Compositional_Neural_Painter
Compositional_Neural_Painter copied to clipboard
[MM 2023] Stroke-based Neural Painting and Stylization with Dynamically Predicted Painting Region
作者您好!我尝试着按照顺序重新训练各个网络,在训练painter network时出现了一些训练时长上的问题想要请教一下。我下载了您推荐的celeba数据集(20万张图片,总大小1.4G),按7比3的比例划分了训练集和测试集。使用了单张3090训练两天,发现只训练到了8万轮左右,请问这个速度正常吗?
作者你好,我想使用 train_renderer-FCN.py 先训练一个可微神经渲染器,但是在源代码 43 行中 train_batch=torch.rand(batch_size, stroke_num).cuda(),train_batch 的 shape 是 batch_size*stroke_num,而 16 行 stroke_num=10,stroke_num 设置为 10,但是 draw_oil 函数需要的笔画参数只有 5 个 x0, y0, w, h, theta,这里 stroke_num 为什么还要设置为 10,stroke_num = 5...
it possible to change the shape of the drawing boundary (from a fixed rectangle) to draw an area of arbitrary shape?
As part of training code, a particular file, `oil_brush.pkl` is mentioned in following places: - [`compositor/DRL/ddpg.py`](https://github.com/sjtuplayer/Compositional_Neural_Painter/blob/f306cab09eb53e1ab99aa619dda558b051be1218/compositor/DRL/ddpg.py#L30) - [`compositor/train_renderer-FCN.py`](https://github.com/sjtuplayer/Compositional_Neural_Painter/blob/f306cab09eb53e1ab99aa619dda558b051be1218/compositor/train_renderer-FCN.py#L26) - [`painter/train.py`](https://github.com/sjtuplayer/Compositional_Neural_Painter/blob/f306cab09eb53e1ab99aa619dda558b051be1218/painter/train.py#L52) From the above references, it appears to be expected at...