def save_features_to_h5(
self,
layer_i_features_fname: str,
indices: tuple,
subset: tuple,
label: tuple,
hidden_states: torch.tensor,
transformer_model_config,
):
"""
A function that saves the indices of the processed examples,
the subset labels (tr/va/te), and the layer features of the transformer model
:param layer_i_features_fname: The name of .h5 file to store the layer features
:param indices: A list of the batch indices
:param subset: A list of the batch subset labels
:param label: A list of the batch target labels
:param hidden_states: A tensor of the batch hidden states
:return:
"""
batch = {
"index": indices.numpy().astype(int),
"subset": np.asarray(list(subset)).astype("S20"),
"label": np.asarray(list(label)).astype("S20"),
"hidden_states": hidden_states.astype(float),
}
if not os.path.isfile(layer_i_features_fname):
with h5py.File(layer_i_features_fname, "w") as h5f:
for field, feats in batch.items():
h5f.create_dataset(
field,
data=feats,
maxshape=(None, transformer_model_config.hidden_size)
if field == "hidden_states" else (None, ),
chunks=True,
)
else:
with h5py.File(layer_i_features_fname, "a") as h5f:
for field, feats in batch.items():
feats_shape = feats.shape[0]
resize_shape = h5f[field].shape[0] + feats_shape
h5f[field].resize(
(resize_shape, transformer_model_config.hidden_size
)) if field == "hidden_states" else h5f[field].resize(
(resize_shape, ))
h5f[field][-feats_shape:] = feats
def run_opencv(bayer: torch.tensor):
bayer = bayer.squeeze().cpu().to(torch.float32).numpy() * 255.0
bayer = bayer.astype(np.uint8)
rgb = cv2.cvtColor(bayer, cv2.COLOR_BAYER_BG2RGB)
return torch.tensor(rgb).permute(2, 0, 1).to(torch.float32) / 255.0
