models.py

import torch.nn.functional as F

from utils.google_utils import *
from utils.parse_config import *
from utils.utils import *

ONNX_EXPORT = False


def create_modules(module_defs, img_size, arc):
    # Constructs module list of layer blocks from module configuration in module_defs

    hyperparams = module_defs.pop(0)
    output_filters = [int(hyperparams['channels'])]
    module_list = nn.ModuleList()
    routs = []  # list of layers which rout to deeper layers
    yolo_index = -1

    for i, mdef in enumerate(module_defs):
        modules = nn.Sequential()
        # if i == 0:
        #     modules.add_module('BatchNorm2d_0', nn.BatchNorm2d(output_filters[-1], momentum=0.1))

        if mdef['type'] == 'convolutional':
            bn = mdef['batch_normalize']
            filters = mdef['filters']
            size = mdef['size']
            stride = mdef['stride'] if 'stride' in mdef else (mdef['stride_y'], mdef['stride_x'])
            modules.add_module('Conv2d', nn.Conv2d(in_channels=output_filters[-1],
                                                   out_channels=filters,
                                                   kernel_size=size,
                                                   stride=stride,
                                                   padding=(size - 1) // 2 if mdef['pad'] else 0,
                                                   groups=mdef['groups'] if 'groups' in mdef else 1,
                                                   bias=not bn))
            if bn:
                modules.add_module('BatchNorm2d', nn.BatchNorm2d(filters, momentum=0.1))
            else:
                routs.append(i)  # detection output (goes into yolo layer)

            if mdef['activation'] == 'leaky':  # activation study https://github.com/ultralytics/yolov3/issues/441
                modules.add_module('activation', nn.LeakyReLU(0.1, inplace=True))
                # modules.add_module('activation', nn.PReLU(num_parameters=1, init=0.10))
            elif mdef['activation'] == 'swish':
                modules.add_module('activation', Swish())

        elif mdef['type'] == 'maxpool':
            size = mdef['size']
            stride = mdef['stride']
            maxpool = nn.MaxPool2d(kernel_size=size, stride=stride, padding=(size - 1) // 2)
            if size == 2 and stride == 1:  # yolov3-tiny
                modules.add_module('ZeroPad2d', nn.ZeroPad2d((0, 1, 0, 1)))
                modules.add_module('MaxPool2d', maxpool)
            else:
                modules = maxpool

        elif mdef['type'] == 'upsample':
            if ONNX_EXPORT:  # explicitly state size, avoid scale_factor
                g = (yolo_index + 1) * 2 / 32  # gain
                modules = nn.Upsample(size=tuple(int(x * g) for x in img_size))  # img_size = (320, 192)
            else:
                modules = nn.Upsample(scale_factor=mdef['stride'])

        elif mdef['type'] == 'route':  # nn.Sequential() placeholder for 'route' layer
            layers = mdef['layers']
            filters = sum([output_filters[i + 1 if i > 0 else i] for i in layers])
            routs.extend([l if l > 0 else l + i for l in layers])
            # if mdef[i+1]['type'] == 'reorg3d':
            #     modules = nn.Upsample(scale_factor=1/float(mdef[i+1]['stride']), mode='nearest')  # reorg3d

        elif mdef['type'] == 'shortcut':  # nn.Sequential() placeholder for 'shortcut' layer
            layers = mdef['from']
            filters = output_filters[-1]
            routs.extend([i + l if l < 0 else l for l in layers])
            modules = weightedFeatureFusion(layers=layers, weight='weights_type' in mdef)

        elif mdef['type'] == 'reorg3d':  # yolov3-spp-pan-scale
            # torch.Size([16, 128, 104, 104])
            # torch.Size([16, 64, 208, 208]) <-- # stride 2 interpolate dimensions 2 and 3 to cat with prior layer
            pass

        elif mdef['type'] == 'yolo':
            yolo_index += 1
            l = mdef['from'] if 'from' in mdef else []
            modules = YOLOLayer(anchors=mdef['anchors'][mdef['mask']],  # anchor list
                                nc=mdef['classes'],  # number of classes
                                img_size=img_size,  # (416, 416)
                                yolo_index=yolo_index,  # 0, 1, 2...
                                layers=l)  # output layers

