According to [SOLVED] KeyError: 'tile_ic' when tuning MobileNet, I found this should be considered as a bug.
The reason of this KeyError is because depthwise conv2d has no tuning space by default. This is not a problem for AutoTVM since it just tunes nothing. However, when we try to apply the best history, it falls here:
1. cfg = dispatch_ctx.query(target, workload)
2. if cfg.is_fallback:
3. if _is_int8_hw_support(data_dtype, kernel_dtype):
4. _get_default_config_int8(cfg, data_tensor, kernel_tensor, strides, padding, out_dtype,
5. is_depthwise, data_layout)
6. else:
7. _get_default_config(cfg, data_tensor, kernel_tensor, strides, padding, out_dtype,
8. is_depthwise, data_layout)
9. # Get the tiling parameters to set the layout names.
10. ic_bn, oc_bn = cfg["tile_ic"].size[-1], cfg["tile_oc"].size[-1]
If cfg comes from the tuning history, then it has no config in the case of depthwise conv2d. Even worst, it is not a fallback config, so line 4 or line 7 won’t be triggered and the key error happens at line 10.
We may have a quick fix that checks if all required entries are available; otherwise we enforce the fallback config.
cc @kevinthesun