def pool_out(parameters, config, gpu_list, _outname):
model = parameters["model"]
dataset = parameters["test_dataset"]
model.eval()
acc_result = None
total_loss = 0
cnt = 0
total_len = len(dataset)
start_time = timer()
output_info = "Pool_Out"
output_time = config.getint("output", "output_time")
save_step = config.getint("output", "save_step")
step = -1
result = []
for step, data in enumerate(dataset):
for key in data.keys():
if isinstance(data[key], torch.Tensor):
if len(gpu_list) > 0:
data[key] = Variable(data[key].cuda())
else:
data[key] = Variable(data[key])
results = model(data, config, gpu_list, acc_result, "poolout")
result = result + results["output"]
cnt += 1
if step % output_time == 0:
delta_t = timer() - start_time
output_value(0, "poolout", "%d/%d" % (step + 1, total_len), "%s/%s" % (
gen_time_str(delta_t), gen_time_str(delta_t * (total_len - step - 1) / (step + 1))),
"%.3lf" % (total_loss / (step + 1)), output_info, '\r', config)
if save_step > 0 and step % save_step == 0:
out_file = open(_outname, 'w', encoding='utf-8')
for item in result:
tmp_dict = {
'id_': item[0],
'res': item[1]
}
out_line = json.dumps(tmp_dict, ensure_ascii=False) + '\n'
out_file.write(out_line)
out_file.close()
if step == -1:
logger.error("There is no data given to the model in this epoch, check your data.")
raise NotImplementedError
delta_t = timer() - start_time
output_info = "Pool_Out"
output_value(0, "poolout", "%d/%d" % (step + 1, total_len), "%s/%s" % (
gen_time_str(delta_t), gen_time_str(delta_t * (total_len - step - 1) / (step + 1))),
"%.3lf" % (total_loss / (step + 1)), output_info, None, config)
return result
def test(parameters, config, gpu_list):
model = parameters["model"]
dataset = parameters["test_dataset"]
model.eval()
acc_result = None
total_loss = 0
cnt = 0
total_len = len(dataset)
start_time = timer()
output_info = "testing"
output_time = config.getint("output", "output_time")
step = -1
result = []
for step, data in enumerate(dataset):
for key in data.keys():
if isinstance(data[key], torch.Tensor):
if len(gpu_list) > 0:
data[key] = Variable(data[key].cuda())
else:
data[key] = Variable(data[key])
results = model(data, config, gpu_list, acc_result, "test")
result = result + results["output"]
cnt += 1
if step % output_time == 0:
delta_t = timer() - start_time
output_value(
0, "test", "%d/%d" % (step + 1, total_len), "%s/%s" %
(gen_time_str(delta_t),
gen_time_str(delta_t * (total_len - step - 1) / (step + 1))),
"%.3lf" % (total_loss / (step + 1)), output_info, '\r', config)
if step == -1:
logger.error(
"There is no data given to the model in this epoch, check your data."
)
raise NotImplementedError
delta_t = timer() - start_time
output_info = "testing"
output_value(
0, "test", "%d/%d" % (step + 1, total_len), "%s/%s" %
(gen_time_str(delta_t),
gen_time_str(delta_t * (total_len - step - 1) / (step + 1))),
"%.3lf" % (total_loss / (step + 1)), output_info, None, config)
res = {}
for x in result:
res[x["id"]] = x["answer"]
return res
def train(parameters, config, gpu_list):
epoch = config.getint("train", "epoch")
batch_size = config.getint("train", "batch_size")
output_time = config.getint("output", "output_time")
test_time = config.getint("output", "test_time")
output_path = os.path.join(config.get("output", "model_path"),
config.get("output", "model_name"))
if os.path.exists(output_path):
logger.warning(
"Output path exists, check whether need to change a name of model")
os.makedirs(output_path, exist_ok=True)
trained_epoch = parameters["trained_epoch"]
model = parameters["model"]
optimizer = parameters["optimizer"]
dataset = parameters["train_dataset"]
global_step = parameters["global_step"]
output_function = parameters["output_function"]
if config.getboolean('train', 'fp16'):
opt_level = 'O2'
model, optimizer = amp.initialize(model,
optimizer,
opt_level=opt_level)
step_size = config.getint("train", "step_size")
gamma = config.getfloat("train", "lr_multiplier")
exp_lr_scheduler = lr_scheduler.StepLR(optimizer,
step_size=step_size,
gamma=gamma)
exp_lr_scheduler.step(trained_epoch)
logger.info("Training start....")