            # Initialize preceding Conv2d() bias (https://arxiv.org/pdf/1708.02002.pdf section 3.3)
            try:
                bo = -4.5  #  obj bias
                bc = math.log(1 / (modules.nc - 0.99))  # cls bias: class probability is sigmoid(p) = 1/nc

                j = l[yolo_index] if 'from' in mdef else -1
                bias_ = module_list[j][0].bias  # shape(255,)
                bias = bias_[:modules.no * modules.na].view(modules.na, -1)  # shape(3,85)
                bias[:, 4] += bo - bias[:, 4].mean()  # obj
                bias[:, 5:] += bc - bias[:, 5:].mean()  # cls, view with utils.print_model_biases(model)
                module_list[j][0].bias = torch.nn.Parameter(bias_, requires_grad=bias_.requires_grad)
            except:
                print('WARNING: smart bias initialization failure.')

        else:
            print('Warning: Unrecognized Layer Type: ' + mdef['type'])

        # Register module list and number of output filters
        module_list.append(modules)
        output_filters.append(filters)

    return module_list, routs


class weightedFeatureFusion(nn.Module):  # weighted sum of 2 or more layers https://arxiv.org/abs/1911.09070
    def __init__(self, layers, weight=False):
        super(weightedFeatureFusion, self).__init__()
        self.layers = layers  # layer indices
        self.weight = weight  # apply weights boolean
        self.n = len(layers) + 1  # number of layers
        if weight:
            self.w = torch.nn.Parameter(torch.zeros(self.n))  # layer weights

    def forward(self, x, outputs):
        # Weights
        if self.weight:
            w = torch.sigmoid(self.w) * (2 / self.n)  # sigmoid weights (0-1)
            x = x * w[0]

        # Fusion
        nc = x.shape[1]  # input channels
        for i in range(self.n - 1):
            a = outputs[self.layers[i]] * w[i + 1] if self.weight else outputs[self.layers[i]]  # feature to add
            ac = a.shape[1]  # feature channels
            dc = nc - ac  # delta channels

            # Adjust channels
            if dc > 0:  # slice input
                x[:, :ac] = x[:, :ac] + a  # or a = nn.ZeroPad2d((0, 0, 0, 0, 0, dc))(a); x = x + a
            elif dc < 0:  # slice feature
                x = x + a[:, :nc]
            else:  # same shape
                x = x + a
        return x


class SwishImplementation(torch.autograd.Function):
    @staticmethod
    def forward(ctx, i):
        ctx.save_for_backward(i)
        return i * torch.sigmoid(i)

    @staticmethod
    def backward(ctx, grad_output):
        sigmoid_i = torch.sigmoid(ctx.saved_variables[0])
        return grad_output * (sigmoid_i * (1 + ctx.saved_variables[0] * (1 - sigmoid_i)))


class MemoryEfficientSwish(nn.Module):
    def forward(self, x):
        return SwishImplementation.apply(x)


class Swish(nn.Module):
    def forward(self, x):
        return x.mul_(torch.sigmoid(x))


class Mish(nn.Module):  # https://github.com/digantamisra98/Mish
    def forward(self, x):
        return x.mul_(F.softplus(x).tanh())


class YOLOLayer(nn.Module):
    def __init__(self, anchors, nc, img_size, yolo_index, layers):
        super(YOLOLayer, self).__init__()
        self.anchors = torch.Tensor(anchors)
        self.index = yolo_index  # index of this layer in layers
        self.layers = layers  # model output layer indices
        self.nl = len(layers)  # number of output layers (3)
        self.na = len(anchors)  # number of anchors (3)
        self.nc = nc  # number of classes (80)
        self.no = nc + 5  # number of outputs (85)
        self.nx = 0  # initialize number of x gridpoints
        self.ny = 0  # initialize number of y gridpoints

        if ONNX_EXPORT:
            stride = [32, 16, 8][yolo_index]  # stride of this layer
            nx = img_size[1] // stride  # number x grid points
            ny = img_size[0] // stride  # number y grid points
            create_grids(self, img_size, (nx, ny))

    def forward(self, p, img_size, out):
        if ONNX_EXPORT:
            bs = 1  # batch size
        else:
            bs, _, ny, nx = p.shape  # bs, 255, 13, 13
            if (self.nx, self.ny) != (nx, ny):
                create_grids(self, img_size, (nx, ny), p.device, p.dtype)

