main_celer.py

import numpy as np
import pandas as pd
import os
from sklearn.model_selection import StratifiedKFold, KFold
import torch
import torch.nn as nn
from torch.utils.data import DataLoader
from torch.optim import Adam, RMSprop
from transformers import BertTokenizerFast
from sklearn.preprocessing import LabelEncoder, OneHotEncoder
from torch.nn.functional import cross_entropy, softmax
from collections import deque
import pickle
import json
import matplotlib.pyplot as plt
import argparse

from Eyettention.utils import *
from Eyettention.model import Eyettention

if __name__ == "__main__":
    parser = argparse.ArgumentParser(description="run Eyettention on CELER dataset")
    parser.add_argument(
        "--test_mode",
        help="New Sentence Split: text, New Reader Split: subject",
        type=str,
        default="text",
    )
    parser.add_argument(
        "--atten_type", help="attention type: global, local, local-g", type=str, default="local-g"
    )
    parser.add_argument(
        "--save_data_folder",
        help="folder path for saving results",
        type=str,
        default="./results/CELER/",
    )
    parser.add_argument(
        "--scanpath_gen_flag", help="whether to generate scanpath", type=int, default=1
    )
    parser.add_argument(
        "--max_pred_len",
        help="if scanpath_gen_flag is True, you can determine the longest scanpath that you want to generate, which should depend on the sentence length",
        type=int,
        default=60,
    )
    parser.add_argument("--gpu", help="gpu index", type=int, default=6)
    args = parser.parse_args()
    gpu = args.gpu

    # use FastTokenizer lead to warning -> The current process just got forked
    os.environ["TOKENIZERS_PARALLELISM"] = "false"
    torch.set_default_tensor_type("torch.FloatTensor")
    availbl = torch.cuda.is_available()
    print(torch.cuda.is_available())
    if availbl:
        device = f"cuda:{gpu}"
    else:
        device = "cpu"
    print(device)
    torch.cuda.set_device(gpu)

    cf = {
        "model_pretrained": "bert-base-cased",
        "lr": 1e-3,
        "max_grad_norm": 10,
        "n_epochs": 1000,
        "n_folds": 5,
        "dataset": "celer",
        "atten_type": args.atten_type,
        "batch_size": 256,
        "max_sn_len": 24,  # max number of words in a sentence, include start token and end token,
        "max_sn_token": 35,  # maximum number of tokens a sentence includes. include start token and end token,
        "max_sp_len": 52,  # max number of words in a scanpath, include start token and end token
        "max_sp_token": 395,  # maximum number of tokens a scanpath includes. include start token and end token
        "norm_type": "z-score",
        "earlystop_patience": 20,
        "max_pred_len": args.max_pred_len,
    }

    # Encode the label into interger categories, setting the exclusive category 'cf["max_sn_len"]-1' as the end sign
    le = LabelEncoder()
    le.fit(np.append(np.arange(-cf["max_sn_len"] + 3, cf["max_sn_len"] - 1), cf["max_sn_len"] - 1))
    # le.classes_

    # load corpus
    word_info_df, _, eyemovement_df = load_corpus(cf["dataset"])

    # only use native speaker
    # Make list with reader index
    reader_list = celer_load_native_speaker()

    # Make list with sentence index
    sn_list = np.unique(
        word_info_df[word_info_df["list"].isin(reader_list)].sentenceid.values
    ).tolist()

    # Split training&test sets by text or reader, depending on configuration
    if args.test_mode == "text":
        print("Start evaluating on new sentences.")
        split_list = sn_list
    elif args.test_mode == "subject":
        print("Start evaluating on new readers.")
        split_list = reader_list

    n_folds = cf["n_folds"]
    kf = KFold(n_splits=n_folds, shuffle=True, random_state=0)
    fold_indx = 0
    # for scanpath generation
    sp_dnn_list = []
    sp_human_list = []
    for train_idx, test_idx in kf.split(split_list):
        loss_dict = {"val_loss": [], "train_loss": [], "test_ll": [], "test_AUC": []}
        list_train = [split_list[i] for i in train_idx]
        list_test = [split_list[i] for i in test_idx]

        # create train validation split for training the models:
        kf_val = KFold(n_splits=n_folds, shuffle=True, random_state=0)
        for train_index, val_index in kf_val.split(list_train):
            # we only evaluate a single fold
            break
        list_train_net = [list_train[i] for i in train_index]
        list_val_net = [list_train[i] for i in val_index]

        if args.test_mode == "text":
            sn_list_train = list_train_net
            sn_list_val = list_val_net
            sn_list_test = list_test
            reader_list_train, reader_list_val, reader_list_test = (
                reader_list,
                reader_list,
                reader_list,
            )

        elif args.test_mode == "subject":
            reader_list_train = list_train_net
            reader_list_val = list_val_net
            reader_list_test = list_test
            sn_list_train, sn_list_val, sn_list_test = sn_list, sn_list, sn_list

