def verify(opt, printargs=None, print_parser=None): if opt['datatype'] == 'train': print("[ note: changing datatype from train to train:ordered ]") opt['datatype'] = 'train:ordered' # create repeat label agent and assign it to the specified task agent = RepeatLabelAgent(opt) world = create_task(opt, agent) log_every_n_secs = opt.get('log_every_n_secs', -1) if log_every_n_secs <= 0: log_every_n_secs = float('inf') log_time = TimeLogger() counts = {} counts['missing_text'] = 0 counts['missing_labels'] = 0 counts['missing_label_candidates'] = 0 counts['empty_string_label_candidates'] = 0 counts['label_candidates_with_missing_label'] = 0 counts['did_not_return_message'] = 0 # Show some example dialogs. while not world.epoch_done(): world.parley() act = world.acts[0] if not isinstance(act, Message): counts['did_not_return_message'] += 1 if 'text' not in act and 'image' not in act: warn("warning: missing text field:\n", act, opt) counts['missing_text'] += 1 if 'labels' not in act and 'eval_labels' not in act: warn("warning: missing labels/eval_labels field:\n", act, opt) counts['missing_labels'] += 1 else: if 'label_candidates' not in act: counts['missing_label_candidates'] += 1 else: labels = act.get('labels', act.get('eval_labels')) is_label_cand = {} for l in labels: is_label_cand[l] = False for c in act['label_candidates']: if c == '': warn("warning: empty string label_candidate:\n", act, opt) counts['empty_string_label_candidates'] += 1 if c in is_label_cand: if is_label_cand[c] is True: warn( "warning: label mentioned twice in candidate_labels:\n", act, opt, ) is_label_cand[c] = True for _, has in is_label_cand.items(): if has is False: warn("warning: label missing in candidate_labels:\n", act, opt) counts['label_candidates_with_missing_label'] += 1 if log_time.time() > log_every_n_secs: text, log = report(world, counts, log_time) if print_parser: print(text) try: # print dataset size if available print('[ loaded {} episodes with a total of {} examples ]'.format( world.num_episodes(), world.num_examples())) except Exception: pass return report(world, counts, log_time)
def verify(opt, printargs=None, print_parser=None): if opt['datatype'] == 'train': print("[ note: changing datatype from train to train:ordered ]") opt['datatype'] = 'train:ordered' # create repeat label agent and assign it to the specified task agent = RepeatLabelAgent(opt) world = create_task(opt, agent) log_every_n_secs = opt.get('log_every_n_secs', -1) if log_every_n_secs <= 0: log_every_n_secs = float('inf') log_time = TimeLogger() dictionary = DictionaryAgent(opt) ignore_tokens = opt.get('ignore_tokens').split(',') counts = {} for t in {'input', 'labels', 'both'}: counts['tokens_in_' + t] = 0 counts['utterances_in_' + t] = 0 counts['avg_utterance_length_in_' + t] = 0 counts['unique_tokens_in_' + t] = 0 counts['unique_utterances_in_' + t] = 0 # for counting the stats.. counts['token_dict_' + t] = {} counts['utterance_dict_' + t] = {} def tokenize(txt): return dictionary.tokenize(txt) def keep_token(t): for s in ignore_tokens: if s != '' and s in t: return False return True # max number of examples to evaluate max_cnt = opt['num_examples'] if opt['num_examples'] > 0 else float('inf') cnt = 0 # Show some example dialogs. while not world.epoch_done() and cnt < max_cnt: cnt += opt.get('batchsize', 1) world.parley() act = world.get_acts()[opt.get('agent')] for itype in {'input', 'labels'}: if itype == 'input': if opt.get('new_line_new_utt'): txts = act.get('text').split('\n') else: txts = [act.get('text')] else: txts = act.get('labels', act.get('eval_labels', [''])) for txt in txts: tokens = tokenize(txt) retxt = [] for t in tokens: if keep_token(t): retxt.append(t) counts['tokens_in_' + itype] += len(retxt) counts['tokens_in_' + 'both'] += len(retxt) counts['utterances_in_' + itype] += 1 counts['utterances_in_' + 'both'] += 1 counts['avg_utterance_length_in_' + itype] = ( counts['tokens_in_' + itype] / counts['utterances_in_' + itype] ) counts['avg_utterance_length_in_' + 'both'] = ( counts['tokens_in_' + 'both'] / counts['utterances_in_' + 'both'] ) for t in retxt: if t not in counts['token_dict_' + itype]: counts['unique_tokens_in_' + itype] += 1 counts['token_dict_' + itype][t] = True if t not in counts['token_dict_' + 'both']: counts['unique_tokens_in_' + 'both'] += 1 counts['token_dict_' + 'both'][t] = True retxt = ' '.join(retxt) if retxt not in counts['utterance_dict_' + itype]: counts['unique_utterances_in_' + itype] += 1 counts['utterance_dict_' + itype][retxt] = True if retxt not in counts['utterance_dict_' + 'both']: counts['unique_utterances_in_' + 'both'] += 1 counts['utterance_dict_' + 'both'][retxt] = True if log_time.time() > log_every_n_secs: text, log = report(world, counts, log_time) if print_parser: print(text) try: # print dataset size if available print( '[ loaded {} episodes with a total of {} examples ]'.format( world.num_episodes(), world.num_examples() ) ) except Exception: pass return report(world, counts, log_time)
