diffcve-2607 / scripts /stream_adapter.py
multimodalart's picture
multimodalart HF Staff
Upload folder using huggingface_hub
d2c5d46 verified
Raw History Blame Contribute Delete
3.01 kB
from dataclasses import dataclass
from typing import Tuple
import numpy as np
from codec.processor import Frame
def normalize_frame(frame):
if frame.rgb is None:
raise ValueError('model adapter requires output_format="rgb24"')
# Preserve the original operation order and float32 arithmetic exactly.
images = frame.rgb.astype(np.float32) / 255.0
images = images[None].transpose(0, 3, 1, 2)
images = (images - 0.5) / 0.5
residues = (frame.res.astype(np.float32) / 255.0)[None, None]
motions = frame.mv.astype(np.float32) / 4.0 / 32.0
motions = motions[None].transpose(0, 3, 1, 2)
return tuple(np.ascontiguousarray(np.clip(value, -1, 1)) for value in (images, residues, motions))
@dataclass
class Segment:
frames: Tuple[Frame, ...] # real frames only, for saving LQ and enhanced PNGs
images: np.ndarray # float32 [T,3,H,W], [-1,1]
residues: np.ndarray # float32 [T,1,H,W], [-1,1]
motions: np.ndarray # float32 [T,2,H,W], [-1,1]
qp: int
@property
def valid_count(self):
return len(self.frames)
def to_torch(self, device='cpu'):
import torch # Optional: use the model server's existing torch environment.
values = [torch.from_numpy(value) for value in (self.images, self.residues, self.motions)]
values[0] = values[0].contiguous(memory_format=torch.channels_last)
return tuple(value.to(device) for value in values)
def iter_segments(frames, segment_size=5, max_frames=0):
if type(segment_size) is not int or segment_size < 1:
raise ValueError('segment_size must be a positive integer')
if type(max_frames) is not int or max_frames < 0:
raise ValueError('max_frames must be a nonnegative integer')
iterator = iter(frames)
previous = None
consumed = 0
while True:
chunk = []
for _ in range(segment_size):
if max_frames and consumed >= max_frames:
break
try:
frame = next(iterator)
except StopIteration:
break
if previous is not None and frame.index != previous + 1:
raise ValueError('noncontiguous model frame indices')
previous = frame.index
chunk.append(frame)
consumed += 1
if not chunk:
if max_frames:
for _ in iterator:
pass
return
qp = chunk[0].qp
if any(frame.qp != qp for frame in chunk):
raise ValueError('mixed QP within a model segment')
padded = chunk + [chunk[-1]] * (segment_size - len(chunk))
normalised = [normalize_frame(frame) for frame in padded]
values = [np.concatenate([value[i] for value in normalised], axis=0) for i in range(3)]
yield Segment(tuple(chunk), *values, qp)
if len(chunk) < segment_size:
if max_frames:
for _ in iterator:
pass
return