        # p.view(bs, 255, 13, 13) -- > (bs, 3, 13, 13, 85)  # (bs, anchors, grid, grid, classes + xywh)
        p = p.view(bs, self.na, self.no, self.ny, self.nx).permute(0, 1, 3, 4, 2).contiguous()  # prediction

        if self.training:
            return p

        elif ONNX_EXPORT:
            # Avoid broadcasting for ANE operations
            m = self.na * self.nx * self.ny
            ng = 1 / self.ng.repeat((m, 1))
            grid_xy = self.grid_xy.repeat((1, self.na, 1, 1, 1)).view(m, 2)
            anchor_wh = self.anchor_wh.repeat((1, 1, self.nx, self.ny, 1)).view(m, 2) * ng

            p = p.view(m, self.no)
            xy = torch.sigmoid(p[:, 0:2]) + grid_xy  # x, y
            wh = torch.exp(p[:, 2:4]) * anchor_wh  # width, height
            p_cls = torch.sigmoid(p[:, 4:5]) if self.nc == 1 else \
                torch.sigmoid(p[:, 5:self.no]) * torch.sigmoid(p[:, 4:5])  # conf
            return p_cls, xy * ng, wh

        else:  # inference
            io = p.clone()  # inference output
            io[..., :2] = torch.sigmoid(io[..., :2]) + self.grid_xy  # xy
            io[..., 2:4] = torch.exp(io[..., 2:4]) * self.anchor_wh  # wh yolo method
            io[..., :4] *= self.stride
            torch.sigmoid_(io[..., 4:])
            return io.view(bs, -1, self.no), p  # view [1, 3, 13, 13, 85] as [1, 507, 85]


class Darknet(nn.Module):
    # YOLOv3 object detection model

    def __init__(self, cfg, img_size=(416, 416), arc='default'):
        super(Darknet, self).__init__()

        self.module_defs = parse_model_cfg(cfg)
        self.module_list, self.routs = create_modules(self.module_defs, img_size, arc)
        self.yolo_layers = get_yolo_layers(self)

        # Darknet Header https://github.com/AlexeyAB/darknet/issues/2914#issuecomment-496675346
        self.version = np.array([0, 2, 5], dtype=np.int32)  # (int32) version info: major, minor, revision
        self.seen = np.array([0], dtype=np.int64)  # (int64) number of images seen during training

    def forward(self, x, verbose=False):
        img_size = x.shape[-2:]
        yolo_out, out = [], []
        if verbose:
            str = ''
            print('0', x.shape)

        for i, (mdef, module) in enumerate(zip(self.module_defs, self.module_list)):
            mtype = mdef['type']
            if mtype in ['convolutional', 'upsample', 'maxpool']:
                x = module(x)
            elif mtype == 'shortcut':  # sum
                if verbose:
                    l = [i - 1] + module.layers  # layers
                    s = [list(x.shape)] + [list(out[i].shape) for i in module.layers]  # shapes
                    str = ' >> ' + ' + '.join(['layer %g %s' % x for x in zip(l, s)])
                x = module(x, out)  # weightedFeatureFusion()
            elif mtype == 'route':  # concat
                layers = mdef['layers']
                if verbose:
                    l = [i - 1] + layers  # layers
                    s = [list(x.shape)] + [list(out[i].shape) for i in layers]  # shapes
                    str = ' >> ' + ' + '.join(['layer %g %s' % x for x in zip(l, s)])
                if len(layers) == 1:
                    x = out[layers[0]]
                else:
                    try:
                        x = torch.cat([out[i] for i in layers], 1)
                    except:  # apply stride 2 for darknet reorg layer
                        out[layers[1]] = F.interpolate(out[layers[1]], scale_factor=[0.5, 0.5])
                        x = torch.cat([out[i] for i in layers], 1)
                    # print(''), [print(out[i].shape) for i in layers], print(x.shape)
            elif mtype == 'yolo':
                yolo_out.append(module(x, img_size, out))
            out.append(x if i in self.routs else [])
            if verbose:
                print('%g/%g %s -' % (i, len(self.module_list), mtype), list(x.shape), str)
                str = ''

        if self.training:  # train
            return yolo_out
        elif ONNX_EXPORT:  # export
            x = [torch.cat(x, 0) for x in zip(*yolo_out)]
            return x[0], torch.cat(x[1:3], 1)  # scores, boxes: 3780x80, 3780x4
        else:  # test
            io, p = zip(*yolo_out)  # inference output, training output
            return torch.cat(io, 1), p

