Checkpoints¶
Download: checkpoint_best.pth, 293.4 MB. Put it in
this directory.
ExplorationConfig.MODEL_CHECKPOINT, the benchmark and intelx.cli.evaluate
all read weights/checkpoint_best.pth by default. To keep checkpoints
somewhere else, redirect the root rather than editing a source file.
What It Is¶
| Architecture | U-Net with a ResNet-34 encoder, intelx.models.unet |
| Parameters | 24,423,009, plus 18,988 BatchNorm buffer values |
| Selection | epoch 31, validation R² 0.357 |
| Format | the training checkpoint as intelx.cli.train writes it |
It is the full training checkpoint, so it carries model_state_dict together
with optimizer_state_dict, epoch, val_r2 and val_loss. The AdamW
moments are two thirds of its size. The model keys carry the _orig_mod.
prefix that torch.compile adds, which
intelx.common.create_model_for_inference strips on load. It contains tensors
and plain values only, so torch.load(path, weights_only=True) reads it.
Tools¶
scripts/inspect_checkpoint.py lists everything a checkpoint carries besides
tensors, without executing its pickle.
scripts/strip_checkpoint.py writes a 49 MB weights-only FP16 copy, should you
want one. It keeps the thirteen BatchNorm running_var buffers that exceed the
FP16 range at FP32. Point MODEL_CHECKPOINT at the copy to use it.
Both need numpy and the standard library only.