genienlp/models/multitask_question_answering_network.py

import os
import math
import numpy as np
import json

import torch
from torch import nn
from torch.nn import functional as F

from util import get_trainable_params, set_seed
from modules import expectedBLEU, expectedMultiBleu, matrixBLEU

from cove import MTLSTM
from allennlp.modules.elmo import Elmo, batch_to_ids

from .common import *


class MultitaskQuestionAnsweringNetwork(nn.Module):

    def __init__(self, field, args):
        super().__init__()
        self.field = field
        self.args = args
        self.pad_idx = self.field.vocab.stoi[self.field.pad_token]
        self.device = set_seed(args)

        def dp(args):
            return args.dropout_ratio if args.rnn_layers > 1 else 0.

        if self.args.glove_and_char:
        
            self.encoder_embeddings = Embedding(field, args.dimension, 
                dropout=args.dropout_ratio, project=not args.cove)
    
            if self.args.cove or self.args.intermediate_cove:
                self.cove = MTLSTM(model_cache=args.embeddings, layer0=args.intermediate_cove, layer1=args.cove)
                cove_params = get_trainable_params(self.cove) 
                for p in cove_params:
                    p.requires_grad = False
                cove_dim = int(args.intermediate_cove) * 600 + int(args.cove) * 600 + 400 # the last 400 is for GloVe and char n-gram embeddings
                self.project_cove = Feedforward(cove_dim, args.dimension)

        if -1 not in self.args.elmo:
            options_file = "https://s3-us-west-2.amazonaws.com/allennlp/models/elmo/2x4096_512_2048cnn_2xhighway/elmo_2x4096_512_2048cnn_2xhighway_options.json"
            weight_file = "https://s3-us-west-2.amazonaws.com/allennlp/models/elmo/2x4096_512_2048cnn_2xhighway/elmo_2x4096_512_2048cnn_2xhighway_weights.hdf5"
            self.elmo = Elmo(options_file, weight_file, 3, dropout=0.0, do_layer_norm=False)
            elmo_params = get_trainable_params(self.elmo)
            for p in elmo_params:
                p.requires_grad = False
            elmo_dim = 1024 * len(self.args.elmo)
            self.project_elmo = Feedforward(elmo_dim, args.dimension)
            if self.args.glove_and_char:
                self.project_embeddings = Feedforward(2 * args.dimension, args.dimension, dropout=0.0)
        
        self.decoder_embeddings = Embedding(field, args.dimension, 
            dropout=args.dropout_ratio, project=True)
    
        self.bilstm_before_coattention = PackedLSTM(args.dimension,  args.dimension,
            batch_first=True, bidirectional=True, num_layers=1, dropout=0)
        self.coattention = CoattentiveLayer(args.dimension, dropout=0.3)
        dim = 2*args.dimension + args.dimension + args.dimension

        self.context_bilstm_after_coattention = PackedLSTM(dim, args.dimension,
            batch_first=True, dropout=dp(args), bidirectional=True, 
            num_layers=args.rnn_layers)
        self.self_attentive_encoder_context = TransformerEncoder(args.dimension, args.transformer_heads, args.transformer_hidden, args.transformer_layers, args.dropout_ratio)
        self.bilstm_context = PackedLSTM(args.dimension, args.dimension,
            batch_first=True, dropout=dp(args), bidirectional=True, 
            num_layers=args.rnn_layers)

        self.question_bilstm_after_coattention = PackedLSTM(dim, args.dimension,
            batch_first=True, dropout=dp(args), bidirectional=True, 
            num_layers=args.rnn_layers)
        self.self_attentive_encoder_question = TransformerEncoder(args.dimension, args.transformer_heads, args.transformer_hidden, args.transformer_layers, args.dropout_ratio)
        self.bilstm_question = PackedLSTM(args.dimension, args.dimension,
            batch_first=True, dropout=dp(args), bidirectional=True, 
            num_layers=args.rnn_layers)

        self.self_attentive_decoder = TransformerDecoder(args.dimension, args.transformer_heads, args.transformer_hidden, args.transformer_layers, args.dropout_ratio)
        self.dual_ptr_rnn_decoder = DualPtrRNNDecoder(args.dimension, args.dimension,
            dropout=args.dropout_ratio, num_layers=args.rnn_layers)

        self.generative_vocab_size = min(len(field.vocab), args.max_generative_vocab)
        self.out = nn.Linear(args.dimension, self.generative_vocab_size)

