Commit 1e6d1b85 authored by Antonio Ruiz's avatar Antonio Ruiz
Browse files

Documented version of the code.

parent b85cf9db

configs/mini_reverse_100_20_20_5_5.yaml

+3 −2
Original line number Diff line number Diff line
@@ -86,6 +86,7 @@ dqn: 
    beam_min: 1

    beam_max: 50

    state_type: 'hidden'

    reward_type: 'hc'

    nu_pretrain: 0

    reward_type: 'bleu_fin'

    nu_pretrain: 50

    non_stop: False

    other_descrip: '_with_best_parameters'

configs/mini_reverse_1_1_1_5_5.yaml

+6 −5
Original line number Diff line number Diff line 
name: "mini_reverse_2000_200_200_5_5"

name: "mini_reverse_1_1_1_5_5"



data:

    src: "src"
@@ -37,7 +37,7 @@ training: 
    validation_freq: 10

    logging_freq: 10

    eval_metric: "bleu"

    model_dir: "mini_reverse_1_1_1_5_5_model"

    model_dir: 'mini_reverse_1_1_1_5_5_lr_bleu_fin_model'

    overwrite: True

    shuffle: True

    use_cuda: False
@@ -85,7 +85,7 @@ model: 
dqn:

    epochs: 5000

    sample_size: 64

    lr: 0.0001

    lr: 0.00001

    egreed_max: 0.9

    egreed_min: 0.001

    gamma: 0.99
@@ -94,8 +94,9 @@ dqn: 
    beam_min: 1

    beam_max: 50

    state_type: 'attention'

    reward_type: 'hc_batch'

    reward_type: 'bleu_fin'

    nu_pretrain: 0

    non_stop: False

    test_variable: 'lr'

    test_range: [0.001, 0.0001, 0.00001]

    test_range: []

    other_descrip: 'lr_var'

configs/mini_reverse_2000_200_200_5_5.yaml

+1 −1
Original line number Diff line number Diff line
@@ -82,7 +82,6 @@ model: 
    # test_range: [0.8, 0.9, 0.99]

    # other_descrip: 'gamma_var'





    #reward_type: 'bleu_fin' or 'hc_batch'

    #state_type: 'attention' or 'hidden'

dqn:
@@ -99,6 +98,7 @@ dqn: 
    state_type: 'attention'

    reward_type: 'hc_batch'

    nu_pretrain: 0

    non_stop: False

    test_variable: 'lr'

    test_range: [0.001, 0.0001, 0.00001]

    other_descrip: 'lr_var'

configs/mini_reverse_gdrive.yaml

deleted100644 → 0
+0 −93
Original line number Diff line number Diff line 
name: "mini_reverse_batch_experiment"



data:

    src: "src"

    trg: "trg"

    # generate data with scripts/generate_reverse_task.py

    train: "test/data/mini_reverse_0/train"

    dev: "test/data/mini_reverse_0/dev"

    test: "test/data/mini_reverse_0/test"

    level: "word"

    lowercase: False

    max_sent_length: 25

    src_voc_min_freq: 0

    src_voc_limit: 100

    trg_voc_min_freq: 0

    trg_voc_limit: 100

    #src_vocab: "mini_reverse_0/src_vocab.txt"

    #trg_vocab: "mini_reverse_0/trg_vocab.txt"



testing:

    beam_size: 1

    alpha: 1.0



training:

    random_seed: 42

    optimizer: "adam"

    learning_rate: 0.001

    learning_rate_min: 0.0002

    weight_decay: 0.0

    clip_grad_norm: 1.0

    batch_size: 1

    batch_type: "sentence"

    scheduling: "plateau"

    patience: 5

    decrease_factor: 0.5

    early_stopping_metric: "eval_metric"

    epochs: 10

    validation_freq: 10

    logging_freq: 10

    eval_metric: "bleu"

    model_dir: "/content/gdrive/My Drive/models/batch_V0"

    overwrite: True

    shuffle: True

    use_cuda: False

    max_output_length: 10

    print_valid_sents: [0, 3, 6]

    keep_last_ckpts: 2



model:

    initializer: "xavier"

    embed_initializer: "normal"

    embed_init_weight: 0.1

    bias_initializer: "zeros"

    init_rnn_orthogonal: False

    lstm_forget_gate: 0.

