File size: 9,075 Bytes
5de1792 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 | #!/usr/bin/env python
# pylint: disable=W0201
import sys
import argparse
import yaml
import numpy as np
from tqdm import tqdm
# torch
import torch
import torch.nn as nn
import torch.optim as optim
import torch.nn.functional as F
# torchlight
import torchlight.torchlight as torchlight
from torchlight.torchlight import str2bool
from torchlight.torchlight import DictAction
from torchlight.torchlight import import_class
from .evaluate_metrics import *
from .processor import Processor
def weights_init(m):
classname = m.__class__.__name__
if classname.find('Conv1d') != -1:
m.weight.data.normal_(0.0, 0.02)
if m.bias is not None:
m.bias.data.fill_(0)
elif classname.find('Conv2d') != -1:
m.weight.data.normal_(0.0, 0.02)
if m.bias is not None:
m.bias.data.fill_(0)
elif classname.find('BatchNorm') != -1:
m.weight.data.normal_(1.0, 0.02)
m.bias.data.fill_(0)
class DT_Processor(Processor):
"""
Processor for Skeleton-based Action Recgnition
"""
def load_model(self):
self.model = self.io.load_model(self.arg.model,
**(self.arg.model_args))
self.model.apply(weights_init)
for name, param in self.model.encoder_q.named_parameters():
if name not in ['fc.weight', 'fc.bias']:
param.requires_grad = False
self.num_grad_layers = 2
self.loss = nn.CrossEntropyLoss()
def load_optimizer(self):
parameters = list(filter(lambda p: p.requires_grad, self.model.parameters()))
assert len(parameters) == self.num_grad_layers
if self.arg.optimizer == 'SGD':
self.optimizer = optim.SGD(
parameters,
lr=self.arg.base_lr,
momentum=0.9,
nesterov=self.arg.nesterov,
weight_decay=self.arg.weight_decay)
elif self.arg.optimizer == 'Adam':
self.optimizer = optim.Adam(
parameters,
lr=self.arg.base_lr,
weight_decay=self.arg.weight_decay)
else:
raise ValueError()
def adjust_lr(self):
if self.arg.optimizer == 'SGD' and self.arg.step:
lr = self.arg.base_lr * (
0.1**np.sum(self.meta_info['epoch'] > np.array(self.arg.step)))
for param_group in self.optimizer.param_groups:
param_group['lr'] = lr
self.lr = lr
else:
self.lr = self.arg.base_lr
def show_topk(self, k):
rank = self.result.argsort()
hit_top_k = [l in rank[i, -k:] for i, l in enumerate(self.label)]
accuracy = sum(hit_top_k) * 1.0 / len(hit_top_k)
self.io.print_log('\tTop{}: {:.2f}%'.format(k, 100 * accuracy))
def show_best(self, k):
rank = self.result.argsort()
hit_top_k = [l in rank[i, -k:] for i, l in enumerate(self.label)]
accuracy = 100 * sum(hit_top_k) * 1.0 / len(hit_top_k)
accuracy = round(accuracy, 5)
self.current_result = accuracy
if self.best_result <= accuracy:
self.best_result = accuracy
self.io.print_log('\tBest Top{}: {:.2f}%'.format(k, self.best_result))
def train(self,epoch):
self.model.train()
self.adjust_lr()
loader = self.data_loader['train']
loss_value = []
process = tqdm(loader)
for data, label, start_pos, video_name in process:
# get data
data = data.float().to(self.dev)
label = label.long().to(self.dev)#N,T
# forward
output = self.model(data)
N,T,_ = output.shape
out4loss = output.reshape(N*T,-1)
loss = self.loss(out4loss, label.reshape(N*T,))
# backward
self.optimizer.zero_grad()
loss.backward()
self.optimizer.step()
# statistics
self.iter_info['loss'] = loss.data.item()
self.iter_info['lr'] = '{:.6f}'.format(self.lr)
loss_value.append(self.iter_info['loss'])
self.show_iter_info()
self.meta_info['iter'] += 1
self.epoch_info['mean_loss']= np.mean(loss_value)
self.show_epoch_info()
self.io.print_timer()
def merge_list(self, vid_list):
group_list = []
idx_per = []
vid_list_uqk = []
for idx, name in enumerate(vid_list):
