def __init__(self, opt, train=True):
super(HurricaneVideoDataset, self).__init__(opt, train=train)
self.nc = 3 if self.opt.dataset == "hurricane" else 6
if self.train:
self.image_path = os.path.join('./dataset/Hurricane/', 'train')
else:
self.image_path = os.path.join('./dataset/Hurricane/', 'test')
threshold = self.window_size if opt.irregular else self.sample_size
self.image_list = remove_files_under_sample_size(
image_path=self.image_path, threshold=threshold)
self.image_list = sorted(self.image_list)
vtrans = [vtransforms.Pad(padding=(1, 0), fill=0)]
if self.train:
# vtrans += [vtransforms.RandomHorizontalFlip()]
# vtrans += [vtransforms.RandomRotation()]
pass
vtrans += [vtransforms.ToTensor(scale=False)]
vtrans += [vtransforms.Normalize(0.5, 0.5)] if opt.input_norm else []
self.vtrans = T.Compose(vtrans)
def VideoSpatialPrediction(vid_name,
target,
net,
num_categories,
num_samples=25,
new_size=299,
batch_size=2):
gc = GradCAM(model=net)
clip_mean = [0.5] * num_samples
clip_std = [0.226] * num_samples
normalize = video_transforms.Normalize(mean=clip_mean, std=clip_std)
val_transform = video_transforms.Compose([
video_transforms.ToTensor(),
normalize,
])
deep = 1
# inception = 299,299, resnet = 224,224
dims = (new_size, new_size, deep, num_samples)
rgb = np.zeros(shape=dims, dtype=np.float64)
rgb_flip = np.zeros(shape=dims, dtype=np.float64)
for i in range(num_samples):
img_file = os.path.join(vid_name, 'vr_{0:02d}.png'.format(i))
img = cv2.imread(img_file, cv2.IMREAD_GRAYSCALE)
rgb[:, :, 0, i] = img
rgb_flip[:, :, 0, i] = img[:, ::-1]
_, _, _, c = rgb.shape
rgb_list = []
for c_index in range(c):
cur_img = rgb[:, :, :, c_index]
cur_img_tensor = val_transform(cur_img)
rgb_list.append(np.expand_dims(cur_img_tensor.numpy(), 0))
rgb_np = np.concatenate(rgb_list, axis=0)
prediction = np.zeros((num_categories, rgb.shape[3]))
index = 50
input_data = rgb_np[index:index + 1, :, :, :]
imgDataTensor = torch.from_numpy(input_data).type(torch.FloatTensor).cuda()
imgDataVar = torch.autograd.Variable(imgDataTensor)
probs, ids = gc.forward(imgDataVar)
ids_ = torch.LongTensor([[target]] * len(imgDataVar)).to(
torch.device("cuda"))
gc.backward(ids=ids_)
regions = gc.generate(target_layer="Mixed_7c")
save_gradcam(vid_name.split("/")[-1] + ".png",
gcam=regions[0, 0],
raw_image=rgb[:, :, :, index])
return prediction
def VideoSpatialPrediction(
vid_name,
net,
num_categories,
num_samples=25,
new_size = 299,
batch_size = 2
):
clip_mean = [0.5]*num_samples
clip_std = [0.226]*num_samples
normalize = video_transforms.Normalize(mean=clip_mean,
std=clip_std)
val_transform = video_transforms.Compose([
video_transforms.ToTensor(),
normalize,
])
deep = 1
# inception = 299,299, resnet = 224,224
dims = (new_size,new_size,deep,num_samples)
rgb = np.zeros(shape=dims, dtype=np.float64)
rgb_flip = np.zeros(shape=dims, dtype=np.float64)
for i in range(num_samples):
img_file = os.path.join(vid_name, 'vr_{0:02d}.png'.format(i))
img = cv2.imread(img_file, cv2.IMREAD_GRAYSCALE)
rgb[:,:,0,i] = img
rgb_flip[:,:,0,i] = img[:,::-1]
_, _, _, c = rgb.shape
rgb_list = []
for c_index in range(c):
cur_img = rgb[:,:,:,c_index]
cur_img_tensor = val_transform(cur_img)
rgb_list.append(np.expand_dims(cur_img_tensor.numpy(), 0))
rgb_np = np.concatenate(rgb_list,axis=0)
prediction = np.zeros((num_categories,rgb.shape[3]))
num_batches = int(math.ceil(float(rgb.shape[3])/batch_size))
for bb in range(num_batches):
span = range(batch_size*bb, min(rgb.shape[3],batch_size*(bb+1)))
input_data = rgb_np[span,:,:,:]
imgDataTensor = torch.from_numpy(input_data).type(torch.FloatTensor).cuda()
imgDataVar = torch.autograd.Variable(imgDataTensor)
output = net(imgDataVar)
result = output.data.cpu().numpy()
prediction[:, span] = np.transpose(result)
return prediction
def get_video_transform(data_name, split_name, opt):
normalizer = video_transforms.Normalize(mean=[0.485, 0.456, 0.406],
std=[0.229, 0.224, 0.225])
t_list = []
if split_name == 'train':
t_list = [
video_transforms.RandomResizedCrop(opt.crop_size),
video_transforms.RandomHorizontalFlip(),
video_transforms.ColorJitter(brightness=0.1, contrast=0.1, hue=0.1)
]
else:
t_list = [
video_transforms.Resize(256),
video_transforms.CenterCrop(opt.crop_size)
]
t_end = [video_transforms.ToTensor(), normalizer]
transform = video_transforms.Compose(t_list + t_end)
return transform
def __init__(self, opt, train=True):
super(VideoDataset, self).__init__(opt, train=train)
# Dataroot & Transform
if opt.dataset == 'mgif':
data_root = './dataset/moving-gif'
vtrans = [vtransforms.Scale(size=128)]
elif opt.dataset == 'kth':
data_root = './dataset/kth_action/'
vtrans = [
vtransforms.CenterCrop(size=120),
vtransforms.Scale(size=128)
]
elif opt.dataset == 'penn':
data_root = './dataset/penn_action/'
vtrans = [vtransforms.Scale(size=128)]
if self.train:
vtrans += [vtransforms.RandomHorizontalFlip()]
vtrans += [vtransforms.RandomRotation()]
vtrans += [vtransforms.ToTensor(scale=True)]
vtrans += [vtransforms.Normalize(0.5, 0.5)] if opt.input_norm else []
self.vtrans = T.Compose(vtrans)
if self.train:
self.image_path = os.path.join(data_root, 'train')
else:
self.image_path = os.path.join(data_root, 'test')
threshold = self.window_size if opt.irregular else self.sample_size
if opt.dataset in ['kth', 'sintel', 'ucf101', 'penn']:
self.image_list = os.listdir(self.image_path)
elif opt.dataset in ['mgif', 'stickman']:
self.image_list = remove_files_under_sample_size(
image_path=self.image_path, threshold=threshold)
self.image_list = sorted(self.image_list)
def main():
global args, best_acc1
args = parser.parse_args()
num_classes = args.num_classes
start_epoch=0
writer = SummaryWriter(args.logdir)
model = build_model(num_classes=num_classes, input_length=args.new_length)
print(model)
# create model
print("Building model ... ")
model = torch.nn.DataParallel(model)
model.cuda()
if not os.path.exists(args.out_dir):
os.makedirs(args.out_dir)
print("Saving everything to directory %s." % (args.out_dir))
# define loss function (criterion) and optimizer
criterion = nn.CrossEntropyLoss().cuda()
optimizer = torch.optim.SGD(model.parameters(), args.lr,
momentum=args.momentum,
weight_decay=args.weight_decay)
scheduler = ReduceLROnPlateau(optimizer, verbose=True, patience=4)
# if resume set to True, load the model and continue training
if args.resume or args.evaluate:
if os.path.isfile(args.model_path):
model, optimizer, start_epoch = load_checkpoint(model, optimizer, args.model_path)
cudnn.benchmark = True
is_color = True
# scale_ratios = [1.0, 0.875, 0.75, 0.66]
clip_mean = {'rgb': [0.485, 0.456, 0.406] * args.new_length, 'flow': [0.9432, 0.9359, 0.9511] *args.new_length,
'skeleton': [0.0071, 0.0078, 0.0079]*args.new_length}
clip_std = {'rgb': [0.229, 0.224, 0.225] * args.new_length, 'flow': [0.0788, 0.0753, 0.0683] * args.new_length,
'skeleton': [0.0581, 0.0621, 0.0623] * args.new_length}
normalize = video_transforms.Normalize(mean=clip_mean,
std=clip_std)
train_transform = video_transforms.Compose([
video_transforms.Resize((args.new_width, args.new_height)),
video_transforms.ToTensor(),
normalize,
])
val_transform = video_transforms.Compose([
video_transforms.Resize((args.new_width, args.new_height)),
video_transforms.ToTensor(),
normalize,
])
train_dataset = datasets.__dict__[args.dataset](root=args.data,
source=args.train_split_file,
phase="train",
is_color=is_color,
new_length=args.new_length,
video_transform=train_transform)
val_dataset = datasets.__dict__[args.dataset](root=args.data,
source=args.test_split_file,
phase="val",
is_color=is_color,
new_length=args.new_length,
video_transform=val_transform,
return_id=True)
print('{} samples found, {} train samples and {} test samples.'.format(len(val_dataset)+len(train_dataset),
len(train_dataset),
len(val_dataset)))
train_loader = torch.utils.data.DataLoader(
train_dataset,
batch_size=args.batch_size, shuffle=True,
num_workers=args.workers, pin_memory=True)
val_loader = torch.utils.data.DataLoader(
val_dataset,
batch_size=args.batch_size, shuffle=False,
num_workers=args.workers, pin_memory=True)
if args.evaluate:
validate(val_loader, model, criterion, epoch=0, writer=writer, classes=val_dataset.classes)
return
for epoch in range(start_epoch, args.epochs):
# train for one epoch
train(train_loader, model, criterion, optimizer, epoch, args, writer)
# evaluate on validation set
acc1, loss = validate(val_loader, model, criterion, epoch, writer)
scheduler.step(loss, epoch=epoch)
# remember best acc@1 and save checkpoint
is_best = acc1 > best_acc1
best_acc1 = max(acc1, best_acc1)
save_checkpoint({
'epoch': epoch + 1,
'arch': 'ThreeStreamTemporal',
'state_dict': model.state_dict(),
'best_acc1': best_acc1,
'optimizer': optimizer.state_dict(),
'scheduler': scheduler.state_dict(),
}, is_best, 'last_checkpoint.pth.tar', args.out_dir)
writer.close()
def flow_test(param_model):
model = param_model
video_list = os.listdir(video)
f = open(video + "flow_result.txt", 'w')
for file in video_list:
if file.endswith("mp4"):
f.write(file + "\n")
# file_data = OrderedDict()
frame_count = 0
clip_mean = [0.5, 0.5] * 10
clip_std = [0.226, 0.226] * 10
normalize = video_transforms.Normalize(mean=clip_mean,
std=clip_std)
# config the transform to match the network's format
transform = video_transforms.Compose([
# video_transforms.Scale((256)),
video_transforms.CenterCrop((224)),
video_transforms.ToTensor(),
normalize,
])
# prepare the translation dictionary label-action
data_handler = UCF101_splitter(
os.getcwd() + '/datasets/ucf101_splits/', None)
data_handler.get_action_index()
class_to_idx = data_handler.action_label
idx_to_class = {v: k for k, v in class_to_idx.items()}
# Start looping on frames received from webcam
vs = cv2.VideoCapture(video + file)
softmax = torch.nn.Softmax()
nn_output = torch.FloatTensor(2 * 10, 224, 224)
count = 0
idx = 0
temp = ''
x = []
sampled_list = []
while (vs.isOpened()):
ret, image = vs.read()
if ret is False:
break
else:
image = cv2.resize(image, (342, 256),
interpolation=cv2.INTER_LINEAR)
x.append(temp)
if count == 11:
sampled_list = []
# input_var = torch.autograd.Variable(clip_input, volatile=True)
temp = ''
input_var = clip_input.view(1, 20, 224, 224).cuda()
output = model(input_var)
output = softmax(output)
output = output.data.cpu().numpy()
preds = output.argsort()[0][-5:][::-1]
pred_classes = [(idx_to_class[str(pred + 1)],
output[0, pred]) for pred in preds]
value = 0
for i in range(5):
if pred_classes[i][0] == label:
value = pred_classes[i][1]
temp += '{} - {:.2f}\n'.format(
pred_classes[i][0], pred_classes[i][1])
f.write(str(value) + "\n")
nn_output = torch.FloatTensor(2 * 10, 224, 224)
count = 1
if count == 0:
old_frame = image.copy()
prev = cv2.cvtColor(old_frame, cv2.COLOR_BGR2GRAY)
else:
next = cv2.cvtColor(image, cv2.COLOR_BGR2GRAY)
flow = cv2.calcOpticalFlowFarneback(
prev, next, 1, 0.5, 3, 15, 3, 5, 1.2, 0)
horz = cv2.normalize(flow[..., 0], None, 0, 255,
cv2.NORM_MINMAX)
vert = cv2.normalize(flow[..., 1], None, 0, 255,
cv2.NORM_MINMAX)
horz = horz.astype('uint8')
vert = vert.astype('uint8')
imgH = Image.fromarray(horz)
