def fvd(generated_embd, reference_embd):
with tf.Graph().as_default():
generated_embd = tf.convert_to_tensor(generated_embd, dtype=np.float32)
reference_embd = tf.convert_to_tensor(reference_embd, dtype=np.float32)
result = calculate_fvd(generated_embd, reference_embd)
with tf.Session() as sess:
sess.run(tf.global_variables_initializer())
sess.run(tf.tables_initializer())
print("FVD is: %.2f." % sess.run(result))
def fvd_session(predictions, target):
# 40, 10, 1, h, w
# with tf.Graph().as_default():
NUMBER_OF_VIDEOS, VIDEO_LENGTH, FRAME_WIDTH, FRAME_HEIGHT, C = predictions.shape
# with values in 0-255
assert C == 3
result = fvd.calculate_fvd(
fvd.create_id3_embedding(fvd.preprocess(predictions, (224, 224))),
fvd.create_id3_embedding(fvd.preprocess(target, (224, 224))))
with tf.Session() as sess:
sess.run(tf.global_variables_initializer())
sess.run(tf.tables_initializer())
print("FVD is: %.2f." % sess.run(result))
return result
def main(argv):
del argv
with tf.Graph().as_default():
first_set_of_videos = tf.zeros(
[NUMBER_OF_VIDEOS, VIDEO_LENGTH, 64, 64, 3])
second_set_of_videos = tf.ones(
[NUMBER_OF_VIDEOS, VIDEO_LENGTH, 64, 64, 3]) * 255
result = calculate_fvd(
create_id3_embedding(preprocess(first_set_of_videos, (224, 224))),
create_id3_embedding(preprocess(second_set_of_videos, (224, 224))))
with tf.Session() as sess:
sess.run(tf.global_variables_initializer())
sess.run(tf.tables_initializer())
print("FVD is: %.2f." % sess.run(result))
def main(argv):
video1 = torch.zeros(16, 15, 64, 64, 3)
video2 = torch.ones(16, 15, 64, 64, 3) * 255
del argv
with tf.Graph().as_default():
first_set_of_videos = tf.convert_to_tensor(video1.numpy(), np.uint8)
# first_set_of_videos = tf.zeros([NUMBER_OF_VIDEOS, VIDEO_LENGTH, 64, 64, 3])
second_set_of_videos = tf.convert_to_tensor(video2.numpy(), np.uint8)
# second_set_of_videos = tf.ones([NUMBER_OF_VIDEOS, VIDEO_LENGTH, 64, 64, 3]) * 255
result = fvd.calculate_fvd(
fvd.create_id3_embedding(
fvd.preprocess(first_set_of_videos, (224, 224))),
fvd.create_id3_embedding(
fvd.preprocess(second_set_of_videos, (224, 224))))
with tf.Session() as sess:
sess.run(tf.global_variables_initializer())
sess.run(tf.tables_initializer())
print("FVD is: %.2f." % sess.run(result))
def fvd(generated_images, reference_images):
with tf.Graph().as_default():
print (generated_images.shape)
print (reference_images.shape)
print (type(generated_images))
print (type(reference_images))
generated_images = tf.convert_to_tensor(generated_images, dtype=np.uint8)
print("Converted to tensor generated_images ")
reference_images = tf.convert_to_tensor(reference_images, dtype=np.uint8)
print("Converted to tensor reference images")
print (tf.shape(generated_images))
print (tf.shape(reference_images))
result = calculate_fvd(
create_id3_embedding(preprocess(generated_images,
(256, 256))),
create_id3_embedding(preprocess(reference_images,
(256, 256))))
with tf.Session() as sess:
sess.run(tf.global_variables_initializer())
sess.run(tf.tables_initializer())
print("FVD is: %.2f." % sess.run(result))
