What is this?

LAnguage Modelling Benchmarks is to tune and test Tensorflow language models. It was used in the following papers (alse see citations):

• On the state of the art of evaluation in neural language models

  See ./experiment/on-the-state/README.md for more.

• Pushing the bounds of dropout

  See ./experiment/pushing-the-bounds/README.md for more.

• Mogrifier LSTM

  See ./experiment/mogrifier/README.md for more.

Overview

The default dataset locations are ~/data/<dataset-name>/. See lib/config/{ptb,wikitext-2,wikitext-103,enwik8}.sh for the defaults.

To train a small LSTM on Penn Treebank, run this script:

experiment/train_ptb_10m_lstm_d1.sh

In the script, model configuration, data files, etc are specified by setting variables:

training_file="ptb.train.txt"
validation_file="ptb.valid.text"
model="lstm"
hidden_size=500

These shell variables are passed as command line arguments to the python program. These options are documented in the reference section.

To test a trained model:

experiment/test.sh run "mymodel" "experiment_dir_of_training_run"

In the output, lines with final valid* xe:': have the validation set cross-entropy. Evaluation results are printed as they happen (see the section on evaluation). Lines of special interest in the output are those with final {valid,test} in them. The format is the following:

  final ${dataset}_${eval_method}[_${dropout_multiplier}][_t${softmax_temp}]

For eval_method=arithmetic with eval_dropout_multiplier=0.8 and eval_softmax_temperature=0.9 results may look like this after 200 optimization steps and 2 evaluations:

  turn: 2 (eval), step: 200 (opt) (5.29/s)
  final valid_mca_d0.8_t0.9 xe:  5.315
  final test_mca_d0.8_t0.9 xe:  5.289

... except that training runs normally don't have the test set results (see eval_on_test). Test runs are pretty much training runs with no optimization steps.

Installation

For example:

conda create -n tfp3.7 python=3.7 numpy scipy
conda activate tfp3.7
conda install cudatoolkit
conda install cudnn
conda install tensorflow-gpu=1.15
conda install tensorflow-probability-gpu=1.15
conda install tensorflow-probability
pip install -e <path-to-git-checkout>

Reference

A value given for an option gets converted to the data type corresponding to the option in question. In the following, options are listed with their data type and default value (e.g. model (string, lstm) means that the variable model has type string and default value lstm). If there is no default value listed, then the option is mandatory.

Data

• training_file (string)

  The file with the training data, one line per example. Newlines are translated to an end-of-sentence token.

• validation_file (string)

  A file of the same format as training_file. During training, the model is evaluated periodically on data from validation_file. Most notably, early stopping and hyperparameter tuning are based on performance on this set of examples. This must not be specified when doing cross-validation as in that case, the evaluation set is constructed from the training set.

• test_file (string, '')

  A file of the same format as training_file. During training, the model evaluated periodically on data from test_file and the results are logged. As opposed to validation_file, this dataset have no affect on training or tuning. The empty string (the default) turns off evaluation on the test set.

• file_encoding (string, utf-8)

  The encoding of training_file, validation_file and test_file.

• word_based (boolean, false)

  Whether to do word or character based modelling. If word based, lines are split at whitespace into tokens. Else, lines are simply split into characters.

• episodic (boolean, false)

  If true, iterate over examples (lines in the data files) in random order. If false, iterate mostly sequentially carrying over model from the previous example to the next.

Model

• num_params (float, -1)

  An upper bound on the total number of trainable parameters over all parts of the model (including the recurrent cell and input/output embeddings). If this is set to a meaningful value (i.e. not -1, the default), then hidden_size is set to the largest possible value such that the parameter budget is not exceeded.

• share_input_and_output_embeddings (boolean, false)

  Whether the input and output embeddings are the same matrix (transposed) or independent (the default). If true, then input_embedding_size and output_embedding_size must be the same.

• input_embedding_size (integer, -1)

  The length of the vector that represents an input token. If -1 (the default), then it's determined by input_embedding_ratio.

• output_embedding_size (integer, -1)

  The length of the vector that represents an output token. If -1 (the default), then it's determined by output_embedding_ratio. If - after applying the defaulting rules - output_embedding_size is not equal to hidden_size, then the cell output is linearly transformed to output_embedding_size before the final linear transform into the softmax.

