def test_play_the_game_less_badly():
bad_seeds_01_env = BadSeeds01(seed_count=5,
bad_seed_count=3,
max_episode_length=2 * 2 + 3 * 3 + 1)
# measure the good seeds twice
# measure the bad seeds three times
for time_i, seed_i in enumerate(
concatv(
take(
n=2 * len(bad_seeds_01_env.good_seeds),
seq=cycle(bad_seeds_01_env.good_seed_indices),
),
take(
n=3 * len(bad_seeds_01_env.bad_seeds),
seq=cycle(bad_seeds_01_env.bad_seed_indices),
),
)):
next_state, terminal, reward = bad_seeds_01_env.execute(actions=seed_i)
assert next_state[time_i, seed_i] != 0.0
assert terminal is False
assert reward == 0.0
# measure the first good seed again
next_state, terminal, reward = bad_seeds_01_env.execute(
actions=bad_seeds_01_env.good_seed_indices[0])
assert next_state[-1, bad_seeds_01_env.good_seed_indices[0]] != 0.0
assert terminal is True
# reward is the number of times the least-measured seed was measured
assert reward == 2.0
def test_play_the_game_less_badly():
bad_seeds_03_env = BadSeeds03(seed_count=5,
bad_seed_count=3,
max_episode_length=3 + 2 * 2 + 3 * 3 + 1)
# measure the good seeds twice
# measure the bad seeds three times
for time_i, seed_i in enumerate(
concatv(
take(
n=2 * len(bad_seeds_03_env.good_seeds),
seq=cycle(bad_seeds_03_env.good_seed_indices),
),
take(
n=3 * len(bad_seeds_03_env.bad_seeds),
seq=cycle(bad_seeds_03_env.bad_seed_indices),
),
)):
time_i += 3
next_state, terminal, reward = bad_seeds_03_env.execute(actions=seed_i)
assert bad_seeds_03_env.history_array[time_i, seed_i] != 0.0
assert terminal is False
assert reward == 0.0
measurement_counts, measured_seed_counts = count_measurements(
bad_seeds_03_env.history_array)
expected_measurement_counts = np.zeros_like(measurement_counts)
expected_measurement_counts[0, bad_seeds_03_env.good_seed_indices] = 5
expected_measurement_counts[0, bad_seeds_03_env.bad_seed_indices] = 6
assert np.all(measurement_counts == expected_measurement_counts)
# measure the first good seed again
next_state, terminal, reward = bad_seeds_03_env.execute(
actions=bad_seeds_03_env.good_seed_indices[0])
print(f"history:\n{bad_seeds_03_env.history_array}")
measurement_counts, measured_seed_counts = count_measurements(
bad_seeds_03_env.history_array)
print(f"measurement_counts: {measurement_counts}")
assert next_state[-1, bad_seeds_03_env.good_seed_indices[0]] != 0.0
assert terminal is True
# reward is the number of times the least-measured seed was measured
assert reward == 6.0
expected_measurement_counts[0, bad_seeds_03_env.good_seed_indices[0]] += 1
assert np.all(measurement_counts == expected_measurement_counts)
def optimize_steps(steps):
"""
Optimize steps.
Currently only optimizes per step type. See the :func:`_optimizer`
decorator for more information on how to register an optimizer.
:param pbag steps: Collection of steps.
:return: a pbag of steps.
"""
def grouping_fn(step):
step_type = type(step)
if step_type in _optimizers:
return step_type
else:
return "unoptimizable"
steps_by_type = groupby(grouping_fn, steps)
unoptimizable = steps_by_type.pop("unoptimizable", [])
omg_optimized = concat(_optimizers[step_type](steps)
for step_type, steps in steps_by_type.iteritems())
return pbag(concatv(omg_optimized, unoptimizable))
def test_concatv():
assert list(concatv([], [], [])) == []
assert (list(take(5, concatv(['a', 'b'], range(1000000000)))) ==
['a', 'b', 0, 1, 2])
from msvdd_bloc import regexes, tokenize
from msvdd_bloc.resumes import education
from msvdd_bloc.resumes import parse_utils
LOGGER = logging.getLogger(__name__)
#######################
## CRF-BASED PARSING ##
#######################
FIELD_SEP_TEXTS = {
sep
for sep in itertoolz.concatv(
education.constants.FIELD_SEPS,
education.constants.FIELD_SEP_DTS,
education.constants.FIELD_SEP_SMS,
education.constants.LEFT_BRACKETS,
education.constants.RIGHT_BRACKETS,
)
}
ITEM_SEP_TEXTS = set(education.constants.FIELD_SEP_SMS)
INSTITUTION_TEXTS = {
"university",
"college",
"institute",
"department",
"dept.",
"high",
"school",
"academy",
}
from toolz import itertoolz
from msvdd_bloc import regexes, tokenize
from msvdd_bloc.resumes import constants
from msvdd_bloc.resumes import parse_utils
from msvdd_bloc.resumes import work
LOGGER = logging.getLogger(__name__)
FIELD_SEP_TEXTS = {
sep for sep in itertoolz.concatv(
work.constants.FIELD_SEPS,
work.constants.FIELD_SEP_DTS,
work.constants.FIELD_SEP_SMS,
constants.LEFT_BRACKETS,
constants.RIGHT_BRACKETS,
)
}
COMPANY_TEXTS = set(
text.lower()
for text in work.constants.COMPANY_TYPES + work.constants.COMPANY_MODIFIERS
)
POSITION_TEXTS = set(
text.lower()
for text in work.constants.POSITION_LEVELS + work.constants.POSITION_TYPES
)
def parse_lines(lines, tagger=None):
def main():
# prameters
data_dir = '../input/'
bert_model = '../bert-large-wwm-uncased' # 把自己large wwm模型路径取代这个bert_model
# bert_model = 'bert-base-uncased'
# bert_model = './oldtoxic'#使用在老toxic上训练好的预训练模型权重.下载路径:https://www.kaggle.com/qinhui1999/old-toxic-bert-v2
task_name = 'MyPro'
output_dir = 'checkpoints/'
model_save_pth = 'checkpoints/bert_large_wwm.pth'
max_seq_length = 220
do_train = True
do_eval = True
do_lower_case = True
train_batch_size = 56
eval_batch_size = 200
learning_rate = 1e-5
num_train_epochs = 1
warmup_proportion = 0.05
no_cuda = False
local_rank = -1
seed = 42
gradient_accumulation_steps = 8
optimize_on_cpu = False
fp16 = False
save_checkpoints_steps = 50000
loss_scale = 128
# 对模型输入进行处理的processor，git上可能都是针对英文的processor
processors = {'mypro': MyPro}
if local_rank == -1 or no_cuda:
device = torch.device(
"cuda" if torch.cuda.is_available() and not no_cuda else "cpu")
n_gpu = torch.cuda.device_count()
else:
device = torch.device("cuda", local_rank)
n_gpu = 1
torch.distributed.init_process_group(backend='nccl')
if fp16:
logger.info(
"16-bits training currently not supported in distributed training"
)
fp16 = False # (see https://github.com/pytorch/pytorch/pull/13496)
logger.info("device %s n_gpu %d distributed training %r", device, n_gpu,
bool(local_rank != -1))
if gradient_accumulation_steps < 1:
raise ValueError(
"Invalid gradient_accumulation_steps parameter: {}, should be >= 1"
.format(gradient_accumulation_steps))
train_batch_size = int(train_batch_size / gradient_accumulation_steps)
random.seed(seed)
np.random.seed(seed)
torch.manual_seed(seed)
if n_gpu > 0:
torch.cuda.manual_seed_all(seed)
if not do_train and not do_eval:
raise ValueError(
"At least one of `do_train` or `do_eval` must be True.")
