Add TransformerLayer, TransformerBlock, C3TR modules (#2333)

* yolotr

* transformer block

* Remove bias in Transformer

* Remove C3T

* Remove a deprecated class

* put the 2nd LayerNorm into the 2nd residual block

* move example model to models/hub, rename to -transformer

* Add module comments and TODOs

* Remove LN in Transformer

* Add comments for Transformer

* Solve the problem of MA with DDP

* cleanup

* cleanup find_unused_parameters

* PEP8 reformat

Co-authored-by: DingYiwei <846414640@qq.com>
Co-authored-by: Glenn Jocher <glenn.jocher@ultralytics.com>

train.py

@@ -218,7 +218,9 @@ def train(hyp, opt, device, tb_writer=None):
 
     # DDP mode
     if cuda and rank != -1:
-        model = DDP(model, device_ids=[opt.local_rank], output_device=opt.local_rank)
+        model = DDP(model, device_ids=[opt.local_rank], output_device=opt.local_rank,
+                    # nn.MultiheadAttention incompatibility with DDP https://github.com/pytorch/pytorch/issues/26698
+                    find_unused_parameters=any(isinstance(layer, nn.MultiheadAttention) for layer in model.modules()))
 
     # Model parameters
     hyp['box'] *= 3. / nl  # scale to layers