def get_predictions_for_8_frames(net, frames):
a_t = time.time()
args_arch = "BNInception"
transform = torchvision.transforms.Compose([
transforms.GroupOverSample(net.input_size, net.scale_size),
transforms.Stack(roll=(args_arch in ['BNInception', 'InceptionV3'])),
transforms.ToTorchFormatTensor(
div=(args_arch not in ['BNInception', 'InceptionV3'])),
transforms.GroupNormalize(net.input_mean, net.input_std),
])
data = transform(frames)
input = data.view(-1, 3, data.size(1), data.size(2)).unsqueeze(0)
with torch.no_grad():
logits = net(input)
torch.onnx.export(net,
input,
"plm.onnx",
verbose=True,
input_names=["input"],
output_names=["output"])
h_x = torch.mean(F.softmax(logits, 1), dim=0).data
probs, idx = h_x.sort(0, True)
b_t = time.time()
print(f'Elapsed: {b_t - a_t}')
# Output the prediction.
for i in range(0, 5):
print('{:.3f} -> {}'.format(probs[i], categories[idx[i]]))
def __init__(self, root,
hierachy=1,
rescale_size=256,
crop_size=224,
seed=1):
if not os.path.exists(root):
raise ValueError('Root doest not exist')
self.root = root
if not hierachy in [1, 2]:
raise ValueError('Bad `hierachy`, must be one of 1, 2.')
self.hierachy = hierachy
if self.hierachy == 1:
self.videos = [os.path.join(self.root, v) for v in os.listdir(self.root)]
else:
cls = os.listdir(self.root)
self.videos = [os.path.join(self.root, c, v) for c in cls for v in os.listdir(os.path.join(self.root, c))]
self.videos = sorted(self.videos, key=str.lower)
self.rescale_size = rescale_size
self.crop_size = crop_size
self.seed = seed
random.seed(self.seed)
np.random.seed(self.seed)
self.offset = 0
# self.trans = torchvision.transforms.Compose([
# torchvision.transforms.Resize(self.rescale_size, interpolation=Image.BILINEAR),
# torchvision.transforms.CenterCrop(self.crop_size),
# torchvision.transforms.ToTensor(),
# torchvision.transforms.Normalize(
# mean=[.485, .456, .406],
# std=[.229, .224, .225])
# ])
self.trans = torchvision.transforms.Compose([
transforms.GroupScale(self.rescale_size, interpolation=Image.BILINEAR),
transforms.GroupCenterCrop(self.crop_size),
transforms.Stack4d(roll=False),
transforms.ToTorchFormatTensor4d(div=True),
transforms.GroupNormalize(
mean=[.5, .5, .5],
std=[.5, .5, .5])
])
del base_dict[key]
#print(base_dict)
"""
#net.load_state_dict(base_dict, strict=False)
net.load_state_dict(checkpoint, strict=True)
#print(net)
#exit(0)
net.eval()
net.cuda()
# Initialize frame transforms.
transform = torchvision.transforms.Compose([
transforms.GroupOverSample(net.module.input_size, net.module.scale_size),
transforms.Stack(roll=(args.arch in ['BNInception', 'InceptionV3'])),
transforms.ToTorchFormatTensor(div=(args.arch not in ['BNInception', 'InceptionV3'])),
transforms.GroupNormalize(net.module.input_mean, net.module.input_std),
])
segments_gt = [0, 0, 1, 1, 0, 0, 0,
0, 0, 1, 1, 1, 1, 0,
1, 0, 0, 0, 0, 0, 0,
1, 1, 1, 0, 0, 0, 0,
1, 1, 1, 0, 0, 1, 1,
2, 2, 0, 0, 1, 1, 1,
0, 0, 0, 0, 2]
pred = [2]* len(segments_gt)
video_dir = 'segments_2_slow/*.mp4'
for video_file_name in sorted(glob.glob(video_dir)):
print('best acc : {}, best cp : {}'.format(best_acc, best_cp))
def main(argv):
# Read arguments passed
(opts, args) = parser.parse_args(argv)
# Reading config
cfg = config(opts.config,
debugging=False,
additionalText="training_ERM_seen_resnet18")
# Use CUDA
# os.environ['CUDA_VISIBLE_DEVICES'] = 1
use_cuda = torch.cuda.is_available()
# If the manual seed is not yet choosen
if cfg.manualSeed == None:
cfg.manualSeed = 1
# Set seed for reproducibility for CPU and GPU randomizaton process
random.seed(cfg.manualSeed)
torch.manual_seed(cfg.manualSeed)
if use_cuda:
torch.cuda.manual_seed_all(cfg.manualSeed)
dataloader_train = None
