stephenrawls commented on a change in pull request #14208: Add support for fast
variable-length LSTM
URL: https://github.com/apache/incubator-mxnet/pull/14208#discussion_r280663659
##########
File path: tests/python/gpu/test_gluon_gpu.py
##########
@@ -225,6 +226,54 @@ def forward(self, inpt):
assert_allclose(net(data).asnumpy(), ref_net(data).asnumpy())
+def check_layer_bidirectional_varseqlen(size, in_size):
+ class RefBiLSTMVarSeqLen(gluon.Block):
+ def __init__(self, size, **kwargs):
+ super(RefBiLSTMVarSeqLen, self).__init__(**kwargs)
+ with self.name_scope():
+ self._lstm_fwd = gluon.rnn.LSTM(size, bidirectional=False,
prefix='l0')
+ self._lstm_bwd = gluon.rnn.LSTM(size, bidirectional=False,
prefix='r0')
+
+ def forward(self, inpt, sequence_length):
+ fwd = self._lstm_fwd(inpt)
+ bwd_inpt = nd.SequenceReverse(inpt,
sequence_length=sequence_length, use_sequence_length=True)
+ bwd = self._lstm_bwd(bwd_inpt)
+ bwd = nd.SequenceReverse(bwd, sequence_length=sequence_length,
use_sequence_length=True)
+ return nd.concat(fwd, bwd, dim=2)
+ weights = {}
+ for d in ['l', 'r']:
+ weights['lstm_{}0_i2h_weight'.format(d)] =
mx.random.uniform(shape=(size*4, in_size))
+ weights['lstm_{}0_h2h_weight'.format(d)] =
mx.random.uniform(shape=(size*4, size))
+ weights['lstm_{}0_i2h_bias'.format(d)] =
mx.random.uniform(shape=(size*4,))
+ weights['lstm_{}0_h2h_bias'.format(d)] =
mx.random.uniform(shape=(size*4,))
+
+ net = gluon.rnn.LSTM(size, bidirectional=True, use_sequence_length=True,
prefix='lstm_')
+ ref_net = RefBiLSTMVarSeqLen(size, prefix='lstm_')
+ net.initialize()
+ ref_net.initialize()
+ net_params = net.collect_params()
+ ref_net_params = ref_net.collect_params()
+ for k in weights:
+ net_params[k].set_data(weights[k])
+ ref_net_params[k.replace('l0', 'l0l0').replace('r0',
'r0l0')].set_data(weights[k])
+
+
+ batch_size = 10
+ num_timesteps = 11
+ data = mx.random.uniform(shape=(num_timesteps, batch_size, in_size))
+
+ # TODO: figure out why int32 doesn't work here
Review comment:
I'll get to the bottom of this in my next PR after this is merged.
(For next PR I'm planning to add support for NTC layout without transposing,
which is supported by cuDNN, however currently gluon transposes NTC -> TNC:
https://github.com/apache/incubator-mxnet/blob/master/python/mxnet/gluon/rnn/rnn_layer.py#L246-L247)
----------------------------------------------------------------
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.
For queries about this service, please contact Infrastructure at:
[email protected]
With regards,
Apache Git Services
