def compute_time_train(model, loss_fun):
"""Computes precise model forward + backward time using dummy data."""
# Use train mode
model.train()
# Generate a dummy mini-batch and copy data to GPU
im_size, batch_size = cfg.TRAIN.IM_SIZE, int(cfg.TRAIN.BATCH_SIZE / cfg.NUM_GPUS)
inputs = torch.rand(batch_size, 3, im_size, im_size).cuda(non_blocking=False)
labels = torch.zeros(batch_size, dtype=torch.int64).cuda(non_blocking=False)
# Cache BatchNorm2D running stats
bns = [m for m in model.modules() if isinstance(m, torch.nn.BatchNorm2d)]
bn_stats = [[bn.running_mean.clone(), bn.running_var.clone()] for bn in bns]
# Compute precise forward backward pass time
fw_timer, bw_timer = Timer(), Timer()
total_iter = cfg.PREC_TIME.NUM_ITER + cfg.PREC_TIME.WARMUP_ITER
for cur_iter in range(total_iter):
# Reset the timers after the warmup phase
if cur_iter == cfg.PREC_TIME.WARMUP_ITER:
fw_timer.reset()
bw_timer.reset()
# Forward
fw_timer.tic()
_, preds, _ = model(inputs)
loss = loss_fun(preds, labels)
torch.cuda.synchronize()
fw_timer.toc()
# Backward
bw_timer.tic()
loss.backward()
torch.cuda.synchronize()
bw_timer.toc()
# Restore BatchNorm2D running stats
for bn, (mean, var) in zip(bns, bn_stats):
bn.running_mean, bn.running_var = mean, var
return fw_timer.average_time, bw_timer.average_time
def compute_time_loader(data_loader):
"""Computes loader time."""
timer = Timer()
loader.shuffle(data_loader, 0)
data_loader_iterator = iter(data_loader)
total_iter = cfg.PREC_TIME.NUM_ITER + cfg.PREC_TIME.WARMUP_ITER
total_iter = min(total_iter, len(data_loader))
for cur_iter in range(total_iter):
if cur_iter == cfg.PREC_TIME.WARMUP_ITER:
timer.reset()
timer.tic()
next(data_loader_iterator)
timer.toc()
return timer.average_time
def compute_time_eval(model):
"""Computes precise model forward test time using dummy data."""
# Use eval mode
model.eval()
# Generate a dummy mini-batch and copy data to GPU
im_size, batch_size = cfg.TRAIN.IM_SIZE, int(cfg.TEST.BATCH_SIZE / cfg.NUM_GPUS)
inputs = torch.zeros(batch_size, 3, im_size, im_size).cuda(non_blocking=False)
# Compute precise forward pass time
timer = Timer()
total_iter = cfg.PREC_TIME.NUM_ITER + cfg.PREC_TIME.WARMUP_ITER
for cur_iter in range(total_iter):
# Reset the timers after the warmup phase
if cur_iter == cfg.PREC_TIME.WARMUP_ITER:
timer.reset()
# Forward
timer.tic()
model(inputs)
torch.cuda.synchronize()
timer.toc()
return timer.average_time
class TrainMeter(object):
"""Measures training stats."""