if hasattr(cfg, "train_mode"):
# Preprocessing (transformation) instantiation for training groupwise
transformation_train = torchvision.transforms.Compose([
transforms.GroupMultiScaleCrop(224, [1, 0.875, 0.75, 0.66]),
transforms.GroupRandomHorizontalFlip(is_flow=False),
transforms.Stack(), # concatenation of images
transforms.ToTorchFormatTensor(), # to torch
transforms.GroupNormalize(mean=[0.485, 0.456, 0.406],
std=[0.229, 0.224,
0.225]), # Normalization
])
if cfg.algo == "ERM" or cfg.algo == "MTGA":
# Loading training Dataset with N segment for TSN
EPICdata_train = EPIC(
mode=cfg.train_mode,
cfg=cfg,
transforms=transformation_train,
)
# Creating validation dataloader
# batch size = 16, num_workers = 8 are best fit for 12 Gb GPU and >= 16 Gb RAM
dataloader_train = DataLoader(
EPICdata_train,
batch_size=cfg.train_batch_size,
shuffle=True,
num_workers=cfg.num_worker_train,
pin_memory=True,
)
elif cfg.algo == "IRM":
df = pd.read_csv(cfg.anno_path)
p_ids = list(set(df["participant_id"].tolist()))
dataloader_train = []
for p_id in p_ids:
tmp_dataset = EPIC(
mode=cfg.train_mode,
cfg=cfg,
transforms=transformation_train,
participant_id=p_id,
)
if tmp_dataset.haveData:
dataloader_train.append(
DataLoader(
tmp_dataset,
batch_size=cfg.train_batch_size,
shuffle=True,
num_workers=cfg.num_worker_train,
pin_memory=True,
))
elif cfg.algo == "FSL":
dataloader_train = {}
# Loading training Dataset with N segment for TSN
EPICdata_train_verb = EPIC(mode=cfg.train_mode,
cfg=cfg,
transforms=transformation_train)
sampler = CategoriesSampler(EPICdata_train_verb.verb_label, 200,
cfg.way, cfg.shot + cfg.query)
dataloader_train["verb"] = DataLoader(
dataset=EPICdata_train_verb,
batch_sampler=sampler,
num_workers=cfg.num_worker_train,
pin_memory=True,
)
EPICdata_train_noun = EPIC(mode=cfg.train_mode,
cfg=cfg,
transforms=transformation_train)
sampler = CategoriesSampler(EPICdata_train_noun.noun_label, 200,
cfg.way, cfg.shot + cfg.query)
dataloader_train["noun"] = DataLoader(
dataset=EPICdata_train_noun,
batch_sampler=sampler,
num_workers=cfg.num_worker_train,
pin_memory=True,
)
dataloader_val = None
if hasattr(cfg, "val_mode") and hasattr(cfg, "train_mode"):
# Preprocessing (transformation) instantiation for validation groupwise
transformation_val = torchvision.transforms.Compose([
transforms.GroupOverSample(
224, 256), # group sampling from images using multiple crops
transforms.Stack(), # concatenation of images
transforms.ToTorchFormatTensor(), # to torch
transforms.GroupNormalize(mean=[0.485, 0.456, 0.406],
std=[0.229, 0.224,
0.225]), # Normalization
])
# Loading validation Dataset with N segment for TSN
EPICdata_val = EPIC(
mode=cfg.val_mode,
cfg=cfg,
transforms=transformation_val,
)
# Creating validation dataloader
dataloader_val = DataLoader(
EPICdata_val,
batch_size=cfg.val_batch_size,
shuffle=False,
num_workers=cfg.num_worker_val,
pin_memory=True,
)
# Loading Models (Resnet50)
model = EPICModel(config=cfg)
if not cfg.feature_extraction:
if hasattr(cfg, "train_mode"):
policies = model.get_optim_policies()
# for group in policies:
# print(('group: {} has {} params, lr_mult: {}, decay_mult: {}'.format(
# group['name'], len(group['params']), group['lr_mult'], group['decay_mult'])))
# Optimizer
# initial lr = 0.01
# momentum = 0.9
# weight_decay = 5e-4
optimizer = torch.optim.SGD(policies,
lr=cfg.lr,
momentum=cfg.momentum,
weight_decay=cfg.weight_decay)
# Loss function (CrossEntropy)
if cfg.algo == "IRM":
criterion = torch.nn.CrossEntropyLoss(reduction="none")
elif cfg.algo == "ERM" or cfg.algo == "MTGA":
criterion = torch.nn.CrossEntropyLoss()
elif cfg.algo == "FSL":
criterion = torch.nn.CrossEntropyLoss()
