from collections import defaultdict import torch.nn as nn from utils.parse_config import * from utils.utils import * ONNX_EXPORT = True def create_modules(module_defs): """ 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() for i, module_def in enumerate(module_defs): modules = nn.Sequential() if module_def['type'] == 'convolutional': bn = int(module_def['batch_normalize']) filters = int(module_def['filters']) kernel_size = int(module_def['size']) pad = (kernel_size - 1) // 2 if int(module_def['pad']) else 0 modules.add_module('conv_%d' % i, nn.Conv2d(in_channels=output_filters[-1], out_channels=filters, kernel_size=kernel_size, stride=int(module_def['stride']), padding=pad, bias=not bn)) if bn: modules.add_module('batch_norm_%d' % i, nn.BatchNorm2d(filters)) if module_def['activation'] == 'leaky': modules.add_module('leaky_%d' % i, nn.LeakyReLU(0.1)) elif module_def['type'] == 'maxpool': kernel_size = int(module_def['size']) stride = int(module_def['stride']) if kernel_size == 2 and stride == 1: modules.add_module('_debug_padding_%d' % i, nn.ZeroPad2d((0, 1, 0, 1))) maxpool = nn.MaxPool2d(kernel_size=kernel_size, stride=stride, padding=int((kernel_size - 1) // 2)) modules.add_module('maxpool_%d' % i, maxpool) elif module_def['type'] == 'upsample': # upsample = nn.Upsample(scale_factor=int(module_def['stride']), mode='nearest') # WARNING: deprecated upsample = Upsample(scale_factor=int(module_def['stride']), mode='nearest') modules.add_module('upsample_%d' % i, upsample) elif module_def['type'] == 'route': layers = [int(x) for x in module_def['layers'].split(',')] filters = sum([output_filters[i + 1 if i > 0 else i] for i in layers]) modules.add_module('route_%d' % i, EmptyLayer()) elif module_def['type'] == 'shortcut': filters = output_filters[int(module_def['from'])] modules.add_module('shortcut_%d' % i, EmptyLayer()) elif module_def['type'] == 'yolo': anchor_idxs = [int(x) for x in module_def['mask'].split(',')] # Extract anchors anchors = [float(x) for x in module_def['anchors'].split(',')] anchors = [(anchors[i], anchors[i + 1]) for i in range(0, len(anchors), 2)] anchors = [anchors[i] for i in anchor_idxs] num_classes = int(module_def['classes']) img_height = int(hyperparams['height']) # Define detection layer yolo_layer = YOLOLayer(anchors, num_classes, img_height, anchor_idxs, cfg=hyperparams['cfg']) modules.add_module('yolo_%d' % i, yolo_layer) # Register module list and number of output filters module_list.append(modules) output_filters.append(filters) return hyperparams, module_list class EmptyLayer(nn.Module): """Placeholder for 'route' and 'shortcut' layers""" def __init__(self): super(EmptyLayer, self).__init__() class Upsample(nn.Module): # Custom Upsample layer (nn.Upsample gives deprecated warning message) def __init__(self, scale_factor=1, mode='nearest'): super(Upsample, self).__init__() self.scale_factor = scale_factor self.mode = mode def forward(self, x): return F.interpolate(x, scale_factor=self.scale_factor, mode=self.mode) class YOLOLayer(nn.Module): def __init__(self, anchors, nC, img_dim, anchor_idxs, cfg): super(YOLOLayer, self).