针对对话式 AI 微调 GPT-2 文本预测

Question

我正在尝试 gpt-2 模型的条件文本生成，以将其调整为一个好的聊天机器人。我正在使用nsheppard 的代码在我的自定义数据集上重新训练它。

我在从 Facebook 数据中提取的自定义对话数据集上训练了我的模型。我将样本长度更改为 20，因为它们是交互式条件生成期间的对话。

数据集看起来像这样：

 How are you 
 Hi Great and you 
 Am also good 
 So you re a graphic designer  
 Yeah 
 How can you contribute to making the game In d graphics aspect 
 Can you show me some of your work if u don t mind  
 Am planning to learn making it a motion type    
 U can go through my photos 
 K 
 Can you make animations for it  
 Flash animations to be specific 
 No please only stable ones 
 Ok

但是，训练结束后，当我尝试与它聊天时，它只是完成我的句子而不是回复它们。

User >>> bye
======================================== SAMPLE 1 ========================================
 and  
 hi 
 are there any positions in khrzh being appointed right now 

我知道 Interactive_conditional_samples.py 是为了根据提示完成句子而构建的，但我认为更改数据集会起作用，并且肯定它不起作用。

火车.py

#!/usr/bin/env python3
# Usage:
#  PYTHONPATH=src ./train --dataset <file|directory|glob>

import argparse
import json
import os
import numpy as np
import tensorflow as tf
import time
import tqdm
from tensorflow.core.protobuf import rewriter_config_pb2

import model, sample, encoder
from load_dataset import load_dataset, Sampler
from accumulate import AccumulatingOptimizer
import memory_saving_gradients

CHECKPOINT_DIR = 'checkpoint'
SAMPLE_DIR = 'samples'


parser = argparse.ArgumentParser(
    description='Fine-tune GPT-2 on your custom dataset.',
    formatter_class=argparse.ArgumentDefaultsHelpFormatter)

parser.add_argument('--dataset', metavar='PATH', type=str, required=True, help='Input file, directory, or glob pattern (utf-8 text, or preencoded .npz files).')
parser.add_argument('--model_name', metavar='MODEL', type=str, default='117M', help='Pretrained model name')
parser.add_argument('--combine', metavar='CHARS', type=int, default=50000, help='Concatenate input files with <|endoftext|> separator into chunks of this minimum size')

parser.add_argument('--batch_size', metavar='SIZE', type=int, default=1, help='Batch size')
parser.add_argument('--learning_rate', metavar='LR', type=float, default=0.00002, help='Learning rate for Adam')
parser.add_argument('--accumulate_gradients', metavar='N', type=int, default=1, help='Accumulate gradients across N minibatches.')
parser.add_argument('--memory_saving_gradients', default=False, action='store_true', help='Use gradient checkpointing to reduce vram usage.')
parser.add_argument('--only_train_transformer_layers', default=False, action='store_true', help='Restrict training to the transformer blocks.')
parser.add_argument('--optimizer', type=str, default='adam', help='Optimizer. <adam|sgd>.')
parser.add_argument('--noise', type=float, default=0.0, help='Add noise to input training data to regularize against typos.')

parser.add_argument('--top_k', type=int, default=40, help='K for top-k sampling.')
parser.add_argument('--top_p', type=float, default=0.0, help='P for top-p sampling. Overrides top_k if set > 0.')

parser.add_argument('--restore_from', type=str, default='latest', help='Either "latest", "fresh", or a path to a checkpoint file')
parser.add_argument('--run_name', type=str, default='run1', help='Run id. Name of subdirectory in checkpoint/ and samples/')
parser.add_argument('--sample_every', metavar='N', type=int, default=100, help='Generate samples every N steps')
parser.add_argument('--sample_length', metavar='TOKENS', type=int, default=1023, help='Sample this many tokens')
parser.add_argument('--sample_num', metavar='N', type=int, default=1, help='Generate this many samples')
parser.add_argument('--save_every', metavar='N', type=int, default=1000, help='Write a checkpoint every N steps')

parser.add_argument('--val_dataset', metavar='PATH', type=str, default=None, help='Dataset for validation loss, defaults to --dataset.')
parser.add_argument('--val_batch_size', metavar='SIZE', type=int, default=2, help='Batch size for validation.')
parser.add_argument('--val_batch_count', metavar='N', type=int, default=40, help='Number of batches for validation.')
parser.add_argument('--val_every', metavar='STEPS', type=int, default=0, help='Calculate validation loss every STEPS steps.')


