求教怎么改能降低显存占用

[nistarlwc]

我已经修改成了用sam2_hiera_small结构来训练。
但是我现在想跑1024尺寸的图像，12GB的显卡只能跑2个batch size。
还能怎么修改能降低些显存占用呢？

[xiongxyowo]

Hi, the following strategies can be adopted:

• Adopt a stronger encoder channel compression strategy. The current RFB compresses the channel to 64, you can try lower values.
• Replace RFB with a simple 1×1 convolution.
• Remove the last decoder block (Up3). In other words, the final result is obtained by 8 times upsampling instead of 4 times.

Note that these strategies may lead to a decrease in segmentation accuracy.

---

同学你好，一般可以采用如下几个策略：

• 采用更激进的编码器通道压缩策略。目前的RFB将通道压缩至了64，你可以尝试更低的值。
• 将RFB替换为简单的1×1卷积。
• 移除最后一个解码器块(Up3)。即，最终结果通过8倍上采样得到而非4倍。

注意这些策略可能导致分割精度的下降。

[Lei0829]

Hi, the following strategies can be adopted:

• Adopt a stronger encoder channel compression strategy. The current RFB compresses the channel to 64, you can try lower values.
• Replace RFB with a simple 1×1 convolution.
• Remove the last decoder block (Up3). In other words, the final result is obtained by 8 times upsampling instead of 4 times.

Note that these strategies may lead to a decrease in segmentation accuracy.

同学你好，一般可以采用如下几个策略：

• 采用更激进的编码器通道压缩策略。目前的RFB将通道压缩至了64，你可以尝试更低的值。
• 将RFB替换为简单的1×1卷积。
• 移除最后一个解码器块(Up3)。即，最终结果通过8倍上采样得到而非4倍。

注意这些策略可能导致分割精度的下降。

如果我想要效果对比图，可是训练出来的像素很低，也是采用上述方法么

[Lei0829]

Hi, the following strategies can be adopted:

• Adopt a stronger encoder channel compression strategy. The current RFB compresses the channel to 64, you can try lower values.
• Replace RFB with a simple 1×1 convolution.
• Remove the last decoder block (Up3). In other words, the final result is obtained by 8 times upsampling instead of 4 times.

Note that these strategies may lead to a decrease in segmentation accuracy.

同学你好，一般可以采用如下几个策略：

• 采用更激进的编码器通道压缩策略。目前的RFB将通道压缩至了64，你可以尝试更低的值。
• 将RFB替换为简单的1×1卷积。
• 移除最后一个解码器块(Up3)。即，最终结果通过8倍上采样得到而非4倍。

注意这些策略可能导致分割精度的下降。

而且我发现训练输出的结果图并不是352352而是45712867

[xiongxyowo]

Hi, we created a wechat discussion group, please refer to wechat.jpg for more details.