| library_name: pytorch | |
| license: mit | |
| tags: | |
| - piscis | |
| - microscopy | |
| - spot-detection | |
| - fluorescence-microscopy | |
| # Piscis ps_20240419_112256 | |
| PyTorch conversion of the legacy Piscis JAX checkpoint | |
| [`wniu/ps_20240419_112256`](https://huggingface.co/wniu/ps_20240419_112256), | |
| for Flax-free inference in NimbusImage's Piscis Predict and Train workers. | |
| ## Conversion validation | |
| The Piscis 1.0.0 JAX-to-PyTorch converter was first validated against the | |
| published `20230905` JAX and PyTorch checkpoint pair. All 583 tensors matched | |
| exactly in key, shape, dtype, and value; the maximum absolute difference was | |
| `0.0`. The same validated converter produced this checkpoint. | |
| - Source checkpoint SHA-256: `28fea47454c108dac5e2899f5c4f84fd489792f28fbd5d6b3e002236516952e2` | |
| - Converted checkpoint SHA-256: `f281a239d1d5075ee2342c4772ba7167c441f878e5345c632436caf3b43d320a` | |
| - Converted checkpoint size: `30,083,712` bytes | |
| - State dictionary: 583 tensors | |
| - Metadata: `adjustment=standardize`, `input_size=(256, 256)`, `dilation_iterations=1`, `channels=1` | |