import train_vae
import utils
# Number of videos must be divisible by 16.
NUMBER_OF_VIDEOS = 16
VIDEO_LENGTH = 32
rn = "golf"
data_loader = dataloader.DataLoader()
rp = data_loader.shuffle_data()
vid, np_idx_new = data_loader.get_batch(0, rp)
videos = vid.permute(0, 1, 3, 4, 2).float()
videos = vid.permute(0, 1, 3, 4, 2).float().cpu().detach().numpy()
videos = tf.convert_to_tensor(videos, dtype=tf.float32)
generated = train_vae.main() #Choose a model
generated_vid = generated.permute(0, 1, 3, 4, 2).float().cpu().detach().numpy()
generated_vid = tf.slice(generated_vid, [0, 0, 0, 0, 0],
[NUMBER_OF_VIDEOS, 32, 64, 64, 3])
generated_vid = tf.convert_to_tensor(generated_vid, dtype=tf.float32)
result = fvd.calculate_fvd(
fvd.create_id3_embedding(fvd.preprocess(videos, (224, 224))),
fvd.create_id3_embedding(fvd.preprocess(generated_vid, (224, 224))))
with tf.Session() as sess:
sess.run(tf.global_variables_initializer())
sess.run(tf.tables_initializer())
print("FVD is: %.2f." % sess.run(result))
def main(argv):
args = argv[0]
# read file lists from directories
dir0_gif_paths = [f for f in os.listdir(args.dir0) if f.endswith('.gif')]
dir0_gif_paths = [os.path.join(args.dir0, f) for f in dir0_gif_paths]
dir1_gif_paths = [f for f in os.listdir(args.dir1) if f.endswith('.gif')]
dir1_gif_paths = [os.path.join(args.dir1, f) for f in dir1_gif_paths]
# assert number of videos to be divisible by 16
remainder_dir0 = len(dir0_gif_paths) % VIDEO_BATCH_SIZE
dir0_gif_paths = dir0_gif_paths[:-remainder_dir0]
remainder_dir1 = len(dir1_gif_paths) % VIDEO_BATCH_SIZE
dir1_gif_paths = dir1_gif_paths[:-remainder_dir1]
# loop over video dirs in batches of 16, compute and assemble activations (id3_embedding)
dir0_embeddings, dir1_embeddings = [], []
# graph0_initialized, graph1_initialized = False, False
dir0_embeddings_file = os.path.join(args.dir0, 'id3_embeddings.npy')
dir1_embeddings_file = os.path.join(args.dir1, 'id3_embeddings.npy')
# --- dir0 ID3 embeddings
if os.path.exists(dir0_embeddings_file):
with open(dir0_embeddings_file, 'rb') as fp:
dir0_embeddings = np.load(fp)
print(
f">>> Found stored ID3 activations for videos in {args.dir0} in {dir0_embeddings_file}."
)
else:
print(f">>> Computing ID3 activations for videos in {args.dir0}...")
for batch_start_idx in tqdm(
range(0, len(dir0_gif_paths), VIDEO_BATCH_SIZE)):
with tf.Graph().as_default():
# load batch of videos from GIFs and represent as tensor
dir0_videos = tf.stack(
[tf.io.decode_gif(tf.io.read_file(f)) \
for f in dir0_gif_paths[batch_start_idx:batch_start_idx+VIDEO_BATCH_SIZE]])
with tf.Session() as sess:
dir0_tensor = sess.run(dir0_videos)
# define placeholder for subsequent feeding
ph_dir0_videos = tf.placeholder(shape=[*dir0_tensor.shape],
dtype=tf.uint8)
# calculate embeddings
id3_embeddings = fvd.create_id3_embedding(
fvd.preprocess(ph_dir0_videos, (224, 224)))
with tf.Session() as sess:
sess.run(tf.global_variables_initializer())
sess.run(tf.tables_initializer())
dir0_embeddings.append(
sess.run(id3_embeddings,
feed_dict={ph_dir0_videos: dir0_tensor}))
dir0_embeddings = np.concatenate(dir0_embeddings, axis=0)
with open(dir0_embeddings_file, 'wb') as fp:
np.save(fp, dir0_embeddings)
print(
f">>> Saved ID3 embeddings for lookup in {dir0_embeddings_file}"
)
print(f">>> Embedding matrix: {dir0_embeddings.shape}")
# --- dir1 ID3 embeddings
if os.path.exists(dir1_embeddings_file):
with open(dir1_embeddings_file, 'rb') as fp:
dir1_embeddings = np.load(fp)
print(
f">>> Found stored ID3 activations for videos in {args.dir1} in {dir1_embeddings_file}."
)
else:
print(f">>> Computing ID3 activations for videos in {args.dir1}...")
for batch_start_idx in tqdm(
range(0, len(dir1_gif_paths), VIDEO_BATCH_SIZE)):
with tf.Graph().as_default():
# load batch of videos from GIFs and represent as tensor
dir1_videos = tf.stack(
[tf.io.decode_gif(tf.io.read_file(f)) \
for f in dir1_gif_paths[batch_start_idx:batch_start_idx+VIDEO_BATCH_SIZE]])
with tf.Session() as sess:
dir1_tensor = sess.run(dir1_videos)
# define placeholder for subsequent feeding
ph_dir1_videos = tf.placeholder(shape=[*dir1_tensor.shape],
dtype=tf.uint8)
# calculate embeddings
id3_embeddings = fvd.create_id3_embedding(
fvd.preprocess(ph_dir1_videos, (224, 224)))
with tf.Session() as sess:
sess.run(tf.global_variables_initializer())
sess.run(tf.tables_initializer())
dir1_embeddings.append(
sess.run(id3_embeddings,
feed_dict={ph_dir1_videos: dir1_tensor}))
dir1_embeddings = np.concatenate(dir1_embeddings, axis=0)
with open(dir1_embeddings_file, 'wb') as fp:
np.save(fp, dir1_embeddings)
print(
f">>> Saved ID3 embeddings for lookup in {dir1_embeddings_file}"
)
print(f">>> Embedding matrix: {dir1_embeddings.shape}")
# --- final FVD
with tf.Graph().as_default():
print(">>> Computing FVD...")
result = fvd.calculate_fvd(dir0_embeddings, dir1_embeddings)
with tf.Session() as sess:
print(">>> FVD is: %.2f." % sess.run(result))