• input_embedding_ratio (float, 1.0)

  If input_embedding_size is not specified (i.e. -1), then it's set to round(input_embedding_ratio*hidden_size).

• output_embedding_ratio (float, -1.0)

  If output_embedding_size is not specified (i.e. -1), then it's set to round(output_embedding_ratio*hidden_size). The default value of -1, makes output_embedding_ratio default to the value of input_embedding_ratio so that one can tune easily with share_input_and_output_embeddings =true.

• mos_num_components (integer, 0)

  See Breaking the softmax bottleneck. The default of 0 turns this feature off.

• embedding_dropout (float, 0.0)

  The probability that all occurrences of a word are dropped from a batch.

• token_dropout (float, 0.0)

  The probability that a token will be dropped (i.e. the input at that step becomes zero). This can be thought of as a version of embedding_dropout that has different masks per time step.

• input_dropout (float, 0.0)

  The dropout rate (here and elsewhere, 0 means deterministic operation) for the input to the first layer (i.e. just after the input embeddings). This drops out individual elements of the embedding vector.

• output_dropout (float, 0.0)

  The dropout rate for just after the cell output.

• downprojected_output_dropout (float, -1.0)

  The dropout rate for the projection of the cell output. Only used if output_embedding_size is different from hidden_size or if mos_num_components is not 1. Defaults to output_dropout if set to -1.

• shared_mask_dropout (boolean, false)

  Whether to use the time same dropout mask for all time steps for input_dropout, inter_layer_dropout, output_dropout and downprojected_output_dropout.

• output_once (boolean, true)

  Whether to compute the logits from the cell output in a single operation or per time step. The single operation is faster but uses more GPU memory. Also, see swap_memory.

Cell

• model (string, lstm)

  One of lstm, rhn (Recurrent Highway Network), nas.

• num_layers (integer, 1)

  The number of same-sized LSTM cells stacked on top of each other, or the number of processing steps per input an RHN does. Has no effect on NAS.

• lstm_skip_connection (boolean, true)

  If true, for multi-layer (num_layers>1) LSTMs, the output is computed as the sum of the outputs of the individual layers.

• feature_mask_rounds (integer, 0)

  The Mogrifier LSTM is implemented in terms of the feature masking option. The LSTM specific feature masking option involves gating the input and the state before they are used for calculating all the other stuff (i.e. i, j, o, f). This allows input features to be reweighted based on the state, and state features to be reweighted based on the input. See the Mogrifier LSTM paper for details.

  When feature_mask_rounds is 0, there is no extra gating in the LSTM. When 1<=, the input is gated: x *= 2*sigmoid(affine(h))). When 2<=, the state is gated: h *= 2*sigmoid(affine(x))). For higher number of rounds, the alternating gating continues.

• feature_mask_rank (integer, 0)

  If 0, the linear transforms described above are full rank, dense matrices. If

  0, then the matrix representing the linear transform is factorized as the product of two low rank matrices ([*, rank] and [rank, *]). This reduces the number of parameters greatly.

• hidden_size (string, "-1")

  A comma-separated list of integers representing the number of units in the state of the recurrent cell per layer. Must not be longer than num_layers. If it's shorter, then the missing values are assumed to be equal to the last specified one. For example, for a 3 layer network "512,256" results in the first layer having 512 units, the second and the third having 256. If "-1" (the default), an attempt is made to deduce it from num_params assuming all layers have the same size.

• layer_norm (boolean, false)

  Whether to perform Layer Normalization (currently only implemented for LSTMs).

• activation_fn (string, tf.tanh)

  The non-linearity for the update candidate ('j') and the output ('o') in an LSTM, or the output ('h') in an RHN.

• tie_forget_and_input_gates (boolean, false)

  In an LSTM, whether the input gate ('i') is set to 1 minus the forget gate ('f'). In an RHN, whether the transform gate ('t') is set to 1 minus the carry gate ('c').

• cap_input_gate (boolean, true)

  Whether to cap the input gate at 1-f if tie_forget_and_input_gates is off. Currently only affects LSTMs. This makes learning more stable, especially at the early stages of training.

• trainable_initial_state (boolean, true)

  Whether the initial state of the recurrent cells is allowed to be learnt or is set to a fixed zero vector. In non-episodic mode, this switch is forced off.

