Learning Rate Calculator: Convert 1e-3, 0.001, and 0.1% in Seconds

The learning rate is the single most influential hyperparameter in deep learning, and yet it hides behind a surprising amount of notational confusion. A paper reports 3e-4, a config file says 0.0003, a tutorial slide shows 0.03%, and a teammate asks, "wait, is that ten times bigger or smaller than what we're running?" Misreading one exponent can silently derail a training run for hours. The Learning Rate Calculator removes that risk entirely: type a rate in any format — decimal, scientific notation, or percentage — and it normalizes the value instantly, shows every equivalent representation, and hands you copy-ready code for PyTorch and TensorFlow.

Beyond conversion, the tool packs in everything you need to reason about learning rates in one place: a table of optimizer defaults, an Adam effective-step-size interpreter, and a side-by-side preview of linear decay versus cosine annealing schedules. It runs 100% client-side, is completely free, and requires no signup.

Whether you are reproducing a paper, porting a model between frameworks, or tuning your first transformer, this guide explains how the calculator works and the concepts behind it.

Why Use the Learning Rate Calculator?

• Eliminate exponent errors. Confusing 1e-4 with 1e-3 is a 10x mistake that often shows up only as a slowly diverging loss curve. Instant normalization across formats makes the magnitude obvious at a glance.
• Parse any input format. Type 0.001, 1e-3, or 0.1% — a trailing % is divided by 100 automatically and scientific notation is parsed natively, so you can copy values straight from papers and slides.
• Check optimizer defaults instantly. A built-in reference table lists the canonical defaults for Adam, AdamW, SGD, Momentum, RMSProp, and Adagrad, so you always know your starting point.
• Preview schedules before committing. Compare linear decay and cosine annealing over 10 steps with the exact formulas used, so you can see how fast each schedule drops the rate.
• Copy framework-ready code. Every value has a copy button, and dedicated snippets show the correct syntax for PyTorch and TensorFlow/Keras — no more second-guessing argument names.
• Private, fast, and free. All computation happens in your browser. Nothing is uploaded, no account is needed, and it works offline once loaded.

Key Features

A few details worth highlighting:

• The conversion panel doesn't stop at the three human formats — it also renders the value exactly as Keras and PyTorch expect it, so the number you paste into your code is guaranteed to match what you verified.
• The Adam-family interpretation panel is a rare feature: it translates a raw learning rate into an intuitive effective step size, including a batch-scaling view, which helps when you apply the common "scale lr with batch size" rule of thumb.
• The schedule preview uses the same formulas you'd implement in a scheduler, so the table is not illustrative — it is the actual math, step by step.

How to Use

1. Enter your learning rate. Type any value into the input field: 0.001, 1e-3, or 0.1%. The tool normalizes it live as you type.
2. Or click a preset. Use one of the six quick preset buttons (from 1 down to 5e-5) to jump straight to a common value.
3. Read the conversion panel. Check the decimal, scientific, and percentage forms, along with the Keras- and PyTorch-ready values. Click any copy button to grab one.
4. Compare against optimizer defaults. Scan the defaults table to see how your value relates to Adam (1e-3), SGD (1e-2), and the rest. If you're using an Adam variant, review the effective step size panel.
5. Preview your schedule. Look at the 10-step table comparing linear decay and cosine annealing for your base rate, then copy the matching framework snippet from the code section into your training script.

Understanding Learning Rate Formats and Schedules

Decimal vs scientific vs percentage

All three formats express the same number:

0.001  =  1e-3  =  0.1%

The decimal form (0.001) is what most framework APIs accept. Scientific notation (1e-3) is what papers and logs prefer because the exponent makes magnitude comparisons trivial. The percentage form (0.1%) appears in slides and older literature, where the learning rate was framed as a fraction of the gradient step. The calculator treats a trailing % as "divide by 100," so 0.1% correctly becomes 0.001.

Why papers write 1e-3

Learning rates routinely span five orders of magnitude — from 1 for carefully tuned SGD on small problems down to 1e-5 for fine-tuning large pretrained models. Counting zeros in 0.00001 is error-prone; reading the exponent in 1e-5 is not. Scientific notation also makes scaling rules legible: "increase the rate 10x" becomes "add 1 to the exponent."

