378 lines
16 KiB
Python
378 lines
16 KiB
Python
import os
|
|
from collections import defaultdict,OrderedDict
|
|
|
|
import torch.nn as nn
|
|
|
|
from utils.parse_config import *
|
|
from utils.utils import *
|
|
from utils.syncbn import SyncBN
|
|
import time
|
|
import math
|
|
|
|
batch_norm=SyncBN #nn.BatchNorm2d
|
|
|
|
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()
|
|
yolo_layer_count = 0
|
|
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, batch_norm(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']))
|
|
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]
|
|
nC = int(module_def['classes']) # number of classes
|
|
img_size = (int(hyperparams['width']),int(hyperparams['height']))
|
|
# Define detection layer
|
|
yolo_layer = YOLOLayer(anchors, nC, hyperparams['nID'], img_size, yolo_layer_count, cfg=hyperparams['cfg'])
|
|
modules.add_module('yolo_%d' % i, yolo_layer)
|
|
yolo_layer_count += 1
|
|
|
|
# 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__()
|
|
|
|
def forward(self, x):
|
|
return x
|
|
|
|
|
|
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, nID, img_size, yolo_layer, cfg):
|
|
super(YOLOLayer, self).__init__()
|
|
self.layer = yolo_layer
|
|
nA = len(anchors)
|
|
self.anchors = torch.FloatTensor(anchors)
|
|
self.nA = nA # number of anchors (3)
|
|
self.nC = nC # number of classes (80)
|
|
self.nID = nID # number of identities
|
|
self.img_size = 0
|
|
self.emb_dim = 512
|
|
|
|
self.SmoothL1Loss = nn.SmoothL1Loss()
|
|
self.SoftmaxLoss = nn.CrossEntropyLoss(ignore_index=-1)
|
|
self.CrossEntropyLoss = nn.CrossEntropyLoss()
|
|
self.IDLoss = nn.CrossEntropyLoss(ignore_index=-1)
|
|
self.s_c = nn.Parameter(1*torch.ones(1)) # -4.15
|
|
self.s_r = nn.Parameter(1*torch.ones(1)) # -4.85
|
|
self.s_id = nn.Parameter(1*torch.ones(1)) # -2.3
|
|
self.emb_scale = math.sqrt(2) * math.log(self.nID-1)
|
|
|
|
|
|
def forward(self, p_cat, img_size, targets=None, classifier=None, test_emb=False):
|
|
p, p_emb = p_cat[:, :24, ...], p_cat[:, 24:, ...]
|
|
nB, nGh, nGw = p.shape[0], p.shape[-2], p.shape[-1]
|
|
|
|
if self.img_size != img_size:
|
|
create_grids(self, img_size, nGh, nGw)
|
|
|
|
if p.is_cuda:
|
|
self.grid_xy = self.grid_xy.cuda()
|
|
self.anchor_wh = self.anchor_wh.cuda()
|
|
|
|
# p.view(bs, 255, 13, 13) -- > (bs, 3, 13, 13, 80) # (bs, anchors, grid, grid, classes + xywh)
|
|
p = p.view(nB, self.nA, self.nC + 5, nGh, nGw).permute(0, 1, 3, 4, 2).contiguous() # prediction
|
|
|
|
p_emb = p_emb.permute(0,2,3,1).contiguous()
|
|
p_box = p[..., :4]
|
|
p_conf = p[..., 4:6].permute(0, 4, 1, 2, 3) # Conf
|
|
|
|
# Training
|
|
if targets is not None:
|
|
if test_emb:
|
|
tconf, tbox, tids = build_targets_max(targets, self.anchor_vec.cuda(), self.nA, self.nC, nGh, nGw)
|
|
else:
|
|
tconf, tbox, tids = build_targets_thres(targets, self.anchor_vec.cuda(), self.nA, self.nC, nGh, nGw)
|
|