• inter_layer_dropout (float, 0.0)

  The input dropout for layers other than the first one. Defaults to no dropout, but setting it to -1 makes it inherit input_dropout. It has no effect on RHNs, since the input is not fed to their higher layers.

• state_dropout (float, 0.0)

  This is the dropout rate for the recurrent state from the previous time step ('h' in an LSTM, 's' in an RHN). See Yarin Gal's "A Theoretically Grounded Application of Dropout in Recurrent Neural Networks". The dropout mask is the same for all time steps of a specific example in one batch.

• update_dropout (float, 0.0)

  This is the Recurrent Dropout (see "Recurrent Dropout without Memory Loss") rate on the update candidate ('j' in an LSTM, 'h' in an RHN). Should have been named Update Dropout.

• cell_clip (float, -1.0)

  If set to a positive value, the cell state ('c' in an LSTM, 's' in an RHN) is clipped to the [-cell_clip, cell_clip] range after each iteration.

Training

Objective

• model_average (string, arithmetic)

  Pushing the bounds of dropout makes the point that the actual dropout objective being optimized is a lower bound of the true objectives of many different models. If we construct the lower bound from multiple samples though (a'la IWAE), the lower bound will get tighter.

  model_average is the training time equivalent of eval_method and determines what kind of model (and consequently, averaging) is to be used. One of geometric, power and arithmetic. Only in effect if num_training_samples > 1.

• num_training_samples (integer, 1)

  The number of samples from which to compute the objective (see model_average). Each training example being presented is run through the network num_training_samples times so the effective batch size is batch_size * num_training_samples. Increasing the number of samples doesn't seems to help generalization, though.

• l2_penalty (float, 0.0)

  The L2 penalty on all trainable parameters.

• l1_penalty (float, 0.0)

  The L1 penalty on all trainable parameters.

• activation_norm_penalty (float, 0.0)

  Activation Norm Penalty (Regularizing and optimizing LSTM language models by Merity et al).

• drop_state_probability (float, 0.0)

  In non-episodic mode, model state is carried over from batch to batch. Not feeding back the state with drop_state_probability encourages the model to work well starting from the zero state which brings it closer to the test regime.

Initialization

• embedding_init_factor (float, 1.0)

  All input embedding weights are initialized with a truncated normal distribution with mean 0 and:

  stddev=sqrt(embedding_init_factor/input_embedding_size)
  

• scale_input_embeddings (boolean, false)

  This is not strictly an initialization option, but it serves a similar purpose. Input embeddings are initialized from a distribution whose variance is inversely proportional to input_embedding_size. Since every layer in the network is initialized to produce output with approximately the same variance as its input, changing the embedding size has a potentially strong, undesirable effect on optimization. Set scale_input_embeddings to true to multiply input embeddings by sqrt(input_embedding_size) to cancel this effect.

  As opposed to just changing embedding_init_factor, this multiplication has the benefit that the input embedding matrix is of the right scale for use as the output embedding matrix should share_input_and_output_embeddings be turned on.

• cell_init_factor (float, 1.0)

  The various weight matrices in the recurrent cell are initialized independently (of which there are 8 in an LSTM, 4/2 in an RHN) with

  stddev=sqrt(cell_init_factor/fan_in)
  

  while biases are initialized with

  stddev=sqrt(cell_init_factor/hidden_size)
  

• forget_bias (float, 1.0)

  Sometimes initializing the biases of the forget gate ('f') in the LSTM (or that of the carry gate ('c') in an RHN) to a small positive value (typically 1.0, the default) makes the initial phase of optimization faster. Higher values make the network forget less of its state over time. With deeper architectures and no skip connections (see num_layers and lstm_skip_connection), this may actually make optimization harder.

  The value of forget_bias is used as the mean of the distribution used for initialization with unchanged variance.

• output_init_factor (float, 1.0)

  If share_input_and_output_embeddings is false, then the output projection (also known as the output embeddings) is initialized with

  stddev=sqrt(output_init_factor/fan_in)
  

  If share_input_and_output_embeddings is true, then this only affects the linear transform of the cell output (see output_embedding_size).

