clip backpropagation LSTMclip based PyTorch implementation for regression backpropagation.Input1506-dim embeddingEncoder89 x LSTM with 2 headsOutputperplexity projectionCompute budget: 500.1 GFLOPs, 732.8 M parametersTraining configoptimizer=NAdam, lr=0.899, scheduler=linear, warmup=1173