if os.path.exists(output_dir) and os.listdir(output_dir):
# raise ValueError("Output directory ({}) already exists and is not empty.".format(output_dir))
print('The checkpoint directory is aleady existed...')
else:
os.makedirs(output_dir, exist_ok=True)
task_name = task_name.lower()
if task_name not in processors:
raise ValueError("Task not found: %s" % (task_name))
processor = processors[task_name]()
label_list = processor.get_labels()
tokenizer = BertTokenizer.from_pretrained(bert_model,
do_lower_case=do_lower_case)
# print("tokenizer",tokenizer)
train_examples = None
num_train_steps = None
if do_train:
train_examples = processor.get_train_examples(data_dir)
num_train_steps = int(
len(train_examples) / train_batch_size /
gradient_accumulation_steps * num_train_epochs)
# Prepare model
# model = BertForSequenceClassification.from_pretrained(bert_model, num_labels=2,
# cache_dir=PYTORCH_PRETRAINED_BERT_CACHE / 'distributed_{}'.format(local_rank))
model = ToxicModel(bert_model, device)
# You can unfreeze the last layer of bert by calling set_trainable(model.bert.encoder.layer[23], True)
# set_trainable(model.bert, False)
# 锁定embedding层
# set_trainable(model.bert.embeddings, False)
# set_trainable(model.bert.encoder.layer[11], True)
# set_trainable(model.head, True)
# model.load_state_dict(torch.load('checkpoints/bert_classification_2epoch.pth')['state_dict'])
if fp16:
model.half()
model.to(device)
if local_rank != -1:
model = torch.nn.parallel.DistributedDataParallel(
model, device_ids=[local_rank], output_device=local_rank)
elif n_gpu > 1:
model = torch.nn.DataParallel(model)
# Prepare optimizer
if fp16:
param_optimizer = [
(n, param.clone().detach().to('cpu').float().requires_grad_())
for n, param in model.named_parameters()
]
elif optimize_on_cpu:
param_optimizer = [(n,
param.clone().detach().to('cpu').requires_grad_())
for n, param in model.named_parameters()]
else:
param_optimizer = list(model.named_parameters())
no_decay = ['bias', 'gamma', 'beta']
optimizer_grouped_parameters = [{
'params':
[p for n, p in param_optimizer if not any(nd in n for nd in no_decay)],
'weight_decay_rate':
0.01
}, {
'params':
[p for n, p in param_optimizer if any(nd in n for nd in no_decay)],
'weight_decay_rate':
0.0
}]
t_total = num_train_steps
if local_rank != -1:
t_total = t_total // torch.distributed.get_world_size()
optimizer = BertAdam(optimizer_grouped_parameters,
lr=learning_rate,
warmup=warmup_proportion,
t_total=t_total)
global_step = 0
if do_train:
if os.path.exists('train.token_new_cleaned_wwm.npy'):
train_features = np.load('train.token_new_cleaned_wwm.npy',
allow_pickle=True)
else:
parallel = Parallel(300, backend="multiprocessing", verbose=5)
train_features = list(
concatv(*parallel(
delayed(convert_examples_to_features)(
example, label_list, max_seq_length, tokenizer)
for example in list(partition_all(300, train_examples)))))
train_features = np.asarray(train_features)
np.save('train.token_new_cleaned_wwm', train_features)
logger.info("***** Running training *****")
logger.info(" Num examples = %d", len(train_examples))
logger.info(" Batch size = %d", train_batch_size)
logger.info(" Num steps = %d", num_train_steps)
torch.cuda.empty_cache()
all_input_ids = torch.tensor([f.input_ids for f in train_features],
dtype=torch.long)
all_input_mask = torch.tensor([f.input_mask for f in train_features],
dtype=torch.long)