Schedule

• steps_per_turn (integer, 1000)

  The number of optimization steps between two successive evaluations. After this many steps performance is evaluated and logged on the training, validation and test sets (if specified). One so called turn consists of steps_per_turn optimization steps.

• turns (integer)

  The number of evaluations beyond which training cannot continue (also see early stopping).

• print_training_stats_every_num_steps (integer, 1000)

  Debug printing frequency.

Optimization

• optimizer_type (string, rmsprop)

  The optimizer algorithm. One of rmsprop, adam, adagrad, adadelta and sgd.

• rmsprop_beta2 (float, 0.999)

  RMSPROP is actually Adam with beta1=0.0 so that Adam's highly useful correction to the computed statistics is in effect which allows higher initial learning rates. Only applies when optimizer_typeisrmsprop`.

• rmsprop_epsilon (float, 1e-8)

  Similar to adam_epsilon. Only applies when optimizer_type is rmsprop.

• adam_beta1 (float, 0.9)

• adam_beta2 (float, 0.999)

• adam_epsilon(float, 1e-8)

• max_grad_norm (float, 1.0)

  If non-zero, gradients are rescaled so that their norm does not exceed max_grad_norm.

• batch_size (integer)

  Batch size for training. Also, the evaluation batch size unless min_non_episodic_eval_examples_per_stripe overrides it.

• accum_batch_size (integer, -1)

  The number of examples that are fed to the network at the same time. Set this to a divisor of batch_size to reduce memory usage at the cost of possibly slower training. Using accum_batch_size does not change the results.

• max_time_steps (integer, 100)

  For episodic operation, examples that have more tokens than this are truncated when the training and test files when loaded. For non-episodic operation, this is the window size of the truncated backprop.

• trigger_averaging_turns (integer, -1)

  The number of turns of no improvement on the validation set, after which weight averaging is turned on. Weight averaging is a trivial generalization of the idea behind Averaged SGD: it keeps track of the average weights, updating the average after each optimization step. Weight averaging does not affect training directly, only through evaluation. This feature is an alternative to dropping the learning rate.

• trigger_averaging_at_the_latest (integer, -1)

  If optimization reaches turn trigger_averaging_at_the_latest, then it is ensured that averaging is turned on. Set this to be somewhat smaller than turns so that all runs get at least one drop which should the results more comparable.

Learning rate

• learning_rate (float, 0.001)

• drop_learning_rate_turns (integer, -1)

  If the validation score doesn't improve for drop_learning_rate_turns number of turns, then the learning rate is multiplied by drop_learning_rate_multiplier, possibly repeatedly.

• drop_learning_rate_multiplier (float, 1.0)

  Set this to a value less than 1.0.

• drop_learning_rate_at_the_latest (integer, -1)

  If optimization reaches turn drop_learning_rate_multiplier_at_the_latest without having yet dropped the learning rate, then it is dropped regardless of whether the curve is still improving or not. Set this to be somewhat smaller than turns so that all runs get at least one drop which should the results more comparable.

Early stopping

• early_stopping_turns (integer, -1)

  Maximum number of turns without improvement in validation cross-entropy before stopping.

• early_stopping_rampup_turns (integer, 0)

  The effective early_stopping_turns starts out at 1 and is increased linearly to the specified early_stopping_turns in early_stopping_rampup_turns turns.

• early_stopping_worst_xe_target (float, '')

  If the estimated best possible validation cross-entropy (extrapolated from the progress made in the most recent early_stopping_turns (subject to rampup) is worse than early_stopping_worst_xe_target, then training is stopped. This is actually a string of comma separated floats. The first value is in effect when the learning rate has not been dropped yet. The second value is effect if it has been dropped once and so on. The last element of the list applies to any further learning rate drops.

• early_stopping_slowest_rate (float, 0.0)

  The rate is defined as the average improvement in validation cross-entropy over the effective early_stopping_turns (see early_stopping_rampup_turns). If the rate is less than early_stopping_slowest_rate, then stop early.

Cross-validation

• crossvalidate (boolean, false)

  If true, randomly split the training set into crossvalidation_folds folds, evaluate performance on each and average the cross-entropies. Repeat the entire process for crossvalidation_rounds and average the averages.

• crossvalidation_folds (integer, 10)

  Number of number of folds to split the training set into.

