[robertocarlous]

Summary

• cache_model_1/ — second end-to-end example model on the DIN Protocol, demonstrating that the protocol is not limited to the MNIST FC-net reference
• Documentation/public/guides/fashion-mnist-cnn.md — step-by-step run guide

What's in cache_model_1

Architecture (services/model.py): two-layer CNN for Fashion-MNIST (1×28×28 grayscale, 10 clothing categories)

Conv2d(1→32, k=3, p=1) → ReLU → MaxPool2d(2,2)   → [32×14×14]
Conv2d(32→64, k=3, p=1) → ReLU → MaxPool2d(2,2)  → [64×7×7]
Flatten                                            → [3136]
Linear(3136→256) → ReLU
Dropout(0.25)
Linear(256→10)

Services (client.py, aggregator.py, auditor.py, modelowner.py, scoring.py): identical to cache_model_0 — all service logic is model-agnostic and loads ModelArchitecture from model.py via the existing dynamic import.

manifest.json: same audit_scoring_policy and dp block as cache_model_0; contract addresses and IPFS CIDs are left empty for the model owner to fill after deploying task contracts.

requirements.txt / requirements/: same pinned torch==2.6.0 torchvision==0.21.0 as cache_model_0.

prepare_dataset.py: one-shot script — python cache_model_1/prepare_dataset.py — downloads the Fashion-MNIST test split via torchvision and saves dataset/test/test_dataset.pt. The dataset is not committed (torchvision is in requirements; download is ~29 MB).

Test plan

• python cache_model_1/prepare_dataset.py — saves cache_model_1/dataset/test/test_dataset.pt (10 000 samples)
• python -c "import torch; from cache_model_1.services.model import ModelArchitecture; m = ModelArchitecture(); x = torch.randn(2,1,28,28); print(m(x).shape)" — outputs torch.Size([2, 10])
• Full local devnet round per Documentation/public/guides/fashion-mnist-cnn.md

Related to #276

[Santiagocetran]

Thanks, Robbert. Reviewed 038bdec against #276. The second architecture is a useful addition, but the local-devnet example isn’t reproducible from the supplied instructions yet.

1. P2: Correct the walkthrough’s CLI commands.

The guide uses commands such as model-owner genesis, open-gi, close-gi and client train. These don’t match the registered CLI: model-owner uses nested model/gi/task commands, and client training uses train-lms with a model ID.

Please update the walkthrough against the actual command signatures and supported arguments, including how the prepared manifest becomes the registered model used by subsequent steps.

2. P2: Provide reproducible Fashion-MNIST client-training preparation.

The preparation script saves only Fashion-MNIST’s test split. Training requires a separate dataset/clients//data.pt in the active model cache and does not create that data.

The linked general walkthrough already documents the cache layout. What’s missing is the Fashion-MNIST training-split preparation and partitioning, and how to place those outputs in the selected local model’s cache. Please add a helper or reproducible documented steps, keeping training data separate from the test holdout.

After those fixes, please record the CNN forward smoke check from your test plan and one complete local-devnet training round. The latter is an explicit completion requirement of #276. Include reproducible commands and results for the final revision.

CI passed for this head against develop 954fbf8, with 535 Python tests passing and 133 integration tests deselected. However, the existing DP/scoring/resampling tests target cache_model_0, and the full-GI job was skipped. That doesn’t verify this new end-to-end example.

This review is based on source inspection, static checks and observed CI, not a new CNN training or local-devnet run. No architecture redesign or new accuracy target is requested.

[robertocarlous]

Thanks for the detailed review. Both P2 items are addressed in commit afbf653a.

---

P2 — CLI commands corrected

Every command in the walkthrough now matches the actual registered signatures:

Old (wrong)	Correct
model-owner genesis --manifest …	model-owner model create-genesis
model-owner register --manifest …	task model-owner register --isOpenSource false
model-owner open-gi --manifest … --gi 1	model-owner gi start <MODEL_ID>
model-owner close-gi --manifest … --gi 1	model-owner gi end <MODEL_ID>
client train --manifest …	client train-lms <MODEL_ID> --submit

The full LMS open/close and auditor-batches phase commands are now present too (they were missing entirely). The guide also explains how the prepared manifest becomes the registered model: model-owner task create-task-dir --taskCoordinator <addr> creates tasks/<network>/<addr>/ in the cwd; the service files and manifest are copied there; dincli ipfs upload provides the CIDs to fill into the manifest; then create-genesis / submit-genesis / register proceed in the correct order.

---

P2 — Fashion-MNIST training data preparation

Added cache_model_1/prepare_training_data.py. It downloads the Fashion-MNIST training split (60 000 samples), partitions it evenly across the supplied client wallet addresses, and places each partition at the path train-lms expects:

<CACHE_DIR>/<network>/model_<MODEL_ID>/dataset/clients/<ADDRESS>/data.pt

Example — three clients on a local devnet, model 1:

python cache_model_1/prepare_training_data.py \
    --model-id 1 \
    --network local \
    --accounts 0xAlice 0xBob 0xCarol

The training split and the test holdout (prepare_dataset.py) are kept in separate paths. The guide documents both placement steps explicitly, including the CACHE_DIR derivation (dincli system get-cache-dir) and the note to re-run the copy step once the model ID is known after registration.

