Issues related to the input dimensions of the MLP model

[wjj19950828]

@merrymercy hi～

In your open source code, the input dimension of the MLP model is 164, which is aligned with Ansor, but in Appendix C of the Tenset paper, the input dimension is set to 324. Did you do anything?

Looking forward your reply~

[merrymercy]

In Appendix C of the paper, we mention that

Where the first 164 elements are from the orignal Ansor paper, the additional 324 - 164 = 160 elements are from the workload embedding.
In our current open source code, we don't use LDP anymore. Instead, we use a simpler approach to get the workload embedding. The related code is

tenset/python/tvm/auto_scheduler/cost_model/xgb_model.py

Lines 79 to 87 in 62f0c20

	def get_workload_embedding(workload_key):
	tags = ['max', 'min', 'add', 'Conv2dOutput', 'conv2d_winograd', 'DepthwiseConv2d',
	'dense', 'softmax', 'compute(b, i, j)']
	dag_str = str(ComputeDAG(workload_key_to_tensors(workload_key)))
	vec = [0] * len(tags)
	for i, tag in enumerate(tags):
	if tag in dag_str:
	vec[i] = 1
	return vec

tenset/python/tvm/auto_scheduler/cost_model/mlp_model.py

Line 333 in 62f0c20

"in_dim": 164 + (10 if use_workload_embedding else 0),

[wjj19950828]

In the paper, the effect of MLP+ranking loss is better than XGB+MSE, but in my experiment, the effect of MLP is not as good as XGB. Do you have any good suggestions for MLP?

[merrymercy]

What's your experiment setting? The results also depend on the dataset and hyperparameters.