crop_yield_prediction/train_cross_location.py [35:129]: - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - n_batches = len(train_dataloader) sum_loss_dic = {} for loss_type in ['loss', 'loss_supervised', 'loss_unsupervised', 'l_n', 'l_d', 'l_nd', 'sn_loss', 'tn_loss', 'norm_loss']: sum_loss_dic[loss_type] = 0 for batch_X, batch_y in train_dataloader: batch_X, batch_y = prep_data(batch_X, batch_y, cuda) # forward optimizer.zero_grad() emb_triplets, pred = model(batch_X, unsup_weight) loss_func = torch.nn.MSELoss() loss_supervised = loss_func(pred, batch_y) if unsup_weight != 0: loss_unsupervised, l_n, l_d, l_nd, sn_loss, tn_loss, norm_loss = triplet_loss(emb_triplets, tilenet_margin, tilenet_l2, tilenet_ltn) loss = (1 - unsup_weight) * loss_supervised + unsup_weight * loss_unsupervised else: loss = loss_supervised loss.backward() optimizer.step() # note keeping sum_loss_dic['loss'] += loss.item() sum_loss_dic['loss_supervised'] += loss_supervised.item() if unsup_weight != 0: sum_loss_dic['loss_unsupervised'] += loss_unsupervised.item() sum_loss_dic['l_n'] += l_n.item() sum_loss_dic['l_d'] += l_d.item() sum_loss_dic['l_nd'] += l_nd.item() sum_loss_dic['sn_loss'] += sn_loss.item() sum_loss_dic['tn_loss'] += tn_loss.item() if tilenet_l2 != 0: sum_loss_dic['norm_loss'] += norm_loss.item() avg_loss_dic = {} for loss_type in sum_loss_dic.keys(): avg_loss_dic[loss_type] = sum_loss_dic[loss_type] / n_batches return avg_loss_dic def cal_performance(prediction, y): rmse = np.around(sqrt(mean_squared_error(y, prediction)), 3) r2 = np.around(r2_score(y, prediction), 3) corr = tuple(map(lambda x: np.around(x, 3), pearsonr(y, prediction)))[0] return rmse, r2, corr def triplet_loss(emb_triplets, margin, l2, ltn): dim = emb_triplets.shape[-1] z_a = emb_triplets[:, :, 0, :] z_tn = emb_triplets[:, :, 1, :] z_sn = emb_triplets[:, :, 2, :] z_d = emb_triplets[:, :, 3, :] # average over timesteps l_n = torch.mean(torch.sqrt(((z_a - z_sn) ** 2).sum(dim=2)), dim=1) l_d = - torch.mean(torch.sqrt(((z_a - z_d) ** 2).sum(dim=2)), dim=1) sn_loss = F.relu(l_n + l_d + margin) tn_loss = torch.mean(torch.sqrt(((z_a - z_tn) ** 2).sum(dim=2)), dim=1) # average by #samples in mini-batch l_n = torch.mean(l_n) l_d = torch.mean(l_d) l_nd = torch.mean(l_n + l_d) sn_loss = torch.mean(sn_loss) tn_loss = torch.mean(tn_loss) loss = (1 - ltn) * sn_loss + ltn * tn_loss norm_loss = 0 if l2 != 0: z_a_norm = torch.sqrt((z_a ** 2).sum(dim=2)) z_sn_norm = torch.sqrt((z_sn ** 2).sum(dim=2)) z_d_norm = torch.sqrt((z_d ** 2).sum(dim=2)) z_tn_norm = torch.sqrt((z_tn ** 2).sum(dim=2)) norm_loss = torch.mean(z_a_norm + z_sn_norm + z_d_norm + z_tn_norm) / (dim ** 0.5) loss += l2 * norm_loss return loss, l_n, l_d, l_nd, sn_loss, tn_loss, norm_loss def eval_epoch(model, validation_dataloader, tilenet_margin, tilenet_l2, tilenet_ltn, unsup_weight, cuda): ''' Epoch operation in evaluation phase ''' model.eval() if cuda: - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - crop_yield_prediction/train_semi_transformer.py [39:133]: - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - n_batches = len(train_dataloader) sum_loss_dic = {} for loss_type in ['loss', 'loss_supervised', 'loss_unsupervised', 'l_n', 'l_d', 'l_nd', 'sn_loss', 'tn_loss', 'norm_loss']: sum_loss_dic[loss_type] = 0 for batch_X, batch_y in train_dataloader: batch_X, batch_y = prep_data(batch_X, batch_y, cuda) # forward optimizer.zero_grad() emb_triplets, pred = model(batch_X, unsup_weight) loss_func = torch.nn.MSELoss() loss_supervised = loss_func(pred, batch_y) if unsup_weight != 0: loss_unsupervised, l_n, l_d, l_nd, sn_loss, tn_loss, norm_loss = triplet_loss(emb_triplets, tilenet_margin, tilenet_l2, tilenet_ltn) loss = (1 - unsup_weight) * loss_supervised + unsup_weight * loss_unsupervised else: loss = loss_supervised loss.backward() optimizer.step() # note keeping sum_loss_dic['loss'] += loss.item() sum_loss_dic['loss_supervised'] += loss_supervised.item() if unsup_weight != 0: sum_loss_dic['loss_unsupervised'] += loss_unsupervised.item() sum_loss_dic['l_n'] += l_n.item() sum_loss_dic['l_d'] += l_d.item() sum_loss_dic['l_nd'] += l_nd.item() sum_loss_dic['sn_loss'] += sn_loss.item() sum_loss_dic['tn_loss'] += tn_loss.item() if tilenet_l2 != 0: sum_loss_dic['norm_loss'] += norm_loss.item() avg_loss_dic = {} for loss_type in sum_loss_dic.keys(): avg_loss_dic[loss_type] = sum_loss_dic[loss_type] / n_batches return avg_loss_dic def cal_performance(prediction, y): rmse = np.around(sqrt(mean_squared_error(y, prediction)), 3) r2 = np.around(r2_score(y, prediction), 3) corr = tuple(map(lambda x: np.around(x, 3), pearsonr(y, prediction)))[0] return rmse, r2, corr def triplet_loss(emb_triplets, margin, l2, ltn): dim = emb_triplets.shape[-1] z_a = emb_triplets[:, :, 0, :] z_tn = emb_triplets[:, :, 1, :] z_sn = emb_triplets[:, :, 2, :] z_d = emb_triplets[:, :, 3, :] # average over timesteps l_n = torch.mean(torch.sqrt(((z_a - z_sn) ** 2).sum(dim=2)), dim=1) l_d = - torch.mean(torch.sqrt(((z_a - z_d) ** 2).sum(dim=2)), dim=1) sn_loss = F.relu(l_n + l_d + margin) tn_loss = torch.mean(torch.sqrt(((z_a - z_tn) ** 2).sum(dim=2)), dim=1) # average by #samples in mini-batch l_n = torch.mean(l_n) l_d = torch.mean(l_d) l_nd = torch.mean(l_n + l_d) sn_loss = torch.mean(sn_loss) tn_loss = torch.mean(tn_loss) loss = (1 - ltn) * sn_loss + ltn * tn_loss norm_loss = 0 if l2 != 0: z_a_norm = torch.sqrt((z_a ** 2).sum(dim=2)) z_sn_norm = torch.sqrt((z_sn ** 2).sum(dim=2)) z_d_norm = torch.sqrt((z_d ** 2).sum(dim=2)) z_tn_norm = torch.sqrt((z_tn ** 2).sum(dim=2)) norm_loss = torch.mean(z_a_norm + z_sn_norm + z_d_norm + z_tn_norm) / (dim ** 0.5) loss += l2 * norm_loss return loss, l_n, l_d, l_nd, sn_loss, tn_loss, norm_loss def eval_epoch(model, validation_dataloader, tilenet_margin, tilenet_l2, tilenet_ltn, unsup_weight, cuda): ''' Epoch operation in evaluation phase ''' model.eval() if cuda: - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -