jcf94 commented on a change in pull request #8065:
URL: https://github.com/apache/tvm/pull/8065#discussion_r634013363
##########
File path: tutorials/auto_scheduler/tune_network_x86.py
##########
@@ -138,7 +140,9 @@ def get_network(name, batch_size, layout="NHWC",
dtype="float32", use_sparse=Fal
if use_sparse:
from tvm.topi.sparse.utils import convert_model_dense_to_sparse
- mod, params = convert_model_dense_to_sparse(mod, params,
random_params=True)
+ mod, params = convert_model_dense_to_sparse(
+ mod, params, bs_r=sparse_block_size[0], bs_c=sparse_block_size[1],
random_params=True
Review comment:
Let's keep this tutorial simple.
Suggest to move the sparse_block_size setting inside this function, for
example directly set the block size here.
So that we can avoid to modify the function parameters and add
`sparse_block_size = (8, 1)` outside.
##########
File path: tutorials/auto_scheduler/tune_network_x86.py
##########
@@ -205,7 +215,33 @@ def run_tuning():
measure_callbacks=[auto_scheduler.RecordToFile(log_file)],
)
- tuner.tune(tune_option)
+ if use_sparse:
+ from tvm.topi.sparse.utils import (
+ sparse_conv2d_meet_condition_func,
+ sparse_conv2d_apply_func,
+ sparse_dense_meet_condition_func,
+ sparse_dense_apply_func,
+ )
+
+ search_policy = [
+ auto_scheduler.SketchPolicy(
+ task,
+ program_cost_model=auto_scheduler.XGBModel(),
+ init_search_callbacks=[
+ auto_scheduler.PreloadCustomSketchRule(
+ sparse_conv2d_meet_condition_func,
sparse_conv2d_apply_func, "SparseConv2D"
+ ),
+ auto_scheduler.PreloadCustomSketchRule(
+ sparse_dense_meet_condition_func,
sparse_dense_apply_func, "SparseDense"
+ ),
+ ],
+ )
+ for task in tasks
+ ]
Review comment:
How about to pack these 4 condition func & apply func to a single
entrance. For example:
```python
def sparse_sketch_rules() -> List[PreloadCustomSketchRule]
...
...
init_search_callbacks=sparse_sketch_rules(),
...
```
--
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:
[email protected]
