def build_writers(self):
"""
Build a list of writers to be used. By default it contains
writers that write metrics to the screen,
a json file, and a tensorboard event file respectively.
If you'd like a different list of writers, you can overwrite it in
your trainer.
Returns:
list[EventWriter]: a list of :class:`EventWriter` objects.
It is now implemented by:
.. code-block:: python
return [
CommonMetricPrinter(self.max_iter),
JSONWriter(os.path.join(self.cfg.OUTPUT_DIR, "metrics.json")),
TensorboardXWriter(self.cfg.OUTPUT_DIR),
]
"""
# Assume the default print/log frequency.
# TODO: customize my writers
return [
# It may not always print what you want to see, since it prints "common" metrics only.
CommonMetricPrinter(self.max_iter),
JSONWriter(os.path.join(self.cfg.OUTPUT_DIR, "metrics.json")),
TensorboardXWriter(self.cfg.OUTPUT_DIR),
]
def build_hooks(self):
r"""
Build a list of default hooks, including timing, evaluation,
checkpointing, lr scheduling, precise BN, writing events.
Returns:
list[HookBase]:
"""
cfg = self.cfg.clone()
cfg.defrost()
ret = [
hooks.IterationTimer(),
hooks.LRScheduler(self.optimizer, self.scheduler),
]
if cfg.MODEL.FREEZE_LAYERS != [''] and cfg.SOLVER.FREEZE_ITERS > 0:
freeze_layers = ",".join(cfg.MODEL.FREEZE_LAYERS)
logger.info(
f'Freeze layer group "{freeze_layers}" training for {cfg.SOLVER.FREEZE_ITERS:d} iterations'
)
ret.append(
hooks.FreezeLayer(
self.model,
self.optimizer,
cfg.MODEL.FREEZE_LAYERS,
cfg.SOLVER.FREEZE_ITERS,
))
def test_and_save_results():
self._last_eval_results = self.test(self.cfg, self.model)
return self._last_eval_results
# Do evaluation after checkpointer, because then if it fails,
# we can use the saved checkpoint to debug.
ret.append(TuneReportHook(cfg.TEST.EVAL_PERIOD, test_and_save_results))
# run writers in the end, so that evaluation metrics are written
ret.append(
hooks.PeriodicWriter([CommonMetricPrinter(self.max_iter)], 200))
return ret
def build_hooks(self):
r"""
Build a list of default hooks, including timing, evaluation,
checkpointing, lr scheduling, precise BN, writing events.
Returns:
list[HookBase]:
"""
cfg = self.cfg.clone()
cfg.defrost()
ret = [
hooks.IterationTimer(),
hooks.LRScheduler(self.optimizer, self.scheduler),
]
ret.append(
hooks.LayerFreeze(
self.model,
cfg.MODEL.FREEZE_LAYERS,
cfg.SOLVER.FREEZE_ITERS,
cfg.SOLVER.FREEZE_FC_ITERS,
))
def test_and_save_results():
self._last_eval_results = self.test(self.cfg, self.model)
return self._last_eval_results
# Do evaluation after checkpointer, because then if it fails,
# we can use the saved checkpoint to debug.
ret.append(TuneReportHook(cfg.TEST.EVAL_PERIOD, test_and_save_results))
if comm.is_main_process():
# run writers in the end, so that evaluation metrics are written
ret.append(
hooks.PeriodicWriter([CommonMetricPrinter(self.max_iter)],
200))
return ret
def do_train(cfg, model, resume=False):
data_loader = build_reid_train_loader(cfg)
model.train()
optimizer = build_optimizer(cfg, model)
iters_per_epoch = len(data_loader.dataset) // cfg.SOLVER.IMS_PER_BATCH
scheduler = build_lr_scheduler(cfg, optimizer, iters_per_epoch)
checkpointer = Checkpointer(model,
cfg.OUTPUT_DIR,
save_to_disk=comm.is_main_process(),
optimizer=optimizer**scheduler)
start_epoch = (checkpointer.resume_or_load(
cfg.MODEL.WEIGHTS, resume=resume).get("epoch", -1) + 1)
iteration = start_iter = start_epoch * iters_per_epoch
max_epoch = cfg.SOLVER.MAX_EPOCH
max_iter = max_epoch * iters_per_epoch
warmup_iters = cfg.SOLVER.WARMUP_ITERS
delay_epochs = cfg.SOLVER.DELAY_EPOCHS
periodic_checkpointer = PeriodicCheckpointer(checkpointer,
cfg.SOLVER.CHECKPOINT_PERIOD,
max_epoch)
writers = ([
CommonMetricPrinter(max_iter),
JSONWriter(os.path.join(cfg.OUTPUT_DIR, "metrics.json")),
TensorboardXWriter(cfg.OUTPUT_DIR)
] if comm.is_main_process() else [])
# compared to "train_net.py", we do not support some hooks, such as
# accurate timing, FP16 training and precise BN here,
# because they are not trivial to implement in a small training loop
logger.info("Start training from epoch {}".format(start_epoch))
with EventStorage(start_iter) as storage:
for epoch in range(start_epoch, max_epoch):
storage.epoch = epoch
for data, _ in zip(data_loader, range(iters_per_epoch)):
storage.iter = iteration
loss_dict = model(data)
losses = sum(loss_dict.values())
assert torch.isfinite(losses).all(), loss_dict
loss_dict_reduced = {
k: v.item()
for k, v in comm.reduce_dict(loss_dict).items()
}
losses_reduced = sum(loss
for loss in loss_dict_reduced.values())
if comm.is_main_process():
storage.put_scalars(total_loss=losses_reduced,
**loss_dict_reduced)
optimizer.zero_grad()
losses.backward()
optimizer.step()
storage.put_scalar("lr",
optimizer.param_groups[0]["lr"],
smoothing_hint=False)
if iteration - start_iter > 5 and (
(iteration + 1) % 200 == 0 or iteration == max_iter - 1):
for writer in writers:
writer.write()
iteration += 1
if iteration <= warmup_iters:
scheduler["warmup_sched"].step()
# Write metrics after each epoch
for writer in writers:
writer.write()
if iteration > warmup_iters and (epoch + 1) >= delay_epochs:
scheduler["lr_sched"].step()
if (cfg.TEST.EVAL_PERIOD > 0
and (epoch + 1) % cfg.TEST.EVAL_PERIOD == 0
and epoch != max_iter - 1):
do_test(cfg, model)
# Compared to "train_net.py", the test results are not dumped to EventStorage
periodic_checkpointer.step(epoch)