[WIP] Add option to handle dynamic shape on reshape node

[disktnk]

Shape size can be changed but ONNX-Chainer cannot deal with it because shape is not represented as computational graph and fixed when passed to functions, like following case:

batch_size = x.shape[0]
x = x.reshape((batch_size, -1))  # batch_size is set to Reshape node as number

To handle this, introduce enable_shape_var option, then chainer.Variable.shape returns as chainer.Variable and make computation graph even shape.

This logic has strong side effect and cannot pass ordinal code such as:

# enable_shape_var = True
shape = x.shape  # returns as chainer.Variable
y = np.empty(shape)  # error because numpy.empty cannot handle Variable type

So pay attention to use this option.

Not set the option:

Set the option

TODO:

I'll separate this PR

• [x] add reshape to deal with dynamic shape -> #194
• [x] add shape converter -> #193
• [ ] add enable_shape_var option

[codecov-io]

Codecov Report

Merging #191 into master will increase coverage by 0.12%. The diff coverage is 88.11%.

@@            Coverage Diff             @@
##           master     #191      +/-   ##
==========================================
+ Coverage   89.57%   89.69%   +0.12%     
==========================================
  Files          24       25       +1     
  Lines        1372     1466      +94     
==========================================
+ Hits         1229     1315      +86     
- Misses        143      151       +8

Impacted Files	Coverage Δ
onnx_chainer/mapping.py	90% <ø> (ø)	:arrow_up:
onnx_chainer/export.py	91.28% <100%> (+1.11%)	:arrow_up:
onnx_chainer/functions/__init__.py	100% <100%> (ø)	:arrow_up:
onnx_chainer/replace_func.py	95.18% <100%> (+3.61%)	:arrow_up:
onnx_chainer/variable.py	85.33% <85.33%> (ø)
onnx_chainer/functions/array.py	90.16% <87.5%> (-0.18%)	:arrow_down:

---

Continue to review full report at Codecov.

Legend - Click here to learn more Δ = absolute <relative> (impact), ø = not affected, ? = missing data Powered by Codecov. Last update ac0963d...cb4fbb0. Read the comment docs.