print("Epoch Stage Iterations Time Usage Loss Output Information")
total_len = len(dataset)
more = ""
if total_len < 10000:
more = "\t"
for epoch_num in range(trained_epoch, epoch):
start_time = timer()
current_epoch = epoch_num
exp_lr_scheduler.step(current_epoch)
acc_result = None
total_loss = 0
output_info = ""
step = -1
try:
train_steps = config.getint('train', 'train_steps')
except:
train_steps = 1
for step, data in enumerate(dataset):
for key in data.keys():
if isinstance(data[key], torch.Tensor):
if len(gpu_list) > 0:
data[key] = Variable(data[key].cuda())
else:
data[key] = Variable(data[key])
data['epoch'] = epoch_num
results = model(data, config, gpu_list, acc_result, "train")
loss, acc_result = results["loss"], results["acc_result"]
loss = loss.mean()
total_loss += float(loss)
loss = loss / train_steps
if config.getboolean('train', 'fp16'):
with amp.scale_loss(loss, optimizer) as scaled_loss:
scaled_loss.backward()
else:
loss.backward()
if step % train_steps == 0:
if config.getboolean('train', 'fp16'):
torch.nn.utils.clip_grad_norm_(
amp.master_params(optimizer), 1.0)
else:
torch.nn.utils.clip_grad_norm_(model.parameters(), 1.0)
optimizer.step()
optimizer.zero_grad()
if step % output_time == 0:
output_info = output_function(acc_result, config, mode='train')
delta_t = timer() - start_time
output_value(
current_epoch, "train", "%d/%d" % (step + 1, total_len),
"%s/%s" % (gen_time_str(delta_t),
gen_time_str(delta_t * (total_len - step - 1) /
(step + 1))),
"%.3lf" % (total_loss / (step + 1)), output_info, '\r',
config)
global_step += 1
output_value(
current_epoch, "train", "%d/%d" % (step + 1, total_len), "%s/%s" %
(gen_time_str(delta_t),
gen_time_str(delta_t * (total_len - step - 1) / (step + 1))),
"%.3lf" % (total_loss / (step + 1)), output_info, None, config)
if step == -1:
logger.error(
"There is no data given to the model in this epoch, check your data."
)
raise NotImplementedError
if config.getboolean('train', 'pre_train'):
save_path = os.path.join(output_path, 'epoch_%d' % current_epoch)
os.makedirs(save_path, exist_ok=True)
model.save_pretrained(save_path)
#else:
checkpoint(os.path.join(output_path, "%d.pkl" % current_epoch), model,
optimizer, current_epoch, config, global_step)
if not config.getboolean('train', 'no_valid'):
if current_epoch % test_time == 0:
with torch.no_grad():
valid(model, parameters["valid_dataset"], current_epoch,
config, gpu_list, output_function)
def train(parameters, config, gpu_list, do_test=False, local_rank=-1):
epoch = config.getint("train", "epoch")
batch_size = config.getint("train", "batch_size")
output_time = config.getint("output", "output_time")
test_time = config.getint("output", "test_time")
output_path = os.path.join(config.get("output", "model_path"), config.get("output", "model_name"))
if os.path.exists(output_path):
logger.warning("Output path exists, check whether need to change a name of model")
os.makedirs(output_path, exist_ok=True)
trained_epoch = parameters["trained_epoch"] + 1
model = parameters["model"]
optimizer = parameters["optimizer"]
dataset = parameters["train_dataset"]
global_step = parameters["global_step"]
output_function = parameters["output_function"]
if do_test:
init_formatter(config, ["test"])
test_dataset = init_test_dataset(config)
if trained_epoch == 0:
shutil.rmtree(
os.path.join(config.get("output", "tensorboard_path"), config.get("output", "model_name")), True)
os.makedirs(os.path.join(config.get("output", "tensorboard_path"), config.get("output", "model_name")),
exist_ok=True)
writer = SummaryWriter(os.path.join(config.get("output", "tensorboard_path"), config.get("output", "model_name")),
config.get("output", "model_name"))
step_size = config.getint("train", "step_size")
gamma = config.getfloat("train", "lr_multiplier")
exp_lr_scheduler = lr_scheduler.StepLR(optimizer, step_size=step_size, gamma=gamma)
exp_lr_scheduler.step(trained_epoch)
logger.info("Training start....")
