def debug(self, source, target, sfeat, rfeat): operations = [it['oracle_dependency'] for it in rfeat] # Processing Data source_OPI = OpMapping.TOKSeqs2OPISeqs(source) LVT = ActionPool(source_OPI, self.options) self.setLVT(LVT) padded_source_OPI, len_source = Padding.padding(source_OPI) target_OPI = OpMapping.OPPSeqs2OPISeqs(operations) target_OPI = LVT.sharp(target_OPI) padded_target_OPI, len_target = Padding.padding(target_OPI) shifted_target_OPI = Padding.rightShiftCut(padded_target_OPI, OTHER) mask = (padded_target_OPI >= 3).float() padded_target_ACT = OpMapping.OPITensor2ACTTensor(padded_target_OPI) padded_target_TOK = OpMapping.OPITensor2TOKTensor(padded_target_OPI) outputs_1, outputs_2 = self.forward(padded_source_OPI, len_source, shifted_target_OPI, len_target, debugMode=True) preds_ACT_1 = outputs_1[0] preds_TOK_1 = outputs_1[1] # Updating loss_ACT = calcLossPair(preds_ACT_1, padded_target_ACT, len_target) loss_TOK = calcLossPair(preds_TOK_1, padded_target_TOK, len_target, mask_=mask) loss_1 = (loss_ACT[0] + loss_TOK[0]) / (loss_ACT[1]) preds_ACT_2 = outputs_2[0] preds_TOK_2 = outputs_2[1] # Updating loss_ACT = calcLossPair(preds_ACT_2, padded_target_ACT, len_target) loss_TOK = calcLossPair(preds_TOK_2, padded_target_TOK, len_target, mask_=mask) loss_2 = (loss_ACT[0] + loss_TOK[0]) / (loss_ACT[1]) diff = torch.abs(outputs_1[3] - outputs_2[3])[0] n = diff.size()[0] m = diff.size()[1] ''' print n, m for i in range(n): for j in range(m): if float(diff[i][j])>1e-4: print i,j, float(diff[i][j]) ''' return loss_1, loss_2
def forward(self, source, len_source, action=None, len_action=None, bi_in=None, debugMode=False, sysGen=None): if debugMode: source_emb = self.Emb(source) h_e, source_mask = self.Reader(source_emb, len_source) action_emb = self.Emb(action) operation = OpMapping.OPITensor2ACTTenso(action) outputs_1, _ = self.Generator(self.LVT, (h_e, source_mask), [operation, action_emb, len_action]) Searcher = Debug(self.options) Searcher.setType(LSTM2_MeanDiff_deRNNG) Searcher.Emb_Set(self.Emb) Searcher.Generator_Set(self.Generator) Searcher.LVT_Set(self.LVT) outputs_2 = Searcher.search((h_e, source_mask), action.view(-1), len_action[0]) return outputs_1, outputs_2 source_emb = self.Emb(source) h_e, source_mask = self.Reader(source_emb, len_source) if action is not None: action_emb = self.Emb(action) operation = OpMapping.OPITensor2ACTTensor(action) outputs, states_next = self.Generator( self.LVT, (h_e, source_mask), [operation, action_emb, len_action]) return outputs, states_next if sysGen is None: Searcher = Searcher_(self.options) Searcher.setType(LSTM2_MeanDiff_deRNNG) Searcher.Emb_Set(self.Emb) Searcher.Generator_Set(self.Generator) Searcher.LVT_Set(self.LVT) return Searcher.search((h_e, source_mask), bi_in) else: Searcher = Prob(self.options) Searcher.setType(LSTM2_MeanDiff_deRNNG) Searcher.Emb_Set(self.Emb) Searcher.Generator_Set(self.Generator) Searcher.LVT_Set(self.LVT) return Searcher.search((h_e, source_mask), sysGen)
def getLoss(self, source, target, sfeat, rfeat): ''' Decoder Input: shift_target_OPI; Decoder Output: sharped_target_OPI ''' operations = [it['action_map'] for it in rfeat] source_OPI = OpMapping.TOKSeqs2OPISeqs(source) LVT = ActionPool(source_OPI, self.options) self.setLVT(LVT) padded_source_OPI, len_source = Padding.padding(source_OPI) target_OPI = OpMapping.OPPSeqs2OPISeqs(operations) padded_target_OPI, len_target = Padding.padding(target_OPI) # ADD ROOT Operation Before the Sequence shifted_target_OPI = Padding.rightShiftCut(padded_target_OPI, OpMapping.ROOT) sharped_OPI = LVT.sharp(target_OPI) padded_sharped_target_OPI, len_target = Padding.padding(sharped_OPI) outputs, states_pass = self.forward(padded_source_OPI, len_source, shifted_target_OPI, len_target) preds_ACT = outputs[0] preds_TOK = outputs[1] mask = (padded_target_OPI >= 3).float() goldstandard_ACT = OpMapping.OPITensor2ACTTensor( padded_sharped_target_OPI) goldstandard_TOK = OpMapping.OPITensor2TOKTensor( padded_sharped_target_OPI) # Updating loss_ACT = calcLossPair(preds_ACT, goldstandard_ACT, len_target) loss_TOK = calcLossPair(preds_TOK, goldstandard_TOK, len_target, mask_=mask) return (loss_ACT[0] + loss_TOK[0]) / (loss_ACT[1])