cbalint13 opened a new pull request #5805:
URL: https://github.com/apache/incubator-tvm/pull/5805
This PR adds ```nn.batch_flatten``` as quantizable layer.
**Description**
* ```nn/batch_flatten``` is commonly used before ```nn.dense``` in final
layers.
* Proposed PR allows it to be included in quantization process avoiding
re-cast to ```float32```.
**Outcome**
* Before
```
%19 = nn.max_pool2d(%18, pool_size=[2, 2], strides=[2, 2], padding=[0, 0,
0, 0]) /* ty=Tensor[(1, 50, 4, 4), int8] */;
%20 = cast(%19, dtype="int8") /* ty=Tensor[(1, 50, 4, 4), int8] */;
%21 = annotation.stop_fusion(%20) /* ty=Tensor[(1, 50, 4, 4), int8] */;
%22 = cast(%21, dtype="float32") /* ty=Tensor[(1, 50, 4, 4), float32] */;
%23 = multiply(%22, 0.0625f /* ty=float32 */) /* ty=Tensor[(1, 50, 4, 4),
float32] */;
%24 = nn.batch_flatten(%23) /* ty=Tensor[(1, 800), float32] */;
%25 = nn.batch_flatten(%24) /* ty=Tensor[(1, 800), float32] */;
%26 = nn.batch_flatten(%25) /* ty=Tensor[(1, 800), float32] */;
%27 = nn.dense(%26, meta[relay.Constant][2] /* ty=Tensor[(512, 800),
float32] */ /* ty=Tensor[(512, 800), float32] */, units=512) /* ty=Tensor[(1,
512), float32] */;
%28 = nn.relu(%27) /* ty=Tensor[(1, 512), float32] */;
%29 = nn.batch_flatten(%28) /* ty=Tensor[(1, 512), float32] */;
%30 = nn.batch_flatten(%29) /* ty=Tensor[(1, 512), float32] */;
nn.dense(%30, meta[relay.Constant][3] /* ty=Tensor[(10, 512), float32] */
/* ty=Tensor[(10, 512), float32] */, units=10) /* ty=Tensor[(1, 10), float32] */
```
* After
```
%19 = nn.max_pool2d(%18, pool_size=[2, 2], strides=[2, 2], padding=[0, 0,
0, 0]) /* ty=Tensor[(1, 50, 4, 4), int8] */;
%20 = cast(%19, dtype="int8") /* ty=Tensor[(1, 50, 4, 4), int8] */;
%21 = annotation.stop_fusion(%20) /* ty=Tensor[(1, 50, 4, 4), int8] */;
%22 = nn.batch_flatten(%21) /* ty=Tensor[(1, 800), int8] */;
%23 = nn.batch_flatten(%22) /* ty=Tensor[(1, 800), int8] */;
%24 = nn.batch_flatten(%23) /* ty=Tensor[(1, 800), int8] */;
%25 = clip(%24, a_min=-127f, a_max=127f) /* ty=Tensor[(1, 800), int8] */;
%26 = nn.dense(%25, meta[relay.Constant][2] /* ty=Tensor[(512, 800), int8]
*/ /* ty=Tensor[(512, 800), int8] */, units=512, out_dtype="int32") /*
ty=Tensor[(1, 512), int32] */;
%27 = nn.relu(%26) /* ty=Tensor[(1, 512), int32] */;
%28 = nn.batch_flatten(%27) /* ty=Tensor[(1, 512), int32] */;
%29 = nn.batch_flatten(%28) /* ty=Tensor[(1, 512), int32] */;
%30 = add(%29, 512 /* ty=int32 */) /* ty=Tensor[(1, 512), int32] */;
%31 = right_shift(%30, 10 /* ty=int32 */) /* ty=Tensor[(1, 512), int32] */;
%32 = clip(%31, a_min=-127f, a_max=127f) /* ty=Tensor[(1, 512), int32] */;
%33 = cast(%32, dtype="int8") /* ty=Tensor[(1, 512), int8] */;
%34 = nn.dense(%33, meta[relay.Constant][3] /* ty=Tensor[(10, 512), int8]
*/ /* ty=Tensor[(10, 512), int8] */, units=10, out_dtype="int32") /*
ty=Tensor[(1, 10), int32] */;
```
@vinx13, @siju-samuel @masahi @FrozenGene @ZihengJiang please help with the
review.
Thank You !
----------------------------------------------------------------
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.
For queries about this service, please contact Infrastructure at:
us...@infra.apache.org
