corechain_with_sltr.py

'''
    create loss function and training data and other necessary utilities

    TODO:
        > Add visaulization so as to understand the interplay of train vs Validation vs test accuracy (not really going to do it)
        > Add logs
            - Need to discuss it before implementing.
'''
from __future__ import print_function

# Torch imports
import torch
import torch.nn as nn
from torch import optim
from torch.utils.data import  DataLoader

# Local imports
import data_loader as dl
import auxiliary as aux
import network as net
from configs import config_loader as cl
from utils import query_graph_to_sparql as sparql_constructor
from utils import dbpedia_interface as db_interface
from utils import embeddings_interface
from utils import natural_language_utilities as nlutils
import network_rdftype as net_rdftype
import network_intent as net_intent
import onefile as one
import lriters
from utils.goodies import *


# Other libs
import numpy as np
import argparse
import json
import time
import os
import sys
import pickle
from pprint import pprint
from progressbar import ProgressBar


if sys.version_info[0] == 3: import configparser as ConfigParser
else: import ConfigParser

#setting up device,model name and loss types.
device = torch.device("cpu")
training_model = 'bilstm_dot'
_dataset = 'lcquad'
pointwise = False
_train_over_validation = False


#Loading relations file.
COMMON_DATA_DIR = 'data/data/common'
_dataset_specific_data_dir = 'data/data/%(dataset)s/' % {'dataset': _dataset}
_inv_relations = aux.load_inverse_relation(COMMON_DATA_DIR)
_word_to_id = aux.load_word_list(COMMON_DATA_DIR)

#onefile stuff
dbp = db_interface.DBPedia(_verbose=True, caching=True)
vocabularize_relation = lambda path: embeddings_interface.vocabularize(nlutils.tokenize(dbp.get_label(path))).tolist()
sparql_constructor.init(embeddings_interface)

# gloveid_to_embeddingid , embeddingid_to_gloveid, word_to_gloveid, gloveid_to_word = aux.load_embeddingid_gloveid()


def load_data(data,parameter_dict,pointwise,schema='default',shuffle = False):
    # Loading training data
    td = dl.TrainingDataGenerator(data,
                                  parameter_dict['max_length'],
                                  parameter_dict['_neg_paths_per_epoch_train'], parameter_dict['batch_size']
                                  , parameter_dict['total_negative_samples'], pointwise=pointwise,schema=schema)
    return DataLoader(td, shuffle=shuffle)


def curatail_padding(data,parameter_dict):

    '''
        Since schema is already implicitly defined/present in the parameter_dict['rel1_pad']
    '''

    data['test_neg_paths'] = data['test_neg_paths'][:, :, :parameter_dict['rel_pad']]
    data['test_pos_paths'] = data['test_pos_paths'][:, :parameter_dict['rel_pad']]

    if parameter_dict['schema'] == 'reldet':
        data['test_neg_paths_rel1_rd'] = data['test_neg_paths_rel1_rd'][:, :, :parameter_dict['rel1_pad']]
        data['test_neg_paths_rel2_rd'] = data['test_neg_paths_rel2_rd'][:, :, :parameter_dict['rel1_pad']]
        data['test_pos_paths_rel1_rd'] = data['test_pos_paths_rel1_rd'][:, :parameter_dict['rel1_pad']]
        data['test_pos_paths_rel2_rd'] = data['test_pos_paths_rel2_rd'][:, :parameter_dict['rel1_pad']]
    elif parameter_dict['schema'] == 'slotptr':
        data['test_neg_paths_rel1_sp'] = data['test_neg_paths_rel1_sp'][:, :, :parameter_dict['rel1_pad']]
        data['test_neg_paths_rel2_sp'] = data['test_neg_paths_rel2_sp'][:, :, :parameter_dict['rel1_pad']]
        data['test_pos_paths_rel1_sp'] = data['test_pos_paths_rel1_sp'][:, :parameter_dict['rel1_pad']]
        data['test_pos_paths_rel2_sp'] = data['test_pos_paths_rel2_sp'][:, :parameter_dict['rel1_pad']]

    data['valid_neg_paths'] = data['valid_neg_paths'][:, :, :parameter_dict['rel_pad']]
    data['valid_pos_paths'] = data['valid_pos_paths'][:, :parameter_dict['rel_pad']]

    if parameter_dict['schema'] == 'reldet':
        data['valid_neg_paths_rel1_rd'] = data['valid_neg_paths_rel1_rd'][:, :, :parameter_dict['rel1_pad']]
        data['valid_neg_paths_rel2_rd'] = data['valid_neg_paths_rel2_rd'][:, :, :parameter_dict['rel1_pad']]
        data['valid_pos_paths_rel1_rd'] = data['valid_pos_paths_rel1_rd'][:, :parameter_dict['rel1_pad']]
        data['valid_pos_paths_rel2_rd'] = data['valid_pos_paths_rel2_rd'][:, :parameter_dict['rel1_pad']]
    elif parameter_dict['schema'] == 'slotptr':
        data['valid_neg_paths_rel1_sp'] = data['valid_neg_paths_rel1_sp'][:, :, :parameter_dict['rel1_pad']]
        data['valid_neg_paths_rel2_sp'] = data['valid_neg_paths_rel2_sp'][:, :, :parameter_dict['rel1_pad']]
        data['valid_pos_paths_rel1_sp'] = data['valid_pos_paths_rel1_sp'][:, :parameter_dict['rel1_pad']]
        data['valid_pos_paths_rel2_sp'] = data['valid_pos_paths_rel2_sp'][:, :parameter_dict['rel1_pad']]