imgV = Image.fromarray(vert)
sampled_list.append(np.expand_dims(imgH, 2))
sampled_list.append(np.expand_dims(imgV, 2))
clip_input = np.concatenate(sampled_list, axis=2)
clip_input = transform(clip_input)
clip_input = clip_input.float().cuda(async=True)
imgH.close()
imgV.close()
prev = next.copy()
count += 1
idx += 1
f.write("----\n")
# file_data[file] = flow_value
# with open('flow.json', 'w', encoding="utf-8") as make_file:
# json.dump(file_data, make_file, ensure_ascii=False, indent="\t")
print(idx)
vs.release()
f.close()
def VideoSpatialPrediction(mode,
vid_name,
net,
num_categories,
start_frame=0,
num_frames=0,
num_samples=25,
index=1,
new_size=299):
if num_frames == 0:
imglist = os.listdir(vid_name)
#imglist = list(filter(lambda x: x[:3]=='img',imglist))
duration = len(imglist)
# print(duration)
else:
duration = num_frames
# selection
if mode == 'rgb':
step = int(math.floor((duration - 1) / (num_samples - 1)))
clip_mean = [0.485, 0.456, 0.406]
clip_std = [0.229, 0.224, 0.225]
else:
clip_mean = [0.5, 0.5]
clip_std = [0.226, 0.226]
normalize = video_transforms.Normalize(mean=clip_mean, std=clip_std)
val_transform = video_transforms.Compose([
video_transforms.ToTensor(),
normalize,
])
deep = 1 if mode == 'rhythm' else 3
# inception = 320,360, resnet = 240, 320
width = 320 if new_size == 299 else 240
height = 360 if new_size == 299 else 320
dims = (width, height, deep, num_samples)
rgb = np.zeros(shape=dims, dtype=np.float64)
rgb_flip = np.zeros(shape=dims, dtype=np.float64)
for i in range(num_samples):
if mode == 'rhythm':
img_file = os.path.join(vid_name,
'visual_rhythm_{0:05d}.png'.format(index))
print(img_file)
img = cv2.imread(img_file, cv2.IMREAD_GRAYSCALE)
img = cv2.resize(img, dims[1::-1])
rgb[:, :, 0, i] = img
rgb_flip[:, :, 0, i] = img[:, ::-1]
else:
img_file = os.path.join(vid_name,
'img_{0:05d}.jpg'.format(i * step + 1))
img = cv2.imread(img_file, cv2.IMREAD_UNCHANGED)
img = cv2.resize(img, dims[1::-1])
img = cv2.cvtColor(img, cv2.COLOR_BGR2RGB)
rgb[:, :, :, i] = img
rgb_flip[:, :, :, i] = img[:, ::-1, :]
# crop 299 = inception, 224 = resnet
size = new_size
corner = [(height - size) // 2, (width - size) // 2]
rgb_1 = rgb[:size, :size, :, :]
rgb_2 = rgb[:size, -size:, :, :]
rgb_3 = rgb[corner[1]:corner[1] + size, corner[0]:corner[0] + size, :, :]
rgb_4 = rgb[-size:, :size, :, :]
rgb_5 = rgb[-size:, -size:, :, :]
rgb_f_1 = rgb_flip[:size, :size, :, :]
rgb_f_2 = rgb_flip[:size, -size:, :, :]
rgb_f_3 = rgb_flip[corner[1]:corner[1] + size,
corner[0]:corner[0] + size, :, :]
rgb_f_4 = rgb_flip[-size:, :size, :, :]
rgb_f_5 = rgb_flip[-size:, -size:, :, :]
rgb = np.concatenate((rgb_1, rgb_2, rgb_3, rgb_4, rgb_5, rgb_f_1, rgb_f_2,
rgb_f_3, rgb_f_4, rgb_f_5),
axis=3)
_, _, _, c = rgb.shape
rgb_list = []
for c_index in range(c):
cur_img = rgb[:, :, :, c_index]
cur_img_tensor = val_transform(cur_img)
rgb_list.append(np.expand_dims(cur_img_tensor.numpy(), 0))
rgb_np = np.concatenate(rgb_list, axis=0)
#batch_size = 25
batch_size = 5
prediction = np.zeros((num_categories, rgb.shape[3]))
num_batches = int(math.ceil(float(rgb.shape[3]) / batch_size))
for bb in range(num_batches):
span = range(batch_size * bb, min(rgb.shape[3], batch_size * (bb + 1)))
input_data = rgb_np[span, :, :, :]
imgDataTensor = torch.from_numpy(input_data).type(
torch.FloatTensor).cuda()
imgDataVar = torch.autograd.Variable(imgDataTensor)
output = net(imgDataVar)
result = output.data.cpu().numpy()
prediction[:, span] = np.transpose(result)
return prediction
def VideoSpatialPrediction3D(vid_name,
net,
num_categories,
architecture_name,
start_frame=0,
num_frames=0,
length=16,
extension='img_{0:05d}.jpg',
ten_crop=False):
if num_frames == 0:
imglist = os.listdir(vid_name)
newImageList = []
if 'rgb' in architecture_name or 'pose' in architecture_name:
for item in imglist:
if 'img' in item:
newImageList.append(item)
elif 'flow' in architecture_name:
for item in imglist:
if 'flow_x' in item:
newImageList.append(item)
duration = len(newImageList)
else:
duration = num_frames
if 'rgb' in architecture_name or 'pose' in architecture_name:
if 'I3D' in architecture_name:
if not 'resnet' in architecture_name:
clip_mean = [0.5, 0.5, 0.5]
clip_std = [0.5, 0.5, 0.5]
else:
clip_mean = [0.45, 0.45, 0.45]
clip_std = [0.225, 0.225, 0.225]
normalize = video_transforms.Normalize(mean=clip_mean,
std=clip_std)
val_transform = video_transforms.Compose([
video_transforms.ToTensor(),
normalize,
])
if '112' in architecture_name:
scale = 0.5
else:
scale = 1
elif 'MFNET3D' in architecture_name:
clip_mean = [0.48627451, 0.45882353, 0.40784314]
clip_std = [0.234, 0.234, 0.234]
normalize = video_transforms.Normalize(mean=clip_mean,
std=clip_std)
val_transform = video_transforms.Compose(
[video_transforms.ToTensor(), normalize])
if '112' in architecture_name:
scale = 0.5
else:
scale = 1
elif 'tsm' in architecture_name:
clip_mean = [0.485, 0.456, 0.406]
clip_std = [0.229, 0.224, 0.225]
normalize = video_transforms.Normalize(mean=clip_mean,
std=clip_std)
val_transform = video_transforms.Compose(
[video_transforms.ToTensor(), normalize])
scale = 1
elif "r2plus1d" in architecture_name:
clip_mean = [0.43216, 0.394666, 0.37645]
clip_std = [0.22803, 0.22145, 0.216989]
normalize = video_transforms.Normalize(mean=clip_mean,
std=clip_std)
val_transform = video_transforms.Compose(
[video_transforms.ToTensor(), normalize])
scale = 0.5
elif 'rep_flow' in architecture_name:
clip_mean = [0.5, 0.5, 0.5]
clip_std = [0.5, 0.5, 0.5]
normalize = video_transforms.Normalize(mean=clip_mean,
std=clip_std)
val_transform = video_transforms.Compose([
video_transforms.ToTensor(),
normalize,
])
scale = 1
elif "slowfast" in architecture_name:
clip_mean = [0.45, 0.45, 0.45]
clip_std = [0.225, 0.225, 0.225]
normalize = video_transforms.Normalize(mean=clip_mean,
std=clip_std)
val_transform = video_transforms.Compose([
video_transforms.ToTensor(),
normalize,
])
scale = 1
else:
scale = 0.5
clip_mean = [114.7748, 107.7354, 99.4750]
clip_std = [1, 1, 1]
normalize = video_transforms.Normalize(mean=clip_mean,
std=clip_std)
val_transform = video_transforms.Compose([
video_transforms.ToTensor2(),
normalize,
])
elif 'flow' in architecture_name:
if 'I3D' in architecture_name:
clip_mean = [0.5] * 2
clip_std = [0.5] * 2
normalize = video_transforms.Normalize(mean=clip_mean,
std=clip_std)
val_transform = video_transforms.Compose([
video_transforms.ToTensor(),
normalize,
])
scale = 1
elif "3D" in architecture_name:
scale = 0.5
clip_mean = [127.5, 127.5]
clip_std = [1, 1]
normalize = video_transforms.Normalize(mean=clip_mean,
std=clip_std)
val_transform = video_transforms.Compose([
video_transforms.ToTensor2(),
normalize,
])
elif "r2plus1d" in architecture_name:
clip_mean = [0.5] * 2
clip_std = [0.226] * 2
normalize = video_transforms.Normalize(mean=clip_mean,
std=clip_std)
val_transform = video_transforms.Compose([
video_transforms.ToTensor(),
normalize,
])
scale = 0.5
if '224' in architecture_name:
scale = 1
if '112' in architecture_name:
scale = 0.5
# selection
#step = int(math.floor((duration-1)/(num_samples-1)))
dims2 = (224, 224, 3, duration)
imageSize = int(224 * scale)
dims = (int(256 * scale), int(340 * scale), 3, duration)
#dims = (int(256 * scale),int(256 * scale),3,duration)
duration = duration - 1
offsets = []
offsetMainIndexes = list(range(1, duration - length, length))
if len(offsetMainIndexes) == 0:
offsets = list(range(1, duration + 2)) * int(
np.floor(length / (duration + 1))) + list(
range(1, length % (duration + 1) + 1))
else:
shift = int((duration - (offsetMainIndexes[-1] + length)) / 2)
for mainOffsetValue in offsetMainIndexes:
for lengthID in range(1, length + 1):
offsets.append(lengthID + mainOffsetValue + shift)
# offsetMainIndexes = list(range(0,duration,length))
# for mainOffsetValue in offsetMainIndexes:
# for lengthID in range(1, length+1):
# loaded_frame_index = lengthID + mainOffsetValue
# moded_loaded_frame_index = loaded_frame_index % (duration + 1)
# if moded_loaded_frame_index == 0:
# moded_loaded_frame_index = (duration + 1)
# offsets.append(moded_loaded_frame_index)
imageList = []
imageList1 = []
imageList2 = []
imageList3 = []
imageList4 = []
imageList5 = []
imageList6 = []
imageList7 = []
imageList8 = []
imageList9 = []
imageList10 = []
imageList11 = []
imageList12 = []
interpolation = cv2.INTER_LINEAR
for index in offsets:
if 'rgb' in architecture_name or 'pose' in architecture_name:
img_file = os.path.join(vid_name, extension.format(index))
img = cv2.imread(img_file, cv2.IMREAD_UNCHANGED)
img = cv2.resize(img, dims[1::-1], interpolation)
#img2 = cv2.resize(img, dims2[1::-1],interpolation)
img = cv2.cvtColor(img, cv2.COLOR_BGR2RGB)
img_flip = img[:, ::-1, :].copy()
elif 'flow' in architecture_name:
flow_x_file = os.path.join(vid_name, extension.format('x', index))
flow_y_file = os.path.join(vid_name, extension.format('y', index))
img_x = cv2.imread(flow_x_file, cv2.IMREAD_GRAYSCALE)
img_y = cv2.imread(flow_y_file, cv2.IMREAD_GRAYSCALE)
img_x = np.expand_dims(img_x, -1)
img_y = np.expand_dims(img_y, -1)
img = np.concatenate((img_x, img_y), 2)
img = cv2.resize(img, dims[1::-1], interpolation)
img_flip = img[:, ::-1, :].copy()
#img_flip2 = img2[:,::-1,:].copy()
#imageList1.append(img[int(16 * scale):int(16 * scale + imageSize), int(16 * scale) : int(16 * scale + imageSize), :])
imageList1.append(img[int(16 * scale):int(16 * scale + imageSize),
int(58 * scale):int(58 * scale + imageSize), :])
imageList2.append(img[:imageSize, :imageSize, :])
imageList3.append(img[:imageSize, -imageSize:, :])
imageList4.append(img[-imageSize:, :imageSize, :])
imageList5.append(img[-imageSize:, -imageSize:, :])
imageList6.append(img_flip[int(16 * scale):int(16 * scale + imageSize),
int(58 * scale):int(58 * scale +
imageSize), :])
imageList7.append(img_flip[:imageSize, :imageSize, :])
imageList8.append(img_flip[:imageSize, -imageSize:, :])
imageList9.append(img_flip[-imageSize:, :imageSize, :])
imageList10.append(img_flip[-imageSize:, -imageSize:, :])
# imageList11.append(img2)
# imageList12.append(img_flip2)
if ten_crop:
imageList = imageList1 + imageList2 + imageList3 + imageList4 + imageList5 + imageList6 + imageList7 + imageList8 + imageList9 + imageList10
else:
imageList = imageList1
#imageList=imageList11+imageList12
rgb_list = []
for i in range(len(imageList)):
cur_img = imageList[i]
cur_img_tensor = val_transform(cur_img)
rgb_list.append(np.expand_dims(cur_img_tensor.numpy(), 0))
input_data = np.concatenate(rgb_list, axis=0)
if 'rgb' in architecture_name or 'pose' in architecture_name:
input_data = input_data.reshape(-1, length, 3, imageSize, imageSize)
elif 'flow' in architecture_name:
input_data = input_data.reshape(-1, length, 2, imageSize, imageSize)
batch_size = 10
result = np.zeros([input_data.shape[0], num_categories])
num_batches = int(math.ceil(float(input_data.shape[0]) / batch_size))
with torch.no_grad():
for bb in range(num_batches):
span = range(batch_size * bb,
min(input_data.shape[0], batch_size * (bb + 1)))
input_data_batched = input_data[span, :, :, :, :]
imgDataTensor = torch.from_numpy(input_data_batched).type(
torch.FloatTensor).cuda()
if 'rgb' in architecture_name or 'pose' in architecture_name:
imgDataTensor = imgDataTensor.view(-1, length, 3, imageSize,
imageSize).transpose(1, 2)
elif 'flow' in architecture_name:
imgDataTensor = imgDataTensor.view(-1, length, 2, imageSize,
imageSize).transpose(1, 2)
if 'bert' in architecture_name or 'pooling' in architecture_name or 'NLB' in architecture_name \
or 'lstm' in architecture_name or 'adamw' in architecture_name:
output, input_vectors, sequenceOut, maskSample = net(
imgDataTensor)
else:
output = net(imgDataTensor)
#span = range(sample_size*bb, min(int(input_data.shape[0]/length),sample_size*(bb+1)))
result[span, :] = output.data.cpu().numpy()
mean_result = np.mean(result, 0)
prediction = np.argmax(mean_result)
top3 = mean_result.argsort()[::-1][:3]
top5 = mean_result.argsort()[::-1][:5]
return prediction, mean_result, top3
def VideoSpatialPrediction(vid_name,
net,
num_categories,
start_frame=0,
num_frames=0,
num_samples=25):
if num_frames == 0:
imglist = os.listdir(vid_name)
duration = len(imglist)
# print(duration)
else:
duration = num_frames
clip_mean = [0.485, 0.456, 0.406]
clip_std = [0.229, 0.224, 0.225]
normalize = video_transforms.Normalize(mean=clip_mean, std=clip_std)
val_transform = video_transforms.Compose([
video_transforms.ToTensor(),
normalize,
])
# selection
step = int(math.floor((duration - 1) / (num_samples - 1)))
dims = (256, 340, 3, num_samples)
rgb = np.zeros(shape=dims, dtype=np.float64)
rgb_flip = np.zeros(shape=dims, dtype=np.float64)
for i in range(num_samples):
img_file = os.path.join(vid_name,
'frame{0:06d}.jpg'.format(i * step + 1))
img = cv2.imread(img_file, cv2.IMREAD_UNCHANGED)
img = cv2.resize(img, dims[1::-1])
img = cv2.cvtColor(img, cv2.COLOR_BGR2RGB)
rgb[:, :, :, i] = img
rgb_flip[:, :, :, i] = img[:, ::-1, :]
# crop
rgb_1 = rgb[:224, :224, :, :]
rgb_2 = rgb[:224, -224:, :, :]
rgb_3 = rgb[16:240, 60:284, :, :]
rgb_4 = rgb[-224:, :224, :, :]
rgb_5 = rgb[-224:, -224:, :, :]
rgb_f_1 = rgb_flip[:224, :224, :, :]
rgb_f_2 = rgb_flip[:224, -224:, :, :]
rgb_f_3 = rgb_flip[16:240, 60:284, :, :]
rgb_f_4 = rgb_flip[-224:, :224, :, :]
rgb_f_5 = rgb_flip[-224:, -224:, :, :]
rgb = np.concatenate((rgb_1, rgb_2, rgb_3, rgb_4, rgb_5, rgb_f_1, rgb_f_2,
rgb_f_3, rgb_f_4, rgb_f_5),
axis=3)
_, _, _, c = rgb.shape
rgb_list = []
for c_index in range(c):
cur_img = rgb[:, :, :, c_index].squeeze()
cur_img_tensor = val_transform(cur_img)
rgb_list.append(np.expand_dims(cur_img_tensor.numpy(), 0))
rgb_np = np.concatenate(rgb_list, axis=0)
# print(rgb_np.shape)
batch_size = 25
prediction = np.zeros((num_categories, rgb.shape[3]))
num_batches = int(math.ceil(float(rgb.shape[3]) / batch_size))
for bb in range(num_batches):
span = range(batch_size * bb, min(rgb.shape[3], batch_size * (bb + 1)))
input_data = rgb_np[span, :, :, :]
imgDataTensor = torch.from_numpy(input_data).type(
torch.FloatTensor).cuda()
imgDataVar = torch.autograd.Variable(imgDataTensor)
output = net(imgDataVar)
result = output.data.cpu().numpy()
prediction[:, span] = np.transpose(result)
return prediction
def VideoTemporalPrediction(mode,
vid_name,
target,
net,
num_categories,
start_frame=0,
num_frames=0,
num_samples=25,
optical_flow_frames=10,
new_size=299,
ext=".jpg"):
gc = GradCAM(model=net)
if num_frames == 0:
imglist = os.listdir(vid_name)
duration = len(imglist)
else:
duration = num_frames
# selection
step = int(math.floor((duration - optical_flow_frames + 1) / num_samples))
clip_mean = [0.5] * optical_flow_frames * 2
clip_std = [0.226] * optical_flow_frames * 2
normalize = video_transforms.Normalize(mean=clip_mean, std=clip_std)
test_transform = video_transforms.Compose(
[video_transforms.ToTensor(), normalize])
# inception = 320,360, resnet = 240, 320
width = 320 if new_size == 299 else 240
height = 360 if new_size == 299 else 320
deep = optical_flow_frames * 2
dims = (width, height, deep, num_samples)
flow = np.zeros(shape=dims, dtype=np.float64)
flow_flip = np.zeros(shape=dims, dtype=np.float64)
for i in range(num_samples):
for j in range(optical_flow_frames):
flow_x_file = os.path.join(
vid_name, mode +
'_x_{0:05d}{1}'.format(i * step + j + 1 + start_frame, ext))
flow_y_file = os.path.join(
vid_name, mode +
'_y_{0:05d}{1}'.format(i * step + j + 1 + start_frame, ext))
img_x = cv2.imread(flow_x_file, cv2.IMREAD_GRAYSCALE)
img_y = cv2.imread(flow_y_file, cv2.IMREAD_GRAYSCALE)
img_x = cv2.resize(img_x, dims[1::-1])
img_y = cv2.resize(img_y, dims[1::-1])
flow[:, :, j * 2, i] = img_x
flow[:, :, j * 2 + 1, i] = img_y
flow_flip[:, :, j * 2, i] = 255 - img_x[:, ::-1]
flow_flip[:, :, j * 2 + 1, i] = img_y[:, ::-1]
# crop 299 = inception, 224 = resnet
size = new_size
corner = [(height - size) // 2, (width - size) // 2]
flow_1 = flow[:size, :size, :, :]
flow_2 = flow[:size, -size:, :, :]
flow_3 = flow[corner[1]:corner[1] + size, corner[0]:corner[0] + size, :, :]
flow_4 = flow[-size:, :size, :, :]
flow_5 = flow[-size:, -size:, :, :]
flow_f_1 = flow_flip[:size, :size, :, :]
flow_f_2 = flow_flip[:size, -size:, :, :]
flow_f_3 = flow_flip[corner[1]:corner[1] + size,
corner[0]:corner[0] + size, :, :]
flow_f_4 = flow_flip[-size:, :size, :, :]
flow_f_5 = flow_flip[-size:, -size:, :, :]
flow = np.concatenate((flow_1, flow_2, flow_3, flow_4, flow_5, flow_f_1,
flow_f_2, flow_f_3, flow_f_4, flow_f_5),
axis=3)
_, _, _, c = flow.shape
flow_list = []
for c_index in range(c):
cur_img = flow[:, :, :, c_index].squeeze()
cur_img_tensor = test_transform(cur_img)
flow_list.append(np.expand_dims(cur_img_tensor.numpy(), 0))
flow_np = np.concatenate(flow_list, axis=0)
prediction = np.zeros((num_categories, flow.shape[3]))
index = 50
input_data = flow_np[index:index + 1, :, :, :]
raw_image_x = flow[:, :, [0, 2, 4, 6, 8], index]
raw_image_y = flow[:, :, [1, 3, 5, 7, 9], index]
print(raw_image_x.shape, raw_image_y.shape)
imgDataTensor = torch.from_numpy(input_data).type(torch.FloatTensor).cuda()
imgDataVar = torch.autograd.Variable(imgDataTensor)
probs, ids = gc.forward(imgDataVar)
ids_ = torch.LongTensor([[target]] * len(imgDataVar)).to(
torch.device("cuda"))
gc.backward(ids=ids_)
regions = gc.generate(target_layer="Mixed_7c")
save_gradcam(vid_name.split("/")[-1] + "_x.png",
gcam=regions[0, 0],
raw_image=flow[:, :, 4:5, index])
save_gradcam(vid_name.split("/")[-1] + "_y.png",
gcam=regions[0, 0],
raw_image=flow[:, :, 5:6, index])
return prediction
def main():
global args, best_prec1, model, writer, best_loss, length, width, height, input_size, scheduler
args = parser.parse_args()
training_continue = args.contine
if '3D' in args.arch:
if 'I3D' in args.arch or 'MFNET3D' in args.arch:
if '112' in args.arch:
scale = 0.5
else:
scale = 1
else:
if '224' in args.arch:
scale = 1
else:
scale = 0.5
elif 'r2plus1d' in args.arch:
scale = 0.5
else:
scale = 1
print('scale: %.1f' % (scale))
input_size = int(224 * scale)
width = int(340 * scale)
height = int(256 * scale)
saveLocation = "./checkpoint/" + args.dataset + "_" + args.arch + "_split" + str(
args.split)
if not os.path.exists(saveLocation):
os.makedirs(saveLocation)
writer = SummaryWriter(saveLocation)
# create model
if args.evaluate:
print("Building validation model ... ")
model = build_model_validate()
optimizer = AdamW(model.parameters(),
lr=args.lr,
weight_decay=args.weight_decay)
elif training_continue:
model, startEpoch, optimizer, best_prec1 = build_model_continue()
for param_group in optimizer.param_groups:
lr = param_group['lr']
#param_group['lr'] = lr
print(
"Continuing with best precision: %.3f and start epoch %d and lr: %f"
% (best_prec1, startEpoch, lr))
else:
print("Building model with ADAMW... ")
model = build_model()
optimizer = AdamW(model.parameters(),
lr=args.lr,
weight_decay=args.weight_decay)
startEpoch = 0
if HALF:
model.half() # convert to half precision
for layer in model.modules():
if isinstance(layer, nn.BatchNorm2d):
layer.float()
print("Model %s is loaded. " % (args.arch))
# define loss function (criterion) and optimizer
criterion = nn.CrossEntropyLoss().cuda()
criterion2 = nn.MSELoss().cuda()
scheduler = lr_scheduler.ReduceLROnPlateau(optimizer,
'min',
patience=5,
verbose=True)
print("Saving everything to directory %s." % (saveLocation))
if args.dataset == 'ucf101':
dataset = './datasets/ucf101_frames'
elif args.dataset == 'hmdb51':
dataset = './datasets/hmdb51_frames'
elif args.dataset == 'smtV2':
dataset = './datasets/smtV2_frames'
elif args.dataset == 'window':
dataset = './datasets/window_frames'
elif args.dataset == 'haa500_basketball':
dataset = './datasets/haa500_basketball_frames'
else:
print("No convenient dataset entered, exiting....")
return 0
cudnn.benchmark = True
modality = args.arch.split('_')[0]
if "3D" in args.arch or 'tsm' in args.arch or 'slowfast' in args.arch or 'r2plus1d' in args.arch:
if '64f' in args.arch:
length = 64
elif '32f' in args.arch:
length = 32
else:
length = 16
else:
length = 1
# Data transforming
if modality == "rgb" or modality == "pose":
is_color = True
scale_ratios = [1.0, 0.875, 0.75, 0.66]
if 'I3D' in args.arch:
if 'resnet' in args.arch:
clip_mean = [0.45, 0.45, 0.45] * args.num_seg * length
clip_std = [0.225, 0.225, 0.225] * args.num_seg * length
else:
clip_mean = [0.5, 0.5, 0.5] * args.num_seg * length
clip_std = [0.5, 0.5, 0.5] * args.num_seg * length
#clip_std = [0.25, 0.25, 0.25] * args.num_seg * length
elif 'MFNET3D' in args.arch:
clip_mean = [0.48627451, 0.45882353, 0.40784314
] * args.num_seg * length
clip_std = [0.234, 0.234, 0.234] * args.num_seg * length
elif "3D" in args.arch:
clip_mean = [114.7748, 107.7354, 99.4750] * args.num_seg * length
clip_std = [1, 1, 1] * args.num_seg * length
elif "r2plus1d" in args.arch:
clip_mean = [0.43216, 0.394666, 0.37645] * args.num_seg * length
clip_std = [0.22803, 0.22145, 0.216989] * args.num_seg * length
elif "rep_flow" in args.arch:
clip_mean = [0.5, 0.5, 0.5] * args.num_seg * length
clip_std = [0.5, 0.5, 0.5] * args.num_seg * length
elif "slowfast" in args.arch:
clip_mean = [0.45, 0.45, 0.45] * args.num_seg * length
clip_std = [0.225, 0.225, 0.225] * args.num_seg * length
else:
clip_mean = [0.485, 0.456, 0.406] * args.num_seg * length
clip_std = [0.229, 0.224, 0.225] * args.num_seg * length
elif modality == "pose":
is_color = True
scale_ratios = [1.0, 0.875, 0.75, 0.66]
clip_mean = [0.485, 0.456, 0.406] * args.num_seg
clip_std = [0.229, 0.224, 0.225] * args.num_seg
elif modality == "flow":
is_color = False
scale_ratios = [1.0, 0.875, 0.75, 0.66]
if 'I3D' in args.arch:
clip_mean = [0.5, 0.5] * args.num_seg * length
clip_std = [0.5, 0.5] * args.num_seg * length
elif "3D" in args.arch:
clip_mean = [127.5, 127.5] * args.num_seg * length
clip_std = [1, 1] * args.num_seg * length
else:
clip_mean = [0.5, 0.5] * args.num_seg * length
clip_std = [0.226, 0.226] * args.num_seg * length
elif modality == "both":
is_color = True
scale_ratios = [1.0, 0.875, 0.75, 0.66]
clip_mean = [0.485, 0.456, 0.406, 0.5, 0.5] * args.num_seg * length
clip_std = [0.229, 0.224, 0.225, 0.226, 0.226] * args.num_seg * length
else:
print("No such modality. Only rgb and flow supported.")