def maketree(path):
    try:
        os.makedirs(path)
    except:
        pass


def randomize(context, hparams, p):
    if p > 0:
        mask = tf.random.uniform(shape=tf.shape(context)) < p
        noise = tf.random.uniform(shape=tf.shape(context), minval=0, maxval=hparams.n_vocab, dtype=tf.int32)
        return tf.where(mask, noise, context)
    else:
        return context


def main():
    args = parser.parse_args()
    enc = encoder.get_encoder(args.model_name)
    hparams = model.default_hparams()
    with open(os.path.join('models', args.model_name, 'hparams.json')) as f:
        hparams.override_from_dict(json.load(f))

    if args.sample_length > hparams.n_ctx:
        raise ValueError(
            "Can't get samples longer than window size: %s" % hparams.n_ctx)

    if args.model_name == '345M':
        args.memory_saving_gradients = True
        if args.optimizer == 'adam':
            args.only_train_transformer_layers = True

    config = tf.ConfigProto()
    config.gpu_options.allow_growth = True
    config.graph_options.rewrite_options.layout_optimizer = rewriter_config_pb2.RewriterConfig.OFF
    with tf.Session(config=config) as sess:
        context = tf.placeholder(tf.int32, [args.batch_size, None])
        context_in = randomize(context, hparams, args.noise)
        output = model.model(hparams=hparams, X=context_in)
        loss = tf.reduce_mean(
            tf.nn.sparse_softmax_cross_entropy_with_logits(
                labels=context[:, 1:], logits=output['logits'][:, :-1]))

        if args.val_every > 0:
            val_context = tf.placeholder(tf.int32, [args.val_batch_size, None])
            val_output = model.model(hparams=hparams, X=val_context)
            val_loss = tf.reduce_mean(
                tf.nn.sparse_softmax_cross_entropy_with_logits(
                    labels=val_context[:, 1:], logits=val_output['logits'][:, :-1]))
            val_loss_summary = tf.summary.scalar('val_loss', val_loss)


        tf_sample = sample.sample_sequence(
            hparams=hparams,
            length=args.sample_length,
            context=context,
            batch_size=args.batch_size,
            temperature=1.0,
            top_k=args.top_k,
            top_p=args.top_p)

        all_vars = [v for v in tf.trainable_variables() if 'model' in v.name]
        train_vars = [v for v in all_vars if '/h' in v.name] if args.only_train_transformer_layers else all_vars

        if args.optimizer == 'adam':
            opt = tf.train.AdamOptimizer(learning_rate=args.learning_rate)
        elif args.optimizer == 'sgd':
            opt = tf.train.GradientDescentOptimizer(learning_rate=args.learning_rate)
        else:
            exit('Bad optimizer:', args.optimizer)

        if args.accumulate_gradients > 1:
            if args.memory_saving_gradients:
                exit("Memory saving gradients are not implemented for gradient accumulation yet.")
            opt = AccumulatingOptimizer(
                opt=opt,
                var_list=train_vars)
            opt_reset = opt.reset()
            opt_compute = opt.compute_gradients(loss)
            opt_apply = opt.apply_gradients()
            summary_loss = tf.summary.scalar('loss', opt_apply)
        else:
            if args.memory_saving_gradients:
                opt_grads = memory_saving_gradients.gradients(loss, train_vars)
            else:
                opt_grads = tf.gradients(loss, train_vars)
            opt_grads = list(zip(opt_grads, train_vars))
            opt_apply = opt.apply_gradients(opt_grads)
            summary_loss = tf.summary.scalar('loss', loss)

