Fine-tune a πŸ€— Transformers modelΒΆ

This notebook is based on an official πŸ€— notebook - β€œHow to fine-tune a model on text classification”. The main aim of this notebook is to show the process of conversion from vanilla πŸ€— to Ray AIR πŸ€— without changing the training logic unless necessary.

In this notebook, we will:

  1. Set up Ray

  2. Load the dataset

  3. Preprocess the dataset with Ray AIR

  4. Run the training with Ray AIR

  5. Predict on test data with Ray AIR

  6. Optionally, share the model with the community

Uncomment and run the following line in order to install all the necessary dependencies (this notebook is being tested with transformers==4.19.1):

#! pip install "datasets" "transformers>=4.19.0" "torch>=1.10.0" "mlflow" "ray[air]>=1.13"

Set up Ray ΒΆ

We will use ray.init() to initialize a local cluster. By default, this cluster will be compromised of only the machine you are running this notebook on. You can also run this notebook on an Anyscale cluster.

Note: this notebook will not run in Ray Client mode.

from pprint import pprint
import ray

ray.init()
2022-08-25 10:09:51,282	INFO worker.py:1223 -- Using address localhost:9031 set in the environment variable RAY_ADDRESS
2022-08-25 10:09:51,697	INFO worker.py:1333 -- Connecting to existing Ray cluster at address: 172.31.80.117:9031...
2022-08-25 10:09:51,706	INFO worker.py:1509 -- Connected to Ray cluster. View the dashboard at https://session-i8ddtfaxhwypbvnyb9uzg7xs.i.anyscaleuserdata-staging.com/auth/?token=agh0_CkcwRQIhAJXwvxwq31GryaWthvXGCXZebsijbuqi7qL2pCa5uROOAiBGjzsyXAJFHLlaEI9zSlNI8ewtghKg5UV3t8NmlxuMcRJmEiCtvjcKE0VPiU7iQx51P9oPQjfpo5g1RJXccVSS5005cBgCIgNuL2E6DAj9xazjBhDwj4veAUIMCP3ClJgGEPCPi94B-gEeChxzZXNfaThERFRmQVhId1lwYlZueWI5dVpnN3hT&redirect_to=dashboard 
2022-08-25 10:09:51,709	INFO packaging.py:342 -- Pushing file package 'gcs://_ray_pkg_3332f64b0a461fddc20be71129115d0a.zip' (0.34MiB) to Ray cluster...
2022-08-25 10:09:51,714	INFO packaging.py:351 -- Successfully pushed file package 'gcs://_ray_pkg_3332f64b0a461fddc20be71129115d0a.zip'.

We can check the resources our cluster is composed of. If you are running this notebook on your local machine or Google Colab, you should see the number of CPU cores and GPUs available on the said machine.

pprint(ray.cluster_resources())
{'CPU': 208.0,
 'GPU': 16.0,
 'accelerator_type:T4': 4.0,
 'memory': 616693614180.0,
 'node:172.31.76.237': 1.0,
 'node:172.31.80.117': 1.0,
 'node:172.31.85.193': 1.0,
 'node:172.31.85.32': 1.0,
 'node:172.31.90.137': 1.0,
 'object_store_memory': 259318055729.0}

In this notebook, we will see how to fine-tune one of the πŸ€— Transformers model to a text classification task of the GLUE Benchmark. We will be running the training using Ray AIR.

You can change those two variables to control whether the training (which we will get to later) uses CPUs or GPUs, and how many workers should be spawned. Each worker will claim one CPU or GPU. Make sure not to request more resources than the resources present!

By default, we will run the training with one GPU worker.

use_gpu = True  # set this to False to run on CPUs
num_workers = 1  # set this to number of GPUs/CPUs you want to use

Fine-tuning a model on a text classification taskΒΆ

The GLUE Benchmark is a group of nine classification tasks on sentences or pairs of sentences. If you would like to learn more, refer to the original notebook.

Each task is named by its acronym, with mnli-mm standing for the mismatched version of MNLI (so same training set as mnli but different validation and test sets):

GLUE_TASKS = ["cola", "mnli", "mnli-mm", "mrpc", "qnli", "qqp", "rte", "sst2", "stsb", "wnli"]

This notebook is built to run on any of the tasks in the list above, with any model checkpoint from the Model Hub as long as that model has a version with a classification head. Depending on your model and the GPU you are using, you might need to adjust the batch size to avoid out-of-memory errors. Set those three parameters, then the rest of the notebook should run smoothly:

task = "cola"
model_checkpoint = "distilbert-base-uncased"
batch_size = 16

Loading the dataset ΒΆ

We will use the πŸ€— Datasets library to download the data and get the metric we need to use for evaluation (to compare our model to the benchmark). This can be easily done with the functions load_dataset and load_metric.

Apart from mnli-mm being a special code, we can directly pass our task name to those functions.

As Ray AIR doesn’t provide integrations for πŸ€— Datasets yet, we will simply run the normal πŸ€— Datasets code to load the dataset from the Hub.

from datasets import load_dataset

actual_task = "mnli" if task == "mnli-mm" else task
datasets = load_dataset("glue", actual_task)

The dataset object itself is DatasetDict, which contains one key for the training, validation, and test set (with more keys for the mismatched validation and test set in the special case of mnli).

We will also need the metric. In order to avoid serialization errors, we will load the metric inside the training workers later. Therefore, now we will just define the function we will use.

from datasets import load_metric

def load_metric_fn():
    return load_metric('glue', actual_task)

The metric is an instance of datasets.Metric.

Preprocessing the data with Ray AIR ΒΆ

Before we can feed those texts to our model, we need to preprocess them. This is done by a πŸ€— Transformers’ Tokenizer, which will (as the name indicates) tokenize the inputs (including converting the tokens to their corresponding IDs in the pretrained vocabulary) and put it in a format the model expects, as well as generate the other inputs that model requires.

To do all of this, we instantiate our tokenizer with the AutoTokenizer.from_pretrained method, which will ensure that:

  • we get a tokenizer that corresponds to the model architecture we want to use,

  • we download the vocabulary used when pretraining this specific checkpoint.

from transformers import AutoTokenizer

tokenizer = AutoTokenizer.from_pretrained(model_checkpoint, use_fast=True)

We pass along use_fast=True to the call above to use one of the fast tokenizers (backed by Rust) from the πŸ€— Tokenizers library. Those fast tokenizers are available for almost all models, but if you got an error with the previous call, remove that argument.

To preprocess our dataset, we will thus need the names of the columns containing the sentence(s). The following dictionary keeps track of the correspondence task to column names:

task_to_keys = {
    "cola": ("sentence", None),
    "mnli": ("premise", "hypothesis"),
    "mnli-mm": ("premise", "hypothesis"),
    "mrpc": ("sentence1", "sentence2"),
    "qnli": ("question", "sentence"),
    "qqp": ("question1", "question2"),
    "rte": ("sentence1", "sentence2"),
    "sst2": ("sentence", None),
    "stsb": ("sentence1", "sentence2"),
    "wnli": ("sentence1", "sentence2"),
}

For Ray AIR, instead of using πŸ€— Dataset objects directly, we will convert them to Ray Datasets. Both are backed by Arrow tables, so the conversion is straightforward. We will use the built-in ray.data.from_huggingface function.

import ray.data

ray_datasets = ray.data.from_huggingface(datasets)
ray_datasets
{'train': Dataset(num_blocks=1, num_rows=8551, schema={sentence: string, label: int64, idx: int32}),
 'validation': Dataset(num_blocks=1, num_rows=1043, schema={sentence: string, label: int64, idx: int32}),
 'test': Dataset(num_blocks=1, num_rows=1063, schema={sentence: string, label: int64, idx: int32})}

We can then write the function that will preprocess our samples. We just feed them to the tokenizer with the argument truncation=True. This will ensure that an input longer than what the model selected can handle will be truncated to the maximum length accepted by the model.

We use a BatchMapper to create a Ray AIR preprocessor that will map the function to the dataset in a distributed fashion. It will run during training and prediction.

import pandas as pd
from ray.data.preprocessors import BatchMapper

def preprocess_function(examples: pd.DataFrame):
    # if we only have one column, we are inferring.
    # no need to tokenize in that case. 
    if len(examples.columns) == 1:
        return examples
    examples = examples.to_dict("list")
    sentence1_key, sentence2_key = task_to_keys[task]
    if sentence2_key is None:
        ret = tokenizer(examples[sentence1_key], truncation=True)
    else:
        ret = tokenizer(examples[sentence1_key], examples[sentence2_key], truncation=True)
    # Add back the original columns
    ret = {**examples, **ret}
    return pd.DataFrame.from_dict(ret)

batch_encoder = BatchMapper(preprocess_function, batch_format="pandas")

Fine-tuning the model with Ray AIR ΒΆ

Now that our data is ready, we can download the pretrained model and fine-tune it.

Since all our tasks are about sentence classification, we use the AutoModelForSequenceClassification class.

We will not go into details about each specific component of the training (see the original notebook for that). The tokenizer is the same as we have used to encoded the dataset before.

The main difference when using the Ray AIR is that we need to create our πŸ€— Transformers Trainer inside a function (trainer_init_per_worker) and return it. That function will be passed to the HuggingFaceTrainer and will run on every Ray worker. The training will then proceed by the means of PyTorch DDP.

Make sure that you initialize the model, metric, and tokenizer inside that function. Otherwise, you may run into serialization errors.

Furthermore, push_to_hub=True is not yet supported. Ray will, however, checkpoint the model at every epoch, allowing you to push it to hub manually. We will do that after the training.