normalize = video_transforms.Normalize(mean=clip_mean, std=clip_std)
if "3D" in args.arch and not ('I3D' in args.arch):
train_transform = video_transforms.Compose([
video_transforms.MultiScaleCrop((input_size, input_size),
scale_ratios),
video_transforms.RandomHorizontalFlip(),
video_transforms.ToTensor2(),
normalize,
])
val_transform = video_transforms.Compose([
video_transforms.CenterCrop((input_size)),
video_transforms.ToTensor2(),
normalize,
])
else:
train_transform = video_transforms.Compose([
video_transforms.MultiScaleCrop((input_size, input_size),
scale_ratios),
video_transforms.RandomHorizontalFlip(),
video_transforms.ToTensor(),
normalize,
])
val_transform = video_transforms.Compose([
video_transforms.CenterCrop((input_size)),
video_transforms.ToTensor(),
normalize,
])
# data loading
train_setting_file = "train_%s_split%d.txt" % (modality, args.split)
train_split_file = os.path.join(args.settings, args.dataset,
train_setting_file)
val_setting_file = "val_%s_split%d.txt" % (modality, args.split)
val_split_file = os.path.join(args.settings, args.dataset,
val_setting_file)
if not os.path.exists(train_split_file) or not os.path.exists(
val_split_file):
print(
"No split file exists in %s directory. Preprocess the dataset first"
% (args.settings))
train_dataset = datasets.__dict__[args.dataset](
root=dataset,
source=train_split_file,
phase="train",
modality=modality,
is_color=is_color,
new_length=length,
new_width=width,
new_height=height,
video_transform=train_transform,
num_segments=args.num_seg)
val_dataset = datasets.__dict__[args.dataset](
root=dataset,
source=val_split_file,
phase="val",
modality=modality,
is_color=is_color,
new_length=length,
new_width=width,
new_height=height,
video_transform=val_transform,
num_segments=args.num_seg)
print('{} samples found, {} train samples and {} test samples.'.format(
len(val_dataset) + len(train_dataset), len(train_dataset),
len(val_dataset)))
train_loader = torch.utils.data.DataLoader(train_dataset,
batch_size=args.batch_size,
shuffle=True,
num_workers=args.workers,
pin_memory=True)
val_loader = torch.utils.data.DataLoader(val_dataset,
batch_size=args.batch_size,
shuffle=False,
num_workers=args.workers,
pin_memory=True)
if args.evaluate:
prec1, prec3, _ = validate(val_loader, model, criterion, criterion2,
modality)
return
for epoch in range(startEpoch, args.epochs):
# if learning_rate_index > max_learning_rate_decay_count:
# break
# adjust_learning_rate(optimizer, epoch)
train(train_loader, model, criterion, criterion2, optimizer, epoch,
modality)
# evaluate on validation set
prec1 = 0.0
lossClassification = 0
if (epoch + 1) % args.save_freq == 0:
prec1, prec3, lossClassification = validate(
val_loader, model, criterion, criterion2, modality)
writer.add_scalar('data/top1_validation', prec1, epoch)
writer.add_scalar('data/top3_validation', prec3, epoch)
writer.add_scalar('data/classification_loss_validation',
lossClassification, epoch)
scheduler.step(lossClassification)
# remember best prec@1 and save checkpoint
is_best = prec1 >= best_prec1
best_prec1 = max(prec1, best_prec1)
# best_in_existing_learning_rate = max(prec1, best_in_existing_learning_rate)
#
# if best_in_existing_learning_rate > prec1 + 1:
# learning_rate_index = learning_rate_index
# best_in_existing_learning_rate = 0
if (epoch + 1) % args.save_freq == 0:
checkpoint_name = "%03d_%s" % (epoch + 1, "checkpoint.pth.tar")
if is_best:
print("Model works well")
save_checkpoint(
{
'epoch': epoch + 1,
'arch': args.arch,
'state_dict': model.state_dict(),
'best_prec1': best_prec1,
'best_loss': best_loss,
'optimizer': optimizer.state_dict(),
}, is_best, checkpoint_name, saveLocation)
checkpoint_name = "%03d_%s" % (epoch + 1, "checkpoint.pth.tar")
save_checkpoint(
{
'epoch': epoch + 1,
'arch': args.arch,
'state_dict': model.state_dict(),
'best_prec1': best_prec1,
'best_loss': best_loss,
'optimizer': optimizer.state_dict(),
}, is_best, checkpoint_name, saveLocation)
writer.export_scalars_to_json("./all_scalars.json")
writer.close()
def main():
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
args = parser.parse_args()
set_logger(log_file=args.log_file, debug_mode=args.debug_mode)
torch.manual_seed(args.random_seed)
torch.cuda.manual_seed(args.random_seed)
cudnn.benchmark = True
mean = [124 / 255, 117 / 255, 104 / 255]
std = [1 / (.0167 * 255)] * 3
normalize = transforms.Normalize(mean=mean, std=std)
train_loader = VideoIterTrain(
dataset_path=args.dataset_path,
annotation_path=args.annotation_path,
clip_length=args.clip_length,
frame_interval=args.train_frame_interval,
video_transform=transforms.Compose([
transforms.Resize((256, 256)),
transforms.RandomCrop((224, 224)),
transforms.ToTensor(),
normalize,
]),
name='train',
return_item_subpath=False,
)
train_iter = torch.utils.data.DataLoader(
train_loader,
batch_size=args.batch_size,
shuffle=False,
num_workers=32, # 4, # change this part accordingly
pin_memory=True)
val_loader = VideoIterTrain(
dataset_path=args.dataset_path,
annotation_path=args.annotation_path_test,
clip_length=args.clip_length,
frame_interval=args.val_frame_interval,
video_transform=transforms.Compose([
transforms.Resize((256, 256)),
transforms.RandomCrop((224, 224)),
transforms.ToTensor(),
normalize,
]),
name='val',
return_item_subpath=False,
)
val_iter = torch.utils.data.DataLoader(
val_loader,
batch_size=args.batch_size,
shuffle=False,
num_workers=32, # 4, # change this part accordingly
pin_memory=True)
network = C3D(pretrained=args.pretrained_3d)
network.to(device)
if not path.exists(features_dir):
mkdir(features_dir)
features_writer = FeaturesWriter()
for i_batch, (data, target, sampled_idx, dirs,
vid_names) in tqdm(enumerate(train_iter)):
data = data.to(device)
with torch.no_grad():
input_var = torch.autograd.Variable(data)
outputs = network(input_var)
for i, (dir, vid_name, start_frame) in enumerate(
zip(dirs, vid_names,
sampled_idx.cpu().numpy())):
dir = path.join(features_dir, dir)
features_writer.write(feature=outputs[i],
video_name=vid_name,
start_frame=start_frame,
dir=dir)
features_writer.dump()
features_writer = FeaturesWriter()
for i_batch, (data, target, sampled_idx, dirs,
vid_names) in tqdm(enumerate(val_iter)):
data = data.to(device)
with torch.no_grad():
input_var = torch.autograd.Variable(data)
outputs = network(input_var)
for i, (dir, vid_name, start_frame) in enumerate(
zip(dirs, vid_names,
sampled_idx.cpu().numpy())):
dir = path.join(features_dir, dir)
features_writer.write(feature=outputs[i],
video_name=vid_name,
start_frame=start_frame,
dir=dir)
features_writer.dump()
def rgb_test(param_model):
model = param_model
f = open(video + "rgb_result.txt", 'w')
video_list = os.listdir(video)
for file in video_list:
if file.endswith("mp4"):
f.write(file + "\n")
frame_count = 2
clip_mean = [0.485, 0.456, 0.406] * 1
clip_std = [0.229, 0.224, 0.225] * 1
normalize = video_transforms.Normalize(mean=clip_mean,
std=clip_std)
# config the transform to match the network's format
transform = video_transforms.Compose([
# video_transforms.Scale((256)),
video_transforms.CenterCrop((224)),
video_transforms.ToTensor(),
normalize,
])
# prepare the translation dictionary label-action
data_handler = UCF101_splitter(
os.getcwd() + '/datasets/ucf101_splits/', None)
data_handler.get_action_index()
class_to_idx = data_handler.action_label
idx_to_class = {v: k for k, v in class_to_idx.items()}
# Start looping on frames received from webcam
vs = cv2.VideoCapture(video + file)
softmax = torch.nn.Softmax()
nn_output = torch.tensor(np.zeros((1, 23)),
dtype=torch.float32).cuda()
sampled_list = []
first_count = 0
while True:
# read each frame and prepare it for feedforward in nn (resize and type)
ret, orig_frame = vs.read()
if ret is False:
break
else:
orig_frame = cv2.resize(orig_frame, (342, 256),
interpolation=cv2.INTER_LINEAR)
frame = cv2.cvtColor(orig_frame, cv2.COLOR_BGR2RGB)
frame = transform(frame).view(1, 3, 224, 224).cuda()
frame = frame.float().cuda(async=True)
# feed the frame to the neural network
nn_output = model(frame)
# vote for class with 25 consecutive frames
if frame_count % 10 == 0:
nn_output = softmax(nn_output)
nn_output = nn_output.data.cpu().numpy()
preds = nn_output.argsort()[0][-5:][::-1]
pred_classes = [(idx_to_class[str(pred + 1)],
nn_output[0, pred]) for pred in preds]
red = (0, 0, 255)
green = (0, 255, 0)
blue = (255, 0, 0)
white = (255, 255, 255)
yellow = (0, 255, 255)
cyan = (255, 255, 0)
magenta = (255, 0, 255)
thickness = 2
center_x = int(342 / 2.0)
center_y = int(256 / 2.0)
location = (center_x - 170, center_y + 80)
fontScale = 1.5
font = cv2.FONT_HERSHEY_SIMPLEX
y0, dy = 180, 20
value = 0
for i in range(5):
y = y0 + i * dy
if pred_classes[i][0] == label:
value = pred_classes[i][1]
# reset the process
f.write(str(value) + "\n")
nn_output = torch.tensor(np.zeros((1, 23)),
dtype=torch.float32).cuda()
# cv2.imwrite("temp/" + str(frame_count) + ".jpg", orig_frame)
# Display the resulting frame and the classified action
frame_count += 1
# When everything done, release the capture
f.write("----\n")
vs.release()
f.close()
def main(args):
############################################################################
# Path to optical flow images
if args.detector == 'yolo':
img_root = './data/yolov3/'
else:
img_root = './data/faster-rcnn/'
# Path to training and testing files
load_path = './data/'
# CPU or GPU?
device = torch.device("cuda")
# Model saving and loading
model_save_path = './data/'
model_load_path = './data/'
# Training settings
epochs = 15
batch_size = 64
learning_rate = 1e-5
num_workers = 8
weight_decay = 1e-2
NUM_FLOW_FRAMES = 9
training_proportion = 100 # How much of the dataset to use? 100 = 100percent
# Transformers for training and validation
transform_train = video_transforms.Compose([
video_transforms.MultiScaleCrop((224, 224), [1.0]),
video_transforms.ToTensor(),
])
transform_val = video_transforms.Compose([
video_transforms.Scale((224)),
video_transforms.ToTensor(),
])
############################################################################
print('################### Training settings ###################')
print('epochs:', epochs, ' batch_size:', batch_size, ' learning_rate:',
learning_rate, ' num_workers:', num_workers, ' NUM_FLOW_FRAMES:',
NUM_FLOW_FRAMES)
results = pd.DataFrame()
print('Training model')
print(args.detector + '_bdd10k_val.pkl')
try:
testset = LocationDatasetBDD(filename='bdd10k_val_' + args.detector +
'.pkl',
root_dir=load_path,
transform=transform_val,
img_root=img_root,
NUM_FLOW_FRAMES=NUM_FLOW_FRAMES)
test_loader = torch.utils.data.DataLoader(testset,
batch_size=batch_size,
shuffle=False,
num_workers=num_workers)
trainset = LocationDatasetBDD(filename='bdd10k_train_' +
args.detector + '.pkl',
root_dir=load_path,
transform=transform_train,
img_root=img_root,
NUM_FLOW_FRAMES=NUM_FLOW_FRAMES,
proportion=training_proportion)
train_loader = torch.utils.data.DataLoader(trainset,
batch_size=batch_size,
shuffle=True,
num_workers=num_workers)
valset = LocationDatasetBDD(filename='bdd10k_val_' + args.detector +
'.pkl',
root_dir=load_path,
transform=transform_val,
img_root=img_root,
NUM_FLOW_FRAMES=NUM_FLOW_FRAMES)
val_loader = torch.utils.data.DataLoader(valset,
batch_size=batch_size,
shuffle=False,
num_workers=num_workers)
except:
sys.exit(
'ERROR: Could not load pkl data file. Check the bdd .pkl files are in the correct path.'
)
model = DynamicTrajectoryPredictor(NUM_FLOW_FRAMES).to(device)
model = model.float()
model = nn.DataParallel(model)
optimizer = optim.Adam(model.parameters(),
lr=learning_rate,
weight_decay=weight_decay)
loss_function = torch.nn.MSELoss()
best_FDE = np.inf
best_MSE = np.inf
best_model = copy.deepcopy(model)
# Begin training
for epoch in range(1, epochs + 1):
train(model, device, train_loader, optimizer, epoch, loss_function)
MSE_5, FDE_5, MSE_10, FDE_10, MSE_15, FDE_15, _, _ = test(
model, device, val_loader, loss_function)
if MSE_15 < best_MSE:
best_MSE = MSE_15
best_model = copy.deepcopy(model)
best_FDE = FDE_15
torch.save(
best_model.state_dict(),
model_save_path + args.detector + '_rn18_bdd10k_flow_css_' +
str(NUM_FLOW_FRAMES) + 'stack_training_proportion_' +
str(training_proportion) + '_shuffled_disp.weights')
print(epoch)
print('Best MSE:', round(best_MSE, 0))
test_mse_5, test_fde_5, test_mse_10, test_fde_10, test_mse_15, test_fde_15, all_outputs, all_targets = test(
best_model, device, test_loader, loss_function)
print('Test mse @ 15:', round(test_mse_15, 0))
# Save the model
torch.save(
best_model.state_dict(),
model_save_path + args.detector + 'bdd10k_rn18_flow_css_' +
str(NUM_FLOW_FRAMES) + 'stack_training_proportion_' +
str(training_proportion) + '_shuffled_disp.weights')
# Save the predictions and the targets
np.save(
'./' + args.detector + '_predictions_rn18_flow_css_' +
str(NUM_FLOW_FRAMES) + 'stack_bdd10k_training_proportion_' +
str(training_proportion) + '_shuffled_disp.npy', all_outputs)
np.save(
'./' + args.detector + '_targets_rn18_flow_css_' +
str(NUM_FLOW_FRAMES) + 'stack_bdd10k__shuffled_disp.npy', all_targets)
# Save the results
result = {
'NUM_FLOW_FRAMES': NUM_FLOW_FRAMES,
'training_proportion': training_proportion,
'val_mse': best_MSE,
'val_fde': best_FDE,
'test_mse_5': test_mse_5,
'test_fde_5': test_fde_5,
'test_mse_10': test_mse_10,
'test_fde_10': test_fde_10,
'test_mse_15': test_mse_15,
'test_fde_15': test_fde_15
}
results = results.append(result, ignore_index=True)
results.to_csv('./' + args.detector + '_results_rn18_bdd10k.csv',
index=False)
def main(args):
############################################################################
# Path to optical flow images
img_root = './data/human-annotated/'
# Path to training and testing files
load_path = './data/'
# CPU or GPU?
device = torch.device("cuda")
# Training settings
epochs = 30
batch_size = 64
learning_rate = 1e-5
num_workers = 8
pretrained = False
weight_decay = 1e-2
NUM_FLOW_FRAMES = 9
model_load_path = args.model_load_path
model_save_path = args.model_save_path
# Transformers for training and validation
transform_train = video_transforms.Compose([
video_transforms.MultiScaleCrop((224, 224), [1.0]),
video_transforms.ToTensor(),
])
transform_val = video_transforms.Compose([
video_transforms.Scale((224)),
video_transforms.ToTensor(),
])
############################################################################
print('################### Training settings ###################')
print('epochs:', epochs, ' batch_size:', batch_size, ' learning_rate:',
learning_rate, ' num_workers:', num_workers, ' model_load_path:',
model_load_path, ' NUM_FLOW_FRAMES:', NUM_FLOW_FRAMES)
results = pd.DataFrame()
testset = LocationDatasetJAAD(filename='jaad_cv_test.pkl',
root_dir=load_path,
transform=transform_val,
img_root=img_root,
NUM_FLOW_FRAMES=NUM_FLOW_FRAMES)
test_loader = torch.utils.data.DataLoader(testset,
batch_size=batch_size,
shuffle=False,
num_workers=num_workers)
trainset = LocationDatasetJAAD(filename='jaad_cv_train_' + str(1) + '.pkl',
root_dir=load_path,
transform=transform_train,
img_root=img_root,
NUM_FLOW_FRAMES=NUM_FLOW_FRAMES)
train_loader = torch.utils.data.DataLoader(trainset,
batch_size=batch_size,
shuffle=True,
num_workers=num_workers)
valset = LocationDatasetJAAD(filename='jaad_cv_val_' + str(1) + '.pkl',
root_dir=load_path,
transform=transform_val,
img_root=img_root,
NUM_FLOW_FRAMES=NUM_FLOW_FRAMES)
val_loader = torch.utils.data.DataLoader(valset,
batch_size=batch_size,
shuffle=False,
num_workers=num_workers)
print('test_loader flow_stack size = ', test_loader['flow_stack'])
def main():
global best_prec1
# create model
print("Building model ... ")
print("Building model ... ", file = f_log)
model = build_model(resume_path = args.resume)
print("Model %s is loaded. " % (args.arch))
print("Model %s is loaded. " % (args.arch), file = f_log)
# define loss function (criterion) and optimizer
criterion = nn.CrossEntropyLoss().cuda()
optimizer = torch.optim.SGD(model.parameters(), args.lr,
momentum=args.momentum,
weight_decay=args.weight_decay)
cudnn.benchmark = True
# Data transforming
if args.modality == "rgb":
is_color = True
scale_ratios = [1.0, 0.875, 0.75, 0.66]
clip_mean = [0.485, 0.456, 0.406]
clip_std = [0.229, 0.224, 0.225]
elif args.modality == "tvl1_flow" or args.modality == "lk_flow":
is_color = False
scale_ratios = [1.0, 0.875, 0.75]
clip_mean = [0.5, 0.5]
clip_std = [0.226, 0.226]
else:
print("No such modality. Only rgb and flow supported.")
print("No such modality. Only rgb and flow supported.", file = f_log)
normalize = video_transforms.Normalize(mean=clip_mean, std=clip_std)
train_transform = video_transforms.Compose([
#video_transforms.Scale((288)),
video_transforms.MultiScaleCrop((256, 256), scale_ratios),
video_transforms.RandomHorizontalFlip(),
video_transforms.ToTensor(),
normalize,
])
val_transform = video_transforms.Compose([
#video_transforms.Scale((288)),
video_transforms.CenterCrop((256)),
video_transforms.ToTensor(),
normalize,
])
# data loading
train_setting_file = "train_%s_split%d.txt" % (args.modality, args.split)
train_split_file = os.path.join(args.settings, args.dataset, train_setting_file)
val_setting_file = "val_%s_split%d.txt" % (args.modality, args.split)
val_split_file = os.path.join(args.settings, args.dataset, val_setting_file)
if not os.path.exists(train_split_file) or not os.path.exists(val_split_file):
print("No split file exists in %s directory. Preprocess the dataset first" % (args.settings))
print("No split file exists in %s directory. Preprocess the dataset first" % (args.settings), file = f_log )
train_dataset = datasets.__dict__[args.dataset](setting=train_split_file, root=args.data, train=True,
new_width=args.new_width, new_height=args.new_height, new_length=args.new_length,
target_width=args.new_width, target_height=args.new_height,
modality=args.modality, num_segments=args.num_segments, transform=train_transform,
name_pattern='frame%06d.jpg')
val_dataset = datasets.__dict__[args.dataset](setting=val_split_file, root=args.data, train=False,
new_width=args.new_width, new_height=args.new_height, new_length=args.new_length,
target_width=args.new_width, target_height=args.new_height,
modality=args.modality, num_segments=args.num_segments, transform=val_transform,
name_pattern='frame%06d.jpg')
print('{} samples found, {} train samples and {} test samples.'.format(len(val_dataset)+len(train_dataset),
len(train_dataset),
len(val_dataset)))
print('{} samples found, {} train samples and {} test samples.'.format(len(val_dataset)+len(train_dataset),
len(train_dataset),
len(val_dataset)), file = f_log )
train_loader = torch.utils.data.DataLoader(
train_dataset,
batch_size=args.batch_size, shuffle=True,
num_workers=args.workers) #, pin_memory=True)
val_loader = torch.utils.data.DataLoader(
val_dataset,
batch_size=args.batch_size, shuffle=True,
num_workers=args.workers) #, pin_memory=True)
if args.evaluate:
validate(val_loader, model, criterion)
return
for epoch in range(args.start_epoch, args.epochs):
adjust_learning_rate(optimizer, epoch)
# train for one epoch
print("start epoch ", epoch)
train(train_loader, model, criterion, optimizer, epoch)
# evaluate on validation set
prec1 = 0.0
if (epoch + 1) % args.save_freq == 0:
prec1 = validate(val_loader, model, criterion)
# remember best prec@1 and save checkpoint
is_best = prec1 > best_prec1
best_prec1 = max(prec1, best_prec1)
if (epoch + 1) % args.save_freq == 0:
checkpoint_name = "%03d_%s" % (epoch + 1, "checkpoint.pth.tar")
save_checkpoint({
'epoch': epoch + 1,
'arch': args.arch,
'state_dict': model.state_dict(),
'best_prec1': best_prec1,
'optimizer' : optimizer.state_dict(),
}, is_best, checkpoint_name, args.save_path)
def main():
global args, prec_list
prec_list = []
args = parser.parse_args()
full_path = logging(args)
print(args.modality + " network trained with the split " +
str(args.split) + ".")
# create model
print("Building model ... ")
exits_model, model = build_model(int(args.start_epoch),
args.pretrain_weights)
if not exits_model:
return
else:
print("Model %s is loaded. " % (args.arch))
# define loss function (criterion) and optimizer
criterion = nn.CrossEntropyLoss().cuda()
optimizer = torch.optim.SGD(model.parameters(),
args.lr,
momentum=args.momentum,
weight_decay=args.weight_decay)
cudnn.benchmark = True
# Data transforming
if args.modality == "rgb" or args.modality == "rgb2":
is_color = True
scale_ratios = [1.0, 0.875, 0.75, 0.66]
clip_mean = [0.485, 0.456, 0.406] * args.new_length
clip_std = [0.229, 0.224, 0.225] * args.new_length
elif args.modality == "flow" or args.modality == "rhythm":
is_color = False
scale_ratios = [1.0, 0.875, 0.75]
clip_mean = [0.5, 0.5] * args.new_length
clip_std = [0.226, 0.226] * args.new_length
else:
print("No such modality. Only rgb and flow supported.")
new_size = 299 if args.arch.find("inception_v3") > 0 else 224
normalize = video_transforms.Normalize(mean=clip_mean, std=clip_std)
train_transform = video_transforms.Compose([
#video_transforms.Scale((256)),
video_transforms.MultiScaleCrop((new_size, new_size), scale_ratios),
video_transforms.RandomHorizontalFlip(),
video_transforms.ToTensor(),
normalize,
])
if args.es:
val_transform = video_transforms.Compose([
# video_transforms.Scale((256)),
video_transforms.CenterCrop((new_size)),
video_transforms.ToTensor(),
normalize,
])
modality_ = "rgb" if (args.modality == "rhythm"
or args.modality[:3] == "rgb") else "flow"
if args.modality == "rgb2":
createNewDataset("train_split%d.txt", "new_train.txt", modality_)
#createNewDataset("val_%s_split%d.txt", "new_val.txt",modality_)
# data loading
train_setting_file = "new_train.txt" if args.modality == "rgb2" else "train_split%d.txt" % (
args.split)
train_split_file = os.path.join(args.settings, args.dataset,
train_setting_file)
if not os.path.exists(
train_split_file): # or not os.path.exists(val_split_file):
print(
"No split file exists in %s directory. Preprocess the dataset first"
% (args.settings))
extension = ".png" if args.dataset == "hmdb51" and args.modality == "rhythm" else ".jpg"
direction_file = "direction.txt" if args.vr_approach == 3 else "direction_video.txt"
direction_path = os.path.join(args.settings, args.dataset, direction_file)
train_dataset = datasets.__dict__['dataset'](
root=args.data,
source=train_split_file,
phase="train",
modality=args.modality,
is_color=is_color,
new_length=args.new_length,
new_width=args.new_width,
new_height=args.new_height,
video_transform=train_transform,
approach_VR=args.vr_approach,
extension=extension,
direction_path=direction_path)
if args.es:
val_setting_file = "val_split%d.txt" % (args.split)
val_split_file = os.path.join(args.settings, args.dataset,
val_setting_file)
if not os.path.exists(val_split_file):
print(
"No split file exists in %s directory. Preprocess the dataset first"
% (args.settings))
val_dataset = datasets.__dict__['dataset'](
root=args.data,
source=val_split_file,
phase="val",
modality=args.modality,
is_color=is_color,
new_length=args.new_length,
new_width=args.new_width,
new_height=args.new_height,
video_transform=val_transform,
approach_VR=args.vr_approach,
extension=extension,
direction_path=direction_path)
print('{} samples found, {} train samples and {} validation samples.'.
format(
len(val_dataset) + len(train_dataset), len(train_dataset),
len(val_dataset)))
else:
print('{} train samples found.'.format(len(train_dataset)))
train_loader = torch.utils.data.DataLoader(train_dataset,
batch_size=args.batch_size,
shuffle=True,
num_workers=args.workers,
pin_memory=True)
if args.es:
val_loader = torch.utils.data.DataLoader(val_dataset,
batch_size=args.batch_size,
shuffle=True,
num_workers=args.workers,
pin_memory=True)
early_stop = EarlyStopping(verbose=True,
log_path=os.path.join(
full_path, "early_stopping.json"))
is_best = False
for epoch in range(args.start_epoch, args.epochs):
adjust_learning_rate(optimizer, epoch)
# train for one epoch
train(train_loader, model, criterion, optimizer, epoch)
if args.es:
# evaluate on validation set
losses = validate(val_loader, model, criterion)
is_best = early_stop(losses.avg, epoch)
if (epoch + 1) % args.save_freq == 0 or is_best:
checkpoint_name = "%03d_%s" % (epoch + 1, "checkpoint_" +
args.modality + "_split_" +
str(args.split) + ".pth.tar")
es_val = float('inf') if not args.es else early_stop.val_loss_min
save_checkpoint(
{
'epoch': epoch + 1,
'arch': args.arch,
'state_dict': model.state_dict(),
'optimizer': optimizer.state_dict(),
'val_loss_min': es_val
}, is_best, checkpoint_name,
os.path.join(full_path, "checkpoints"))
prec_name = "%03d_%s" % (epoch + 1,
"prec_split_" + str(args.split) + ".txt")
save_precision(prec_name, os.path.join(full_path, "precision"))
if args.es and early_stop.early_stop:
break
if not args.es: # Final model
checkpoint_name = "%03d_%s" % (epoch + 1, "checkpoint_" +
args.modality + "_split_" +
str(args.split) + ".pth.tar")
save_checkpoint(
{
'epoch': epoch + 1,
'arch': args.arch,
'state_dict': model.state_dict(),
'optimizer': optimizer.state_dict(),
'val_loss_min': float('inf')
}, True, checkpoint_name, os.path.join(full_path, "checkpoints"))
def VideoTemporalPrediction(
mode,
vid_name,
net,
num_categories,
start_frame=0,
num_frames=0,
num_samples=25,
optical_flow_frames=10,
new_size = 299,
ext = ".jpg"
):
if num_frames == 0:
imglist = os.listdir(vid_name)
duration = len(imglist)
else:
duration = num_frames
# selection
step = int(math.floor((duration-optical_flow_frames+1)/num_samples))
clip_mean = [0.5] * optical_flow_frames * 2
clip_std = [0.226] * optical_flow_frames * 2
normalize = video_transforms.Normalize(mean=clip_mean,
std=clip_std)
test_transform = video_transforms.Compose([
video_transforms.ToTensor(),
normalize
])
# inception = 320,360, resnet = 240, 320
width = 320 if new_size==299 else 240
height = 360 if new_size==299 else 320
deep = optical_flow_frames*2
dims = (width,height,deep,num_samples)
flow = np.zeros(shape=dims, dtype=np.float64)
flow_flip = np.zeros(shape=dims, dtype=np.float64)
for i in range(num_samples):
for j in range(optical_flow_frames):
flow_x_file = os.path.join(vid_name, mode+'_x_{0:05d}{1}'.format(i*step+j+1 + start_frame, ext))
flow_y_file = os.path.join(vid_name, mode+'_y_{0:05d}{1}'.format(i*step+j+1 + start_frame, ext))
img_x = cv2.imread(flow_x_file, cv2.IMREAD_GRAYSCALE)
img_y = cv2.imread(flow_y_file, cv2.IMREAD_GRAYSCALE)
img_x = cv2.resize(img_x, dims[1::-1])
img_y = cv2.resize(img_y, dims[1::-1])
flow[:,:,j*2 ,i] = img_x
flow[:,:,j*2+1,i] = img_y
flow_flip[:,:,j*2 ,i] = 255 - img_x[:, ::-1]
flow_flip[:,:,j*2+1,i] = img_y[:, ::-1]
# crop 299 = inception, 224 = resnet
size = new_size
corner = [(height-size)//2, (width-size)//2]
flow_1 = flow[:size, :size, :,:]
flow_2 = flow[:size, -size:, :,:]
flow_3 = flow[corner[1]:corner[1]+size, corner[0]:corner[0]+size, :,:]
flow_4 = flow[-size:, :size, :,:]
flow_5 = flow[-size:, -size:, :,:]
flow_f_1 = flow_flip[:size, :size, :,:]
flow_f_2 = flow_flip[:size, -size:, :,:]
flow_f_3 = flow_flip[corner[1]:corner[1]+size, corner[0]:corner[0]+size, :,:]
flow_f_4 = flow_flip[-size:, :size, :,:]
flow_f_5 = flow_flip[-size:, -size:, :,:]
flow = np.concatenate((flow_1,flow_2,flow_3,flow_4,flow_5,flow_f_1,flow_f_2,flow_f_3,flow_f_4,flow_f_5), axis=3)
_, _, _, c = flow.shape
flow_list = []
for c_index in range(c):
cur_img = flow[:,:,:,c_index].squeeze()
cur_img_tensor = test_transform(cur_img)
flow_list.append(np.expand_dims(cur_img_tensor.numpy(), 0))
flow_np = np.concatenate(flow_list,axis=0)
batch_size = 15
prediction = np.zeros((num_categories,flow.shape[3]))
num_batches = int(math.ceil(float(flow.shape[3])/batch_size))
for bb in range(num_batches):
span = range(batch_size*bb, min(flow.shape[3],batch_size*(bb+1)))
input_data = flow_np[span,:,:,:]
imgDataTensor = torch.from_numpy(input_data).type(torch.FloatTensor).cuda()
imgDataVar = torch.autograd.Variable(imgDataTensor)
output = net(imgDataVar)
result = output.data.cpu().numpy()
prediction[:, span] = np.transpose(result)
return prediction
def main():
global args, best_prec1
args = parser.parse_args()
print(args.modality + " network trained whith the split " +
str(args.split) + ".")
