Hugging Face CIFAR-100 Embeddings Example

Open in GitHub

In this notebook we will see how to use a pre-trained Vision Transformers (ViT) model to collect embeddings on the CIFAR-100 dataset.

This notebook demonstrates:

• Registering the CIFAR-100 dataset from Hugging Face.

• Computing image embeddings with transformers and reducing them to 2D with UMAP.

• Adding the computed embeddings as metrics to a 3LC Run.

Project Setup

PROJECT_NAME = "CIFAR-100"
RUN_NAME = "Collect Image Embeddings"
DESCRIPTION = "Collect image embeddings from ViT model on CIFAR-100"
DEVICE = "cuda:0"
TRAIN_DATASET_NAME = "hf-cifar-100-train"
TEST_DATASET_NAME = "hf-cifar-100-test"
MODEL = "google/vit-base-patch16-224"
BATCH_SIZE = 32
TRANSIENT_DATA_PATH = "../transient_data"
TLC_PUBLIC_EXAMPLES_DEVELOPER_MODE = True
INSTALL_DEPENDENCIES = False

%%capture
if INSTALL_DEPENDENCIES:
    %pip --quiet install datasets transformers
    %pip --quiet install torch --index-url https://download.pytorch.org/whl/cu118
    %pip --quiet install torchvision --index-url https://download.pytorch.org/whl/cu118
    %pip --quiet install tlc[umap]

Imports

import logging

import datasets
import tlc

logging.getLogger("transformers.modeling_utils").setLevel(logging.ERROR)  # Reduce model loading logs
datasets.utils.logging.disable_progress_bar()

Prepare the data

To read the data into 3LC, we use tlc.Table.from_hugging_face() available under the Hugging Face integration. This returns a Table that works similarly to a Hugging Face datasets.Dataset.

cifar100_train = tlc.Table.from_hugging_face(
    "cifar100",
    split="train",
    table_name="train",
    project_name=PROJECT_NAME,
    dataset_name=TRAIN_DATASET_NAME,
    description="CIFAR-100 training dataset",
    if_exists="overwrite",
)

cifar100_test = tlc.Table.from_hugging_face(
    "cifar100",
    split="test",
    table_name="test",
    project_name=PROJECT_NAME,
    dataset_name=TEST_DATASET_NAME,
    description="CIFAR-100 test dataset",
    if_exists="overwrite",
)

cifar100_train[0]["img"]

Compute the data

We then use the transformers library to compute embeddings and umap-learn to reduce the embeddings to two dimensions.

import torch
from torch.utils.data import DataLoader
from tqdm.auto import tqdm
from transformers import ViTImageProcessor, ViTModel

device = torch.device(DEVICE)
feature_extractor = ViTImageProcessor.from_pretrained(MODEL)
model = ViTModel.from_pretrained(MODEL).to(device)

The cache for model files in Transformers v4.22.0 has been updated. Migrating your old cache. This is a one-time only operation. You can interrupt this and resume the migration later on by calling `transformers.utils.move_cache()`.

def extract_feature(sample):
    return feature_extractor(images=sample["img"], return_tensors="pt")

def infer_on_dataset(dataset):
    activations = []
    dataloader = DataLoader(dataset, batch_size=BATCH_SIZE, shuffle=False)
    for inputs in tqdm(dataloader, total=len(dataloader)):
        inputs["pixel_values"] = inputs["pixel_values"].squeeze()
        inputs = inputs.to(DEVICE)
        outputs = model(**inputs)
        activations.append(outputs.last_hidden_state[:, 0, :].detach().cpu())

    return activations

activations = []
model.eval()

for dataset in (cifar100_train, cifar100_test):
    dataset = dataset.map(extract_feature)
    activations.extend(infer_on_dataset(dataset))

activations = torch.cat(activations).numpy()
activations.shape

(60000, 768)

import umap

reducer = umap.UMAP(n_components=2)
embeddings_2d = reducer.fit_transform(activations)

Collect the embeddings as 3LC metrics

In this example the metrics are contained in a numpy.ndarray object. We can specify the schema of this data and provide it directly to 3LC using Run.add_metrics_data().

run = tlc.init(
    project_name=PROJECT_NAME,
    run_name=RUN_NAME,
    description=DESCRIPTION,
    if_exists="overwrite",
)

embeddings_2d_train = embeddings_2d[: len(cifar100_train)]
embeddings_2d_test = embeddings_2d[len(cifar100_train) :]

embeddings_2d_train = embeddings_2d[: len(cifar100_train)]
embeddings_2d_test = embeddings_2d[len(cifar100_train) :]

for dataset, embeddings in ((cifar100_train, embeddings_2d_train), (cifar100_test, embeddings_2d_test)):
    run.add_metrics_data(
        {"embeddings": list(embeddings)},
        override_column_schemas={"embeddings": tlc.FloatVector2Schema()},
        input_table_url=dataset.url,
    )

run.set_status_completed()