---

name: image-projection description: "Project an image into iGAN latent space and reconstruct it with iGAN_predict.py, safe command builders, AlexNet planning, and projection troubleshooting." disable-model-invocation: true metadata: disco-role: operating license: MIT

iGAN Image Projection Router

Use this sub-skill when the task is to project a real image into the latent space of an iGAN DCGAN model and save the reconstructed image.

This router covers the iGAN_predict.py workflow, the solver choices cnn, opt, and cnn_opt, the AlexNet conv4 feature dependency used by the published command, and the model-file requirements for predictor parameters.

Do not use this sub-skill for random sampling, downloading DCGAN model-zoo artifacts, interactive editing, non-UI constraint maps, or training the predictor network. Route those tasks to sibling skills instead.

Read First

• Projection workflow recipes: references/projection-workflows.md
• CLI and internal API facts: references/api-reference.md
• Projection failure modes: references/troubleshooting.md
• Dry projection command builder: scripts/build_projection_command.py
• AlexNet URL/target planner: scripts/igan_alexnet_urls.py

When To Use

Use this sub-skill for tasks such as:

• Build the exact iGAN_predict.py command for a given input image.
• Choose between cnn, opt, and cnn_opt projection solvers.
• Explain where the output reconstruction image will be written.
• Preflight required artifacts for projection without running Theano.
• Plan the AlexNet conv4 pickle target without performing a download.
• Diagnose missing AlexNet, missing DCGAN model files, or predictor-param errors.

Route Elsewhere

• Random sample grids or DCGAN model-zoo setup: use ../model-inference/SKILL.md.
• Predictor training, HDF5 datasets, batchnorm refresh, or model packing: use ../training-data/SKILL.md.
• Color, mask, edge, or sketch constraints for headless generation: use ../constraint-generation/SKILL.md.
• Interactive PyQt4 drawing workflows: use ../interactive-ui/SKILL.md.

Projection Preconditions

Projection is a legacy native workflow. Before proposing a runtime run, confirm:

1. The user has an iGAN runtime checkout or equivalent files containing iGAN_predict.py.
2. The Python environment can import Theano, Lasagne, NumPy, SciPy, PIL/Pillow, and iGAN local modules.
3. The intended backend is compatible with old Theano CUDA/cuDNN expectations, or the user accepts a slow/experimental CPU attempt.
4. The DCGAN model file exists at models/<model_name>.<model_type> or the user supplies --model_file.
5. The model file includes generator parameters and, for the current projection script, predictor parameters and predictor batchnorm arrays.
6. The AlexNet feature pickle exists at models/caffe_reference_conv4.pkl for the default feature layer.
7. The input image is readable by PIL/Pillow; the script resizes it internally to the model resolution and then back to the original dimensions.

If any runtime artifact is missing, use the bundled dry-run helpers and references instead of attempting a native GPU run.

Safe Command Construction

For most tasks, first build a dry command:

python sub-skills/image-projection/scripts/build_projection_command.py \
  --model-name shoes_64 \
  --input-image pics/shoes_test.png \
  --solver cnn_opt

For a custom pure optimization command with an explicit output path:

python sub-skills/image-projection/scripts/build_projection_command.py \
  --model-name shoes_64 \
  --input-image inputs/shoe.png \
  --solver opt \
  --output-image outputs/shoe_opt.png

The helper prints a command only. It never imports Theano, opens images, uses a GPU, starts training, or downloads files.

AlexNet Artifact Planning

The documented projection command uses AlexNet conv4 features. Plan the URL and local target without network side effects:

python sub-skills/image-projection/scripts/igan_alexnet_urls.py --layer conv4

The planner only reports URLs and expected targets. It does not download AlexNet, DCGAN models, or predictor parameters.

Native Runtime Command Pattern

When all artifacts and the legacy runtime are available, the command pattern is:

THEANO_FLAGS='device=gpu0,floatX=float32,nvcc.fastmath=True' \
python iGAN_predict.py \
  --model_name shoes_64 \
  --input_image pics/shoes_test.png \
  --solver cnn_opt

If --output_image is omitted and the input filename contains .png, the script writes beside the input as <input_stem>_<solver>.png. For example, pics/shoes_test.png with cnn_opt becomes pics/shoes_test_cnn_opt.png.

Solver Selection

• cnn: fastest feed-forward projection through the trained predictor network; requires usable predictor params.
• opt: L-BFGS-B latent optimization against pixel and feature reconstruction losses; slower and still compiled through the current predictor-aware setup.
• cnn_opt: default hybrid; starts from the predictor and refines with L-BFGS-B; best documented quality.

Prefer cnn_opt unless the user explicitly requests speed (cnn) or a custom optimization baseline (opt). Reject unknown solver strings before runtime; the source script does not validate them robustly.

Output Contract

A successful run prints parsed arguments, reads one image, compiles Theano functions, loads the model and AlexNet feature network, reconstructs one image, and saves the result to --output_image or the default solver-suffixed PNG path.

The output image is a reconstructed RGB image resized back to the original input dimensions. The script does not save the optimized latent vector by default.

Validation Checklist

Before telling a user a projection is ready to run, verify or explicitly mark unknown:

• iGAN_predict.py is available in the working runtime checkout.
• models/<model_name>.<model_type> or the supplied --model_file exists.
• models/caffe_reference_conv4.pkl exists for the default projection path.
• lib/ilsvrc_2012_mean.npy exists with the local iGAN library files.
• The input image path exists and is a PIL-readable image.
• The environment has Theano/Lasagne/SciPy/PIL and a backend acceptable to the user.
• The model contains predictor params if using cnn or cnn_opt, and preferably even for opt because the original script compiles the predictor unconditionally.

Troubleshooting Entry Points

If native execution fails, classify the symptom first:

• Import errors or syntax/runtime incompatibilities: see references/troubleshooting.md.
• Missing files or wrong targets: see references/troubleshooting.md.
• Predictor-param or solver problems: see references/troubleshooting.md.
• Theano/CUDA/cuDNN compile failures: see references/troubleshooting.md.

Boundary Reminders

Keep this sub-skill focused on image projection only. It may refer to DCGAN model files and predictor parameters as required inputs, but it does not own model-zoo downloads, predictor training, or model packing. It may mention AlexNet conv4 because the projection loss depends on it, but network downloads must remain a manual user action planned by the bundled URL helper.