def __init__(self, dialog, ctx_gen, corpus, sv_config, sys_model, usr_model, rl_config, dialog_eval, ctx_gen_eval): self.dialog = dialog self.ctx_gen = ctx_gen self.corpus = corpus self.sv_config = sv_config self.sys_model = sys_model self.usr_model = usr_model self.rl_config = rl_config self.dialog_eval = dialog_eval self.ctx_gen_eval = ctx_gen_eval # training data for supervised learning train_dial, val_dial, test_dial = self.corpus.get_corpus() self.train_data = DealDataLoaders('Train', train_dial, self.sv_config) self.val_data = DealDataLoaders('Val', val_dial, self.sv_config) self.test_data = DealDataLoaders('Test', test_dial, self.sv_config) # training func for supervised learning self.train_func = train_single_batch # recording func self.record_func = record if self.rl_config.record_freq > 0: self.ppl_exp_file = open(os.path.join(self.rl_config.record_path, 'ppl.tsv'), 'w') self.rl_exp_file = open(os.path.join(self.rl_config.record_path, 'rl.tsv'), 'w') self.learning_exp_file = open(os.path.join(self.rl_config.record_path, 'learning.tsv'), 'w') # evaluation self.validate_func = validate self.evaluator = evaluators.BleuEvaluator('Deal') self.generate_func = generate
def main(config): prepare_dirs_loggers(config, os.path.basename(__file__)) corpus_client = corpora.DailyDialogCorpus(config) dial_corpus = corpus_client.get_corpus() train_dial, valid_dial, test_dial = dial_corpus['train'],\ dial_corpus['valid'],\ dial_corpus['test'] evaluator = evaluators.BleuEvaluator("CornellMovie") # create data loader that feed the deep models train_feed = data_loaders.DailyDialogLoader("Train", train_dial, config) valid_feed = data_loaders.DailyDialogLoader("Valid", valid_dial, config) test_feed = data_loaders.DailyDialogLoader("Test", test_dial, config) model = sent_models.DiVAE(corpus_client, config) if config.forward_only: test_file = os.path.join(config.log_dir, config.load_sess, "{}-test-{}.txt".format(get_time(), config.gen_type)) dump_file = os.path.join(config.log_dir, config.load_sess, "{}-z.pkl".format(get_time())) model_file = os.path.join(config.log_dir, config.load_sess, "model") else: test_file = os.path.join(config.session_dir, "{}-test-{}.txt".format(get_time(), config.gen_type)) dump_file = os.path.join(config.session_dir, "{}-z.pkl".format(get_time())) model_file = os.path.join(config.session_dir, "model") if config.use_gpu: model.cuda() if config.forward_only is False: try: engine.train(model, train_feed, valid_feed, test_feed, config, evaluator, gen=utt_utils.generate) except KeyboardInterrupt: print("Training stopped by keyboard.") config.batch_size = 50 model.load_state_dict(torch.load(model_file)) engine.validate(model, test_feed, config) utt_utils.sweep(model, test_feed, config, num_batch=50) with open(os.path.join(dump_file), "wb") as f: print("Dumping test to {}".format(dump_file)) utt_utils.dump_latent(model, test_feed, config, f, num_batch=None) with open(os.path.join(test_file), "wb") as f: print("Saving test to {}".format(test_file)) utt_utils.generate(model, test_feed, config, evaluator, num_batch=None, dest_f=f)
def main(config): corpus_client = getattr(corpora, config.corpus_client)(config) corpus_client.vocab, corpus_client.rev_vocab, corpus_client.unk_id = load_vocab( config.vocab) prepare_dirs_loggers(config, os.path.basename(__file__)) dial_corpus = corpus_client.get_corpus() train_dial, valid_dial, test_dial = (dial_corpus['train'], dial_corpus['valid'], dial_corpus['test']) evaluator = evaluators.BleuEvaluator("CornellMovie") # create data loader that feed the deep models train_feed = data_loaders.SMDDialogSkipLoader("Train", train_dial, config) valid_feed = data_loaders.SMDDialogSkipLoader("Valid", valid_dial, config) test_feed = data_loaders.SMDDialogSkipLoader("Test", test_dial, config) model = dialog_models.VAE(corpus_client, config) if config.forward_only: test_file = os.path.join( config.log_dir, config.load_sess, "{}-test-{}.txt".format(get_time(), config.gen_type)) dump_file = os.path.join(config.log_dir, config.load_sess, "{}-z.pkl".format(get_time())) model_file = os.path.join(config.log_dir, config.load_sess, "model") else: test_file = os.path.join( config.session_dir, "{}-test-{}.txt".format(get_time(), config.gen_type)) dump_file = os.path.join(config.session_dir, "{}-z.pkl".format(get_time())) model_file = os.path.join(config.session_dir, "model") if config.use_gpu: model.cuda() if config.forward_only is False: try: engine.train(model, train_feed, valid_feed, test_feed, config, evaluator, gen=dialog_utils.generate_vae) except KeyboardInterrupt: print("Training stopped by keyboard.")