def build_dict(opt, skip_if_built=False): if isinstance(opt, ParlaiParser): logging.error('Should be passed opt not Parser') opt = opt.parse_args() if not opt.get('dict_file'): logging.error( 'Tried to build dictionary but `--dict-file` is not set. Set ' 'this param so the dictionary can be saved.' ) return if skip_if_built and PathManager.exists(opt['dict_file']): # Dictionary already built, skip all loading or setup logging.debug("dictionary already built.") return None if opt.get('dict_class'): # Custom dictionary class dictionary = str2class(opt['dict_class'])(opt) else: # Default dictionary class dictionary = DictionaryAgent(opt) if PathManager.exists(opt['dict_file']) or ( hasattr(dictionary, 'is_prebuilt') and dictionary.is_prebuilt() ): # Dictionary already built, return loaded dictionary agent logging.debug("dictionary already built.") return dictionary if is_distributed(): raise ValueError('Dictionaries should be pre-built before distributed train.') ordered_opt = copy.deepcopy(opt) cnt = 0 # we use train set to build dictionary ordered_opt['batchsize'] = 1 # Set this to none so that image features are not calculated when Teacher is # instantiated while building the dict ordered_opt['image_mode'] = 'no_image_model' ordered_opt.log() datatypes = ['train:ordered:stream'] if opt.get('dict_include_valid'): datatypes.append('valid:stream') if opt.get('dict_include_test'): datatypes.append('test:stream') cnt = 0 for dt in datatypes: ordered_opt['datatype'] = dt world_dict = create_task(ordered_opt, dictionary) # pass examples to dictionary log_time = TimeLogger() total = world_dict.num_examples() if opt['dict_maxexs'] >= 0: total = min(total, opt['dict_maxexs']) log_every_n_secs = opt.get('log_every_n_secs', None) if log_every_n_secs: pbar = tqdm.tqdm( total=total, desc='Building dictionary', unit='ex', unit_scale=True ) else: pbar = None while not world_dict.epoch_done(): cnt += 1 if cnt > opt['dict_maxexs'] and opt['dict_maxexs'] >= 0: logging.info('Processed {} exs, moving on.'.format(opt['dict_maxexs'])) # don't wait too long... break world_dict.parley() if pbar: pbar.update(1) if pbar: pbar.close() dictionary.save(opt['dict_file'], sort=True) logging.info( f'dictionary built with {len(dictionary)} tokens ' f'in {log_time.total_time():.1f}s' ) return dictionary
def detect(opt, printargs=None, print_parser=None): """ Checks a task for offensive language. """ if print_parser is not None: if print_parser is True and isinstance(opt, ParlaiParser): print_parser = opt elif print_parser is False: print_parser = None random.seed(42) # Create model and assign it to the specified task agent = create_agent(opt, requireModelExists=True) world = create_task(opt, agent) if opt['safety'] == 'string_matcher' or opt['safety'] == 'all': offensive_string_matcher = OffensiveStringMatcher() if opt['safety'] == 'classifier' or opt['safety'] == 'all': offensive_classifier = OffensiveLanguageClassifier() if print_parser: # Show arguments after loading model print_parser.opt = agent.opt print_parser.print_args() log_every_n_secs = opt.get('log_every_n_secs', -1) if log_every_n_secs <= 0: log_every_n_secs = float('inf') log_time = TimeLogger() stats = { 'bad_words': [], 'bad_words_cnt': 0, 'string_offensive': 0, 'classifier_offensive': 0, 'total_offensive': 0, 'total': 0, } def report(world, stats): report = world.report() log = { 'word_offenses': stats['bad_words_cnt'], 'classifier_offenses%': 100 * (stats['classifier_offensive'] / stats['total']), 'string_offenses%': 100 * (stats['string_offensive'] / stats['total']), 'total_offenses%': 100 * (stats['total_offensive'] / stats['total']), } text, log = log_time.log(report['exs'], world.num_examples(), log) logging.info(text) def classify(text, stats): offensive = False stats['total'] += 1 if opt['safety'] == 'string_matcher' or opt['safety'] == 'all': bad_words = offensive_string_matcher.contains_offensive_language( text) if bad_words: stats['string_offensive'] += 1 offensive = True stats['bad_words'].append(bad_words) if