Text2sql model for medical sql generation

I’m trying to deploy this model for generating SQL from text and establish an interactive interface where users can input natural language queries and receive SQL outputs. You will find below main.py and model.py. this model is taken from TREQS repository
this is main.py

import argparse
import torch
from model import modelABS
from LeafNATS.utils.utils import str2bool
from end2end_large import End2EndBase
parser = argparse.ArgumentParser()
'''
Use in the framework and cannot remove.
'''
parser.add_argument('--task', default='train', help='train | validate | test | evaluate')

parser.add_argument('--data_dir', default='mimicsql_data/mimicsql_natural', help='directory that store the data.')
parser.add_argument('--file_train', default='train.json', help='Training')
parser.add_argument('--file_val', default='dev.json', help='validation')
parser.add_argument('--file_test', default='test.json', help='test data')

parser.add_argument('--n_epoch', type=int, default=20, help='number of epochs.')
parser.add_argument('--batch_size', type=int, default=16, help='batch size.')
parser.add_argument('--checkpoint', type=int, default=100, help='How often you want to save model?')
parser.add_argument('--val_num_batch', type=int, default=100, help='how many batches')
parser.add_argument('--nbestmodel', type=int, default=5, help='How many models you want to keep?')

parser.add_argument('--continue_training', type=str2bool, default=True, help='Do you want to continue?')
parser.add_argument('--train_base_model', type=str2bool, default=False, help='True: Use Pretrained Param | False: Transfer Learning')
parser.add_argument('--use_move_avg', type=str2bool, default=False, help='move average')
parser.add_argument('--use_optimal_model', type=str2bool, default=True, help='Do you want to use the best model?')
parser.add_argument('--model_optimal_key', default='0,0', help='epoch,batch')
parser.add_argument('--is_lower', type=str2bool, default=True, help='convert all tokens to lower case?')
'''
User specified parameters.
'''
parser.add_argument('--device', default=torch.device("cuda:0"), help='device')
parser.add_argument('--file_vocab', default='vocab', help='file store training vocabulary.')

parser.add_argument('--max_vocab_size', type=int, default=50000, help='max number of words in the vocabulary.')
parser.add_argument('--word_minfreq', type=int, default=5, help='min word frequency')

parser.add_argument('--emb_dim', type=int, default=128, help='source embedding dimension')
parser.add_argument('--src_hidden_dim', type=int, default=256, help='encoder hidden dimension')
parser.add_argument('--trg_hidden_dim', type=int, default=256, help='decoder hidden dimension')
parser.add_argument('--src_seq_len', type=int, default=400, help='length of source documents.')
parser.add_argument('--trg_seq_len', type=int, default=100, help='length of target documents.')
parser.add_argument('--nLayers', type=int, default=1, help='Encoder RNN layers')

parser.add_argument('--learning_rate', type=float, default=0.0005, help='learning rate.')
parser.add_argument('--grad_clip', type=float, default=2.0, help='clip the gradient norm.')

parser.add_argument('--file_output', default='output.json', help='test output file')
parser.add_argument('--beam_size', type=int, default=5, help='beam size.')
parser.add_argument('--copy_words', type=str2bool, default=True, help='Do you want to copy words?')
# scheduler
parser.add_argument('--step_size', type=int, default=2, help='---')
parser.add_argument('--step_decay', type=float, default=0.8, help='---')

args = parser.parse_args()

if args.task == 'train' or args.task == 'validate' or args.task == 'test':
    from model import modelABS
    model = modelABS(args)
if args.task == "train":
    model.train()
if args.task == "validate":
    model.validate()
if args.task == "test":
    model.test()
if args.task == "evaluate":
    from LeafNATS.eval_scripts.eval_pyrouge_v2 import run_pyrouge
    run_pyrouge(args)

and this is model.py


import os
import time

import torch
from torch.autograd import Variable

from seq2sql.model_seq2seq_base import modelSeq2SeqBase
from LeafNATS.data.seq2sql.process_batch_cqa_v1 import process_batch

from LeafNATS.modules.embedding.nats_embedding import natsEmbedding
from LeafNATS.modules.encoder.encoder_rnn import EncoderRNN
from LeafNATS.modules.encoder2decoder.nats_encoder2decoder import natsEncoder2Decoder
from LeafNATS.modules.attention.nats_attention_encoder import AttentionEncoder
from LeafNATS.modules.attention.nats_attention_decoder import AttentionDecoder

# from LeafNATS.utils.utils import *

class modelABS(modelSeq2SeqBase):
    
    def __init__(self, args):
        super().__init__(args=args)
        
    def build_scheduler(self, optimizer):
        '''
        Schedule Learning Rate
        '''
        scheduler = torch.optim.lr_scheduler.StepLR(
            optimizer=optimizer, step_size=self.args.step_size, 
            gamma=self.args.step_decay)
        
        return scheduler
        
    def build_batch(self, batch_id):
        '''
        get batch data
        '''
        output = process_batch(
            batch_id=batch_id, 
            path_=os.path.join('..', 'nats_results'), 
            fkey_=self.args.task, 
            batch_size=self.args.batch_size, 
            vocab2id=self.batch_data['vocab2id'], 
            max_lens=[self.args.src_seq_len, self.args.trg_seq_len])
            
        self.batch_data['ext_id2oov'] = output['ext_id2oov']
        self.batch_data['src_var'] = output['src_var'].to(self.args.device)
        self.batch_data['batch_size'] = self.batch_data['src_var'].size(0)
        self.batch_data['src_seq_len'] = self.batch_data['src_var'].size(1)
        self.batch_data['src_mask_pad'] = output['src_mask_pad'].to(self.args.device)
        
        if self.args.task == 'train' or self.args.task == 'validate':
            self.batch_data['trg_input'] = output['trg_input_var'].to(self.args.device)
            # different from seq2seq models.
            self.batch_data['trg_output'] = output['trg_output_var'].to(self.args.device)
            self.batch_data['trg_seq_len'] = self.batch_data['trg_input'].size(1)
        else:
            self.batch_data['src_mask_unk'] = output['src_mask_unk'].to(self.args.device)
            self.batch_data['src_txt'] = output['src_txt']
            self.batch_data['trg_txt'] = output['trg_txt']
            self.batch_data['trg_seq_len'] = 1
        
    def build_models(self):
        '''
        build all models.
        in this model source and target share embeddings
        '''
        self.train_models['embedding'] = natsEmbedding(
            vocab_size = self.batch_data['vocab_size'],
            emb_dim = self.args.emb_dim,
            share_emb_weight = True
        ).to(self.args.device)
        
        self.train_models['encoder'] = EncoderRNN(
            self.args.emb_dim, self.args.src_hidden_dim,
            self.args.nLayers, 'lstm',
            device = self.args.device
        ).to(self.args.device)
        
        self.train_models['encoder2decoder'] = natsEncoder2Decoder(
            src_hidden_size = self.args.src_hidden_dim,
            trg_hidden_size = self.args.trg_hidden_dim,
            rnn_network = 'lstm',
            device = self.args.device
        ).to(self.args.device)
        
        self.train_models['decoderRNN'] = torch.nn.LSTMCell(
            self.args.emb_dim+self.args.trg_hidden_dim, 
            self.args.trg_hidden_dim
        ).to(self.args.device)
        
        self.train_models['attnEncoder'] = AttentionEncoder(
            self.args.src_hidden_dim,
            self.args.trg_hidden_dim,
            attn_method='luong_general',
            repetition='temporal'
        ).to(self.args.device)
        
        self.train_models['attnDecoder'] = AttentionDecoder(
            self.args.trg_hidden_dim,
            attn_method='luong_general'
        ).to(self.args.device)
        
        self.train_models['wrapDecoder'] = torch.nn.Linear(
            self.args.src_hidden_dim*2+self.args.trg_hidden_dim*2,
            self.args.trg_hidden_dim, bias=True
        ).to(self.args.device)
        
        self.train_models['genPrb'] = torch.nn.Linear(
            self.args.emb_dim+self.args.src_hidden_dim*2+self.args.trg_hidden_dim, 1
        ).to(self.args.device)
        
        # decoder to vocab
        self.train_models['decoder2proj'] = torch.nn.Linear(
            self.args.trg_hidden_dim, self.args.emb_dim, bias=False
        ).to(self.args.device)
                    
    def build_encoder(self):
        '''
        Encoder Pipeline
        self.pipe_data = {
            'encoder': {},
            'decoderA': {}}
            'decoderB': {'accu_attn': [], 'last_word': word}}
        '''
        src_emb = self.train_models['embedding'].get_embedding(
            self.batch_data['src_var'])
        src_enc, hidden_encoder = self.train_models['encoder'](src_emb)
        trg_hidden0 = self.train_models['encoder2decoder'](hidden_encoder)
        
