def main_iit_v2c():
# Parameters
config = FEConfig()
model_names = ['resnet50']
annotation_files = ['train.txt', 'test.txt']
for annotation_file in annotation_files:
annotations = iit_v2c.load_annotations(config.DATASET_PATH, annotation_file)
# Get torch.dataset object
clips, targets, vocab, config = iit_v2c.parse_dataset(config,
annotation_file,
numpy_features=False)
config.display()
transform = transforms.Compose([transforms.Resize((224, 224)),
transforms.ToTensor(),
transforms.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225])])
image_dataset = iit_v2c.FeatureDataset(clips,
targets,
numpy_features=False,
transform=transform)
for model_name in model_names:
extract(config.DATASET_PATH, image_dataset, model_name)
ROOT_DIR = ROOT_DIR
CHECKPOINT_PATH = os.path.join(ROOT_DIR, 'checkpoints')
DATASET_PATH = os.path.join(ROOT_DIR, 'datasets', 'IIT-V2C')
MAXLEN = 10
# Setup configuration class
config = TestConfig()
# Setup tf.dataset object
vocab = pickle.load(
open(os.path.join(config.CHECKPOINT_PATH, 'vocab.pkl'), 'rb'))
annotation_file = config.MODE + '.txt'
clips, targets, _, config = iit_v2c.parse_dataset(config,
annotation_file,
vocab=vocab)
test_dataset = iit_v2c.FeatureDataset(clips, targets)
test_loader = data.DataLoader(test_dataset,
batch_size=config.BATCH_SIZE,
shuffle=False,
num_workers=config.WORKERS)
config.display()
# Setup and build video2command training inference
v2c_model = Video2Command(config)
v2c_model.build()
# Safely create prediction dir if non-exist
if not os.path.exists(os.path.join(config.CHECKPOINT_PATH, 'prediction')):
os.makedirs(os.path.join(config.CHECKPOINT_PATH, 'prediction'))
# Start evaluating
DATASET_PATH = os.path.join(ROOT_DIR, 'datasets', 'IIT-V2C')
MAXLEN = 10
# Test configuration
config = TrainConfig()
config.display()
print()
# Test parse_dataset
annotation_file = config.MODE + '.txt'
clips, targets, vocab, config = iit_v2c.parse_dataset(config, annotation_file, numpy_features=False)
config.display()
print('Vocabulary:')
print(vocab.word2idx)
print('length ("<pad>" included):', len(vocab))
print('dataset:', len(clips), len(targets))
print()
transform = transforms.Compose([transforms.Resize(224),
transforms.ToTensor()])
train_dataset = iit_v2c.FeatureDataset(clips,
targets,
numpy_features=False,
transform=transform)
# Test torch dataloader object
for i, (Xv, S, clip_name) in enumerate(train_dataset):
print(Xv.shape, S.shape, clip_name)
break
config = TrainConfig()
config.display()
print()
# Test parse_dataset
annotation_file = config.MODE + '.txt'
clips, targets, vocab, config = iit_v2c.parse_dataset(config, annotation_file)
config.display()
print('Vocabulary:')
print(vocab.word2idx)
print('length ("<#SPECIAL>" included):', len(vocab))
print('dataset:', len(clips), len(targets))
print()
train_dataset = iit_v2c.FeatureDataset(clips, targets)
train_loader = data.DataLoader(train_dataset,
batch_size=config.BATCH_SIZE,
shuffle=True,
num_workers=config.WORKERS)
# Test torch dataloader object
for i, (Xv, S, clip_name) in enumerate(train_loader):
print(Xv.shape, S.shape, clip_name)
break
# Test parse_dataset on test_dataset
config.MODE = 'test'
annotation_file = config.MODE + '.txt'
clips_test, targets_test, vocab_test, config = iit_v2c.parse_dataset(config, annotation_file, vocab=vocab)
print(vocab == vocab_test)
test_dataset = iit_v2c.FeatureDataset(clips_test, targets_test)