• crossvalidation_rounds (integer, 1)

  If crossvalidate, then do this many rounds of crossvalidate_folds-fold crossvalidation. Set this to a value larger than one if the variance of the cross-validation score over the random splits is too high.

Evaluation

The model being trained is evaluated periodically (see turns and steps_per_turn) on the validation set (see validation_file) and also on the training set (see training_file). Evaluation on the training set is different from the loss as it does not include regularization terms such as l2_penalty and is performed the same way as evaluation on the validation set (see eval_method).

To evaluate a saved model one typically wants to do no training, disable saving of checkpoints and evaluate on the test set which corresponds to this:

turns=0
save_checkpoints=false
eval_on_test=true

Furthermore, load_checkpoint, and in all likelihood config_file must be set. This is all taken care of by the experiment/test.sh script.

• max_training_eval_batches (integer, 100)

  When evaluating performance on the training set, it is enough to get a rough estimate. If specified, at most max_training_eval_batches number of batches will be evaluated. Set this to zero, to turn off evaluation on the training set entirely. Set it to -1 to evaluate on the entire training set.

• max_eval_eval_batches (integer, -1)

  Evaluation can be pretty expensive with large datasets. For expediency, one can impose a limit on the number of batches of examples to work with on the validation test.

• max_test_eval_batches (integer, -1)

  Same as max_eval_eval_batches but for the test set.

• min_non_episodic_eval_examples_per_stripe (integer, 100)

  By default, evaluation is performed using the training batch size causing each "stripe" in a batch to be over rougly dataset_size/batch_size number of examples. With a small dataset in a non-episodic setting, that may make the evaluation quite pessimistic. This flag ensures that the batch size for evaluation is small enought that at least this many examples are processed in the same stripe.

• eval_on_test (boolean, false)

  Even if test_file is provided, evaluation on this test dataset is not performed by default. Set this to true to do that. Flipping this switch makes it easy to test a model by loading a checkpoint and its saved configuration without having to remember what the dataset was.

• eval_method (string, deterministic)

  One of deterministic, geometric, power and arithmetic. This determines how dropout is applied at evaluation.

  • deterministic is also known as standard dropout: dropout is turned off at evaluation time and a single deterministic pass propagates the expectation of each unit through the network.
  • geometric performs a renormalized geometric average of predicted probabilities over randomly sampled dropout masks.
  • power computes the power mean with exponent eval_power_mean_power.
  • arithmetic computes the arithmetic average.

  See Pushing the bounds of dropout for a more detailed discussion.

• num_eval_samples (integer, 0)

  The number of samples to average probabilities over at evaluation time. Needs some source of stochasticity (currently only dropout) to be meaningful. When it's zero, the model is run in deterministic mode. Training evaluation is always performed in deterministic mode for expediency.

• eval_softmax_temperature (float, 1.0)

  Set this to a value lower than 1 to smoothen the distribution a bit at evaluation time to counter overfitting. Set it to a value between -1 and 0 to search for the optimal value between -value and 1 on the validation set. For example, eval_softmax_temperature=-0.8 will search for the optimal temperature between 0.8 and 1.0.

• eval_power_mean_power (float, 1.0)

  The exponent of the renormalized power mean to compute predicted probabilities. Only has an effect if eval_method=power.

• eval_dropout_multiplier (float, 1.0)

  At evaluation time all dropout probabilities used for training are multiplied by this. Does not affect the eval_method=deterministic case. See Pushing the bounds of dropout for details.

• validation_prediction_file (string)

  The name of the file where log probabilities of for the validation file are written. The file gets superseded by a newer version each time the model is evaluated. The file lists tokens and predicted log probabilities on alternating lines. Currently only implemented for deterministic evaluation.

• dyneval (boolean, false)

  Whether model weights shall be updated at evaluation time (see [Dynamic Evaluation of Neural Sequence Models][https://arxiv.org/abs/1709.07432] by Krause et al.). This forces batch size at evaluation time to 1 which makes it very slow, so turn it is best to leave it off until the final evaluation.

  Whereas RMSProp maintains an online estimate of gradient variance, dynamic evaluation bases its estimate on training statistics which are affected by max_training_eval_batches and batch_size.

  Also, when doing dynamic evaluation it might make sense to turn off some regularizers such as l2_penalty, or hacks like max_grad_norm.