tconf, tbox, tids = tconf.cuda(), tbox.cuda(), tids.cuda()
|
|
mask = tconf > 0
|
|
|
|
# Compute losses
|
|
nT = sum([len(x) for x in targets]) # number of targets
|
|
nM = mask.sum().float() # number of anchors (assigned to targets)
|
|
nP = torch.ones_like(mask).sum().float()
|
|
if nM > 0:
|
|
lbox = self.SmoothL1Loss(p_box[mask], tbox[mask])
|
|
else:
|
|
FT = torch.cuda.FloatTensor if p_conf.is_cuda else torch.FloatTensor
|
|
lbox, lconf = FT([0]), FT([0])
|
|
lconf = self.SoftmaxLoss(p_conf, tconf)
|
|
lid = torch.Tensor(1).fill_(0).squeeze().cuda()
|
|
emb_mask,_ = mask.max(1)
|
|
|
|
# For convenience we use max(1) to decide the id, TODO: more reseanable strategy
|
|
tids,_ = tids.max(1)
|
|
tids = tids[emb_mask]
|
|
embedding = p_emb[emb_mask].contiguous()
|
|
embedding = self.emb_scale * F.normalize(embedding)
|
|
nI = emb_mask.sum().float()
|
|
|
|
if test_emb:
|
|
if np.prod(embedding.shape)==0 or np.prod(tids.shape) == 0:
|
|
return torch.zeros(0, self. emb_dim+1).cuda()
|
|
emb_and_gt = torch.cat([embedding, tids.float()], dim=1)
|
|
return emb_and_gt
|
|
|
|
if len(embedding) > 1:
|
|
logits = classifier(embedding).contiguous()
|
|
lid = self.IDLoss(logits, tids.squeeze())
|
|
|
|
# Sum loss components
|
|
#loss = torch.exp(-self.s_r)*lbox + torch.exp(-self.s_c)*lconf + torch.exp(-self.s_id)*lid + \
|
|
# (self.s_r + self.s_c + self.s_id)
|
|
#loss *= 0.5
|
|
loss = lbox + lconf + lid
|
|
|
|
return loss, loss.item(), lbox.item(), lconf.item(), lid.item(), nT
|
|
|
|
else:
|
|
p_conf = torch.softmax(p_conf, dim=1)[:,1,...].unsqueeze(-1)
|
|
p_emb = p_emb.unsqueeze(1).repeat(1,self.nA,1,1,1).contiguous()
|
|
p_cls = torch.zeros(nB,self.nA,nGh,nGw,1).cuda() # Temp
|
|
p = torch.cat([p_box, p_conf, p_cls, p_emb], dim=-1)
|
|
p[..., :4] = decode_delta_map(p[..., :4], self.anchor_vec.to(p))
|
|
p[..., :4] *= self.stride
|
|
|
|
# reshape from [nB, nA, nGh, nGw, 5 + nD] to [nB, -1, 5+nD]
|
|
return p.view(nB, -1, p.shape[-1])
|
|
|
|
|
|
class Darknet(nn.Module):
|
|
"""YOLOv3 object detection model"""
|
|
|
|
def __init__(self, cfg_path, img_size=(1088, 608), nID=1591, test_emb=False):
|
|
super(Darknet, self).__init__()
|
|
|
|
self.module_defs = parse_model_cfg(cfg_path)
|
|
self.module_defs[0]['cfg'] = cfg_path
|
|
self.module_defs[0]['nID'] = nID
|
|
self.hyperparams, self.module_list = create_modules(self.module_defs)
|
|
self.img_size = img_size
|
|
self.loss_names = ['loss', 'box', 'conf', 'id', 'nT']
|
|
self.losses = OrderedDict()
|
|
for ln in self.loss_names:
|
|
self.losses[ln] = 0
|
|
self.emb_dim = 512
|
|
self.classifier = nn.Linear(self.emb_dim, nID)
|
|
self.test_emb=test_emb
|
|
|
|
|
|
def forward(self, x, targets=None, targets_len=None):
|
|
self.losses = OrderedDict()
|
|
for ln in self.loss_names:
|
|
self.losses[ln] = 0
|
|
is_training = (targets is not None) and (not self.test_emb)
|
|
#img_size = x.shape[-1]
|
|