    def fuse(self):
        # Fuse Conv2d + BatchNorm2d layers throughout model
        fused_list = nn.ModuleList()
        for a in list(self.children())[0]:
            if isinstance(a, nn.Sequential):
                for i, b in enumerate(a):
                    if isinstance(b, nn.modules.batchnorm.BatchNorm2d):
                        # fuse this bn layer with the previous conv2d layer
                        conv = a[i - 1]
                        fused = torch_utils.fuse_conv_and_bn(conv, b)
                        a = nn.Sequential(fused, *list(a.children())[i + 1:])
                        break
            fused_list.append(a)
        self.module_list = fused_list
        # model_info(self)  # yolov3-spp reduced from 225 to 152 layers


def get_yolo_layers(model):
    return [i for i, x in enumerate(model.module_defs) if x['type'] == 'yolo']  # [82, 94, 106] for yolov3


def create_grids(self, img_size=416, ng=(13, 13), device='cpu', type=torch.float32):
    nx, ny = ng  # x and y grid size
    self.img_size = max(img_size)
    self.stride = self.img_size / max(ng)

    # build xy offsets
    yv, xv = torch.meshgrid([torch.arange(ny), torch.arange(nx)])
    self.grid_xy = torch.stack((xv, yv), 2).to(device).type(type).view((1, 1, ny, nx, 2))

    # build wh gains
    self.anchor_vec = self.anchors.to(device) / self.stride
    self.anchor_wh = self.anchor_vec.view(1, self.na, 1, 1, 2).to(device).type(type)
    self.ng = torch.Tensor(ng).to(device)
    self.nx = nx
    self.ny = ny


def load_darknet_weights(self, weights, cutoff=-1):
    # Parses and loads the weights stored in 'weights'

    # Establish cutoffs (load layers between 0 and cutoff. if cutoff = -1 all are loaded)
    file = Path(weights).name
    if file == 'darknet53.conv.74':
        cutoff = 75
    elif file == 'yolov3-tiny.conv.15':
        cutoff = 15

    # Read weights file
    with open(weights, 'rb') as f:
        # Read Header https://github.com/AlexeyAB/darknet/issues/2914#issuecomment-496675346
        self.version = np.fromfile(f, dtype=np.int32, count=3)  # (int32) version info: major, minor, revision
        self.seen = np.fromfile(f, dtype=np.int64, count=1)  # (int64) number of images seen during training

        weights = np.fromfile(f, dtype=np.float32)  # the rest are weights

    ptr = 0
    for i, (mdef, module) in enumerate(zip(self.module_defs[:cutoff], self.module_list[:cutoff])):
        if mdef['type'] == 'convolutional':
            conv = module[0]
            if mdef['batch_normalize']:
                # Load BN bias, weights, running mean and running variance
                bn = module[1]
                nb = bn.bias.numel()  # number of biases
                # Bias
                bn.bias.data.copy_(torch.from_numpy(weights[ptr:ptr + nb]).view_as(bn.bias))
                ptr += nb
                # Weight
                bn.weight.data.copy_(torch.from_numpy(weights[ptr:ptr + nb]).view_as(bn.weight))
                ptr += nb
                # Running Mean
                bn.running_mean.data.copy_(torch.from_numpy(weights[ptr:ptr + nb]).view_as(bn.running_mean))
                ptr += nb
                # Running Var
                bn.running_var.data.copy_(torch.from_numpy(weights[ptr:ptr + nb]).view_as(bn.running_var))
                ptr += nb
            else:
                # Load conv. bias
                nb = conv.bias.numel()
                conv_b = torch.from_numpy(weights[ptr:ptr + nb]).view_as(conv.bias)
                conv.bias.data.copy_(conv_b)
                ptr += nb
            # Load conv. weights
            nw = conv.weight.numel()  # number of weights
            conv.weight.data.copy_(torch.from_numpy(weights[ptr:ptr + nw]).view_as(conv.weight))
            ptr += nw


def save_weights(self, path='model.weights', cutoff=-1):
    # Converts a PyTorch model to Darket format (*.pt to *.weights)
    # Note: Does not work if model.fuse() is applied
    with open(path, 'wb') as f:
        # Write Header https://github.com/AlexeyAB/darknet/issues/2914#issuecomment-496675346
        self.version.tofile(f)  # (int32) version info: major, minor, revision
        self.seen.tofile(f)  # (int64) number of images seen during training