If you wish to use thrid party logging libraries, such as MLflow or Weights&Biases, do not set them in TrainingArguments (they will be automatically disabled) - instead, you should pass Ray AIR callbacks to HuggingFaceTrainer’s run_config. In this example, we will use MLflow.

from transformers import AutoModelForSequenceClassification, TrainingArguments, Trainer
import numpy as np
import torch

num_labels = 3 if task.startswith("mnli") else 1 if task=="stsb" else 2
metric_name = "pearson" if task == "stsb" else "matthews_correlation" if task == "cola" else "accuracy"
model_name = model_checkpoint.split("/")[-1]
validation_key = "validation_mismatched" if task == "mnli-mm" else "validation_matched" if task == "mnli" else "validation"
name = f"{model_name}-finetuned-{task}"

def trainer_init_per_worker(train_dataset, eval_dataset = None, **config):
    print(f"Is CUDA available: {torch.cuda.is_available()}")
    metric = load_metric_fn()
    tokenizer = AutoTokenizer.from_pretrained(model_checkpoint, use_fast=True)
    model = AutoModelForSequenceClassification.from_pretrained(model_checkpoint, num_labels=num_labels)
    args = TrainingArguments(
        name,
        evaluation_strategy="epoch",
        save_strategy="epoch",
        logging_strategy="epoch",
        learning_rate=config.get("learning_rate", 2e-5),
        per_device_train_batch_size=batch_size,
        per_device_eval_batch_size=batch_size,
        num_train_epochs=config.get("epochs", 2),
        weight_decay=config.get("weight_decay", 0.01),
        push_to_hub=False,
        disable_tqdm=True,  # declutter the output a little
        no_cuda=not use_gpu,  # you need to explicitly set no_cuda if you want CPUs
    )

    def compute_metrics(eval_pred):
        predictions, labels = eval_pred
        if task != "stsb":
            predictions = np.argmax(predictions, axis=1)
        else:
            predictions = predictions[:, 0]
        return metric.compute(predictions=predictions, references=labels)

    trainer = Trainer(
        model,
        args,
        train_dataset=train_dataset,
        eval_dataset=eval_dataset,
        tokenizer=tokenizer,
        compute_metrics=compute_metrics
    )

    print("Starting training")
    return trainer

With our trainer_init_per_worker complete, we can now instantiate the HuggingFaceTrainer. Aside from the function, we set the scaling_config, controlling the amount of workers and resources used, and the datasets we will use for training and evaluation.

We specify the MlflowLoggerCallback inside the run_config, and pass the preprocessor we have defined earlier as an argument. The preprocessor will be included with the returned Checkpoint, meaning it will also be applied during inference.

from ray.train.huggingface import HuggingFaceTrainer
from ray.air.config import RunConfig, ScalingConfig, CheckpointConfig
from ray.air.integrations.mlflow import MLflowLoggerCallback

trainer = HuggingFaceTrainer(
    trainer_init_per_worker=trainer_init_per_worker,
    scaling_config=ScalingConfig(num_workers=num_workers, use_gpu=use_gpu),
    datasets={"train": ray_datasets["train"], "evaluation": ray_datasets[validation_key]},
    run_config=RunConfig(
        callbacks=[MLflowLoggerCallback(experiment_name=name)],
        checkpoint_config=CheckpointConfig(num_to_keep=1, checkpoint_score_attribute="eval_loss", checkpoint_score_order="min"),
    ),
    preprocessor=batch_encoder,
)

Finally, we call the fit method to start training with Ray AIR. We will save the Result object to a variable so we can access metrics and checkpoints.

result = trainer.fit()
== Status ==
Current time: 2022-08-25 10:14:09 (running for 00:04:06.45)
Memory usage on this node: 4.3/62.0 GiB
Using FIFO scheduling algorithm.
Resources requested: 0/208 CPUs, 0/16 GPUs, 0.0/574.34 GiB heap, 0.0/241.51 GiB objects (0.0/4.0 accelerator_type:T4)
Result logdir: /home/ray/ray_results/HuggingFaceTrainer_2022-08-25_10-10-02
Number of trials: 1/1 (1 TERMINATED)
Trial name status loc iter total time (s) loss learning_rate epoch
HuggingFaceTrainer_c1ff5_00000TERMINATED172.31.90.137:947 2 200.2170.3886 0 2