        # initialize tokenizer
        tokenizer = BertTokenizerFast.from_pretrained(cf["model_pretrained"])
        # Preparing batch data
        dataset_train = celerdataset(
            word_info_df, eyemovement_df, cf, reader_list_train, sn_list_train, tokenizer
        )
        train_dataloaderr = DataLoader(
            dataset_train, batch_size=cf["batch_size"], shuffle=True, drop_last=True
        )

        dataset_val = celerdataset(
            word_info_df, eyemovement_df, cf, reader_list_val, sn_list_val, tokenizer
        )
        val_dataloaderr = DataLoader(
            dataset_val, batch_size=cf["batch_size"], shuffle=False, drop_last=True
        )

        dataset_test = celerdataset(
            word_info_df, eyemovement_df, cf, reader_list_test, sn_list_test, tokenizer
        )
        test_dataloaderr = DataLoader(
            dataset_test, batch_size=cf["batch_size"], shuffle=False, drop_last=False
        )

        # z-score normalization for gaze features
        fix_dur_mean, fix_dur_std = calculate_mean_std(
            dataloader=train_dataloaderr, feat_key="sp_fix_dur", padding_value=0, scale=1000
        )
        landing_pos_mean, landing_pos_std = calculate_mean_std(
            dataloader=train_dataloaderr, feat_key="sp_landing_pos", padding_value=0
        )
        sn_word_len_mean, sn_word_len_std = calculate_mean_std(
            dataloader=train_dataloaderr, feat_key="sn_word_len"
        )

        # load model
        dnn = Eyettention(cf)

        # training
        episode = 0
        optimizer = Adam(dnn.parameters(), lr=cf["lr"])
        dnn.train()
        dnn.to(device)
        av_score = deque(maxlen=100)
        old_score = 1e10
        save_ep_couter = 0
        print("Start training")
        for episode_i in range(episode, cf["n_epochs"] + 1):
            dnn.train()
            print("episode:", episode_i)
            counter = 0
            for batchh in train_dataloaderr:
                counter += 1
                batchh.keys()
                sn_input_ids = batchh["sn_input_ids"].to(device)
                sn_attention_mask = batchh["sn_attention_mask"].to(device)
                word_ids_sn = batchh["word_ids_sn"].to(device)
                sn_word_len = batchh["sn_word_len"].to(device)

                sp_input_ids = batchh["sp_input_ids"].to(device)
                sp_attention_mask = batchh["sp_attention_mask"].to(device)
                word_ids_sp = batchh["word_ids_sp"].to(device)

                sp_pos = batchh["sp_pos"].to(device)
                sp_landing_pos = batchh["sp_landing_pos"].to(device)
                sp_fix_dur = (batchh["sp_fix_dur"] / 1000).to(device)

                # normalize gaze features
                mask = ~torch.eq(sp_fix_dur, 0)
                sp_fix_dur = (sp_fix_dur - fix_dur_mean) / fix_dur_std * mask
                sp_landing_pos = (sp_landing_pos - landing_pos_mean) / landing_pos_std * mask
                sp_fix_dur = torch.nan_to_num(sp_fix_dur)
                sp_landing_pos = torch.nan_to_num(sp_landing_pos)
                sn_word_len = (sn_word_len - sn_word_len_mean) / sn_word_len_std
                sn_word_len = torch.nan_to_num(sn_word_len)

                # zero old gradients
                optimizer.zero_grad()
                # predict output with DNN
                dnn_out, atten_weights = dnn(
                    sn_emd=sn_input_ids,
                    sn_mask=sn_attention_mask,
                    sp_emd=sp_input_ids,
                    sp_pos=sp_pos,
                    word_ids_sn=word_ids_sn,
                    word_ids_sp=word_ids_sp,
                    sp_fix_dur=sp_fix_dur,
                    sp_landing_pos=sp_landing_pos,
                    sn_word_len=sn_word_len,
                )

                dnn_out = dnn_out.permute(0, 2, 1)  # [batch, dec_o_dim, step]

                # prepare label and mask
                pad_mask, label = load_label(sp_pos, cf, le, device)
                loss = nn.CrossEntropyLoss(reduction="none")
                batch_error = torch.mean(torch.masked_select(loss(dnn_out, label), ~pad_mask))