    return data


def training_loop(training_model, parameter_dict,modeler,train_loader,
                  optimizer,loss_func, data, dataset, device, test_every, validate_every , pointwise = False, problem='core_chain',lrschedule=None,curtail_padding_rel=True):

    model_save_location = aux.save_location(problem, training_model, dataset)
    aux_save_information = {
        'epoch' : 0,
        'test_accuracy':0.0,
        'validation_accuracy':0.0,
        'parameter_dict':parameter_dict
    }
    train_loss = []
    valid_accuracy = []
    test_accuracy = []
    best_validation_accuracy = 0
    best_test_accuracy = 0

    if parameter_dict['schema'] == 'reldet':
        parameter_dict['rel1_pad'] =  parameter_dict['relrd_pad']
    elif parameter_dict['schema'] == 'slotptr':
        parameter_dict['rel1_pad'] = parameter_dict['relsp_pad']

        ###############
    # Training Loop
    ###############


    #Makes test data of appropriate shape
    print("the dataset is ", dataset)
    if curtail_padding_rel and dataset == 'lcquad':
        data = curatail_padding(data, parameter_dict)

    try:

        for epoch in range(parameter_dict['epochs']):

            # Epoch start print
            print("Epoch: ", epoch, "/", parameter_dict['epochs'])

            # Bookkeeping variables
            epoch_loss = []
            epoch_time = time.time()

            # Loop for one batch
            # tqdm_loop = tqdm(enumerate(train_loader))
            for i_batch, sample_batched in enumerate(train_loader):

                # Bookkeeping and data preparation
                batch_time = time.time()

                if lr_schedule:
                    update_lr(optimizer, lrschedule.get())

                if not pointwise:
                    ques_batch = torch.tensor(np.reshape(sample_batched[0][0], (-1, parameter_dict['max_length'])),
                                              dtype=torch.long, device=device)
                    pos_batch = torch.tensor(np.reshape(sample_batched[0][1], (-1, parameter_dict['max_length'])),
                                             dtype=torch.long, device=device)
                    neg_batch = torch.tensor(np.reshape(sample_batched[0][2], (-1, parameter_dict['max_length'])),
                                             dtype=torch.long, device=device)

                    data['dummy_y'] = torch.ones(ques_batch.shape[0], device=device)

                    if parameter_dict['schema'] != 'default':
                        pos_rel1_batch = torch.tensor(np.reshape(sample_batched[0][3], (-1, parameter_dict['max_length'])),
                                                      dtype=torch.long, device=device)
                        pos_rel2_batch = torch.tensor(np.reshape(sample_batched[0][4], (-1, parameter_dict['max_length'])),
                                                      dtype=torch.long, device=device)
                        neg_rel1_batch = torch.tensor(np.reshape(sample_batched[0][5], (-1, parameter_dict['max_length'])),
                                                      dtype=torch.long, device=device)
                        neg_rel2_batch = torch.tensor(np.reshape(sample_batched[0][6], (-1, parameter_dict['max_length'])),
                                                      dtype=torch.long, device=device)


                        data_batch = {
                            'ques_batch': ques_batch,
                            'pos_batch': pos_batch[:,:parameter_dict['rel_pad']],
                            'neg_batch': neg_batch[:,:parameter_dict['rel_pad']],
                            'y_label': data['dummy_y'],
                            'pos_rel1_batch': pos_rel1_batch[:,:parameter_dict['rel1_pad']],
                            'pos_rel2_batch':pos_rel2_batch[:,:parameter_dict['rel1_pad']],
                            'neg_rel1_batch':neg_rel1_batch[:,:parameter_dict['rel1_pad']],
                            'neg_rel2_batch' : neg_rel2_batch[:,:parameter_dict['rel1_pad']]
                        }

                    else:


                        data_batch = {
                            'ques_batch': ques_batch,
                            'pos_batch': pos_batch[:,:parameter_dict['rel_pad']],
                            'neg_batch': neg_batch[:,:parameter_dict['rel_pad']],
                            'y_label': data['dummy_y']}

                else:
                    ques_batch = torch.tensor(np.reshape(sample_batched[0][0], (-1, parameter_dict['max_length'])),
                                              dtype=torch.long, device=device)
                    path_batch = torch.tensor(np.reshape(sample_batched[0][1], (-1, parameter_dict['max_length'])),
                                             dtype=torch.long, device=device)
                    y = torch.tensor(sample_batched[1],dtype = torch.float,device=device).view(-1)


                    if parameter_dict['schema'] != 'default':
                        path_rel1_batch = torch.tensor(np.reshape(sample_batched[0][2], (-1, parameter_dict['max_length'])),
                                                      dtype=torch.long, device=device)
                        path_rel2_batch = torch.tensor(np.reshape(sample_batched[0][3], (-1, parameter_dict['max_length'])),
                                                      dtype=torch.long, device=device)

