Source code for examples.attn.mltr30k

# Copyright 2020 Reid Swanson
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
"""
Documentation
"""
# Python Modules
import argparse
import logging
import os
import shutil

# from typing import Dict, Any, Optional

# 3rd Party Modules
import numpy as np
import tensorflow as tf
import deletor.tfutils as tfutils
from examples.utils import train, evaluate, make_optimizer, make_loss
from deletor.random.sample import IndependentMultiOutputSampler

tfutils.grow_memory()

# Project Modules
from examples.pipeline import load_dataset, is_valid_query, truncate_document_list, \
    make_padded_shapes, make_padding_values, apply_map, N_FEATURES
from deletor.models.attn import ModelParameter, \
    MultiHeadAttention, GroupwiseMultiHeadAttentionNetwork
from deletor.metrics import NormalizedDiscountedCumulativeGain


np.set_printoptions(precision=6, suppress=True, edgeitems=10, linewidth=10000)
logging.basicConfig(
    level=logging.DEBUG,
    format='%(asctime)-15s [%(name)s]:%(lineno)d %(levelname)s %(message)s'
)
log = logging.getLogger('gsf/mltr30k')

AUTOTUNE = tf.data.experimental.AUTOTUNE


[docs]class AttentionSchedule(tf.keras.optimizers.schedules.LearningRateSchedule): def __init__(self, n_features: int, warmup_steps: int = 4000): super().__init__() self.n_features = tf.cast(n_features, tf.float32) self.warmup_steps = warmup_steps def __call__(self, step): arg_1 = tf.math.rsqrt(step) arg_2 = step * (self.warmup_steps ** -1.5) return tf.math.rsqrt(self.n_features) * tf.math.minimum(arg_1, arg_2)
[docs] def get_config(self): pass
[docs]class DampedSineSchedule(tf.keras.optimizers.schedules.LearningRateSchedule): def __init__( self, base_learning_rate: float = 0.0005, min_learning_rate: float = 0.00001, amplitude: float = 50.0, amplitude_decay: float = 80000, angular_frequency: float = 1.0 / 45000.0, phase: float = -np.pi / 6.0, learning_rate_decay: float = 550000 ): super().__init__() self.base_learning_rate = base_learning_rate self.min_learning_rate = min_learning_rate self.amplitude = amplitude self.amplitude_decay = amplitude_decay self.angular_frequency = angular_frequency self.phase = phase self.learning_rate_decay = learning_rate_decay def __call__(self, step): step = tf.cast(step, tf.float32) amp = self.amplitude / (self.amplitude_decay + step) sin_value = tf.math.sin(self.angular_frequency * step + self.phase) lr_offset = self.base_learning_rate * tf.exp(-step / self.learning_rate_decay) lr = amp * sin_value + lr_offset return tf.maximum(lr, self.min_learning_rate)
[docs] def get_config(self): pass
# noinspection DuplicatedCode
[docs]def prepare_data(args: argparse.Namespace): list_size = args.list_size group_size = args.group_size multiples = args.multiples n_heads = args.n_heads train_bsz = args.training_batch_size eval_bsz = args.evaluation_batch_size drop_remainder = args.drop_remainder train_data = load_dataset(args.train_file, args.scaler) valid_data = load_dataset(args.valid_file, args.scaler) test_data = load_dataset(args.test_file, args.scaler) train_data = train_data.filter(is_valid_query) valid_data = valid_data.filter(is_valid_query) test_data = test_data.filter(is_valid_query) if list_size: train_data = train_data.map(lambda x, y: truncate_document_list(x, y, list_size)) valid_data = valid_data.map(lambda x, y: truncate_document_list(x, y, list_size)) test_data = test_data.map(lambda x, y: truncate_document_list(x, y, list_size)) train_data = train_data.cache() valid_data = valid_data.cache() test_data = test_data.cache() train_data = train_data.shuffle(1000, args.random_seed, reshuffle_each_iteration=True) # The code for bucketing by sequence length should be in branch v0.2 # It doesn't really seem to improve the accuracy and any efficiency gains # appear to be small. The primary reason for not using it is that it # seems to break when using an unbounded list_size (i.e., None). It's # possible this could be fixed, but given the