        self.dropout = nn.Dropout(0.4)

    def set_embeddings(self, embeddings):
        self.encoder_embeddings.set_embeddings(embeddings)
        self.decoder_embeddings.set_embeddings(embeddings)

    def forward(self, batch, iteration):
        context, context_lengths, context_limited, context_elmo    = batch.context,  batch.context_lengths,  batch.context_limited, batch.context_elmo
        question, question_lengths, question_limited, question_elmo = batch.question, batch.question_lengths, batch.question_limited, batch.question_elmo
        answer, answer_lengths, answer_limited       = batch.answer,   batch.answer_lengths,   batch.answer_limited
        oov_to_limited_idx, limited_idx_to_full_idx  = batch.oov_to_limited_idx, batch.limited_idx_to_full_idx

        def map_to_full(x):
            return limited_idx_to_full_idx[x]
        self.map_to_full = map_to_full

        if -1 not in self.args.elmo:
            def elmo(z, layers, device):
                e = self.elmo(batch_to_ids(z).to(device))['elmo_representations']
                return torch.cat([e[x] for x in layers], -1)
            context_elmo =  self.project_elmo(elmo(context_elmo, self.args.elmo, context.device).detach())
            question_elmo = self.project_elmo(elmo(question_elmo, self.args.elmo, question.device).detach())

        if self.args.glove_and_char:
            context_embedded = self.encoder_embeddings(context)
            question_embedded = self.encoder_embeddings(question)
            if self.args.cove:
                context_embedded = self.project_cove(torch.cat([self.cove(context_embedded[:, :, -300:], context_lengths), context_embedded], -1).detach())
                question_embedded = self.project_cove(torch.cat([self.cove(question_embedded[:, :, -300:], question_lengths), question_embedded], -1).detach())
            if -1 not in self.args.elmo:
                context_embedded = self.project_embeddings(torch.cat([context_embedded, context_elmo], -1))
                question_embedded = self.project_embeddings(torch.cat([question_embedded, question_elmo], -1))
        else:
            context_embedded, question_embedded = context_elmo, question_elmo 

        context_encoded = self.bilstm_before_coattention(context_embedded, context_lengths)[0]
        question_encoded = self.bilstm_before_coattention(question_embedded, question_lengths)[0]

        context_padding = context.data == self.pad_idx
        question_padding = question.data == self.pad_idx

        coattended_context, coattended_question = self.coattention(context_encoded, question_encoded, context_padding, question_padding)

        context_summary = torch.cat([coattended_context, context_encoded, context_embedded], -1)
        condensed_context, _ = self.context_bilstm_after_coattention(context_summary, context_lengths)
        self_attended_context = self.self_attentive_encoder_context(condensed_context, padding=context_padding)
        final_context, (context_rnn_h, context_rnn_c) = self.bilstm_context(self_attended_context[-1], context_lengths)
        context_rnn_state = [self.reshape_rnn_state(x) for x in (context_rnn_h, context_rnn_c)]

        question_summary = torch.cat([coattended_question, question_encoded, question_embedded], -1)
        condensed_question, _ = self.question_bilstm_after_coattention(question_summary, question_lengths)
        self_attended_question = self.self_attentive_encoder_question(condensed_question, padding=question_padding)
        final_question, (question_rnn_h, question_rnn_c) = self.bilstm_question(self_attended_question[-1], question_lengths)
        question_rnn_state = [self.reshape_rnn_state(x) for x in (question_rnn_h, question_rnn_c)]

        context_indices = context_limited if context_limited is not None else context
        question_indices = question_limited if question_limited is not None else question
        answer_indices = answer_limited if answer_limited is not None else answer

        pad_idx = self.field.decoder_stoi[self.field.pad_token]
        context_padding = context_indices.data == pad_idx
        question_padding = question_indices.data == pad_idx

        self.dual_ptr_rnn_decoder.applyMasks(context_padding, question_padding)

        if self.training:
            answer_padding = (answer_indices.data == pad_idx)[:, :-1]
            answer_embedded = self.decoder_embeddings(answer)
            self_attended_decoded = self.self_attentive_decoder(answer_embedded[:, :-1].contiguous(), self_attended_context, context_padding=context_padding, answer_padding=answer_padding, positional_encodings=True)
            decoder_outputs = self.dual_ptr_rnn_decoder(self_attended_decoded, 
                final_context, final_question, hidden=context_rnn_state)
            rnn_output, context_attention, question_attention, context_alignment, question_alignment, vocab_pointer_switch, context_question_switch, rnn_state = decoder_outputs