                        data_batch = {
                            'ques_batch': ques_batch,
                            'path_batch': path_batch[:,:parameter_dict['rel_pad']],
                            'y_label': y,
                            'path_rel1_batch': path_rel1_batch[:,:parameter_dict['rel1_pad']],
                            'path_rel2_batch': path_rel2_batch[:,:parameter_dict['rel1_pad']]
                        }
                    else:
                        data_batch = {
                            'ques_batch': ques_batch,
                            'path_batch': path_batch[:,:parameter_dict['rel_pad']],
                            'y_label': y
                        }




                loss = modeler.train(data=data_batch,
                                  optimizer=optimizer,
                                  loss_fn=loss_func,
                                  device=device)

                # Bookkeep the training loss
                epoch_loss.append(loss.item())

                # tqdm_loop.desc("#"+str(i_batch)+"\tLoss:" + str(loss.item())[:min(5, len(str(loss.item())))])

                print("Batch:\t%d" % i_batch, "/%d\t: " % (parameter_dict['batch_size']),
                      "%s" % (time.time() - batch_time),
                      "\t%s" % (time.time() - epoch_time),
                      "\t%s" % (str(loss.item())),
                      end=None if i_batch + 1 == int(int(i_batch) / parameter_dict['batch_size']) else "\n")

            # EPOCH LEVEL

            # Track training loss
            train_loss.append(epoch_loss)

            # test_every = False
            if test_every:
                # Run on test set
                if epoch%test_every == 0:
                    if parameter_dict['schema'] != 'default':
                        if parameter_dict['schema']  == 'slotptr':
                            test_accuracy.append(aux.validation_accuracy(data['test_questions'], data['test_pos_paths'],
                                                             data['test_neg_paths'],modeler, device, data['test_pos_paths_rel1_sp'],data['test_pos_paths_rel2_sp'],
                                                                 data['test_neg_paths_rel1_sp'],data['test_neg_paths_rel2_sp']))
                        else:
                            test_accuracy.append(aux.validation_accuracy(data['test_questions'], data['test_pos_paths'],
                                                                         data['test_neg_paths'], modeler, device,
                                                                         data['test_pos_paths_rel1_rd'],
                                                                         data['test_pos_paths_rel2_rd'],
                                                                         data['test_neg_paths_rel1_rd'],
                                                                         data['test_neg_paths_rel2_rd']))
                    else:
                        test_accuracy.append(aux.validation_accuracy(data['test_questions'], data['test_pos_paths'],
                                                                     data['test_neg_paths'], modeler, device))
                    if test_accuracy[-1] >= best_test_accuracy:
                        best_test_accuracy = test_accuracy[-1]
                        aux_save_information['test_accuracy'] = best_test_accuracy
            else:
                test_accuracy.append(0)
                best_test_accuracy = 0

            # Run on validation set
            if validate_every:
                if epoch%validate_every == 0:
                    if parameter_dict['schema'] != 'default':
                        if parameter_dict['schema'] == 'slotptr':
                            valid_accuracy.append(aux.validation_accuracy(data['valid_questions'], data['valid_pos_paths'],
                                                              data['valid_neg_paths'],  modeler, device, data['valid_pos_paths_rel1_sp'],data['valid_pos_paths_rel2_sp'],
                                                                 data['valid_neg_paths_rel1_sp'],data['valid_neg_paths_rel2_sp']))
                        else:
                            valid_accuracy.append(aux.validation_accuracy(data['valid_questions'][:-1], data['valid_pos_paths'][:-1],
                                                                          data['valid_neg_paths'][:-1], modeler, device,
                                                                          data['valid_pos_paths_rel1_rd'][:-1],
                                                                          data['valid_pos_paths_rel2_rd'][:-1],
                                                                          data['valid_neg_paths_rel1_rd'][:-1],
                                                                          data['valid_neg_paths_rel2_rd'][:-1]))
                    else:
                        valid_accuracy.append(aux.validation_accuracy(data['valid_questions'], data['valid_pos_paths'],
                                                                      data['valid_neg_paths'], modeler, device))
                    if valid_accuracy[-1] > best_validation_accuracy:
                        print("MODEL WEIGHTS RIGHT NOW: ", modeler.get_parameter_sum())
                        best_validation_accuracy = valid_accuracy[-1]
                        aux_save_information['epoch'] = epoch
                        aux_save_information['validation_accuracy'] = best_validation_accuracy
                        aux.save_model(model_save_location, modeler, model_name='model.torch'
                                   , epochs=epoch, optimizer=optimizer, accuracy=best_validation_accuracy, aux_save_information=aux_save_information)

