", 'bert_model.ckpt') e.train_crs_classify(e_config, data_loader, load_id)
def train_rte(): hp = hyperparams.HPBert() e = Experiment(hp) vocab_filename = "bert_voca.txt" #load_id = ("uncased_L-12_H-768_A-12", 'bert_model.ckpt') load_id = ("tlm_simple", "model.ckpt-15000") e_config = ExperimentConfig() e_config.name = "RTE_{}".format("tlm_simple_15000") e_config.num_epoch = 10 e_config.save_interval = 30 * 60 # 30 minutes e_config.load_names = ['bert'] data_loader = rte.DataLoader(hp.seq_max, vocab_filename, True) e.train_rte(e_config, data_loader, load_id)
def train_adhoc_with_reinforce(): hp = hyperparams.HPAdhoc() e = Experiment(hp) e_config = ExperimentConfig() e_config.name = "Adhoc_{}".format("E") e_config.num_epoch = 4 e_config.save_interval = 30 * 60 # 30 minutes e_config.load_names = ['bert'] vocab_size = 30522 vocab_filename = "bert_voca.txt" data_loader = ws.DataLoader(hp.seq_max, vocab_filename, vocab_size) load_id = ("uncased_L-12_H-768_A-12", 'bert_model.ckpt') e.train_adhoc(e_config, data_loader, load_id)
def train_nil(): hp = HP() e = Experiment(hp) nli_setting = NLI() nli_setting.vocab_size = 30522 nli_setting.vocab_filename = "bert_voca.txt" e_config = ExperimentConfig() e_config.name = "NLI_only_{}".format("512") e_config.num_epoch = 2 e_config.save_interval = 30 * 60 # 30 minutes data_loader = nli.DataLoader(hp.seq_max, nli_setting.vocab_filename, True) load_id = ("uncased_L-12_H-768_A-12", 'bert_model.ckpt') e.train_nli_ex_0(nli_setting, e_config, data_loader, load_id, False)
def train_nil_cold(): print('train_nil_cold') hp = hyperparams.HPBert() e = Experiment(hp) nli_setting = NLI() nli_setting.vocab_size = 30522 nli_setting.vocab_filename = "bert_voca.txt" e_config = ExperimentConfig() e_config.name = "NLI_Cold" e_config.num_epoch = 2 e_config.save_interval = 30 * 60 # 30 minutes data_loader = nli.DataLoader(hp.seq_max, nli_setting.vocab_filename, True) saved = e.train_nli_ex_0(nli_setting, e_config, data_loader, None, False) e.test_acc2(nli_setting, e_config, data_loader, saved)
def train_nli_with_premade(explain_tag): hp = hyperparams.HPBert() e = Experiment(hp) nli_setting = NLI() nli_setting.vocab_size = 30522 nli_setting.vocab_filename = "bert_voca.txt" e_config = ExperimentConfig() e_config.name = "NLIEx_{}".format("Premade_"+explain_tag) e_config.num_epoch = 1 e_config.save_interval = 30 * 60 # 30 minutes e_config.load_names = ['bert'] #, 'cls_dense'] #, 'aux_conflict'] data_loader = nli.DataLoader(hp.seq_max, nli_setting.vocab_filename, True) load_id = ("uncased_L-12_H-768_A-12", 'bert_model.ckpt') e.train_nli_ex_with_premade_data(nli_setting, e_config, data_loader, load_id, explain_tag)
def train_adhoc_fad(): hp = hyperparams.HPFAD() hp.batch_size = 16 e = Experiment(hp) e_config = ExperimentConfig() e_config.name = "Adhoc_{}".format("FAD") e_config.num_epoch = 4 e_config.save_interval = 10 * 60 # 60 minutes e_config.load_names = ['bert'] #, 'reg_dense'] vocab_size = 30522 data_loader = data_sampler.DataLoaderFromFile(hp.batch_size, vocab_size) load_id = ("uncased_L-12_H-768_A-12", 'bert_model.ckpt') #load_id = ("Adhoc_I2", 'model-290') e.train_adhoc2(e_config, data_loader, load_id)
def bert_lm_test(): hp = hyperparams.HPQL() e = Experiment(hp) e_config = ExperimentConfig() e_config.name = "Contrv_{}".format("B") e_config.num_epoch = 4 e_config.save_interval = 30 * 60 # 30 minuteslm_protest e_config.load_names = ['bert', 'cls'] vocab_size = 30522 vocab_filename = "bert_voca.txt" data_loader = ws.DataLoader(hp.seq_max, vocab_filename, vocab_size) load_id = ("uncased_L-12_H-768_A-12", 'bert_model.ckpt') e.bert_lm_pos_neg(e_config, data_loader, load_id)