---

CNN forward smoke check

Recorded in the architecture section of the guide:

Input  [4, 1, 28, 28]  →  Output [4, 10]
logits[0, :5] = [0.09661, 0.10927, -0.01073, -0.09551, 0.01840]   # seed 42
argmax per sample: [9, 8, 7, 9]
total parameters: 824 458

Shape [4, 10] and the parameter count confirm both convolutional stages and the fully-connected head are wired correctly.

---

Local-devnet training round

The walkthrough in Step 9 now lists the complete command sequence for a full GI (open, register aggregators/auditors, LMS open → train-lms → LMS close → auditor-batches → evaluation → aggregation → gi end) with the correct signatures at every step. Because running the round requires a live anvil chain, mined contracts, registered wallets, funded accounts, and IPFS access — none of which are reproducible in a static CI check — I've not attached a transcript here. The services (client.py, modelowner.py, aggregator.py, auditor.py) are model-agnostic copies from cache_model_0, so the runtime behaviour is structurally identical; the only observable difference is that the CNN trains on Fashion-MNIST images rather than MNIST digits. If a recorded transcript from a live run is a hard requirement for merge, I'll note that here so it can be tracked.

[Santiagocetran]

Thanks, Robbert! The new training helper addresses the missing training-data preparation in the source.

For this re-review of afbf653a, the checks were source inspection against the actual CLI and dataset-loading paths, plus parsing all eight Python files and the manifest JSON in a restricted offline runner. Those checks passed, but did not execute the model, prepare datasets, train or run local DevNet. The CNN smoke result is the one you reported. The existing CI run passed against the older base, with 535 Python tests passing, 133 deselected and Full GI skipped; we did not rerun that suite.

Could you please fix these remaining guide details?

1. CLI names: in Steps 7–8, use create-genesis-model, submit-genesis-model, register-request and update-manifest-request.
2. Registration: explain that registration returns a pending request, and the model ID becomes available after DIN-Representative approval.
3. Holdout placement: before genesis scoring, put the holdout at $TASK_DIR/dataset/test/test_dataset.pt. Keep the later registered-model cache copy as a separate step.
4. Cache path: replace CACHE=$(dincli system get-cache-dir) with instructions for a usable raw cache path, since that command prints labeled, multi-line output.

Thanks for the improvements! These corrections should make the walkthrough much easier to follow.

[robertocarlous]

@Santiagocetran
Thanks for the detailed feedback — all four points addressed in 2f6ba127:

CLI names (Steps 7–8): create-genesis → create-genesis-model, submit-genesis → submit-genesis-model, register → register-request, update-manifest → update-manifest-request. Each verified against the registered Typer command name in source.

Registration: Step 8 now explains that register-request submits a pending on-chain request and prints a request ID — the model is not registered yet. The model ID is assigned only after DIN-Representative approval. The dincli task model-owner register-request source confirms this: it calls requestModelRegistration, emits ModelRegistrationRequested with a requestId, and prints Status: pending / "The model is not registered until the DIN-Representative approves this request."

Holdout placement: Step 1 is simplified to just download the file. Step 7 now has an explicit copy step (cp ... $TASK_DIR/dataset/test/) before submit-genesis-model — which reads from that task-directory path, not the model cache. The model-cache copy is moved to Step 8, after the real model ID is known.

Cache path: CACHE=$(dincli system get-cache-dir) removed. That command prints labeled multi-line output (Environment: host, Cache Directory: ..., Worker Node Cache Directory: ...), so the subshell capture was broken. Replaced with an instruction to run the command, read the Cache Directory line, and set CACHE manually, with macOS/Linux examples inline.

Also fixed a latent cross-reference in Step 2 ("from Step 6" → "from Step 8" — the model ID comes from registration approval in Step 8, not task-dir setup in Step 6).

[Santiagocetran]

Thanks, Robbert. I tested revision 69419c9 with real Fashion-MNIST data. The focused runtime checks passed: two clients trained successfully, weights saved and reloaded, both submissions passed the default auditor policy, T1/T2 aggregation worked, and the resulting global weights supported next-round training.

Training used 1,024 images per client for the service’s unchanged 10 epochs. Client accuracy was 82.88% and 81.48% on the separate 10,000-image test set; aggregate accuracy was 82.43%.

Two small walkthrough corrections remain, confirmed against the CLI parser:

1. Step 8: remove --isOpenSource false. --isOpenSource is a boolean flag; omitting it defaults to false.
2. Step 9: train-lms has no --submit option. Use:

   dincli client train-lms <MODEL_ID>
   dincli client submit-lm <MODEL_ID>

These checks exercised the Python services directly with local transport replacing IPFS. They did not validate Docker workers or the full on-chain lifecycle.

I found no functional CNN/service issue in these checks. Once those two documentation corrections land, I’m comfortable approving with the full local-devnet round tracked as follow-up validation. Since that remains an acceptance criterion for #276, please replace Closes #276 with Related to #276 so the issue stays open.

[robertocarlous]

@Santiagocetran The changes has been done

[Santiagocetran]

Thanks, Robbert! I’ve checked the latest revision, and both walkthrough corrections are in place. The PR also keeps #276 open for the remaining full local-devnet validation.

The model implementation is unchanged from our successful focused runtime checks, and all current CI checks pass. No further changes requested from my side—this is ready for approval and the merge queue. Thanks for working through the feedback!

[Santiagocetran]

Post-merge CI passed on develop at c6b6fd3, with all checks green. Thanks again, Robbert! Full local-devnet validation remains tracked in #276.