            # Resample new negative paths per epoch and shuffle all data
            train_loader.dataset.shuffle()

            # Epoch level prints
            print("Time: %s\t" % (time.time() - epoch_time),
                  "Loss: %s\t" % (sum(epoch_loss)),
                  "Valdacc: %s\t" % (valid_accuracy[-1]),
                    "Testacc: %s\n" % (test_accuracy[-1]),
                  "BestValidAcc: %s\n" % (best_validation_accuracy),
                  "BestTestAcc: %s\n" % (best_test_accuracy))

        return train_loss, modeler, valid_accuracy, test_accuracy, model_save_location

    except KeyboardInterrupt:
        print('-' * 89)
        return train_loss, modeler, valid_accuracy, test_accuracy, model_save_location



def evaluate(device,pointwise,dataset,training_model,training_model_number,finetune=False):




    # Reading and setting up config parser
    config = ConfigParser.ConfigParser()
    config.readfp(open('configs/macros.cfg'))

    # setting up device,model name and loss types.
    device = device
    training_model = training_model
    _dataset = dataset
    pointwise = pointwise

    # 19 is performing the best
    training_model_number = training_model_number
    _debug = False

    # Loading relations file.
    COMMON_DATA_DIR = 'data/data/common'
    INTENTS = ['count', 'ask', 'list']
    RDFTYPES = ['x', 'uri', 'none']

    _dataset_specific_data_dir = 'data/data/%(dataset)s/' % {'dataset': _dataset}
    # glove_id_sf_to_glove_id_rel = dl.create_relation_lookup_table(COMMON_DATA_DIR)

    # Model specific paramters
    # #Model specific paramters
    if pointwise:
        training_config = 'pointwise'
    else:
        training_config = 'pairwise'

    parameter_dict = cl.runtime_parameters(dataset=_dataset, training_model=training_model,
                                           training_config=training_config, config_file='configs/macros.cfg')

    if training_model == 'cnn_dot':
        parameter_dict['output_dim'] = int(config.get(_dataset, 'output_dim'))

    # Update parameters
    parameter_dict['_dataset_specific_data_dir'] = _dataset_specific_data_dir
    parameter_dict['_model_dir'] = './data/models/'

    parameter_dict['corechainmodel'] = training_model
    parameter_dict['corechainmodelnumber'] = str(training_model_number)

    parameter_dict['intentmodel'] = 'bilstm_dense'
    parameter_dict['intentmodelnumber'] = '4'

    parameter_dict['rdftypemodel'] = 'bilstm_dense'
    parameter_dict['rdftypemodelnumber'] = '3'

    parameter_dict['rdfclassmodel'] = 'bilstm_dot'
    parameter_dict['rdfclassmodelnumber'] = '4'

    parameter_dict['dataset'] = _dataset


    TEMP = aux.data_loading_parameters(_dataset, parameter_dict, runtime=True)

    _dataset_specific_data_dir, _model_specific_data_dir, _file, \
    _max_sequence_length, _neg_paths_per_epoch_train, \
    _neg_paths_per_epoch_validation, _training_split, _validation_split, _index = TEMP

    _data,  _vectors = dl.create_dataset_runtime(file=_file, _dataset=_dataset,
                                                                         _dataset_specific_data_dir=_dataset_specific_data_dir,
                                                                         split_point=.80)

    parameter_dict['vectors'] = _vectors

    # For interpretability's sake
    # word_to_gloveid, gloveid_to_word = aux.load_embeddingid_gloveid()

    """
        Different counters and metrics to store accuracy of diff modules

            Core chain accuracy counter counts the number of time the core chain predicated is same as 
            positive path. This also includes for ask query.
            The counter might confuse the property and the ontology. 

            Similar functionality with rdf_type and intent

            **word vector accuracy counter**: 
                Counts the number of times just using  word2vec similarity, 
                the best path came the most similar. 
                This will only work if CANDIDATE_SPACE is not none.

    """

    '''
        c_flag  is true if the core_chain was correctly predicted. 
        same is the case for i_flag and r_flag, rt_flag (correct candidate for rdf type)
    '''
    c_flag, i_flag, r_flag, rt_flag = False, False, False, False

    '''
        Stores tuple of (fmeasure,precision,recall)
    '''
    results = []

    Logging = parameter_dict.copy()
    Logging['runtime'] = []

    quesans = one.QuestionAnswering(parameter_dict, pointwise, _word_to_id, device, _dataset, False)

    # Some logs which run during runtime, not after.
    core_chain_acc_log = []
    core_chain_mrr_log = []

    startindex = 0

    for index, data in enumerate(_data[startindex:]):
        print (index)
        if index == 16 or index == 25:
            continue
        print(data.keys())
        index += startindex

        log, metrics = one.answer_question(qa=quesans,
                                       index=index,
                                       data=data,
                                       relations=_inv_relations,
                                       parameter_dict=parameter_dict)