opt['safety'] == 'classifier' or opt['safety'] == 'all': if text in offensive_classifier: stats['classifier_offensive'] += 1 offensive = True if offensive: stats['total_offensive'] += 1 while not world.epoch_done(): world.parley() stats['bad_words'] = [] for a in world.acts: text = a.get('text', '') classify(text, stats) labels = a.get('labels', a.get('eval_labels', '')) for l in labels: classify(l, stats) if len(stats['bad_words']) > 0 and opt['display_examples']: logging.info(world.display()) logging.info("Offensive words detected: {}".format(', '.join( stats['bad_words']))) stats['bad_words_cnt'] += len(stats['bad_words']) if log_time.time() > log_every_n_secs: report(world, stats) if world.epoch_done(): logging.info("epoch done") report(world, stats) return world.report()
def learn_arora(opt): """ Go through ConvAI2 data and collect word counts, thus compute the unigram probability distribution. Use those probs to compute weighted sentence embeddings for all utterances, thus compute first principal component. Save all info to arora.pkl file. """ arora_file = os.path.join(opt['datapath'], 'controllable_dialogue', 'arora.pkl') opt['task'] = 'fromfile:parlaiformat' opt['log_every_n_secs'] = 2 print('Getting word counts from ConvAI2 train set...') opt['datatype'] = 'train:ordered' opt['fromfile_datapath'] = os.path.join(opt['datapath'], 'controllable_dialogue', 'ConvAI2_parlaiformat', 'train.txt') # Do include inputs because ConvAI2 train set reverses every convo: word_counter_train, total_count_train, all_utts_train = get_word_counts( opt, count_inputs=False) print('Getting word counts from ConvAI2 val set...') opt['datatype'] = 'valid' opt['fromfile_datapath'] = os.path.join(opt['datapath'], 'controllable_dialogue', 'ConvAI2_parlaiformat', 'valid.txt') # Don't include inputs because ConvAI2 val set doesn't reverses convos: word_counter_valid, total_count_valid, all_utts_valid = get_word_counts( opt, count_inputs=True) # Merge word counts word_counter = word_counter_train for word, count in word_counter_valid.items(): word_counter[word] += count total_count = total_count_train + total_count_valid # Merge all_utts all_utts = all_utts_train + all_utts_valid # Compute unigram prob for every word print("Computing unigram probs for all words...") word2prob = {w: c / total_count for w, c in word_counter.items()} # Settings for sentence embedder arora_a = 0.0001 glove_name = '840B' glove_dim = 300 glove_cache = modelzoo_path(opt['datapath'], 'models:glove_vectors') # Embed every sentence, without removing first singular value print('Embedding all sentences...') sent_embedder = SentenceEmbedder( word2prob, arora_a, glove_name, glove_dim, first_sv=None, glove_cache=glove_cache, ) utt_embs = [] log_timer = TimeLogger() for n, utt in enumerate(all_utts): utt_emb = sent_embedder.embed_sent(utt.split(), rem_first_sv=False) utt_embs.append(utt_emb) if log_timer.time() > opt['log_every_n_secs']: text, _log = log_timer.log(n, len(all_utts)) print(text) # Use SVD to calculate singular vector # https://docs.scipy.org/doc/numpy-1.13.0/reference/generated/numpy.linalg.svd.html print('Calculating SVD...') utt_embs = np.stack(utt_embs, axis=0) # shape (num_embs, glove_dim) U, s, V = np.linalg.svd(utt_embs, full_matrices=False) first_sv = V[0, :] # first row of V. shape (glove_dim) # Remove singular vector from all embs to get complete Arora-style sent embs print('Removing singular vec from all sentence embeddings...') utt_embs_adj = [ remove_first_sv(torch.Tensor(emb), torch.Tensor(first_sv)).numpy() for emb in utt_embs ] # list of np arrays shape (glove_dim) # Make dict mapping ConvAI2 dataset utterances to Arora sent emb # We save this to file for convenience (e.g. if you want to inspect) utt2emb = {utt: emb for (utt, emb) in zip(all_utts, utt_embs_adj)} # Save unigram distribution, first singular value, hyperparameter value for a, # info about GloVe vectors used, and full dict of utt->emb to file print("Saving Arora embedding info to %s..." % arora_file) with open(arora_file, "wb") as f: pickle.dump( { 'word2prob': word2prob, # dict: string to float between 0 and 1 'first_sv': first_sv, # np array shape (glove_dim) 'arora_a': arora_a, # float, 0.0001 'glove_name': glove_name, # string, '840B' 'glove_dim': glove_dim, # int, 300 'utt2emb': utt2emb, # dict: string to np array shape (glove_dim) }, f, )
