use same network in TrainOneStepCell

This commit is contained in:
huangbingjian 2021-03-18 17:22:02 +08:00
parent eaecc83ec2
commit d925490301
6 changed files with 12 additions and 22 deletions

View File

@ -46,8 +46,6 @@ class LossCallBack(Callback):
self._per_print_times = per_print_times
self.count = 0
self.rpn_loss_sum = 0
self.rpn_cls_loss_sum = 0
self.rpn_reg_loss_sum = 0
self.rank_id = rank_id
global time_stamp_init, time_stamp_first
@ -57,14 +55,10 @@ class LossCallBack(Callback):
def step_end(self, run_context):
cb_params = run_context.original_args()
rpn_loss = cb_params.net_outputs[0].asnumpy()
rpn_cls_loss = cb_params.net_outputs[1].asnumpy()
rpn_reg_loss = cb_params.net_outputs[2].asnumpy()
rpn_loss = cb_params.net_outputs.asnumpy()
self.count += 1
self.rpn_loss_sum += float(rpn_loss)
self.rpn_cls_loss_sum += float(rpn_cls_loss)
self.rpn_reg_loss_sum += float(rpn_reg_loss)
cur_step_in_epoch = (cb_params.cur_step_num - 1) % cb_params.batch_num + 1
@ -72,12 +66,10 @@ class LossCallBack(Callback):
global time_stamp_first
time_stamp_current = time.time()
rpn_loss = self.rpn_loss_sum / self.count
rpn_cls_loss = self.rpn_cls_loss_sum / self.count
rpn_reg_loss = self.rpn_reg_loss_sum / self.count
loss_file = open("./loss_{}.log".format(self.rank_id), "a+")
loss_file.write("%lu epoch: %s step: %s ,rpn_loss: %.5f, rpn_cls_loss: %.5f, rpn_reg_loss: %.5f"%
loss_file.write("%lu epoch: %s step: %s rpn_loss: %.5f"%
(time_stamp_current - time_stamp_first, cb_params.cur_epoch_num, cur_step_in_epoch,
rpn_loss, rpn_cls_loss, rpn_reg_loss))
rpn_loss))
loss_file.write("\n")
loss_file.close()
@ -123,18 +115,16 @@ class TrainOneStepCell(nn.Cell):
Args:
network (Cell): The training network.
network_backbone (Cell): The forward network.
optimizer (Cell): Optimizer for updating the weights.
sens (Number): The adjust parameter. Default value is 1.0.
reduce_flag (bool): The reduce flag. Default value is False.
mean (bool): Allreduce method. Default value is False.
degree (int): Device number. Default value is None.
"""
def __init__(self, network, network_backbone, optimizer, sens=1.0, reduce_flag=False, mean=True, degree=None):
def __init__(self, network, optimizer, sens=1.0, reduce_flag=False, mean=True, degree=None):
super(TrainOneStepCell, self).__init__(auto_prefix=False)
self.network = network
self.network.set_grad()
self.backbone = network_backbone
self.weights = ParameterTuple(network.trainable_params())
self.optimizer = optimizer
self.grad = C.GradOperation(get_by_list=True,
@ -146,8 +136,8 @@ class TrainOneStepCell(nn.Cell):
def construct(self, x, gt_bbox, gt_label, gt_num, img_shape=None):
weights = self.weights
rpn_loss, _, _, rpn_cls_loss, rpn_reg_loss = self.backbone(x, gt_bbox, gt_label, gt_num, img_shape)
loss = self.network(x, gt_bbox, gt_label, gt_num, img_shape)
grads = self.grad(self.network, weights)(x, gt_bbox, gt_label, gt_num, img_shape, self.sens)
if self.reduce_flag:
grads = self.grad_reducer(grads)
return F.depend(rpn_loss, self.optimizer(grads)), rpn_cls_loss, rpn_reg_loss
return F.depend(loss, self.optimizer(grads))

View File

@ -100,10 +100,10 @@ if __name__ == '__main__':
weight_decay=config.weight_decay, loss_scale=config.loss_scale)
net_with_loss = WithLossCell(net, loss)
if args_opt.run_distribute:
net = TrainOneStepCell(net_with_loss, net, opt, sens=config.loss_scale, reduce_flag=True,
net = TrainOneStepCell(net_with_loss, opt, sens=config.loss_scale, reduce_flag=True,
mean=True, degree=device_num)
else:
net = TrainOneStepCell(net_with_loss, net, opt, sens=config.loss_scale)
net = TrainOneStepCell(net_with_loss, opt, sens=config.loss_scale)
time_cb = TimeMonitor(data_size=dataset_size)
loss_cb = LossCallBack(rank_id=rank)

View File

@ -69,7 +69,7 @@ class LossCallBack(Callback):
total_loss = self.loss_sum / self.count
loss_file = open("./loss_{}.log".format(self.rank_id), "a+")
loss_file.write("%lu epoch: %s step: %s ,total_loss: %.5f" %
loss_file.write("%lu epoch: %s step: %s total_loss: %.5f" %
(time_stamp_current - time_stamp_first, cb_params.cur_epoch_num, cur_step_in_epoch,
total_loss))
loss_file.write("\n")

View File

@ -69,7 +69,7 @@ class LossCallBack(Callback):
total_loss = self.loss_sum / self.count
loss_file = open("./loss_{}.log".format(self.rank_id), "a+")
loss_file.write("%lu epoch: %s step: %s ,total_loss: %.5f" %
loss_file.write("%lu epoch: %s step: %s total_loss: %.5f" %
(time_stamp_current - time_stamp_first, cb_params.cur_epoch_num, cur_step_in_epoch,
total_loss))
loss_file.write("\n")

View File

@ -68,7 +68,7 @@ class LossCallBack(Callback):
total_loss = self.loss_sum / self.count
loss_file = open("./loss_{}.log".format(self.rank_id), "a+")
loss_file.write("%lu epoch: %s step: %s ,total_loss: %.5f" %
loss_file.write("%lu epoch: %s step: %s total_loss: %.5f" %
(time_stamp_current - time_stamp_first, cb_params.cur_epoch_num, cur_step_in_epoch,
total_loss))
loss_file.write("\n")

View File

@ -97,7 +97,7 @@ class LossCallBack(Callback):
total_loss = self.loss_sum/self.count
loss_file = open("./loss_{}.log".format(self.rank_id), "a+")
loss_file.write("%lu epoch: %s step: %s ,total_loss: %.5f" %
loss_file.write("%lu epoch: %s step: %s total_loss: %.5f" %
(time_stamp_current - time_stamp_first, cb_params.cur_epoch_num, cur_step_in_epoch,
total_loss))
loss_file.write("\n")