(RayTrainWorker pid=1114, ip=172.31.90.137) 2022-08-25 10:10:44,617	INFO config.py:71 -- Setting up process group for: env:// [rank=0, world_size=4]
(RayTrainWorker pid=1114, ip=172.31.90.137) Is CUDA available: True
(RayTrainWorker pid=1116, ip=172.31.90.137) Is CUDA available: True
(RayTrainWorker pid=1117, ip=172.31.90.137) Is CUDA available: True
(RayTrainWorker pid=1115, ip=172.31.90.137) Is CUDA available: True
Downloading builder script: 5.76kB [00:00, 6.45MB/s]                   
Downloading builder script: 5.76kB [00:00, 6.91MB/s]                   
Downloading builder script: 5.76kB [00:00, 6.44MB/s]                   
Downloading builder script: 5.76kB [00:00, 6.94MB/s]                   
Downloading tokenizer_config.json: 100%|β–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆ| 28.0/28.0 [00:00<00:00, 30.5kB/s]
Downloading config.json: 100%|β–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆ| 483/483 [00:00<00:00, 817kB/s]
Downloading vocab.txt:   0%|          | 0.00/226k [00:00<?, ?B/s]
Downloading vocab.txt:  18%|β–ˆβ–Š        | 41.0k/226k [00:00<00:00, 353kB/s]
Downloading vocab.txt: 100%|β–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆ| 226k/226k [00:00<00:00, 773kB/s] 
Downloading tokenizer.json:   0%|          | 0.00/455k [00:00<?, ?B/s]
Downloading tokenizer.json:   6%|β–Œ         | 28.0k/455k [00:00<00:01, 227kB/s]
Downloading tokenizer.json:  24%|β–ˆβ–ˆβ–       | 111k/455k [00:00<00:00, 488kB/s] 
Downloading tokenizer.json:  42%|β–ˆβ–ˆβ–ˆβ–ˆβ–     | 191k/455k [00:00<00:00, 559kB/s]
Downloading tokenizer.json:  67%|β–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–‹   | 303k/455k [00:00<00:00, 694kB/s]
Downloading tokenizer.json: 100%|β–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆ| 455k/455k [00:00<00:00, 815kB/s]
Downloading pytorch_model.bin:   0%|          | 0.00/256M [00:00<?, ?B/s]
Downloading pytorch_model.bin:   0%|          | 1.20M/256M [00:00<00:21, 12.6MB/s]
Downloading pytorch_model.bin:   2%|▏         | 6.02M/256M [00:00<00:07, 34.9MB/s]
Downloading pytorch_model.bin:   6%|β–Œ         | 15.0M/256M [00:00<00:04, 62.0MB/s]
Downloading pytorch_model.bin:   9%|β–‰         | 24.0M/256M [00:00<00:03, 74.8MB/s]
Downloading pytorch_model.bin:  13%|β–ˆβ–Ž        | 33.1M/256M [00:00<00:02, 82.3MB/s]
Downloading pytorch_model.bin:  17%|β–ˆβ–‹        | 42.2M/256M [00:00<00:02, 86.7MB/s]
Downloading pytorch_model.bin:  20%|β–ˆβ–ˆ        | 51.4M/256M [00:00<00:02, 89.8MB/s]
Downloading pytorch_model.bin:  24%|β–ˆβ–ˆβ–Ž       | 60.6M/256M [00:00<00:02, 91.8MB/s]
Downloading pytorch_model.bin:  27%|β–ˆβ–ˆβ–‹       | 69.8M/256M [00:00<00:02, 93.3MB/s]
Downloading pytorch_model.bin:  31%|β–ˆβ–ˆβ–ˆ       | 78.9M/256M [00:01<00:01, 94.2MB/s]
Downloading pytorch_model.bin:  34%|β–ˆβ–ˆβ–ˆβ–      | 88.0M/256M [00:01<00:01, 94.6MB/s]
Downloading pytorch_model.bin:  38%|β–ˆβ–ˆβ–ˆβ–Š      | 97.2M/256M [00:01<00:01, 95.1MB/s]
Downloading pytorch_model.bin:  42%|β–ˆβ–ˆβ–ˆβ–ˆβ–     | 106M/256M [00:01<00:01, 95.6MB/s] 
Downloading pytorch_model.bin:  45%|β–ˆβ–ˆβ–ˆβ–ˆβ–Œ     | 116M/256M [00:01<00:01, 96.0MB/s]
Downloading pytorch_model.bin:  49%|β–ˆβ–ˆβ–ˆβ–ˆβ–‰     | 125M/256M [00:01<00:01, 96.2MB/s]
Downloading pytorch_model.bin:  52%|β–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–    | 134M/256M [00:01<00:01, 96.0MB/s]
Downloading pytorch_model.bin:  56%|β–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–Œ    | 143M/256M [00:01<00:01, 96.1MB/s]
Downloading pytorch_model.bin:  60%|β–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–‰    | 152M/256M [00:01<00:01, 96.0MB/s]
Downloading pytorch_model.bin:  63%|β–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–Ž   | 162M/256M [00:01<00:01, 96.2MB/s]
Downloading pytorch_model.bin:  67%|β–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–‹   | 171M/256M [00:02<00:00, 96.1MB/s]
Downloading pytorch_model.bin:  70%|β–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆ   | 180M/256M [00:02<00:00, 96.2MB/s]
Downloading pytorch_model.bin:  74%|β–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–  | 189M/256M [00:02<00:00, 96.2MB/s]
Downloading pytorch_model.bin:  78%|β–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–Š  | 198M/256M [00:02<00:00, 96.2MB/s]
Downloading pytorch_model.bin:  81%|β–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆ  | 208M/256M [00:02<00:00, 95.9MB/s]
Downloading pytorch_model.bin:  85%|β–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ– | 217M/256M [00:02<00:00, 95.9MB/s]
Downloading pytorch_model.bin:  88%|β–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–Š | 226M/256M [00:02<00:00, 96.2MB/s]
Downloading pytorch_model.bin:  92%|β–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–| 235M/256M [00:02<00:00, 96.1MB/s]
Downloading pytorch_model.bin:  96%|β–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–Œ| 244M/256M [00:02<00:00, 96.1MB/s]
Downloading pytorch_model.bin: 100%|β–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆ| 256M/256M [00:02<00:00, 91.6MB/s]
(RayTrainWorker pid=1117, ip=172.31.90.137) Some weights of the model checkpoint at distilbert-base-uncased were not used when initializing DistilBertForSequenceClassification: ['vocab_projector.weight', 'vocab_transform.bias', 'vocab_layer_norm.weight', 'vocab_projector.bias', 'vocab_transform.weight', 'vocab_layer_norm.bias']
(RayTrainWorker pid=1117, ip=172.31.90.137) - This IS expected if you are initializing DistilBertForSequenceClassification from the checkpoint of a model trained on another task or with another architecture (e.g. initializing a BertForSequenceClassification model from a BertForPreTraining model).
(RayTrainWorker pid=1117, ip=172.31.90.137) - This IS NOT expected if you are initializing DistilBertForSequenceClassification from the checkpoint of a model that you expect to be exactly identical (initializing a BertForSequenceClassification model from a BertForSequenceClassification model).
(RayTrainWorker pid=1117, ip=172.31.90.137) Some weights of DistilBertForSequenceClassification were not initialized from the model checkpoint at distilbert-base-uncased and are newly initialized: ['pre_classifier.bias', 'classifier.bias', 'classifier.weight', 'pre_classifier.weight']
(RayTrainWorker pid=1117, ip=172.31.90.137) You should probably TRAIN this model on a down-stream task to be able to use it for predictions and inference.
(RayTrainWorker pid=1114, ip=172.31.90.137) Some weights of the model checkpoint at distilbert-base-uncased were not used when initializing DistilBertForSequenceClassification: ['vocab_layer_norm.weight', 'vocab_projector.bias', 'vocab_layer_norm.bias', 'vocab_transform.bias', 'vocab_projector.weight', 'vocab_transform.weight']
(RayTrainWorker pid=1114, ip=172.31.90.137) - This IS expected if you are initializing DistilBertForSequenceClassification from the checkpoint of a model trained on another task or with another architecture (e.g. initializing a BertForSequenceClassification model from a BertForPreTraining model).
(RayTrainWorker pid=1114, ip=172.31.90.137) - This IS NOT expected if you are initializing DistilBertForSequenceClassification from the checkpoint of a model that you expect to be exactly identical (initializing a BertForSequenceClassification model from a BertForSequenceClassification model).
(RayTrainWorker pid=1114, ip=172.31.90.137) Some weights of DistilBertForSequenceClassification were not initialized from the model checkpoint at distilbert-base-uncased and are newly initialized: ['pre_classifier.bias', 'pre_classifier.weight', 'classifier.weight', 'classifier.bias']
(RayTrainWorker pid=1114, ip=172.31.90.137) You should probably TRAIN this model on a down-stream task to be able to use it for predictions and inference.
(RayTrainWorker pid=1116, ip=172.31.90.137) Some weights of the model checkpoint at distilbert-base-uncased were not used when initializing DistilBertForSequenceClassification: ['vocab_layer_norm.bias', 'vocab_transform.bias', 'vocab_projector.bias', 'vocab_layer_norm.weight', 'vocab_transform.weight', 'vocab_projector.weight']
(RayTrainWorker pid=1116, ip=172.31.90.137) - This IS expected if you are initializing DistilBertForSequenceClassification from the checkpoint of a model trained on another task or with another architecture (e.g. initializing a BertForSequenceClassification model from a BertForPreTraining model).
(RayTrainWorker pid=1116, ip=172.31.90.137) - This IS NOT expected if you are initializing DistilBertForSequenceClassification from the checkpoint of a model that you expect to be exactly identical (initializing a BertForSequenceClassification model from a BertForSequenceClassification model).
(RayTrainWorker pid=1116, ip=172.31.90.137) Some weights of DistilBertForSequenceClassification were not initialized from the model checkpoint at distilbert-base-uncased and are newly initialized: ['classifier.bias', 'pre_classifier.weight', 'pre_classifier.bias', 'classifier.weight']
(RayTrainWorker pid=1116, ip=172.31.90.137) You should probably TRAIN this model on a down-stream task to be able to use it for predictions and inference.
(RayTrainWorker pid=1115, ip=172.31.90.137) Some weights of the model checkpoint at distilbert-base-uncased were not used when initializing DistilBertForSequenceClassification: ['vocab_projector.bias', 'vocab_projector.weight', 'vocab_transform.bias', 'vocab_layer_norm.bias', 'vocab_transform.weight', 'vocab_layer_norm.weight']
(RayTrainWorker pid=1115, ip=172.31.90.137) - This IS expected if you are initializing DistilBertForSequenceClassification from the checkpoint of a model trained on another task or with another architecture (e.g. initializing a BertForSequenceClassification model from a BertForPreTraining model).
(RayTrainWorker pid=1115, ip=172.31.90.137) - This IS NOT expected if you are initializing DistilBertForSequenceClassification from the checkpoint of a model that you expect to be exactly identical (initializing a BertForSequenceClassification model from a BertForSequenceClassification model).
(RayTrainWorker pid=1115, ip=172.31.90.137) Some weights of DistilBertForSequenceClassification were not initialized from the model checkpoint at distilbert-base-uncased and are newly initialized: ['pre_classifier.weight', 'classifier.weight', 'classifier.bias', 'pre_classifier.bias']
(RayTrainWorker pid=1115, ip=172.31.90.137) You should probably TRAIN this model on a down-stream task to be able to use it for predictions and inference.
(RayTrainWorker pid=1114, ip=172.31.90.137) Starting training
(RayTrainWorker pid=1116, ip=172.31.90.137) Starting training
(RayTrainWorker pid=1117, ip=172.31.90.137) Starting training
(RayTrainWorker pid=1115, ip=172.31.90.137) Starting training
(RayTrainWorker pid=1114, ip=172.31.90.137) ***** Running training *****
(RayTrainWorker pid=1114, ip=172.31.90.137)   Num examples = 8551
(RayTrainWorker pid=1114, ip=172.31.90.137)   Num Epochs = 2
(RayTrainWorker pid=1114, ip=172.31.90.137)   Instantaneous batch size per device = 16
(RayTrainWorker pid=1114, ip=172.31.90.137)   Total train batch size (w. parallel, distributed & accumulation) = 64
(RayTrainWorker pid=1114, ip=172.31.90.137)   Gradient Accumulation steps = 1
(RayTrainWorker pid=1114, ip=172.31.90.137)   Total optimization steps = 1070
(RayTrainWorker pid=1114, ip=172.31.90.137) The following columns in the training set don't have a corresponding argument in `DistilBertForSequenceClassification.forward` and have been ignored: sentence, idx. If sentence, idx are not expected by `DistilBertForSequenceClassification.forward`,  you can safely ignore this message.
(RayTrainWorker pid=1114, ip=172.31.90.137) {'loss': 0.5437, 'learning_rate': 1e-05, 'epoch': 1.0}
(RayTrainWorker pid=1114, ip=172.31.90.137) ***** Running Evaluation *****
(RayTrainWorker pid=1114, ip=172.31.90.137)   Num examples = 1043
(RayTrainWorker pid=1114, ip=172.31.90.137)   Batch size = 16
(RayTrainWorker pid=1114, ip=172.31.90.137) The following columns in the evaluation set don't have a corresponding argument in `DistilBertForSequenceClassification.forward` and have been ignored: sentence, idx. If sentence, idx are not expected by `DistilBertForSequenceClassification.forward`,  you can safely ignore this message.
(RayTrainWorker pid=1114, ip=172.31.90.137) {'eval_loss': 0.5794203281402588, 'eval_matthews_correlation': 0.3293676852500821, 'eval_runtime': 0.9804, 'eval_samples_per_second': 277.441, 'eval_steps_per_second': 5.1, 'epoch': 1.0}
(RayTrainWorker pid=1114, ip=172.31.90.137) Saving model checkpoint to distilbert-base-uncased-finetuned-cola/checkpoint-535
(RayTrainWorker pid=1114, ip=172.31.90.137) Configuration saved in distilbert-base-uncased-finetuned-cola/checkpoint-535/config.json
(RayTrainWorker pid=1114, ip=172.31.90.137) Model weights saved in distilbert-base-uncased-finetuned-cola/checkpoint-535/pytorch_model.bin
(RayTrainWorker pid=1114, ip=172.31.90.137) tokenizer config file saved in distilbert-base-uncased-finetuned-cola/checkpoint-535/tokenizer_config.json
(RayTrainWorker pid=1114, ip=172.31.90.137) Special tokens file saved in distilbert-base-uncased-finetuned-cola/checkpoint-535/special_tokens_map.json
Result for HuggingFaceTrainer_c1ff5_00000:
  _time_this_iter_s: 90.87123560905457
  _timestamp: 1661447540
  _training_iteration: 1
  date: 2022-08-25_10-12-20
  done: false
  epoch: 1.0
  eval_loss: 0.5794203281402588
  eval_matthews_correlation: 0.3293676852500821
  eval_runtime: 0.9804
  eval_samples_per_second: 277.441
  eval_steps_per_second: 5.1
  experiment_id: 592e02b25b254bd1a3743904313dc85b
  hostname: ip-172-31-90-137
  iterations_since_restore: 1
  learning_rate: 1.0e-05
  loss: 0.5437
  node_ip: 172.31.90.137
  pid: 947
  should_checkpoint: true
  step: 535
  time_since_restore: 103.24057936668396
  time_this_iter_s: 103.24057936668396
  time_total_s: 103.24057936668396
  timestamp: 1661447540
  timesteps_since_restore: 0
  training_iteration: 1
  trial_id: c1ff5_00000
  warmup_time: 0.003858327865600586
  
