diff --git a/Networks.py b/Networks.py index 3f4518f16cfd6b0eee17c3d7baadd873afe0b411..4543af8e635b02d649baf743fcf5c1815e9e5d3e 100644 --- a/Networks.py +++ b/Networks.py @@ -54,3 +54,69 @@ class BaseNet(nn.Module): ################################################################################ +################################################################################ +class LSTMNet(nn.Module): + def __init__(self, dicts, outputSize, incremental) : + super().__init__() + self.dummyParam = nn.Parameter(torch.empty(0), requires_grad=False) + + self.incremental = incremental + self.featureFunctionLSTM = "b.-2 b.-1 b.0 b.1 b.2" + self.featureFunction = "s.0 s.1 s.2 s.0.0 s.0.-1 s.0.1 s.1.0 s.1.-1 s.1.1 s.2.0 s.2.-1 s.2.1" + self.historyNb = 5 + self.columns = ["UPOS", "FORM"] + + self.embSize = 64 + self.nbInputLSTM = len(self.featureFunctionLSTM.split()) + self.nbInputBase = len(self.featureFunction.split()) + self.nbTargets = self.nbInputBase + self.nbInputLSTM + self.inputSize = len(self.columns)*self.nbTargets+self.historyNb + self.outputSize = outputSize + for name in dicts.dicts : + self.add_module("emb_"+name, nn.Embedding(len(dicts.dicts[name]), self.embSize)) + self.lstmFeat = nn.LSTM(len(self.columns)*self.embSize, len(self.columns)*int(self.embSize/2), 1, batch_first=True, bidirectional = True) + self.lstmHist = nn.LSTM(self.embSize, int(self.embSize/2), 1, batch_first=True, bidirectional = True) + self.fc1 = nn.Linear(self.inputSize * self.embSize, 1600) + self.fc2 = nn.Linear(1600, outputSize) + self.dropout = nn.Dropout(0.3) + + self.apply(self.initWeights) + + def forward(self, x) : + embeddings = [] + embeddingsLSTM = [] + for i in range(len(self.columns)) : + embeddings.append(getattr(self, "emb_"+self.columns[i])(x[...,i*self.nbInputBase:(i+1)*self.nbInputBase])) + for i in range(len(self.columns)) : + embeddingsLSTM.append(getattr(self, "emb_"+self.columns[i])(x[...,len(self.columns)*self.nbInputBase+i*self.nbInputLSTM:len(self.columns)*self.nbInputBase+(i+1)*self.nbInputLSTM])) + + z = torch.cat(embeddingsLSTM,-1) + z = self.lstmFeat(z)[0] + z = z.reshape(x.size(0), -1) + y = torch.cat(embeddings,-1).reshape(x.size(0),-1) + y = torch.cat([y,z], -1) + if self.historyNb > 0 : + historyEmb = getattr(self, "emb_HISTORY")(x[...,len(self.columns)*self.nbTargets:len(self.columns)*self.nbTargets+self.historyNb]) + historyEmb = self.lstmHist(historyEmb)[0] + historyEmb = historyEmb.reshape(x.size(0), -1) + y = torch.cat([y, historyEmb],-1) + y = self.dropout(y) + y = F.relu(self.dropout(self.fc1(y))) + y = self.fc2(y) + return y + + def currentDevice(self) : + return self.dummyParam.device + + def initWeights(self,m) : + if type(m) == nn.Linear: + torch.nn.init.xavier_uniform_(m.weight) + m.bias.data.fill_(0.01) + + def extractFeatures(self, dicts, config) : + colsValuesBase = Features.extractColsFeatures(dicts, config, self.featureFunction, self.columns, self.incremental) + colsValuesLSTM = Features.extractColsFeatures(dicts, config, self.featureFunctionLSTM, self.columns, self.incremental) + historyValues = Features.extractHistoryFeatures(dicts, config, self.historyNb) + return torch.cat([colsValuesBase, colsValuesLSTM, historyValues]) +################################################################################ +