def _eval_single_world(opt, agent, task):
print(
'[ Evaluating task {} using datatype {}. ] '.format(
task, opt.get('datatype', 'N/A')
)
)
task_opt = opt.copy() # copy opt since we're editing the task
task_opt['task'] = task
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
while not world.epoch_done() and cnt < max_cnt:
cnt += opt.get('batchsize', 1)
world.parley()
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['exs'], world.num_examples(), report)
print(text)
report = world.report()
world.reset()
return report
def train_lambda(opt, printargs=None, print_parser=None):
"""Trains lambda for pre-fit model components on training data.
:param opt: tells the evaluation function how to run
:param bool print_parser: if provided, prints the options that are set within the
model after loading the model
:return: the final result of calling report()
"""
if printargs is not None:
print('[ Deprecated Warning: eval_model no longer uses `printargs` ]')
print_parser = printargs
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: eval_model should be passed opt not Parser ]'
)
opt = opt.parse_args()
random.seed(42)
# Create model and assign it to the specified task
agent = create_agent(opt, requireModelExists=True)
world = create_task(opt, agent)
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()
# Show some example dialogs:
cnt = 0
while not world.epoch_done():
cnt += opt.get('batchsize', 1)
world.parley()
if opt['display_examples']:
print(world.display() + "\n~~")
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('finished evaluating task {} using datatype {}'.format(
opt['task'], opt.get('datatype', 'N/A')))
report = world.report()
print(report)
return report
def eval_model(opt, printargs=None, print_parser=None):
"""Evaluates a model.
Arguments:
opt -- tells the evaluation function how to run
print_parser -- if provided, prints the options that are set within the
model after loading the model
"""
if printargs is not None:
print('[ Deprecated Warning: eval_model no longer uses `printargs` ]')
print_parser = printargs
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: eval_model should be passed opt not Parser ]'
)
opt = opt.parse_args()
random.seed(42)
# Create model and assign it to the specified task
agent = create_agent(opt, requireModelExists=True)
world = create_task(opt, agent)
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()
# Show some example dialogs:
cnt = 0
while not world.epoch_done():
cnt += 1
world.parley()
if opt['display_examples']:
print(world.display() + "\n~~")
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")
report = world.report()
print(report)
return report
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)
bad = OffensiveLanguageDetector()
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()
# Show some example dialogs:
cnt = 0
while not world.epoch_done():
world.parley()
words = []
for a in world.acts:
offensive = bad.contains_offensive_language(a.get('text', ''))
if offensive:
words.append(offensive)
labels = a.get('labels', a.get('eval_labels', ''))
for l in labels:
offensive = bad.contains_offensive_language(l)
if offensive:
words.append(offensive)
if len(words) > 0 and opt['display_examples']:
print(world.display())
print("[Offensive words detected:]", ', '.join(words))
print("\n~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~\n")
cnt += len(words)
if log_time.time() > log_every_n_secs:
report = world.report()
log = {'offenses': cnt}
text, log = log_time.log(report['exs'], world.num_examples(), log)
print(text)
if world.epoch_done():
print("EPOCH DONE")
print(
str(cnt) + " offensive messages found out of " +
str(world.num_examples()) + " messages.")
return world.report()
def get_word_counts(opt, count_inputs):
"""Goes through the dataset specified in opt, returns word counts and all utterances
Inputs:
count_inputs: If True, include both input and reply when counting words and
utterances. Otherwise, only include reply text.
Returns:
word_counter: a Counter mapping each word to the total number of times it appears
total_count: int. total word count, i.e. the sum of the counts for each word
all_utts: list of strings. all the utterances that were used for counting words
"""
# Create repeat label agent and assign it to the specified task
agent = RepeatLabelAgent(opt)
world = create_task(opt, agent)
# Count word frequency for all words in dataset
word_counter = Counter()
total_count = 0
all_utts = []
log_timer = TimeLogger()
while True:
world.parley()
# Count words in reply
reply = world.acts[0].get('labels',
world.acts[0].get('eval_labels'))[0]
words = reply.split()
word_counter.update(words)
total_count += len(words)
all_utts.append(reply)
# Optionally count words in input text
if count_inputs:
input = world.acts[0]['text']
input = input.split('\n')[
-1] # e.g. in ConvAI2, this removes persona
words = input.split()
word_counter.update(words)
total_count += len(words)
all_utts.append(input)
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():
print('EPOCH DONE')
break
assert total_count == sum(word_counter.values())
return word_counter, total_count, all_utts
def get_word_counts(opt, count_inputs):
"""Goes through the dataset specified in opt and gets word counts.