def main(config): corpus_client = getattr(corpora, config.corpus_client)(config) corpus_client.vocab, corpus_client.rev_vocab, corpus_client.unk_id = load_vocab( config.vocab) prepare_dirs_loggers(config, os.path.basename(__file__)) dial_corpus = corpus_client.get_corpus() train_dial, valid_dial, test_dial = (dial_corpus['train'], dial_corpus['valid'], dial_corpus['test']) evaluator = evaluators.BleuEvaluator("CornellMovie") # create data loader that feed the deep models train_feed = data_loaders.SMDDialogSkipLoader("Train", train_dial, config) valid_feed = data_loaders.SMDDialogSkipLoader("Valid", valid_dial, config) test_feed = data_loaders.SMDDialogSkipLoader("Test", test_dial, config) model = dialog_models.StED(corpus_client, config) if config.forward_only: test_file = os.path.join( config.log_dir, config.load_sess, "{}-test-{}.txt".format(get_time(), config.gen_type)) dump_file = os.path.join(config.log_dir, config.load_sess, "{}-z.pkl".format(get_time())) model_file = os.path.join(config.log_dir, config.load_sess, "model") else: test_file = os.path.join( config.session_dir, "{}-test-{}.txt".format(get_time(), config.gen_type)) dump_file = os.path.join(config.session_dir, "{}-z.pkl".format(get_time())) model_file = os.path.join(config.session_dir, "model") if config.use_gpu: model.cuda() if not config.forward_only: try: engine.train(model, train_feed, valid_feed, test_feed, config, evaluator, gen=dialog_utils.generate_with_adv) except KeyboardInterrupt: print("Training stopped by keyboard.") config.batch_size = 10 model.load_state_dict(torch.load(model_file)) engine.validate(model, valid_feed, config) engine.validate(model, test_feed, config) dialog_utils.generate_with_adv(model, test_feed, config, None, num_batch=0) selected_clusters = utt_utils.latent_cluster(model, train_feed, config, num_batch=None) selected_outs = dialog_utils.selective_generate(model, test_feed, config, selected_clusters) print(len(selected_outs)) with open(os.path.join(dump_file + '.json'), 'wb') as f: json.dump(selected_clusters, f, indent=2) with open(os.path.join(dump_file + '.out.json'), 'wb') as f: json.dump(selected_outs, f, indent=2) with open(os.path.join(dump_file), "wb") as f: print("Dumping test to {}".format(dump_file)) dialog_utils.dump_latent(model, test_feed, config, f, num_batch=None) with open(os.path.join(test_file), "wb") as f: print("Saving test to {}".format(test_file)) dialog_utils.gen_with_cond(model, test_feed, config, num_batch=None, dest_f=f) with open(os.path.join(test_file + '.txt'), "wb") as f: print("Saving test to {}".format(test_file)) dialog_utils.generate(model, test_feed, config, evaluator, num_batch=None, dest_f=f)
def main(config): prepare_dirs_loggers(config, os.path.basename(__file__)) manualSeed = config.seed random.seed(manualSeed) torch.manual_seed(manualSeed) np.random.seed(manualSeed) sample_shape = config.batch_size, config.state_noise_dim, config.action_noise_dim evaluator = evaluators.BleuEvaluator(os.path.basename(__file__)) train_feed = WoZGanDataLoaders("train", config) valid_feed = WoZGanDataLoaders("val", config) test_feed = WoZGanDataLoaders("test", config) # action2name = load_action2name(config) action2name = None corpus_client = None # model = GanAgent_AutoEncoder(corpus_client, config, action2name) # model = GanAgent_AutoEncoder_Encode(corpus_client, config, action2name) # model = GanAgent_AutoEncoder_State(corpus_client, config, action2name) if config.gan_type == 'wgan': model = WGanAgent_VAE_State(corpus_client, config, action2name) else: model = GanAgent_VAE_State(corpus_client, config, action2name) logger.info(summary(model, show_weights=False)) model.discriminator.apply(weights_init) model.generator.apply(weights_init) model.vae.apply(weights_init) if config.forward_only: test_file = os.path.join( config.log_dir, config.load_sess, "{}-test-{}.txt".format(get_time(), config.gen_type)) dump_file = os.path.join(config.log_dir, config.load_sess, "{}-z.pkl".format(get_time())) model_file = os.path.join(config.log_dir, config.load_sess, "model") else: test_file = os.path.join( config.session_dir, "{}-test-{}.txt".format(get_time(), config.gen_type)) dump_file = os.path.join(config.session_dir, "{}-z.pkl".format(get_time())) model_file = os.path.join(config.session_dir, "model") vocab_file = os.path.join(config.session_dir, "vocab.json") if config.use_gpu: model.cuda() pred_list = [] generator_samples = [] print("Evaluate initial model on Validate set") model.eval() # policy_validate_for_human(model,valid_feed, config, sample_shape) disc_validate(model, valid_feed, config, sample_shape) _, sample_batch = gen_validate(model, valid_feed, config, sample_shape, -1) generator_samples.append([-1, sample_batch]) machine_data, human_data = build_fake_data(model, valid_feed, config, sample_shape) model.train() print("Start VAE training") # # this is for the training of VAE. If you already have a pretrained model, you can skip this step. # if config.forward_only is False: # try: # engine.vae_train(model, train_feed, valid_feed, test_feed, config) # except KeyboardInterrupt: # print("Training stopped by keyboard.") # print("AutoEncoder Training Done ! ") # load_model_vae(model, config) # this is a pretrained vae model, you can load it to the current model. TODO: move path todata_args path = './logs/2019-09-06T10:50:18.034181-mwoz_gan_vae.py' load_model_vae(model, path) print("Start GAN training") if config.forward_only is False: try: engine.gan_train(model, machine_data, train_feed, valid_feed, test_feed, config, evaluator, pred_list, generator_samples) except KeyboardInterrupt: print("Training stopped by keyboard.") print("Reward Model Training Done ! ") print("Saved path: {}".format(model_file))
def main(config): prepare_dirs_loggers(config, os.path.basename(__file__)) manualSeed=config.seed random.seed(manualSeed) torch.manual_seed(manualSeed) np.random.seed(manualSeed) sample_shape = config.batch_size, config.state_noise_dim, config.action_noise_dim evaluator = evaluators.BleuEvaluator(os.path.basename(__file__)) train_feed = WoZGanDataLoaders("train", config) valid_feed = WoZGanDataLoaders("val", config) test_feed = WoZGanDataLoaders("test", config) # action2name = load_action2name(config) action2name = None corpus_client = None if config.gan_type=='gan' and config.input_type=='sat': model = GanAgent_SAT_WoZ(corpus_client, config, action2name) else: raise ValueError("No such GAN types: {}".format(config.gan_type)) logger.info(summary(model, show_weights=False)) model.discriminator.apply(weights_init) model.generator.apply(weights_init) if config.state_type=='rnn': load_context_encoder(model, os.path.join(config.log_dir, config.encoder_sess, "model_lirl")) if config.forward_only: test_file = os.path.join(config.log_dir, config.load_sess, "{}-test-{}.txt".format(get_time(), config.gen_type)) dump_file = os.path.join(config.log_dir, config.load_sess, "{}-z.pkl".format(get_time())) model_file = os.path.join(config.log_dir, config.load_sess, "model") else: test_file = os.path.join(config.session_dir, "{}-test-{}.txt".format(get_time(), config.gen_type)) dump_file = os.path.join(config.session_dir, "{}-z.pkl".format(get_time())) model_file = os.path.join(config.session_dir, "model") vocab_file = os.path.join(config.session_dir, "vocab.json") if config.use_gpu: model.cuda() pred_list = [] generator_samples = [] print("Evaluate initial model on Validate set") model.eval() # policy_validate_for_human(model,valid_feed, config, sample_shape) disc_validate(model, valid_feed, config, sample_shape) _, sample_batch = gen_validate(model,valid_feed, config, sample_shape, -1) generator_samples.append([-1, sample_batch]) machine_data, human_data = build_fake_data(model, valid_feed, config, sample_shape) model.train() print("Start training") if config.forward_only is False: try: engine.gan_train(model, machine_data, train_feed, valid_feed, test_feed, config, evaluator, pred_list, generator_samples) except KeyboardInterrupt: print("Training stopped by