                # backpropagate loss
                batch_error.backward()
                # clip gradients
                gradient_clipping(dnn, cf["max_grad_norm"])

                # learn
                optimizer.step()
                av_score.append(batch_error.to("cpu").detach().numpy())
                print("counter:", counter)
                print(
                    "\rSample {}\tAverage Error: {:.10f} ".format(counter, np.mean(av_score)),
                    end=" ",
                )
            loss_dict["train_loss"].append(np.mean(av_score))

            val_loss = []
            dnn.eval()
            for batchh in val_dataloaderr:
                with torch.no_grad():
                    sn_input_ids_val = batchh["sn_input_ids"].to(device)
                    sn_attention_mask_val = batchh["sn_attention_mask"].to(device)
                    word_ids_sn_val = batchh["word_ids_sn"].to(device)
                    sn_word_len_val = batchh["sn_word_len"].to(device)

                    sp_input_ids_val = batchh["sp_input_ids"].to(device)
                    sp_attention_mask_val = batchh["sp_attention_mask"].to(device)
                    word_ids_sp_val = batchh["word_ids_sp"].to(device)

                    sp_pos_val = batchh["sp_pos"].to(device)
                    sp_landing_pos_val = batchh["sp_landing_pos"].to(device)
                    sp_fix_dur_val = (batchh["sp_fix_dur"] / 1000).to(device)

                    # normalize gaze features
                    mask = ~torch.eq(sp_fix_dur_val, 0)
                    sp_fix_dur_val = (sp_fix_dur_val - fix_dur_mean) / fix_dur_std * mask
                    sp_landing_pos_val = (
                        (sp_landing_pos_val - landing_pos_mean) / landing_pos_std * mask
                    )
                    sp_fix_dur_val = torch.nan_to_num(sp_fix_dur_val)
                    sp_landing_pos_val = torch.nan_to_num(sp_landing_pos_val)
                    sn_word_len_val = (sn_word_len_val - sn_word_len_mean) / sn_word_len_std
                    sn_word_len_val = torch.nan_to_num(sn_word_len_val)

                    dnn_out_val, atten_weights_val = dnn(
                        sn_emd=sn_input_ids_val,
                        sn_mask=sn_attention_mask_val,
                        sp_emd=sp_input_ids_val,
                        sp_pos=sp_pos_val,
                        word_ids_sn=word_ids_sn_val,
                        word_ids_sp=word_ids_sp_val,
                        sp_fix_dur=sp_fix_dur_val,
                        sp_landing_pos=sp_landing_pos_val,
                        sn_word_len=sn_word_len_val,
                    )
                    dnn_out_val = dnn_out_val.permute(0, 2, 1)  # [batch, dec_o_dim, step

                    # prepare label and mask
                    pad_mask_val, label_val = load_label(sp_pos_val, cf, le, device)
                    batch_error_val = torch.mean(
                        torch.masked_select(loss(dnn_out_val, label_val), ~pad_mask_val)
                    )
                    val_loss.append(batch_error_val.detach().to("cpu").numpy())
            print("\nvalidation loss is {} \n".format(np.mean(val_loss)))
            loss_dict["val_loss"].append(np.mean(val_loss))

            if np.mean(val_loss) < old_score:
                # save model if val loss is smallest
                torch.save(
                    dnn.state_dict(),
                    "{}/CELoss_CELER_{}_eyettention_{}_newloss_fold{}.pth".format(
                        args.save_data_folder, args.test_mode, args.atten_type, fold_indx
                    ),
                )
                old_score = np.mean(val_loss)
                print("\nsaved model state dict\n")
                save_ep_couter = episode_i
            else:
                # early stopping
                if episode_i - save_ep_couter >= cf["earlystop_patience"]:
                    break

        # evaluation
        dnn.eval()
        res_llh = []
        dnn.load_state_dict(
            torch.load(
                os.path.join(
                    args.save_data_folder,
                    f"CELoss_CELER_{args.test_mode}_eyettention_{args.atten_type}_newloss_fold{fold_indx}.pth",
                ),
                map_location="cpu",
            )
        )
        dnn.to(device)
        batch_indx = 0
        for batchh in test_dataloaderr:
            with torch.no_grad():
                sn_input_ids_test = batchh["sn_input_ids"].to(device)
                sn_attention_mask_test = batchh["sn_attention_mask"].to(device)
                word_ids_sn_test = batchh["word_ids_sn"].to(device)
                sn_word_len_test = batchh["sn_word_len"].to(device)

                sp_input_ids_test = batchh["sp_input_ids"].to(device)
                sp_attention_mask_test = batchh["sp_attention_mask"].to(device)
                word_ids_sp_test = batchh["word_ids_sp"].to(device)

                sp_pos_test = batchh["sp_pos"].to(device)
                sp_landing_pos_test = batchh["sp_landing_pos"].to(device)
                sp_fix_dur_test = (batchh["sp_fix_dur"] / 1000).to(device)

