def evaluate(args, model, tokenizer, processor, label_list, device, mode="test"):
num_labels = len(label_list) + 1
eval_data = load_examples(args, tokenizer, processor, label_list, mode)
# Run prediction for full data
eval_sampler = SequentialSampler(eval_data)
eval_dataloader = DataLoader(eval_data, sampler=eval_sampler, batch_size=args.eval_batch_size)
model.eval()
y_true = []
y_pred = []
label_map = {i: label for i, label in enumerate(label_list, 1)}
label_map[0] = 'unknown'
nb_tr_examples, nb_tr_steps = 0, 0
for batch in tqdm(eval_dataloader, desc="Evaluating"):
batch = tuple(t.to(device) for t in batch)
input_ids, input_mask, segment_ids, label_ids, valid_ids, label_mask, b_use_valid_filter,\
adj_matrix, dep_matrix = batch
with torch.no_grad():
logits = model(input_ids=input_ids, token_type_ids=segment_ids, attention_mask=input_mask,
valid_ids=valid_ids, adjacency_matrix=adj_matrix)
nb_tr_examples += input_ids.size(0)
nb_tr_steps += 1
logits = torch.argmax(F.log_softmax(logits, dim=2), dim=2)
logits = logits.detach().cpu().numpy()
label_ids = label_ids.detach().cpu().numpy()
for i, label in enumerate(label_ids):
temp_1 = []
temp_2 = []
for j, m in enumerate(label):
if j == 0:
continue
elif label_ids[i][j] == num_labels - 1:
y_true.append(temp_1)
y_pred.append(temp_2)
break
else:
temp_1.append(label_map[label_ids[i][j]])
temp_2.append(label_map[logits[i][j]])
logger.info("nb_tr_examples: {}, nb_tr_steps: {}".format(nb_tr_examples, nb_tr_steps))
result = evaluate_ote(y_true, y_pred)
logging.info(result)
return {
"precision": result[0],
"recall": result[1],
"f1": result[2]
}
def train(data_dir='data/memes/',
dim_proj=512,
maxlen=30,
batch_size=256,
keep_ratio=1.,
shuffle_data=True,
learning_rate=0.001,
global_steps=50000,
disp_freq=100,
save_freq=1000,
test_freq=1000,
saveto_file='params.npz',
weight_decay=0.0005,
reload_model=False,
train=True):
"""
Topo-LSTM model training.
"""
options = locals().copy()
saveto = data_dir + saveto_file
# loads graph
G, node_index = data_utils.load_graph(data_dir)
print nx.info(G)
options['n_words'] = len(node_index)
print options
# creates and initializes shared variables.
print 'Initializing variables...'
params = init_params(options)
if reload_model:
print 'reusing saved model.'
load_params(saveto, params)
tparams = init_tparams(params)
# builds Topo-LSTM model
print 'Building model...'
model = tprnn_model.build_model(tparams, options)
print 'Loading test data...'
test_examples = data_utils.load_examples(data_dir,
dataset='test',
node_index=node_index,
maxlen=maxlen,
G=G)
test_loader = data_utils.Loader(test_examples, options=options)
print 'Loaded %d test examples' % len(test_examples)
if train:
# prepares training data.
print 'Loading train data...'
train_examples = data_utils.load_examples(
data_dir,
dataset='train',
keep_ratio=options['keep_ratio'],
node_index=node_index,
maxlen=maxlen,
G=G)
train_loader = data_utils.Loader(train_examples, options=options)
print 'Loaded %d training examples.' % len(train_examples)
# compiles updates.
optimizer = downhill.build(algo='adam',
loss=model['cost'],
params=tparams.values(),
inputs=model['data'])
updates = optimizer.get_updates(max_gradient_elem=5.,
learning_rate=learning_rate)
f_update = theano.function(model['data'],
model['cost'],
updates=list(updates))
# training loop.
start_time = timeit.default_timer()
# downhill.minimize(
# loss=cost,
# algo='adam',
# train=train_loader,
# # inputs=input_list + [labels],
# # params=tparams.values(),
# # patience=0,
# max_gradient_clip=1,
# # max_gradient_norm=1,
# learning_rate=learning_rate,
# monitors=[('cost', cost)],
# monitor_gradients=False)
n_examples = len(train_examples)
batches_per_epoch = n_examples // options['batch_size'] + 1
n_epochs = global_steps // batches_per_epoch + 1
global_step = 0
cost_history = []
for _ in range(n_epochs):
for _ in range(batches_per_epoch):
cost = f_update(*train_loader())
cost_history += [cost]
if global_step % disp_freq == 0:
print 'global step %d, cost: %f' % (global_step, cost)
