lingbot-sn80-workspace / 1gpu /eval /dump_train.py
salpu's picture
1-GPU box: eval harness wired+corrected, LoRA export merge fix, utilisation profiling, findings
3922a1b verified
Raw History Blame Contribute Delete
314 Bytes
import sys, numpy as np
from torchcodec.decoders import VideoDecoder
import imageio
d = VideoDecoder(sys.argv[1])
f = d[0].permute(1,2,0).numpy().astype("uint8")
print("train frame shape:", f.shape, "top/bottom mean:", f[:f.shape[0]//2].mean(), f[f.shape[0]//2:].mean())
imageio.imwrite("/out/train_frame.png", f)