pooler.py

import torch 
import torch.nn as nn 
import math
import random


class Pooler(nn.Module):
    def __init__(self, dim_in, dim_out, pool_out_size):
        super().__init__()
        assert isinstance(pool_out_size, str)
        self.pool_out_size = pool_out_size.split(",")
        print("pool_out_size: {}".format(self.pool_out_size))

        self.mlp = nn.Sequential(
            nn.Linear(dim_in, dim_out),
            nn.GELU(),
            nn.Linear(dim_out, dim_out)
        )
                
    def forward(self, x):
        """
        Args:
            x (torch.Tensor): image features
                shape (b, v, D)
        Returns:
            shape (b, n, D) where n is self.num_latents
        """
        b, v, d = x.shape
        s = int(math.sqrt(v -1))
        x = x[:, 1:, :]           # remove cls_token
        x_in = x.reshape(b, s, s, d)

        pool_out_size = random.choice(self.pool_out_size)

        if '+' in pool_out_size:        # "16+32" means ensemble the pool size of 16 and 32
            pool_out_size_list = [int(p) for p in pool_out_size.split('+')]
        else:
            pool_out_size_list = [int(pool_out_size)]
        pool_out_size_list.sort(reverse=True)

        x_out = []
        for pool_out_size in pool_out_size_list:
            assert s % pool_out_size == 0 
            x = x_in.reshape(b, pool_out_size, s//pool_out_size, pool_out_size, s//pool_out_size, d)
            x = x.permute([0, 1, 3, 5, 2, 4]).reshape(b, pool_out_size * pool_out_size, d, -1).mean(-1)
            x = self.mlp(x)                 # [b, h*w, d]
            x_out.append(x)
        x_out = torch.cat(x_out, dim=-2)
        return x_out