train_speaker.py

import torch
from torch import optim

import os
import os.path
import time
import numpy as np
import pandas as pd
from collections import defaultdict
import argparse

import utils
from utils import read_vocab, Tokenizer, timeSince, try_cuda
from env import R2RBatch, ImageFeatures
from model import SpeakerEncoderLSTM, SpeakerDecoderLSTM
from speaker import Seq2SeqSpeaker
import eval_speaker

from vocab import SUBTRAIN_VOCAB, TRAIN_VOCAB, TRAINVAL_VOCAB

MAX_INSTRUCTION_LENGTH = 80

batch_size = 100
max_episode_len = 10
word_embedding_size = 300
glove_path = 'tasks/R2R/data/train_glove.npy'
action_embedding_size = 2048+128
hidden_size = 512
bidirectional = False
dropout_ratio = 0.5
feedback_method = 'teacher'  # teacher or sample
learning_rate = 0.0001
weight_decay = 0.0005
FEATURE_SIZE = 2048+128
n_iters = 20000
log_every = 100
save_every = 1000


def get_model_prefix(args, image_feature_list):
    image_feature_name = "+".join(
        [featurizer.get_name() for featurizer in image_feature_list])
    model_prefix = 'speaker_{}_{}'.format(
        feedback_method, image_feature_name)
    if args.use_train_subset:
        model_prefix = 'trainsub_' + model_prefix
    return model_prefix


def eval_model(agent, results_path, use_dropout, feedback, allow_cheat=False):
    agent.results_path = results_path
    agent.test(
        use_dropout=use_dropout, feedback=feedback, allow_cheat=allow_cheat)


def filter_param(param_list):
    return [p for p in param_list if p.requires_grad]


def train(args, train_env, agent, log_every=log_every, val_envs=None):
    ''' Train on training set, validating on both seen and unseen. '''

    if val_envs is None:
        val_envs = {}

    print('Training with %s feedback' % feedback_method)
    encoder_optimizer = optim.Adam(
        filter_param(agent.encoder.parameters()), lr=learning_rate,
        weight_decay=weight_decay)
    decoder_optimizer = optim.Adam(
        filter_param(agent.decoder.parameters()), lr=learning_rate,
        weight_decay=weight_decay)

    data_log = defaultdict(list)
    start = time.time()

    split_string = "-".join(train_env.splits)

    def make_path(n_iter):
        return os.path.join(
            args.SNAPSHOT_DIR, '%s_%s_iter_%d' % (
                get_model_prefix(args, train_env.image_features_list),
                split_string, n_iter))

    best_metrics = {}
    last_model_saved = {}
    for idx in range(0, args.n_iters, log_every):
        agent.env = train_env

        interval = min(log_every, args.n_iters-idx)
        iter = idx + interval
        data_log['iteration'].append(iter)

        # Train for log_every interval
        agent.train(encoder_optimizer, decoder_optimizer, interval,
                    feedback=feedback_method)
        train_losses = np.array(agent.losses)
        assert len(train_losses) == interval
        train_loss_avg = np.average(train_losses)
        data_log['train loss'].append(train_loss_avg)
        loss_str = 'train loss: %.4f' % train_loss_avg

        save_log = []
        # Run validation
        for env_name, (val_env, evaluator) in sorted(val_envs.items()):
            agent.env = val_env
            # Get validation loss under the same conditions as training
            agent.test(use_dropout=True, feedback=feedback_method,
                       allow_cheat=True)
            val_losses = np.array(agent.losses)
            val_loss_avg = np.average(val_losses)
            data_log['%s loss' % env_name].append(val_loss_avg)

            agent.results_path = '%s%s_%s_iter_%d.json' % (
                args.RESULT_DIR, get_model_prefix(
                    args, train_env.image_features_list),
                env_name, iter)

            # Get validation distance from goal under evaluation conditions
            results = agent.test(use_dropout=False, feedback='argmax')
            if not args.no_save:
                agent.write_results()
            print("evaluating on {}".format(env_name))
            score_summary, _ = evaluator.score_results(results, verbose=True)

            loss_str += ', %s loss: %.4f' % (env_name, val_loss_avg)
            for metric, val in score_summary.items():
                data_log['%s %s' % (env_name, metric)].append(val)
                if metric in ['bleu']:
                    loss_str += ', %s: %.3f' % (metric, val)

                    key = (env_name, metric)
                    if key not in best_metrics or best_metrics[key] < val:
                        best_metrics[key] = val
                        if not args.no_save:
                            model_path = make_path(iter) + "_%s-%s=%.3f" % (
                                env_name, metric, val)
                            save_log.append(
                                "new best, saved model to %s" % model_path)
                            agent.save(model_path)
                            if key in last_model_saved:
                                for old_model_path in \
                                        agent._encoder_and_decoder_paths(
                                            last_model_saved[key]):
                                    os.remove(old_model_path)
                            last_model_saved[key] = model_path