        #     log, metrics = answer_question(qa=None,
        #                                    index=None,
        #                                    data=None,
        #                                    gloveid_to_embeddingid=None,
        #                                    embeddingid_to_gloveid=None,
        #                                    relations=None,
        #                                    parameter_dict=None)

        sparql = one.create_sparql(log=log,
                               data=data,
                               embeddings_interface=embeddings_interface,
                               relations=_inv_relations)
        # sparql = ""
        # metrics = eval(data, log, metrics)

        # Update logs
        Logging['runtime'].append({'log': log, 'metrics': metrics,
                                   'pred_sparql': sparql, 'true_sparql': data['parsed-data']['node']['sparql_query']})

        # Update metrics
        core_chain_acc_log.append(metrics['core_chain_accuracy_counter'])
        core_chain_mrr_log.append(metrics['core_chain_mrr_counter'])

        # Make shit interpretable
        # question = aux.id_to_word(log['question'], gloveid_to_word, remove_pad=True)
        # true_path = aux.id_to_word(log['true_path'], gloveid_to_word, remove_pad=True)
        # pred_path = aux.id_to_word(log['pred_path'], gloveid_to_word, remove_pad=True)

        print("#%s" % index, "\t\bAcc: ", np.mean(core_chain_acc_log))

        #     print("\t\bQues: ", question)
        #     print("\t\bTPath: ", true_path, "\n\t\bPPath: ", pred_path)

        #     print("\t\bTIntent: ", log['true_intent'])
        #     print("\t\bPIntent: ", log['pred_intent'])
        #     print("\t\bPRdftype: ", log['true_rdf_type'])
        #     print("\t\bTRdftype: ", log['pred_rdf_type'])
        #     print("\t\bPRdfclass: ", log['true_rdf_class'])
        #     print("\t\bTRdfclass: ", log['pred_rdf_class'])

        #     print("")
        #     pprint(log)
        #     print("")
        pprint(metrics)
        #     print("\n",sparql)
        print("\n################################\n")

    Logging = one.evaluate(Logging,dbp)
    if not finetune:
        model_path = os.path.join(parameter_dict['_model_dir'], 'core_chain')
        model_path = os.path.join(model_path, parameter_dict['corechainmodel'])
        model_path = os.path.join(model_path, parameter_dict['dataset'])
        model_path = os.path.join(model_path, parameter_dict['corechainmodelnumber'])
        pickle.dump(Logging,open(model_path+'/result.pickle','wb+'))
    else:
        return np.mean(core_chain_acc_log)



if __name__ == "__main__":

    parser = argparse.ArgumentParser()

    parser.add_argument('-dataset', action='store', dest='dataset',
                        help='dataset includes lcquad,qald',default = 'lcquad')

    parser.add_argument('-model', action='store', dest='model',
                        help='name of the model to use',default='cnn_dot')

    parser.add_argument('-pointwise', action='store', dest='pointwise',
                        help='to use pointwise training procedure make it true',default=True)

    parser.add_argument('-train_valid', action='store', dest='train_over_validation',
                        help='train over validation', default=False)

    parser.add_argument('-device', action='store', dest='device',
                        help='cuda for gpu else cpu', default='cpu')

    parser.add_argument('-finetune', action='store', dest='finetune',
                        help='train over validation', default=False)

    parser.add_argument('-bidirectional', action='store', dest='bidirectional',
                        help='train over validation', default=True)

    parser.add_argument('-evals', action='store', dest='evals',
                        help='train over validation', default=True)

    parser.add_argument('-ilr', action='store', dest='ilr',
                        help='init lr', default=0.001)


    args = parser.parse_args()

    # setting up device,model namenpp and loss types.
    device = torch.device(args.device)
    training_model = args.model
    _dataset = args.dataset
    pointwise = aux.to_bool(args.pointwise)
    _train_over_validation = aux.to_bool(args.train_over_validation)
    finetune = aux.to_bool(args.finetune)
    bidirectional = aux.to_bool(args.bidirectional)
    evals = aux.to_bool(args.evals)
    ilr = float(args.ilr)

    # #Model specific paramters
    if pointwise:
        training_config = 'pointwise'
    else:
        training_config = 'pairwise'

    parameter_dict = cl.corechain_parameters(dataset=_dataset,training_model=training_model,
                                             training_config=training_config,config_file='configs/macros.cfg')


    if _dataset == 'lcquad':
        test_every = parameter_dict['test_every']
    else:
        test_every = False
    validate_every = parameter_dict['validate_every']


    data = aux.load_data(_dataset=_dataset, _train_over_validation = _train_over_validation,
                         _parameter_dict=parameter_dict, _relations =  _inv_relations, _pointwise=pointwise, _device=device,k=-1)

    if training_model == 'reldet':
        schema = 'reldet'
    elif training_model == 'slotptr' or training_model == 'slotptr_common_encoder' or training_model == 'slotptrortho'\
            or training_model == 'ulmfit_slotptr':
        schema = 'slotptr'
    elif training_model == 'bilstm_dot_multiencoder':
        schema = 'default'
    else:
        schema = 'default'