def __init__(self, epoch_iters):
self.epoch_iters = epoch_iters
self.max_iter = cfg.OPTIM.MAX_EPOCH * epoch_iters
self.iter_timer = Timer()
self.desc_loss = ScalarMeter(cfg.LOG_PERIOD)
self.desc_loss_total = 0.0
self.att_loss = ScalarMeter(cfg.LOG_PERIOD)
self.att_loss_total = 0.0
self.lr = None
# Current minibatch errors (smoothed over a window)
self.mb_top1_err = ScalarMeter(cfg.LOG_PERIOD)
self.mb_top5_err = ScalarMeter(cfg.LOG_PERIOD)
self.mb_att_top1_err = ScalarMeter(cfg.LOG_PERIOD)
self.mb_att_top5_err = ScalarMeter(cfg.LOG_PERIOD)
# Number of misclassified examples
self.num_top1_mis = 0
self.num_top5_mis = 0
self.num_att_top1_mis = 0
self.num_att_top5_mis = 0
self.num_samples = 0
def reset(self, timer=False):
if timer:
self.iter_timer.reset()
self.desc_loss.reset()
self.att_loss.reset()
self.desc_loss_total = 0.0
self.att_loss_total = 0.0
self.lr = None
self.mb_top1_err.reset()
self.mb_top5_err.reset()
self.mb_att_top1_err = ScalarMeter(cfg.LOG_PERIOD)
self.mb_att_top5_err = ScalarMeter(cfg.LOG_PERIOD)
self.num_top1_mis = 0
self.num_top5_mis = 0
self.num_att_top1_mis = 0
self.num_att_top5_mis = 0
self.num_samples = 0
def iter_tic(self):
self.iter_timer.tic()
def iter_toc(self):
self.iter_timer.toc()
def update_stats(self, desc_top1_err, desc_top5_err, att_top1_err,
att_top5_err, desc_loss, att_loss, lr, mb_size):
# Current minibatch stats
self.mb_top1_err.add_value(desc_top1_err)
self.mb_top5_err.add_value(desc_top5_err)
self.desc_loss.add_value(desc_loss)
self.mb_att_top1_err.add_value(att_top1_err)
self.mb_att_top5_err.add_value(att_top5_err)
self.att_loss.add_value(att_loss)
self.lr = lr
# Aggregate stats
self.num_top1_mis += desc_top1_err * mb_size
self.num_top5_mis += desc_top5_err * mb_size
self.num_att_top1_mis += att_top1_err * mb_size
self.num_att_top5_mis += att_top5_err * mb_size
self.desc_loss_total += desc_loss * mb_size
self.att_loss_total += att_loss * mb_size
self.num_samples += mb_size
def get_iter_stats(self, cur_epoch, cur_iter):
cur_iter_total = cur_epoch * self.epoch_iters + cur_iter + 1
eta_sec = self.iter_timer.average_time * (self.max_iter -
cur_iter_total)
mem_usage = gpu_mem_usage()
stats = {
"epoch": "{}/{}".format(cur_epoch + 1, cfg.OPTIM.MAX_EPOCH),
"iter": "{}/{}".format(cur_iter + 1, self.epoch_iters),
"time_avg": self.iter_timer.average_time,
"time_diff": self.iter_timer.diff,
"eta": time_string(eta_sec),
"desc_top1_err": self.mb_top1_err.get_win_median(),
"desc_top5_err": self.mb_top5_err.get_win_median(),
"desc_loss": self.desc_loss.get_win_median(),
"att_top1_err": self.mb_att_top1_err.get_win_median(),
"att_top5_err": self.mb_att_top5_err.get_win_median(),
"att_loss": self.att_loss.get_win_median(),
"lr": self.lr,
"mem": int(np.ceil(mem_usage)),
}
return stats
def log_iter_stats(self, cur_epoch, cur_iter):
if (cur_iter + 1) % cfg.LOG_PERIOD != 0:
return
stats = self.get_iter_stats(cur_epoch, cur_iter)
logger.info(logging.dump_log_data(stats, "train_iter"))
def get_epoch_stats(self, cur_epoch):
cur_iter_total = (cur_epoch + 1) * self.epoch_iters
eta_sec = self.iter_timer.average_time * (self.max_iter -
cur_iter_total)
mem_usage = gpu_mem_usage()
desc_top1_err = self.num_top1_mis / self.num_samples
desc_top5_err = self.num_top5_mis / self.num_samples
desc_avg_loss = self.desc_loss_total / self.num_samples
att_top1_err = self.num_att_top1_mis / self.num_samples
att_top5_err = self.num_att_top5_mis / self.num_samples
att_avg_loss = self.att_loss_total / self.num_samples
stats = {
"epoch": "{}/{}".format(cur_epoch + 1, cfg.OPTIM.MAX_EPOCH),
"time_avg": self.iter_timer.average_time,
"eta": time_string(eta_sec),
"desc_top1_err": desc_top1_err,
"desc_top5_err": desc_top5_err,
"desc_loss": desc_avg_loss,
"att_top1_err": att_top1_err,
"att_top5_err": att_top5_err,
"att_loss": att_avg_loss,
"lr": self.lr,
"mem": int(np.ceil(mem_usage)),
}
return stats
def log_epoch_stats(self, cur_epoch):
stats = self.get_epoch_stats(cur_epoch)
logger.info(logging.dump_log_data(stats, "train_epoch"))
class TestMeter(object):
"""Measures testing stats (only desc)."""