# dump model parameters.
if global_step % save_freq == 0:
params = unzip(tparams)
np.savez(saveto, **params)
pickle.dump(options, open('%s.pkl' % saveto, 'wb'), -1)
# evaluate on test data.
if global_step % test_freq == 0:
scores = evaluate(model['f_prob'], test_loader)
print 'eval scores: ', scores
end_time = timeit.default_timer()
print 'time used: %d seconds.' % (end_time - start_time)
global_step += 1
scores = evaluate(model['f_prob'], test_loader)
pprint.pprint(scores)
def train_pet(args) -> List:
# Load configs
model_config, train_config, eval_config = load_pet_configs(args)
# Load dataset
train_data = load_examples(args.task_name, args.data_dir, TRAIN_SET,
num_examples=args.train_examples, split_examples_evenly=args.split_examples_evenly)
eval_data = load_examples(args.task_name, args.data_dir, TEST_SET if args.eval_set == 'test' else DEV_SET,
num_examples=args.eval_examples, split_examples_evenly=args.split_examples_evenly)
dev_data = load_examples(args.task_name, args.data_dir, DEV32_SET,
num_examples=args.dev_examples, split_examples_evenly=args.split_examples_evenly)
set_seed(args.seed)
# Record all evaluation results on dev & eval set
all_result = []
dev_result_all = defaultdict(lambda: defaultdict(list))
eval_result_all = defaultdict(lambda: defaultdict(list))
# In 2 stage training, the 1st stage evaluations should also be recorded
if args.do_train and args.do_eval and args.two_stage_train:
dev_stage1_all = defaultdict(lambda: defaultdict(list))
eval_stage1_all = defaultdict(lambda: defaultdict(list))
# Iterates through all patterns
for pattern_id in args.pattern_ids:
# Repeat training
for iteration in range(args.pet_repetitions):
results_dict = {}
model_config.pattern_id = pattern_id
pattern_iter_output_dir = "{}/p{}-i{}".format(
args.output_dir, pattern_id, iteration)
if os.path.exists(pattern_iter_output_dir):
logger.warning(
f"Path {pattern_iter_output_dir} already exists, skipping it...")
continue
os.makedirs(pattern_iter_output_dir)
# Init wrapper model
assert model_config.pattern_id is not None, 'A pattern_id must be set for initializing a new PET model'
wrapper = TransformerModelWrapper(model_config)
# Training
logger.info('--- Start iteration %d ---' % iteration)
if args.do_train:
if not args.two_stage_train:
# Single stage training
logger.info('=== Start training ===')
results_dict.update(train_single_model(train_data, eval_data, dev_data, pattern_iter_output_dir,
wrapper, train_config, eval_config))
evaluate_single_model(pattern_id, pattern_iter_output_dir, eval_data,
dev_data, eval_config, results_dict, dev_result_all, eval_result_all)
with open(os.path.join(pattern_iter_output_dir, 'results.json'), 'w') as fh:
json.dump(results_dict, fh)
else:
# Two stage training
# 1. Only train prompts and label tokens
logger.info('=== Start training stage 1 ===')
results_dict.update(train_single_model(train_data, eval_data, dev_data, pattern_iter_output_dir,
wrapper, train_config, eval_config, stage=1))
evaluate_single_model(pattern_id, pattern_iter_output_dir, eval_data,
dev_data, eval_config, results_dict, dev_stage1_all, eval_stage1_all)
with open(os.path.join(pattern_iter_output_dir, 'results_stage1.json'), 'w') as fh:
json.dump(results_dict, fh)
# 2. Train full model
logger.info('=== Start training stage 2 ===')
results_dict.update(train_single_model(train_data, eval_data, dev_data, pattern_iter_output_dir,
wrapper, train_config, eval_config, stage=2))
evaluate_single_model(pattern_id, pattern_iter_output_dir, eval_data,
dev_data, eval_config, results_dict, dev_result_all, eval_result_all)
with open(os.path.join(pattern_iter_output_dir, 'results.json'), 'w') as fh:
json.dump(results_dict, fh)
# Save configs
train_config.save(os.path.join(
pattern_iter_output_dir, 'train_config.json'))
eval_config.save(os.path.join(
pattern_iter_output_dir, 'eval_config.json'))
logger.info("Saving complete")
# Do evaluation only
elif args.do_eval:
evaluate_single_model(pattern_id, pattern_iter_output_dir, eval_data,
dev_data, eval_config, results_dict, dev_result_all, eval_result_all)
# Write overall results
with open(os.path.join(pattern_iter_output_dir, 'results.json'), 'w') as fh:
json.dump(results_dict, fh)
# Clear cache
wrapper.model = None
wrapper = None
torch.cuda.empty_cache()
# Calculate average results of current pattern
if args.do_eval:
logger.info("=== OVERALL RESULTS ===")
if args.do_train and args.do_eval and args.two_stage_train:
# Store stage 1 results first
all_result.extend(write_results(os.path.join(
args.output_dir, 'result_stage1.txt'), dev_stage1_all, eval_stage1_all))
all_result.extend(write_results(os.path.join(
args.output_dir, 'result.txt'), dev_result_all, eval_result_all))
return all_result
def train(data_dir='data/memes/',
dim_proj=256,
dim_att=128,
maxlen=30,
batch_size=256,
keep_ratio=1.,
shuffle_data=True,
learning_rate=0.001,
global_steps=50000,
disp_freq=100,
save_freq=100,
test_freq=100,
saveto_file='params.npz',
tmsaveto_file='timeparams.npz',
weight_decay=0.0005,
sigmasqr=1,
tdim=1.,
reload_model=False,
train=True):
"""
Topo-LSTM model training.