lack of accuracy/efficiency # improvements it's probably not worth it. padded_shapes = make_padded_shapes(list_size) padding_values = make_padding_values() train_data = train_data.padded_batch(train_bsz, padded_shapes, padding_values, drop_remainder) valid_data = valid_data.padded_batch(eval_bsz, padded_shapes, padding_values, drop_remainder) test_data = test_data.padded_batch(eval_bsz, padded_shapes, padding_values, drop_remainder) sampler = IndependentMultiOutputSampler(group_size, multiple=multiples) make_attention_mask = MultiHeadAttention.make_attention_mask data = [train_data, valid_data, test_data] data = apply_map(sampler, *data) # data = apply_map(pad_groups, *data, **{'group_size': group_size}) data = apply_map(lambda x, y: make_attention_mask(x, y, multi_head=n_heads > 1), *data) data = [d.prefetch(AUTOTUNE) for d in data] return data[0], data[1], data[2]
[docs]def setup_model(args: argparse.Namespace): model_params = { ModelParameter.N_LAYERS: args.n_layers, ModelParameter.N_FEATURES: N_FEATURES, ModelParameter.N_MODEL: args.n_model, ModelParameter.GROUP_SIZE: args.group_size, ModelParameter.N_HEADS: args.n_heads, ModelParameter.N_FF_UNITS: args.n_feed_forward_units, ModelParameter.USE_AVERAGE: args.use_average, ModelParameter.SHARE_WEIGHTS: args.share_weights, ModelParameter.USE_LAYER_NORM: args.use_layer_norm, ModelParameter.DROPOUT_RATE: args.dropout_rate } # model = GroupwiseAttentionNetwork(model_params) model = GroupwiseMultiHeadAttentionNetwork(model_params) optimizer = make_optimizer(args) loss = make_loss(args) metrics = [ NormalizedDiscountedCumulativeGain(k=1), NormalizedDiscountedCumulativeGain(k=5), NormalizedDiscountedCumulativeGain(k=10), ] model.compile(optimizer=optimizer, loss=loss, metrics=metrics) return model
# noinspection PyTypeChecker
[docs]def main(args: argparse.Namespace): max_epochs = args.max_epochs datasets = prepare_data(args) train_data, valid_data, test_data = datasets if os.path.exists(args.checkpoint_dir): log.info(f"Removing existing checkpoint directory: {args.checkpoint_dir}") shutil.rmtree(args.checkpoint_dir, ignore_errors=True) tf.config.experimental_run_functions_eagerly(args.run_eagerly) model = setup_model(args) model(tf.data.experimental.get_single_element(train_data.take(1))[0], training=True) model.summary(print_fn=log.info) # Custom loop train_meta = { 'sample_pre_batch': args.sample_pre_batch, 'max_epochs': tf.constant(max_epochs), 'step': tf.Variable(0), 'elapsed_time': tf.Variable(0., tf.float32), 'train_time': tf.Variable(0., tf.float32), 'valid_time': tf.Variable(0., tf.float32), 'secs_step': tf.Variable(0., tf.float32), 'train_loss': tf.keras.metrics.Mean(), 'metrics': {k: tf.keras.metrics.Mean() for k in (1, 5, 10)}, 'best_result': tf.Variable(0., tf.float32) } ckpt = tf.train.Checkpoint(epoch=tf.Variable(0), optimizer=model.optimizer, model=model) manager = tf.train.CheckpointManager(ckpt, args.checkpoint_dir, max_to_keep=1) start_time = tf.timestamp() for epoch in range(train_meta['max_epochs']): train_meta['train_loss'].reset_states() train(model, train_data, train_meta) evaluate(model, valid_data, train_meta) flag_best_result = '' if train_meta['metrics'][5].result() > train_meta['best_result']: manager.save() train_meta['best_result'].assign(train_meta['metrics'][5].result()) flag_best_result = ' *' train_meta['elapsed_time'].assign(tf.cast(tf.timestamp() - start_time, tf.float32)) ckpt.epoch.assign(epoch) log.info( f"epoch: {epoch+1:5d} " f"step: {train_meta['step'].numpy():8d} " f"elapsed time: {train_meta['elapsed_time'].numpy():8.2f}s " f"train time: {train_meta['train_time'].numpy():6.2f}s " f"secs/step: {train_meta['secs_step'].numpy():6.3f} " f"val time: {train_meta['valid_time'].numpy():6.2f} " f"train/loss: {train_meta['train_loss'].result():10.4f} " f"val/ndcg@01: {train_meta['metrics'][1].result():10.4f} " f"val/ndcg@05: {train_meta['metrics'][5].result():10.4f} " f"val/ndcg@10: {train_meta['metrics'][10].result():10.4f}" f"{flag_best_result}" ) # Evaluate on the test data using the best model during training ckpt = tf.train.Checkpoint(epoch=tf.Variable(0), model=model, optimizer=model.optimizer) ckpt.restore(tf.train.latest_checkpoint(args.checkpoint_dir)) evaluate(model, test_data, train_meta) log.info( f"test/ndcg@01: {train_meta['metrics'][1].result():10.4f} " f"test/ndcg@05: {train_meta['metrics'][5].result():10.4f} " f"test/ndcg@10: {train_meta['metrics'][10].result():10.4f}" )
