modify

xt/model/dqn/dqn_mlp_ms.py

@@ -21,7 +21,7 @@
 from xt.model.model_ms import XTModel_MS
 from zeus.common.util.common import import_config
 from zeus.common.util.register import Registers
-from xt.model.ms_compat import Dense, Adam, MSELoss, Cell, Model, ms
+from xt.model.ms_compat import Dense, Adam, MSELoss, Cell, Model, DynamicLossScaleUpdateCell, ms
 import mindspore.ops as ops
 from xt.model.ms_utils import MSVariables
 from xt.model.dqn.dqn_cnn_ms import MyTrainOneStepCell
@@ -47,7 +47,12 @@ def create_model(self, model_info):
  adam = Adam(params=self.net.trainable_params(), learning_rate=self.learning_rate)
  # model = Model(self.net, loss_fn=loss, optimizer=adam)
  loss_net = ms.nn.WithLossCell(self.net, loss_fn)
- model = MyTrainOneStepCell(loss_net, adam, scale_sense=1, grad_clip=True)
+ device_target = ms.get_context("device_target")
+ if device_target == 'Ascend':
+ manager = DynamicLossScaleUpdateCell(loss_scale_value=2 ** 12, scale_factor=2, scale_window=1000)
+ model = MyTrainOneStepCell(loss_net, adam, manager, grad_clip=True, clipnorm=10.)
+ else:
+ model = MyTrainOneStepCell(loss_net, adam, grad_clip=True, clipnorm=10.)
  self.actor_var = MSVariables(self.net)
  return model
 

xt/model/model_ms.py

@@ -52,15 +52,13 @@ def predict(self, state):
  :param state:
  :return: output tensor ref to policy.model
  """
- state = ms.Tensor(state)
- return self.model.predict(state).asnumpy()
+ return self.model.predict(state)
 
  def train(self, state, label):
  """Train the model."""
  state = ms.Tensor(state, dtype=ms.float32)
  label = ms.Tensor(label, dtype=ms.float32)
- loss = self.model.network(state, label)
- self.actor_var = MSVariables(self.model.network)
+ loss = self.model(state, label)
  return loss.asnumpy().item()
 
  def set_weights(self, weights):

xt/model/ms_compat.py

@@ -46,12 +46,6 @@ def import_ms_compact():
  from mindspore.ops import Cast, MultitypeFuncGraph, ReduceSum, ReduceMax, ReduceMean
  from mindspore import History
 
-def loss_to_val(loss):
- """Make keras instance into value."""
- if isinstance(loss, History):
- loss = loss.history.get("loss")[0]
- return loss
-
 
 DTYPE_MAP = {
  "float32": ms.float32,