            probs = self.probs(self.out, rnn_output, vocab_pointer_switch, context_question_switch, 
                context_attention, question_attention, 
                context_indices, question_indices, 
                oov_to_limited_idx)


            if self.args.use_bleu_loss and iteration >= self.args.loss_switch * max(self.args.train_iterations):
                max_order = 4
                targets = answer_indices[:, 1:].contiguous()
                batch_size = targets.size(0)
                reference_lengths = [l-1 for l in answer_lengths]
                translation_len = max(reference_lengths)
                translation_lengths = torch.tensor([translation_len] * batch_size, device=self.device)

                bleu_loss_smoothed = expectedMultiBleu.bleu(probs, targets, translation_lengths, reference_lengths, max_order=max_order, smooth=True)
                loss = -1 * bleu_loss_smoothed[0]

            elif self.args.use_maxmargin_loss:
                targets = answer_indices[:, 1:].contiguous()
                loss = max_margin_loss(probs, targets, pad_idx=pad_idx)

            else:
                probs, targets = mask(answer_indices[:, 1:].contiguous(), probs.contiguous(), pad_idx=pad_idx)
                loss = F.nll_loss(probs.log(), targets)
            return loss, None

        else:
            return None, self.greedy(self_attended_context, final_context, final_question, 
                context_indices, question_indices,
                oov_to_limited_idx, rnn_state=context_rnn_state).data
 
    def reshape_rnn_state(self, h):
        return h.view(h.size(0) // 2, 2, h.size(1), h.size(2)) \
                .transpose(1, 2).contiguous() \
                .view(h.size(0) // 2, h.size(1), h.size(2) * 2).contiguous()

    def probs(self, generator, outputs, vocab_pointer_switches, context_question_switches, 
        context_attention, question_attention, 
        context_indices, question_indices, 
        oov_to_limited_idx):

        size = list(outputs.size())

        size[-1] = self.generative_vocab_size
        scores = generator(outputs.view(-1, outputs.size(-1))).view(size)
        p_vocab = F.softmax(scores, dim=scores.dim()-1)
        scaled_p_vocab = vocab_pointer_switches.expand_as(p_vocab) * p_vocab

        effective_vocab_size = self.generative_vocab_size + len(oov_to_limited_idx)
        if self.generative_vocab_size < effective_vocab_size:
            size[-1] = effective_vocab_size - self.generative_vocab_size
            buff = scaled_p_vocab.new_full(size, EPSILON)
            scaled_p_vocab = torch.cat([scaled_p_vocab, buff], dim=buff.dim()-1)

        # p_context_ptr
        scaled_p_vocab.scatter_add_(scaled_p_vocab.dim()-1, context_indices.unsqueeze(1).expand_as(context_attention), 
            (context_question_switches * (1 - vocab_pointer_switches)).expand_as(context_attention) * context_attention)

        # p_question_ptr
        scaled_p_vocab.scatter_add_(scaled_p_vocab.dim()-1, question_indices.unsqueeze(1).expand_as(question_attention), 
            ((1 - context_question_switches) * (1 - vocab_pointer_switches)).expand_as(question_attention) * question_attention)

        return scaled_p_vocab


    def greedy(self, self_attended_context, context, question, context_indices, question_indices, oov_to_limited_idx, rnn_state=None):
        B, TC, C = context.size()
        T = self.args.max_output_length
        outs = context.new_full((B, T), self.field.decoder_stoi['<pad>'], dtype=torch.long)
        hiddens = [self_attended_context[0].new_zeros((B, T, C))
                   for l in range(len(self.self_attentive_decoder.layers) + 1)]
        hiddens[0] = hiddens[0] + positional_encodings_like(hiddens[0])
        eos_yet = context.new_zeros((B, )).byte()
    
