Skip to content

FORMAT.md

Latest commit

 

History

History
66 lines (47 loc) · 1.89 KB

FORMAT.md

File metadata and controls

66 lines (47 loc) · 1.89 KB
Raw

PyTorch .pth format notes

These notes record the subset reverse-engineered for PTHrs 0.2.0.

Container

Modern PyTorch checkpoints begin with a ZIP local-file header (PK\x03\x04). All relevant paths normally share one generated prefix:

<prefix>/data.pkl
<prefix>/byteorder
<prefix>/version
<prefix>/data/0
<prefix>/data/1
...

data.pkl is small relative to the model. Each data/<id> entry is a raw typed storage byte array. PyTorch normally stores these entries without ZIP compression and adds alignment padding to the local-header extra field.

Pickle connection

Tensor records use torch._utils._rebuild_tensor_v2 with arguments equivalent to:

(storage, storage_offset, shape, stride, requires_grad, backward_hooks)

The storage value is supplied by the pickle BINPERSID opcode. Its persistent ID is normally:

("storage", torch.<DType>Storage, "<id>", "<device>", element_count)

PTHrs maps <id> to <prefix>/data/<id>. The device string is retained but does not affect reading.

Tensor views

The storage offset and stride are measured in elements, not bytes. For a tensor coordinate i, the storage element is:

storage_offset + sum(i[axis] * stride[axis])

read_tensor returns contiguous row-major bytes. If the saved view is already contiguous, PTHrs performs one bounded range read. Otherwise it reads the smallest enclosing storage span and gathers elements in logical order.

Observed voice-conversion checkpoint fields

Common inference checkpoints contain:

  • weight: state dictionary from parameter names to tensors
  • config: positional network configuration
  • info: training checkpoint label or epoch
  • sr: sampling-rate label such as 40k
  • f0: whether pitch guidance is enabled
  • version: application architecture version such as v1 or v2

These remain generic Value data so callers can handle future variations.