How to get the middle features of the effecientnet for FCN/FPN?

[lartpang]

No description provided.

[lartpang]

I use the method to do it:

class EfficientNet(nn.Module):
    ...
    def extract_features_midconv(self, inputs):
        """ Returns output of the middle convolution layers """
        out_feats = []
        
        # Stem
        x = self._swish(self._bn0(self._conv_stem(inputs)))
        prev_x_size = x.size()[-1]
        out_feats.append(x)
        
        # Blocks
        for idx, block in enumerate(self._blocks):
            drop_connect_rate = self._global_params.drop_connect_rate
            if drop_connect_rate:
                drop_connect_rate *= float(idx) / len(self._blocks)
            x = block(x, drop_connect_rate=drop_connect_rate)
            if x.size()[-1] != prev_x_size:
                prev_x_size = x.size()[-1]
                out_feats.append(x)
            else:
                out_feats[-1] = x
        
        # Head
        x = self._swish(self._bn1(self._conv_head(x)))
        if x.size()[-1] != prev_x_size:
            out_feats.append(x)
        else:
            out_feats[-1] = x
        
        return out_feats

I think this treatment is a little inelegant, do you have a better way?

[fmahoudeau]

What about this solution ?
It returns features at stride 4, 8, 16 and 32.

def extract_features_midconv(self, inputs):
    """ Returns output of the middle convolution layers """
    out = []

    # Stem
    x = self._swish(self._bn0(self._conv_stem(inputs)))

    # Blocks
    for idx, block in enumerate(self._blocks):
        drop_connect_rate = self._global_params.drop_connect_rate
        if drop_connect_rate:
            drop_connect_rate *= float(idx) / len(self._blocks)
        y = block(x, drop_connect_rate=drop_connect_rate)
        if y.size()[-1] != x.size()[-1]:
            out.append(x)
        x = y
    
    # Head
    x = self._swish(self._bn1(self._conv_head(x)))
    out.append(x)

    return out[1:]

[lartpang]

@fmahoudeau Oh..yeah! This is good, and better than my code! 👍

[ahundt]

#63 also does this