# If multiple GPUs are available (and bridged)
# if torch.cuda.device_count() > 1:
# print("Let's use", torch.cuda.device_count(), "GPUs!")
# model = torch.nn.DataParallel(model)
# Convert model and loss function to GPU if available for faster computation
if use_cuda:
model = model.cuda()
criterion = criterion.cuda()
# Loading Trainer
experiment = Experiment(
cfg=cfg,
model=model,
loss=criterion,
optimizer=optimizer,
use_cuda=use_cuda,
data_train=dataloader_train,
data_val=dataloader_val,
debugging=False,
)
# Train the model
experiment.train()
else:
# Load Model Checkpoint
checkpoint = torch.load(cfg.checkpoint_filename_final)
model.load_state_dict(checkpoint["model_state_dict"])
if use_cuda:
model = model.cuda()
transformation = torchvision.transforms.Compose([
transforms.GroupOverSample(
224,
256), # group sampling from images using multiple crops
transforms.Stack(), # concatenation of images
transforms.ToTorchFormatTensor(), # to torch
transforms.GroupNormalize(mean=[0.485, 0.456, 0.406],
std=[0.229, 0.224,
0.225]), # Normalization
])
# Loading Predictor
experiment = Experiment(cfg=cfg,
model=model,
use_cuda=use_cuda,
debugging=False)
filenames = ["seen.json", "unseen.json"]
for filename in filenames:
EPICdata = EPIC(
mode=cfg.val_mode,
cfg=cfg,
transforms=transformation,
test_mode=filename[:-5],
)
data_loader = torch.utils.data.DataLoader(EPICdata,
batch_size=8,
shuffle=False,
num_workers=4,
pin_memory=True)
experiment.data_val = data_loader
experiment.predict(filename)
else:
# Load Model Checkpoint
checkpoint = torch.load(cfg.checkpoint_filename_final)
model.load_state_dict(checkpoint["model_state_dict"])
if use_cuda:
model = model.cuda()
model.eval()
transformation = torchvision.transforms.Compose([
transforms.GroupOverSample(
224, 256), # group sampling from images using multiple crops
transforms.Stack(), # concatenation of images
transforms.ToTorchFormatTensor(), # to torch
transforms.GroupNormalize(mean=[0.485, 0.456, 0.406],
std=[0.229, 0.224,
0.225]), # Normalization
])
# Loading Predictor
experiment = Experiment(cfg=cfg,
model=model,
use_cuda=use_cuda,
debugging=False)
with torch.no_grad():
modes = ["train-unseen", "val-unseen"]
for mode in modes:
data = np.empty((1, 2050))
EPICdata = EPIC(
mode=mode,
cfg=cfg,
transforms=transformation,
)
data_loader = torch.utils.data.DataLoader(EPICdata,
batch_size=1,
shuffle=False,
num_workers=0,
pin_memory=True)
for i, sample_batch in enumerate(data_loader):
output = experiment.extract_features(sample_batch)
verb_ann = sample_batch["verb_id"].data.item()
noun_ann = sample_batch["noun_id"].data.item()
out = np.append(np.mean(output, 0), verb_ann)
out = np.append(out, noun_ann)
data = np.concatenate((data, np.expand_dims(out, 0)), 0)
np.save(mode, data)
t2 = t_stack(t1)
print('t2', len(t2), t2.shape)
#
#----------------------------
t_start = time.time()
t_torch = transforms.ToTorchFormatTensor(div=False) # chw
t3 = t_torch(t2)
print('t3', t3.shape)
t_end = time.time()
t_diff = t_end - t_start
print('cost {}s'.format(t_diff)) #0.055351972579956055s
#----------------------------
t_norm = transforms.GroupNormalize(mean=[104, 117, 128], std=[1])
t4 = t_norm(t3)
print('t4', t4.shape)
#imshow(t2[:,:,:3])
#imshow(t1[0])
#print(t3[:3,:8,:8])
#plt.imshow(np.transpose(t3[:3,:,:], (1, 2,0)))
print(t3.size(0))
#---------------------------------------------------------------------------------
#print('-'*10,'pickle','-'*20)
#
xy_train = EpicLoad.read_pickle('epic_kitchens/train.pkl')
xy_test = EpicLoad.read_pickle('epic_kitchens/test.pkl')
xy_val = EpicLoad.read_pickle('epic_kitchens/val.pkl')
# Show all entrypoints and their help strings
for entrypoint in torch.hub.list(repo):
print(entrypoint)
print(torch.hub.help(repo, entrypoint))
cap = cv2.VideoCapture(videofile)
P_VAL = 0.85
# Initialize frame transforms.
transform = torchvision.transforms.Compose([
transforms.GroupScale(tsn.scale_size),
transforms.GroupCenterCrop(tsn.input_size),
transforms.Stack(roll=(args.arch in ['BNInception', 'InceptionV3'])),
transforms.ToTorchFormatTensor(
div=(args.arch not in ['BNInception', 'InceptionV3'])),
transforms.GroupNormalize(tsn.input_mean, tsn.input_std),
])
def load_frames(frames, num_frames=8):
"""
Convert video frame images to list of tensors
"""
if len(frames) >= num_frames:
return frames[::int(np.ceil(len(frames) / float(num_frames)))]
else:
raise (ValueError(
'Video must have at least {} frames'.format(num_frames)))
# check if CUDA is available
