Operator requests

[Hakuyume]

operator	requiring models
Sigmoid	YOLOv2, (NStep)LSTM
Transpose	YOLOv2, Super resolution (#56 (comment)), GGNN (#56 (comment))
Reshape	YOLOv2
Mul	SSD, (NStep)LSTM
Conv (dilation > 1)	SSD
Upsample	YOLOv3*, FPN*
Gather	GGNN (#56 (comment))
Identity	GGNN (#56 (comment))
Conv (group > 1)	MobileNet

[msakai]

+1 for Transpose to run super resolution example of MXNet.

[msakai]

I want following operators to run chainer-chemistry implementation of GGNN.

• Gather operator and int32 tensors (c.f. add int32 and int64 dtypes and replace int with int32_t #41)
  • The use of them can be avoided if we input one-hot vectors instead of int32 tensors, and I take this approach for ONNX file I attached to Non-deterministic and suspicious errors #89. But int32 tensors is more concise and desirable.
• Transpose operator
• Identity operator
  • Chainer's copy function is exported as Identity operator (c.f. support Copy chainer/onnx-chainer#27 ). But the use of Identity could eliminated if we transform ONNX file.

[okdshin]

@Hakuyume Sigmoid, Transpose, Mul, Conv(dilation > 1) are now available on mkldnn_with_generic_fallback backend. Please check.