---

name: deep-models description: "Operate CausalML optional deep neural estimators: DragonNet and CEVAE across TensorFlow, Torch/Pyro, and JAX backends." disable-model-invocation: true metadata: disco-role: operating license: NOASSERTION

deep-models

Use this sub-skill for CausalML 0.17.0 workflows that need optional neural causal estimators:

• TensorFlow causalml.inference.tf.DragonNet
• Torch/Pyro causalml.inference.torch.CEVAE
• JAX causalml.inference.jax.DragonNet
• JAX causalml.inference.jax.CEVAE

Prefer keyword arguments (X=, treatment=, y=) for all fit and fit_predict calls. The current neural-estimator order is X, treatment, y, p; using keywords avoids argument-order migration warnings and makes the treatment/outcome roles explicit.

Route here when

• The user asks for DragonNet, CEVAE, neural causal estimators, hidden-confounder CEVAE, or optional tf, torch, or jax backends.
• The user needs tiny CPU examples for backend smoke checks before running larger experiments.
• The user asks how to save/load TensorFlow or JAX DragonNet checkpoints, or JAX CEVAE checkpoints.
• The user sees optional-backend import errors, CUDA warnings, Pyro/JAX runtime issues, or shape errors in neural CATE estimation.

For non-neural meta-learners, TMLE, IV/DRIV, or serialization helpers, use ../causal-estimation/SKILL.md. For AUUC/Qini/cumulative-gain scoring and decision optimization after producing ITE/CATE predictions, use ../analysis-and-decision/SKILL.md.

Reference map

• references/backend-setup.md: extras, imports, CPU/GPU expectations, and backend smoke checks.
• references/dragonnet.md: TensorFlow and JAX DragonNet API recipes, outputs, and save/load.
• references/cevae.md: Torch/Pyro and JAX CEVAE API recipes, tiny settings, and checkpoint differences.
• references/troubleshooting.md: common import, runtime, training, shape, and checkpoint failures.

Minimal operating sequence

1. Choose the backend and install the matching optional extra; do not assume one deep backend implies the others are installed.
2. Build numeric X with shape (n_samples, n_features), binary treatment with shape (n_samples,), and numeric y with shape (n_samples,).
3. Start with a tiny CPU configuration to prove imports, shapes, and method availability.
4. Fit with keyword arguments and inspect only finite ITE/propensity outputs before using metrics or policy decisions.
5. Save only where the wrapper supports it: TensorFlow DragonNet, JAX DragonNet, and JAX CEVAE. Torch/Pyro CEVAE exposes no CausalML save/load wrapper.