keyboard.") # save_data_for_tsne(human_data, machine_data, generator_samples, pred_list, config) print("Training Done ! ") ''' model.load_state_dict(torch.load(model_file)) print("Evaluate final model on Validate set") model.eval() policy_validate_for_human(model,valid_feed, config, sample_shape) disc_validate(model, valid_feed, config, sample_shape) gen_validate(model,valid_feed, config, sample_shape) print("Evaluate final model on Test set") policy_validate_for_human(model,test_feed, config, sample_shape) disc_validate(model, test_feed, config, sample_shape) gen_validate(model,test_feed, config, sample_shape) ''' # dialog_utils.generate_with_adv(model, test_feed, config, None, num_batch=None) # selected_clusters, index_cluster_id_train = utt_utils.latent_cluster(model, train_feed, config, num_batch=None) # _, index_cluster_id_test = utt_utils.latent_cluster(model, test_feed, config, num_batch=None) # _, index_cluster_id_valid = utt_utils.latent_cluster(model, valid_feed, config, num_batch=None) # selected_outs = dialog_utils.selective_generate(model, test_feed, config, selected_clusters) # print(len(selected_outs)) '''
def main(config): prepare_dirs_loggers(config, os.path.basename(__file__)) manualSeed = config.seed random.seed(manualSeed) torch.manual_seed(manualSeed) np.random.seed(manualSeed) sample_shape = config.batch_size, config.state_noise_dim, config.action_noise_dim evaluator = evaluators.BleuEvaluator(os.path.basename(__file__)) train_feed = WoZGanDataLoaders_StateActionEmbed("train", config) valid_feed = WoZGanDataLoaders_StateActionEmbed("val", config) test_feed = WoZGanDataLoaders_StateActionEmbed("test", config) # action2name = load_action2name(config) action2name = None corpus_client = None # model = GanAgent_VAE_StateActioneEmbed(corpus_client, config, action2name) model = GanAgent_StateVaeActionSeg(corpus_client, config, action2name) logger.info(summary(model, show_weights=False)) model.discriminator.apply(weights_init) model.generator.apply(weights_init) model.vae.apply(weights_init) if config.forward_only: test_file = os.path.join( config.log_dir, config.load_sess, "{}-test-{}.txt".format(get_time(), config.gen_type)) dump_file = os.path.join(config.log_dir, config.load_sess, "{}-z.pkl".format(get_time())) model_file = os.path.join(config.log_dir, config.load_sess, "model") else: test_file = os.path.join( config.session_dir, "{}-test-{}.txt".format(get_time(), config.gen_type)) dump_file = os.path.join(config.session_dir, "{}-z.pkl".format(get_time())) model_file = os.path.join(config.session_dir, "model") vocab_file = os.path.join(config.session_dir, "vocab.json") if config.use_gpu: model.cuda() pred_list = [] generator_samples = [] print("Evaluate initial model on Validate set") model.eval() # policy_validate_for_human(model,valid_feed, config, sample_shape) disc_validate(model, valid_feed, config, sample_shape) _, sample_batch = gen_validate(model, valid_feed, config, sample_shape, -1) generator_samples.append([-1, sample_batch]) machine_data, human_data = build_fake_data(model, valid_feed, config, sample_shape) model.train() print("Start training") if config.forward_only is False: try: engine.vae_train(model, train_feed, valid_feed, test_feed, config) except KeyboardInterrupt: print("Training stopped by keyboard.") print("AutoEncoder Training Done ! ") load_model_vae(model, config) # path='logs/2019-09-18T12:20:26.063708-mwoz_gan_vae_StateActionEmbed.py' # this is embed version # path='logs/2019-09-18T12:24:45.517636-mwoz_gan_vae_StateActionEmbed.py' # this is embed_merged version # path='logs/2019-09-18T17:21:35.420069-mwoz_gan_vae_StateActionEmbed.py' # this is state_vae action_seg version, without hotel domain # load_model_vae(model, path) if config.forward_only is False: try: engine.gan_train(model, machine_data, train_feed, valid_feed, test_feed, config, evaluator, pred_list, generator_samples) except KeyboardInterrupt: print("Training stopped by keyboard.") print("Reward Model Training Done ! ") print("Saved path: {}".format(model_file))