Inputs:
count_inputs: If True, include both input and reply when counting words
and utterances. Otherwise, only include reply text.
Returns:
word_counter_per_sent: a Counter mapping each word to the number of
utterances in which it appears.
num_sents: int. number of utterances counted
"""
# Create repeat label agent and assign it to the specified task
agent = RepeatLabelAgent(opt)
world = create_task(opt, agent)
# Count word frequency for all words in dataset
word_counter_per_sent = Counter()
num_sents = 0
count = 0
log_timer = TimeLogger()
while True:
count += 1
world.parley()
reply = world.acts[0].get('labels',
world.acts[0].get('eval_labels'))[0]
words = reply.split()
words_no_dups = list(set(words)) # remove duplicates
word_counter_per_sent.update(words_no_dups)
num_sents += 1
# Optionally count words in input text
if count_inputs:
input = world.acts[0]['text']
words = input.split()
words_no_dups = list(set(words)) # remove duplicates
word_counter_per_sent.update(words_no_dups)
num_sents += 1
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():
print('EPOCH DONE')
break
return word_counter_per_sent, num_sents
def verify(opt, printargs=None, print_parser=None):
# 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_label_candidates'] = 0
# Show some example dialogs.
while not world.epoch_done():
world.parley()
act = world.acts[0]
if 'text' not in act:
print("warning: missing text field")
counts['missing_text'] += 1
if 'labels' not in act and 'eval_labels' not in act:
print("warning: missing labels/eval_labels field")
counts['missing_labels'] += 1
else:
if 'label_candidates' not in act:
counts['missing_label_candidates'] += 1
else:
for c in act['label_candidates']:
if c == '':
print("warning: empty string label_candidate")
counts['empty_label_candidates'] += 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 build_cands(opt):
# create repeat label agent and assign it to the specified task
if opt['numthreads'] > 1:
# Broken in hogwild mode. Just fall back to single processing mode
opt['numthreads'] = 1
agent = RepeatLabelAgent(opt)
world = create_task(opt, agent)
if opt['outfile'] is None:
outfile = tempfile.mkstemp(prefix='{}_{}_'.format(
opt['task'], opt['datatype']),
suffix='.txt')[1]
else:
outfile = opt['outfile']
if opt.get('num_examples', -1) == -1:
num_examples = world.num_examples()
else:
num_examples = opt['num_examples']
log_timer = TimeLogger()
print('[ starting to build candidates from task.. (ex:' +
str(num_examples) + ')]')
print('[ saving output to {} ]'.format(outfile))
cands = []
for _ in range(num_examples):
world.parley()
# We get the acts of the first agent, which is the teacher.
acts = world.get_acts()[0]
if isinstance(acts, dict):
