A method to copy a tensor with stride and dimension. It is useful
for Crop, Concat, etc.

Add soft-label support for cross-entropy operator.

Fix compile error in prelu_op.

Add dropout operator.

Prelu with forward, backward and python test passed

Fix a ssd bug

Add a space

Refactorization overview

Fix format

Feature/transform ctx

Remove lazy-initialization in device_context

Using DeviceContext, not Place to get stream

Update

Add MKLDNN pool

Check and only check the output varibles specified by self.outputs

* Also use `const DeviceContext&` all the time, to prevent `const_cast`

Fix #4169
Fix #3468
Fix #3475

Change type of dropout_prob to template typename.

refine avg-pooling, which is exclusive. refine related code.