apeforest commented on a change in pull request #14491: Enable slice embedding
concat split fuse
URL: https://github.com/apache/incubator-mxnet/pull/14491#discussion_r273351087
##########
File path: tests/python/mkl/test_subgraph.py
##########
@@ -647,6 +647,42 @@ def test_neg_fc_relu():
syms, attrs, excluded_attrs = neg_fc_relu(no_bias, dshape, flatten)
check_neg_fusion(syms, attrs, excluded_attrs, dshape, name='fc')
+@with_seed()
+def test_SliceSplitEmbeddingConcatFuse():
+ DATA_SHAPE = [(1000, 26000)]
+ num_embed_features = 26
+ num_cont_features = 13
+ input_dims = 1000
+ hidden_units = 32
+
+ dns_data = mx.symbol.Variable("dns_data", shape=DATA_SHAPE[0])
+ x = mx.symbol.slice(data=dns_data, begin=(0, 0),
+ end=(None, num_embed_features))
+ embeds = mx.symbol.split(data=x, num_outputs=num_embed_features,
squeeze_axis=1)
+ x = mx.symbol.slice(data=dns_data, begin=(0, num_embed_features),
+ end=(None, num_embed_features + num_cont_features))
+ features = [x]
+ for i, embed in enumerate(embeds):
+ embed_weight = mx.symbol.Variable('embed_%d_weight' % i,
stype='row_sparse')
+ features.append(mx.symbol.sparse.Embedding(data=embed,
weight=embed_weight,
+ input_dim=input_dims, output_dim=hidden_units,
sparse_grad=True))
+ hidden = mx.symbol.concat(*features, dim=1)
+ sym = mx.symbol.SoftmaxOutput(hidden, name='model')
+
+ sym_sg = sym.get_backend_symbol('MKLDNN_WIDE_AND_DEEP_INPUT_FUSE')
+ assert
''.join(sym_sg.attr_dict().keys()).find('SliceSplitEmbeddingConcatFuse') != -1
+
+ arg_shapes, _, aux_shapes = sym.infer_shape()
+ arg_array = [mx.nd.random.uniform(-1, 1, shape=shape) for shape in
arg_shapes]
+ aux_array = [mx.nd.random.uniform(shape=shape) for shape in aux_shapes]
+ exe = sym.bind(ctx=mx.current_context(), args=arg_array,
aux_states=aux_array, grad_req='null')
+ exe.forward()
+ os.environ['MXNET_SUBGRAPH_BACKEND'] = 'MKLDNN_WIDE_AND_DEEP_INPUT_FUSE'
Review comment:
is this environment variable newly introduced? where is the documentation?
----------------------------------------------------------------
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
