mirrors
/
PaddleClas
mirror of https://github.com/PaddlePaddle/PaddleClas.git
1
0
Fork 0
Code Issues Projects Releases Wiki Activity

Support training without amp

pull/2671/head
tianyi1997 2023-02-17 15:09:23 +08:00 committed by HydrogenSulfate
parent 4553d22cf1
commit f313a6d873
1 changed files with 17 additions and 9 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

26
ppcls/engine/train/train_metabin.py
View File

@ -98,7 +98,6 @@ def train_epoch_metabin(engine, epoch_id, print_batch_step):
for key, value in mtest_loss_dict.items()} for key, value in mtest_loss_dict.items()}
} }
# step lr (by iter) # step lr (by iter)
# the last lr_sch is cyclic_lr
for i in range(len(engine.lr_sch)): for i in range(len(engine.lr_sch)):
if not getattr(engine.lr_sch[i], "by_epoch", False): if not getattr(engine.lr_sch[i], "by_epoch", False):
engine.lr_sch[i].step() engine.lr_sch[i].step()
@ -117,7 +116,6 @@ def train_epoch_metabin(engine, epoch_id, print_batch_step):
tic = time.time() tic = time.time()
# step lr(by epoch) # step lr(by epoch)
# the last lr_sch is cyclic_lr
for i in range(len(engine.lr_sch)): for i in range(len(engine.lr_sch)):
if getattr(engine.lr_sch[i], "by_epoch", False) and \ if getattr(engine.lr_sch[i], "by_epoch", False) and \
type_name(engine.lr_sch[i]) != "ReduceOnPlateau": type_name(engine.lr_sch[i]) != "ReduceOnPlateau":
@ -191,10 +189,16 @@ def get_meta_data(meta_dataloader_iter, num_domain):
def forward(engine, batch, loss_func): def forward(engine, batch, loss_func):
batch_info = defaultdict() batch_info = defaultdict()
batch_info = {"label": batch[1], "domain": batch[2]} batch_info = {"label": batch[1], "domain": batch[2]}
amp_level = engine.config["AMP"].get("level", "O1").upper() if engine.amp:
with paddle.amp.auto_cast( amp_level = engine.config["AMP"].get("level", "O1").upper()
custom_black_list={"flatten_contiguous_range", "greater_than"}, with paddle.amp.auto_cast(
level=amp_level): custom_black_list={
"flatten_contiguous_range", "greater_than"
},
level=amp_level):
out = engine.model(batch[0], batch[1])
loss_dict = loss_func(out, batch_info)
else:
out = engine.model(batch[0], batch[1]) out = engine.model(batch[0], batch[1])
loss_dict = loss_func(out, batch_info) loss_dict = loss_func(out, batch_info)
return out, loss_dict return out, loss_dict
@ -202,9 +206,13 @@ def forward(engine, batch, loss_func):
def backward(engine, loss, optimizer): def backward(engine, loss, optimizer):
optimizer.clear_grad() optimizer.clear_grad()
scaled = engine.scaler.scale(loss) if engine.amp:
scaled.backward() scaled = engine.scaler.scale(loss)
engine.scaler.minimize(optimizer, scaled) scaled.backward()
engine.scaler.minimize(optimizer, scaled)
else:
loss.backward()
optimizer.step()
for name, layer in engine.model.backbone.named_sublayers(): for name, layer in engine.model.backbone.named_sublayers():
if "gate" == name.split('.')[-1]: if "gate" == name.split('.')[-1]:
layer.clip_gate() layer.clip_gate()