def eval_wordstat(opt, print_parser=None): """ Evaluates a model. :param opt: tells the evaluation function how to run :param print_parser: if provided, prints the options that are set within the model after loading the model """ random.seed(42) # Create model and assign it to the specified task agent = create_agent(opt, requireModelExists=True) world = create_task(opt, agent) if opt.get('external_dict'): print('[ Using external dictionary from: {} ]'.format( opt['external_dict'])) dict_opt = copy.deepcopy(opt) dict_opt['dict_file'] = opt['external_dict'] dictionary = DictionaryAgent(dict_opt) else: print('[ Using model bundled dictionary ]') dictionary = agent.dict batch_size = opt['batchsize'] if print_parser: # Show arguments after loading model print_parser.opt = agent.opt print_parser.print_args() log_every_n_secs = opt.get('log_every_n_secs', -1) if log_every_n_secs <= 0: log_every_n_secs = float('inf') log_time = TimeLogger() cnt = 0 word_statistics = { 'mean_wlength': [], 'mean_clength': [], 'freqs_cnt': Counter(), 'word_cnt': 0, 'pred_list': [], 'pure_pred_list': [], 'context_list': [], 'unique_words': set(), } bins = [int(i) for i in opt['freq_bins'].split(',')] def process_prediction(prediction, word_statistics): normalized = normalize_answer(prediction) word_statistics['pred_list'].append(normalized) freqs, _cnt, wlength, clength = get_word_stats(prediction, dictionary, bins=bins) word_statistics['word_cnt'] += _cnt word_statistics['mean_wlength'].append(wlength) word_statistics['mean_clength'].append(clength) word_statistics['freqs_cnt'] += Counter(freqs) word_statistics['unique_words'] |= set(normalized.split(" ")) return word_statistics while not world.epoch_done(): world.parley() if batch_size == 1: cnt += 1 prediction = world.acts[-1]['text'] word_statistics['context_list'].append(world.acts[0]['text']) word_statistics['pure_pred_list'].append(prediction) word_statistics = process_prediction(prediction, word_statistics) else: for w in world.worlds: try: if 'text' not in w.acts[-1]: continue prediction = w.acts[-1]['text'] word_statistics['context_list'].append(w.acts[0]['text']) word_statistics['pure_pred_list'].append(prediction) except IndexError: continue cnt += 1 word_statistics = process_prediction(prediction, word_statistics) if log_time.time() > log_every_n_secs: report = world.report() text, report = log_time.log(report['exs'], world.num_examples(), report) print(text) stat_str = 'total_words: {}, '.format(word_statistics['word_cnt']) stat_str += ', '.join([ '<{}:{} ({:.{prec}f}%)'.format( b, word_statistics['freqs_cnt'].get(b, 0), (word_statistics['freqs_cnt'].get(b, 0) / word_statistics['word_cnt']) * 100, prec=2, ) for b in bins ]) print("Word statistics: {}, avg_word_length: {:.{prec}f}, " "avg_char_length: {:.{prec}f}".format( stat_str, numpy.array(word_statistics['mean_wlength']).mean(), numpy.array(word_statistics['mean_clength']).mean(), prec=2, )) if opt['num_examples'] > 0 and cnt >= opt['num_examples']: break if world.epoch_done(): print("EPOCH DONE") if opt['compute_unique'] is True: unique_list = [] cntr = Counter(word_statistics['pred_list']) for k, v in cntr.items(): if v == 1: unique_list.append(k) print("Unique responses: {:.{prec}f}%".format( len(unique_list) / len(word_statistics['pred_list']) * 100, prec=2)) print("Total unique tokens:", len(word_statistics['unique_words'])) if opt['dump_predictions_path'] is not None: with open(opt['dump_predictions_path'], 'w') as f: f.writelines([ 'CONTEXT: {}\nPREDICTION:{}\n\n'.format(c, p) for c, p in zip( word_statistics['context_list'], word_statistics['pure_pred_list'], ) ]) if opt['compute_unique'] is True: with open(opt['dump_predictions_path'] + '_unique', 'w') as f: f.writelines(['{}\n'.format(i) for i in unique_list]) stat_str = 'total_words: {}, '.format(word_statistics['word_cnt']) stat_str += ', '.join([ '<{}:{} ({:.{prec}f}%)'.format( b, word_statistics['freqs_cnt'].get(b, 0), (word_statistics['freqs_cnt'].get(b, 0) / word_statistics['word_cnt']) * 100, prec=2, ) for b in bins ]) print("Word statistics: {}, avg_word_length: {:.{prec}f}, " "avg_char_length: {:.{prec}f}".format( stat_str, numpy.array(word_statistics['mean_wlength']).mean(), numpy.array(word_statistics['mean_clength']).mean(), prec=2, )) report = world.report() print(report) return report