(RayTrainWorker pid=1114, ip=172.31.90.137) Saving model checkpoint to distilbert-base-uncased-finetuned-cola/checkpoint-1070
(RayTrainWorker pid=1114, ip=172.31.90.137) Configuration saved in distilbert-base-uncased-finetuned-cola/checkpoint-1070/config.json
(RayTrainWorker pid=1114, ip=172.31.90.137) Model weights saved in distilbert-base-uncased-finetuned-cola/checkpoint-1070/pytorch_model.bin
(RayTrainWorker pid=1114, ip=172.31.90.137) tokenizer config file saved in distilbert-base-uncased-finetuned-cola/checkpoint-1070/tokenizer_config.json
(RayTrainWorker pid=1114, ip=172.31.90.137) Special tokens file saved in distilbert-base-uncased-finetuned-cola/checkpoint-1070/special_tokens_map.json
(RayTrainWorker pid=1114, ip=172.31.90.137) {'loss': 0.3886, 'learning_rate': 0.0, 'epoch': 2.0}
(RayTrainWorker pid=1114, ip=172.31.90.137) ***** Running Evaluation *****
(RayTrainWorker pid=1114, ip=172.31.90.137)   Num examples = 1043
(RayTrainWorker pid=1114, ip=172.31.90.137)   Batch size = 16
(RayTrainWorker pid=1114, ip=172.31.90.137) The following columns in the evaluation set don't have a corresponding argument in `DistilBertForSequenceClassification.forward` and have been ignored: sentence, idx. If sentence, idx are not expected by `DistilBertForSequenceClassification.forward`,  you can safely ignore this message.
(RayTrainWorker pid=1114, ip=172.31.90.137) {'eval_loss': 0.6215357184410095, 'eval_matthews_correlation': 0.42957017514952434, 'eval_runtime': 0.9956, 'eval_samples_per_second': 273.204, 'eval_steps_per_second': 5.022, 'epoch': 2.0}
(RayTrainWorker pid=1114, ip=172.31.90.137) Saving model checkpoint to distilbert-base-uncased-finetuned-cola/checkpoint-1070
(RayTrainWorker pid=1114, ip=172.31.90.137) Configuration saved in distilbert-base-uncased-finetuned-cola/checkpoint-1070/config.json
(RayTrainWorker pid=1114, ip=172.31.90.137) Model weights saved in distilbert-base-uncased-finetuned-cola/checkpoint-1070/pytorch_model.bin
(RayTrainWorker pid=1114, ip=172.31.90.137) tokenizer config file saved in distilbert-base-uncased-finetuned-cola/checkpoint-1070/tokenizer_config.json
(RayTrainWorker pid=1114, ip=172.31.90.137) Special tokens file saved in distilbert-base-uncased-finetuned-cola/checkpoint-1070/special_tokens_map.json
(RayTrainWorker pid=1114, ip=172.31.90.137) {'train_runtime': 174.4696, 'train_samples_per_second': 98.023, 'train_steps_per_second': 6.133, 'train_loss': 0.4661755713346963, 'epoch': 2.0}
(RayTrainWorker pid=1114, ip=172.31.90.137) 
(RayTrainWorker pid=1114, ip=172.31.90.137) 
(RayTrainWorker pid=1114, ip=172.31.90.137) Training completed. Do not forget to share your model on huggingface.co/models =)
(RayTrainWorker pid=1114, ip=172.31.90.137) 
(RayTrainWorker pid=1114, ip=172.31.90.137) 
Result for HuggingFaceTrainer_c1ff5_00000:
  _time_this_iter_s: 96.96447467803955
  _timestamp: 1661447637
  _training_iteration: 2
  date: 2022-08-25_10-13-57
  done: false
  epoch: 2.0
  eval_loss: 0.6215357184410095
  eval_matthews_correlation: 0.42957017514952434
  eval_runtime: 0.9956
  eval_samples_per_second: 273.204
  eval_steps_per_second: 5.022
  experiment_id: 592e02b25b254bd1a3743904313dc85b
  hostname: ip-172-31-90-137
  iterations_since_restore: 2
  learning_rate: 0.0
  loss: 0.3886
  node_ip: 172.31.90.137
  pid: 947
  should_checkpoint: true
  step: 1070
  time_since_restore: 200.21722102165222
  time_this_iter_s: 96.97664165496826
  time_total_s: 200.21722102165222
  timestamp: 1661447637
  timesteps_since_restore: 0
  train_loss: 0.4661755713346963
  train_runtime: 174.4696
  train_samples_per_second: 98.023
  train_steps_per_second: 6.133
  training_iteration: 2
  trial_id: c1ff5_00000
  warmup_time: 0.003858327865600586
  
Result for HuggingFaceTrainer_c1ff5_00000:
  _time_this_iter_s: 96.96447467803955
  _timestamp: 1661447637
  _training_iteration: 2
  date: 2022-08-25_10-13-57
  done: true
  epoch: 2.0
  eval_loss: 0.6215357184410095
  eval_matthews_correlation: 0.42957017514952434
  eval_runtime: 0.9956
  eval_samples_per_second: 273.204
  eval_steps_per_second: 5.022
  experiment_id: 592e02b25b254bd1a3743904313dc85b
  experiment_tag: '0'
  hostname: ip-172-31-90-137
  iterations_since_restore: 2
  learning_rate: 0.0
  loss: 0.3886
  node_ip: 172.31.90.137
  pid: 947
  should_checkpoint: true
  step: 1070
  time_since_restore: 200.21722102165222
  time_this_iter_s: 96.97664165496826
  time_total_s: 200.21722102165222
  timestamp: 1661447637
  timesteps_since_restore: 0
  train_loss: 0.4661755713346963
  train_runtime: 174.4696
  train_samples_per_second: 98.023
  train_steps_per_second: 6.133
  training_iteration: 2
  trial_id: c1ff5_00000
  warmup_time: 0.003858327865600586
  
2022-08-25 10:14:09,300	INFO tune.py:758 -- Total run time: 246.67 seconds (246.44 seconds for the tuning loop).

You can use the returned Result object to access metrics and the Ray AIR Checkpoint associated with the last iteration.

result
Result(metrics={'loss': 0.3886, 'learning_rate': 0.0, 'epoch': 2.0, 'step': 1070, 'eval_loss': 0.6215357184410095, 'eval_matthews_correlation': 0.42957017514952434, 'eval_runtime': 0.9956, 'eval_samples_per_second': 273.204, 'eval_steps_per_second': 5.022, 'train_runtime': 174.4696, 'train_samples_per_second': 98.023, 'train_steps_per_second': 6.133, 'train_loss': 0.4661755713346963, '_timestamp': 1661447637, '_time_this_iter_s': 96.96447467803955, '_training_iteration': 2, 'should_checkpoint': True, 'done': True, 'trial_id': 'c1ff5_00000', 'experiment_tag': '0'}, error=None, log_dir=PosixPath('/home/ray/ray_results/HuggingFaceTrainer_2022-08-25_10-10-02/HuggingFaceTrainer_c1ff5_00000_0_2022-08-25_10-10-04'))

Tune hyperparameters with Ray AIR ΒΆ

If we would like to tune any hyperparameters of the model, we can do so by simply passing our HuggingFaceTrainer into a Tuner and defining the search space.

We can also take advantage of the advanced search algorithms and schedulers provided by Ray Tune. In this example, we will use an ASHAScheduler to aggresively terminate underperforming trials.

from ray import tune
from ray.tune import Tuner
from ray.tune.schedulers.async_hyperband import ASHAScheduler

tune_epochs = 4
tuner = Tuner(
    trainer,
    param_space={
        "trainer_init_config": {
            "learning_rate": tune.grid_search([2e-5, 2e-4, 2e-3, 2e-2]),
            "epochs": tune_epochs,
        }
    },
    tune_config=tune.TuneConfig(
        metric="eval_loss",
        mode="min",
        num_samples=1,
        scheduler=ASHAScheduler(
            max_t=tune_epochs,
        )
    ),
    run_config=RunConfig(
        checkpoint_config=CheckpointConfig(num_to_keep=1, checkpoint_score_attribute="eval_loss", checkpoint_score_order="min")
    ),
)
tune_results = tuner.fit()
== Status ==
Current time: 2022-08-25 10:20:13 (running for 00:06:01.75)
Memory usage on this node: 4.4/62.0 GiB
Using AsyncHyperBand: num_stopped=4 Bracket: Iter 4.000: -0.8064090609550476 | Iter 1.000: -0.6378736793994904
Resources requested: 0/208 CPUs, 0/16 GPUs, 0.0/574.34 GiB heap, 0.0/241.51 GiB objects (0.0/4.0 accelerator_type:T4)
Current best trial: 5654d_00001 with eval_loss=0.6492420434951782 and parameters={'trainer_init_config': {'learning_rate': 0.0002, 'epochs': 4}}
Result logdir: /home/ray/ray_results/HuggingFaceTrainer_2022-08-25_10-14-11
Number of trials: 4/4 (4 TERMINATED)
Trial name status loc trainer_init_conf... iter total time (s) loss learning_rate epoch
HuggingFaceTrainer_5654d_00000TERMINATED172.31.90.137:1729 2e-05 4 347.171 0.1958 0 4
HuggingFaceTrainer_5654d_00001TERMINATED172.31.76.237:1805 0.0002 1 95.24920.6225 0.00015 1
HuggingFaceTrainer_5654d_00002TERMINATED172.31.85.32:1322 0.002 1 93.76130.6463 0.0015 1
HuggingFaceTrainer_5654d_00003TERMINATED172.31.85.193:1060 0.02 1 99.36770.926 0.015 1