# We turn into a batch of 1 example, in case batching is being used.
acts = [acts]
for a in acts:
candidate = a.get('labels', a.get('eval_labels', None))
if candidate is not None:
candidate = candidate[0]
cands.append(candidate)
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():
print('EPOCH DONE')
break
fw = open(outfile, 'w')
fw.write('\n'.join(cands))
fw.close()
def dump_data(opt):
# create repeat label agent and assign it to the specified task
agent = RepeatLabelAgent(opt)
world = create_task(opt, agent)
if opt['outfile'] is None:
outfile = tempfile.mkstemp(prefix='{}_{}_'.format(
opt['task'], opt['datatype']),
suffix='.txt')[1]
else:
outfile = opt['outfile']
if opt['num_examples'] == -1:
num_examples = world.num_examples()
else:
num_examples = opt['num_examples']
log_timer = TimeLogger()
print('[ starting to convert.. ]')
print('[ saving output to {} ]'.format(outfile))
fw = open(outfile, 'w')
text = ''
for _ in range(num_examples):
world.parley()
world.acts[0]['labels'] = world.acts[0].get(
'labels', world.acts[0].pop('eval_labels', None))
samp = world.acts[0]
text += samp["text"].replace("\n", " ") + " "
fw.write("__label__%s %s\n" %
(samp["labels"][0].replace(' ', '_'), text))
if world.acts[0].get('episode_done', False):
text = ''
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():
print('EPOCH DONE')
break
fw.close()
def dump_data(opt):
# create repeat label agent and assign it to the specified task
agent = RepeatLabelAgent(opt)
world = create_task(opt, agent)
ignorefields = opt.get('ignore_fields', '')
if opt['outfile'] is None:
outfile = tempfile.mkstemp(prefix='{}_{}_'.format(
opt['task'], opt['datatype']),
suffix='.txt')[1]
else:
outfile = opt['outfile']
if opt['num_examples'] == -1:
num_examples = world.num_examples()
else:
num_examples = opt['num_examples']
log_timer = TimeLogger()
print('[ starting to convert.. ]')
print('[ saving output to {} ]'.format(outfile))
fw = open(outfile, 'w')
for _ in range(num_examples):
world.parley()
world.acts[0]['labels'] = world.acts[0].get(
'labels', world.acts[0].pop('eval_labels', None))
txt = msg_to_str(world.acts[0], ignore_fields=ignorefields)
fw.write(txt + '\n')
if world.acts[0].get('episode_done', False):
fw.write('\n')
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():
print('EPOCH DONE')
break
fw.close()
def build_dict(opt, skip_if_built=False):
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 opt.get('dict_class'):
# Custom dictionary class
dictionary = str2class(opt['dict_class'])(opt)
else:
# Default dictionary class
dictionary = DictionaryAgent(opt)
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
ordered_opt['image_mode'] = 'none'
if ordered_opt['task'] == 'pytorch_teacher':
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
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_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()
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 log_time.time() > log_every_n_secs:
sys.stdout.write('\r')
text, _log = log_time.log(cnt, max(opt.get('dict_maxexs', 0),
world_dict.num_examples()))
sys.stdout.write(text)
sys.stdout.flush()
dictionary.save(opt['dict_file'], sort=True)
print('[ dictionary built with {} tokens ]'.format(len(dictionary)))
return dictionary
def _probe_single_world(opt, agent, task):
print('[ Evaluating task {} using datatype {}. ] '.format(
task, opt.get('datatype', 'N/A')))
task_opt = opt.copy() # copy opt since we're editing the task
task_opt['task'] = task
world = create_task(task_opt, agent) # create worlds for tasks
if task_opt['batchsize'] == 1:
raise ValueError('Batch size must be greater than 1. '
'Use the --batchsize flag to set batch size.')
# 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
while not world.epoch_done() and cnt < max_cnt:
cnt += opt.get('batchsize', 1)
world.parley()
if opt['display_examples']:
# display examples
print(world.display() + '\n~~')
if log_time.time() > log_every_n_secs:
report = world.report()
_report = {'exs': cnt}
text, report = log_time.log(cnt, world.num_examples(), _report)
print(text)
# Create save folder for probing outputs
task_name = world.opt['task'].split('.')[-2]
model_dir = Path(world.opt['model_file']).parent
probing_dir = model_dir.joinpath('probing')
probing_module_dir = probing_dir.joinpath(world.opt['probe'])
task_dir = probing_module_dir.joinpath(task_name)
save_path = task_dir.joinpath(task_name + '.pkl')
if not probing_dir.exists():
print("*" * 10, "\n", "*" * 10)
print(f"Creating dir to save probing outputs at {probing_dir}")
print("*" * 10, "\n", "*" * 10)
probing_dir.mkdir()
if not probing_module_dir.exists():
print("*" * 10, "\n", "*" * 10)
print(
f"Creating dir to save {world.opt['probe']} probing outputs at {probing_module_dir}"
)
print("*" * 10, "\n", "*" * 10)
probing_module_dir.mkdir()
if not task_dir.exists():
print("*" * 10, "\n", "*" * 10)
print(
f"Creating dir to save {task_name} probing outputs at {task_dir}")
print("*" * 10, "\n", "*" * 10)
task_dir.mkdir()
if save_path.exists():
warnings.warn(
f"\nVector representations for probing already exists at {save_path}!!\n"
"They will be overwritten.", RuntimeWarning)
print("*" * 10, "\n", "*" * 10)
print(
f"Creating pickle file to save {task_name} probing outputs at {save_path}"
)
print("*" * 10, "\n", "*" * 10)
# Save probing outputs
try:
pickle.dump(world.world.agents[1].probing_outputs,
open(save_path, 'wb'))
except:
pickle.dump(world.world.agents[1].probing_outputs,
open(save_path, 'wb'),
protocol=4)
report = world.report()
world.reset()
return report
def eval_model(opt, printargs=None, print_parser=None):
"""Evaluates a model.
