Marcus Gawronsky
Create trainer.py
9afa7a5
# We want to train a classification model on our own data
from transformers import AutoImageProcessor, AutoModelForImageClassification, TrainingArguments, Trainer
from sklearn.preprocessing import LabelEncoder, OneHotEncoder
import numpy as np
import pandas as pd
from datasets import load_dataset, load_from_disk
import torch
from joblib import cpu_count
from fire import Fire
def get_feature_function(preprocessor, encoder):
def feature_extraction_function(examples):
data = {}
data["pixel_values"] = preprocessor(examples["image"]).pixel_values
data['label'] = np.eye(len(encoder.classes_))[encoder.transform(examples['product_subcategory_name'])]
return data
return feature_extraction_function
def train(
dataset_id,
hub_model_id,
model_id = 'facebook/convnextv2-atto-1k-224',
run_id = 'convextv2-atto-dataset',
logging_steps = 100
)
dataset = load_dataset(dataset_id)
preprocessor = AutoImageProcessor.from_pretrained(model_id)
labels = np.unique(dataset['train']['product_subcategory_name'])
encoder = LabelEncoder().fit(y=labels)
model = AutoModelForImageClassification.from_pretrained(model_id,
ignore_mismatched_sizes=True,
num_labels=len(encoder.classes_),
id2label={i: label for i, label in enumerate(encoder.classes_)},
label2id={label: i for i, label in enumerate(encoder.classes_)}
)
dataset.set_transform(get_feature_function(preprocessor, encoder))
training_args = TrainingArguments(
output_dir=f"results/{run_id}",
remove_unused_columns=False,
learning_rate=5e-5,
per_device_train_batch_size=32,
gradient_accumulation_steps=4,
per_device_eval_batch_size=16,
num_train_epochs=3,
warmup_ratio=0.1,
report_to=['tensorboard'],
run_name=run_id,
logging_steps=logging_steps,
eval_steps=logging_steps,
save_steps=logging_steps,
save_total_limit=1,
save_strategy="steps",
evaluation_strategy="steps",
skip_memory_metrics=False,
logging_first_step=True,
push_to_hub=True,
hub_model_id=hub_model_id,
hub_private_repo=True,
hub_strategy="every_save",
save_safetensors=True,
# memory
dataloader_num_workers=cpu_count()//4, # we have to prefetch the data to ensure efficient and stable GPU utilization
dataloader_pin_memory=True
)
trainer = Trainer(
model=model,
args=training_args,
train_dataset=dataset["train"],
eval_dataset=dataset["test"]
)
trainer.train()
if __name__ == '__main__':
Fire(train)