Neural Network Trainer Calculator

It depends on the architecture, and the answer at the defaults surprises people: with 784 inputs and 256 hidden units, the input-to-first-hidden matrix alone is 200,704 of the 269,322 total parameters, so Input Size dominates. Hidden Units per Layer enters quadratically in each hidden-to-hidden matrix -- doubling it roughly quadruples those -- so it takes over once the network is deep enough, or the input narrow enough, for those matrices to outweigh the input projection.

No -- Batch Size has zero effect on Total Parameters, since parameter count depends only on the network's architecture (layer sizes), not on how many samples you process at once. Batch Size does change Total Memory, because larger batches require more memory to hold activations during the forward and backward pass.

Learning Rate Schedule projects how your learning rate decays over training (constant, step, exponential, or cosine annealing) and recommends an optimizer; Batch Size & Training turns measured per-step timings and per-sample memory into batches per epoch, total training time and the largest batch your GPU can hold -- it asks you for those measurements rather than guessing them from the architecture; Regularization scores dropout, L2 weight decay, batch normalization, and data augmentation into an overall regularization strength and overfitting risk; and Activation Functions compares ReLU, Sigmoid, Tanh, GELU, and Swish by formula, gradient behavior, and relative compute cost.

Complexity Rating is a simple threshold classification of Total Parameters: Simple at 10,000 parameters or fewer, Medium above 10,000 up to 1 million, Large above 1 million up to 1 billion, and "Very Large (LLM scale)" above 1 billion parameters. It's a rough sizing label, not a claim that your specific architecture matches a real large language model's design.

Entirely from what you enter. Training time is simply batches per epoch x epochs x your measured seconds-per-batch, and Max Batch Size is your GPU memory minus your model and optimizer memory, divided by your measured per-sample activation memory. The calculator does not try to predict step time or activation footprint from layer sizes -- those depend on the GPU, the precision, the kernels and the framework, and any figure derived from architecture alone would be wrong by orders of magnitude. Time twenty real training steps and measure peak memory at batch size 1, then enter those.