tdim: scale time down by how many times
"""
options = locals().copy()
#savedstep = '0'
saveto = data_dir + saveto_file
tmsaveto = data_dir + tmsaveto_file
# loads graph
Gp, node_index = data_utils.load_graph(data_dir)
#print nx.info(G)
options['n_events'] = len(node_index)
print options
# creates and initializes shared variables.
print 'Initializing variables...'
params = init_params(options)
if reload_model:
print 'reusing saved model.'
load_params(saveto, params)
tparams = init_tparams(params)
timeparams = init_timeparams(options)
if reload_model:
print 'reusing saved model.'
load_params(tmsaveto, timeparams)
timetparams = init_tparams(timeparams)
# builds Topo-LSTM model
print 'Building model...'
model = tpgru_model.build_model(tparams, timetparams, options)
print 'Loading test data...'
test_examples = data_utils.load_examples(data_dir,
dataset='test',
node_index=node_index,
maxlen=maxlen,
Gp=Gp)
test_loader = data_utils.Loader(test_examples, options=options)
print 'Loaded %d test examples' % len(test_examples)
if train:
# prepares training data.
print 'Loading train data...'
train_examples = data_utils.load_examples(
data_dir,
dataset='train',
keep_ratio=options['keep_ratio'],
node_index=node_index,
maxlen=maxlen,
Gp=Gp)
train_loader = data_utils.Loader(train_examples, options=options)
print 'Loaded %d training examples.' % len(train_examples)
# compiles updates.
optimizer = downhill.build(algo='adam',
loss=model['cost'],
params=tparams.values(),
inputs=model['data'])
updates = optimizer.get_updates(max_gradient_elem=5.,
learning_rate=learning_rate)
f_update = theano.function(model['data'],
model['cost'],
updates=list(updates))
toptimizer = downhill.build(algo='adam',
loss=model['timecost'],
params=timetparams.values(),
inputs=model['timedata'])
tupdates = toptimizer.get_updates(max_gradient_elem=5.,
learning_rate=0.005)
f_t_update = theano.function(model['timedata'],
model['timecost'],
updates=list(tupdates))
# training loop.
start_time = timeit.default_timer()
n_examples = len(train_examples)
batches_per_epoch = n_examples // options['batch_size'] + 1
n_epochs = global_steps // batches_per_epoch + 1
global_step = 0
#cost_history = []
for _ in range(n_epochs):
for _ in range(batches_per_epoch):
batch_data = train_loader()
cost = f_update(*(batch_data[:-3] + (batch_data[-2], )))
#cost_history += [cost]
timecost = f_t_update(*(batch_data[:-2] + (batch_data[-1], )))
if global_step % disp_freq == 0:
print 'global step %d, cost: %f' % (global_step, cost)
print 'timecost: %f' % (timecost)
# dump model parameters.
if global_step % save_freq == 0:
params = unzip(tparams)
np.savez(data_dir + saveto_file, **params)
pickle.dump(
options, open('%s.pkl' % (data_dir + saveto_file),
'wb'), -1)
timeparams = unzip(timetparams)
np.savez(data_dir + tmsaveto_file, **timeparams)
