simplify trainer quickstart (official converter) + ONNX export config

README.md

@@ -51,27 +51,50 @@ Pick a model and run it:
 
 ### As a trainer (want to train your own)
 
+NERVE is a normal [traiNNer-redux](https://github.com/the-database/traiNNer-redux)
+architecture — no special steps.
+
 ```bash
-# 1. Put the architecture in traiNNer-redux
+# 1. Add the architecture (auto-registers: traiNNer scans *_arch.py)
 cp nerve_arch.py <traiNNer-redux>/traiNNer/archs/
 
-# 2. Train (copy a config from configs/ and point the paths at your data)
-python train.py -opt configs/4x_NERVE_release.yml --auto_resume
+# 2. Copy a training config and point it at your data
+cp configs/4x_NERVE_release.yml <traiNNer-redux>/options/
+#    -> edit dataroot_gt / dataroot_lq in that file
+
+# 3. Train
+cd <traiNNer-redux>
+python train.py -opt options/4x_NERVE_release.yml --auto_resume
 
-# 3. Export to dynamic ONNX
-python scripts/export_dynamic.py --dim 64 --n-blocks 24 \
-    --checkpoint <run>/models/net_g_ema_latest.safetensors --scale 4
+# 4. Export to optimized dynamic ONNX (official converter)
+cp configs/4x_NERVE_onnx.yml <traiNNer-redux>/options/
+#    -> set path.pretrain_network_g to your trained checkpoint
+python convert_to_onnx.py -opt options/4x_NERVE_onnx.yml
 ```
 
-That's the whole workflow. The net is a single file, the config is a single
-file, and the checkpoint is a single file — no fusion step, no paired
-fused/unfused weights.
+That is the whole workflow. The net is one file, the config is one file, the
+checkpoint is one file — **no fusion step, no paired fused/unfused weights.**
+
+- `4x_NERVE_onnx.yml` exports dynamic-shape ONNX (opset 20) and can emit
+  `fp32`, `fp16`, or `bf16` (`onnx.dtype`), with OnnxSlim optimization and a
+  built-in PyTorch-vs-ONNX verify pass.
+- Prefer no extra dependencies? `scripts/export_dynamic.py` is a tiny
+  standalone exporter (torch + onnxruntime only).
+
+**Start from a released pretrain** (skip the expensive early training) by
+setting `path.pretrain_network_g` in your training config to a file from
+`models/`:
+
+| training stage | warm-start from |
+|---|---|
+| `4x_NERVE_release.yml` | `models/4x_NERVE_release.safetensors` (or the 2x one with `strict_load_g: false`) |
+| `4x_NERVE_OTF_fidelity.yml` | `models/4x_NERVE_release.safetensors` |
+| `4x_NERVE_OTF_gan.yml` | your OTF fidelity checkpoint |
 
 **Training tips** (all NERVE models follow these):
 - bf16 AMP is fine; no gradient clipping needed (norm-free net).
 - `--auto_resume` is recommended for long runs.
-- For real-world input, continue from the clean release into the OTF fidelity
-  then the OTF GAN config — that's the chain the released models use.
+- For real-world input, follow the chain: clean release -> OTF fidelity -> OTF GAN.
 
 ---
 
@@ -190,7 +213,7 @@ because it has no attention ops.
 
 ```
 nerve_arch.py                    the architecture (single file)
-configs/                         example training configs (release, OTF fidelity, OTF GAN, 2x)
+configs/                         example training configs (+ 4x_NERVE_onnx.yml for ONNX export)
 models/                          pretrained checkpoints (see table above)
 onnx/                            dynamic fp32 ONNX (opset 20, onnxslim-optimized)
 ncnn/                            ncnn fp16 param/bin (fixed 256x256 input)

configs/4x_NERVE_onnx.yml

@@ -0,0 +1,33 @@
+# yaml-language-server: $schema=https://raw.githubusercontent.com/the-database/traiNNer-redux/refs/heads/master/schemas/redux-config.schema.json
+####################
+# General Settings
+####################
+name: 4x_NERVE
+scale: 4  # 1, 2, 3, 4
+num_gpu: auto
+
+#####################
+# Network Settings
+#####################
+network_g:
+  type: nerve
+
+############################
+# Pretrain and Resume Paths
+############################
+path:
+  # Set this to the checkpoint you trained (or a released model).
+  pretrain_network_g: experiments/4x_NERVE_release/models/net_g_ema_latest.safetensors
+
+#########################
+# ONNX conversion options
+#########################
+onnx:
+  dynamo: false  # legacy exporter; NERVE also exports with dynamo: true if you prefer.
+  dtype: fp32  # fp32, fp16, or bf16.
+  bf16_exclude_depthwise: true
+  opset: 20
+  shape: 1x3xHxW  # dynamic H/W. Use 1x3x256x256 for a fully static model.
+  wrap_5d_to_4d: true
+  verify: true  # verify ONNX vs PyTorch after export
+  optimize: true  # run OnnxSlim on the exported model