        summary_lr = tf.summary.scalar('learning_rate', args.learning_rate)
        summaries = tf.summary.merge([summary_lr, summary_loss])

        summary_log = tf.summary.FileWriter(
            os.path.join(CHECKPOINT_DIR, args.run_name))

        saver = tf.train.Saver(
            var_list=all_vars,
            max_to_keep=5,
            keep_checkpoint_every_n_hours=2)
        sess.run(tf.global_variables_initializer())

        if args.restore_from == 'latest':
            ckpt = tf.train.latest_checkpoint(
                os.path.join(CHECKPOINT_DIR, args.run_name))
            if ckpt is None:
                # Get fresh GPT weights if new run.
                ckpt = tf.train.latest_checkpoint(
                    os.path.join('models', args.model_name))
        elif args.restore_from == 'fresh':
            ckpt = tf.train.latest_checkpoint(
                os.path.join('models', args.model_name))
        else:
            ckpt = tf.train.latest_checkpoint(args.restore_from)
        print('Loading checkpoint', ckpt)
        saver.restore(sess, ckpt)

        print('Loading dataset...')
        chunks = load_dataset(enc, args.dataset, args.combine)
        data_sampler = Sampler(chunks)
        if args.val_every > 0:
            val_chunks = load_dataset(enc, args.val_dataset, args.combine) if args.val_dataset else chunks
        print('dataset has', data_sampler.total_size, 'tokens')
        print('Training...')

        if args.val_every > 0:
            # Sample from validation set once with fixed seed to make
            # it deterministic during training as well as across runs.
            val_data_sampler = Sampler(val_chunks, seed=1)
            val_batches = [[val_data_sampler.sample(1024) for _ in range(args.val_batch_size)]
                           for _ in range(args.val_batch_count)]

        counter = 1
        counter_path = os.path.join(CHECKPOINT_DIR, args.run_name, 'counter')
        if os.path.exists(counter_path):
            # Load the step number if we're resuming a run
            # Add 1 so we don't immediately try to save again
            with open(counter_path, 'r') as fp:
                counter = int(fp.read()) + 1

        def save():
            maketree(os.path.join(CHECKPOINT_DIR, args.run_name))
            print(
                'Saving',
                os.path.join(CHECKPOINT_DIR, args.run_name,
                             'model-{}').format(counter))
            saver.save(
                sess,
                os.path.join(CHECKPOINT_DIR, args.run_name, 'model'),
                global_step=counter)
            with open(counter_path, 'w') as fp:
                fp.write(str(counter) + '\n')

        def generate_samples():
            print('Generating samples...')
            context_tokens = data_sampler.sample(1)
            all_text = []
            index = 0
            while index < args.sample_num:
                out = sess.run(
                    tf_sample,
                    feed_dict={context: args.batch_size * [context_tokens]})
                for i in range(min(args.sample_num - index, args.batch_size)):
                    text = enc.decode(out[i])
                    text = '======== SAMPLE {} ========\n{}\n'.format(
                        index + 1, text)
                    all_text.append(text)
                    index += 1
            print(text)
            maketree(os.path.join(SAMPLE_DIR, args.run_name))
            with open(
                    os.path.join(SAMPLE_DIR, args.run_name,
                                 'samples-{}').format(counter), 'w') as fp:
                fp.write('\n'.join(all_text))

        def validation():
            print('Calculating validation loss...')
            losses = []
            for batch in tqdm.tqdm(val_batches):
                losses.append(sess.run(val_loss, feed_dict={val_context: batch}))
            v_val_loss = np.mean(losses)
            v_summary = sess.run(val_loss_summary, feed_dict={val_loss: v_val_loss})
            summary_log.add_summary(v_summary, counter)
            summary_log.flush()
            print(
                '[{counter} | {time:2.2f}] validation loss = {loss:2.2f}'
                .format(
                    counter=counter,
                    time=time.time() - start_time,
                    loss=v_val_loss))