layer_outputs = []
|
|
output = []
|
|
|
|
for i, (module_def, module) in enumerate(zip(self.module_defs, self.module_list)):
|
|
mtype = module_def['type']
|
|
if mtype in ['convolutional', 'upsample', 'maxpool']:
|
|
x = module(x)
|
|
elif mtype == 'route':
|
|
layer_i = [int(x) for x in module_def['layers'].split(',')]
|
|
if len(layer_i) == 1:
|
|
x = layer_outputs[layer_i[0]]
|
|
else:
|
|
x = torch.cat([layer_outputs[i] for i in layer_i], 1)
|
|
elif mtype == 'shortcut':
|
|
layer_i = int(module_def['from'])
|
|
x = layer_outputs[-1] + layer_outputs[layer_i]
|
|
elif mtype == 'yolo':
|
|
if is_training: # get loss
|
|
targets = [targets[i][:int(l)] for i,l in enumerate(targets_len)]
|
|
x, *losses = module[0](x, self.img_size, targets, self.classifier)
|
|
for name, loss in zip(self.loss_names, losses):
|
|
self.losses[name] += loss
|
|
elif self.test_emb:
|
|
targets = [targets[i][:int(l)] for i,l in enumerate(targets_len)]
|
|
x = module[0](x, self.img_size, targets, self.classifier, self.test_emb)
|
|
else: # get detections
|
|
x = module[0](x, self.img_size)
|
|
output.append(x)
|
|
layer_outputs.append(x)
|
|
|
|
if is_training:
|
|
self.losses['nT'] /= 3
|
|
output = [o.squeeze() for o in output]
|
|
return sum(output), torch.Tensor(list(self.losses.values())).cuda()
|
|
elif self.test_emb:
|
|
return torch.cat(output, 0)
|
|
return torch.cat(output, 1)
|
|
|
|
|
|
def create_grids(self, img_size, nGh, nGw):
|
|
self.stride = img_size[0]/nGw
|
|
assert self.stride == img_size[1] / nGh
|
|
|
|
# build xy offsets
|
|
grid_x = torch.arange(nGw).repeat((nGh, 1)).view((1, 1, nGh, nGw)).float()
|
|
grid_y = torch.arange(nGh).repeat((nGw, 1)).transpose(0,1).view((1, 1, nGh, nGw)).float()
|
|
#grid_y = grid_x.permute(0, 1, 3, 2)
|
|
self.grid_xy = torch.stack((grid_x, grid_y), 4)
|
|
|
|
# build wh gains
|
|
self.anchor_vec = self.anchors / self.stride
|
|
self.anchor_wh = self.anchor_vec.view(1, self.nA, 1, 1, 2)
|
|
|
|
|
|
def load_darknet_weights(self, weights, cutoff=-1):
|
|
# Parses and loads the weights stored in 'weights'
|
|
# cutoff: save layers between 0 and cutoff (if cutoff = -1 all are saved)
|
|
weights_file = weights.split(os.sep)[-1]
|
|
|
|
# Try to download weights if not available locally
|
|
if not os.path.isfile(weights):
|
|
try:
|
|
os.system('wget https://pjreddie.com/media/files/' + weights_file + ' -O ' + weights)
|
|
except IOError:
|
|
print(weights + ' not found')
|
|
|
|
# Establish cutoffs
|
|
if weights_file == 'darknet53.conv.74':
|
|
cutoff = 75
|
|
elif weights_file == 'yolov3-tiny.conv.15':
|
|
cutoff = 15
|
|
|
|
# Open the weights file
|
|
fp = open(weights, '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] # number of images seen during training
|
|
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 # number of images seen during training
|
|
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()
|