__init__() anchors = [(a_w, a_h) for a_w, a_h in anchors] # (pixels) nA = len(anchors) self.anchors = anchors self.nA = nA # number of anchors (3) self.nC = nC # number of classes (80) self.bbox_attrs = 5 + nC self.img_dim = img_dim # from hyperparams in cfg file, NOT from parser if anchor_idxs[0] == (nA * 2): # 6 stride = 32 elif anchor_idxs[0] == nA: # 3 stride = 16 else: stride = 8 if cfg.endswith('yolov3-tiny.cfg'): stride *= 2 # Build anchor grids nG = int(self.img_dim / stride) # number grid points self.grid_x = torch.arange(nG).repeat(nG, 1).view([1, 1, nG, nG]).float() self.grid_y = torch.arange(nG).repeat(nG, 1).t().view([1, 1, nG, nG]).float() self.anchor_wh = torch.FloatTensor([(a_w / stride, a_h / stride) for a_w, a_h in anchors]) # scale anchors self.anchor_w = self.anchor_wh[:, 0].view((1, nA, 1, 1)) self.anchor_h = self.anchor_wh[:, 1].view((1, nA, 1, 1)) self.weights = class_weights() self.loss_means = torch.ones(6) self.tx, self.ty, self.tw, self.th = [], [], [], [] self.yolo_layer = anchor_idxs[0] / nA # 2, 1, 0 self.stride = stride if ONNX_EXPORT: # use fully populated and reshaped tensors self.anchor_w = self.anchor_w.repeat((1, 1, nG, nG)).view(1, -1, 1) self.anchor_h = self.anchor_h.repeat((1, 1, nG, nG)).view(1, -1, 1) self.grid_x = self.grid_x.repeat(1, nA, 1, 1).view(1, -1, 1) self.grid_y = self.grid_y.repeat(1, nA, 1, 1).view(1, -1, 1) self.grid_xy = torch.cat((self.grid_x, self.grid_y), 2) self.anchor_wh = torch.cat((self.anchor_w, self.anchor_h), 2) / nG def forward(self, p, targets=None, batch_report=False, var=None): FT = torch.cuda.FloatTensor if p.is_cuda else torch.FloatTensor bs = p.shape[0] # batch size nG = p.shape[2] # number of grid points if p.is_cuda and not self.weights.is_cuda: self.grid_x, self.grid_y = self.grid_x.cuda(), self.grid_y.cuda() self.anchor_w, self.anchor_h = self.anchor_w.cuda(), self.anchor_h.cuda() self.weights, self.loss_means = self.weights.cuda(), self.loss_means.cuda() # p.view(12, 255, 13, 13) -- > (12, 3, 13, 13, 80) # (bs, anchors, grid, grid, classes + xywh) p = p.view(bs, self.nA, self.bbox_attrs, nG, nG).permute(0, 1, 3, 4, 2).contiguous() # prediction # Training if targets is not None: MSELoss = nn.MSELoss() BCEWithLogitsLoss = nn.BCEWithLogitsLoss() CrossEntropyLoss = nn.CrossEntropyLoss() # Get outputs x = torch.sigmoid(p[..., 0]) # Center x y = torch.sigmoid(p[..., 1]) # Center y p_conf = p[..., 4] # Conf p_cls = p[..., 5:] # Class # Width and height (yolo method) w = p[..., 2] # Width h = p[..., 3] # Height width = torch.exp(w.data) * self.anchor_w height = torch.exp(h.data) * self.anchor_h # Width and height (power method) # w = torch.sigmoid(p[..., 2]) # Width # h = torch.sigmoid(p[..., 3]) # Height # width = ((w.data * 2) ** 2) * self.anchor_w # height = ((h.data * 2) ** 2) * self.anchor_h p_boxes = None if batch_report: # Predicted boxes: add offset and scale with anchors (in grid space, i.e. 0-13) gx = x.data + self.grid_x[:, :, :nG, :nG] gy = y.data + self.grid_y[:, :, :nG, :nG] p_boxes = torch.stack((gx - width / 2, gy - height / 2, gx + width / 2, gy + height / 2), 4) # x1y1x2y2 tx, ty, tw, th, mask, tcls, TP, FP, FN, TC = \ build_targets(p_boxes, p_conf, p_cls, targets, self.anchor_wh, self.nA, self.nC, nG, batch_report) tcls = tcls[mask] if x.is_cuda: tx, ty, tw, th, mask, tcls = tx.cuda(), ty.cuda(), tw.cuda(), th.cuda(), mask.cuda(), tcls.cuda() # Compute losses nT = sum([len(x) for x in targets]) # number of targets nM = mask.sum().float() # number of anchors (assigned to targets) nB = len(targets) # batch size k = nM / nB if nM > 0: lx = k * MSELoss(x[mask], tx[mask]) ly = k * MSELoss(y[mask], ty[mask]) lw = k * MSELoss(w[mask], tw[mask]) lh = k * MSELoss(h[mask], th[mask]) # self.tx.extend(tx[mask].data.numpy()) # self.ty.extend(ty[mask].data.numpy()) # self.tw.extend(tw[mask].data.numpy()) # self.th.extend(th[mask].data.numpy()) # print([np.mean(self.tx), np.std(self.tx)],[np.mean(self.ty), np.std(self.ty)],[np.mean(self.tw), np.std(self.tw)],[np.mean(self.th), np.std(self.th)]) # [0.5040668, 0.2885492] [0.51384246, 0.28328574] [-0.4754091, 0.57951087] [-0.25998235, 0.44858757] # [0.50184494, 0.2858976] [0.51747805, 