if idx == len(vid_list)-1:
last_name = ''
else:
last_name = vid_list[idx+1]
if name != last_name:
idx_per.append(idx)
group_list.append(np.asarray(idx_per).astype(int) )
vid_list_uqk.append(name)
idx_per = []
else:
idx_per.append(idx)
return group_list, vid_list_uqk
def test(self, epoch):
self.model.eval()
loader = self.data_loader['test']
loss_value = []
result_frag = []
label_frag = []
start_frag = []
video_name_frag = []
process = tqdm(loader)
sfm = nn.Softmax(dim=-1)
for data, label,start_pos,video_name in process:
# get data
data = data.float().to(self.dev)
label = label.long().to(self.dev)
# inference
with torch.no_grad():
output = self.model(data)
N,T,_ = output.shape
out4loss = output.reshape(N*T,-1)
loss = self.loss(out4loss, label.reshape(N*T,))
output = sfm(output)
result_frag.append(output)
label_frag.append(label)
start_frag = start_frag + start_pos.cpu().numpy().tolist()
#print(video_name)
video_name_frag = video_name_frag + list(video_name)
group_list, vid_list_uqk = self.merge_list(video_name_frag)
start_list = np.array(start_frag).astype(int)
prob_val = torch.cat(result_frag,dim=0)
gt_dict = {}
res_dict = {}
num_seq = 200 # the sequence of a clip, defined in the data_preprocess stage, detection/tools/pku_part1_gendata_fordetect.py
for idx, vid_name in zip((group_list), vid_list_uqk):
vid_name = vid_name.split('_')[0]
prob_seq = np.zeros((int(np.max(start_list[idx])) + num_seq, 53))
for idx2 in idx:
#print(prob_seq[start_list[idx2]:start_list[idx2]+num_seq].shape,prob_val[idx2].shape)
prob_seq[start_list[idx2]:start_list[idx2]+num_seq] = \
prob_seq[start_list[idx2]:start_list[idx2]+num_seq] + prob_val[idx2].cpu().numpy()
prob_seq = F.normalize(torch.from_numpy(prob_seq), dim=1, p=1).numpy()
# print vid_name, prob_seq.shape
pred_labels = get_interval_frm_frame_predict(prob_seq)
label_path = '/mnt/netdisk/Datasets/088-PKUMMD/PKUMMDv1/Train_Label_PKU_final/'
labels = np.loadtxt(os.path.join(label_path, vid_name+'.txt'),delimiter=',').astype(int)#T,4 (label,start,end,confidence)
# NOTE: for ground truth, action labels starts from 0, but for our detection, 0 indicates empty action
labels[:,0] = labels[:,0] + 1
gt_dict[vid_name] = labels
res_dict[vid_name] = pred_labels
for thresh in [0.1, 0.3, 0.5]:
metrics = eval_detect_mAP(gt_dict, res_dict, minoverlap=thresh)
print('thresh: ',thresh, metrics['map'])
self.current_result = metrics['map']
self.best_result = max(self.best_result,self.current_result)
print (metrics['map'])
@staticmethod
def get_parser(add_help=False):
# parameter priority: command line > config > default
parent_parser = Processor.get_parser(add_help=False)
parser = argparse.ArgumentParser(
add_help=add_help,
parents=[parent_parser],
description='Spatial Temporal Graph Convolution Network')
# region arguments yapf: disable
# evaluation
parser.add_argument('--show_topk', type=int, default=[1, 5], nargs='+', help='which Top K accuracy will be shown')
parser.add_argument('--warm_up_epoch', type=int, default=0, help='which Top K accuracy will be shown')
# optim
parser.add_argument('--base_lr', type=float, default=0.01, help='initial learning rate')
parser.add_argument('--step', type=int, default=[], nargs='+', help='the epoch where optimizer reduce the learning rate')
parser.add_argument('--optimizer', default='SGD', help='type of optimizer')
parser.add_argument('--nesterov', type=str2bool, default=True, help='use nesterov or not')
parser.add_argument('--weight_decay', type=float, default=0.0001, help='weight decay for optimizer')
# endregion yapf: enable
return parser
|