        # set up pipe_data pass to decoder
        self.pipe_data['encoder'] = {}
        self.pipe_data['encoder']['src_emb'] = src_emb
        self.pipe_data['encoder']['src_enc'] = src_enc
        
        self.pipe_data['decoderB'] = {}
        self.pipe_data['decoderB']['hidden'] = trg_hidden0        
        self.pipe_data['decoderB']['h_attn'] = Variable(torch.zeros(
            self.batch_data['batch_size'], self.args.trg_hidden_dim
        )).to(self.args.device)
        self.pipe_data['decoderB']['past_attn'] = Variable(torch.ones(
            self.batch_data['batch_size'], self.batch_data['src_seq_len']
        )/float(self.batch_data['src_seq_len'])).to(self.args.device)
        self.pipe_data['decoderB']['past_dech'] = Variable(torch.zeros(
            1, 1)).to(self.args.device)
        self.pipe_data['decoderB']['accu_attn'] = []
        
        self.pipe_data['decoderFF'] = {}
        self.pipe_data['decoderFF']['h_attn'] = []
        self.pipe_data['decoderFF']['attn'] = []
        self.pipe_data['decoderFF']['genPrb'] = []
        # when training get target embedding at the same time.
        if self.args.task == 'train' or self.args.task == 'validate':
            trg_emb = self.train_models['embedding'].get_embedding(
                self.batch_data['trg_input'])
            self.pipe_data['decoderFF']['trg_seq_emb'] = trg_emb
        
    def build_decoder_one_step(self, k=0):
        '''
        Decoder one-step
        '''
        # embedding at current decoding step
        if self.args.task == 'train' or self.args.task == 'validate':
            self.pipe_data['decoderA'] = self.pipe_data['decoderB']
            word_emb = self.pipe_data['decoderFF']['trg_seq_emb'][:, k]
        else:
            word_emb = self.train_models['embedding'].get_embedding(
                self.pipe_data['decoderA']['last_word'])

        h_attn = self.pipe_data['decoderA']['h_attn']
        dec_input = torch.cat((word_emb, h_attn), 1)
        hidden = self.pipe_data['decoderA']['hidden']
        past_attn = self.pipe_data['decoderA']['past_attn']
        accu_attn = self.pipe_data['decoderA']['accu_attn']
        past_dech = self.pipe_data['decoderA']['past_dech']
        
        hidden = self.train_models['decoderRNN'](dec_input, hidden)
        ctx_enc, attn, attn_ee = self.train_models['attnEncoder'](
            hidden[0], self.pipe_data['encoder']['src_enc'], 
            past_attn, self.batch_data['src_mask_pad'])
        # temporal attention
        past_attn = past_attn + attn_ee
        # decoder attention
        if k == 0:
            ctx_dec = Variable(torch.zeros(
                self.batch_data['batch_size'], self.args.trg_hidden_dim
            )).to(self.args.device)
        else:
            ctx_dec, _ = self.train_models['attnDecoder'](
                hidden[0], past_dech)
        past_dech = past_dech.transpose(0, 1) # seqL*batch*hidden
        dec_idx = past_dech.size(0)
        if k == 0:
            past_dech = hidden[0].unsqueeze(0) # seqL*batch*hidden
            past_dech = past_dech.transpose(0, 1) # batch*seqL*hidden
        else:
            past_dech = past_dech.contiguous().view(
                -1, self.args.trg_hidden_dim) # seqL*batch**hidden
            past_dech = torch.cat((past_dech, hidden[0]), 0) # (seqL+1)*batch**hidden
            past_dech = past_dech.view(
                dec_idx+1, self.batch_data['batch_size'], self.args.trg_hidden_dim
            ) # (seqL+1)*batch*hidden
            past_dech = past_dech.transpose(0, 1) # batch*(seqL+1)*hidden
        # wrap up.
        h_attn = self.train_models['wrapDecoder'](torch.cat((ctx_enc, ctx_dec, hidden[0]), 1))
        # pointer generator
        pt_input = torch.cat((word_emb, hidden[0], ctx_enc), 1)
        genPrb = torch.sigmoid(self.train_models['genPrb'](pt_input))
        
