Calculation results
Architecture flow
Layer-by-layer summary
| # | Layer | Type | Input | Output | Weights | Biases | Trainable | Non-trainable | Total | Share | Memory |
|---|
Parameter distribution
Memory estimate
Formula used
Dense: (input features × output units) + output biases Convolution: ((input channels ÷ groups) × kernel elements × filters) + filter biases Depthwise convolution: input channels × depth multiplier × kernel elements + biases Separable convolution: depthwise parameters + pointwise parameters Embedding: vocabulary size × embedding dimension Simple RNN: hidden × (input + hidden + bias terms) GRU: 3 × hidden × (input + hidden + bias terms) LSTM: 4 × hidden × (input + hidden + bias terms) Attention: Q, K, V projections + output projection LoRA: rank × (input dimension + output dimension) × adapted modules Convolution output: floor((N + 2P - D(K - 1) - 1) ÷ S + 1)
How to use
Set model-wide precision and memory assumptions first. Choose an optimiser and batch size. Add activation estimates only when needed.
Add each neural network layer in order. Enter dimensions, kernels, heads, or hidden sizes. Freeze layers when their weights will not train.
Press Calculate parameters to build the report. Review layer shares and memory estimates. Export the result using one available format.
Example data
| Layer | Configuration | Parameters |
|---|---|---|
| Input | 784 features | 0 |
| Dense 1 | 784 × 128 + 128 | 100,480 |
| Dense 2 | 128 × 64 + 64 | 8,256 |
| Output | 64 × 10 + 10 | 650 |
| Total | — | 109,386 |
Frequently asked questions
What is a trainable parameter?
A trainable parameter is updated during optimisation. Weights and enabled biases are common examples. Frozen values remain unchanged during training.
Why do convolution layers use fewer parameters?
Convolution layers reuse kernels across spatial positions. Weight sharing reduces parameter growth. Large feature maps can still require substantial activation memory.
Are batch-normalisation statistics trainable?
Scale and offset values are usually trainable. Moving mean and variance are usually non-trainable. Framework implementations may differ slightly.
Does batch size change parameter count?
Batch size does not change learned parameters. It changes activation memory and throughput. Larger batches can require much more training memory.
How is Adam memory estimated?
Adam commonly stores two optimiser-state tensors. Gradients require another parameter-sized tensor. Mixed precision may add master weights.
What does pruning percentage mean?
Pruning removes or masks selected weights. Sparse storage may reduce effective memory. Actual speed gains depend on hardware and libraries.
How are bidirectional recurrent layers counted?
Bidirectional layers maintain two recurrent directions. Their parameters are approximately doubled. Output features also commonly double.
Are preset results exact?
Presets demonstrate common layer patterns. Named architectures vary across implementations. Confirm exact settings against your framework model summary.
Are parameters the same as FLOPs?
Parameters measure stored learned values. FLOPs estimate computation for a forward or training pass. Neither metric alone describes total performance.