y4m 格式介绍：https://wiki.multimedia.cx/index.php/YUV4MPEG2
y4m 与 yuv（yuv420 8bit planar） 互转命令：
y4mtoyuv: ffmpeg -i xx.y4m -vsync 0 xx.yuv -y
yuvtoy4m: ffmpeg -s 1920x1080 -i xx.yuv -vsync 0 xx.y4m -y
y4m 与 png 互转命令：
y4mtobmp: ffmpeg -i xx.y4m -vsync 0 xx%3d.bmp -y
bmptoy4m: ffmpeg -i xx%3d.bmp -pix_fmt yuv420p -vsync 0 xx.y4m -y
y4m 每25帧抽样命令：
ffmpeg -i xxx.y4m -vf select='not(mod(n,25))' -vsync 0 -y xxx_sub25.y4m

初赛训练数据下载链接

round1_train_input:
http://tianchi-media.oss-cn-beijing.aliyuncs.com/231711_youku/round1/train/input/youku_00000_00049_l.zip
http://tianchi-media.oss-cn-beijing.aliyuncs.com/231711_youku/round1/train/input/youku_00050_00099_l.zip
http://tianchi-media.oss-cn-beijing.aliyuncs.com/231711_youku/round1/train/input/youku_00100_00149_l.zip

round1_train_label:
http://tianchi-media.oss-cn-beijing.aliyuncs.com/231711_youku/round1/train/label/youku_00000_00049_h_GT.zip
http://tianchi-media.oss-cn-beijing.aliyuncs.com/231711_youku/round1/train/label/youku_00050_00099_h_GT.zip
http://tianchi-media.oss-cn-beijing.aliyuncs.com/231711_youku/round1/train/label/youku_00100_00149_h_GT.zip

初赛验证数据下载链接

round1_val_input:
http://tianchi-media.oss-cn-beijing.aliyuncs.com/231711_youku/round1/train/input/youku_00150_00199_l.zip

round1_val_label:
http://tianchi-media.oss-cn-beijing.aliyuncs.com/231711_youku/round1/train/label/youku_00150_00199_h_GT.zip

####################
2019/6/3
简单转化为bmp格式，上传到百度网盘
链接：https://pan.baidu.com/s/1y1VbT5GsKIS8CJEO4CWnWQ
提取码：emw0

####################
2019/6/3 因为图片过大 模型改成半精度进行训练
uploaded latest model trained with automated mixed precision
install NVIDIA APEX dependency first (https://github.com/nvidia/apex)
$ git clone https://github.com/NVIDIA/apex
$ cd apex
$ pip install -v --no-cache-dir --global-option="--cpp_ext" --global-option="--cuda_ext" .

####################
TODO 已提issue: 1.用SSIM loss（已上传 但没merge进模型）替换原有的criterion_content ref:https://github.com/Po-Hsun-Su/pytorch-ssim
2.简化模型 或者使用小patch（切片等）降低显存占用 使普通GPU也能训练高分辨率图片 trade off between acc and FLOPs

####################
2019/6/5
upload pretrained model and code for prediction
链接：https://pan.baidu.com/s/1p-RJg8pgjaf88PNL4P8t7g 提取码：o1ew