models.py

import copy
import torch
from torch import nn
import numpy as np

import utils #
import sys
import traceback


class SummarisationModel(nn.Module):

    def __init__(self, gpu):

        super().__init__()

        self.cnns = nn.ModuleList([nn.Conv1d(200, 50, i+1, padding="valid", groups=1) for i in range(7)])

        self.document_encoder = nn.LSTM(350, 600) 

        self.sentence_extractor = nn.LSTM(350, 600)

        self.projector = nn.Linear(600, 2)

        self.softmax = nn.Softmax(dim=-1)

        self.device = torch.device("cuda" if gpu else "cpu")

        self.to(self.device)

    def encode_sentences(self, document):

        convolutions = []
        for cnn in self.cnns:
            convolutions.append(cnn(document.transpose(1,2)).amax(dim=2))

        return torch.cat(convolutions, dim=1)

    def encode_document(self, encoded_sentences):

        _, (hidden_state, cell_state) = self.document_encoder(encoded_sentences.flip(dims=(0,)))

        return hidden_state, cell_state

    def encode(self, document):

        encoded_sentences = self.encode_sentences(document)

        return encoded_sentences, self.encode_document(encoded_sentences)

    def forward(self, document, k=3):

        encoded_sentences, states = self.encode(document)

        logits = self.projector(self.sentence_extractor(encoded_sentences, states)[0])
        
        probs = self.softmax(logits)[:,0]
        
        if k < len(probs):
            return probs.topk(k).indices, probs # handle doc weniger als 3 sents? 
        return torch.arange(len(probs)), probs

    def test(self, dataset):

        running_rouge_1 = 0.0
        running_rouge_2 = 0.0
        running_rouge_l = 0.0
        
        self.eval()
        with torch.no_grad():

            for datapoint in dataset:
                
                if len(datapoint.raw_document) == 0 or len(datapoint.raw_summary) == 0: 
                    print("Warning in Testing! This datapoint has an empty document or an empty summary")
                    continue
                
                if self.gpu: 
                    datapoint = datapoint.to(torch.device("cuda"))

                top_indices, probs = self.__call__(datapoint.document)
                r_1, r_2, r_l = utils.rouge(utils.select_elements(datapoint.raw_document, top_indices), datapoint.raw_summary, verbose=True)  
                running_rouge_1 += r_1
                running_rouge_2 += r_2
                running_rouge_l += r_l
            
        epoch_rouge_1 = running_rouge_1 / len(dataset)
        epoch_rouge_2 = running_rouge_2 / len(dataset)  
        epoch_rouge_l = running_rouge_l / len(dataset)

        return epoch_rouge_1, epoch_rouge_2, epoch_rouge_l
        
    def validation(self, dataset):

        return sum(self.test(dataset)) / 3.0




class ActorOnlySummarisationModel(SummarisationModel):
    
    def __init__(self, gpu):

        super().__init__(gpu)

        self.optimizer = torch.optim.Adam(self.parameters(), lr=0.001)

    def training_epoch(self, dataloader, learning_rate=None): # def scheduler? or global variable?

        if learning_rate != None:
            for g in self.optimizer.param_groups:
                g['lr'] = learning_rate 
        
        self.train()

        epoch_loss = 0.0
        epoch_rouge = 0.0
            
        for batch in dataloader:

            self.optimizer.zero_grad()

            for datapoint in batch:
                # check if dp to gpu is OK 
                datapoint = datapoint.to(self.device) # device definiert in main_ActorOnly.py

                try:    # Prevent breakdown for inapt datapoints
                    # documents with empty content!
                    if len(datapoint.raw_document) == 0 or len(datapoint.raw_summary) == 0: 
                        print("Warning! This datapoint has an empty document or an empty summary")
                        continue

                    _, probs = self.__call__(datapoint.document)

                    o = datapoint.p_searchspace @ torch.log(probs) + datapoint.n_searchspace @ torch.log(1 - probs)

                    idx_sample = torch.argmax(o)

                    loss = - datapoint.top_rouge[idx_sample] * o[idx_sample]

                    loss.backward()

                    epoch_loss += loss.item()
                    epoch_rouge += datapoint.top_rouge[idx_sample]
                
                except Exception as e:
                    traceback.print_exception(*sys.exc_info())
                    continue
            
            self.optimizer.step()
        
        return epoch_loss / len(dataloader.dataset), epoch_rouge / len(dataloader.dataset)
        

class SummarisationModelWithCrossEntropyLoss(SummarisationModel):
    
    def __init__(self, gpu):

        super().__init__(gpu)

        self.loss_fn = nn.BCELoss(reduction='sum')
        self.optimizer = torch.optim.Adam(self.parameters(), lr=0.001)

    def training_epoch(self, dataloader, learning_rate=None):

        if learning_rate != None:
            for g in self.optimizer.param_groups:
                g['lr'] = learning_rate 

        self.train()

        epoch_loss = 0.0
            
        for batch in dataloader:

            self.optimizer.zero_grad()

            for datapoint in batch:
                