0.2896323] [0.12962963, 0.6263085] [-0.2722081, 0.61574113] # [0.5032071, 0.28825334] [0.5063132, 0.2808862] [0.21124361, 0.44760725] [0.35445485, 0.6427766] # import matplotlib.pyplot as plt # plt.hist(self.x) # lconf = k * BCEWithLogitsLoss(p_conf[mask], mask[mask].float()) lcls = (k / 4) * CrossEntropyLoss(p_cls[mask], torch.argmax(tcls, 1)) # lcls = (k * 10) * BCEWithLogitsLoss(p_cls[mask], tcls.float()) else: lx, ly, lw, lh, lcls, lconf = FT([0]), FT([0]), FT([0]), FT([0]), FT([0]), FT([0]) # lconf += k * BCEWithLogitsLoss(p_conf[~mask], mask[~mask].float()) lconf = (k * 64) * BCEWithLogitsLoss(p_conf, mask.float()) # Sum loss components balance_losses_flag = False if balance_losses_flag: k = 1 / self.loss_means.clone() loss = (lx * k[0] + ly * k[1] + lw * k[2] + lh * k[3] + lconf * k[4] + lcls * k[5]) / k.mean() self.loss_means = self.loss_means * 0.99 + \ FT([lx.data, ly.data, lw.data, lh.data, lconf.data, lcls.data]) * 0.01 else: loss = lx + ly + lw + lh + lconf + lcls # Sum False Positives from unassigned anchors FPe = torch.zeros(self.nC) if batch_report: i = torch.sigmoid(p_conf[~mask]) > 0.5 if i.sum() > 0: FP_classes = torch.argmax(p_cls[~mask][i], 1) FPe = torch.bincount(FP_classes, minlength=self.nC).float().cpu() # extra FPs return loss, loss.item(), lx.item(), ly.item(), lw.item(), lh.item(), lconf.item(), lcls.item(), \ nT, TP, FP, FPe, FN, TC else: if ONNX_EXPORT: p = p.view(1, -1, 85) xy = torch.sigmoid(p[..., 0:2]) + self.grid_xy # x, y width_height = torch.exp(p[..., 2:4]) * self.anchor_wh # width, height p_conf = torch.sigmoid(p[..., 4:5]) # Conf ## p_cls = torch.sigmoid(p[..., 5:85]) # Class p_cls = F.softmax(p[..., 5:85], 2) * p_conf # SSD-like conf # p_cls = torch.exp(p[..., 5:85]) / torch.exp(p[..., 5:85]).sum(2).unsqueeze(2) #* p_conf # F.softmax() equivalent return torch.cat((xy / nG, width_height, p_conf, p_cls), 2).squeeze().t() p[..., 0] = torch.sigmoid(p[..., 0]) + self.grid_x # x p[..., 1] = torch.sigmoid(p[..., 1]) + self.grid_y # y p[..., 2] = torch.exp(p[..., 2]) * self.anchor_w # width p[..., 3] = torch.exp(p[..., 3]) * self.anchor_h # height p[..., 4] = torch.sigmoid(p[..., 4]) # p_conf p[..., :4] *= self.stride # reshape from [1, 3, 13, 13, 85] to [1, 507, 85] return p.view(bs, -1, 5 + self.nC) class Darknet(nn.Module): """YOLOv3 object detection model""" def __init__(self, cfg_path, img_size=416): super(Darknet, self).__init__() self.module_defs = parse_model_config(cfg_path) self.module_defs[0]['cfg'] = cfg_path self.module_defs[0]['height'] = img_size self.hyperparams, self.module_list = create_modules(self.module_defs) self.img_size = img_size self.loss_names = ['loss', 'x', 'y', 'w', 'h', 'conf', 'cls', 'nT', 'TP', 'FP', 'FPe', 'FN', 'TC'] def forward(self, x, targets=None, batch_report=False, var=0): self.losses = defaultdict(float) is_training = targets is not None layer_outputs = [] output = [] for i, (module_def, module) in enumerate(zip(self.module_defs, self.module_list)): if module_def['type'] in ['convolutional', 'upsample', 'maxpool']: x = module(x) elif module_def['type'] == 'route': layer_i = [int(x) for x in module_def['layers'].split(',')] x = torch.cat([layer_outputs[i] for i in layer_i], 1) elif module_def['type'] == 'shortcut': layer_i = int(module_def['from']) x = layer_outputs[-1] + layer_outputs[layer_i] elif module_def['type'] == 'yolo': # Train phase: get loss if is_training: x, *losses = module[0](x, targets, batch_report, var) for name, loss in zip(self.loss_names, losses): self.losses[name] += loss # Test phase: Get detections else: x = module(x) output.append(x) layer_outputs.append(x) if is_training: if