[Relay][Pass] Add a relay pass to extract fake quantized ops #10089
Conversation
margaretqian
requested review from
MarisaKirisame,
ZihengJiang,
anijain2305,
areusch,
comaniac,
icemelon,
jroesch,
junrushao,
jwfromm,
mbrookhart,
merrymercy,
slyubomirsky,
tqchen,
vinx13,
wweic,
yzhliu and
zhiics
as code owners
|
I will take a look tomorrow |
AndrewZhaoLuo
left a comment
AndrewZhaoLuo
left a comment
There was a problem hiding this comment.
Thanks for the PR! A few comments
| mod = tvm.IRModule.from_expr(op) | ||
| fake_quantized_op_freqs = relay.analysis.list_fake_quantized_op_freqs(mod) | ||
|
|
||
| assert len(fake_quantized_op_freqs) == 1 |
There was a problem hiding this comment.
I think you can just do direct equality throughout this file
dict(fake_quantized_op_freqs) == {"nn.conv2d": 1}
|
|
||
| using ExprSet = std::unordered_set<Expr, ObjectPtrHash, ObjectPtrEqual>; | ||
|
|
||
| class FakeQuantizedRegionExtractor : public ExprVisitor { |
There was a problem hiding this comment.
Can you just reuse SubgraphExtractor in src/relay/transforms/fake_quantization_to_integer.cc?
anwang2009
left a comment
anwang2009
left a comment
There was a problem hiding this comment.
tests look great! very comprehensive.
| * \file src/relay/transforms/fake_quantization_to_integer.h | ||
| * \brief Extract subgraph of a fake quantized region. | ||
| * | ||
| * https://llvm.org/doxygen/CallGraph_8h_source.html |
There was a problem hiding this comment.
This line is probably copypasta?
| if (op != dequantize_op_) { | ||
| if (fake_quantized_op_freqs_.find(op_name) != fake_quantized_op_freqs_.end()) { | ||
| fake_quantized_op_freqs_.Set(op_name, | ||
| int64_t(fake_quantized_op_freqs_.at(op_name)) + 1); |
There was a problem hiding this comment.
Why do you need to cast here?
There was a problem hiding this comment.
i was getting compile-time errors that fake_quantized_op_freqs_.at(op_name) + 1 is a PrimExpr instead of a tvm::Integer and it seemed like casting worked around the issue -- lmk if there's a better way around this?
…10089) * add relay pass to collect fake quantized ops * add more tests * more tests * lint * lint * remove unused imports * update comment * lint * reuse SubgraphExtractor and update test assertions * remove print * lint * remove unneeded comment Co-authored-by: Margaret Qian <[email protected]>
3 participants
Add a relay pass to collect fake quantized ops and frequencies from within fake quantized regions.