MLToolslabel · build · tune · ship

Fine-tuning

Learning Rate Schedule Visualizer

A schedule is easier to reason about as a picture. Set the total steps from epochs and batches, the peak learning rate, warmup and the schedule type, and see the curve, the rate at any step, and the values as CSV or a code snippet for the framework you use.

Loading the tool…

How it works

01

Set the run

Examples, batch size, epochs, gradient accumulation give total steps; or type steps directly.

02

Choose the schedule

Constant, linear decay, cosine, cosine with restarts, step decay, one-cycle, with linear warmup steps or ratio and a minimum rate.

03

Export

CSV of step and rate, or the equivalent transformers get_scheduler / torch.optim.lr_scheduler call.

Formats, limits and the numbers that matter

Warmup

3-10 percent of steps for transformers; more when the batch is small or the model is fresh. Warmup avoids the early divergence a high rate causes on uncentred gradients.

Cosine vs linear

Both end near zero; cosine keeps the rate high longer, which tends to help when the dataset is small and the run short.

Minimum rate

Ending at 10 percent of peak rather than zero keeps late steps useful for LoRA runs of a few epochs.

Common questions

How does the step count relate to epochs?

Steps = ceil(examples / (batch x accumulation)) x epochs; the tool shows the arithmetic.

Which schedule for YOLO?

Ultralytics uses linear by default with cos_lr as an option; both are plotted with its warmup_epochs convention.

Can I overlay two?

Yes; add a second schedule to compare.

Next in the workflow

Fine-tuning Script GeneratorGenerate a runnable training script and config for YOLO detection, Hugging Face LoRA SFT, or an image classifier from your paths and settings.Training Time & Cost EstimatorEstimate hours and cloud cost from dataset size, epochs, throughput and GPU price, for LLM and vision runs.GPU Memory (VRAM) CalculatorEstimate the GPU memory a fine-tune or inference run needs from parameters, precision, LoRA, batch size and sequence length.Model Size & Quantization CalculatorFile size and inference memory for a model at fp32, fp16, int8, int4 and GGUF quantisation levels, plus KV cache per token.