        # setup piped_data
        self.pipe_data['decoderB'] = {}
        self.pipe_data['decoderB']['h_attn'] = h_attn
        self.pipe_data['decoderB']['past_attn'] = past_attn
        self.pipe_data['decoderB']['hidden'] = hidden
        self.pipe_data['decoderB']['past_dech'] = past_dech
        self.pipe_data['decoderB']['accu_attn'] = [a for a in accu_attn]
        self.pipe_data['decoderB']['accu_attn'].append(attn)
        
        if self.args.task == 'train' or self.args.task == 'validate':
            self.pipe_data['decoderFF']['h_attn'].append(h_attn)
            self.pipe_data['decoderFF']['attn'].append(attn)
            self.pipe_data['decoderFF']['genPrb'].append(genPrb)
            if k == self.batch_data['trg_seq_len']-1:
                self.pipe_data['decoderFF']['h_attn'] = \
                torch.cat(self.pipe_data['decoderFF']['h_attn'], 0).view(
                    self.batch_data['trg_seq_len'], 
                    self.batch_data['batch_size'], 
                    self.args.trg_hidden_dim).transpose(0,1)
                
                self.pipe_data['decoderFF']['attn'] = \
                torch.cat(self.pipe_data['decoderFF']['attn'], 0).view(
                    self.batch_data['trg_seq_len'], 
                    self.batch_data['batch_size'], 
                    self.args.src_seq_len).transpose(0,1)
                
                self.pipe_data['decoderFF']['genPrb'] = \
                torch.cat(self.pipe_data['decoderFF']['genPrb'], 0).view(
                    self.batch_data['trg_seq_len'], 
                    self.batch_data['batch_size']).transpose(0,1)
        else:
            self.pipe_data['decoderFF']['h_attn'] = h_attn
            self.pipe_data['decoderFF']['attn'] = attn.unsqueeze(0)
            self.pipe_data['decoderFF']['genPrb'] = genPrb

    def build_vocab_distribution(self):
        '''
        Data flow from input to output.
        '''
        trg_out = self.pipe_data['decoderFF']['h_attn']
        trg_out = self.train_models['decoder2proj'](trg_out)
        trg_out = self.train_models['embedding'].get_decode2vocab(trg_out)
        trg_out = trg_out.view(
            self.batch_data['batch_size'], self.batch_data['trg_seq_len'], -1)
        prb = torch.softmax(trg_out, dim=2)
        
        vocab_size = self.batch_data['vocab_size']
        batch_size = self.batch_data['batch_size']
        # trg_seq_len = self.batch_data['trg_seq_len']
        src_seq_len = self.batch_data['src_seq_len']
        
        # pointer-generator calculate index matrix
        pt_idx = Variable(torch.FloatTensor(torch.zeros(1, 1, 1))).to(self.args.device)
        pt_idx = pt_idx.repeat(batch_size, src_seq_len, vocab_size)
        pt_idx.scatter_(2, self.batch_data['src_var'].unsqueeze(2), 1.0)
               
        p_gen = self.pipe_data['decoderFF']['genPrb']
        attn_ = self.pipe_data['decoderFF']['attn']
        
        prb_output = p_gen.unsqueeze(2)*prb + \
                     (1.0-p_gen.unsqueeze(2))*torch.bmm(attn_, pt_idx)
                
        return prb_output + 1e-20
    
    def build_pipelines(self):
        '''
        Build pipeline from input to output.
        Output is loss.
        Input is word one-hot encoding.
        '''
        self.build_encoder()
        for k in range(self.args.trg_seq_len):
            self.build_decoder_one_step(k)
        prb = self.build_vocab_distribution()
        
        pad_mask = torch.ones(self.batch_data['vocab_size']).to(self.args.device)
        pad_mask[self.batch_data['vocab2id']['<pad>']] = 0
        self.loss_criterion = torch.nn.NLLLoss(pad_mask).to(self.args.device)
        
        prb = torch.log(prb)
        loss = self.loss_criterion(
            prb.reshape(-1, self.batch_data['vocab_size']),
            self.batch_data['trg_output'].view(-1))
        
        return loss