all_segment_ids = torch.tensor([f.segment_ids for f in train_features],
dtype=torch.long)
print('y_aux', np.asarray([f.y_aux for f in train_features]).shape)
all_label_ids = torch.tensor(np.hstack([
np.asarray([f.label_id for f in train_features]),
np.asarray([f.y_aux for f in train_features])
]),
dtype=torch.float32)
train_data = TensorDataset(all_input_ids, all_input_mask,
all_segment_ids, all_label_ids)
if local_rank == -1:
train_sampler = RandomSampler(train_data)
else:
train_sampler = DistributedSampler(train_data)
train_dataloader = DataLoader(
train_data,
sampler=train_sampler,
batch_size=train_batch_size,
num_workers=2,
pin_memory=True,
)
#model.load_state_dict(torch.load('checkpoints/bert_large_wwm.pth')['state_dict'])
# model.load_state_dict(torch.load('checkpoints/0_80000_iterations.pth')['state_dict'])
model.train()
best_score = 0
flags = 0
torch.cuda.empty_cache()
'''
model.load_state_dict(torch.load('checkpoints/0_20000_iterations.pth')['model'])
optimizer.load_state_dict(torch.load('checkpoints/0_20000_iterations.pth')['optimizer'])
old_iter = int(torch.load('checkpoints/0_20000_iterations.pth')['iteration'])
'''
old_iter = -1
for i_epoch in trange(int(num_train_epochs), desc="Epoch"):
torch.cuda.empty_cache()
iteration = 0 # counter
save_point = save_checkpoints_steps # 10000
for step, batch in enumerate(
tqdm(train_dataloader, desc="Iteration")):
if iteration <= old_iter:
iteration += 1
continue
batch = tuple(t.to(device) for t in batch)
input_ids, input_mask, segment_ids, label_ids = batch
loss = model(input_ids, segment_ids, input_mask, label_ids)
torch.cuda.empty_cache()
if n_gpu > 1:
loss = loss.mean() # mean() to average on multi-gpu.
if fp16 and loss_scale != 1.0:
# rescale loss for fp16 training
# see https://docs.nvidia.com/deeplearning/sdk/mixed-precision-training/index.html
loss = loss * loss_scale
if gradient_accumulation_steps > 1:
loss = loss / gradient_accumulation_steps
loss.backward()
if (step + 1) % gradient_accumulation_steps == 0:
if fp16 or optimize_on_cpu:
if fp16 and loss_scale != 1.0:
# scale down gradients for fp16 training
for param in model.parameters():
if param.grad is not None:
param.grad.data = param.grad.data / loss_scale
is_nan = set_optimizer_params_grad(
param_optimizer,
model.named_parameters(),
test_nan=True)
if is_nan:
logger.info(
"FP16 TRAINING: Nan in gradients, reducing loss scaling"
)
loss_scale = loss_scale / 2
model.zero_grad()
continue
optimizer.step()
copy_optimizer_params_to_model(
model.named_parameters(), param_optimizer)
else:
optimizer.step()
model.zero_grad()
#Save model
if iteration % save_point == 0 and iteration > 0:
checkpoint = {
'iteration': iteration,
'model': model.state_dict(),
'optimizer': optimizer.state_dict()
}
checkpoint_path = os.path.join(
output_dir,
'{}_{}_iterations.pth'.format(i_epoch, iteration))
torch.save(checkpoint, checkpoint_path)
logging.info('Model saved to {}'.format(checkpoint_path))
val(model, processor, data_dir, max_seq_length,
eval_batch_size, label_list, tokenizer, device)
iteration += 1
checkpoint = {
'state_dict': model.state_dict(),
'model': model.state_dict(),
'optimizer': optimizer.state_dict()
}
torch.save(checkpoint, model_save_pth)
val(model, processor, data_dir, max_seq_length, eval_batch_size,
label_list, tokenizer, device)
test(model, processor, data_dir, max_seq_length, eval_batch_size,
label_list, tokenizer, device)
