| 64 | |
| 65 | template<class MyDevice> |
| 66 | void Hinge::backward_dev_impl(const MyDevice & dev, |
| 67 | const vector<const Tensor*>& xs, |
| 68 | const Tensor& fx, |
| 69 | const Tensor& dEdf, |
| 70 | unsigned i, |
| 71 | Tensor& dEdxi) const { |
| 72 | DYNET_ASSERT(i == 0, "Failed dimension check in Hinge::backward"); |
| 73 | // TODO: Can we do this on device? |
| 74 | if(pelement != nullptr) { |
| 75 | if(as_scalar(fx)) { // there was some loss |
| 76 | const float d = as_scalar(dEdf); |
| 77 | Tensor eloss(xs[0]->d, static_cast<float*>(aux_mem), fx.device, DeviceMempool::FXS); |
| 78 | // TODO: The > comparison should not be calculated twice. Keep it in auxiliary memory? |
| 79 | tvec(dEdxi).device(*dev.edevice) += (tvec(eloss) > 0.f).cast<float>() * d; |
| 80 | #if defined(__CUDACC__) && defined(EIGEN_NO_MALLOC) |
| 81 | DYNET_RUNTIME_ERR("CUDA memory allocation in hinge"); |
| 82 | #endif |
| 83 | tvec(dEdxi).chip<0>(*pelement).device(*dev.edevice) -= (tvec(eloss) > 0.f).template cast<float>().sum() * d; |
| 84 | } |
| 85 | } else { |
| 86 | DYNET_ASSERT(pelements != nullptr, "Hinge::backward has neither pointer to single element nor vector"); |
| 87 | vector<float> fx_vec = as_vector(fx); |
| 88 | vector<float> d_vec = as_vector(dEdf); |
| 89 | Tensor eloss(xs[0]->d, static_cast<float*>(aux_mem), fx.device, DeviceMempool::FXS); |
| 90 | for(size_t b = 0; b < fx.d.bd; b++) { |
| 91 | if(fx_vec[b]) { // there was some loss |
| 92 | tb<1>(dEdxi).chip<1>(b).device(*dev.edevice) += (tb<1>(eloss).chip<1>(b) > 0.f).cast<float>() * d_vec[b]; |
| 93 | #if defined(__CUDACC__) && defined(EIGEN_NO_MALLOC) |
| 94 | DYNET_RUNTIME_ERR("CUDA memory allocation in hinge"); |
| 95 | #endif |
| 96 | tb<1>(dEdxi).chip<1>(b).chip<0>((*pelements)[b]).device(*dev.edevice) -= (tb<1>(eloss).chip<1>(b) > 0.f).template cast<float>().sum() * d_vec[b]; |
| 97 | } |
| 98 | } |
| 99 | } |
| 100 | } |
| 101 | DYNET_NODE_INST_DEV_IMPL(Hinge) |
| 102 | |
| 103 | // ************* HingeDim ************* |