examples/NLG/src/gpt2_beam.py

#  ------------------------------------------------------------------------------------------
#  Copyright (c) Microsoft Corporation. All rights reserved.
#  Licensed under the MIT License (MIT). See LICENSE in the repo root for license information.
#  ------------------------------------------------------------------------------------------

# python -m torch.distributed.launch --nproc_per_node=1 src/gpt2_beam.py \
#     --data ./data/e2e/test.jsonl \
#     --batch_size 1 \
#     --seq_len 512 \
#     --eval_len 64 \
#     --model_card gpt2.md \
#     --platform local \
#     --beam 10 \
#     --length_penalty 0.8 \
#     --no_repeat_ngram_size 4 \
#     --repetition_penalty 1.0 \
#     --eos_token_id 628 \
#     --lora_dim 4 \
#     --lora_alpha 32 \
#     --work_dir ./trained_models/GPT2_M/e2e \
#     --output_file predict.26290.jsonl \
#     --init_checkpoint ./trained_models/GPT2_M/e2e/model.26290.pt


import argparse
import time
import math
import os, sys
import json
import itertools
from typing import Callable, Dict, Iterable, List, Optional, Tuple

import torch
from torch import Tensor, device, dtype, nn
from torch.nn import CrossEntropyLoss
from torch.nn import functional as F
from torch.utils.data import DataLoader
import torch.nn.functional as F
torch.set_printoptions(threshold=100000)

import numpy as np

from gpu import (
    add_gpu_params, 
    parse_gpu, 
    distributed_opt, 
    distributed_gather, 
    distributed_sync, 
    cleanup
)

from exp_utils import create_exp_dir

from data_utils import FT_Dataset 
from model import GPT2Config, GPT2LMModel


parser = argparse.ArgumentParser(description='PyTorch GPT2 beam decoding')

add_gpu_params(parser)

parser.add_argument('--data', type=str, default='../data/wikitext-103',
                    help='location of the data corpus')

parser.add_argument('--batch_size', type=int, default=10,
                    help='batch size')

parser.add_argument('--seq_len', type=int, default=512,
                    help='number of tokens to predict')

parser.add_argument('--eval_len', type=int, default=256,
                    help='evaluation length')

parser.add_argument('--min_length', type=int, default=0,
                    help='minimum generation length')

parser.add_argument('--model_card', default='gpt2.sm', choices=['gpt2.sm', 'gpt2.md', 'gpt2.lg'],
                    help='model names')

parser.add_argument('--init_checkpoint', default=None, type=str, help='initial checkpoint')

parser.add_argument('--lora_dim', type=int, default=0, help='lora attn dimension')

parser.add_argument('--lora_alpha', type=int, default=128, help='lora attn alpha')

parser.add_argument('--work_dir', type=str, default=os.getenv('PT_OUTPUT_DIR', 'gpt2_model'), 
                    help='working folder')

parser.add_argument('--beam', type=int, default=1, help='beam search size')

parser.add_argument('--length_penalty', type=float, default=1.0, help='length penalty')

parser.add_argument('--no_repeat_ngram_size', type=int, default=4, help='no_repeat_ngram_size')

parser.add_argument('--repetition_penalty', type=float, default=1.0, help='repetition_penalty')

parser.add_argument('--eos_token_id', action='append', type=int, default=[50256], 
                    help='eos token id')

parser.add_argument('--output_file', type=str, default='beam_prediction.jsonl', 
                    help='output file name')


def print_args(args):
    if args.rank == 0:
        print('=' * 100)
        for k, v in args.__dict__.items():
            print('        - {} : {}'.format(k, v))
        print('=' * 100)


def _reorder_cache(past: Tuple, beam_idx: Tensor) -> Tuple[Tensor]:
    return tuple(layer_past.index_select(1, beam_idx).contiguous().detach() for layer_past in past)


def _calc_banned_ngram_tokens(
    prev_input_ids: Tensor, 
    num_hypos: int, 
    no_repeat_ngram_size: int, 
    cur_len: int
) -> None:
    """Copied from fairseq for no_repeat_ngram in beam_search"""
    if cur_len + 1 < no_repeat_ngram_size:
        # return no banned tokens if we haven't generated no_repeat_ngram_size tokens yet
        return [[] for _ in range(num_hypos)]
    
    generated_ngrams = [{} for _ in range(num_hypos)]
    for idx in range(num_hypos):
        gen_tokens = prev_input_ids[idx].tolist()
        generated_ngram = generated_ngrams[idx]
        for ngram in zip(*[gen_tokens[i:] for i in range(no_repeat_ngram_size)]):
            prev_ngram_tuple = tuple(ngram[:-1])
            generated_ngram[prev_ngram_tuple] = generated_ngram.get(prev_ngram_tuple, []) + [ngram[-1]]

    def _get_generated_ngrams(hypo_idx):
        # Before decoding the next token, prevent decoding of ngrams that have already appeared
        start_idx = cur_len + 1 - no_repeat_ngram_size
        ngram_idx = tuple(prev_input_ids[hypo_idx, start_idx:cur_len].tolist())
        return generated_ngrams[hypo_idx].get(ngram_idx, [])