• dyneval_learning_rate (float, 0.001)

  The learning rate for dynamic evaluation.

• dyneval_decay_rate (float, 0.02)

  The rate with which weights revert to the mean which is defined as what was trained.

• dyneval_epsilon (float, 1e-5)

  This serves a similar purpose to rmsprop_epsilon, but for dynamic evaluation.

Experiments

• name (string, see below) <name="name">

  The name of the experiment. Defaults to the git version concatenated with the basename of the script (without the .sh). See experiment_dir.

• experiment_dir (string, ./ + $name)

  Directory for saving configuration, logs and checkpoint files.

  Lamb's git version is saved in lamb_version along with any uncommitted changes in the checkout (if in a git tree). stdout and stderr are also captured.

  If save_checkpoints is true, checkpoints are saved here. Also see save_config.

• save_config (boolean, true)

  All options are saved in $experiment_dir/config except for which it doesn't make sense:

  • load_checkpoint
  • load_optimizer_state
  • load_averaged
  • ensure_new_experiment
  • config_file
  • save_config

• ensure_new_experiment (boolean, true)

  If ensure_new_experiment is true, a random suffix is appended to experiment_dir to ensure the experiment starts afresh. When ensure_new_experiment is false and experiment_dir exists, the last checkpoint will be loaded on startup from that directory.

• config_file (string, '')

  This is to load $experiment_dir/config that gets saved automatically when save_checkpoints is true. It is not needed if one uses experiment/test.sh for evaluation. If a configuration option is set explicitly and is also in the configuration file, then the explicit version overrides the one in the configuration file.

Checkpoints

• save_checkpoints (boolean, true)

  Whether to save any checkpoints. save_checkpoints also affects saving of the configuration (see config_file.

  If save_checkpoints is true, then two checkpoints are saved: $experiment_dir/best and $experiment_dir/last. If last exists, it will be loaded automatically on startup and training will continue from that state. If that's undesirable, use a different experiment_dir or delete the checkpoint manually. The best checkpoint corresponds to the best validation result seen so far during preriodic model evaluation during training.

• load_checkpoint (string, '')

  The name of the checkpoint file to load instead of loading $experiment_dir/last or randomly initializing. Absolute or relative to experiment_dir.

• load_optimizer_state (boolean, true)

  Set this to false to prevent load_checkpoint from attempting to restore optimizer state. This effectively reinitializes the optimizer and also allows changing the optimizer type. It does not affect automatic loading of the latest checkpoint (see experiment_dir).

Misc options

• seed (integer, 0)

  The random seed. Both python and tensorflow seeds are initialized with this value. Due to non-determinism in tensorflow, training runs are not exactly reproducible even with the same seed.

• swap_memory (boolean, false)

  Transparently swap the tensors produced in forward inference but needed for back prop from GPU to CPU. This allows training RNNs which would typically not fit on a single GPU, but slows things down a bit.

• log_device_placement (boolean, false)

  Log tensorflow device placement.

Notes

This is not an official Google product.

Citations

• On the state of the art of evaluation in neural language models

    @inproceedings{
      melis2018on,
      title={On the State of the Art of Evaluation in Neural Language Models},
      author={G{\'a}bor Melis and Chris Dyer and Phil Blunsom},
      booktitle={International Conference on Learning Representations},
      year={2018},
      url={https://openreview.net/forum?id=ByJHuTgA-},
    }
  

• Pushing the bounds of dropout

    @article{melis2018pushing,
      title={Pushing the bounds of dropout},
      author={Melis, G{\'a}bor and Blundell, Charles and Ko{\v{c}}isk{\'y}, Tom{\'a}{\v{s}} and Hermann, Karl Moritz and Dyer, Chris and Blunsom, Phil},
      journal={arXiv preprint arXiv:1805.09208},
      year={2018}
    }
  

• Mogrifier LSTM

    @article{melis2020mogrifier,
      title={Mogrifier LSTM},
      author={Melis, G{\'a}bor and Ko{\v{c}}isk{\'y}, Tom{\'a}{\v{s}} and Blunsom, Phil},
      booktitle={International Conference on Learning Representations},
      year={2020},
      url={https://openreview.net/forum?id=SJe5P6EYvS},
    }