# evaluate on test data.
if global_step % test_freq == 0:
scores = evaluate(model['f_prob'], test_loader,
model['f_tprob'], options['tdim'])
print 'eval scores: ', scores
end_time = timeit.default_timer()
print 'time used: %d seconds.' % (end_time - start_time)
global_step += 1
scores = evaluate(model['f_prob'], test_loader, model['f_tprob'],
options['tdim'])
pprint.pprint(scores)
def main_train(args, model, tokenizer, processor, label_list, device, n_gpu):
train_examples = processor.get_train_examples(args.data_dir)
num_train_optimization_steps = int(
len(train_examples) / args.train_batch_size / args.gradient_accumulation_steps) * args.num_train_epochs
if args.local_rank != -1:
num_train_optimization_steps = num_train_optimization_steps // torch.distributed.get_world_size()
# Prepare optimizer
param_optimizer = list(model.named_parameters())
no_decay = ['bias', 'LayerNorm.bias', 'LayerNorm.weight']
optimizer_grouped_parameters = [
{'params': [p for n, p in param_optimizer if not any(nd in n for nd in no_decay)], 'weight_decay': 0.01},
{'params': [p for n, p in param_optimizer if any(nd in n for nd in no_decay)], 'weight_decay': 0.0}
]
if args.fp16:
print("using fp16")
try:
from apex.optimizers import FusedAdam
except ImportError:
raise ImportError(
"Please install apex from https://www.github.com/nvidia/apex to use distributed and fp16 training.")
optimizer = FusedAdam(optimizer_grouped_parameters,
lr=args.learning_rate,
bias_correction=False)
if args.loss_scale == 0:
model, optimizer = amp.initialize(model, optimizer, opt_level="O2", keep_batchnorm_fp32=False,
loss_scale="dynamic")
else:
model, optimizer = amp.initialize(model, optimizer, opt_level="O2", keep_batchnorm_fp32=False,
loss_scale=args.loss_scale)
scheduler = LinearWarmUpScheduler(optimizer, warmup=args.warmup_proportion,
total_steps=num_train_optimization_steps)
else:
print("using fp32")
optimizer = BertAdam(optimizer_grouped_parameters,
lr=args.learning_rate,
warmup=args.warmup_proportion,
t_total=num_train_optimization_steps)
if args.local_rank != -1:
try:
from apex.parallel import DistributedDataParallel as DDP
except ImportError:
raise ImportError(
"Please install apex from https://www.github.com/nvidia/apex to use distributed and fp16 training.")
model = DDP(model)
elif n_gpu > 1:
model = torch.nn.DataParallel(model)
global_step = 0
nb_tr_steps = 0
tr_loss = 0
average_loss = 0
print("data prep")
train_data = load_examples(args, tokenizer, processor, label_list, "train")
if args.local_rank == -1:
train_sampler = RandomSampler(train_data)
else:
train_sampler = DistributedSampler(train_data)
train_dataloader = DataLoader(train_data, sampler=train_sampler, batch_size=args.train_batch_size)
model.train()
nb_tr_examples = 0
for epoch_num in trange(int(args.num_train_epochs), desc="Epoch"):
if args.max_steps > 0 and global_step > args.max_steps:
break
for step, batch in enumerate(tqdm(train_dataloader, desc="Iteration")):
if args.max_steps > 0 and global_step > args.max_steps:
break
batch = tuple(t.to(device) for t in batch)
input_ids, input_mask, segment_ids, label_ids, valid_ids, label_mask, b_use_valid_filter, \
adj_matrix, dep_matrix = batch
loss = model(input_ids=input_ids, token_type_ids=segment_ids, attention_mask=input_mask, labels=label_ids,
valid_ids=valid_ids, adjacency_matrix=adj_matrix)
if n_gpu > 1:
loss = loss.mean() # mean() to average on multi-gpu.
if args.gradient_accumulation_steps > 1:
loss = loss / args.gradient_accumulation_steps
if args.fp16:
with amp.scale_loss(loss, optimizer) as scaled_loss:
scaled_loss.backward()
else:
loss.backward()
tr_loss += loss.item()
average_loss += loss
nb_tr_examples += input_ids.size(0)
nb_tr_steps += 1
if (step + 1) % args.gradient_accumulation_steps == 0:
if args.fp16:
torch.nn.utils.clip_grad_norm_(amp.master_params(optimizer), 1.0)
else:
torch.nn.utils.clip_grad_norm_(model.parameters(), 1.0)
if args.fp16:
# modify learning rate with special warm up for BERT which FusedAdam doesn't do
scheduler.step()
optimizer.step()
optimizer.zero_grad()
global_step += 1
logging.info("Global Steps:{} Final Loss = {}".format(global_step, average_loss))
average_loss = 0
if args.local_rank == -1 or torch.distributed.get_rank() == 0 or args.world_size <= 1:
# Save model checkpoint
output_dir = os.path.join(args.output_dir, "epoch-{}".format(epoch_num))
if not os.path.exists(output_dir):
os.makedirs(output_dir)
save_zen_model(output_dir, model, args)
loss = tr_loss / nb_tr_steps if args.do_train else None
return loss, global_step