(RayTrainWorker pid=1789, ip=172.31.90.137) 2022-08-25 10:14:23,379	INFO config.py:71 -- Setting up process group for: env:// [rank=0, world_size=4]
(RayTrainWorker pid=1792, ip=172.31.90.137) Is CUDA available: True
(RayTrainWorker pid=1790, ip=172.31.90.137) Is CUDA available: True
(RayTrainWorker pid=1791, ip=172.31.90.137) Is CUDA available: True
(RayTrainWorker pid=1789, ip=172.31.90.137) Is CUDA available: True
(RayTrainWorker pid=1974, ip=172.31.76.237) 2022-08-25 10:14:29,354	INFO config.py:71 -- Setting up process group for: env:// [rank=0, world_size=4]
(RayTrainWorker pid=1977, ip=172.31.76.237) Is CUDA available: True
(RayTrainWorker pid=1976, ip=172.31.76.237) Is CUDA available: True
(RayTrainWorker pid=1975, ip=172.31.76.237) Is CUDA available: True
(RayTrainWorker pid=1974, ip=172.31.76.237) Is CUDA available: True
(RayTrainWorker pid=1483, ip=172.31.85.32) 2022-08-25 10:14:35,313	INFO config.py:71 -- Setting up process group for: env:// [rank=0, world_size=4]
(RayTrainWorker pid=1790, ip=172.31.90.137) Starting training
(RayTrainWorker pid=1792, ip=172.31.90.137) Starting training
(RayTrainWorker pid=1791, ip=172.31.90.137) Starting training
(RayTrainWorker pid=1789, ip=172.31.90.137) Starting training
(RayTrainWorker pid=1789, ip=172.31.90.137) ***** Running training *****
(RayTrainWorker pid=1789, ip=172.31.90.137)   Num examples = 8551
(RayTrainWorker pid=1789, ip=172.31.90.137)   Num Epochs = 4
(RayTrainWorker pid=1789, ip=172.31.90.137)   Instantaneous batch size per device = 16
(RayTrainWorker pid=1789, ip=172.31.90.137)   Total train batch size (w. parallel, distributed & accumulation) = 64
(RayTrainWorker pid=1789, ip=172.31.90.137)   Gradient Accumulation steps = 1
(RayTrainWorker pid=1789, ip=172.31.90.137)   Total optimization steps = 2140
(RayTrainWorker pid=1483, ip=172.31.85.32) Is CUDA available: True
(RayTrainWorker pid=1485, ip=172.31.85.32) Is CUDA available: True
(RayTrainWorker pid=1486, ip=172.31.85.32) Is CUDA available: True
(RayTrainWorker pid=1484, ip=172.31.85.32) Is CUDA available: True
(RayTrainWorker pid=1977, ip=172.31.76.237) Starting training
(RayTrainWorker pid=1976, ip=172.31.76.237) Starting training
(RayTrainWorker pid=1975, ip=172.31.76.237) Starting training
(RayTrainWorker pid=1974, ip=172.31.76.237) Starting training
(RayTrainWorker pid=1974, ip=172.31.76.237) ***** Running training *****
(RayTrainWorker pid=1974, ip=172.31.76.237)   Num examples = 8551
(RayTrainWorker pid=1974, ip=172.31.76.237)   Num Epochs = 4
(RayTrainWorker pid=1974, ip=172.31.76.237)   Instantaneous batch size per device = 16
(RayTrainWorker pid=1974, ip=172.31.76.237)   Total train batch size (w. parallel, distributed & accumulation) = 64
(RayTrainWorker pid=1974, ip=172.31.76.237)   Gradient Accumulation steps = 1
(RayTrainWorker pid=1974, ip=172.31.76.237)   Total optimization steps = 2140
(RayTrainWorker pid=1483, ip=172.31.85.32) Starting training
(RayTrainWorker pid=1485, ip=172.31.85.32) Starting training
(RayTrainWorker pid=1486, ip=172.31.85.32) Starting training
(RayTrainWorker pid=1484, ip=172.31.85.32) Starting training
(RayTrainWorker pid=1483, ip=172.31.85.32) ***** Running training *****
(RayTrainWorker pid=1483, ip=172.31.85.32)   Num examples = 8551
(RayTrainWorker pid=1483, ip=172.31.85.32)   Num Epochs = 4
(RayTrainWorker pid=1483, ip=172.31.85.32)   Instantaneous batch size per device = 16
(RayTrainWorker pid=1483, ip=172.31.85.32)   Total train batch size (w. parallel, distributed & accumulation) = 64
(RayTrainWorker pid=1483, ip=172.31.85.32)   Gradient Accumulation steps = 1
(RayTrainWorker pid=1483, ip=172.31.85.32)   Total optimization steps = 2140
(RayTrainWorker pid=1223, ip=172.31.85.193) 2022-08-25 10:14:48,193	INFO config.py:71 -- Setting up process group for: env:// [rank=0, world_size=4]
(RayTrainWorker pid=1223, ip=172.31.85.193) Is CUDA available: True
(RayTrainWorker pid=1224, ip=172.31.85.193) Is CUDA available: True
(RayTrainWorker pid=1226, ip=172.31.85.193) Is CUDA available: True
(RayTrainWorker pid=1225, ip=172.31.85.193) Is CUDA available: True
Downloading builder script: 5.76kB [00:00, 6.59MB/s]                   
Downloading builder script: 5.76kB [00:00, 6.52MB/s]                   
Downloading builder script: 5.76kB [00:00, 6.07MB/s]                   
Downloading builder script: 5.76kB [00:00, 6.81MB/s]                   
Downloading tokenizer_config.json: 100%|β–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆ| 28.0/28.0 [00:00<00:00, 46.0kB/s]
Downloading config.json: 100%|β–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆ| 483/483 [00:00<00:00, 766kB/s]
Downloading vocab.txt:   0%|          | 0.00/226k [00:00<?, ?B/s]
Downloading vocab.txt:  32%|β–ˆβ–ˆβ–ˆβ–      | 72.0k/226k [00:00<00:00, 624kB/s]
Downloading vocab.txt: 100%|β–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆ| 226k/226k [00:00<00:00, 966kB/s] 
Downloading tokenizer.json:   0%|          | 0.00/455k [00:00<?, ?B/s]
Downloading tokenizer.json:   6%|β–‹         | 29.0k/455k [00:00<00:01, 233kB/s]
Downloading tokenizer.json:  30%|β–ˆβ–ˆβ–‰       | 136k/455k [00:00<00:00, 600kB/s] 
Downloading tokenizer.json: 100%|β–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆ| 455k/455k [00:00<00:00, 1.44MB/s]
Downloading pytorch_model.bin:   0%|          | 0.00/256M [00:00<?, ?B/s]
Downloading pytorch_model.bin:   1%|          | 2.32M/256M [00:00<00:10, 24.4MB/s]
Downloading pytorch_model.bin:   4%|▍         | 11.0M/256M [00:00<00:04, 63.4MB/s]
Downloading pytorch_model.bin:   8%|β–Š         | 20.0M/256M [00:00<00:03, 77.7MB/s]
Downloading pytorch_model.bin:  11%|β–ˆβ–        | 29.1M/256M [00:00<00:02, 84.8MB/s]
Downloading pytorch_model.bin:  15%|β–ˆβ–        | 38.2M/256M [00:00<00:02, 88.5MB/s]
Downloading pytorch_model.bin:  18%|β–ˆβ–Š        | 47.3M/256M [00:00<00:02, 90.7MB/s]
Downloading pytorch_model.bin:  22%|β–ˆβ–ˆβ–       | 56.4M/256M [00:00<00:02, 92.4MB/s]
Downloading pytorch_model.bin:  26%|β–ˆβ–ˆβ–Œ       | 65.5M/256M [00:00<00:02, 93.4MB/s]
Downloading pytorch_model.bin:  29%|β–ˆβ–ˆβ–‰       | 74.7M/256M [00:00<00:02, 94.2MB/s]
Downloading pytorch_model.bin:  33%|β–ˆβ–ˆβ–ˆβ–Ž      | 83.8M/256M [00:01<00:01, 94.8MB/s]
Downloading pytorch_model.bin:  36%|β–ˆβ–ˆβ–ˆβ–‹      | 93.0M/256M [00:01<00:01, 95.1MB/s]
Downloading pytorch_model.bin:  40%|β–ˆβ–ˆβ–ˆβ–‰      | 102M/256M [00:01<00:01, 95.4MB/s] 
Downloading pytorch_model.bin:  44%|β–ˆβ–ˆβ–ˆβ–ˆβ–Ž     | 111M/256M [00:01<00:01, 95.6MB/s]
Downloading pytorch_model.bin:  47%|β–ˆβ–ˆβ–ˆβ–ˆβ–‹     | 120M/256M [00:01<00:01, 95.7MB/s]
Downloading pytorch_model.bin:  51%|β–ˆβ–ˆβ–ˆβ–ˆβ–ˆ     | 130M/256M [00:01<00:01, 95.8MB/s]
Downloading pytorch_model.bin:  54%|β–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–    | 139M/256M [00:01<00:01, 95.8MB/s]
Downloading pytorch_model.bin:  58%|β–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–Š    | 148M/256M [00:01<00:01, 95.9MB/s]
Downloading pytorch_model.bin:  61%|β–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–   | 157M/256M [00:01<00:01, 96.1MB/s]
Downloading pytorch_model.bin:  65%|β–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–Œ   | 166M/256M [00:01<00:00, 96.1MB/s]
Downloading pytorch_model.bin:  69%|β–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–Š   | 175M/256M [00:02<00:00, 96.1MB/s]
Downloading pytorch_model.bin:  72%|β–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–  | 185M/256M [00:02<00:00, 96.2MB/s]
Downloading pytorch_model.bin:  76%|β–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–Œ  | 194M/256M [00:02<00:00, 96.2MB/s]
Downloading pytorch_model.bin:  79%|β–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–‰  | 203M/256M [00:02<00:00, 96.1MB/s]
Downloading pytorch_model.bin:  83%|β–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–Ž | 212M/256M [00:02<00:00, 96.1MB/s]
Downloading pytorch_model.bin:  87%|β–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–‹ | 221M/256M [00:02<00:00, 96.2MB/s]
Downloading pytorch_model.bin:  90%|β–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆ | 231M/256M [00:02<00:00, 96.2MB/s]
Downloading pytorch_model.bin:  94%|β–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–| 240M/256M [00:02<00:00, 96.1MB/s]
Downloading pytorch_model.bin:  97%|β–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–‹| 249M/256M [00:02<00:00, 96.0MB/s]
Downloading pytorch_model.bin: 100%|β–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆ| 256M/256M [00:02<00:00, 93.2MB/s]
(RayTrainWorker pid=1223, ip=172.31.85.193) Starting training
(RayTrainWorker pid=1226, ip=172.31.85.193) Starting training
(RayTrainWorker pid=1225, ip=172.31.85.193) Starting training
(RayTrainWorker pid=1224, ip=172.31.85.193) Starting training
(RayTrainWorker pid=1223, ip=172.31.85.193) ***** Running training *****
(RayTrainWorker pid=1223, ip=172.31.85.193)   Num examples = 8551
(RayTrainWorker pid=1223, ip=172.31.85.193)   Num Epochs = 4
(RayTrainWorker pid=1223, ip=172.31.85.193)   Instantaneous batch size per device = 16
(RayTrainWorker pid=1223, ip=172.31.85.193)   Total train batch size (w. parallel, distributed & accumulation) = 64
(RayTrainWorker pid=1223, ip=172.31.85.193)   Gradient Accumulation steps = 1
(RayTrainWorker pid=1223, ip=172.31.85.193)   Total optimization steps = 2140
(RayTrainWorker pid=1789, ip=172.31.90.137) ***** Running Evaluation *****
(RayTrainWorker pid=1789, ip=172.31.90.137)   Num examples = 1043
(RayTrainWorker pid=1789, ip=172.31.90.137)   Batch size = 16
(RayTrainWorker pid=1789, ip=172.31.90.137) {'loss': 0.5458, 'learning_rate': 1.5000000000000002e-05, 'epoch': 1.0}
(RayTrainWorker pid=1789, ip=172.31.90.137) The following columns in the evaluation set don't have a corresponding argument in `DistilBertForSequenceClassification.forward` and have been ignored: sentence, idx. If sentence, idx are not expected by `DistilBertForSequenceClassification.forward`,  you can safely ignore this message.
(RayTrainWorker pid=1789, ip=172.31.90.137) {'eval_loss': 0.6037685871124268, 'eval_matthews_correlation': 0.3654892178274207, 'eval_runtime': 0.9847, 'eval_samples_per_second': 276.225, 'eval_steps_per_second': 5.078, 'epoch': 1.0}
(RayTrainWorker pid=1789, ip=172.31.90.137) Saving model checkpoint to distilbert-base-uncased-finetuned-cola/checkpoint-535
(RayTrainWorker pid=1789, ip=172.31.90.137) Configuration saved in distilbert-base-uncased-finetuned-cola/checkpoint-535/config.json
(RayTrainWorker pid=1789, ip=172.31.90.137) Model weights saved in distilbert-base-uncased-finetuned-cola/checkpoint-535/pytorch_model.bin
(RayTrainWorker pid=1789, ip=172.31.90.137) tokenizer config file saved in distilbert-base-uncased-finetuned-cola/checkpoint-535/tokenizer_config.json
(RayTrainWorker pid=1789, ip=172.31.90.137) Special tokens file saved in distilbert-base-uncased-finetuned-cola/checkpoint-535/special_tokens_map.json
Result for HuggingFaceTrainer_5654d_00000:
  _time_this_iter_s: 85.01727724075317
  _timestamp: 1661447753
  _training_iteration: 1
  date: 2022-08-25_10-15-53
  done: false
  epoch: 1.0
  eval_loss: 0.6037685871124268
  eval_matthews_correlation: 0.3654892178274207
  eval_runtime: 0.9847
  eval_samples_per_second: 276.225
  eval_steps_per_second: 5.078
  experiment_id: cee1b96afcf344e89482e3c5e298a412
  hostname: ip-172-31-90-137
  iterations_since_restore: 1
  learning_rate: 1.5000000000000002e-05
  loss: 0.5458
  node_ip: 172.31.90.137
  pid: 1729
  should_checkpoint: true
  step: 535
  time_since_restore: 94.93232989311218
  time_this_iter_s: 94.93232989311218
  time_total_s: 94.93232989311218
  timestamp: 1661447753
  timesteps_since_restore: 0
  training_iteration: 1
  trial_id: 5654d_00000
  warmup_time: 0.0037021636962890625
  
