ethanrom/helm

1# YOLOv5 🚀 by Ultralytics, GPL-3.0 license2"""3Experimental modules4"""5import math6 7import numpy as np8import torch9import torch.nn as nn10 11from utils.downloads import attempt_download12 13 14class Sum(nn.Module):15    # Weighted sum of 2 or more layers https://arxiv.org/abs/1911.0907016    def __init__(self, n, weight=False):  # n: number of inputs17        super().__init__()18        self.weight = weight  # apply weights boolean19        self.iter = range(n - 1)  # iter object20        if weight:21            self.w = nn.Parameter(-torch.arange(1.0, n) / 2, requires_grad=True)  # layer weights22 23    def forward(self, x):24        y = x[0]  # no weight25        if self.weight:26            w = torch.sigmoid(self.w) * 227            for i in self.iter:28                y = y + x[i + 1] * w[i]29        else:30            for i in self.iter:31                y = y + x[i + 1]32        return y33 34 35class MixConv2d(nn.Module):36    # Mixed Depth-wise Conv https://arxiv.org/abs/1907.0959537    def __init__(self, c1, c2, k=(1, 3), s=1, equal_ch=True):  # ch_in, ch_out, kernel, stride, ch_strategy38        super().__init__()39        n = len(k)  # number of convolutions40        if equal_ch:  # equal c_ per group41            i = torch.linspace(0, n - 1E-6, c2).floor()  # c2 indices42            c_ = [(i == g).sum() for g in range(n)]  # intermediate channels43        else:  # equal weight.numel() per group44            b = [c2] + [0] * n45            a = np.eye(n + 1, n, k=-1)46            a -= np.roll(a, 1, axis=1)47            a *= np.array(k) ** 248            a[0] = 149            c_ = np.linalg.lstsq(a, b, rcond=None)[0].round()  # solve for equal weight indices, ax = b50 51        self.m = nn.ModuleList([52            nn.Conv2d(c1, int(c_), k, s, k // 2, groups=math.gcd(c1, int(c_)), bias=False) for k, c_ in zip(k, c_)])53        self.bn = nn.BatchNorm2d(c2)54        self.act = nn.SiLU()55 56    def forward(self, x):57        return self.act(self.bn(torch.cat([m(x) for m in self.m], 1)))58 59 60class Ensemble(nn.ModuleList):61    # Ensemble of models62    def __init__(self):63        super().__init__()64 65    def forward(self, x, augment=False, profile=False, visualize=False):66        y = [module(x, augment, profile, visualize)[0] for module in self]67        # y = torch.stack(y).max(0)[0]  # max ensemble68        # y = torch.stack(y).mean(0)  # mean ensemble69        y = torch.cat(y, 1)  # nms ensemble70        return y, None  # inference, train output71 72 73def attempt_load(weights, device=None, inplace=True, fuse=True):74    # Loads an ensemble of models weights=[a,b,c] or a single model weights=[a] or weights=a75    from models.yolo import Detect, Model76 77    model = Ensemble()78    for w in weights if isinstance(weights, list) else [weights]:79        ckpt = torch.load(attempt_download(w), map_location='cpu')  # load80        ckpt = (ckpt.get('ema') or ckpt['model']).to(device).float()  # FP32 model81 82        # Model compatibility updates83        if not hasattr(ckpt, 'stride'):84            ckpt.stride = torch.tensor([32.])85        if hasattr(ckpt, 'names') and isinstance(ckpt.names, (list, tuple)):86            ckpt.names = dict(enumerate(ckpt.names))  # convert to dict87 88        model.append(ckpt.fuse().eval() if fuse and hasattr(ckpt, 'fuse') else ckpt.eval())  # model in eval mode89 90    # Module compatibility updates91    for m in model.modules():92        t = type(m)93        if t in (nn.Hardswish, nn.LeakyReLU, nn.ReLU, nn.ReLU6, nn.SiLU, Detect, Model):94            m.inplace = inplace  # torch 1.7.0 compatibility95            if t is Detect and not isinstance(m.anchor_grid, list):96                delattr(m, 'anchor_grid')97                setattr(m, 'anchor_grid', [torch.zeros(1)] * m.nl)98        elif t is nn.Upsample and not hasattr(m, 'recompute_scale_factor'):99            m.recompute_scale_factor = None  # torch 1.11.0 compatibility100 101    # Return model102    if len(model) == 1:103        return model[-1]104 105    # Return detection ensemble106    print(f'Ensemble created with {weights}\n')107    for k in 'names', 'nc', 'yaml':108        setattr(model, k, getattr(model[0], k))109    model.stride = model[torch.argmax(torch.tensor([m.stride.max() for m in model])).int()].stride  # max stride110    assert all(model[0].nc == m.nc for m in model), f'Models have different class counts: {[m.nc for m in model]}'111    return model112