        rnn_output, context_alignment, question_alignment = None, None, None
        for t in range(T):
            if t == 0:
                embedding = self.decoder_embeddings(
                    self_attended_context[-1].new_full((B, 1), self.field.vocab.stoi['<init>'], dtype=torch.long), [1]*B)
            else:
                embedding = self.decoder_embeddings(outs[:, t - 1].unsqueeze(1), [1]*B)
            hiddens[0][:, t] = hiddens[0][:, t] + (math.sqrt(self.self_attentive_decoder.d_model) * embedding).squeeze(1)
            for l in range(len(self.self_attentive_decoder.layers)):
                hiddens[l + 1][:, t] = self.self_attentive_decoder.layers[l].feedforward(
                    self.self_attentive_decoder.layers[l].attention(
                    self.self_attentive_decoder.layers[l].selfattn(hiddens[l][:, t], hiddens[l][:, :t + 1], hiddens[l][:, :t + 1])
                  , self_attended_context[l], self_attended_context[l]))
            decoder_outputs = self.dual_ptr_rnn_decoder(hiddens[-1][:, t].unsqueeze(1),
                context, question, 
                context_alignment=context_alignment, question_alignment=question_alignment,
                hidden=rnn_state, output=rnn_output)
            rnn_output, context_attention, question_attention, context_alignment, question_alignment, vocab_pointer_switch, context_question_switch, rnn_state = decoder_outputs
            probs = self.probs(self.out, rnn_output, vocab_pointer_switch, context_question_switch, 
                context_attention, question_attention, 
                context_indices, question_indices, 
                oov_to_limited_idx)
            pred_probs, preds = probs.max(-1)
            preds = preds.squeeze(1)
            eos_yet = eos_yet | (preds == self.field.decoder_stoi['<eos>'])
            outs[:, t] = preds.cpu().apply_(self.map_to_full)
            if eos_yet.all():
                break
        return outs



class DualPtrRNNDecoder(nn.Module):

    def __init__(self, d_in, d_hid, dropout=0.0, num_layers=1):
        super().__init__()
        self.d_hid = d_hid
        self.d_in = d_in
        self.num_layers = num_layers
        self.dropout = nn.Dropout(dropout)

        self.input_feed = True
        if self.input_feed:
            d_in += 1 * d_hid

        self.rnn = LSTMDecoder(self.num_layers, d_in, d_hid, dropout)
        self.context_attn = LSTMDecoderAttention(d_hid, dot=True)
        self.question_attn = LSTMDecoderAttention(d_hid, dot=True)

        self.vocab_pointer_switch = nn.Sequential(Feedforward(2 * self.d_hid + d_in, 1), nn.Sigmoid())
        self.context_question_switch = nn.Sequential(Feedforward(2 * self.d_hid + d_in, 1), nn.Sigmoid())

    def forward(self, input, context, question, output=None, hidden=None, context_alignment=None, question_alignment=None):
        context_output = output.squeeze(1) if output is not None else self.make_init_output(context)
        context_alignment = context_alignment if context_alignment is not None else self.make_init_output(context)
        question_alignment = question_alignment if question_alignment is not None else self.make_init_output(question)

        context_outputs, vocab_pointer_switches, context_question_switches, context_attentions, question_attentions, context_alignments, question_alignments = [], [], [], [], [], [], []
        for emb_t in input.split(1, dim=1):
            emb_t = emb_t.squeeze(1)
            context_output = self.dropout(context_output)
            if self.input_feed:
                emb_t = torch.cat([emb_t, context_output], 1)
            dec_state, hidden = self.rnn(emb_t, hidden)
            context_output, context_attention, context_alignment = self.context_attn(dec_state, context)
            question_output, question_attention, question_alignment = self.question_attn(dec_state, question)
            vocab_pointer_switch = self.vocab_pointer_switch(torch.cat([dec_state, context_output, emb_t], -1))
            context_question_switch = self.context_question_switch(torch.cat([dec_state, question_output, emb_t], -1))
            context_output = self.dropout(context_output)
            context_outputs.append(context_output)
            vocab_pointer_switches.append(vocab_pointer_switch)
            context_question_switches.append(context_question_switch)
            context_attentions.append(context_attention)
            context_alignments.append(context_alignment)
            question_attentions.append(question_attention)
            question_alignments.append(question_alignment)

        context_outputs, vocab_pointer_switches, context_question_switches, context_attention, question_attention = [self.package_outputs(x) for x in [context_outputs, vocab_pointer_switches, context_question_switches, context_attentions, question_attentions]]
        return context_outputs, context_attention, question_attention, context_alignment, question_alignment, vocab_pointer_switches, context_question_switches, hidden


    def applyMasks(self, context_mask, question_mask):
        self.context_attn.applyMasks(context_mask)
        self.question_attn.applyMasks(question_mask)

    def make_init_output(self, context):
        batch_size = context.size(0)
        h_size = (batch_size, self.d_hid)
        return context.new_zeros(h_size)

    def package_outputs(self, outputs):
        outputs = torch.stack(outputs, dim=1)
        return outputs