batch_report: self.losses['TC'] /= 3 # target category metrics = torch.zeros(3, len(self.losses['FPe'])) # TP, FP, FN ui = np.unique(self.losses['TC'])[1:] for i in ui: j = self.losses['TC'] == float(i) metrics[0, i] = (self.losses['TP'][j] > 0).sum().float() # TP metrics[1, i] = (self.losses['FP'][j] > 0).sum().float() # FP metrics[2, i] = (self.losses['FN'][j] == 3).sum().float() # FN metrics[1] += self.losses['FPe'] self.losses['TP'] = metrics[0].sum() self.losses['FP'] = metrics[1].sum() self.losses['FN'] = metrics[2].sum() self.losses['metrics'] = metrics else: self.losses['TP'] = 0 self.losses['FP'] = 0 self.losses['FN'] = 0 self.losses['nT'] /= 3 self.losses['TC'] = 0 if ONNX_EXPORT: # Produce a single-layer *.onnx model (upsample ops not working in PyTorch 1.0 export yet) output = output[0] # first layer reshaped to 85 x 507 return output[5:85].t(), output[:4].t() # ONNX scores, boxes return sum(output) if is_training else torch.cat(output, 1) def load_weights(self, weights_path, cutoff=-1): # Parses and loads the weights stored in 'weights_path' # @:param cutoff - save layers between 0 and cutoff (cutoff = -1 -> all are saved) if weights_path.endswith('darknet53.conv.74'): cutoff = 75 # Open the weights file fp = open(weights_path, 'rb') header = np.fromfile(fp, dtype=np.int32, count=5) # First five are header values # Needed to write header when saving weights self.header_info = header self.seen = header[3] weights = np.fromfile(fp, dtype=np.float32) # The rest are weights fp.close() ptr = 0 for i, (module_def, module) in enumerate(zip(self.module_defs[:cutoff], self.module_list[:cutoff])): if module_def['type'] == 'convolutional': conv_layer = module[0] if module_def['batch_normalize']: # Load BN bias, weights, running mean and running variance bn_layer = module[1] num_b = bn_layer.bias.numel() # Number of biases # Bias bn_b = torch.from_numpy(weights[ptr:ptr + num_b]).view_as(bn_layer.bias) bn_layer.bias.data.copy_(bn_b) ptr += num_b # Weight bn_w = torch.from_numpy(weights[ptr:ptr + num_b]).view_as(bn_layer.weight) bn_layer.weight.data.copy_(bn_w) ptr += num_b # Running Mean bn_rm = torch.from_numpy(weights[ptr:ptr + num_b]).view_as(bn_layer.running_mean) bn_layer.running_mean.data.copy_(bn_rm) ptr += num_b # Running Var bn_rv = torch.from_numpy(weights[ptr:ptr + num_b]).view_as(bn_layer.running_var) bn_layer.running_var.data.copy_(bn_rv) ptr += num_b else: # Load conv. bias num_b = conv_layer.bias.numel() conv_b = torch.from_numpy(weights[ptr:ptr + num_b]).view_as(conv_layer.bias) conv_layer.bias.data.copy_(conv_b) ptr += num_b # Load conv. weights num_w = conv_layer.weight.numel() conv_w = torch.from_numpy(weights[ptr:ptr + num_w]).view_as(conv_layer.weight) conv_layer.weight.data.copy_(conv_w) ptr += num_w """ @:param path - path of the new weights file @:param cutoff - save layers between 0 and cutoff (cutoff = -1 -> all are saved) """ def save_weights(self, path, cutoff=-1): fp = open(path, 'wb') self.header_info[3] = self.seen self.header_info.tofile(fp) # Iterate through layers for i, (module_def, module) in enumerate(zip(self.module_defs[:cutoff], self.module_list[:cutoff])): if module_def['type'] == 'convolutional': conv_layer = module[0] # If batch norm, load bn first if module_def['batch_normalize']: bn_layer = module[1] bn_layer.bias.data.cpu().numpy().tofile(fp) bn_layer.weight.data.cpu().numpy().tofile(fp) bn_layer.running_mean.data.cpu().numpy().tofile(fp) bn_layer.running_var.data.cpu().numpy().tofile(fp) # Load conv bias else: conv_layer.bias.data.cpu().numpy().tofile(fp) # Load conv weights conv_layer.weight.data.cpu().numpy().tofile(fp) fp.close()