[RELAY] Port from_nnvm to NNVM as to_relay #2144
Conversation
|
Very good idea, lets us get lots of Relay test cases for free |
jroesch
force-pushed
the
nnvm_to_relay
branch
from
f1f8f69 to
2a66ea2
Compare
jroesch
force-pushed
the
nnvm_to_relay
branch
from
27b2b18 to
d8d6371
Compare
jroesch
changed the title
|
@ZihengJiang @tqchen review would be good now, after rebasing I fixed a few bugs, tests are green, just need to fix linting and clean up a few things. |
jroesch
force-pushed
the
nnvm_to_relay
branch
from
241425e to
c3489fa
Compare
|
@siju-samuel, @eqy could you review this @tqchen said it would be good to have you do review. |
nnvm/python/nnvm/to_relay.py
Outdated
|
|
||
| if op_name == "null": | ||
| v = var(node_name, shape=oshape, dtype=odtype) | ||
| params.append(v) |
There was a problem hiding this comment.
This need to be updated to support the multiple output cases, consider directly reuse logic(we can refactor the function to take convert_map as argument) https://github.com/dmlc/tvm/blob/master/python/tvm/relay/frontend/mxnet.py#L469
There was a problem hiding this comment.
There is code below to handle multiple outputs, it will generate a tuple of all the outputs unless I'm missing something.
There was a problem hiding this comment.
I see that now, the two code handled it differently. In the mxnet converter code relay_map is asked to always record a list, tuple, or TuppleWrapper, and when we get the child we always fetch the corresponding index. This is going to be helpful for cases when we need post proc for multiple outputs(which can directly return a list).
The reason that I thought it only works with single output is due to the fact that when you get oshape, you only get output 0(instead of a list of all of them), so for functions like split, you only get the first output
nnvm/python/nnvm/to_relay.py
Outdated
| return op.divide(left, right) | ||
|
|
||
|
|
||
| def _to_rshift(children, attrs, odtype): |
There was a problem hiding this comment.
consider summarizing function as _rbinop_scalar(new_op) as in https://github.com/dmlc/tvm/blob/master/python/tvm/relay/frontend/mxnet.py#L130
nnvm/python/nnvm/to_relay.py
Outdated
| return dense | ||
|
|
||
|
|
||
| def _to_relu(children, attrs, odtype): |
There was a problem hiding this comment.
Likely we can use _rename in https://github.com/dmlc/tvm/blob/master/python/tvm/relay/frontend/mxnet.py#L28
|
This is a little random but I noticed that we are using fancy python3.6 string features here, did we already declare that we are dropping support for python3.5? |
|
We do need to keep it compatible by not using the python3.6 features here |
| def check_model(sym, shapes, dtypes): | ||
| net = nnvm.graph.create(sym) | ||
| graph_json, mod, params = nnvm.compiler.build(net, 'llvm', shape=shapes, dtype=dtypes) | ||
| relay_model = to_relay.from_nnvm(net, shapes, dtypes) |
There was a problem hiding this comment.
Am I missing an import here? When I try to run this it can't find from_nnvm in to_relay.
jroesch
force-pushed
the
nnvm_to_relay
branch
4 times, most recently
from
36b75b5 to
9fe2e8f
Compare
Repair imports First test passes Fix bug in axes passing Add implementations for upsampling and pad Add support for bias with conv ops Rebase fixup Borrow the reshape changes while waiting on apache#2159 Debugging tests Repair after rebase Refactor to use common functionality Fix helper func Remove final eval Remove some debuggin Remove unecessary ws change Fix multiple outputs Tests pass Enable all tests Rebase repair Fix last test case Fix last test case All tests but RNNs pass Fix tests Remove final eval Reformat Fix bugs in MxNet Fix 3.5 support in Relay, and rename Factor common functionality into one file Apply most of code review Fix linting Fix NNVM linting Fix test error Fix MLP test Fix linting error One more linting issue Clean up diff and fix bugs Retrigger disable MLP test Retrigger Retrigger Roll back NNVM change Retrigger Retrigger Retrigger
jroesch
force-pushed
the
nnvm_to_relay
branch
from
9fe2e8f to
f94a529
Compare
5 participants
Add code which converts and NNVM model to a Relay program.