[Relay][Frontend] Add MXNet test example for relay

[icemelon]

• Add MXNet test examples to Relay
• Fix a bug in BiasAddSimplifier. Thanks to @merrymercy for the patch.

[merrymercy]

I already did type correction in
https://github.com/dmlc/tvm/blob/838e7181dd5e16aa5c7a87bbfbf11b5b42a80f1a/src/relay/pass/canonicalize_ops.cc#L29
And the following multiple bias_add example works. Could you give me a broken case?

import tvm
from tvm import relay

x = relay.var("x", shape=(1, 64, 56, 56))
bias = relay.var("bias")
bias2 = relay.var("bias")
weight = relay.var("weight")
y = relay.nn.conv2d(x, weight, channels=64, kernel_size=(3, 3), padding=(1, 1))
y = relay.nn.bias_add(y, bias)
y = relay.nn.bias_add(y, bias2)
y = relay.Function(relay.ir_pass.free_vars(y), y)

a = relay.ir_pass.infer_type(y)
a = relay.ir_pass.canonicalize_ops(a)

print(a.astext())

[icemelon]

If you add another stmt in between these two bias_add, it will fail.

import tvm
from tvm import relay

x = relay.var("x", shape=(1, 64, 56, 56))
bias = relay.var("bias")
bias2 = relay.var("bias")
weight = relay.var("weight")
weight2 = relay.var("weight")
y = relay.nn.conv2d(x, weight, channels=64, kernel_size=(3, 3), padding=(1, 1))
y = relay.nn.bias_add(y, bias)
y = relay.nn.conv2d(y, weight2, channels=128, kernel_size=(3, 3), padding=(1, 1))
y = relay.nn.bias_add(y, bias2)
y = relay.Function(relay.ir_pass.free_vars(y), y)

a = relay.ir_pass.infer_type(y)
a = relay.ir_pass.canonicalize_ops(a)

print(a.astext())

[merrymercy]

Try this one line change!
Other changes in this pr look good to me.

diff --git a/src/relay/pass/canonicalize_ops.cc b/src/relay/pass/canonicalize_ops.cc
index 77cd59e2..4482dc39 100644
--- a/src/relay/pass/canonicalize_ops.cc
+++ b/src/relay/pass/canonicalize_ops.cc
@@ -22,7 +22,7 @@ class BiasAddSimplifier : public ExprMutator {
       CHECK_EQ(call->args.size(), 2);
       const BiasAddAttrs* param = call->attrs.as<BiasAddAttrs>();
 
-      auto ttype = call->args[0]->type_as<TensorTypeNode>();
+      auto ttype = n->args[0]->type_as<TensorTypeNode>();
       size_t n_dim = ttype->shape.size();
       Expr expanded_bias = ExpandBiasToMatchAxis(call->args[1], n_dim, {param->axis});
       Expr ret = Add(call->args[0], expanded_bias);

[icemelon]

Thanks @merrymercy. This change fixed the issue. Updated it this PR.

[tqchen]

This PR seems to be duplicated with https://github.com/dmlc/tvm/tree/master/tests/python/frontend/mxnet should we consolidate everything into one location?