    banned_tokens = [_get_generated_ngrams(hypo_idx) for hypo_idx in range(num_hypos)]
    return banned_tokens


def _enforce_repetition_penalty_(
    lprobs, 
    batch_size, 
    num_beams, 
    prev_output_tokens, 
    repetition_penalty
):
    """repetition penalty (from CTRL paper https://arxiv.org/abs/1909.05858). """

    for i in range(batch_size * num_beams):
        print('prev_output_tokens.shape', prev_output_tokens.shape)
        print('prev_output_tokens[i].shape', prev_output_tokens[i].shape)

        for previous_token in set(prev_output_tokens[i].tolist()):
            # if score < 0 then repetition penalty has to multiplied to reduce the previous token probability
            if lprobs[i, previous_token] < 0:
                lprobs[i, previous_token] *= repetition_penalty
            else:
                lprobs[i, previous_token] /= repetition_penalty

def _postprocess_next_token_scores(
    scores,
    history,
    cur_len,
    batch_size,
    num_beams,
    repetition_penalty=1.0,                                
    no_repeat_ngram_size=4,
    bad_words_ids=None,
    min_length=0,
    max_length=100,
    eos_token_id=None,
):
    # repetition penalty (from CTRL paper https://arxiv.org/abs/1909.05858)
    if repetition_penalty != 1.0 and history is not None:
        _enforce_repetition_penalty_(scores, batch_size, num_beams, history, repetition_penalty)

    # score: batch_size * beam, vocab
    # set eos token prob to zero if min_length is not reached
    if eos_token_id is not None and cur_len < min_length:
        for eos in eos_token_id:
            scores[:, eos] = -float("inf")

    if no_repeat_ngram_size > 0 and history is not None:
        # calculate a list of banned tokens to prevent repetitively generating the same ngrams
        num_batch_hypotheses = batch_size * num_beams
        # from fairseq: https://github.com/pytorch/fairseq/blob/a07cb6f40480928c9e0548b737aadd36ee66ac76/fairseq/sequence_generator.py#L345
        banned_batch_tokens = _calc_banned_ngram_tokens(
                history, num_batch_hypotheses, no_repeat_ngram_size, cur_len
        )

        for i, banned_tokens in enumerate(banned_batch_tokens):
            scores[i, banned_tokens] = -float("inf")

    return scores


def _add_beam_candidate(
    best_score, 
    best_sequence, 
    batch_size, 
    num_beams, 
    beam_scores, 
    history, 
    eos_token_id=None
):
    last_tokens = history[:, -1]
    for _i in range(batch_size * num_beams):
        if eos_token_id is None or last_tokens[_i] in eos_token_id:
            cur_len = history.shape[-1]
            _score = beam_scores.view(-1)[_i] / cur_len ** args.length_penalty

            batch_id = _i // num_beams

            if not batch_id in best_score or best_score[batch_id] < _score:
                best_score[batch_id] = _score
                best_sequence[batch_id][:cur_len] = history[_i]

            beam_scores.view(-1)[_i] = -float("inf")


def beam(model, data_iter, args):
    model.eval()
    total_loss = 0.
    start_time = time.time()

    all_predictions = {}
    with torch.no_grad():
        for idx, data in enumerate(data_iter):
            data = {key: value for key, value in data.items()}

            _id = data['id'].to(args.device)
            _query = data['query'].to(args.device)
            _query_len = data['query_len'].to(args.device)

            ## local adaptation start.

            ## local adaptation end.


            output = None
            score = None

            batch_size = _id.size(0)
            num_beams = args.beam
            length_penalty = args.length_penalty

            _batch = torch.arange(0, _id.size(0), device=args.device, dtype=torch.long)
            
            past = None
            len_past = None

            _query = _query.repeat(1, num_beams).view(batch_size * num_beams, -1)
            _query_len = _query_len.unsqueeze(-1).repeat(1, num_beams).view(-1)

            _bbatch = _batch.unsqueeze(-1).repeat(1, num_beams).view(-1)
            
            # scores for each sentence in the beam
            beam_scores = torch.zeros(
                (batch_size, num_beams), dtype=torch.float, device=_query.device
            )

            best_sequence = torch.zeros(
                (batch_size, args.eval_len), dtype=torch.long, device=_query.device
            )
            best_score = {}

            history = None
            with torch.no_grad():
                for i in range(0, args.eval_len):
                    if i == 0:
                        logits, past = model(_query) 
                        logits = logits[_bbatch, (_query_len-1).long(), :] # batch_size * beam, vocab
                    else:
                        #print('token_id.shape', token_id.shape, token_id)
                        #print('past.shape', past[0].shape)
                        #print('len_past.shape', len_past.shape, len_past)
                        