# create model
print("Building model ... ")
exits_model, model = build_model(int(args.start_epoch))
if not exits_model:
return
else:
print("Model %s is loaded. " % (args.arch))
# define loss function (criterion) and optimizer
criterion = nn.CrossEntropyLoss().cuda()
optimizer = torch.optim.SGD(model.parameters(),
args.lr,
momentum=args.momentum,
weight_decay=args.weight_decay)
# create file where we allocate the models by each args.save_freq epochs
if not os.path.exists(args.resume):
os.makedirs(args.resume)
print("Saving everything to directory %s." % (args.resume))
cudnn.benchmark = True
# Data transforming
if args.modality == "rgb" or args.modality == "rhythm" or args.modality == "history":
is_color = True
scale_ratios = [1.0, 0.875, 0.75, 0.66]
clip_mean = [0.485, 0.456, 0.406] * args.new_length
clip_std = [0.299, 0.224, 0.225] * args.new_length
elif args.modality == "flow":
is_color = False
scale_ratios = [1.0, 0.875, 0.75]
clip_mean = [0.5, 0.5] * args.new_length
clip_std = [0.226, 0.226] * args.new_length
else:
print("No such modality. Only rgb and flow supported.")
new_size = 299 if args.arch == 'rgb_inception_v3' else 224
normalize = video_transforms.Normalize(mean=clip_mean, std=clip_std)
train_transform = video_transforms.Compose([
#video_transforms.Scale((256)),
video_transforms.MultiScaleCrop((new_size, new_size), scale_ratios),
video_transforms.RandomHorizontalFlip(),
video_transforms.ToTensor(),
normalize,
])
val_transform = video_transforms.Compose([
# video_transforms.Scale((256)),
video_transforms.CenterCrop((new_size)),
video_transforms.ToTensor(),
normalize,
])
#createNewDataset("train_%s_split%d.txt" , "new_train.txt")
#createNewDataset("val_%s_split%d.txt", "new_test.txt")
# data loading
#train_setting_file = 'new_train.txt'
modality_ = "rgb" if (args.modality == "rhythm"
or args.modality == "history") else args.modality
train_setting_file = "train_%s_split%d.txt" % (modality_, args.split)
train_split_file = os.path.join(args.settings, args.dataset,
train_setting_file)
#val_setting_file = 'new_test.txt'
val_setting_file = "val_%s_split%d.txt" % (modality_, args.split)
val_split_file = os.path.join(args.settings, args.dataset,
val_setting_file)
if not os.path.exists(train_split_file) or not os.path.exists(
val_split_file):
print(
"No split file exists in %s directory. Preprocess the dataset first"
% (args.settings))
train_dataset = datasets.__dict__['dataset'](
root=args.data,
source=train_split_file,
phase="train",
modality=args.modality,
is_color=is_color,
new_length=args.new_length,
new_width=args.new_width,
new_height=args.new_height,
video_transform=train_transform)
val_dataset = datasets.__dict__['dataset'](root=args.data,
source=val_split_file,
phase="val",
modality=args.modality,
is_color=is_color,
new_length=args.new_length,
new_width=args.new_width,
new_height=args.new_height,
video_transform=val_transform)
print('{} samples found, {} train samples and {} test samples.'.format(
len(val_dataset) + len(train_dataset), len(train_dataset),
len(val_dataset)))
train_loader = torch.utils.data.DataLoader(train_dataset,
batch_size=args.batch_size,
shuffle=True,
num_workers=args.workers,
pin_memory=True)
val_loader = torch.utils.data.DataLoader(val_dataset,
batch_size=args.batch_size,
shuffle=True,
num_workers=args.workers,
pin_memory=True)
if args.evaluate:
validate(val_loader, model, criterion)
return
for epoch in range(args.start_epoch, args.epochs):
adjust_learning_rate(optimizer, epoch)
# train for one epoch
train(train_loader, model, criterion, optimizer, epoch)
# evaluate on validation set
prec1 = 0.0
if (epoch + 1) % args.save_freq == 0:
prec1 = validate(val_loader, model, criterion)
# remember best prec@1 and save checkpoint
is_best = prec1 > best_prec1
best_prec1 = max(prec1, best_prec1)
if (epoch + 1) % args.save_freq == 0:
checkpoint_name = "%03d_%s" % (epoch + 1, "checkpoint_" +
args.modality + "_split_" +
str(args.split) + ".pth.tar")
save_checkpoint(
{
'epoch': epoch + 1,
'arch': args.arch,
'state_dict': model.state_dict(),
'best_prec1': best_prec1,
'optimizer': optimizer.state_dict(),
}, is_best, checkpoint_name, args.resume)
def VideoTemporalPrediction(vid_name,
net,
num_categories,
start_frame=0,
num_frames=0,
num_samples=5,
optical_flow_frames=25):
if num_frames == 0:
# print(vid_name)
imglist = glob.glob(os.path.join(vid_name, '*flow_x*.jpg'))
duration = len(imglist)
else:
duration = num_frames
clip_mean = [0.5] * 20
clip_std = [0.226] * 20
normalize = video_transforms.Normalize(mean=clip_mean, std=clip_std)
val_transform = video_transforms.Compose([
video_transforms.ToTensor(),
normalize,
])
# selection
step = int(math.floor((duration - optical_flow_frames + 1) / num_samples))
dims = (256, 340, optical_flow_frames * 2, num_samples)
flow = np.zeros(shape=dims, dtype=np.float64)
flow_flip = np.zeros(shape=dims, dtype=np.float64)
for i in range(num_samples):
for j in range(optical_flow_frames):
flow_x_file = os.path.join(
vid_name,
'flow_x_{0:04d}.jpg'.format(i * step + j + 1 + start_frame))
flow_y_file = os.path.join(
vid_name,
'flow_y_{0:04d}.jpg'.format(i * step + j + 1 + start_frame))
img_x = cv2.imread(flow_x_file, cv2.IMREAD_GRAYSCALE)
img_y = cv2.imread(flow_y_file, cv2.IMREAD_GRAYSCALE)
img_x = cv2.resize(img_x, dims[1::-1])
img_y = cv2.resize(img_y, dims[1::-1])
flow[:, :, j * 2, i] = img_x
flow[:, :, j * 2 + 1, i] = img_y
flow_flip[:, :, j * 2, i] = 255 - img_x[:, ::-1]
flow_flip[:, :, j * 2 + 1, i] = img_y[:, ::-1]
# crop
flow_1 = flow[:224, :224, :, :]
flow_2 = flow[:224, -224:, :, :]
flow_3 = flow[16:240, 60:284, :, :]
flow_4 = flow[-224:, :224, :, :]
flow_5 = flow[-224:, -224:, :, :]
flow_f_1 = flow_flip[:224, :224, :, :]
flow_f_2 = flow_flip[:224, -224:, :, :]
flow_f_3 = flow_flip[16:240, 60:284, :, :]
flow_f_4 = flow_flip[-224:, :224, :, :]
flow_f_5 = flow_flip[-224:, -224:, :, :]
flow = np.concatenate((flow_1, flow_2, flow_3, flow_4, flow_5, flow_f_1,
flow_f_2, flow_f_3, flow_f_4, flow_f_5),
axis=3)
_, _, _, c = flow.shape
flow_list = []
for c_index in range(c):
cur_img = flow[:, :, :, c_index].squeeze()
cur_img_tensor = val_transform(cur_img)
flow_list.append(np.expand_dims(cur_img_tensor.numpy(), 0))
flow_np = np.concatenate(flow_list, axis=0)
batch_size = 25
prediction = np.zeros((num_categories, flow.shape[3]))
num_batches = int(math.ceil(float(flow.shape[3]) / batch_size))
for bb in range(num_batches):
span = range(batch_size * bb, min(flow.shape[3],
batch_size * (bb + 1)))
input_data = flow_np[span, :, :, :]
imgDataTensor = torch.from_numpy(input_data).type(
torch.FloatTensor).cuda()
imgDataVar = torch.autograd.Variable(imgDataTensor)
output = net(imgDataVar)
result = output.data.cpu().numpy()
prediction[:, span] = np.transpose(result)
return prediction
def main():
'''
导入模型
'''
device = 'cuda' if torch.cuda.is_available() else 'cpu'
model = DynamicTrajectoryPredictor(9).to(device)
model = model.float()
model = nn.DataParallel(model)
# summary(model, input_size=(18, 224, 224))
model.load_state_dict(
torch.load(
'./data/yolomyvideo_rn50_flow_css_9stack_training_proportion_100_shuffled_disp.weights'
), False)
model.eval()
load_path = './data/'
img_root = '../../flow_result/'
# Training settings
epochs = 15
batch_size = 1
learning_rate = 1e-5
num_workers = 8
weight_decay = 1e-2
NUM_FLOW_FRAMES = 9
training_proportion = 100
# Transformers
transform_val = video_transforms.Compose([
video_transforms.Scale((224)),
video_transforms.ToTensor(),
])
for fold_type in ['train', 'val', 'test']:
for fold_num in range(1, 4):
result.clear()
valset = LocationDatasetBDD(filename=fold_type + str(fold_num) +
'_myvideo_location_features_yolo.pkl',
root_dir=load_path,
transform=transform_val,
img_root=img_root,
NUM_FLOW_FRAMES=NUM_FLOW_FRAMES)
val_loader = torch.utils.data.DataLoader(valset,
batch_size=batch_size,
shuffle=False,
num_workers=num_workers)
for param in model.parameters():
param.requires_grad = False
start_time = time.time()
for batch_idx, data in enumerate(val_loader):
if batch_idx % 100 == 0:
end_time = time.time()
print(fold_type + ':', fold_num, ' Batch ', batch_idx,
' of ', len(val_loader), ' Cost time: ',
end_time - start_time)
start_time = end_time
# break
# if batch_idx == 20:
# break
flow = data['flow_stack'].to(device)
flow = flow.float()
output = model(flow)
# print('Processing: ', batch_idx)
ans = np.array(result).reshape(-1, 2048)
print(ans.shape)
with open('record_extract.txt', 'w') as f:
f.write(fold_type + ' ' + str(fold_num) + ' ' + str(ans.shape))
np.save(
'./data/sted_feature/fold_' + str(fold_num) + '_' + fold_type +
'_dtp_features.npy', ans)
def main(args):
############################################################################
# Path to optical flow images
img_root = './data/human-annotated/'
# Path to training and testing files
load_path = './data/'
# CPU or GPU?
device = torch.device("cuda")
# Training settings
epochs = 30
batch_size = 64
learning_rate = 1e-5
num_workers = 8
pretrained = False
weight_decay = 1e-2
NUM_FLOW_FRAMES = 9
model_load_path = args.model_load_path
model_save_path = args.model_save_path
# Transformers for training and validation
transform_train = video_transforms.Compose([
video_transforms.MultiScaleCrop((224, 224), [1.0]),
video_transforms.ToTensor(),
])
transform_val = video_transforms.Compose([
video_transforms.Scale((224)),
video_transforms.ToTensor(),
])
############################################################################
print('################### Training settings ###################')
print('epochs:', epochs, ' batch_size:', batch_size, ' learning_rate:',
learning_rate, ' num_workers:', num_workers, ' model_load_path:',
model_load_path, ' NUM_FLOW_FRAMES:', NUM_FLOW_FRAMES)
results = pd.DataFrame()
for fold in [1, 2, 3, 4, 5]:
if pretrained:
learning_rate = 1e-6
epochs = 30
else:
learning_rate = 1e-5
epochs = 40
print('Training on fold ' + str(fold))
try:
testset = LocationDatasetJAAD(filename='jaad_cv_test.pkl',
root_dir=load_path,
transform=transform_val,
img_root=img_root,
NUM_FLOW_FRAMES=NUM_FLOW_FRAMES)
test_loader = torch.utils.data.DataLoader(testset,
batch_size=batch_size,
shuffle=False,
num_workers=num_workers)
trainset = LocationDatasetJAAD(filename='jaad_cv_train_' +
str(fold) + '.pkl',
root_dir=load_path,
transform=transform_train,
img_root=img_root,
NUM_FLOW_FRAMES=NUM_FLOW_FRAMES)
train_loader = torch.utils.data.DataLoader(trainset,
batch_size=batch_size,
shuffle=True,
num_workers=num_workers)
valset = LocationDatasetJAAD(filename='jaad_cv_val_' + str(fold) +
'.pkl',
root_dir=load_path,
transform=transform_val,
img_root=img_root,
NUM_FLOW_FRAMES=NUM_FLOW_FRAMES)
val_loader = torch.utils.data.DataLoader(valset,
batch_size=batch_size,
shuffle=False,
num_workers=num_workers)
except:
sys.exit(
'ERROR: Could not load pkl data file. Check the jaad .pkl files are in the correct path.'