print("Epoch Stage Iterations Time Usage Loss Output Information")
total_len = len(dataset)
more = ""
if total_len < 10000:
more = "\t"
for epoch_num in range(trained_epoch, epoch):
start_time = timer()
current_epoch = epoch_num
exp_lr_scheduler.step(current_epoch)
acc_result = None
total_loss = 0
output_info = ""
step = -1
for step, data in enumerate(dataset):
for key in data.keys():
if isinstance(data[key], torch.Tensor):
if len(gpu_list) > 0:
data[key] = Variable(data[key].cuda())
else:
data[key] = Variable(data[key])
optimizer.zero_grad()
results = model(data, config, gpu_list, acc_result, "train")
loss, acc_result = results["loss"], results["acc_result"]
total_loss += float(loss)
loss.backward()
optimizer.step()
if step % output_time == 0 and local_rank <= 0:
output_info = output_function(acc_result, config)
delta_t = timer() - start_time
output_value(current_epoch, "train", "%d/%d" % (step + 1, total_len), "%s/%s" % (
gen_time_str(delta_t), gen_time_str(delta_t * (total_len - step - 1) / (step + 1))),
"%.3lf" % (total_loss / (step + 1)), output_info, '\r', config)
global_step += 1
writer.add_scalar(config.get("output", "model_name") + "_train_iter", float(loss), global_step)
# break
if local_rank <= 0:
output_info = output_function(acc_result, config)
delta_t = timer() - start_time
output_value(current_epoch, "train", "%d/%d" % (step + 1, total_len), "%s/%s" % (
gen_time_str(delta_t), gen_time_str(delta_t * (total_len - step - 1) / (step + 1))),
"%.3lf" % (total_loss / (step + 1)), output_info, None, config)
if step == -1:
logger.error("There is no data given to the model in this epoch, check your data.")
raise NotImplementedError
if local_rank <= 0:
checkpoint(os.path.join(output_path, "%d.pkl" % current_epoch), model, optimizer, current_epoch, config,
global_step)
writer.add_scalar(config.get("output", "model_name") + "_train_epoch", float(total_loss) / (step + 1),
current_epoch)
if current_epoch % test_time == 0:
with torch.no_grad():
valid(model, parameters["valid_dataset"], current_epoch, writer, config, gpu_list, output_function)
if do_test:
valid(model, test_dataset, current_epoch, writer, config, gpu_list, output_function, mode="test")
if local_rank >= 0:
torch.distributed.barrier()
def train(parameters, config, gpu_list):
epoch = config.getint("train", "epoch")
batch_size = config.getint("train", "batch_size")
output_file = config.get("output", "output_file") if config.get(
"output", "output_file") != "None" else None
if output_file is not None:
if os.path.exists(output_file):
os.remove(output_file)
output_time = config.getint("output", "output_time")
test_time = config.getint("output", "test_time")
output_path = os.path.join(config.get("output", "model_path"),
config.get("output", "model_name"))
if os.path.exists(output_path):
logger.warning(
"Output path exists, check whether need to change a name of model")
os.makedirs(output_path, exist_ok=True)
trained_epoch = parameters["trained_epoch"]
model = parameters["model"]
optimizer = parameters["optimizer"]
dataset = parameters["train_dataset"]
global_step = parameters["global_step"]
output_function = parameters["output_function"]
step_size = config.getint("train", "step_size")
gamma = config.getfloat("train", "lr_multiplier")
exp_lr_scheduler = lr_scheduler.StepLR(optimizer,
step_size=step_size,
gamma=gamma)
exp_lr_scheduler.step(trained_epoch)
logger.info("Training start....")