                # normalize gaze features
                mask = ~torch.eq(sp_fix_dur_test, 0)
                sp_fix_dur_test = (sp_fix_dur_test - fix_dur_mean) / fix_dur_std * mask
                sp_landing_pos_test = (
                    (sp_landing_pos_test - landing_pos_mean) / landing_pos_std * mask
                )
                sp_fix_dur_test = torch.nan_to_num(sp_fix_dur_test)
                sp_landing_pos_test = torch.nan_to_num(sp_landing_pos_test)
                sn_word_len_test = (sn_word_len_test - sn_word_len_mean) / sn_word_len_std
                sn_word_len_test = torch.nan_to_num(sn_word_len_test)

                dnn_out_test, atten_weights_test = dnn(
                    sn_emd=sn_input_ids_test,
                    sn_mask=sn_attention_mask_test,
                    sp_emd=sp_input_ids_test,
                    sp_pos=sp_pos_test,
                    word_ids_sn=word_ids_sn_test,
                    word_ids_sp=word_ids_sp_test,
                    sp_fix_dur=sp_fix_dur_test,
                    sp_landing_pos=sp_landing_pos_test,
                    sn_word_len=sn_word_len_test,
                )

                # We do not use nn.CrossEntropyLoss here to calculate the likelihood because it combines nn.LogSoftmax and nn.NLL,
                # while nn.LogSoftmax returns a log value based on e, we want 2 instead
                # m = nn.LogSoftmax(dim=2) -- base e, we want base 2
                m = nn.Softmax(dim=2)
                dnn_out_test = m(dnn_out_test).detach().to("cpu").numpy()

                # prepare label and mask
                pad_mask_test, label_test = load_label(sp_pos_test, cf, le, "cpu")
                pred = dnn_out_test.argmax(axis=2)
                # compute log likelihood for the batch samples
                res_batch = eval_log_llh(dnn_out_test, label_test, pad_mask_test)
                res_llh.append(np.array(res_batch))

                if bool(args.scanpath_gen_flag) == True:
                    sn_len = (
                        (torch.max(torch.nan_to_num(word_ids_sn_test), dim=1)[0] + 1 - 2)
                        .detach()
                        .to("cpu")
                        .numpy()
                    )
                    # compute the scan path generated from the model when the first few fixed points are given
                    sp_dnn = dnn.scanpath_generation(
                        sn_emd=sn_input_ids_test,
                        sn_mask=sn_attention_mask_test,
                        word_ids_sn=word_ids_sn_test,
                        sn_word_len=sn_word_len_test,
                        le=le,
                        max_pred_len=cf["max_pred_len"],
                    )

                    sp_dnn, sp_human = prepare_scanpath(
                        sp_dnn.detach().to("cpu").numpy(), sn_len, sp_pos_test, cf
                    )
                    sp_dnn_list.extend(sp_dnn)
                    sp_human_list.extend(sp_human)

                batch_indx += 1

        res_llh = np.concatenate(res_llh).ravel()
        loss_dict["test_ll"].append(res_llh)
        loss_dict["fix_dur_mean"] = fix_dur_mean
        loss_dict["fix_dur_std"] = fix_dur_std
        loss_dict["landing_pos_mean"] = landing_pos_mean
        loss_dict["landing_pos_std"] = landing_pos_std
        loss_dict["sn_word_len_mean"] = sn_word_len_mean
        loss_dict["sn_word_len_std"] = sn_word_len_std
        print("\nTest likelihood is {} \n".format(np.mean(res_llh)))
        # save results
        with open(
            "{}/res_CELER_{}_eyettention_{}_Fold{}.pickle".format(
                args.save_data_folder, args.test_mode, args.atten_type, fold_indx
            ),
            "wb",
        ) as handle:
            pickle.dump(loss_dict, handle, protocol=pickle.HIGHEST_PROTOCOL)
        fold_indx += 1

    if bool(args.scanpath_gen_flag) == True:
        # save results
        dic = {"sp_dnn": sp_dnn_list, "sp_human": sp_human_list}
        with open(
            os.path.join(
                args.save_data_folder,
                f"CELER_scanpath_generation_eyettention_{args.test_mode}_{args.atten_type}.pickle",
            ),
            "wb",
        ) as handle:
            pickle.dump(dic, handle, protocol=pickle.HIGHEST_PROTOCOL)