                        logits, past = model(token_id, past=past, len_past=len_past) 
                        logits = logits[:, -1, :]    # batch_size * beam, vocab

                    logits = _postprocess_next_token_scores(           
                        logits,
                        history,
                        i,
                        batch_size,
                        num_beams,
                        repetition_penalty=args.repetition_penalty,                                
                        no_repeat_ngram_size=args.no_repeat_ngram_size,
                        min_length=args.min_length,
                        eos_token_id=args.eos_token_id,
                    )

                    softmax_probs = F.softmax(logits, dim=-1)
                    ##_prob, _w_idx = torch.topk(softmax_probs, num_beams) # batch_size, beam

                    vocab_size = softmax_probs.shape[-1] 
                    

                    _logprob = torch.log(softmax_probs) # batch_size * beam, vocab
                    if i == 0:
                        next_scores = _logprob.view(batch_size, num_beams, -1)[:, 0, :] # batch_size, vocab
                        
                    else:
                        next_scores = beam_scores.unsqueeze(-1) + _logprob.view(batch_size, num_beams, -1)
                        next_scores = next_scores.view(batch_size, -1) # batch_size, beam * vocab

                    next_scores, next_tokens = torch.topk(
                        next_scores, num_beams, dim=1, largest=True, sorted=True
                    )     # batch_size, num_beams
                    
                    beam_id = (next_tokens // vocab_size).view(-1)    # batch_size * num_beams
                    token_id = (next_tokens % vocab_size).view(-1).unsqueeze(-1) # batch_size, num_beams

                    beam_idx = beam_id.view(batch_size, num_beams) + (_batch * num_beams).unsqueeze(-1)
                    past = _reorder_cache(past, beam_idx.view(-1))                
                    beam_scores = next_scores # batch_size, num_beams
                    len_past = (_query_len + i).long()

                    if history is None:
                        history = token_id.detach()
                    else:
                        history = torch.cat((history[beam_idx.view(-1)], token_id.detach()), dim=1).detach()

                    _add_beam_candidate(
                        best_score, best_sequence, batch_size, num_beams, beam_scores, history, 
                        eos_token_id=args.eos_token_id
                    )
                
                _add_beam_candidate(
                    best_score, best_sequence, batch_size, num_beams, beam_scores, history
                )


            with torch.no_grad():
                _id = distributed_gather(args, _id)
                output = distributed_gather(args, best_sequence)
                #score = distributed_gather(args, score)
                distributed_sync(args)

            if args.rank == 0:
                _id = _id.view(-1).cpu()
                output = output.view(-1, output.shape[-1]).cpu()
                #score = score.view(-1, score.shape[-1]).cpu()

                for _b in range(0, _id.shape[-1]):
                    _i = int(_id[_b].item())
                    all_predictions[_i] = {}
                    all_predictions[_i]['id'] = _i
                    all_predictions[_i]['predict'] = output[_b].tolist()
                    #all_predictions[_i]['score'] = score[_b].tolist()

                if idx % 10 == 0:
                    print('inference samples', idx)
                    # pred_file = os.path.join(args.work_dir, args.output_file) 
                    # print('saving prediction file', pred_file)
                    # with open(pred_file, 'w') as writer:
                    #     for _i in all_predictions:
                    #         writer.write(json.dumps(all_predictions[_i]) + '\n')

    if args.rank == 0:
        pred_file = os.path.join(args.work_dir, args.output_file) 
        print('saving prediction file', pred_file)
        with open(pred_file, 'w') as writer:
            for _i in all_predictions:
                writer.write(json.dumps(all_predictions[_i]) + '\n')


if __name__ == '__main__':
    args = parser.parse_args()
    parse_gpu(args)
    print_args(args)
    
    if args.rank == 0:
        args.logging = create_exp_dir(args.work_dir)

    valid_data = FT_Dataset(
        args.data, args.batch_size, args.seq_len, args.eval_len, 
    )
    valid_data = valid_data.get_item_list(0, 1000)
    valid_sampler = torch.utils.data.distributed.DistributedSampler(valid_data)
    valid_loader = DataLoader(
        valid_data, batch_size=args.batch_size, num_workers=0, shuffle=False, 
        pin_memory=False, drop_last=False, sampler=valid_sampler
    )

    if args.model_card == 'gpt2.sm':
        config = GPT2Config(
            n_embd=768, n_layer=12, n_head=12, 
            lora_attn_dim=args.lora_dim, lora_attn_alpha=args.lora_alpha,
        )
    elif args.model_card == 'gpt2.md':
        config = GPT2Config(
            n_embd=1024, n_layer=24, n_head=16, 
            lora_attn_dim=args.lora_dim, lora_attn_alpha=args.lora_alpha,
        )
    elif args.model_card == 'gpt2.lg':
        config = GPT2Config(
            n_embd=1280, n_layer=36, n_head=20, 
            lora_attn_dim=args.lora_dim, lora_attn_alpha=args.lora_alpha,
        )

    lm_net = GPT2LMModel(config)
    if args.init_checkpoint is not None:
        print('loading model pretrained weight.')
        cp = torch.load(args.init_checkpoint, map_location=torch.device('cpu'))
        lm_net.load_weight(cp)    
    lm_net = lm_net.cuda()
    print(lm_net.transformer.h[0].mlp)

    print('model sampling ...')
    beam(lm_net, valid_loader, args)
    distributed_sync(args)
    print('cleanup dist ...')
    cleanup(args)