Added new type of model bs_roformer_low_mem (fully compatible with …

…`bs_roformer`). It allows to train BSRoformer model with batch size x20 times larger. Changes proposed by `unwa`.

README.md

@@ -12,7 +12,7 @@ Available models for training:
 * Demucs4HT [[Paper](https://arxiv.org/abs/2211.08553)]. Key: `htdemucs`.
 * VitLarge23 based on [Segmentation Models Pytorch](https://github.com/qubvel/segmentation_models.pytorch). Key: `segm_models`.
 * TorchSeg based on [TorchSeg module](https://github.com/qubvel/segmentation_models.pytorch). Key: `torchseg`.
-* Band Split RoFormer [[Paper](https://arxiv.org/abs/2309.02612), [Repository](https://github.com/lucidrains/BS-RoFormer)] . Key: `bs_roformer`.
+* Band Split RoFormer [[Paper](https://arxiv.org/abs/2309.02612), [Repository](https://github.com/lucidrains/BS-RoFormer)] . Key: `bs_roformer` or `bs_roformer_low_mem`.
 * Mel-Band RoFormer [[Paper](https://arxiv.org/abs/2310.01809), [Repository](https://github.com/lucidrains/BS-RoFormer)]. Key: `mel_band_roformer` or `mel_band_roformer_low_mem`.
 * Swin Upernet [[Paper](https://arxiv.org/abs/2103.14030)] Key: `swin_upernet`.
 * BandIt Plus [[Paper](https://arxiv.org/abs/2309.02539), [Repository](https://github.com/karnwatcharasupat/bandit)] Key: `bandit`.