    encoder:

        rnn_type: "lstm"

        embeddings:

            embedding_dim: 16

            scale: False

        hidden_size: 24

        bidirectional: True

        dropout: 0.1

        num_layers: 1

    decoder:

        rnn_type: "lstm"

        embeddings:

            embedding_dim: 16

            scale: False

        hidden_size: 24

        dropout: 0.1

        hidden_dropout: 0.1

        num_layers: 1

        input_feeding: True

        init_hidden: "bridge"

        attention: "luong"



dqn:

    epochs: 2000

    sample_size: 256

    lr: 0.00005

    egreed_max: 0.9

    egreed_min: 0.001

    gamma_max: 0.9

    gamma_min: 0.3

    nu_iter: 300

    mem_cap: 5000

    beam_min: 1

    beam_max: 50

    state_type: 'attention'

    reward_type: 'hc_batch'

    nu_pretrain: 40

    other_descrip: '_mini_reverse_with_batch_w_attention'

configs/mini_reverse_mega_tiny.yaml

deleted100644 → 0
+0 −95
Original line number Diff line number Diff line 
name: "mini_reverse_100_10_10_5_5"



data:

    src: "src"

    trg: "trg"

    # generate data with scripts/generate_reverse_task.py

    train: "test/data/mini_reverse_100_10_10_5_5/train"

    dev: "test/data/mini_reverse_100_10_10_5_5/dev"

    test: "test/data/mini_reverse_100_10_10_5_5/test"

    level: "word"

    lowercase: False

    max_sent_length: 25

    src_voc_min_freq: 0

    src_voc_limit: 100

    trg_voc_min_freq: 0

    trg_voc_limit: 100

    #src_vocab: "mini_reverse_100_10_10_5_5_model/src_vocab.txt"

    #trg_vocab: "mini_reverse_100_10_10_5_5_model/trg_vocab.txt"



testing:

    beam_size: 1

    alpha: 1.0



training:

    random_seed: 42

    optimizer: "adam"

    learning_rate: 0.01

    learning_rate_min: 0.00002

    weight_decay: 0.0

    clip_grad_norm: 1.0

    batch_size: 10

    batch_type: "sentence"

    scheduling: "plateau"

    patience: 5

    decrease_factor: 0.5

    early_stopping_metric: "eval_metric"

    epochs: 50

    validation_freq: 10

    logging_freq: 10

    eval_metric: "bleu"

    model_dir: "mini_reverse_100_10_10_5_5_model"

    # model_dir: "/content/gdrive/My Drive/models/batch_V0"

    overwrite: True

    shuffle: True

    use_cuda: False

    max_output_length: 10

    print_valid_sents: [0, 3, 6]

    keep_last_ckpts: 2



model:

    initializer: "xavier"

    embed_initializer: "normal"

    embed_init_weight: 0.1

    bias_initializer: "zeros"

    init_rnn_orthogonal: False

    lstm_forget_gate: 0.

    encoder:

        rnn_type: "lstm"

        embeddings:

            embedding_dim: 16

            scale: False

        hidden_size: 24

        bidirectional: True

        dropout: 0.1

        num_layers: 1

    decoder:

        rnn_type: "lstm"

        embeddings:

            embedding_dim: 16

            scale: False

        hidden_size: 24

        dropout: 0.1

        hidden_dropout: 0.1

        num_layers: 1

        input_feeding: True

        init_hidden: "bridge"

        attention: "luong"

dqn:

    epochs: 1000

    sample_size: 256

    lr: 0.00005

    egreed_max: 0.9

    egreed_min: 0.001

    gamma_max: 0.9

    gamma_min: 0.3

    nu_iter: 100

    mem_cap: 5000

    beam_min: 1

    beam_max: 50

    state_type: 'attention'

    test_variable: 'lr'

    test_range: [0.5, 0.7, 0.4]

    reward_type: 'hc_batch'

    nu_pretrain: 0

    other_descrip: 'lr__00005_rep_results'