def __init__(self, max_iter):
self.max_iter = max_iter
self.iter_timer = Timer()
# Current minibatch errors (smoothed over a window)
self.mb_top1_err = ScalarMeter(cfg.LOG_PERIOD)
self.mb_top5_err = ScalarMeter(cfg.LOG_PERIOD)
# Min errors (over the full test set)
self.min_top1_err = 100.0
self.min_top5_err = 100.0
# Number of misclassified examples
self.num_top1_mis = 0
self.num_top5_mis = 0
self.num_samples = 0
def reset(self, min_errs=False):
if min_errs:
self.min_top1_err = 100.0
self.min_top5_err = 100.0
self.iter_timer.reset()
self.mb_top1_err.reset()
self.mb_top5_err.reset()
self.num_top1_mis = 0
self.num_top5_mis = 0
self.num_samples = 0
def iter_tic(self):
self.iter_timer.tic()
def iter_toc(self):
self.iter_timer.toc()
def update_stats(self, top1_err, top5_err, mb_size):
self.mb_top1_err.add_value(top1_err)
self.mb_top5_err.add_value(top5_err)
self.num_top1_mis += top1_err * mb_size
self.num_top5_mis += top5_err * mb_size
self.num_samples += mb_size
def get_iter_stats(self, cur_epoch, cur_iter):
mem_usage = gpu_mem_usage()
iter_stats = {
"epoch": "{}/{}".format(cur_epoch + 1, cfg.OPTIM.MAX_EPOCH),
"iter": "{}/{}".format(cur_iter + 1, self.max_iter),
"time_avg": self.iter_timer.average_time,
"time_diff": self.iter_timer.diff,
"top1_err": self.mb_top1_err.get_win_median(),
"top5_err": self.mb_top5_err.get_win_median(),
"mem": int(np.ceil(mem_usage)),
}
return iter_stats
def log_iter_stats(self, cur_epoch, cur_iter):
if (cur_iter + 1) % cfg.LOG_PERIOD != 0:
return
stats = self.get_iter_stats(cur_epoch, cur_iter)
logger.info(logging.dump_log_data(stats, "test_iter"))
def get_epoch_stats(self, cur_epoch):
top1_err = self.num_top1_mis / self.num_samples
top5_err = self.num_top5_mis / self.num_samples
self.min_top1_err = min(self.min_top1_err, top1_err)
self.min_top5_err = min(self.min_top5_err, top5_err)
mem_usage = gpu_mem_usage()
stats = {
"epoch": "{}/{}".format(cur_epoch + 1, cfg.OPTIM.MAX_EPOCH),
"time_avg": self.iter_timer.average_time,
"top1_err": top1_err,
"top5_err": top5_err,
"min_top1_err": self.min_top1_err,
"min_top5_err": self.min_top5_err,
"mem": int(np.ceil(mem_usage)),
}
return stats
def log_epoch_stats(self, cur_epoch):
stats = self.get_epoch_stats(cur_epoch)
logger.info(logging.dump_log_data(stats, "test_epoch"))