if config.get("model", "model_name") in ["Pipeline"]:
print(
"Epoch Stage Iterations Time Usage Loss \tOutput Information"
)
else:
print(
"Epoch Stage Iterations Time Usage Loss Output Information"
)
total_len = len(dataset)
more = ""
if total_len < 10000:
more = "\t"
for epoch_num in range(trained_epoch, epoch):
start_time = timer()
current_epoch = epoch_num
exp_lr_scheduler.step(current_epoch)
acc_result = [None, None]
total_loss = None
output_info = ""
step = -1
for step, data in enumerate(dataset):
for key in data.keys():
if isinstance(data[key], torch.Tensor):
if len(gpu_list) > 0:
data[key] = Variable(data[key].cuda())
else:
data[key] = Variable(data[key])
optimizer.zero_grad()
results = model(data, config, gpu_list, acc_result, "train")
loss, acc_result = results["loss"], results["acc_result"]
if type(loss) == list:
loss[0].backward()
loss[1].backward()
if total_loss == None:
total_loss = [0, 0]
total_loss[0] += float(loss[0])
total_loss[1] += float(loss[1])
else:
loss.backward()
if total_loss == None:
total_loss = 0
total_loss += float(loss)
optimizer.step()
if step % output_time == 0:
output_stage_1 = output_function(acc_result[0], config)
output_stage_2 = output_function(acc_result[1], config)
delta_t = timer() - start_time
if type(loss) == list:
output_value(
current_epoch, "train",
"%d/%d" % (step + 1, total_len), "%s/%s" %
(gen_time_str(delta_t),
gen_time_str(delta_t * (total_len - step - 1) /
(step + 1))), "%.3lf, %.3lf\t" %
(total_loss[0] / (step + 1), total_loss[1] /
(step + 1)), output_stage_1 + output_stage_2, '\r',
config)
else:
output_value(
current_epoch, "train",
"%d/%d" % (step + 1, total_len), "%s/%s" %
(gen_time_str(delta_t),
gen_time_str(delta_t * (total_len - step - 1) /
(step + 1))),
"%.3lf" % (total_loss / (step + 1)),
output_stage_1 + output_stage_2, '\r', config)
global_step += 1
if type(loss) == list:
output_value(
current_epoch, "train", "%d/%d" % (step + 1, total_len),
"%s/%s" % (gen_time_str(delta_t),
gen_time_str(delta_t * (total_len - step - 1) /
(step + 1))), "%.3lf, %.3lf\t" %
(total_loss[0] / (step + 1), total_loss[1] / (step + 1)),
output_stage_1 + output_stage_2, None, config)
else:
output_value(
current_epoch, "train", "%d/%d" % (step + 1, total_len),
"%s/%s" % (gen_time_str(delta_t),
gen_time_str(delta_t * (total_len - step - 1) /
(step + 1))),
"%.3lf" % (total_loss / (step + 1)), output_info, None, config)
if output_file is not None:
f = open(output_file, "a")
f.write("epoch = %d" % epoch_num +
json.dumps(print_IoU(acc_result[1])) + "\n")
f.close()
if step == -1:
logger.error(
"There is no data given to the model in this epoch, check your data."
)
raise NotImplementedError
checkpoint(os.path.join(output_path, "%d.pkl" % current_epoch), model,
optimizer, current_epoch, config, global_step)
if current_epoch % test_time == 0:
with torch.no_grad():
valid(model, parameters["valid_dataset"], current_epoch, None,
config, gpu_list, output_function)