(RayTrainWorker pid=1974, ip=172.31.76.237) {'loss': 0.6225, 'learning_rate': 0.00015000000000000001, 'epoch': 1.0}
(RayTrainWorker pid=1974, ip=172.31.76.237) ***** Running Evaluation *****
(RayTrainWorker pid=1974, ip=172.31.76.237)   Num examples = 1043
(RayTrainWorker pid=1974, ip=172.31.76.237)   Batch size = 16
(RayTrainWorker pid=1974, ip=172.31.76.237) The following columns in the evaluation set don't have a corresponding argument in `DistilBertForSequenceClassification.forward` and have been ignored: idx, sentence. If idx, sentence are not expected by `DistilBertForSequenceClassification.forward`,  you can safely ignore this message.
(RayTrainWorker pid=1974, ip=172.31.76.237) {'eval_loss': 0.6492420434951782, 'eval_matthews_correlation': 0.0, 'eval_runtime': 1.0157, 'eval_samples_per_second': 267.792, 'eval_steps_per_second': 4.923, 'epoch': 1.0}
(RayTrainWorker pid=1974, ip=172.31.76.237) Saving model checkpoint to distilbert-base-uncased-finetuned-cola/checkpoint-535
(RayTrainWorker pid=1974, ip=172.31.76.237) Configuration saved in distilbert-base-uncased-finetuned-cola/checkpoint-535/config.json
(RayTrainWorker pid=1974, ip=172.31.76.237) Model weights saved in distilbert-base-uncased-finetuned-cola/checkpoint-535/pytorch_model.bin
(RayTrainWorker pid=1974, ip=172.31.76.237) tokenizer config file saved in distilbert-base-uncased-finetuned-cola/checkpoint-535/tokenizer_config.json
(RayTrainWorker pid=1974, ip=172.31.76.237) Special tokens file saved in distilbert-base-uncased-finetuned-cola/checkpoint-535/special_tokens_map.json
Result for HuggingFaceTrainer_5654d_00001:
  _time_this_iter_s: 84.79700112342834
  _timestamp: 1661447759
  _training_iteration: 1
  date: 2022-08-25_10-16-00
  done: true
  epoch: 1.0
  eval_loss: 0.6492420434951782
  eval_matthews_correlation: 0.0
  eval_runtime: 1.0157
  eval_samples_per_second: 267.792
  eval_steps_per_second: 4.923
  experiment_id: 88145f9344584715a4bd7d018f751b12
  hostname: ip-172-31-76-237
  iterations_since_restore: 1
  learning_rate: 0.00015000000000000001
  loss: 0.6225
  node_ip: 172.31.76.237
  pid: 1805
  should_checkpoint: true
  step: 535
  time_since_restore: 95.24916434288025
  time_this_iter_s: 95.24916434288025
  time_total_s: 95.24916434288025
  timestamp: 1661447760
  timesteps_since_restore: 0
  training_iteration: 1
  trial_id: 5654d_00001
  warmup_time: 0.003660917282104492
  
(RayTrainWorker pid=1483, ip=172.31.85.32) {'loss': 0.6463, 'learning_rate': 0.0015, 'epoch': 1.0}
(RayTrainWorker pid=1483, ip=172.31.85.32) ***** Running Evaluation *****
(RayTrainWorker pid=1483, ip=172.31.85.32)   Num examples = 1043
(RayTrainWorker pid=1483, ip=172.31.85.32)   Batch size = 16
(RayTrainWorker pid=1483, ip=172.31.85.32) The following columns in the evaluation set don't have a corresponding argument in `DistilBertForSequenceClassification.forward` and have been ignored: sentence, idx. If sentence, idx are not expected by `DistilBertForSequenceClassification.forward`,  you can safely ignore this message.
(RayTrainWorker pid=1483, ip=172.31.85.32) {'eval_loss': 0.6586529612541199, 'eval_matthews_correlation': 0.0, 'eval_runtime': 0.9576, 'eval_samples_per_second': 284.05, 'eval_steps_per_second': 5.222, 'epoch': 1.0}
(RayTrainWorker pid=1483, ip=172.31.85.32) Saving model checkpoint to distilbert-base-uncased-finetuned-cola/checkpoint-535
(RayTrainWorker pid=1483, ip=172.31.85.32) Configuration saved in distilbert-base-uncased-finetuned-cola/checkpoint-535/config.json
(RayTrainWorker pid=1483, ip=172.31.85.32) Model weights saved in distilbert-base-uncased-finetuned-cola/checkpoint-535/pytorch_model.bin
(RayTrainWorker pid=1483, ip=172.31.85.32) tokenizer config file saved in distilbert-base-uncased-finetuned-cola/checkpoint-535/tokenizer_config.json
(RayTrainWorker pid=1483, ip=172.31.85.32) Special tokens file saved in distilbert-base-uncased-finetuned-cola/checkpoint-535/special_tokens_map.json
Result for HuggingFaceTrainer_5654d_00002:
  _time_this_iter_s: 84.01720070838928
  _timestamp: 1661447764
  _training_iteration: 1
  date: 2022-08-25_10-16-04
  done: true
  epoch: 1.0
  eval_loss: 0.6586529612541199
  eval_matthews_correlation: 0.0
  eval_runtime: 0.9576
  eval_samples_per_second: 284.05
  eval_steps_per_second: 5.222
  experiment_id: 5f8ab183779d40379d59ea615f9d5411
  hostname: ip-172-31-85-32
  iterations_since_restore: 1
  learning_rate: 0.0015
  loss: 0.6463
  node_ip: 172.31.85.32
  pid: 1322
  should_checkpoint: true
  step: 535
  time_since_restore: 93.76131749153137
  time_this_iter_s: 93.76131749153137
  time_total_s: 93.76131749153137
  timestamp: 1661447764
  timesteps_since_restore: 0
  training_iteration: 1
  trial_id: 5654d_00002
  warmup_time: 0.004533290863037109
  