    train_loader = load_data(data, parameter_dict, pointwise, schema=schema)
    # if training_model == 'bilstm_dot_ulmfit':
    #     _, data['vectors'] = vocab_master.load_ulmfit()
    parameter_dict['vectors'] = data['vectors']
    parameter_dict['schema'] = schema
    if not bidirectional:
        parameter_dict['bidirectional'] = False


    if training_model == 'bilstm_dot':
        if not finetune:
            print("********", parameter_dict['bidirectional'])
            modeler = net.BiLstmDot( _parameter_dict = parameter_dict,_word_to_id=_word_to_id,
                                                 _device=device,_pointwise=pointwise, _debug=False)

            optimizer = optim.Adam(list(filter(lambda p: p.requires_grad, modeler.encoder.parameters())))
        else:
            print("^^^^^^^^^^^^^^^^^^finetuning^^^^^^^^^^^^^^^^")
            if pointwise:
                model_path = 'data/models/core_chain/bilstm_dot_pointwise/lcquad/19/model.torch'
            else:
                model_path = 'data/models/core_chain/bilstm_dot/lcquad/13/model.torch'
            modeler = net.BiLstmDot(_parameter_dict=parameter_dict, _word_to_id=_word_to_id,
                                    _device=device, _pointwise=pointwise, _debug=False)

            modeler.load_from(model_path)
            optimizer = optim.Adam(list(filter(lambda p: p.requires_grad, modeler.encoder.parameters())),lr=ilr)

    if training_model == 'bilstm_dense':
        modeler = net.BiLstmDense( _parameter_dict = parameter_dict,_word_to_id=_word_to_id,
                                             _device=device,_pointwise=pointwise, _debug=False)

        optimizer = optim.Adam(list(filter(lambda p: p.requires_grad, modeler.encoder.parameters()))+
                               list(filter(lambda p: p.requires_grad, modeler.dense.parameters())))

    if training_model == 'bilstm_densedot':
        modeler = net.BiLstmDenseDot( _parameter_dict = parameter_dict,_word_to_id=_word_to_id,
                                             _device=device,_pointwise=pointwise, _debug=False)

        optimizer = optim.Adam(list(filter(lambda p: p.requires_grad, modeler.encoder.parameters())) +
                               list(filter(lambda p: p.requires_grad, modeler.dense.parameters())))

    if training_model == 'cnn_dot':
        modeler = net.CNNDot( _parameter_dict = parameter_dict,_word_to_id=_word_to_id,
                                             _device=device,_pointwise=pointwise, _debug=False)

        optimizer = optim.Adam(list(filter(lambda p: p.requires_grad, modeler.encoder.parameters())))

    if training_model == 'decomposable_attention':
        if not finetune:
            modeler = net.DecomposableAttention(_parameter_dict=parameter_dict, _word_to_id=_word_to_id,
                                                _device=device, _pointwise=pointwise, _debug=False)

            optimizer = optim.Adam(list(filter(lambda p: p.requires_grad, modeler.encoder.parameters())) +
                                   list(filter(lambda p: p.requires_grad, modeler.scorer.parameters())))
        else:
            if not pointwise:
                model_path = 'data/models/core_chain/decomposable_attention/lcquad/71/model.torch'
            else:
                model_path = 'data/models/core_chain/decomposable_attention_pointwise/lcquad/8/model.torch'

            modeler = net.DecomposableAttention(_parameter_dict=parameter_dict, _word_to_id=_word_to_id,
                                                _device=device, _pointwise=pointwise, _debug=False)

            modeler.load_from(model_path)
            optimizer = optim.Adam(list(filter(lambda p: p.requires_grad, modeler.encoder.parameters())) +
                                   list(filter(lambda p: p.requires_grad, modeler.scorer.parameters())),lr=0.0001)

    if training_model == 'reldet':
        modeler = net.RelDetection(_parameter_dict=parameter_dict, _word_to_id=_word_to_id,
                                            _device=device, _pointwise=pointwise, _debug=False)
        optimizer = optim.Adam(list(filter(lambda p: p.requires_grad, modeler.encoder.parameters())))
    # if False:
    # # if training_model == 'slotptr':00
    #     modeler = net.SlotPointerModel(_parameter_dict=parameter_dict, _word_to_id=_word_to_id,
    #                                         _device=device, _pointwise=pointwise, _debug=False)
    #     optimizer = optim.Adam(list(filter(lambda p: p.requires_grad, modeler.encoder_q.parameters())) +
    #                            list(filter(lambda p: p.requires_grad, modeler.encoder_p.parameters())) +
    #                            list(filter(lambda p: p.requires_grad, modeler.comparer.parameters())))
    if training_model == 'slotptr':

