Modification about random combine

[yaozengwei]

This PR modifies the random combining of different encoder layers.

• It combines from 1/3 layers, i.e., layer-4 (counts from 0).
• It removes linear layers in RandomCombine class.

Experimental results can be found on #431.

[pkufool]

Do we need mod_inputs any more?

[danpovey]

We have to watch out, here, because this would invalidate already-trained models. If we have uploaded a trained model from this recipe, I suggest we could re-upload a new one.

[yaozengwei]

We have to watch out, here, because this would invalidate already-trained models. If we have uploaded a trained model from this recipe, I suggest we could re-upload a new one.

We have to watch out, here, because this would invalidate already-trained models. If we have uploaded a trained model from this recipe, I suggest we could re-upload a new one.

OK, I will re-upload a new one.

[csukuangfj]

OK, I will re-upload a new one.

Have you finished uploading the pre-trained models? Please also update RESULTS.md

[csukuangfj]

please also update the test code

icefall/egs/librispeech/ASR/pruned_transducer_stateless5/conformer.py

Line 1270 in 8761452

num_channels=num_channels,

[csukuangfj]

Please also remove the following doc:

icefall/egs/librispeech/ASR/pruned_transducer_stateless5/conformer.py

Lines 1067 to 1069 in 8761452

	All but the last input will have a linear transform before we
	randomly combine them; these linear transforms will be initialized
	to the identity transform.

[yaozengwei]

OK, I will re-upload a new one.

Have you finished uploading the pre-trained models? Please also update RESULTS.md

Thanks for the reminder. See #467.