(RayTrainWorker pid=1223, ip=172.31.85.193) ***** Running Evaluation *****
(RayTrainWorker pid=1223, ip=172.31.85.193)   Num examples = 1043
(RayTrainWorker pid=1223, ip=172.31.85.193)   Batch size = 16
(RayTrainWorker pid=1223, ip=172.31.85.193) The following columns in the evaluation set don't have a corresponding argument in `DistilBertForSequenceClassification.forward` and have been ignored: idx, sentence. If idx, sentence are not expected by `DistilBertForSequenceClassification.forward`,  you can safely ignore this message.
(RayTrainWorker pid=1223, ip=172.31.85.193) {'loss': 0.926, 'learning_rate': 0.015, 'epoch': 1.0}
(RayTrainWorker pid=1223, ip=172.31.85.193) {'eval_loss': 0.6529427766799927, 'eval_matthews_correlation': 0.0, 'eval_runtime': 0.9428, 'eval_samples_per_second': 288.51, 'eval_steps_per_second': 5.303, 'epoch': 1.0}
(RayTrainWorker pid=1223, ip=172.31.85.193) Saving model checkpoint to distilbert-base-uncased-finetuned-cola/checkpoint-535
(RayTrainWorker pid=1223, ip=172.31.85.193) Configuration saved in distilbert-base-uncased-finetuned-cola/checkpoint-535/config.json
(RayTrainWorker pid=1223, ip=172.31.85.193) Model weights saved in distilbert-base-uncased-finetuned-cola/checkpoint-535/pytorch_model.bin
(RayTrainWorker pid=1223, ip=172.31.85.193) tokenizer config file saved in distilbert-base-uncased-finetuned-cola/checkpoint-535/tokenizer_config.json
(RayTrainWorker pid=1223, ip=172.31.85.193) Special tokens file saved in distilbert-base-uncased-finetuned-cola/checkpoint-535/special_tokens_map.json
Result for HuggingFaceTrainer_5654d_00003:
  _time_this_iter_s: 89.4301290512085
  _timestamp: 1661447782
  _training_iteration: 1
  date: 2022-08-25_10-16-22
  done: true
  epoch: 1.0
  eval_loss: 0.6529427766799927
  eval_matthews_correlation: 0.0
  eval_runtime: 0.9428
  eval_samples_per_second: 288.51
  eval_steps_per_second: 5.303
  experiment_id: 8495977eeefd405fa4d9c1ea8fa735e1
  hostname: ip-172-31-85-193
  iterations_since_restore: 1
  learning_rate: 0.015
  loss: 0.926
  node_ip: 172.31.85.193
  pid: 1060
  should_checkpoint: true
  step: 535
  time_since_restore: 99.36774587631226
  time_this_iter_s: 99.36774587631226
  time_total_s: 99.36774587631226
  timestamp: 1661447782
  timesteps_since_restore: 0
  training_iteration: 1
  trial_id: 5654d_00003
  warmup_time: 0.004132509231567383
  
(RayTrainWorker pid=1789, ip=172.31.90.137) ***** Running Evaluation *****
(RayTrainWorker pid=1789, ip=172.31.90.137)   Num examples = 1043
(RayTrainWorker pid=1789, ip=172.31.90.137)   Batch size = 16
(RayTrainWorker pid=1789, ip=172.31.90.137) The following columns in the evaluation set don't have a corresponding argument in `DistilBertForSequenceClassification.forward` and have been ignored: sentence, idx. If sentence, idx are not expected by `DistilBertForSequenceClassification.forward`,  you can safely ignore this message.
(RayTrainWorker pid=1789, ip=172.31.90.137) {'loss': 0.3841, 'learning_rate': 1e-05, 'epoch': 2.0}
(RayTrainWorker pid=1789, ip=172.31.90.137) {'eval_loss': 0.5994958281517029, 'eval_matthews_correlation': 0.4573244914254411, 'eval_runtime': 0.9442, 'eval_samples_per_second': 288.066, 'eval_steps_per_second': 5.295, 'epoch': 2.0}
(RayTrainWorker pid=1789, ip=172.31.90.137) Saving model checkpoint to distilbert-base-uncased-finetuned-cola/checkpoint-1070
(RayTrainWorker pid=1789, ip=172.31.90.137) Configuration saved in distilbert-base-uncased-finetuned-cola/checkpoint-1070/config.json
(RayTrainWorker pid=1789, ip=172.31.90.137) Model weights saved in distilbert-base-uncased-finetuned-cola/checkpoint-1070/pytorch_model.bin
(RayTrainWorker pid=1789, ip=172.31.90.137) tokenizer config file saved in distilbert-base-uncased-finetuned-cola/checkpoint-1070/tokenizer_config.json
(RayTrainWorker pid=1789, ip=172.31.90.137) Special tokens file saved in distilbert-base-uncased-finetuned-cola/checkpoint-1070/special_tokens_map.json
Result for HuggingFaceTrainer_5654d_00000:
  _time_this_iter_s: 76.82565689086914
  _timestamp: 1661447830
  _training_iteration: 2
  date: 2022-08-25_10-17-10
  done: false
  epoch: 2.0
  eval_loss: 0.5994958281517029
  eval_matthews_correlation: 0.4573244914254411
  eval_runtime: 0.9442
  eval_samples_per_second: 288.066
  eval_steps_per_second: 5.295
  experiment_id: cee1b96afcf344e89482e3c5e298a412
  hostname: ip-172-31-90-137
  iterations_since_restore: 2
  learning_rate: 1.0e-05
  loss: 0.3841
  node_ip: 172.31.90.137
  pid: 1729
  should_checkpoint: true
  step: 1070
  time_since_restore: 171.76071190834045
  time_this_iter_s: 76.82838201522827
  time_total_s: 171.76071190834045
  timestamp: 1661447830
  timesteps_since_restore: 0
  training_iteration: 2
  trial_id: 5654d_00000
  warmup_time: 0.0037021636962890625
  
(RayTrainWorker pid=1789, ip=172.31.90.137) ***** Running Evaluation *****
(RayTrainWorker pid=1789, ip=172.31.90.137)   Num examples = 1043
(RayTrainWorker pid=1789, ip=172.31.90.137)   Batch size = 16
(RayTrainWorker pid=1789, ip=172.31.90.137) The following columns in the evaluation set don't have a corresponding argument in `DistilBertForSequenceClassification.forward` and have been ignored: sentence, idx. If sentence, idx are not expected by `DistilBertForSequenceClassification.forward`,  you can safely ignore this message.
(RayTrainWorker pid=1789, ip=172.31.90.137) {'loss': 0.2687, 'learning_rate': 5e-06, 'epoch': 3.0}
(RayTrainWorker pid=1789, ip=172.31.90.137) {'eval_loss': 0.6935313940048218, 'eval_matthews_correlation': 0.5300538425561, 'eval_runtime': 1.0176, 'eval_samples_per_second': 267.305, 'eval_steps_per_second': 4.914, 'epoch': 3.0}
(RayTrainWorker pid=1789, ip=172.31.90.137) Saving model checkpoint to distilbert-base-uncased-finetuned-cola/checkpoint-1605
(RayTrainWorker pid=1789, ip=172.31.90.137) Configuration saved in distilbert-base-uncased-finetuned-cola/checkpoint-1605/config.json
(RayTrainWorker pid=1789, ip=172.31.90.137) Model weights saved in distilbert-base-uncased-finetuned-cola/checkpoint-1605/pytorch_model.bin
(RayTrainWorker pid=1789, ip=172.31.90.137) tokenizer config file saved in distilbert-base-uncased-finetuned-cola/checkpoint-1605/tokenizer_config.json
(RayTrainWorker pid=1789, ip=172.31.90.137) Special tokens file saved in distilbert-base-uncased-finetuned-cola/checkpoint-1605/special_tokens_map.json
Result for HuggingFaceTrainer_5654d_00000:
  _time_this_iter_s: 76.47252488136292
  _timestamp: 1661447906
  _training_iteration: 3
  date: 2022-08-25_10-18-26
  done: false
  epoch: 3.0
  eval_loss: 0.6935313940048218
  eval_matthews_correlation: 0.5300538425561
  eval_runtime: 1.0176
  eval_samples_per_second: 267.305
  eval_steps_per_second: 4.914
  experiment_id: cee1b96afcf344e89482e3c5e298a412
  hostname: ip-172-31-90-137
  iterations_since_restore: 3
  learning_rate: 5.0e-06
  loss: 0.2687
  node_ip: 172.31.90.137
  pid: 1729
  should_checkpoint: true
  step: 1605
  time_since_restore: 248.23273348808289
  time_this_iter_s: 76.47202157974243
  time_total_s: 248.23273348808289
  timestamp: 1661447906
  timesteps_since_restore: 0
  training_iteration: 3
  trial_id: 5654d_00000
  warmup_time: 0.0037021636962890625
  
(RayTrainWorker pid=1789, ip=172.31.90.137) Saving model checkpoint to distilbert-base-uncased-finetuned-cola/checkpoint-2140
(RayTrainWorker pid=1789, ip=172.31.90.137) Configuration saved in distilbert-base-uncased-finetuned-cola/checkpoint-2140/config.json
(RayTrainWorker pid=1789, ip=172.31.90.137) Model weights saved in distilbert-base-uncased-finetuned-cola/checkpoint-2140/pytorch_model.bin
(RayTrainWorker pid=1789, ip=172.31.90.137) tokenizer config file saved in distilbert-base-uncased-finetuned-cola/checkpoint-2140/tokenizer_config.json
(RayTrainWorker pid=1789, ip=172.31.90.137) Special tokens file saved in distilbert-base-uncased-finetuned-cola/checkpoint-2140/special_tokens_map.json
(RayTrainWorker pid=1789, ip=172.31.90.137) ***** Running Evaluation *****
(RayTrainWorker pid=1789, ip=172.31.90.137)   Num examples = 1043
(RayTrainWorker pid=1789, ip=172.31.90.137)   Batch size = 16
(RayTrainWorker pid=1789, ip=172.31.90.137) The following columns in the evaluation set don't have a corresponding argument in `DistilBertForSequenceClassification.forward` and have been ignored: sentence, idx. If sentence, idx are not expected by `DistilBertForSequenceClassification.forward`,  you can safely ignore this message.
(RayTrainWorker pid=1789, ip=172.31.90.137) {'loss': 0.1958, 'learning_rate': 0.0, 'epoch': 4.0}
(RayTrainWorker pid=1789, ip=172.31.90.137) Saving model checkpoint to distilbert-base-uncased-finetuned-cola/checkpoint-2140
(RayTrainWorker pid=1789, ip=172.31.90.137) Configuration saved in distilbert-base-uncased-finetuned-cola/checkpoint-2140/config.json
(RayTrainWorker pid=1789, ip=172.31.90.137) {'eval_loss': 0.8064090609550476, 'eval_matthews_correlation': 0.5322860764824153, 'eval_runtime': 1.0006, 'eval_samples_per_second': 271.827, 'eval_steps_per_second': 4.997, 'epoch': 4.0}
(RayTrainWorker pid=1789, ip=172.31.90.137) Model weights saved in distilbert-base-uncased-finetuned-cola/checkpoint-2140/pytorch_model.bin
(RayTrainWorker pid=1789, ip=172.31.90.137) tokenizer config file saved in distilbert-base-uncased-finetuned-cola/checkpoint-2140/tokenizer_config.json
(RayTrainWorker pid=1789, ip=172.31.90.137) Special tokens file saved in distilbert-base-uncased-finetuned-cola/checkpoint-2140/special_tokens_map.json
(RayTrainWorker pid=1789, ip=172.31.90.137) 
(RayTrainWorker pid=1789, ip=172.31.90.137) 
(RayTrainWorker pid=1789, ip=172.31.90.137) Training completed. Do not forget to share your model on huggingface.co/models =)
(RayTrainWorker pid=1789, ip=172.31.90.137) 
(RayTrainWorker pid=1789, ip=172.31.90.137) 
(RayTrainWorker pid=1789, ip=172.31.90.137) {'train_runtime': 329.1948, 'train_samples_per_second': 103.902, 'train_steps_per_second': 6.501, 'train_loss': 0.34860724689804506, 'epoch': 4.0}
Result for HuggingFaceTrainer_5654d_00000:
  _time_this_iter_s: 98.92064905166626
  _timestamp: 1661448005
  _training_iteration: 4
  date: 2022-08-25_10-20-05
  done: true
  epoch: 4.0
  eval_loss: 0.8064090609550476
  eval_matthews_correlation: 0.5322860764824153
  eval_runtime: 1.0006
  eval_samples_per_second: 271.827
  eval_steps_per_second: 4.997
  experiment_id: cee1b96afcf344e89482e3c5e298a412
  hostname: ip-172-31-90-137
  iterations_since_restore: 4
  learning_rate: 0.0
  loss: 0.1958
  node_ip: 172.31.90.137
  pid: 1729
  should_checkpoint: true
  step: 2140
  time_since_restore: 347.1705844402313
  time_this_iter_s: 98.93785095214844
  time_total_s: 347.1705844402313
  timestamp: 1661448005
  timesteps_since_restore: 0
  train_loss: 0.34860724689804506
  train_runtime: 329.1948
  train_samples_per_second: 103.902
  train_steps_per_second: 6.501
  training_iteration: 4
  trial_id: 5654d_00000
  warmup_time: 0.0037021636962890625
  
2022-08-25 10:20:13,409	INFO tune.py:758 -- Total run time: 361.90 seconds (361.74 seconds for the tuning loop).

We can view the results of the tuning run as a dataframe, and obtain the best result.

tune_results.get_dataframe().sort_values("eval_loss")
loss learning_rate epoch step eval_loss eval_matthews_correlation eval_runtime eval_samples_per_second eval_steps_per_second _timestamp ... pid hostname node_ip time_since_restore timesteps_since_restore iterations_since_restore warmup_time config/trainer_init_config/epochs config/trainer_init_config/learning_rate logdir
1 0.6225 0.00015 1.0 535 0.649242 0.000000 1.0157 267.792 4.923 1661447759 ... 1805 ip-172-31-76-237 172.31.76.237 95.249164 0 1 0.003661 4 0.00020 /home/ray/ray_results/HuggingFaceTrainer_2022-...
3 0.9260 0.01500 1.0 535 0.652943 0.000000 0.9428 288.510 5.303 1661447782 ... 1060 ip-172-31-85-193 172.31.85.193 99.367746 0 1 0.004133 4 0.02000 /home/ray/ray_results/HuggingFaceTrainer_2022-...
2 0.6463 0.00150 1.0 535 0.658653 0.000000 0.9576 284.050 5.222 1661447764 ... 1322 ip-172-31-85-32 172.31.85.32 93.761317 0 1 0.004533 4 0.00200 /home/ray/ray_results/HuggingFaceTrainer_2022-...
0 0.1958 0.00000 4.0 2140 0.806409 0.532286 1.0006 271.827 4.997 1661448005 ... 1729 ip-172-31-90-137 172.31.90.137 347.170584 0 4 0.003702 4 0.00002 /home/ray/ray_results/HuggingFaceTrainer_2022-...

4 rows Γ— 33 columns

best_result = tune_results.get_best_result()

Predict on test data with Ray AIR ΒΆ

You can now use the checkpoint to run prediction with HuggingFacePredictor, which wraps around πŸ€— Pipelines. In order to distribute prediction, we use BatchPredictor. While this is not necessary for the very small example we are using (you could use HuggingFacePredictor directly), it will scale well to a large dataset.

from ray.train.huggingface import HuggingFacePredictor
from ray.train.batch_predictor import BatchPredictor
import pandas as pd

predictor = BatchPredictor.from_checkpoint(
    checkpoint=best_result.checkpoint,
    predictor_cls=HuggingFacePredictor,
    task="text-classification",
    device=0 if use_gpu else -1,  # -1 is CPU, otherwise device index
)
prediction = predictor.predict(ray_datasets["test"].map_batches(lambda x: x[["sentence"]]), num_gpus_per_worker=int(use_gpu))
prediction.show()
Map_Batches: 100%|β–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆ| 1/1 [00:00<00:00, 12.41it/s]
Map_Batches: 100%|β–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆ| 1/1 [00:00<00:00,  7.46it/s]
Map Progress (1 actors 1 pending): 100%|β–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆ| 1/1 [00:18<00:00, 18.46s/it]
{'label': 'LABEL_1', 'score': 0.6822417974472046}
{'label': 'LABEL_1', 'score': 0.6822402477264404}
{'label': 'LABEL_1', 'score': 0.6822407841682434}
{'label': 'LABEL_1', 'score': 0.6822386980056763}
{'label': 'LABEL_1', 'score': 0.6822428107261658}
{'label': 'LABEL_1', 'score': 0.6822453737258911}
{'label': 'LABEL_1', 'score': 0.6822437047958374}
{'label': 'LABEL_1', 'score': 0.6822428703308105}
{'label': 'LABEL_1', 'score': 0.6822431683540344}
{'label': 'LABEL_1', 'score': 0.6822426915168762}
{'label': 'LABEL_1', 'score': 0.6822447776794434}
{'label': 'LABEL_1', 'score': 0.6822456121444702}
{'label': 'LABEL_1', 'score': 0.6822471022605896}
{'label': 'LABEL_1', 'score': 0.6822477579116821}
{'label': 'LABEL_1', 'score': 0.682244598865509}
{'label': 'LABEL_1', 'score': 0.6822422742843628}
{'label': 'LABEL_1', 'score': 0.6822470426559448}
{'label': 'LABEL_1', 'score': 0.6822417378425598}
{'label': 'LABEL_1', 'score': 0.6822449564933777}
{'label': 'LABEL_1', 'score': 0.682239294052124}

Share the model ΒΆ

To be able to share your model with the community, there are a few more steps to follow.

We have conducted the training on the Ray cluster, but share the model from the local enviroment - this will allow us to easily authenticate.

First you have to store your authentication token from the Hugging Face website (sign up here if you haven’t already!) then execute the following cell and input your username and password:

from huggingface_hub import notebook_login

notebook_login()

Then you need to install Git-LFS. Uncomment the following instructions:

# !apt install git-lfs

Now, load the model and tokenizer locally, and recreate the πŸ€— Transformers Trainer:

from ray.train.huggingface import HuggingFaceCheckpoint

checkpoint = HuggingFaceCheckpoint.from_checkpoint(result.checkpoint)
hf_trainer = checkpoint.get_model(model=AutoModelForSequenceClassification)

You can now upload the result of the training to the Hub, just execute this instruction:

hf_trainer.push_to_hub()

You can now share this model with all your friends, family, favorite pets: they can all load it with the identifier "your-username/the-name-you-picked" so for instance:

from transformers import AutoModelForSequenceClassification

model = AutoModelForSequenceClassification.from_pretrained("sgugger/my-awesome-model")