        def sample_batch():
            return [data_sampler.sample(1024) for _ in range(args.batch_size)]


        avg_loss = (0.0, 0.0)
        start_time = time.time()

        try:
            while True:
                if counter % args.save_every == 0:
                    save()
                if counter % args.sample_every == 0:
                    generate_samples()
                if args.val_every > 0 and (counter % args.val_every == 0 or counter == 1):
                    validation()

                if args.accumulate_gradients > 1:
                    sess.run(opt_reset)
                    for _ in range(args.accumulate_gradients):
                        sess.run(
                            opt_compute, feed_dict={context: sample_batch()})
                    (v_loss, v_summary) = sess.run((opt_apply, summaries))
                else:
                    (_, v_loss, v_summary) = sess.run(
                        (opt_apply, loss, summaries),
                        feed_dict={context: sample_batch()})

                summary_log.add_summary(v_summary, counter)

                avg_loss = (avg_loss[0] * 0.99 + v_loss,
                            avg_loss[1] * 0.99 + 1.0)

                print(
                    '[{counter} | {time:2.2f}] loss={loss:2.2f} avg={avg:2.2f}'
                    .format(
                        counter=counter,
                        time=time.time() - start_time,
                        loss=v_loss,
                        avg=avg_loss[0] / avg_loss[1]))

                counter += 1
        except KeyboardInterrupt:
            print('interrupted')
            save()


if __name__ == '__main__':
    main()

样本.py

import tensorflow as tf

import model

def top_k_logits(logits, k):
    if k == 0:
        # no truncation
        return logits

    def _top_k():
        values, _ = tf.nn.top_k(logits, k=k)
        min_values = values[:, -1, tf.newaxis]
        return tf.where(
            logits < min_values,
            tf.ones_like(logits, dtype=logits.dtype) * -1e10,
            logits,
        )
    return tf.cond(
       tf.equal(k, 0),
       lambda: logits,
       lambda: _top_k(),
    )


def top_p_logits(logits, p):
    with tf.variable_scope('top_p_logits'):
        logits_sort = tf.sort(logits, direction='DESCENDING')
        probs_sort = tf.nn.softmax(logits_sort)
        probs_sums = tf.cumsum(probs_sort, axis=1, exclusive=True)
        logits_masked = tf.where(probs_sums < p, logits_sort, tf.ones_like(logits_sort)*1000) # [batchsize, vocab]
        min_logits = tf.reduce_min(logits_masked, axis=1, keepdims=True) # [batchsize, 1]
        return tf.where(
            logits < min_logits,
            tf.ones_like(logits, dtype=logits.dtype) * -1e10,
            logits,
        )


def sample_sequence(*, hparams, length, start_token=None, batch_size=None, context=None, temperature=1, top_k=0, top_p=0.0):
    if start_token is None:
        assert context is not None, 'Specify exactly one of start_token and context!'
    else:
        assert context is None, 'Specify exactly one of start_token and context!'
        context = tf.fill([batch_size, 1], start_token)

    def step(hparams, tokens, past=None):
        lm_output = model.model(hparams=hparams, X=tokens, past=past, reuse=tf.AUTO_REUSE)

        logits = lm_output['logits'][:, :, :hparams.n_vocab]
        presents = lm_output['present']
        presents.set_shape(model.past_shape(hparams=hparams, batch_size=batch_size))
        return {
            'logits': logits,
            'presents': presents,
        }

    with tf.name_scope('sample_sequence'):
        # Don't feed the last context token -- leave that to the loop below
        # TODO: Would be slightly faster if we called step on the entire context,
        # rather than leaving the last token transformer calculation to the while loop.
        context_output = step(hparams, context[:, :-1])

