Extrusion Output Calculator for AI Layers

Calculator Form

Operation

Input Height

Input Width

Input Channels

Filters or Output Channels

Repeated Layers

Kernel Height

Kernel Width

Stride Height

Stride Width

Padding Height

Padding Width

Dilation Height

Dilation Width

Output Padding Height

Output Padding Width

Groups

Batch Size

Bytes Per Value

Include Bias

Example Data Table

Case	Operation	Input	Kernel / Stride / Pad	Filters	Repeats	Output	Parameters	MACs Per Batch
Vision Stem	Convolution 2D	224 × 224 × 3	7×7 / 2×2 / 3×3	32	1	112 × 112 × 32	4,736	59,006,976
Residual Stack	Convolution 2D	56 × 56 × 64	3×3 / 1×1 / 1×1	128	3	56 × 56 × 128	368,640	1,156,055,040
Decoder Upsample	Transposed Convolution 2D	28 × 28 × 128	4×4 / 2×2 / 1×1	64	2	112 × 112 × 64	196,736	1,233,125,376
Pooling Pyramid	Max Pooling 2D	64 × 64 × 64	2×2 / 2×2 / 0×0	64	2	16 × 16 × 64	0	327,680

Formula Used

For convolution and pooling, output height uses floor(((input height + 2 × padding height − effective kernel height) ÷ stride height) + 1). Width uses the same structure.

Effective kernel height equals dilation height × (kernel height − 1) + 1. Effective kernel width follows the same rule.

For transposed convolution, output height equals ((input height − 1) × stride height) − (2 × padding height) + effective kernel height + output padding height.

Trainable parameters for grouped convolution equal filters × (input channels ÷ groups) × kernel height × kernel width, plus bias when enabled.

Estimated MACs per batch equal output height × output width × output channels × kernel height × kernel width × grouped input channels. Pooling uses window operations instead of learned parameters.

Approximate output memory equals batch size × output height × output width × output channels × bytes per value.

How to Use This Calculator

Choose the layer operation you want to inspect.
Enter input height, input width, and input channels.
Set filters, kernel size, stride, padding, and dilation.
Use output padding only for transposed convolution.
Set groups for grouped or depthwise style designs.
Pick batch size and numeric precision for memory estimates.
Use repeated layers to test stacked blocks quickly.
Submit the form and review summary, layer table, graph, and export files.

FAQs

1. What does this calculator estimate?

It estimates output tensor shape, trainable parameters, approximate MACs, activation memory, effective kernel size, and repeated-layer shape changes for common 2D operations.

2. When should I use transposed convolution?

Use it when you need learned upsampling in decoders, generators, or segmentation heads. It expands spatial dimensions while still applying trainable kernels.

3. Why do grouped layers matter?

Groups split channels into smaller paths. They reduce parameter count and compute, and they support depthwise or channel-partitioned designs.

4. Why can output size become invalid?

A large kernel, heavy dilation, or small input can push the computed spatial size below one. Adjust kernel, stride, padding, or input size.

5. Are MACs and FLOPs the same?

Not exactly. Many practitioners treat one MAC as roughly two FLOPs for multiply and add. This page reports MACs directly.

6. Does memory include optimizer state?

No. The estimate focuses on output activations and input activations. Training usually needs more memory for gradients, parameters, and optimizer tensors.

7. Can I model pooling with this tool?

Yes. Max pooling and average pooling both use the same output-size logic here, but they add no trainable parameters.

8. Why repeat layers in one run?

Repeated layers help you inspect stacked blocks quickly. You can see shrinking or expanding shapes, cumulative parameters, and total compute in one pass.