aniunflow/decoder.py

from __future__ import annotations
import torch
import torch.nn as nn
import torch.nn.functional as F
from .utils import warp

class MSRecurrentDecoder(nn.Module):
    def __init__(self, chs=[64,128,192], iters_per_level=4, residual_scale=0.1):
        super().__init__()
        self.iters = iters_per_level
        self.residual_scale = residual_scale  # Scale down residual updates to prevent flow collapse
        
        self.refine1 = nn.Sequential(
            nn.Conv2d(chs[1]+2+chs[1], chs[1], 3, 1, 1), 
            nn.GELU(), 
            nn.Conv2d(chs[1], 2, 3, 1, 1)
        )
        self.refine0 = nn.Sequential(
            nn.Conv2d(chs[0]+2+chs[0], chs[0], 3, 1, 1), 
            nn.GELU(), 
            nn.Conv2d(chs[0], 2, 3, 1, 1)
        )
        
        # Initialize output layers with small weights to prevent large residuals at start
        self._init_weights()
    
    def _init_weights(self):
        """Initialize output conv layers with small weights for stable training."""
        for refine in [self.refine1, self.refine0]:
            # Last conv layer should output near-zero initially
            last_conv = refine[-1]
            nn.init.zeros_(last_conv.weight)
            nn.init.zeros_(last_conv.bias)
    
    def forward(self, coarse_flows, feats_levels, latent_levels, attn_bias=None):
        lvl1, lvl2, lvl3 = feats_levels
        outs = []
        for k, f8 in enumerate(coarse_flows):
            cur = f8
            # refine at 1/8
            for _ in range(self.iters):
                x = torch.cat([lvl2[k], cur, warp(lvl2[k+1], cur)], dim=1)
                residual = self.refine1(x)
                cur = cur + self.residual_scale * residual  # Scaled residual update
            # upsample to 1/4 and refine
            cur2 = F.interpolate(cur, scale_factor=2.0, mode='bilinear', align_corners=True) * 2.0  # Scale flow with resolution
            for _ in range(self.iters):
                x = torch.cat([lvl1[k], cur2, warp(lvl1[k+1], cur2)], dim=1)
                residual = self.refine0(x)
                cur2 = cur2 + self.residual_scale * residual  # Scaled residual update
            outs.append(cur2)
        return outs