Falconsai

/

nsfw_image_detection

like 290

A basic training arch similar to the one we used:
#---------------------------------------------------------------------------------------------------------
training_args = TrainingArguments(
output_dir="nsfw_model",
remove_unused_columns=False,
evaluation_strategy="epoch",
save_strategy="epoch",
learning_rate=5e-5,
per_device_train_batch_size=btch_sz,
gradient_accumulation_steps=4,
per_device_eval_batch_size=btch_sz,
num_train_epochs=EPOCHS,
warmup_ratio=0.1,
logging_steps=10,
load_best_model_at_end=True,
metric_for_best_model="accuracy",
# Set to 2 so as not to save every checkpoint
save_total_limit=2,
push_to_hub=False,
)
trainer = Trainer(
model=model,
args=training_args,
data_collator=data_collator,
train_dataset=ds["train"],
eval_dataset=ds["test"],
tokenizer=image_processor,
compute_metrics=compute_metrics,
)

trainer.train()

#---------------------------------------------------------------------------------------------------------

Apologies, but based on the nature of the content, we cannot publish the dataset.

Hi
Thankyou for replying,
I understand the data cannot be shared.

Am i correct in assuming that dataset is a simple image, text pair where text is probably just the binary classes, nsfw / safe?
Did you finetune all the layers or used freezing of some layers?

Yes, all.