Keras Layer Stack to TensorFlow Serving Graph (2D)

Architecture

Hidden layers 3

Neurons / layer 9

Deployment

INT8 quantization

Live Readout

Inference latency

— ms

Training loss

—

Graph nodes

—

Parameters

—

How it works

The same feed-forward network is drawn two ways. In Keras eager mode every neuron is its own node and every matmul, bias-add and ReLU dispatches separately. Switching to the TF serving graph fuses each hidden layer's matmul + bias + activation into one solid block (TensorFlow's Grappler optimizer does this automatically on export), which is why the node count and latency both drop. Enabling INT8 quantization shrinks the fused blocks further and drops latency again, the same way TensorFlow Lite / TF-TRT quantize weights for edge and server deployment.

Latency model — base cost scales with depth × width; fusion cuts it to ~42%, quantization on top of fusion cuts it to ~55% of that.
Parameter count — sum of weights + biases between consecutive layers, exactly as Keras' model.summary() would report.
Pulses travelling left→right show a forward pass; press train to animate a loss curve while the network "trains" before the graph is frozen for export.

KERAS EAGER

Keras op / weight Fused TF block INT8 quantized