def self_chat(opt): random.seed(opt['seed']) partner = opt['partner_model_file'] partner_opt_file = opt.get('partner_opt_file') # Create agents agent1 = create_agent(opt, requireModelExists=True) agent1.opt.log("Agent 1 Opt") if partner is None: # Self chat with same model agent2 = agent1.clone() else: # Self chat with different models if partner_opt_file: print(f"WARNING: Loading override opts from: {partner_opt_file}") with PathManager.open(partner_opt_file) as f: partner_opt = json.load(f) else: partner_opt = {} partner_opt['interactive_mode'] = opt.get('interactive_mode', True) print( f"WARNING: Setting partner interactive mode to: {partner_opt['interactive_mode']}" ) agent2 = create_agent_from_model_file(partner, partner_opt) agent2.opt.log("Agent 2 Opt") # Set IDs agent1.id = agent1.id + "_1" agent2.id = agent2.id + "_2" model_id = agent1.id + "_" + agent2.id world = create_task(opt, user_agents=[agent1, agent2]) # Set up world logging logger = WorldLogger(opt) log_time = TimeLogger() # Run some self chats. for i in range(opt['num_self_chats']): _run_self_chat_episode(opt, world, logger) report = world.report() text, report = log_time.log(i + 1, opt['num_self_chats'], report) logging.info(text) # Save chats if opt['outfile'] is None: outfile = '/tmp/{}_selfchat'.format(model_id) else: outfile = opt['outfile'] if opt['save_format'] == 'conversations' and hasattr(world, 'write'): # use self chat specific world to write conversation # this might be useful for logging extra contextual # information (like personas) world.write(logger, outfile) else: # use default logger write function logger.write(outfile, world, opt['save_format']) return logger.get_logs()
def self_chat(opt, print_parser=None): if print_parser is not None: if print_parser is True and isinstance(opt, ParlaiParser): print_parser = opt elif print_parser is False: print_parser = None if isinstance(opt, ParlaiParser): print( '[ Deprecated Warning: self_chat should be passed opt not Parser ]' ) opt = opt.parse_args() random.seed(opt['seed']) # Create models agent1 = create_agent(opt, requireModelExists=True) agent2 = agent1.clone() if hasattr(agent2, 'id'): agent2.id = agent2.id + "2" # Check for `selfchat` in the task name if 'selfchat' not in opt['task']: warn_once( 'You are using self chat with task {}. '.format(opt['task']) + 'If your task has an existing self chat world, then run with ' '-t {}:selfchat'.format(opt['task'])) world = create_task(opt, [agent1, agent2]) if print_parser: # Show arguments after loading model print_parser.opt = agent1.opt print_parser.print_args() # set up logging log_every_n_secs = opt.get('log_every_n_secs', -1) if log_every_n_secs <= 0: log_every_n_secs = float('inf') log_time = TimeLogger() logger = WorldLogger(opt) # Run some self chats. max_cnt = int(opt['num_examples'] * opt.get('selfchat_max_turns') / opt.get('batchsize')) cnt = 0 for _ in tqdm.trange(max_cnt): cnt += opt.get('batchsize', 1) world.parley() logger.log(world) if opt.get('display_examples'): print(world.display()) if log_time.time() > log_every_n_secs: text = log_time.log(cnt, max_cnt) print(text) if opt.get('display_examples'): print('-- end of episode --') logger.reset_world() # flush last episode indent = opt['indent'] if opt['indent'] >= 0 else None logger.write(opt['outfile'], opt['format'], indent=indent) return logger.get_logs()
def dump_data(opt): """ Dump task data to ACUTE-Eval. """ # create repeat label agent and assign it to the specified task agent = RepeatLabelAgent(opt) world = create_task(opt, agent) task = opt.get('task') speaker_0_id = opt.get('speaker_0_id') or f'{task}_as_human' speaker_1_id = opt.get('speaker_1_id') or f'{task}_as_model' if opt['outfile'] is None: outfile = tempfile.mkstemp(prefix='{}_{}_'.format( opt['task'], opt['datatype']), suffix='.txt')[1] else: outfile = opt['outfile'] num_episodes = (world.num_episodes() if opt['num_episodes'] == -1 else min( opt['num_episodes'], world.num_episodes())) log_timer = TimeLogger() print(f'[ starting to convert, saving output to {outfile} ]') dialogues = [] for _ in range(num_episodes): episode = [] episode_done = False while not episode_done: world.parley() acts = world.get_acts() text = acts[0].get('text') split_text = text.split('\n') label = random.choice(acts[0].get('labels', acts[0].pop('eval_labels', None))) if not episode and opt.get('prepended_context'): # first turn context = split_text[:-1] text = split_text[-1] context_turn = [{ 'text': context, 'episode_done': False, 'id': 'context' } for _ in range(2)] episode.append(context_turn) turn = [ { 'text': text, 'episode_done': False, 'id': speaker_0_id }, { 'text': label, 'episode_done': False, 'id': speaker_1_id }, ] episode.append(turn) if acts[0].get('episode_done', False): episode[-1][-1]['episode_done'] = True episode_done = True dialogues.append(episode) if log_timer.time() > opt['log_every_n_secs']: text, _log = log_timer.log(world.total_parleys, world.num_examples()) print(text) if world.epoch_done(): break Conversations.save_conversations(dialogues, outfile, opt)