        # Iterate through layers
        for i, (mdef, module) in enumerate(zip(self.module_defs[:cutoff], self.module_list[:cutoff])):
            if mdef['type'] == 'convolutional':
                conv_layer = module[0]
                # If batch norm, load bn first
                if mdef['batch_normalize']:
                    bn_layer = module[1]
                    bn_layer.bias.data.cpu().numpy().tofile(f)
                    bn_layer.weight.data.cpu().numpy().tofile(f)
                    bn_layer.running_mean.data.cpu().numpy().tofile(f)
                    bn_layer.running_var.data.cpu().numpy().tofile(f)
                # Load conv bias
                else:
                    conv_layer.bias.data.cpu().numpy().tofile(f)
                # Load conv weights
                conv_layer.weight.data.cpu().numpy().tofile(f)


def convert(cfg='cfg/yolov3-spp.cfg', weights='weights/yolov3-spp.weights'):
    # Converts between PyTorch and Darknet format per extension (i.e. *.weights convert to *.pt and vice versa)
    # from models import *; convert('cfg/yolov3-spp.cfg', 'weights/yolov3-spp.weights')

    # Initialize model
    model = Darknet(cfg)

    # Load weights and save
    if weights.endswith('.pt'):  # if PyTorch format
        model.load_state_dict(torch.load(weights, map_location='cpu')['model'])
        save_weights(model, path='converted.weights', cutoff=-1)
        print("Success: converted '%s' to 'converted.weights'" % weights)

    elif weights.endswith('.weights'):  # darknet format
        _ = load_darknet_weights(model, weights)

        chkpt = {'epoch': -1,
                 'best_fitness': None,
                 'training_results': None,
                 'model': model.state_dict(),
                 'optimizer': None}

        torch.save(chkpt, 'converted.pt')
        print("Success: converted '%s' to 'converted.pt'" % weights)

    else:
        print('Error: extension not supported.')


def attempt_download(weights):
    # Attempt to download pretrained weights if not found locally
    msg = weights + ' missing, try downloading from https://drive.google.com/open?id=1LezFG5g3BCW6iYaV89B2i64cqEUZD7e0'

    if weights and not os.path.isfile(weights):
        d = {'yolov3-spp.weights': '16lYS4bcIdM2HdmyJBVDOvt3Trx6N3W2R',
             'yolov3.weights': '1uTlyDWlnaqXcsKOktP5aH_zRDbfcDp-y',
             'yolov3-tiny.weights': '1CCF-iNIIkYesIDzaPvdwlcf7H9zSsKZQ',
             'yolov3-spp.pt': '1f6Ovy3BSq2wYq4UfvFUpxJFNDFfrIDcR',
             'yolov3.pt': '1SHNFyoe5Ni8DajDNEqgB2oVKBb_NoEad',
             'yolov3-tiny.pt': '10m_3MlpQwRtZetQxtksm9jqHrPTHZ6vo',
             'darknet53.conv.74': '1WUVBid-XuoUBmvzBVUCBl_ELrzqwA8dJ',
             'yolov3-tiny.conv.15': '1Bw0kCpplxUqyRYAJr9RY9SGnOJbo9nEj',
             'ultralytics49.pt': '158g62Vs14E3aj7oPVPuEnNZMKFNgGyNq',
             'ultralytics68.pt': '1Jm8kqnMdMGUUxGo8zMFZMJ0eaPwLkxSG',
             'yolov3-spp-ultralytics.pt': '1UcR-zVoMs7DH5dj3N1bswkiQTA4dmKF4'}

        file = Path(weights).name
        if file in d:
            r = gdrive_download(id=d[file], name=weights)
        else:  # download from pjreddie.com
            url = 'https://pjreddie.com/media/files/' + file
            print('Downloading ' + url)
            r = os.system('curl -f ' + url + ' -o ' + weights)

        # Error check
        if not (r == 0 and os.path.exists(weights) and os.path.getsize(weights) > 1E6):  # weights exist and > 1MB
            os.system('rm ' + weights)  # remove partial downloads
            raise Exception(msg)