:param opt: tells the evaluation function how to run
:param bool print_parser: if provided, prints the options that are set within the
model after loading the model
:return: the final result of calling report()
"""
if printargs is not None:
print('[ Deprecated Warning: eval_model no longer uses `printargs` ]')
print_parser = printargs
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: eval_model should be passed opt not Parser ]'
)
opt = opt.parse_args()
random.seed(42)
# Create model and assign it to the specified task
agent = create_agent(opt, requireModelExists=True)
world = create_task(opt, agent)
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()
# Show some example dialogs:
cnt = 0
file_name = "results" + str(time.time()) + "wow" + ".txt"
print("Writing to " + file_name)
with open(file_name, "w+") as results:
while not world.epoch_done():
cnt += opt.get('batchsize', 1)
world.parley()
if opt['display_examples']:
print(world.display() + "\n~~")
results.write(world.display() + "\n")
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('finished evaluating task {} using datatype {}'.format(
opt['task'], opt.get('datatype', 'N/A')))
report = world.report()
print(report)
print("\n".join([
"", "*" * 80,
"Thank you for using ParlAI! We are conducting a user survey.",
"Please consider filling it out at https://forms.gle/uEFbYGP7w6hiuGQT9",
"*" * 80, ""
]))
return report
def eval_wordstat(opt, print_parser=None):
"""Evaluates a model.
Arguments:
opt -- tells the evaluation function how to run
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': []
}
bins = [int(i) for i in opt['freq_bins'].split(',')]
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
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:
prediction = w.acts[-1]['text']
word_statistics['context_list'].append(w.acts[0]['text'])
word_statistics['pure_pred_list'].append(prediction)
except:
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']) + ', '.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))
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])
report = world.report()
print(report)
return 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 make_dataset(opt):
# Initialize control information so we can compute sentence attributes.
# Here we set build_task=False so we don't download data/controllable_dialogue
# (because we're trying to create it instead).
initialize_control_information(opt, build_task=False)
# Create repeat label agent and assign it to the specified task
agent = RepeatLabelAgent(opt)
world = create_task(opt, agent)
ignorefields = opt.get('ignore_fields', '')
outfile = opt['outfile']
# Number of examples to process
if opt['num_examples'] == -1:
num_examples = world.num_examples()
else:
num_examples = opt['num_examples']
# List of controls to include:
controls = opt['controls'].split(',') if opt['controls'] != '' else []
print('[ starting to convert.. ]')
print('[ saving output to {} ]'.format(outfile))
fw = open(outfile, 'w')
log_timer = TimeLogger()
for _ in range(num_examples):
world.parley()
world.acts[0]['labels'] = world.acts[0].get(
'labels', world.acts[0].pop('eval_labels', None))
# Need to get history in order to compute control values
hist = ConvAI2History(world.acts[0]['text'], assume_persontokens=False)
response = world.acts[0]['labels'][0]
# Compute control values
for ctrl in controls:
ctrl_val = eval_attr(response, hist, ctrl)
if ctrl == 'avg_nidf':
assert ctrl_val >= 0
assert ctrl_val <= 1
elif ctrl == 'question':
assert ctrl_val in [0, 1]
elif ctrl == 'lastuttsim':
if ctrl_val is not None:
assert ctrl_val >= -1
assert ctrl_val <= 1
else:
raise Exception('unexpected ctrl name: %s' % ctrl)
world.acts[0][ctrl] = ctrl_val # add control value to act
# Write to file
txt = msg_to_str(world.acts[0], ignore_fields=ignorefields)
fw.write(txt + '\n')
if world.acts[0].get('episode_done', False):
fw.write('\n')
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():
print('EPOCH DONE')
break
fw.close()
def build_dict(opt, skip_if_built=False):
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
dictionary = str2class(opt['dict_class'])(opt)
else:
# Default dictionary class
dictionary = DictionaryAgent(opt)
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
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
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, 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)
# 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']
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()
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 world_idx, w in enumerate(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 open(outfile, 'w') as f:
json.dump(data, f)