def verify(opt): if opt['datatype'] == 'train': logging.warn('changing datatype from train to train:ordered') opt['datatype'] = 'train:ordered' # create repeat label agent and assign it to the specified task opt['fixed_response'] = None agent = FixedResponseAgent(opt) world = create_task(opt, agent) opt.log() log_every_n_secs = opt.get('log_every_n_secs', -1) if log_every_n_secs <= 0: log_every_n_secs = float('inf') log_time = TimeLogger() dictionary = DictionaryAgent(opt) ignore_tokens = opt.get('ignore_tokens').split(',') counts = {} for t in {'input', 'labels', 'both'}: counts[f'{t}/tokens'] = 0 counts[f'{t}/utterances'] = 0 counts[f'{t}/avg_utterance_length'] = None counts[f'{t}/unique_tokens'] = 0 counts[f'{t}/unique_utterances'] = 0 # for counting the stats.. counts[f'{t}/token_dict'] = {} counts[f'{t}/utterance_dict'] = {} def tokenize(txt): return dictionary.tokenize(txt) def keep_token(t): for s in ignore_tokens: if s != '' and s in t: return False return True # max number of examples to evaluate max_cnt = opt['num_examples'] if opt['num_examples'] > 0 else float('inf') cnt = 0 # Show some example dialogs. while not world.epoch_done() and world.total_exs < max_cnt: world.parley() act = world.get_acts()[opt.get('agent')] for itype in {'input', 'labels'}: if itype == 'input': if opt.get('new_line_new_utt'): txts = act.get('text').split('\n') else: txts = [act.get('text')] else: txts = act.get('labels', act.get('eval_labels', [''])) for txt in txts: tokens = tokenize(txt) retxt = [t for t in tokens if keep_token(t)] counts[f'{itype}/tokens'] += len(retxt) counts['both/tokens'] += len(retxt) counts[f'{itype}/utterances'] += 1 counts['both/utterances'] += 1 counts[f'{itype}/avg_utterance_length'] += AverageMetric( len(retxt), 1) counts[f'both/avg_utterance_length'] += AverageMetric( len(retxt), 1) for t in retxt: if t not in counts[f'{itype}/token_dict']: counts[f'{itype}/unique_tokens'] += 1 counts[f'{itype}/token_dict'][t] = True if t not in counts['both/token_dict']: counts['both/unique_tokens'] += 1 counts['both/token_dict'][t] = True retxt = ' '.join(retxt) if retxt not in counts[f'{itype}/utterance_dict']: counts[f'{itype}/unique_utterances'] += 1 counts[f'{itype}/utterance_dict'][retxt] = True if retxt not in counts['both/utterance_dict']: counts['both/unique_utterances'] += 1 counts['both/utterance_dict'][retxt] = True if log_time.time() > log_every_n_secs: report = _report(world, counts) cnt = report.pop('exs') text, log = log_time.log(cnt, world.num_examples(), report) logging.info(text) try: # print dataset size if available logging.info(f'loaded {world.num_episodes()} episodes with a total ' f'of {world.num_examples()} examples') except AttributeError: pass retval = _report(world, counts) retval.pop('exs') return retval
def build_dict(opt, skip_if_built=False): print("||||||||||||||||||||| -==================") if isinstance(opt, ParlaiParser): print('[ Deprecated Warning: should be passed opt not Parser ]') opt = opt.parse_args() if not opt.get('dict_file'): print('Tried to build dictionary but `--dict-file` is not set. Set ' + 'this param so the dictionary can be saved.') return if skip_if_built and os.path.isfile(opt['dict_file']): # Dictionary already built, skip all loading or setup print("[ dictionary already built .]") return None if is_distributed(): raise ValueError( 'Dictionaries should be pre-built before distributed train.') if opt.get('dict_class'): # Custom dictionary class print("\tdict_class: " + opt['dict_class']) dictionary = str2class(opt['dict_class'])(opt) else: # Default dictionary class dictionary = DictionaryAgent(opt) print("\tdict_class: " + type(dictionary).