)
model = DynamicTrajectoryPredictor(NUM_FLOW_FRAMES).to(device)
model = model.float()
model = nn.DataParallel(model)
if model_load_path is not None:
print('loading model from', model_load_path)
model.load_state_dict(torch.load(model_load_path))
optimizer = optim.Adam(model.parameters(),
lr=learning_rate,
weight_decay=weight_decay)
loss_function = torch.nn.MSELoss()
best_FDE = np.inf
best_MSE = np.inf
best_model = copy.deepcopy(model)
# Begin training
for epoch in range(1, epochs + 1):
# Set learning rate to 1e-6 after 30 epochs
if epoch > 30:
optimizer = optim.Adam(model.parameters(),
lr=1e-6,
weight_weight_decay=decay)
train(model, device, train_loader, optimizer, epoch, loss_function)
MSE_5, FDE_5, MSE_10, FDE_10, MSE_15, FDE_15, _, _ = test(
model, device, val_loader, loss_function)
if MSE_15 < best_MSE:
best_MSE = MSE_15
best_model = copy.deepcopy(model)
best_FDE = FDE_15
print(epoch)
print('Best MSE:', round(best_MSE, 0))
test_mse_5, test_fde_5, test_mse_10, test_fde_10, test_mse_15, test_fde_15, all_outputs, all_targets = test(
best_model, device, test_loader, loss_function)
print('Test mse @ 15:', round(test_mse_15, 0))
# Save the model
torch.save(
best_model.state_dict(), model_save_path + 'rn18_flow_css_' +
str(NUM_FLOW_FRAMES) + 'stack_fold_' + str(fold) + '_pretrained-' +
str(pretrained) + '_disp.weights')
# Save the predictions and the targets
np.save(
'./predictions_rn18_flow_css_' + str(NUM_FLOW_FRAMES) +
'stack_jaad_fold_' + str(fold) + 'pretrained-' + str(pretrained) +
'_disp.npy', all_outputs)
np.save(
'./targets_rn18_flow_css_' + str(NUM_FLOW_FRAMES) +
'stack_jaad_fold_' + str(fold) + 'pretrained-' + str(pretrained) +
'_disp.npy', all_targets)
# Save the results
result = {
'NUM_FLOW_FRAMES': NUM_FLOW_FRAMES,
'fold': fold,
'val_mse': best_MSE,
'val_fde': best_FDE,
'test_mse_5': test_mse_5,
'test_fde_5': test_fde_5,
'test_mse_10': test_mse_10,
'test_fde_10': test_fde_10,
'test_mse_15': test_mse_15,
'test_fde_15': test_fde_15,
'pretrained': pretrained
}
results = results.append(result, ignore_index=True)
results.to_csv('./results_rn18_jaad.csv', index=False)
def main():
global args, best_prec1
args = parser.parse_args()
# create model
print("Building model ... ")
model = build_model()
if torch.cuda.is_available():
model = torch.nn.DataParallel(model)
print("Model %s is loaded. " % (args.arch))
# define loss function (criterion) and optimizer
criterion = nn.CrossEntropyLoss().cuda()
optimizer = torch.optim.SGD(model.parameters(),
args.lr,
momentum=args.momentum,
weight_decay=args.weight_decay)
if not os.path.exists(args.resume):
os.makedirs(args.resume)
print("Saving everything to directory %s." % (args.resume))
cudnn.benchmark = True
# Data transforming
if args.modality == "rgb":
is_color = True
scale_ratios = [1.0, 0.875, 0.75, 0.66]
clip_mean = [0.485, 0.456, 0.406] * args.new_length
clip_std = [0.229, 0.224, 0.225] * args.new_length
elif args.modality == "flow":
is_color = False
scale_ratios = [1.0, 0.875, 0.75]
clip_mean = [0.5, 0.5] * args.new_length
clip_std = [0.226, 0.226] * args.new_length
else:
print("No such modality. Only rgb and flow supported.")
normalize = video_transforms.Normalize(mean=clip_mean, std=clip_std)
train_transform = video_transforms.Compose([
# video_transforms.Scale((256)),
video_transforms.MultiScaleCrop((224, 224), scale_ratios),
video_transforms.RandomHorizontalFlip(),
video_transforms.ToTensor(),
normalize,
])
val_transform = video_transforms.Compose([
# video_transforms.Scale((256)),
video_transforms.CenterCrop((224)),
video_transforms.ToTensor(),
normalize,
])
# data loading
# train_setting_file = "train_%s_split%d.txt" % (args.modality, args.split)
# train_split_file = os.path.join(args.settings, args.dataset, train_setting_file)
# val_setting_file = "val_%s_split%d.txt" % (args.modality, args.split)
# val_split_file = os.path.join(args.settings, args.dataset, val_setting_file)
# if not os.path.exists(train_split_file) or not os.path.exists(val_split_file):
# print("No split file exists in %s directory. Preprocess the dataset first" % (args.settings))
train_split_file = './datasets/settings/train_set_detail.csv'
val_split_file = './datasets/settings/val_set_detail.csv'
train_dataset = datasets.__dict__[args.dataset](
root=args.data, #neet to change
source=train_split_file,
phase="train",
modality=args.modality,
is_color=is_color,
new_length=args.new_length,
new_width=args.new_width,
new_height=args.new_height,
video_transform=train_transform,
name_pattern="frame%06d.jpg") # frame000001
val_dataset = datasets.__dict__[args.dataset](
root=args.data,
source=val_split_file,
phase="val",
modality=args.modality,
is_color=is_color,
new_length=args.new_length,
new_width=args.new_width,
new_height=args.new_height,
video_transform=val_transform,
name_pattern="frame%06d.jpg")
print('{} samples found, {} train samples and {} test samples.'.format(
len(val_dataset) + len(train_dataset), len(train_dataset),
len(val_dataset)))
train_loader = torch.utils.data.DataLoader(train_dataset,
batch_size=args.batch_size,
shuffle=True,
num_workers=args.workers,
pin_memory=True)
val_loader = torch.utils.data.DataLoader(val_dataset,
batch_size=args.batch_size,
shuffle=True,
num_workers=args.workers,
pin_memory=True)
if args.evaluate:
validate(val_loader, model, criterion)
return
model_path = '/home/thl/Desktop/challeng/checkpoints/Mulity_100step_900epoch_batch80/model_best.pth.tar'
params = torch.load(model_path)
model.load_state_dict(params['state_dict'])
for epoch in range(args.start_epoch, args.epochs):
adjust_learning_rate(optimizer, epoch)
# train for one epoch
train(train_loader, model, criterion, optimizer, epoch)
# evaluate on validation set
prec1 = 0.0
if (epoch + 1) % args.save_freq == 0:
prec1 = validate(val_loader, model, criterion)
# remember best prec@1 and save checkpoint
is_best = prec1 > best_prec1
best_prec1 = max(prec1, best_prec1)
if (epoch + 1) % args.save_freq == 0:
checkpoint_name = "%03d_%s" % (epoch + 1, "checkpoint.pth.tar")
save_checkpoint(
{
'epoch': epoch + 1,
'arch': args.arch,
'state_dict': model.state_dict(),
'best_prec1': best_prec1,
'optimizer': optimizer.state_dict(),
}, is_best, checkpoint_name, args.resume)
import math
import numpy as np
from PIL import Image
from tsn_dataset import TSNDataSet
from p3d_model import P3D199,get_optim_policies
import video_transforms
from tsn_models import TSN
from torch.nn.utils import clip_grad_norm
val_transform=video_transforms.Compose(
[
video_transforms.Resize((182,242)),
video_transforms.CenterCrop(160),
video_transforms.ToTensor(),
video_transforms.Normalize((0.485,0.456,0.406),
(0.229,0.224,0.225))]
)
val_loader=torch.utils.data.DataLoader(
TSNDataSet("","tsntest_01.lst",
num_segments=2,
new_length=16,
modality="RGB",
image_tmpl="frame{:06d}.jpg",
transform=val_transform,
random_shift=False),
batch_size=1,
shuffle=False,
num_workers=1,
def VideoSpatialPrediction(
vid_name,
net,
num_categories,
num_frames=0,
ext_batch_sz=100,
int_batch_sz=5,
new_size = 299
):
if num_frames == 0:
imglist = os.listdir(vid_name)
duration = len(imglist)
else:
duration = num_frames
clip_mean = [0.485, 0.456, 0.406]
clip_std = [0.229, 0.224, 0.225]
normalize = video_transforms.Normalize(mean=clip_mean,
std=clip_std)
val_transform = video_transforms.Compose([
video_transforms.ToTensor(),
normalize,
])
deep = 3
# inception = 320,360, resnet = 240, 320
width = 320 if new_size==299 else 240
height = 360 if new_size==299 else 320
predictions = []
for i in range(len(num_categories)):
predictions.append(np.zeros((num_categories[i],num_frames*10)))
#control memory (RAM) usage
num_ext_batch = int(math.ceil(float(num_frames)/ext_batch_sz))
for i in range(num_ext_batch):
start = i*ext_batch_sz
end = min(start+ext_batch_sz, num_frames)
dims = (width,height,deep,end-start)
rgb = np.zeros(shape=dims, dtype=np.float64)
rgb_flip = np.zeros(shape=dims, dtype=np.float64)
for j in range(end-start):
img_file = os.path.join(vid_name, 'img_{0:05d}.jpg'.format(j+start+1))
img = cv2.imread(img_file, cv2.IMREAD_UNCHANGED)
img = cv2.resize(img, dims[1::-1])
img = cv2.cvtColor(img, cv2.COLOR_BGR2RGB)
rgb[:,:,:,j] = img
rgb_flip[:,:,:,j] = img[:,::-1,:]
# crop 299 = inception, 224 = resnet
size = new_size
corner = [(height-size)//2, (width-size)//2]
rgb_1 = rgb[:size, :size, :,:]
rgb_2 = rgb[:size, -size:, :,:]
rgb_3 = rgb[corner[1]:corner[1]+size, corner[0]:corner[0]+size, :,:]
rgb_4 = rgb[-size:, :size, :,:]
rgb_5 = rgb[-size:, -size:, :,:]
rgb_f_1 = rgb_flip[:size, :size, :,:]
rgb_f_2 = rgb_flip[:size, -size:, :,:]
rgb_f_3 = rgb_flip[corner[1]:corner[1]+size, corner[0]:corner[0]+size, :,:]
rgb_f_4 = rgb_flip[-size:, :size, :,:]
rgb_f_5 = rgb_flip[-size:, -size:, :,:]
rgb = np.concatenate((rgb_1,rgb_2,rgb_3,rgb_4,rgb_5,rgb_f_1,rgb_f_2,rgb_f_3,rgb_f_4,rgb_f_5), axis=3)
rgb_1, rgb_2, rgb_3, rgb_4, rgb_5 = [],[],[],[],[]
rgb_f_1, rgb_f_2, rgb_f_3, rgb_f_4, rgb_f_5 = [],[],[],[],[]
rgb_flip = []
_, _, _, c = rgb.shape
rgb_list = []
for c_index in range(c):
cur_img = rgb[:,:,:,c_index]
cur_img_tensor = val_transform(cur_img)
rgb_list.append(np.expand_dims(cur_img_tensor.numpy(), 0))
rgb_shape = rgb.shape
rgb = []
rgb_np = np.concatenate(rgb_list,axis=0)
#control memory (GPU) usage
num_int_batches = int(math.ceil(float(rgb_shape[3])/int_batch_sz))
rgb_list = []
for bb in range(num_int_batches):
span = range(int_batch_sz*bb, min(rgb_shape[3],int_batch_sz*(bb+1)))
input_data = rgb_np[span,:,:,:]
imgDataTensor = torch.from_numpy(input_data).type(torch.FloatTensor).cuda()
imgDataVar = torch.autograd.Variable(imgDataTensor)
output = net(imgDataVar)
for ii in range(len(output)):
output_ = output[ii].reshape(-1, num_categories[ii])
result = output_.data.cpu().numpy()
pos = [ x%(end-start) + start + int(x/(end-start))*num_frames for x in span ]
predictions[ii][:, pos] = np.transpose(result)
rgb_np = []
result = []
for ii in range(len(predictions)):
result.append(np.split(predictions[ii],10,axis=1))
return result
def main():
device = 'cuda' if torch.cuda.is_available() else 'cpu'
model = DynamicTrajectoryPredictor(9).to(device)
model = model.float()
model = nn.DataParallel(model)
# 输出网络结构
# summary(model, input_size=(18, 224, 224))
model.load_state_dict(torch.load('./model.weights'), False)
load_path = './data_inference/'
img_root = '../../../flow_result/'
# Training settings
epochs = 15
batch_size = 1
learning_rate = 1e-5
num_workers = 8
weight_decay = 1e-2
NUM_FLOW_FRAMES = 9
training_proportion = 100
# Transformers
transform_val = video_transforms.Compose([
video_transforms.Scale((224)),
video_transforms.ToTensor(),
])
valset = LocationDatasetBDD(filename='myvideo_val_yolo_0.pkl',
root_dir=load_path,
transform=transform_val,
img_root=img_root,
NUM_FLOW_FRAMES=NUM_FLOW_FRAMES)
val_loader = torch.utils.data.DataLoader(valset,
batch_size=batch_size,
shuffle=False,
num_workers=num_workers)
# for param in model.parameters():
# param.requires_grad = False
model.eval()
tmp_result = []
start_time = time.time()
with torch.no_grad():
for batch_idx, data in enumerate(val_loader):
if batch_idx % 100 == 0:
end_time = time.time()
print(' Batch ', batch_idx, ' of ', len(val_loader),
' Cost time: ', end_time - start_time)
start_time = end_time
flow = data['flow_stack'].to(device)
flow = flow.float()
output = model(flow).detach().cpu().numpy()
tmp_result.append(output)
# print(output.shape)
# if batch_idx == 1:
# print(batch_idx, ' : \n', output)
# print(batch_idx, ' : \n', output)
if batch_idx == 0:
break
ans = np.array(tmp_result).reshape(-1, 120)
print(ans)
print(ans.shape)
# with open('record_extract.txt', 'w') as f:
# f.write(fold_type + ' ' + str(fold_num) + ' ' + str(ans.shape))
# np.save('./data_inference/model_val_inference_result.npy', ans)
np.save('./data_inference/val_prediction.npy', ans)
def VideoSpatialPrediction3D_bert(
vid_name,
net,
num_categories,
architecture_name,
start_frame=0,
num_frames=0,
num_seg=4,
length = 16,
extension = 'img_{0:05d}.jpg',
ten_crop = False
):
if num_frames == 0:
imglist = os.listdir(vid_name)
newImageList=[]
if 'rgb' in architecture_name or 'pose' in architecture_name:
for item in imglist:
if 'img' in item:
newImageList.append(item)
elif 'flow' in architecture_name:
for item in imglist:
if 'flow_x' in item:
newImageList.append(item)
duration = len(newImageList)
else:
duration = num_frames
if 'rgb' in architecture_name:
if 'I3D' in architecture_name:
if not 'resnet' in architecture_name:
clip_mean = [0.5, 0.5, 0.5]
clip_std = [0.5, 0.5, 0.5]
else:
clip_mean = [0.45, 0.45, 0.45]
clip_std = [0.225, 0.225, 0.225]
normalize = video_transforms.Normalize(mean=clip_mean,
std=clip_std)
val_transform = video_transforms.Compose([
video_transforms.ToTensor(),
normalize,
])
if '112' in architecture_name:
scale = 0.5
else:
scale = 1
