Skip to content

Latest commit

 

History

History
224 lines (170 loc) · 9.73 KB

eval.md

File metadata and controls

224 lines (170 loc) · 9.73 KB

Evaluating a Model

Introduction

This page outlines the steps to evaluate a model with T5X on downstream tasks defined with SeqIO.

Refer to this tutorial when you have an existing model that you want to evaluate. If you would like to fine-tune your model before evaluation, please refer to the fine-tuning tutorial. You can run evals as part of your fine-tuning run as well.

Overview

Evaluating a model with T5X consists of the following steps:

  1. Choose the model to evaluate.
  2. Choose the SeqIO Task/Mixture to evaluate the model on.
  3. Write a Gin file that configures the model, SeqIO Task/Mixture and other details of your eval run.
  4. Launch your experiment locally or on XManager.
  5. Monitor your experiment and parse metrics.

These steps are explained in detail in the following sections. An example run that evaluates a fine-tuned T5-1.1-Small checkpoint on the (Open Domain) Natural Questions benchmark is also showcased.

Step 1: Choose a model

To evaluate a model, you need a Gin config file that defines the model params, and the model checkpoint to load from. For this example, a T5-1.1-Small model fine-tuned on the natural_questions_open_test SeqIO Task will be used:

If you would like to fine-tune your model before evaluation, please follow the fine-tuning tutorial, and continue to Step 2.

Step 2: Choose a SeqIO Task/Mixture

A SeqIO Task encapsulates the data source, the preprocessing logic to be performed on the data before querying the model, the postprocessing logic to be performed on model outputs, and the metrics to be computed given the postprocessed outputs and targets. A SeqIO Mixture denotes a collection of Tasks and enables fine-tuning a model on multiple Tasks simultaneously.

Many common datasets and benchmarks, e.g. GLUE, SuperGLUE, WMT, SQUAD, CNN/Daily Mail, etc. have been implemented as SeqIO Tasks/Mixtures and can be used directly. These Tasks/Mixtures are defined in t5/data/tasks.py and t5/data/mixtures.py.

For the example run, you will evaluate the model on the Natural Questions benchmark, which has been implemented as the natural_questions_open Task in /third_party/google_research/google_research/t5_closed_book_qa/t5_cbqa/tasks.py. Here's an example of a single row of preprocessed data from this Task:

{
    'inputs_pretokenized': 'nq question: what was the main motive of salt march',
    'inputs': [3, 29, 1824, 822, 10, 125, 47,  8, 711, 10280, 13, 3136, 10556, 1]
    'targets_pretokenized': 'challenge to British authority',
    'targets': [1921, 12, 2390, 5015, 1],
    'answers': ['challenge to British authority']
}

Step 3: Write a Gin Config

After choosing the model and SeqIO Task/Mixture for your run, the next step is to configure your run using Gin. If you're not familiar with Gin, reading the T5X Gin Primer is recommended.

T5X provides a Gin file that configures the T5X eval job (located at t5x/configs/runs/eval.gin), and expects a few params from you. These params can be specified in a separate Gin file, or via commandline flags. Following are the required params:

  • CHECKPOINT_PATH: This is the path to the model checkpoint (from Step 1). For the example run, set this to 'gs://t5-data/pretrained_models/cbqa/small_ssm_nq/model.ckpt-1110000'.
  • MIXTURE_OR_TASK_NAME: This is the SeqIO Task or Mixture name to run eval on (from Step 2). For the example run, set this to 'natural_questions_open'.
  • EVAL_OUTPUT_DIR: A path to write eval outputs to. When launching using XManager, this path is automatically set and can be accessed from the XManager Artifacts page. When running locally using Blaze, you can explicitly pass a directory using a flag. Launch commands are provided in the next step.

In addition to the above params, you will need to import eval.gin and the Gin file for the model, which for the example run is t5_1_1_small.gin.

include 'runs/eval.gin'
include 'models/t5_small.gin'

Note that the include statements use relative paths in this example. You will pass an appropriate gin_search_paths flag to locate these files when launching your run. Absolute paths to Gin files can also be used, e.g.

include 't5x/configs/runs/eval.gin'
include 't5x/google/examples/flaxformer_t5/configs/models/t5_1_1_small.gin'

You will also need to import the Python module(s) that register SeqIO Tasks and Mixtures used in your run. For the example run, we add import google_research.t5_closed_book_qa.t5_cbqa.tasks since it is where 'glue_v002_proportional' is registered.

If you choose a module that is not included as a dependency in the T5X trainer binary, or if you have defined your gin config file in a location other than the T5X config directory, you will need to follow the instructions in the Advanced Topics section to link in the custom gin file and/or task definition.

Note that for most common Task/Mixtures, such as the glue_v002_proportional used in this tutorial, the necessary modules are already included. It is also possible to skip writing a Gin file and instead pass the params as flags when launching the eval job (see instructions in Step 4).

Finally, your Gin file should look like this:

include 't5x/configs/runs/eval.gin'
include 't5x/google/examples/flaxformer_t5/configs/models/t5_1_1_small.gin'

# Register necessary SeqIO Tasks/Mixtures.
import google_research.t5_closed_book_qa.t5_cbqa.tasks

CHECKPOINT_PATH = 'gs://t5-data/pretrained_models/cbqa/small_ssm_nq/model.ckpt-1110000'
MIXTURE_OR_TASK_NAME = 'natural_questions_open'

See t5_1_1_small_cbqa_natural_questions.gin for this example.

In this example, we run the evaluation on one checkpoint. It is common to evaluate with multiple checkpoints. We provide an easy way to do so without having to recompile the model graph for each checkpoints. This is simply done by adding utils.RestoreCheckpointConfig.mode = "all" to a gin file. Our t5x/configs/runs/eval.gin uses "specific" mode.

Step 4: Launch your experiment

To launch your experiment locally (for debugging only; larger checkpoints may cause issues), run the following on commandline:

EVAL_OUTPUT_DIR="/tmp/model-eval/"
python -m t5x.eval \
  --gin_file=t5x/google/examples/flaxformer_t5/configs/examples/eval/t5_1_1_small_cbqa_natural_questions.gin \
  --gin.EVAL_OUTPUT_DIR=\"${EVAL_OUTPUT_DIR}\" \
  --alsologtostderr

Note that relative paths can be used to locate the gin files. For that, multiple comma-separated paths can be passed to the gin_search_paths flag, and these paths should contain all Gin files used or included in your experiment.

You can have a look inside eval.gin to see other useful parameters that it is possible to pass in, including dataset split, batch size, and random seed.

Step 5: Monitor your experiment and parse metrics

After evaluation has completed, you can parse metrics into CSV format using the following script:

EVAL_OUTPUT_DIR= # from Step 4 if running locally, from XManager Artifacts otherwise
VAL_DIR="$EVAL_OUTPUT_DIR/inference_eval"
python -m t5.scripts.parse_tb \
  --summary_dir="$VAL_DIR" \
  --seqio_summaries \
  --out_file="$VAL_DIR/results.csv" \
  --alsologtostderr

Next Steps

Now that you have successfully evaluated a model on the Natural Questions benchmark, here are some topics you might want to explore next:

We also touch upon a few advanced topics related to evaluations below that might be useful, especially when customizing your eval job.

Advanced Topics

Defining a custom SeqIO Task/Mixture to evaluate on {.no-toc}

Refer to SeqIO documentation.

Defining a custom metric to evaluate

The best way to define a custom metric is to define a new SeqIO Task/Mixture that contains this custom metric. Please refer to the SeqIO Documentation on custom metrics.