        if not finetune:
            modeler = net.QelosSlotPointerModel(_parameter_dict=parameter_dict, _word_to_id=_word_to_id,
                                                _device=device, _pointwise=pointwise, _debug=False)
            optimizer = optim.Adam(list(filter(lambda p: p.requires_grad, modeler.encoder_q.parameters())) +
                                   list(filter(lambda p: p.requires_grad, modeler.encoder_p.parameters())), weight_decay=0.0001)
        else:
            print("^^^^^^^^^^^^finetuning^^^^^^^")
            if pointwise:
                model_path = 'data/models/core_chain/slotptr_pointwise/lcquad/8/model.torch'
            else:
                model_path = 'data/models/core_chain/slotptr/transfer-b/0/model.torch'
            modeler = net.QelosSlotPointerModel(_parameter_dict=parameter_dict, _word_to_id=_word_to_id,
                                                               _device=device, _pointwise=pointwise, _debug=False)
            optimizer = optim.Adam(list(filter(lambda p: p.requires_grad, modeler.encoder_q.parameters())) +
                                   list(filter(lambda p: p.requires_grad, modeler.encoder_p.parameters())),
                                   weight_decay=0.0001,lr=ilr)
            modeler.load_from(model_path)


    if training_model == 'slotptrortho':

        if not finetune:
            modeler = net.QelosSlotPointerModelOrthogonal(_parameter_dict=parameter_dict, _word_to_id=_word_to_id,
                                                _device=device, _pointwise=pointwise, _debug=False)
            optimizer = optim.Adam(list(filter(lambda p: p.requires_grad, modeler.encoder_q.parameters())) +
                                   list(filter(lambda p: p.requires_grad, modeler.encoder_p.parameters())), weight_decay=0.0001)
        else:
            model_path = 'data/models/core_chain/slotptr/lcquad/94/model.torch'
            modeler = net.QelosSlotPointerModelOrthogonal(_parameter_dict=parameter_dict, _word_to_id=_word_to_id,
                                                               _device=device, _pointwise=pointwise, _debug=False)
            optimizer = optim.Adam(list(filter(lambda p: p.requires_grad, modeler.encoder_q.parameters())) +
                                   list(filter(lambda p: p.requires_grad, modeler.encoder_p.parameters())),
                                   weight_decay=0.0001,lr=0.0001)
            modeler.load_from(model_path)


    if training_model == 'bilstm_dot_skip':
        modeler = net.BiLstmDot_skip( _parameter_dict = parameter_dict,_word_to_id=_word_to_id,
                                             _device=device,_pointwise=pointwise, _debug=False)

        optimizer = optim.Adam(list(filter(lambda p: p.requires_grad, modeler.encoder.parameters())))

    if training_model == 'bilstm_dot_multiencoder':
        print(schema)
        modeler = net.BiLstmDot_multiencoder( _parameter_dict = parameter_dict,_word_to_id=_word_to_id,
                                             _device=device,_pointwise=pointwise, _debug=False)

        optimizer = optim.Adam(list(filter(lambda p: p.requires_grad, modeler.encoder_p.parameters())) +
                               list(filter(lambda p: p.requires_grad, modeler.encoder_q.parameters())))

    if training_model == 'bilstm_dot_ulmfit':
        modeler = net.BiLstmDot_ulmfit( _parameter_dict = parameter_dict,_word_to_id=_word_to_id,
                                             _device=device,_pointwise=pointwise, _debug=False)

        modeler.freeze_layer(modeler.encoder)
        modeler.unfreeze_layer(modeler.encoder.rnns[-1])
        # +
        # list(filter(lambda p: p.requires_grad, modeler.linear.parameters()))
        optimizer = optim.Adam(list(filter(lambda p: p.requires_grad, modeler.encoder.parameters())),lr=0.001)

    if training_model == 'slotptr_common_encoder':
        print("*************",parameter_dict['bidirectional'])


        if not finetune:
            modeler = net.QelosSlotPointerModel_common_encoder(_parameter_dict=parameter_dict, _word_to_id=_word_to_id,
                                                _device=device, _pointwise=pointwise, _debug=False)
            optimizer = optim.Adam(list(filter(lambda p: p.requires_grad, modeler.encoder_q.parameters())) +
                                   list(filter(lambda p: p.requires_grad, modeler.encoder_p.parameters())), weight_decay=0.0001)
        else:
            print("^^^^^^^^^^^^finetuning^^^^^^^")
            if pointwise:
                model_path = 'data/models/core_chain/slotptr_common_encoder_pointwise/lcquad/1/model.torch'
            else:
                model_path = 'data/models/core_chain/slotptr_common_encoder/lcquad/5/model.torch'
            modeler = net.QelosSlotPointerModel_common_encoder(_parameter_dict=parameter_dict, _word_to_id=_word_to_id,
                                                               _device=device, _pointwise=pointwise, _debug=False)
            optimizer = optim.Adam(list(filter(lambda p: p.requires_grad, modeler.encoder_q.parameters())) +
                                   list(filter(lambda p: p.requires_grad, modeler.encoder_p.parameters())),
                                   weight_decay=0.0001)
            modeler.load_from(model_path)

    if training_model == 'ulmfit_slotptr':
        modeler = net.ULMFITQelosSlotPointerModel( _parameter_dict = parameter_dict,_word_to_id=_word_to_id,
                                             _device=device,_pointwise=pointwise, _debug=False)

        optimizer = optim.Adam(list(filter(lambda p: p.requires_grad, modeler.encoder_q.parameters())) +
                               list(filter(lambda p: p.requires_grad, modeler.encoder_p.parameters())), weight_decay=0.0001, lr=0.01)


