def main():
args = parser.parse_args()
output_path = logging(args)
model_path = args.model_path
data_dir = args.data_dir
start_frame = 0
num_samples = 25 if args.modality[:3] == 'rgb' else 1
ext = ".png" if args.modality == "rhythm" and args.dataset == "hmdb51" else ".jpg"
num_categories = 51 if args.dataset == 'hmdb51' else 101
new_size = 224
model_start_time = time.time()
params = torch.load(model_path)
if args.architecture == "inception_v3":
new_size = 299
if args.modality == "rhythm":
spatial_net = models.flow_inception_v3(pretrained=False,
channels=1,
num_classes=num_categories)
else:
spatial_net = models.rgb_inception_v3(pretrained=False,
channels=3,
num_classes=num_categories)
else:
if args.modality == "rhythm":
spatial_net = models.flow_resnet152(pretrained=False,
channels=1,
num_classes=num_categories)
else:
spatial_net = models.rgb_resnet152(pretrained=False,
channels=3,
num_classes=num_categories)
spatial_net.load_state_dict(params['state_dict'])
spatial_net.cuda()
spatial_net.eval()
model_end_time = time.time()
model_time = model_end_time - model_start_time
print("Action recognition spatial model is loaded in %4.4f seconds." %
(model_time))
test_path = os.path.join(args.settings, args.dataset)
test_file = os.path.join(test_path,
"dataset_list.txt") if args.w else os.path.join(
test_path, "test_split%d.txt" % (args.split))
print(test_file)
f_test = open(test_file, "r")
test_list = f_test.readlines()
print("we got %d videos" % len(test_list))
line_id = 1
match_count = 0
result_list = []
lines = []
if args.vr_approach == 3:
direction_path = os.path.join(args.settings, args.dataset,
"direction.txt")
lines = [int(line.rstrip('\n')) for line in open(direction_path)]
elif args.vr_approach == 4:
direction_path = os.path.join(args.settings, args.dataset,
"direction_video.txt")
lines = {
line.split()[0]: int(line.split()[1])
for line in open(direction_path)
}
for line in test_list:
line_info = line.split(" ")
clip_path = os.path.join(data_dir, line_info[0])
num_frames = int(line_info[1])
input_video_label = int(line_info[2])
video_name = clip_path.split("/")[-1]
index = lines[input_video_label] if args.vr_approach == 3 else (
lines[video_name] if args.vr_approach == 4 else args.vr_approach)
spatial_prediction = VideoSpatialPrediction(args.modality,
clip_path,
spatial_net,
num_categories,
start_frame,
num_frames,
num_samples,
index,
new_size=new_size,
ext=ext)
avg_spatial_pred_fc8 = np.mean(spatial_prediction, axis=1)
result_list.append(avg_spatial_pred_fc8)
pred_index = np.argmax(avg_spatial_pred_fc8)
print(args.modality + " split " + str(args.split) +
", sample %d/%d: GT: %d, Prediction: %d ==> correct: %d" %
(line_id, len(test_list), input_video_label, pred_index,
match_count))
if pred_index == input_video_label:
match_count += 1
line_id += 1
print(match_count)
print(len(test_list))
print("Accuracy is: %4.4f" % (float(match_count) / len(test_list)))
npy_name = args.dataset + "_" + args.modality + "_" + args.architecture + "_s" + str(
args.split) + ".npy"
npy_path = os.path.join(output_path, npy_name)
np.save(npy_path, np.array(result_list))
def main():
args = parser.parse_args()
#model_path = '../../parameters/'+args.architecture+"/"+args.modality+'_s'+str(args.split)+'.pth.tar'
#data_dir = '../datasets/'+args.dataset+'_frames'
#data_dir = '/home/Datasets/UCF-101-OF_CPU'
model_path = args.model_path
data_dir = args.data_dir
start_frame = 0
if args.modality[:3] == 'rgb':
num_samples = 25
else:
num_samples = 1
num_categories = 51 if args.dataset == 'hmdb51' else 101
model_start_time = time.time()
params = torch.load(model_path)
new_size = 224
if args.architecture == "inception_v3":
new_size = 299
if args.modality == "rhythm":
spatial_net = models.flow_inception_v3(pretrained=False,
channels=1,
num_classes=num_categories)
else:
spatial_net = models.rgb_inception_v3(pretrained=False,
channels=3,
num_classes=num_categories)
else:
if args.modality == "rhythm":
spatial_net = models.flow_resnet152(pretrained=False,
channels=1,
num_classes=num_categories)
else:
spatial_net = models.rgb_resnet152(pretrained=False,
channels=3,
num_classes=num_categories)
spatial_net.load_state_dict(params['state_dict'])
spatial_net.cuda()
spatial_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))
val_file = "./splits/" + args.dataset + "/val_split%d.txt" % (args.split)
#val_file = 'spatial_testlist01_with_labels.txt'
f_val = open(val_file, "r")
val_list = f_val.readlines()
print("we got %d test videos" % len(val_list))
line_id = 1
match_count = 0
result = []
lines = [
int(line.rstrip('\n'))
for line in open('../datasets/settings/' + args.dataset +
'/direction.txt')
]
for line in val_list:
line_info = line.split(" ")
clip_path = os.path.join(data_dir, line_info[0])
num_frames = int(line_info[1])
input_video_label = int(line_info[2])
spatial_prediction = VideoSpatialPrediction(
args.modality, clip_path, spatial_net, num_categories, start_frame,