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
# Show some example dialogs.
while not world.epoch_done():
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 bucket_data(opt):
# create repeat label agent and assign it to the specified task
agent = RepeatLabelAgent(opt)
world = create_task(opt, agent)
if opt['num_examples'] == -1:
num_examples = world.num_examples()
else:
num_examples = opt['num_examples']
log_timer = TimeLogger()
assert opt['control'] != ''
ctrl = opt['control']
num_buckets = opt['num_buckets']
ctrl_vals = [] # list of floats
for _ in range(num_examples):
world.parley()
world.acts[0]['labels'] = world.acts[0].get(
'labels', world.acts[0].pop('eval_labels', None))
if ctrl not in world.acts[0].keys():
raise Exception('Error: control %s isn\'t in the data. available keys: %s'
% (ctrl, ', '.join(world.acts[0].keys())))
ctrl_val = world.acts[0][ctrl]
if ctrl_val == "None":
assert ctrl == 'lastuttsim'
ctrl_val = None
else:
ctrl_val = float(ctrl_val)
if ctrl == 'avg_nidf':
assert ctrl_val >= 0
assert ctrl_val <= 1
elif ctrl == 'question':
assert ctrl_val in [0, 1]
elif ctrl == 'lastuttsim':
if ctrl_val is not None:
assert ctrl_val >= -1
assert ctrl_val <= 1
else:
raise Exception('Unexpected ctrl name: %s' % ctrl)
ctrl_vals.append(ctrl_val)
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():
print('EPOCH DONE')
break
if ctrl == 'lastuttsim':
num_nones = len([v for v in ctrl_vals if v is None])
ctrl_vals = [v for v in ctrl_vals if v is not None]
print("Have %i Nones for lastuttsim; these have been removed "
"for bucket calculation" % num_nones)
print('Collected %i control vals between %.6f and %.6f'
% (len(ctrl_vals), min(ctrl_vals), max(ctrl_vals)))
# Calculate bucket lower bounds
print('Calculating lowerbounds for %i buckets...' % num_buckets)
ctrl_vals = sorted(ctrl_vals)
lb_indices = [int(len(ctrl_vals)*i/num_buckets) for i in range(num_buckets)]
lbs = [ctrl_vals[idx] for idx in lb_indices]
print('\nBucket lowerbounds for control %s: ' % ctrl)
print(lbs)
# Calculate the actual bucket sizes
bucket_sizes = Counter()
bucket_ids = [sort_into_bucket(ctrl_val, lbs) for ctrl_val in ctrl_vals]
bucket_sizes.update(bucket_ids)
print('\nBucket sizes: ')
for bucket_id in sorted(bucket_sizes.keys()):
print("%i: %i" % (bucket_id, bucket_sizes[bucket_id]))
def eval_wordstat(opt, print_parser=None):
"""Evaluates a model.
Arguments:
opt -- tells the evaluation function how to run
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['external_dict'] is not None:
print('[ Using external dictionary from: {} ]'.format(
opt['external_dict']))
dictionary = DictionaryAgent(opt)
dictionary.load(opt['external_dict'])
else:
print('[ Using model bundled dictionary ]')
dictionary = agent.dict
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
mean_wlength = []
mean_clength = []
freqs_cnt = Counter()
word_cnt = 0
bins = [int(i) for i in opt['freq_bins'].split(',')]
while not world.epoch_done():
cnt += 1
world.parley()
prediction = world.acts[-1]['text']
freqs, _cnt, wlength, clength = get_word_stats(prediction,
dictionary,
bins=bins)
word_cnt += _cnt
mean_wlength.append(wlength)
mean_clength.append(clength)
freqs_cnt += Counter(freqs)
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_cnt) + ', '.join([
'<{}:{} ({:.{prec}f}%)'.format(
b,
freqs_cnt.get(b, 0),
(freqs_cnt.get(b, 0) / 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(mean_wlength).mean(),
numpy.array(mean_clength).mean(),
prec=2))
if opt['num_examples'] > 0 and cnt >= opt['num_examples']:
break
if world.epoch_done():
print("EPOCH DONE")
report = world.report()
print(report)
return report
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:
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)