__name__) if os.path.isfile(opt['dict_file']): # Dictionary already built, return loaded dictionary agent print("[ dictionary already built .]") return dictionary ordered_opt = copy.deepcopy(opt) cnt = 0 # we use train set to build dictionary ordered_opt['numthreads'] = 1 ordered_opt['batchsize'] = 1 # Set this to none so that image features are not calculated when Teacher is # instantiated while building the dict # TODO: change 'none' to 'no_image_model' ordered_opt['image_mode'] = 'none' ordered_opt['pytorch_teacher_batch_sort'] = False if ordered_opt['task'] == 'pytorch_teacher' or not ordered_opt['task']: pytorch_teacher_task = ordered_opt.get('pytorch_teacher_task', '') if pytorch_teacher_task != '': ordered_opt['task'] = pytorch_teacher_task datatypes = ['train:ordered:stream'] if opt.get('dict_include_valid'): datatypes.append('valid:stream') if opt.get('dict_include_test'): datatypes.append('test:stream') cnt = 0 print("|||||||||||||||||| datatypes : " + str(datatypes)) for dt in datatypes: ordered_opt['datatype'] = dt world_dict = create_task(ordered_opt, dictionary) # pass examples to dictionary print('[ running dictionary over data.. ]') log_time = TimeLogger() total = world_dict.num_examples() if opt['dict_maxexs'] >= 0: total = min(total, opt['dict_maxexs']) log_every_n_secs = opt.get('log_every_n_secs', None) if log_every_n_secs: pbar = tqdm.tqdm(total=total, desc='Building dictionary', unit='ex', unit_scale=True) else: pbar = None while not world_dict.epoch_done(): cnt += 1 if cnt > opt['dict_maxexs'] and opt['dict_maxexs'] >= 0: print('Processed {} exs, moving on.'.format( opt['dict_maxexs'])) # don't wait too long... break world_dict.parley() if pbar: pbar.update(1) if pbar: pbar.close() dictionary.save(opt['dict_file'], sort=True) print('[ dictionary built with {} tokens in {}s ]'.format( len(dictionary), round(log_time.total_time(), 2))) return dictionary
def eval_wordstat(opt): """ Evaluates a model. :param opt: tells the evaluation function how to run """ random.seed(42) # Setup control information initialize_control_information(opt) # Create model and assign it to the specified task agent = create_agent(opt, requireModelExists=True) world = create_task(opt, agent) if opt.get('external_dict'): print('[ Using external dictionary from: {} ]'.format( opt['external_dict'])) dict_opt = copy.deepcopy(opt) dict_opt['dict_file'] = opt['external_dict'] dictionary = DictionaryAgent(dict_opt) else: print('[ Using model bundled dictionary ]') dictionary = agent.dict batch_size = opt['batchsize'] log_every_n_secs = opt.get('log_every_n_secs', -1) if log_every_n_secs <= 0: log_every_n_secs = float('inf') log_time = TimeLogger() data = {} # This will be written to the output json file data['opt'] = agent.opt # Save the opt to json # Determine the output filename if opt['gold_response']: # Special output file for gold response model_dir, _ = os.path.split(opt.get('model_file')) outfile = os.path.join(model_dir, 'goldresponse') if opt['use_reply'] != 'label': raise ValueError( 'You should set --use-reply label (not --use-reply model) ' 'when measuring goldresponse stats') else: outfile = "%s.%s.%s.%s" % ( opt.get('model_file'), opt.get('datatype'), "use%sreply" % agent.opt['use_reply'], "beam%i" % agent.opt['beam_size'], ) if agent.opt['beam_size'] > 1: outfile += ".beamminnbest%i" % agent.opt['beam_min_n_best'] if len(agent.control_settings) > 0: outfile += ".setcontrols:" + "_".join([ "%s%s" % (c, str(agent.control_settings[c]['set_value'])) for c in sorted(agent.control_settings.keys()) ]) if agent.opt['beam_reorder'] not in ['none', False]: outfile += ".beamreorder_%s" % agent.opt['beam_reorder'] if len(agent.wd_features) > 0: sorted_bfw = sorted(list(zip(agent.wd_features, agent.wd_wts)), key=lambda x: x[0]) outfile += ".WDfeatures:" + "_".join( ["%s%s" % (f, str(w)) for f, w in sorted_bfw]) if opt['num_examples'] != -1: outfile += ".numex%i" % opt['num_examples'] outfile += ".wordstats.json" print("\nOutfile: %s\n" % outfile) cnt = 0 word_statistics = { 'mean_wlength': [], # list of length (in words) of utterances 'mean_clength': [], # list of length (in chars) of utterances 'freqs_cnt': Counter(), # Counter for word frequencies, bucketed 'word_cnt': 0, # total number of words in all utterances 'pred_list': [], # list of generated utterances after applying normalize_answer 'pure_pred_list': [], # list of generated utterances 'context_list': [], # list of text inputs (persona and conversation history) } bins = [int(i) for i in opt['freq_bins'].split(',')] # This dictionary records all the sentence-level controllable attributes # For each attribute, we have a list of all the values sent_attrs = {attr: [] for attr in ATTR2SENTSCOREFN.keys()} # str to list of floats # histories will be a list of ConvAI2History objects histories = [] def process_prediction(prediction, word_statistics): word_statistics['pred_list'].append(normalize_answer(prediction)) freqs, _cnt, wlength, clength = get_word_stats(prediction, dictionary, bins=bins) word_statistics['word_cnt'] += _cnt word_statistics['mean_wlength'].append(wlength) word_statistics['mean_clength'].append(clength) word_statistics['freqs_cnt'] += Counter(freqs) return