Update FAQ

Error "slow_conv2d_cpu" not implemented for 'Half'
ScrollRack · May 5, 2022 · 0786521 · 0786521
1 parent fc8c8de
commit 0786521
Showing 1 changed file with 5 additions and 2 deletions.
diff --git a/docs/FAQ.md b/docs/FAQ.md
@@ -1,7 +1,10 @@
 # FAQ
 
-1. **How to select models?**<br>
+1. **Q: How to select models?**<br>
 A: Please refer to [docs/model_zoo.md](docs/model_zoo.md)
 
-1. **Can `face_enhance` be used for anime images/animation videos?**<br>
+1. **Q: Can `face_enhance` be used for anime images/animation videos?**<br>
 A: No, it can only be used for real faces. It is recommended not to use this option for anime images/animation videos to save GPU memory.
+
+1. **Q: Error "slow_conv2d_cpu" not implemented for 'Half'**<br>
+A: In order to save GPU memory consumption and speed up inference, Real-ESRGAN uses half precision (fp16) during inference by default. However, some operators for half inference are not implemented in CPU mode. You need to add **`--fp32` option** for the commands. For example, `python inference_realesrgan.py -n RealESRGAN_x4plus.pth -i inputs --fp32`.