szha 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_r281824733
##########
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
+ sequence_length = nd.random.randint(1, num_timesteps+1,
shape=(batch_size)).astype("float")
+
+ net_output = net(data, sequence_length=sequence_length).asnumpy()
+ ref_net_output = ref_net(data, sequence_length).asnumpy()
+ sequence_length_np = sequence_length.asnumpy().astype("int32")
+
+ # Only compare the valid sections for each batch entry
+ for b in range(batch_size):
+ assert_allclose(net_output[:sequence_length_np[b], b],
ref_net_output[:sequence_length_np[b], b])
Review comment:
I'm fine for it to be in a follow-up PR.
----------------------------------------------------------------
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
