Fitted SIREN corpora of MNIST: shared and independent initializations

# Fitted SIREN corpora of MNIST: shared and independent initializations Data deposit for "Generic identifiability of two-layer sine networks and the role of parametersymmetry in learning from INR weights" (M. D. Güven). Code, registrations and results:https://github.com/ITheClixs/project-siren-gap ## Contents | path | what ||---|---|| `P-shared-det/` | 70,000 SIRENs, one per MNIST image, all fitted from one shared initialization || `P-random/` | 70,000 SIRENs of the same images, each fitted from its own initialization || `shared_theta0_canonical.safetensors` | the shared initialization of `P-shared-det`, in canonical form | Every comparison of readers in Table 8 of the paper is made on these two corpora. ## Networks Two hidden layers of width 32, input dimension 2, output dimension 1, sine activations. Stored incanonical form: SIREN's frequency factor omega_0 = 30 is absorbed into the weights and biases of bothhidden layers, so a network computes h1 = sin(W1 x + b1), h2 = sin(W2 h1 + b2), f(x) = W3 h2 + b3 on coordinates x in [-1, 1]^2 (pixel centres of the 28 x 28 grid), with pixel values scaled to[-1, 1]. ## Files Each corpus holds `shard_<start>.safetensors` files, each a batch of networks with tensors`hidden.0.W` [B, 32, 2], `hidden.0.b` [B, 32], `hidden.1.W` [B, 32, 32], `hidden.1.b` [B, 32],`w_out` [B, 1, 32], `b_out` [B, 1], and a row-aligned `shard_<start>.parquet` with one row per network:`image_id`, `label`, `split` (train/val/test: 55,000/5,000/10,000), `protocol`, `init_seed`,`fit_seed`, `steps`, `lr`, `final_psnr`, `final_loss`, `wallclock_s`, `code_version`.`metadata.parquet` concatenates the rows; `config.json` records the fitting configuration. ## Fitting Adam, learning rate 1e-3, constant, 300 full-batch steps on the squared error, on the pixel grid.Initialization: SIREN's, with first-layer weights uniform on +-1/2, hidden weights uniform on+-sqrt(6/32)/30 and every bias uniform on +-1/sqrt(fan-in), before omega_0 is absorbed.`P-shared-det` uses initialization seed 0 for every network; `P-random` uses seed 1,000,000 +image_id for training and validation images and 3,000,000 + image_id for test images(`scripts/03_generate_inrbench.py` in the repository regenerates both). ## Loading ```pythonfrom safetensors.torch import load_fileimport pandas as pdshard = load_file("P-random/shard_000000.safetensors")meta = pd.read_parquet("P-random/shard_000000.parquet")``` `src/sirengap/data/schema.py::load_corpus` in the repository loads a whole corpus. ## License Creative Commons Attribution-ShareAlike 4.0 International (CC BY-SA 4.0). The networks are fitted toMNIST images, which are commonly distributed under CC BY-SA 3.0.