                _, probs = self.__call__(datapoint.document)

                loss = self.loss_fn(probs, datapoint.bin_summary)

                loss.backward()

                epoch_loss += loss.item()
            
            self.optimizer.step()
        
        return epoch_loss / len(dataloader.dataset)

            

class ActorCriticSummarisationModel(SummarisationModel):

    def __init__(self, actor_wts, critic_wts, gpu):

        super().__init__(gpu)
        # move to main!, critic wird loss_fn
        self.critic = Critic(self) # ?
        critic.load_state_dict(critic_wts)

        # actor 
        self.load_state_dict(actor_wts)

        
        

    def forward()

    def train()

    def test()



# gpu, set_parameters
class Critic(nn.Module):

    def __init__(self, model, steepness=8, denoise=100):

        super().__init__()

        self.steepness = steepness
        self.denoise = denoise
        # eventuell move to main
        self.optimizer = torch.optim.Adam(self.parameters(), lr= 0.001)
        self.loss_fn = nn.MSELoss()

        model = copy.deepcopy(model)
        #model.eval()
        for param in model.parameters():
            param.requires_grad = False

        self.document_encoder = model.encode_document # encode, encode_document
 
        self.layer_1 = nn.Linear(1200, 600)
        self.layer_2 = nn.Linear(600, 600)
        self.layer_3 = nn.Linear(600, 1)

        W_1 = torch.cat((torch.eye(600), -torch.eye(600)), 1)
        W_2 = torch.eye(600)
        W_3 = torch.ones(600)

        self.layer_1.weight.data /= self.denoise
        self.layer_1.weight.data += W_1
        self.layer_2.weight.data /= self.denoise
        self.layer_2.weight.data += W_2
        self.layer_3.weight.data /= self.denoise
        self.layer_3.weight.data += W_3
        self.layer_3.weight.data /= 600

    def forward(self, encoded_sentences_1, encoded_sentences_2):

        _, document_vec_1 = self.document_encoder(encoded_sentences_1)
        _, document_vec_2 = self.document_encoder(encoded_sentences_2)

        double_document = torch.cat((torch.squeeze(document_vec_1), torch.squeeze(document_vec_2)), dim=-1)

        return torch.tanh(self.steepness*nn.functional.relu(self.layer_3(
            nn.functional.relu(self.layer_2(
            utils.gaussian(self.layer_1(double_document)))))))
    

    def training_epoch(self, dataloader, learning_rate=None):

        if learning_rate != None:
            for g in self.optimizer.param_groups:
                g['lr'] = learning_rate 

        self.train()
        pos_samples= 0.5 

        epoch_loss = 0.0
        for batch in train_dataloader:

            self.optimizer.zero_grad()
        
            for datapoint in batch:

                r = np.random.random()
                if r > pos_samples:

                    k = np.random.choice(len(datapoint.p_searchspace)) 
                    sample = datapoint.sent_vecs.masked_select(datapoint.p_searchspace[k].bool())  # not padded sent embeddngs

                    score = self.__call__(sample, datapoint.gold_sent_vecs) 
                    loss = self.loss_fn(score, datapoint.top_rouge[k])

                else:
                    if len(datapoint.sent_vecs) >= 3: 
                        narray = np.random.choice(len(datapoint.sent_vecs), 3, replace = False) 
                        narray.sort()
                        sample = datapoint.sent_vecs[narray]
                    else:
                        continue # handle len(sent_vecs) < 3 

                    score = self.__call__(sample, datapoint.gold_sent_vecs)
                    loss = self.loss_fn(score, utils.rouge(datapoint.raw_document[narray], datapoint.raw_summary))
                    # rouge score berechnen für negative sample => besser wäre externes berechnen und speichern?

                epoch_loss += loss.item()
                
                loss.backward()
            
            self.optimizer.step()
        
        return epoch_loss / len(dataloader.dataset)

    def test(self, dataset):
        self.eval()
        pos_samples= 0.5 

        running_diff = 0.0
        with torch.no_grad():

            for datapoint in dataset:
            
                r = np.random.random()
                if r > pos_samples:

                    k = np.random.choice(len(datapoint.p_searchspace)) 
                    sample = datapoint.sent_vecs.masked_select(datapoint.p_searchspace[k].bool())  # not padded sent embeddngs

                    score = self.__call__(sample, datapoint.gold_sent_vecs) 
                    score_diff = score - datapoint.top_rouge[k] # tensor
                    

                else:
                    if len(datapoint.sent_vecs) >= 3: 
                        narray = np.random.choice(len(datapoint.sent_vecs), 3, replace = False) 
                        narray.sort()
                        sample = datapoint.sent_vecs[narray]
                    else:
                        continue # handle len(sent_vecs) < 3 

                    score = self.__call__(sample, datapoint.gold_sent_vecs)
                    score_diff = score - utils.rouge(datapoint.raw_document[narray], datapoint.raw_summary)
                    # rouge score berechnen für negative sample => besser wäre externes berechnen und speichern?

                running_diff += abs(score_diff.item())
    
        return running_diff / len(dataset)