Train a instance classifier on a 3LC Table¶

In this tutorial, we will fine-tune a classifier using instances (segmentations or bounding boxes) from a 3LC Table.

We will load the COCO128 table from an earlier notebook and use it to create a torch.utils.Dataset of bounding box crops. These cropped images will be used to fine-tune a classifier. In a later tutorial, we will use this trained model to generate embeddings and predicted labels.

Install dependencies¶

[ ]:

%pip install 3lc[pacmap]
%pip install git+https://github.com/3lc-ai/3lc-examples.git
%pip install timm

Imports¶

[ ]:

import tlc

from tlc_tools.augment_bbs.finetune_on_crops import train_model
from tlc_tools.common import infer_torch_device
from tlc_tools.split import split_table

Project setup¶

[ ]:

EPOCHS = 10
DOWNLOAD_PATH = "../../../transient_data"
MODEL_NAME = "efficientnet_b0"
NUM_WORKERS = 0

[ ]:

MODEL_CHECKPOINT = DOWNLOAD_PATH + "/instance_classifier.pth"

Set device¶

[ ]:

DEVICE = infer_torch_device()
print(f"Using device: {DEVICE}")

Load input Table¶

We will reuse the table created in the notebook create-table-from-coco-detection.ipynb.

[ ]:

input_table = tlc.Table.from_names(
    "initial-segmentation",
    "COCO128",
    "3LC Tutorials - COCO128",
)

Split the Table¶

[ ]:

# Create splits for training and validation
splits = split_table(input_table, {"train": 0.8, "val": 0.2}, if_exists="reuse")

train_table = splits["train"]
val_table = splits["val"]

print(f"Using table {train_table} for training")
print(f"Using table {val_table} for validation")

Train model¶

[ ]:

model, checkpoint_path = train_model(
    train_table_url=train_table.url,
    val_table_url=val_table.url,
    model_checkpoint=MODEL_CHECKPOINT,
    epochs=EPOCHS,
    num_workers=NUM_WORKERS,
)