# noinspection DuplicatedCode
[docs]def make_command_line_options(): cli = argparse.ArgumentParser() cli.add_argument( '--train-file', required=True, type=str, help="The training tfrecords file." ) cli.add_argument( '--valid-file', required=True, type=str, help="The validation tfrecords file." ) cli.add_argument( '--test-file', required=True, type=str, help="The test tfrecords file." ) cli.add_argument( '--checkpoint-dir', required=True, type=str, help="The directory where model checkpoints will be saved." ) cli.add_argument( '--run-eagerly', required=False, action='store_true' ) cli.add_argument( '--scaler', required=False, type=str, nargs=2, help=( "This argument requires two parameters. The first is the path to " "a scaler file created with the build dataset script. The second " "is the name of the scaler to use. Choose one of: " "minmax, standard, robust, power." ) ) cli.add_argument( '--max-epochs', required=False, type=int, default=500, help="The maximum number of epochs before the training terminates no matter what." ) cli.add_argument( '--optimizer', required=False, type=str, default='adagrad', choices=['adagrad', 'adam', 'sgd', 'nesterov', 'rmsprop'] ) cli.add_argument( '--learning-rate', required=False, type=float, default=0.001 ) cli.add_argument( '--loss', required=False, type=str, choices=['ndcg', 'softmax', 'cross_entropy', 'mse'], default='ndcg' ) cli.add_argument( '--list-size', required=False, type=int, default=None, help="The maximum number of documents per query or no maximum if not set." ) cli.add_argument( '--group-size', required=False, type=int, default=16, help="The group size to use." ) cli.add_argument( '--sample-pre-batch', required=False, action='store_true', default=False, help=( "If this flag is set then the alternate form of training will be " "performed where documents are sampled before training." ) ) cli.add_argument( '--multiples', required=False, type=int, default=1, help="The sampling multiplier." ) cli.add_argument( '--training-batch-size', required=False, type=int, default=128 ) cli.add_argument( '--evaluation-batch-size', required=False, type=int, default=128 ) cli.add_argument( '--use-average', required=False, action='store_true', default=False, help=( "According to the paper, when a document is sampled more than once its scores are " "summed. When this option is set the scores are averaged over the number of times " "each document is seen instead." ) ) cli.add_argument( '--share-weights', required=False, action='store_true', default=False, help="Apply each document through a shared dense layer before concatenating them." ) cli.add_argument( '--n-model', required=False, type=int, default=128 ) cli.add_argument( '--use-layer-norm', required=False, action='store_true' ) cli.add_argument( '--n-layers', required=False, type=int, default=4 ) cli.add_argument( '--n-heads', required=False, type=int, default=8 ) cli.add_argument( '--n-feed-forward-units', required=False, type=int, default=128 ) cli.add_argument( '--dropout-rate', required=False, type=float, default=0.0 ) cli.add_argument( '--drop-remainder', action='store_true', default=False, help="This is necessary when using the keras training/eval loops." ) cli.add_argument( '--random-seed', required=False, type=int, help="The random seed to use for sampling query results." ) cli.set_defaults(func=main) return cli
if __name__ == '__main__': clo = make_command_line_options() cli_args = clo.parse_args() cli_args.func(cli_args)