Adam's effective step size

For adaptive optimizers, the raw learning rate is not the whole story. Adam normalizes gradients by the running average of squared gradients, so the effective update is approximately:

effective step ≈ lr / sqrt(v̂)

When the second-moment estimate v̂ equals 1, the effective step equals the learning rate itself — which is why 1e-3 "feels like" a step of 1e-3 in well-conditioned settings. The calculator also shows the value scaled by sqrt(batch_size = 256), giving intuition for the common linear-scaling rule when you change batch sizes.

Linear decay vs cosine annealing

The preview table evaluates both schedules over 10 steps using these exact formulas, where frac = i / (steps − 1):

linear decay:      lr_i = lr × (1 − frac)
cosine annealing:  lr_i = 0.5 × lr × (1 + cos(π × i / (steps − 1)))

Linear decay drops the rate at a constant pace. Cosine annealing starts slowly, falls steeply through the middle of training, and flattens out near zero — a shape that often yields better final accuracy because the model spends more time at both high and low rates. Comparing the two side by side for your base rate makes the choice concrete instead of theoretical.

Practical Use Cases

Reproducing a paper's hyperparameters

A paper reports "AdamW with lr 3e-4 and cosine decay." Type 3e-4 into the calculator, confirm it equals 0.0003 (or 0.03%), verify it against the AdamW default of 1e-3 (it's 3x smaller — typical for transformers), and preview what cosine annealing does to it over your planned number of steps before writing a single line of code.

Porting a model between frameworks

Moving from PyTorch to Keras means re-expressing every optimizer argument. The conversion panel shows both framework-ready values at once, and the snippet section gives you the exact constructor syntax, so optim.Adam(model.parameters(), lr=1e-3) becomes tf.keras.optimizers.Adam(learning_rate=1e-3) without ambiguity.

Fine-tuning a pretrained model

Fine-tuning typically runs 10–100x below pretraining rates — 5e-5 is a classic BERT-era choice. Click the 5e-5 preset, check the effective step interpretation for your Adam variant, and compare linear versus cosine decay to decide how aggressively to anneal during a short 3-epoch fine-tune.

Teaching and team communication

When reviewing configs with teammates or students, converting 1e-3 to "one tenth of one percent" instantly communicates scale. The optimizer defaults table doubles as a cheat sheet for anyone learning why Adam's default differs from SGD's by exactly one order of magnitude.

Best Practices

• Start from the optimizer default, then move in powers of ten. The defaults table encodes decades of community experience; sweep 1e-2 → 1e-3 → 1e-4 rather than arbitrary values.
• Double-check exponents when copying from papers. Paste the value into the calculator and confirm the decimal form before it enters your config.
• Match the schedule to training length. Cosine annealing shines over longer runs; for very short runs, the 10-step preview shows you whether either schedule changes the rate meaningfully at all.
• Scale deliberately. If you change batch size, use the batch-scaling view to sanity-check the adjusted rate instead of guessing a multiplier.
• Keep units consistent in configs. Store rates in decimal or scientific notation in code; reserve percentages for presentations. Mixing formats in one config file is a common source of bugs.
• Verify after pasting. Even with copy-ready snippets, re-read the pasted value once — clipboard workflows are reliable, but a quick visual check costs nothing.

Start Tuning With Confidence

Learning rate work is equal parts math and bookkeeping, and the bookkeeping is where mistakes happen. The Learning Rate Calculator handles the notation, the optimizer references, and the schedule math in one fast, private page — so you can spend your attention on the experiment, not on counting zeros. Open it, type your rate, and get back to training.

Related Tools You Might Like

• Epoch Batch Converter — convert batches of Unix timestamps to readable dates, with timezone selection and CSV export.
• Matrix Calculator — multiply, invert, and decompose matrices for linear algebra and model math.
• Statistics Calculator — compute mean, variance, and other summary statistics for your metrics and data.

Happy tuning!