    if not pointwise:
        loss_func = nn.MarginRankingLoss(margin=1,size_average=False)
    else:
        loss_func = nn.BCEWithLogitsLoss()
        training_model += '_pointwise'

    for iterations, sample_batched in enumerate(train_loader):
        pass

    parameter_dict['epochs'] = 100
    # lr_args = {'iterations': (iterations + 1) * parameter_dict['epochs']}
    # lr_args = {'iterations': (iterations + 1) * parameter_dict['epochs'],'cycles':2}
    lr_args = {'iterations': (iterations + 1) * parameter_dict['epochs'], 'cut_frac': 0.1, 'ratio': 32}
    # lr_schedule = lriters.LearningRateScheduler(optimizer, lr_args, lriters.CosineAnnealingLR)
    # lr_schedule = lriters.LearningRateScheduler(optimizer, lr_args, lriters.ConstantLR)
    lr_schedule = lriters.LearningRateScheduler(optimizer, lr_args, lriters.SlantedTriangularLR)

    train_loss, modeler, valid_accuracy, test_accuracy,model_save_location = training_loop(training_model = training_model,
                                                                               parameter_dict = parameter_dict,
                                                                               modeler = modeler,
                                                                               train_loader = train_loader,
                                                                               optimizer=optimizer,
                                                                               loss_func=loss_func,
                                                                               data=data,
                                                                               dataset=parameter_dict['dataset'],
                                                                               device=device,
                                                                               test_every=test_every,
                                                                               validate_every=validate_every,
                                                                                pointwise=pointwise,
                                                                                lrschedule = lr_schedule,
                                                                               problem='core_chain')

    print(valid_accuracy)
    print(test_accuracy)
    print("validation accuracy is , ", max(valid_accuracy))
    print("maximum test accuracy is , ", max(test_accuracy))
    print("correct test accuracy i.e test accuracy where validation is highest is ", test_accuracy[valid_accuracy.index(max(valid_accuracy))])
    print("model saved at, " , model_save_location )
    msl = model_save_location.split('/')
    print(f"model save locaton info {msl}")
    json.dump(train_loss,open(f"{model_save_location}/loss.json",'w+'))

    if evals:
        evaluate(device=device,
                 pointwise=pointwise,
                 dataset=msl[-2],
                 training_model=msl[-3],
                 training_model_number=msl[-1])
    # rsync -avz --progress corechain.py qrowdgpu+titan:/shared/home/GauravMaheshwari/new_kranti/KrantikariQA/
    # rsync -avz --progress auxiliary.py qrowdgpu+titan:/shared/home/GauravMaheshwari/new_kranti/KrantikariQA/
    # rsync -avz --progress network.py qrowdgpu+titan:/shared/home/GauravMaheshwari/new_kranti/KrantikariQA/
    # rsync -avz --progress components.py qrowdgpu+titan:/shared/home/GauravMaheshwari/new_kranti/KrantikariQA/
    # rsync -avz --progress data_loader.py qrowdgpu+titan:/shared/home/GauravMaheshwari/new_kranti/KrantikariQA/

     # rsync -avz --progress corechain.py priyansh@sda-srv04:/data/priyansh/new_kranti
     # rsync -avz --progress auxiliary.py priyansh@sda-srv04:/data/priyansh/new_kranti
     # rsync -avz --progress components.py priyansh@sda-srv04:/data/priansh/new_kranti
     # rsync -avz --progress network.py priyansh@sda-srv04:/data/priyansh/new_kranti

    # [0.16, 0.19, 0.24, 0.24, 0.28, 0.3, 0.25, 0.27, 0.29, 0.27, 0.26, 0.28, 0.27, 0.3, 0.32, 0.29, 0.31, 0.3, 0.31,
    #  0.28, 0.31, 0.32, 0.29, 0.31, 0.32, 0.34, 0.32, 0.31, 0.32, 0.33, 0.33, 0.32, 0.33, 0.32, 0.32, 0.33, 0.32, 0.33,
    #  0.32, 0.37, 0.35, 0.34, 0.33, 0.34, 0.34, 0.36, 0.39]
    # [0.2, 0.235, 0.25, 0.29, 0.3, 0.325, 0.28, 0.315, 0.29, 0.325, 0.305, 0.33, 0.35, 0.315, 0.335, 0.315, 0.35, 0.345,
    #  0.325, 0.365, 0.335, 0.35, 0.36, 0.385, 0.37, 0.355, 0.35, 0.375, 0.345, 0.345, 0.36, 0.385, 0.4, 0.375, 0.38,
    #  0.35, 0.395, 0.385, 0.38, 0.39, 0.365, 0.4, 0.395, 0.39, 0.39, 0.4, 0.4]
    # validation
    # accuracy is, 0.39
    # maximum
    # test
    # accuracy is, 0.4
    # correct
    # test
    # accuracy
    # i.e
    # test
    # accuracy
    # where
    # validation is highest is 0.4