word_statistics t0 = time.time() while not world.epoch_done(): world.parley() # orig eval_wordstat.py handles bsz=1 but for simplicity we assume bsz>1 assert batch_size != 1 for w in world.worlds: try: try: response_act = w.acts[-1] prediction = response_act['text'] except KeyError: continue if opt['gold_response']: # If we're measuring gold response, use eval_label as prediction prediction = w.acts[0]['eval_labels'][0] response_act = {'text': prediction} word_statistics['context_list'].append(w.acts[0]['text']) word_statistics['pure_pred_list'].append(prediction) except IndexError: continue cnt += 1 word_statistics = process_prediction(prediction, word_statistics) # Compute and record sentence-level attributes history = ConvAI2History(w.acts[0]['text']) histories.append(history) sent_attrs = update_sent_attr_stats(sent_attrs, history, prediction) # Periodically log some info if log_time.time() > log_every_n_secs: report = world.report() text, report = log_time.log(report['exs'], world.num_examples(), report) print(text) if opt['num_examples'] > 0 and cnt >= opt['num_examples']: break if world.epoch_done(): print("EPOCH DONE") print("Time to process %i examples: %f seconds" % (cnt, time.time() - t0)) # Compute percent unique # Note this is w.r.t. normalized pred_list not original pure_pred_list unique_list = [] cntr = Counter(word_statistics['pred_list']) for k, v in cntr.items(): if v == 1: unique_list.append(k) unique_percent = len(unique_list) / len(word_statistics['pred_list']) * 100 # Print a final report report = world.report() if opt['gold_response']: report['ppl'] = 0.0 # For gold responses, overwrite the perplexity print(report) # Put all information in data dict data[ 'unique_percent'] = unique_percent # percent of all responses that are unique data[ 'word_statistics'] = word_statistics # word stats, as in orig eval_wordstat data['report'] = report # the final report data['histories'] = [(hist.persona_lines, hist.partner_utts, hist.own_utts) for hist in histories] # history for each example data[ 'sent_attrs'] = sent_attrs # all sentence attribute values for responses # Write data to outfile print("Writing to %s..." % outfile) with PathManager.open(outfile, 'w') as f: json.dump(data, f)
def _eval_single_world(opt, agent, task): logging.info( f'Evaluating task {task} using datatype {opt.get("datatype")}.') # set up world logger task_opt = opt.copy() # copy opt since we're editing the task task_opt['task'] = task # add task suffix in case of multi-tasking if opt['world_logs']: task_opt['world_logs'] = get_task_world_logs( task, task_opt['world_logs'], is_multitask=len(opt['task'].split(',')) > 1) world_logger = WorldLogger(task_opt) if task_opt['world_logs'] else None world = create_task(task_opt, agent) # create worlds for tasks # set up logging log_every_n_secs = opt.get('log_every_n_secs', -1) if log_every_n_secs <= 0: log_every_n_secs = float('inf') log_time = TimeLogger() # max number of examples to evaluate max_cnt = opt['num_examples'] if opt['num_examples'] > 0 else float('inf') cnt = 0 total_cnt = world.num_examples() if is_distributed(): logging.warning('Progress bar is approximate in distributed mode.') while not world.epoch_done() and cnt < max_cnt: cnt += opt.get('batchsize', 1) world.parley() if world_logger is not None: world_logger.log(world) if opt['display_examples']: # display examples print(world.display() + '\n~~') if log_time.time() > log_every_n_secs: report = world.report() text, report = log_time.log(report.get('exs', 0), min(max_cnt, total_cnt), report) logging.info(text) if world_logger is not None: # dump world acts to file world_logger.reset() # add final acts to logs if is_distributed(): rank = get_rank() base_outfile, extension = os.path.splitext(task_opt['world_logs']) outfile = base_outfile + f'_{rank}' + extension else: outfile = task_opt['world_logs'] world_logger.write(outfile, world, file_format=opt['save_format']) report = aggregate_unnamed_reports(all_gather_list(world.report())) if isinstance(world.agents, list) and len(world.agents) > 1: classifier_agent = world.agents[CLASSIFIER_AGENT] if hasattr(classifier_agent, 'calc_auc') and classifier_agent.calc_auc: for class_indices, curr_auc in zip( classifier_agent.auc_class_indices, classifier_agent.aucs): report[ f'AUC_{classifier_agent.class_list[class_indices]}'] = curr_auc classifier_agent.reset_auc() # for safety measures agent.reset_auc() world.reset() return report