num_frames, num_samples, args.vr_approach
if args.vr_approach != 3 else lines[input_video_label], new_size)
avg_spatial_pred_fc8 = np.mean(spatial_prediction, axis=1)
result.append(avg_spatial_pred_fc8)
# avg_spatial_pred = softmax(avg_spatial_pred_fc8)
pred_index = np.argmax(avg_spatial_pred_fc8)
print(args.modality + " split " + str(args.split) +
", sample %d/%d: GT: %d, Prediction: %d ==> correct: %d" %
(line_id, len(val_list), input_video_label, pred_index,
match_count))
if pred_index == input_video_label:
match_count += 1
line_id += 1
print(match_count)
print(len(val_list))
print("Accuracy is : %4.4f" % ((float(match_count) / len(val_list))))
np.save(
args.dataset + "_" + args.modality + "_" + args.architecture + "_s" +
str(args.split) + ".npy", np.array(result))
def main():
args = parser.parse_args()
output_path = logging(args)
model_path = args.model_path
data_dir = args.data_dir
start_frame = 0
num_samples = 10
num_categories = 51 if args.dataset == 'hmdb51' else 101
model_start_time = time.time()
new_size = 224
if args.architecture == "inception_v3":
new_size = 299
spatial_net = models.flow_inception_v3(pretrained=False,
channels=1,
num_classes=num_categories)
else:
spatial_net = models.flow_resnet152(pretrained=False,
channels=1,
num_classes=num_categories)
params = torch.load(model_path)
spatial_net.load_state_dict(params['state_dict'])
spatial_net.cuda()
spatial_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))
test_path = os.path.join(args.settings, args.dataset)
test_file = os.path.join(test_path,
"dataset_list.txt") if args.w else os.path.join(
test_path, "test_split%d.txt" % (args.split))
f_test = open(test_file, "r")
test_list = f_test.readlines()
print("we got %d test videos" % len(test_list))
line_id = 1
match_count = 0
result = []
for line in test_list:
line_info = line.split(" ")
clip_path = os.path.join(data_dir, line_info[0])
num_frames = int(line_info[1])
input_video_label = int(line_info[2])
spatial_prediction = VideoSpatialPrediction(clip_path, spatial_net,
num_categories,
num_samples, new_size,
args.batch_size)
avg_spatial_pred_fc8 = np.mean(spatial_prediction, axis=1)
result.append(avg_spatial_pred_fc8)
# avg_spatial_pred = softmax(avg_spatial_pred_fc8)
pred_index = np.argmax(avg_spatial_pred_fc8)
print("Rhythm split " + str(args.split) +
", sample %d/%d: GT: %d, Prediction: %d ==> correct: %d" %
(line_id, len(test_list), input_video_label, pred_index,
match_count))
if pred_index == input_video_label:
match_count += 1
line_id += 1
print(match_count)
print(len(test_list))
print("Accuracy is : %4.4f" % ((float(match_count) / len(test_list))))
npy_name = args.dataset + "_rhythm_" + args.architecture + "_s" + str(
args.split) + ".npy"
npy_path = os.path.join(output_path, npy_name)
np.save(npy_path, np.array(result))
def main():
args = parser.parse_args()
#model_path = '../../parameters/'+args.architecture+'/'+args.modality+'_s'+str(args.split)+'.pth.tar'
#data_path = '/home/Datasets/UCF-101-OF_CPU'
model_path = args.model_path
data_path = args.data_dir
start_frame = 0
num_categories = 51 if args.dataset == 'hmdb51' else 101
new_size = 224
model_start_time = time.time()
params = torch.load(model_path)
if args.architecture == "inception_v3":
new_size = 299
temporal_net = models.flow_inception_v3(pretrained=False,
channels=20,
num_classes=num_categories)
else:
temporal_net = models.flow_resnet152(pretrained=False,
channels=20,
num_classes=num_categories)
temporal_net.load_state_dict(params['state_dict'])
temporal_net.cuda()
temporal_net.eval()
model_end_time = time.time()
model_time = model_end_time - model_start_time
print("Action recognition temporal model is loaded in %4.4f seconds." %
(model_time))
val_file = "./splits/" + args.dataset + "/val_split%d.txt" % (args.split)
#val_file = 'spatial_testlist01_with_labels.txt'
f_val = open(val_file, "r")
val_list = f_val.readlines()
print("we got %d test videos" % len(val_list))
line_id = 1
match_count = 0
result_list = []
for line in val_list:
line_info = line.split(" ")
clip_path = os.path.join(data_path, line_info[0])
num_frames = int(line_info[1])
input_video_label = int(line_info[2])
spatial_prediction = VideoTemporalPrediction(args.modality,
clip_path,
temporal_net,
num_categories,
start_frame,
num_frames,
new_size=new_size)
avg_spatial_pred_fc8 = np.mean(spatial_prediction, axis=1)
# print(avg_spatial_pred_fc8.shape)
result_list.append(avg_spatial_pred_fc8)
# avg_spatial_pred = softmax(avg_spatial_pred_fc8)
pred_index = np.argmax(avg_spatial_pred_fc8)
print(args.modality + " split " + str(args.split) +
", sample %d/%d: GT: %d, Prediction: %d" %
(line_id, len(val_list), input_video_label, pred_index))
if pred_index == input_video_label:
match_count += 1
line_id += 1
print(match_count)
print(len(val_list))
print("Accuracy is %4.4f" % (float(match_count) / len(val_list)))
np.save(
args.dataset + "_" + args.modality + "_resnet152_s" + str(args.split) +
".npy", np.array(result_list))