lack F.interpolate in bisenetv1.py

[qixi666666]

trafficstars

in bisenetv1.py class ContextPath lack F.interpolate about avg in line 141, so the result is wrong

[CoinCheung]

What result is wrong? Would you be more specific?

[qixi666666]

avg should be avg_up = F.interpolate(avg, (H32, W32), mode='nearest'), then feat32_sum = feat32_arm + avg_up, if lack F.interpolate(avg, (H32, W32), mode='nearest'), avg only 1*1, feat32_sum = feat32_arm + avg will not report wrong, but the result is not true.

[CoinCheung]

Have you compared the values?

[qixi666666]

你可以直接看看原始的old / model.py ContextPath 中这部分的定义，确实少一块，上升的过程

[CoinCheung]

This should bring no difference, since input shape is (1,1).

[qixi666666]

not 1*1 , should be H32, W32 , both feat32_arm and avg should be H32, W32, then +

[CoinCheung]

The feature will be broadcast to H32 and W32 before adding.