szha commented on a change in pull request #11482: make gluon rnn layers hybrid
blocks
URL: https://github.com/apache/incubator-mxnet/pull/11482#discussion_r199281972
##########
File path: python/mxnet/gluon/rnn/rnn_layer.py
##########
@@ -173,67 +177,63 @@ def begin_state(self, batch_size=0, func=ndarray.zeros,
**kwargs):
states.append(func(name='%sh0_%d'%(self.prefix, i), **info))
return states
- def forward(self, inputs, states=None):
- batch_size = inputs.shape[self._layout.find('N')]
+ def hybrid_forward(self, F, inputs, states=None, **kwargs):
+ if F is ndarray:
+ batch_size = inputs.shape[self._layout.find('N')]
+ if self._input_size == 0:
+ for i in range(self._dir):
+ self.i2h_weight[i].shape = (self._gates*self._hidden_size,
inputs.shape[2])
+ self.i2h_weight[i]._finish_deferred_init()
skip_states = states is None
if skip_states:
- states = self.begin_state(batch_size, ctx=inputs.context)
- if isinstance(states, ndarray.NDArray):
+ if F is ndarray:
+ states = self.begin_state(batch_size, ctx=inputs.context)
+ else:
+ states = self.begin_state(0, func=symbol.zeros)
+ if isinstance(states, (ndarray.NDArray, symbol.Symbol)):
states = [states]
- for state, info in zip(states, self.state_info(batch_size)):
- if state.shape != info['shape']:
- raise ValueError(
- "Invalid recurrent state shape. Expecting %s, got %s."%(
- str(info['shape']), str(state.shape)))
- if self._input_size == 0:
- for i in range(self._dir):
- self.i2h_weight[i].shape = (self._gates*self._hidden_size,
inputs.shape[2])
- self.i2h_weight[i]._finish_deferred_init()
- if inputs.context.device_type == 'gpu' or \
- self._mode in ['lstm', 'gru'] and not self._dropout:
- out = self._forward_kernel(inputs, states)
- else:
- out = self._forward(inputs, states)
+ if F is ndarray:
+ for state, info in zip(states, self.state_info(batch_size)):
+ if state.shape != info['shape']:
+ raise ValueError(
+ "Invalid recurrent state shape. Expecting %s, got
%s."%(
+ str(info['shape']), str(state.shape)))
+ out = self._forward_kernel(F, inputs, states, **kwargs)
# out is (output, state)
return out[0] if skip_states else out
- def _forward(self, inputs, states):
- """forward using gluon cell"""
- ns = len(states)
- axis = self._layout.find('T')
- states = sum(zip(*((j for j in i) for i in states)), ())
- outputs, states = self._unfused.unroll(
- inputs.shape[axis], inputs, states,
- layout=self._layout, merge_outputs=True)
- new_states = []
- for i in range(ns):
- state = ndarray.concat(*(j.reshape((1,)+j.shape) for j in
states[i::ns]), dim=0)
- new_states.append(state)
-
- return outputs, new_states
-
- def _forward_kernel(self, inputs, states):
+ def __call__(self, inputs, *states):
Review comment:
this is not possible due to the inverse shape inference in concat.
----------------------------------------------------------------
This is an automated message from the Apache Git Service.
To respond to the message, please log on GitHub and use the
URL above to go to the specific comment.
For queries about this service, please contact Infrastructure at:
us...@infra.apache.org
With regards,
Apache Git Services
