示例#1
0
        sum(param.nelement() for param in nli_model.encoder.parameters()
            if param.requires_grad)))

    best_nli_dev_acc = 0
    best_nli_train_acc = -np.infty

    # load trained model from checkpoint
    if len(sys.argv) > 1:
        checkpoint_dir = sys.argv[1]
        print('loading from checkpoint in {}'.format(checkpoint_dir))
        model_pipeline_pytorch.load_checkpoint(nli_model,
                                               checkpoint=checkpoint_dir)
        nli_state['lr'] = 0.01
        print('resetting lr as {}'.format(nli_state['lr']))

    nli_model.embed.weight.data = load_embeddings.load_embeddings(
        nli_dm.vocab, constants.EMBED_DATA_PATH, nli_args.embedding_size)
    squad_model.embed.weight.data = load_embeddings.load_embeddings(
        squad_dm.vocab, constants.EMBED_DATA_PATH, nli_args.embedding_size)

    nli_criterion = nn.NLLLoss()
    squad_criterion = nn.NLLLoss()

    for epoch in range(NumEpochs):
        nli_dm.shuffle_train_data()
        squad_dm.shuffle_train_data()

        print('lr {}'.format(nli_state['lr']))
        nli_optimizer = optim.SGD(
            [param for param in nli_model.parameters() if param.requires_grad],
            lr=nli_state['lr'])
        squad_optimizer = optim.SGD([
示例#2
0
})
squad_state = {k: v for k, v in squad_args.items()}


if __name__ == "__main__":
    print(args)
    checkpoint = sys.argv[1]
    print('found checkpoint dir {}'.format(checkpoint))

    dm = wrangle.DataManager(args)
    if args.add_squad:  # add squad to vocab to match checkpoint
        squad_dm = SquadDataManager(squad_args, vocab=dm.vocab)
    args.n_embed = dm.vocab.n_words
    if args.type == 'siamese':
        model = siamese_pytorch.SiameseClassifier(config=args)
        model.embed.weight.data = load_embeddings.load_embeddings(
            dm.vocab, constants.EMBED_DATA_PATH, args.embedding_size)
    elif args.type == 'decomposable':
        model = decomposable_pytorch.SNLIClassifier(config=args)
        model.encoder.embedding.weight.data = load_embeddings.load_embeddings(
            dm.vocab, constants.EMBED_DATA_PATH, args.embedding_size)
    else:
        model = Seq2SeqPytorch(args=args, vocab=dm.vocab)
        model.encoder.embedding.weight.data = load_embeddings.load_embeddings(
            dm.vocab, constants.EMBED_DATA_PATH, args.embedding_size)

    model_pipeline_pytorch.load_checkpoint(model, checkpoint=checkpoint)

    model.eval()
    if args.cuda:
        model = model.cuda()