        print(('%s (%d %d%%) %s' % (
            timeSince(start, float(iter)/args.n_iters),
            iter, float(iter)/args.n_iters*100, loss_str)))
        for s in save_log:
            print(s)

        if not args.no_save:
            if save_every and iter % save_every == 0:
                agent.save(make_path(iter))

            df = pd.DataFrame(data_log)
            df.set_index('iteration')
            df_path = '%s%s_log.csv' % (
                args.PLOT_DIR, get_model_prefix(
                    args, train_env.image_features_list))
            df.to_csv(df_path)


def setup():
    torch.manual_seed(1)
    torch.cuda.manual_seed(1)

def make_speaker(args):
    enc_hidden_size = hidden_size//2 if bidirectional else hidden_size
    glove = np.load(glove_path)
    feature_size = FEATURE_SIZE
    vocab = read_vocab(TRAIN_VOCAB)
    encoder = try_cuda(SpeakerEncoderLSTM(
        action_embedding_size, feature_size, enc_hidden_size, dropout_ratio,
        bidirectional=bidirectional))
    decoder = try_cuda(SpeakerDecoderLSTM(
        len(vocab), word_embedding_size, hidden_size, dropout_ratio,
        glove=glove))
    agent = Seq2SeqSpeaker(
        None, "", encoder, decoder, MAX_INSTRUCTION_LENGTH)
    return agent

def make_env_and_models(args, train_vocab_path, train_splits, test_splits,
                        test_instruction_limit=None):
    setup()
    image_features_list = ImageFeatures.from_args(args)
    vocab = read_vocab(train_vocab_path)
    tok = Tokenizer(vocab=vocab)
    train_env = R2RBatch(image_features_list, batch_size=batch_size,
                         splits=train_splits, tokenizer=tok)

    enc_hidden_size = hidden_size//2 if bidirectional else hidden_size
    glove = np.load(glove_path)
    feature_size = FEATURE_SIZE
    encoder = try_cuda(SpeakerEncoderLSTM(
        action_embedding_size, feature_size, enc_hidden_size, dropout_ratio,
        bidirectional=bidirectional))
    decoder = try_cuda(SpeakerDecoderLSTM(
        len(vocab), word_embedding_size, hidden_size, dropout_ratio,
        glove=glove))

    test_envs = {
        split: (R2RBatch(image_features_list, batch_size=batch_size,
                         splits=[split], tokenizer=tok,
                         instruction_limit=test_instruction_limit),
                eval_speaker.SpeakerEvaluation(
                    [split], instructions_per_path=test_instruction_limit))
        for split in test_splits}

    return train_env, test_envs, encoder, decoder


def train_setup(args):
    train_splits = ['train']
    # val_splits = ['train_subset', 'val_seen', 'val_unseen']
    val_splits = ['val_seen', 'val_unseen']
    vocab = TRAIN_VOCAB

    if args.use_train_subset:
        train_splits = ['sub_' + split for split in train_splits]
        val_splits = ['sub_' + split for split in val_splits]
        vocab = SUBTRAIN_VOCAB

    train_env, val_envs, encoder, decoder = make_env_and_models(
        args, vocab, train_splits, val_splits)
    agent = Seq2SeqSpeaker(
        train_env, "", encoder, decoder, MAX_INSTRUCTION_LENGTH)
    return agent, train_env, val_envs


# Test set prediction will be handled separately
# def test_setup(args):
#     train_env, test_envs, encoder, decoder = make_env_and_models(
#         args, TRAINVAL_VOCAB, ['train', 'val_seen', 'val_unseen'], ['test'])
#     agent = Seq2SeqSpeaker(
#         None, "", encoder, decoder, MAX_INSTRUCTION_LENGTH,
#         max_episode_len=max_episode_len)
#     return agent, train_env, test_envs


def train_val(args):
    ''' Train on the training set, and validate on seen and unseen splits. '''
    agent, train_env, val_envs = train_setup(args)
    train(args, train_env, agent, val_envs=val_envs)


# Test set prediction will be handled separately
# def test_submission(args):
#     ''' Train on combined training and validation sets, and generate test
#     submission. '''
#     agent, train_env, test_envs = test_setup(args)
#     train(args, train_env, agent)
#
#     test_env = test_envs['test']
#     agent.env = test_env
#
#     agent.results_path = '%s%s_%s_iter_%d.json' % (
#         args.RESULT_DIR, get_model_prefix(args, train_env.image_features_list),
#         'test', n_iters)
#     agent.test(use_dropout=False, feedback='argmax')
#     if not args.no_save:
#         agent.write_results()


def make_arg_parser():
    parser = argparse.ArgumentParser()
    ImageFeatures.add_args(parser)
    parser.add_argument(
        "--use_train_subset", action='store_true',
        help="use a subset of the original train data for validation")
    parser.add_argument("--n_iters", type=int, default=20000)
    parser.add_argument("--no_save", action='store_true')
    return parser


if __name__ == "__main__":
    utils.run(make_arg_parser(), train_val)