elif 'MFNET3D' in architecture_name:
clip_mean = [0.48627451, 0.45882353, 0.40784314]
clip_std = [0.234, 0.234, 0.234]
normalize = video_transforms.Normalize(mean=clip_mean,
std=clip_std)
val_transform = video_transforms.Compose([
video_transforms.ToTensor(),
normalize])
if '112' in architecture_name:
scale = 0.5
else:
scale = 1
elif 'tsm' in architecture_name:
clip_mean = [0.485, 0.456, 0.406]
clip_std = [0.229, 0.224, 0.225]
normalize = video_transforms.Normalize(mean=clip_mean,
std=clip_std)
val_transform = video_transforms.Compose([
video_transforms.ToTensor(),
normalize])
scale = 1
elif "r2plus1d" in architecture_name:
clip_mean = [0.43216, 0.394666, 0.37645]
clip_std = [0.22803, 0.22145, 0.216989]
normalize = video_transforms.Normalize(mean=clip_mean,
std=clip_std)
val_transform = video_transforms.Compose([
video_transforms.ToTensor(),
normalize])
scale = 0.5
elif 'rep_flow' in architecture_name:
clip_mean = [0.5, 0.5, 0.5]
clip_std = [0.5, 0.5, 0.5]
normalize = video_transforms.Normalize(mean=clip_mean,
std=clip_std)
val_transform = video_transforms.Compose([
video_transforms.ToTensor(),
normalize,
])
scale = 1
elif "slowfast" in architecture_name:
clip_mean = [0.45, 0.45, 0.45]
clip_std = [0.225, 0.225, 0.225]
normalize = video_transforms.Normalize(mean=clip_mean,
std=clip_std)
val_transform = video_transforms.Compose([
video_transforms.ToTensor(),
normalize,
])
scale = 1
else:
scale = 0.5
clip_mean = [114.7748, 107.7354, 99.4750]
clip_std = [1, 1, 1]
normalize = video_transforms.Normalize(mean=clip_mean,
std=clip_std)
val_transform = video_transforms.Compose([
video_transforms.ToTensor2(),
normalize,
])
elif 'flow' in architecture_name:
if 'I3D' in architecture_name:
clip_mean = [0.5] * 2
clip_std = [0.5] * 2
normalize = video_transforms.Normalize(mean=clip_mean,
std=clip_std)
val_transform = video_transforms.Compose([
video_transforms.ToTensor(),
normalize,
])
scale = 1
else:
scale = 0.5
clip_mean = [127.5, 127.5]
clip_std = [1, 1]
normalize = video_transforms.Normalize(mean=clip_mean,
std=clip_std)
val_transform = video_transforms.Compose([
video_transforms.ToTensor2(),
normalize,
])
# selection
#step = int(math.floor((duration-1)/(num_samples-1)))
if '224' in architecture_name:
scale = 1
if '112' in architecture_name:
scale = 0.5
imageSize=int(224 * scale)
dims = (int(256 * scale),int(340 * scale),3,duration)
duration = duration - 1
average_duration = int(duration / num_seg)
offsetMainIndexes = []
offsets = []
for seg_id in range(num_seg):
if average_duration >= length:
offsetMainIndexes.append(int((average_duration - length + 1)/2 + seg_id * average_duration))
elif duration >=length:
average_part_length = int(np.floor((duration-length)/num_seg))
offsetMainIndexes.append(int((average_part_length*(seg_id) + average_part_length*(seg_id+1))/2))
else:
increase = int(duration / num_seg)
offsetMainIndexes.append(0 + seg_id * increase)
for mainOffsetValue in offsetMainIndexes:
for lengthID in range(1, length+1):
loaded_frame_index = lengthID + mainOffsetValue
moded_loaded_frame_index = loaded_frame_index % (duration + 1)
if moded_loaded_frame_index == 0:
moded_loaded_frame_index = (duration + 1)
offsets.append(moded_loaded_frame_index)
imageList=[]
imageList1=[]
imageList2=[]
imageList3=[]
imageList4=[]
imageList5=[]
imageList6=[]
imageList7=[]
imageList8=[]
imageList9=[]
imageList10=[]
imageList11=[]
imageList12=[]
interpolation = cv2.INTER_LINEAR
for index in offsets:
if 'rgb' in architecture_name or 'pose' in architecture_name:
img_file = os.path.join(vid_name, extension.format(index))
img = cv2.imread(img_file, cv2.IMREAD_UNCHANGED)
img = cv2.resize(img, dims[1::-1],interpolation)
#img2 = cv2.resize(img, dims2[1::-1],interpolation)
img = cv2.cvtColor(img, cv2.COLOR_BGR2RGB)
img_flip = img[:,::-1,:].copy()
elif 'flow' in architecture_name:
flow_x_file = os.path.join(vid_name, extension.format('x',index))
flow_y_file = os.path.join(vid_name, extension.format('y',index))
img_x = cv2.imread(flow_x_file, cv2.IMREAD_GRAYSCALE)
img_y = cv2.imread(flow_y_file, cv2.IMREAD_GRAYSCALE)
img_x = np.expand_dims(img_x,-1)
img_y = np.expand_dims(img_y,-1)
img = np.concatenate((img_x,img_y),2)
img = cv2.resize(img, dims[1::-1],interpolation)
img_flip = img[:,::-1,:].copy()
#img_flip2 = img2[:,::-1,:].copy()
imageList1.append(img[int(16 * scale):int(16 * scale + imageSize), int(58 * scale) : int(58 * scale + imageSize), :])
imageList2.append(img[:imageSize, :imageSize, :])
imageList3.append(img[:imageSize, -imageSize:, :])
imageList4.append(img[-imageSize:, :imageSize, :])
imageList5.append(img[-imageSize:, -imageSize:, :])
imageList6.append(img_flip[int(16 * scale):int(16 * scale + imageSize), int(58 * scale) : int(58 * scale + imageSize), :])
imageList7.append(img_flip[:imageSize, :imageSize, :])
imageList8.append(img_flip[:imageSize, -imageSize:, :])
imageList9.append(img_flip[-imageSize:, :imageSize, :])
imageList10.append(img_flip[-imageSize:, -imageSize:, :])
# imageList11.append(img2)
# imageList12.append(img_flip2)
if ten_crop:
imageList=imageList1+imageList2+imageList3+imageList4+imageList5+imageList6+imageList7+imageList8+imageList9+imageList10
else:
imageList=imageList1
#imageList=imageList11+imageList12
rgb_list=[]
for i in range(len(imageList)):
cur_img = imageList[i]
cur_img_tensor = val_transform(cur_img)
rgb_list.append(np.expand_dims(cur_img_tensor.numpy(), 0))
input_data=np.concatenate(rgb_list,axis=0)
with torch.no_grad():
imgDataTensor = torch.from_numpy(input_data).type(torch.FloatTensor).cuda()
if 'rgb' in architecture_name or 'pose' in architecture_name:
imgDataTensor = imgDataTensor.view(-1,length,3,imageSize,imageSize).transpose(1,2)
elif 'flow' in architecture_name:
imgDataTensor = imgDataTensor.view(-1,length,2,imageSize,imageSize).transpose(1,2)
if 'bert' in architecture_name or 'pooling' in architecture_name:
output, input_vectors, sequenceOut, maskSample = net(imgDataTensor)
else:
output = net(imgDataTensor)
# outputSoftmax=soft(output)
result = output.data.cpu().numpy()
mean_result=np.mean(result,0)
prediction=np.argmax(mean_result)
top3 = mean_result.argsort()[::-1][:3]
return prediction, mean_result, top3
def main():
global args, best_prec1
args = parser.parse_args()
# create model
print("Building model ... ")
model = build_model()
print("Model %s is loaded. " % (args.modality + "_" + args.arch))
if not os.path.exists(args.resume):
os.makedirs(args.resume)
print("Saving everything to directory %s." % (args.resume))
# define loss function (criterion) and optimizer
criterion = nn.CrossEntropyLoss().cuda()
optimizer = torch.optim.SGD(model.parameters(),
args.lr,
momentum=args.momentum,
weight_decay=args.weight_decay)
cudnn.benchmark = True
# Data transforming
clip_mean = [0.485, 0.456, 0.406] * args.new_length
clip_std = [0.229, 0.224, 0.225] * args.new_length
normalize = video_transforms.Normalize(mean=clip_mean, std=clip_std)
if args.modality == "rgb":
scale_ratios = [1.0, 0.875, 0.75, 0.66]
elif args.modality == "flow":
scale_ratios = [1.0, 0.875, 0.75]
else:
print("No such modality. Only rgb and flow supported.")
train_transform = video_transforms.Compose([
video_transforms.Scale((256)),
video_transforms.MultiScaleCrop((224, 224), scale_ratios),
video_transforms.RandomHorizontalFlip(),
video_transforms.ToTensor(),
normalize,
])
val_transform = video_transforms.Compose([
video_transforms.Scale((256)),
video_transforms.CenterCrop((224)),
video_transforms.ToTensor(),
normalize,
])
# data loading
train_setting_file = "train_%s_split%d.txt" % (args.modality, args.split)
train_split_file = os.path.join(args.settings, args.dataset,
train_setting_file)
val_setting_file = "val_%s_split%d.txt" % (args.modality, args.split)
val_split_file = os.path.join(args.settings, args.dataset,
val_setting_file)
if not os.path.exists(train_split_file) or not os.path.exists(
val_split_file):
print(
"No split file exists in %s directory. Preprocess the dataset first"
% (args.settings))
train_dataset = datasets.__dict__[args.dataset](
args.data,
train_split_file,
"train",
args.new_length,
video_transform=train_transform)
val_dataset = datasets.__dict__[args.dataset](
args.data,
val_split_file,
"val",
args.new_length,
video_transform=val_transform)
print('{} samples found, {} train samples and {} test samples.'.format(
len(val_dataset) + len(train_dataset), len(train_dataset),
len(val_dataset)))
train_loader = torch.utils.data.DataLoader(train_dataset,
batch_size=args.batch_size,
shuffle=True,
num_workers=args.workers,
pin_memory=True)
val_loader = torch.utils.data.DataLoader(val_dataset,
batch_size=args.batch_size,
shuffle=True,
num_workers=args.workers,
pin_memory=True)
if args.evaluate:
validate(val_loader, model, criterion)
return
for epoch in range(args.start_epoch, args.epochs):
adjust_learning_rate(optimizer, epoch)
# train for one epoch
train(train_loader, model, criterion, optimizer, epoch)
# evaluate on validation set
prec1 = validate(val_loader, model, criterion)
# remember best prec@1 and save checkpoint
is_best = prec1 > best_prec1
best_prec1 = max(prec1, best_prec1)
if (epoch + 1) % args.save_freq == 0:
checkpoint_name = "%03d_%s" % (epoch + 1, "checkpoint.pth.tar")
save_checkpoint(
{
'epoch': epoch + 1,
'arch': args.arch,
'state_dict': model.state_dict(),
'best_prec1': best_prec1,
'optimizer': optimizer.state_dict(),
}, is_best, checkpoint_name, args.resume)
def main():
global args
args = parser.parse_args()
if args.dataset == 'ucf101':
num_categories = 101
elif args.dataset == 'hmdb51':
num_categories = 51
elif args.dataset == 'kinetics':
num_categories = 400
else:
raise ValueError('Unknown dataset ' + args.dataset)
start_frame = 0
model_start_time = time.time()
params = torch.load(args.weights)
#hard code
net = models.__dict__[args.arch](pretrained=False,
num_classes=num_categories)
net.load_state_dict(params['state_dict'])
net.cuda()
net.eval()
model_end_time = time.time()
model_time = model_end_time - model_start_time
print("Action recognition model is loaded in %4.4f seconds." %
(model_time))
###
if args.modality == "rgb":
new_length = 1
is_color = True
scale_ratios = [1.0, 0.875, 0.75, 0.66]
clip_mean = [0.485, 0.456, 0.406] * new_length
clip_std = [0.229, 0.224, 0.225] * new_length
elif args.modality == "flow":
new_length = 10
is_color = False
scale_ratios = [1.0, 0.875, 0.75]
clip_mean = [0.5, 0.5] * new_length
clip_std = [0.226, 0.226] * new_length
normalize = video_transforms.Normalize(mean=clip_mean, std=clip_std)
val_transform = video_transforms.Compose([
video_transforms.GroupCenterCrop(net.input_size),
video_transforms.CenterCrop((224)),
video_transforms.ToTensor(),
normalize,
])
dataset = datasets.load_clip(root=args.data,
source=args.test_list,
phase="val",
modality=args.modality,
is_color=is_color,
new_length=new_length,
new_width=args.new_width,
new_height=args.new_height,
video_transform=val_transform)
data_loader = torch.utils.data.DataLoader(dataset,
batch_size=1,
shuffle=False,
num_workers=args.workers * 2,
pin_memory=True)
data_gen = enumerate(data_loader)
total_num = len(data_loader.dataset)
output = []
def eval_video(video_data):
i, data, label = video_data
num_crop = 1
if args.modality == 'rgb':
length = 3
elif args.modality == 'flow':
length = 10
elif args.modality == 'RGBDiff':
length = 18
else:
raise ValueError("Unknown modality " + args.modality)
input_var = torch.autograd.Variable(data.view(-1, length, data.size(2),
data.size(3)),
volatile=True)
input_var = input_var.type(torch.FloatTensor).cuda()
rst = net(input_var).data.cpu().numpy().copy()
return i, rst.reshape(
(num_crop, args.test_segments,
num_categories)).mean(axis=0).reshape(
(args.test_segments, 1, num_categories)), label[0]
proc_start_time = time.time()
max_num = args.max_num if args.max_num > 0 else len(data_loader.dataset)
for i, (data, label) in data_gen:
data = data.float().cuda(async=True)
label = label.cuda(async=True)
input_var = torch.autograd.Variable(data, volatile=True)
target_var = torch.autograd.Variable(label, volatile=True)
rst = net(input_var).data.cpu().numpy()
#avg_pred_fc8 = np.mean(rst, axis=1)
# print(avg_spatial_pred_fc8.shape)
#result_list.append(avg_pred_fc8)
# avg_spatial_pred = softmax(avg_spatial_pred_fc8)
pred_index = np.argmax(rst)
#print (label.cpu().numpy())
#print (pred_index)
# print(rst)
# if i >= max_num:
# break
# rst = eval_video((i, data, label))
output.append(rst)
if label.cpu().numpy()[0] == pred_index:
match_count += 1
cnt_time = time.time() - proc_start_time
print('video {} done, total {}/{}, average {} sec/video'.format(
i, i + 1, total_num,
float(cnt_time) / (i + 1)))
print(match_count)
print(total_num)
print("Accuracy is %4.4f" % (float(match_count) / float(total_num)))
np.save("{}_sX_{}_{}.npy".format(args.dataset, args.modality, args.arch),
np.array(output)) #hard code