        def body(past, prev, output):
            next_outputs = step(hparams, prev[:, tf.newaxis], past=past)
            logits = next_outputs['logits'][:, -1, :]  / tf.to_float(temperature)
            if top_p > 0.0:
                logits = top_p_logits(logits, p=top_p)
            else:
                logits = top_k_logits(logits, k=top_k)
            samples = tf.multinomial(logits, num_samples=1, output_dtype=tf.int32)
            return [
                tf.concat([past, next_outputs['presents']], axis=-2),
                tf.squeeze(samples, axis=[1]),
                tf.concat([output, samples], axis=1),
            ]

        def cond(*args):
            return True

        _, _, tokens = tf.while_loop(
            cond=cond, body=body,
            maximum_iterations=length,
            loop_vars=[
                context_output['presents'],
                context[:, -1],
                context,
            ],
            shape_invariants=[
                tf.TensorShape(model.past_shape(hparams=hparams, batch_size=batch_size)),
                tf.TensorShape([batch_size]),
                tf.TensorShape([batch_size, None]),
            ],
            back_prop=False,
        )

        return tokens

Interactive_conditional_samples.py

#!/usr/bin/env python3

import fire
import json
import os
import numpy as np
import tensorflow as tf

import model, sample, encoder

def interact_model(
    model_name='chatbot',
    seed=None,
    nsamples=1,
    batch_size=1,
    length=20,
    temperature=1,
    top_k=0,
    top_p=0.0
):
    """
    Interactively run the model
    :model_name=chatbot : String, which model to use
    :seed=None : Integer seed for random number generators, fix seed to reproduce
     results
    :nsamples=1 : Number of samples to return total
    :batch_size=1 : Number of batches (only affects speed/memory).  Must divide nsamples.
    :length=None : Number of tokens in generated text, if None (default), is
     determined by model hyperparameters
    :temperature=1 : Float value controlling randomness in boltzmann
     distribution. Lower temperature results in less random completions. As the
     temperature approaches zero, the model will become deterministic and
     repetitive. Higher temperature results in more random completions.
    :top_k=0 : Integer value controlling diversity. 1 means only 1 word is
     considered for each step (token), resulting in deterministic completions,
     while 40 means 40 words are considered at each step. 0 (default) is a
     special setting meaning no restrictions. 40 generally is a good value.
    :top_p=0.0 : Float value controlling diversity. Implements nucleus sampling,
     overriding top_k if set to a value > 0. A good setting is 0.9.
    """
    if batch_size is None:
        batch_size = 1
    assert nsamples % batch_size == 0

    enc = encoder.get_encoder(model_name)
    hparams = model.default_hparams()
    with open(os.path.join('models', model_name, 'hparams.json')) as f:
        hparams.override_from_dict(json.load(f))

    if length is None:
        length = hparams.n_ctx // 2
    elif length > hparams.n_ctx:
        raise ValueError("Can't get samples longer than window size: %s" % hparams.n_ctx)

    with tf.Session(graph=tf.Graph()) as sess:
        context = tf.placeholder(tf.int32, [batch_size, None])
        np.random.seed(seed)
        tf.set_random_seed(seed)
        output = sample.sample_sequence(
            hparams=hparams, length=length,
            context=context,
            batch_size=batch_size,
            temperature=temperature, top_k=top_k, top_p=top_p
        )

        s

    

Answer 1

我知道这是一个老问题了，但我已经成功地在 GPT-2 上调整了许多问答风格的数据集，并提出了一个对未来发现这个问题的人有用的建议。

GPT-2 读取非结构化文本数据，但它非常擅长推断和遵守该数据中的结构。您的问题基本上是您没有使用 GPT-2 理解的标识符来终止输入行，因此它会继续该句子。

解决此问题的一个简单方法是注释您的数据集。实际上，任何带有停止/启动标记的东西都可以工作，但您还应该注释说话者的身份。我只会做这样的事情：

A: How are you <EOL>
B: Hi Great and you <EOL>
A: Am also good <EOL>
B: So you re a graphic designer <EOL>
B: Another line from B <EOL>

这种方法的另一个好处是 GPT-2 将学习多行输入/输出，以及两个熟悉者的不同身份。