Post-Training Quantization (PTQ) Scheduling

This initial stage involves curating a representative dataset, distinct from training data, to profile the model's activation statistics. The calibration set must capture the operational data distribution to ensure accurate range estimation for weights and activations.

• Purpose: Estimate dynamic ranges for quantization scaling factors.
• Key Consideration: Dataset size is typically small (100-1000 samples) to avoid computational overhead.
• Common Practice: Use a random subset of the validation dataset that preserves class distribution.

The model is executed in inference mode on the calibration data to collect statistical profiles of each layer's activations. The min/max ranges or histograms are recorded to determine the quantization parameters (scale and zero-point).

• Methods: Min-Max, Moving Average Min-Max, or Entropy-based (KL Divergence) calibration.
• Output: A set of quantization parameters (scale, zero_point) for each quantizable tensor.
• Critical Detail: Asymmetric quantization often requires separate min/max for positive and negative activations.

In this stage, the pre-trained FP32 weights are discretized into lower-bit integer values (e.g., INT8). This is a static process as weights are constant. The quantization parameters are derived directly from the weight tensor's distribution.

• Process: weight_int8 = clamp(round(weight_fp32 / scale) + zero_point)
• Bit-Widths: Common targets are 8-bit (INT8) or 4-bit (INT4) for weights.
• Per-Channel vs. Per-Tensor: Per-channel quantization uses separate scale/zero-point for each output channel, offering higher accuracy at the cost of more parameters.

Also known as fake quantization, this optional diagnostic step inserts quantization and dequantization (QDQ) nodes into the model graph. The model runs in FP32, but the operations simulate the rounding and clamping effects of integer math.

• Purpose: Validate the chosen quantization parameters and estimate accuracy loss before full deployment.
• Tool Support: Frameworks like TensorRT, PyTorch FX, and ONNX Runtime provide simulation modes.
• Output: A quantized computation graph and a baseline accuracy metric for the quantized model.

The computational graph is optimized for the target integer kernel. This involves fusing common operation sequences (e.g., Conv + BatchNorm + ReLU) into a single, quantizable operator. This reduces memory movement and leverages optimized low-precision kernels.

• Example Fusion: Conv2D -> BatchNorm -> ReLU becomes a single QLinearConv operator.
• Hardware Dependency: The valid fusion patterns are dictated by the supported operators of the target inference engine (e.g., TFLite, Core ML).

If simulation reveals unacceptable accuracy loss, a limited fine-tuning loop may be initiated. This is a hybrid of PTQ and QAT where the quantized model is briefly trained with a low learning rate to recover performance.

• Trigger: Typically used when accuracy drop exceeds a pre-defined threshold (e.g., >1%).
• Process: Uses Straight-Through Estimator (STE) to approximate gradients through the non-differentiable quantization operation.
• Resource Note: Adds computational cost, moving PTQ closer to a limited retraining scenario.

A QAT schedule outlines the phased integration of quantization simulation during training, not after. This contrasts with PTQ, which is purely post-hoc.

• Key Phases: Typically includes a full-precision warmup, followed by the introduction of fake quantization nodes to simulate rounding error during forward/backward passes, and concludes with fine-tuning.
• Objective: To produce a model whose weights are already robust to low-precision representation, often achieving higher accuracy than PTQ for the same target bit-width.
• Trade-off: Requires access to the original training pipeline and computational resources for additional training cycles.

A pruning schedule is a predefined strategy dictating the timing, rate, and criteria for removing parameters from a neural network. It is the structural analogue to a PTQ schedule.

• Core Components: Defines the sparsity target, the pruning criterion (e.g., weight magnitude), the frequency of pruning steps, and any fine-tuning intervals.
• Common Strategies: One-shot pruning removes weights in a single step; iterative magnitude pruning uses repeated prune-and-fine-tune cycles; gradual pruning increases sparsity smoothly over many training steps.
• Scheduling Goal: To minimize disruptive loss spikes and allow the network to adapt to its sparser architecture.

Layer-wise sensitivity measures how much a model's accuracy degrades when a specific layer is compressed (quantized or pruned). This analysis is foundational to intelligent PTQ and pruning schedules.

• Purpose: Informs non-uniform compression policies. A schedule may apply aggressive 4-bit quantization to insensitive layers while preserving 8-bit or FP16 precision in critical, sensitive layers.
• Measurement: Typically evaluated by compressing one layer at a time on a calibration set and observing the change in loss or accuracy.
• Outcome: Enables mixed-precision quantization schedules that allocate bit-widths per layer to optimize the accuracy-efficiency Pareto frontier.

Automated Model Compression is a framework that uses search algorithms, like reinforcement learning, to automatically determine the optimal compression policy (e.g., per-layer bit-width for PTQ) without manual scheduling.

• Mechanism: An RL agent takes the model architecture and a hardware constraint (e.g., latency target) as input. It iteratively samples compression actions (e.g., "quantize Layer 3 to 4-bit"), receives a reward based on the resulting accuracy and speed, and learns an optimal policy.
• Contrast with Manual Scheduling: AMC discovers complex, non-uniform schedules that might be non-intuitive for a human engineer, often achieving better performance.
• Scope: Can jointly schedule quantization and pruning decisions.

Multi-stage compression is a scheduling paradigm where different compression techniques are applied in separate, sequential phases. A PTQ schedule is often the final stage in such a pipeline.

• Typical Pipeline: 1) Pruning to remove redundant parameters, 2) Fine-tuning to recover accuracy, 3) Quantization (PTQ or QAT) to reduce precision, 4) Optional second fine-tuning.
• Rationale: Applying techniques in isolation simplifies the optimization problem. Pruning first reduces the parameter count, which can sometimes make the model more amenable to subsequent quantization.
• Scheduling Challenge: Requires careful calibration of each stage's intensity and recovery time to avoid compounding errors.

HW-NAS is a search methodology that discovers efficient network architectures by directly incorporating target hardware metrics into the search objective. It represents a pre-training alternative to post-training compression scheduling.

• Objective Function: Doesn't just search for high-accuracy models. It searches for models that maximize accuracy while minimizing latency, energy consumption, or memory footprint on specific hardware (e.g., a mobile NPU).
• Contrast with PTQ Scheduling: HW-NAS designs a model that is inherently efficient, while PTQ scheduling takes an existing model and adapts it. They are complementary strategies.
• Outcome: Produces a compact, hardware-optimized model that can then be further refined with PTQ or QAT schedules.