hzwer / hzwer/Practical-RIFE

RIFE4.25微调和一些代码疑问

Open
#157 1 comment 0 reactions 0 assignees View on GitHub

Nobody has claimed this yet.

Dominant language
Python
Stars
1k
Forks
131
PR merge metrics
No merged PRs in 30d

Description

作者您好,我在预训练的Rife v4.25模型基础上,用 vimeo-90k 随机采样 25% 约 13000 个样本作为训练数据,学习率[1e-7,1e-5],微调了3~5 个 epoch。
然后用inference_video插帧视频时边缘会有水波状扭曲,以及运动物体重影,看起来模型的细节处理和运动光流估计能力都被破坏了。
train过程中的loss一直震荡不收敛,validate的lpips逐渐下降
是训练轮数不够还是数据集不足,或者是训练策略的问题?

以及关于RIFE4.25训练代码的一些问题:

  • dataset.py中64到67行
    if img0.shape[0] < 448:
    img0 = np.concatenate((img0, img0[:, :, ::-1].copy()), 0)
    img1 = np.concatenate((img1, img1[:, :, ::-1].copy()), 0)
    gt = np.concatenate((gt, gt[:, :, ::-1].copy()), 0)
    这里为什么要交换颜色通道?似乎是笔误想翻转高度?而且这样拼接导致画面不连续了

  • model.py中Model的update函数返回值
    return merged[-1], {
    'merged_tea': teacher_res[0][0],
    'mask': mask,
    'mask_tea': mask,
    'flow': flow[3][:, :2],
    为什么返回的是倒数第二层的flow?最后一层的不是更加直观吗
    既然mask和mask_tea都是同一个mask,还有必要在train里写入tensorboard对比吗?还是说只是占个位

  • flownet.py的FlownetCas中
    for i in range(5):
    if flow is not None:
    flow_d, mask, feat = stu[i](torch.cat((warped_img0, warped_img1, warped_f0, warped_f1, timestep, mask, feat), 1), flow, scale=scale[i])
    flow = flow + flow_d
    else:
    flow, mask, feat = stu[i](torch.cat((img0, img1, f0, f1, timestep), 1), None, scale=scale[i])
    mask_list.append(mask)
    flow_list.append(flow)
    feat_list.append(feat[:, :1])
    后续只用了feat第一层的置信度做融合,后面的层不是浪费了吗

Contributor guide

No contributing guide indexed for this repository

First steps

  1. Read the whole issue, then the project's contributing guide.
  2. Comment on the issue to say you are picking it up — it saves two people doing the same work.
  3. Fork the repository and make your change on a branch.
  4. Open a pull request that references the issue number.

Research direction

Read dataset.py lines 64-67, model.py Model.update, and flownet.py FlownetCas; reproduce the fine-tuning with inference_video and compare training loss with validation LPIPS. The issue is complete when each reported behavior is explained and any confirmed code or training problem has a scoped fix and verification.

Written by the indexing model from the issue text.

Assessment

Tech stack
python, pytorch
Domain
computer-vision, machine-learning
Issue type
Bug
Difficulty
5/5
Estimated time
Over a week
Activity status
Active
Clarity
Needs clarification
Newbie friendliness
35/100

Get new issues in your inbox

A short digest of beginner-friendly GitHub issues.