View on TensorFlow.org | Run in Google Colab | View on GitHub | Download notebook | See TF Hub model |
Yunpeng Li, Dominik Roblek, and Marco Tagliasacchi. From Here to There: Video Inbetweening Using Direct 3D Convolutions, 2019.
https://arxiv.org/abs/1905.10240
Current Hub characteristics:
- has models for BAIR Robot pushing videos and KTH action video dataset (though this colab uses only BAIR)
- BAIR dataset already available in Hub. However, KTH videos need to be supplied by the users themselves.
- only evaluation (video generation) for now
- batch size and frame size are hard-coded
Setup
Since tfds.load('bair_robot_pushing_small', split='test')
would download a 30GB archive that also contains the training data, we download a separated archive that only contains the 190MB test data. The used dataset has been published by this paper and is licensed as Creative Commons BY 4.0.
import tensorflow as tf
import matplotlib.pyplot as plt
import numpy as np
import seaborn as sns
import tensorflow_hub as hub
import tensorflow_datasets as tfds
from tensorflow_datasets.core import SplitGenerator
from tensorflow_datasets.video.bair_robot_pushing import BairRobotPushingSmall
import tempfile
import pathlib
TEST_DIR = pathlib.Path(tempfile.mkdtemp()) / "bair_robot_pushing_small/softmotion30_44k/test/"
# Download the test split to $TEST_DIR
mkdir -p $TEST_DIR
wget -nv https://storage.googleapis.com/download.tensorflow.org/data/bair_test_traj_0_to_255.tfrecords -O $TEST_DIR/traj_0_to_255.tfrecords
2024-03-09 12:55:09 URL:https://storage.googleapis.com/download.tensorflow.org/data/bair_test_traj_0_to_255.tfrecords [189852160/189852160] -> "/tmpfs/tmp/tmprci2g912/bair_robot_pushing_small/softmotion30_44k/test/traj_0_to_255.tfrecords" [1]
# Since the dataset builder expects the train and test split to be downloaded,
# patch it so it only expects the test data to be available
builder = BairRobotPushingSmall()
test_generator = SplitGenerator(name='test', gen_kwargs={"filedir": str(TEST_DIR)})
builder._split_generators = lambda _: [test_generator]
builder.download_and_prepare()
WARNING:absl:DEPRECATED! Do not use a DatasetBuilder class directly, but call `tfds.builder_cls('bair_robot_pushing_small')`.
BAIR: Demo based on numpy array inputs
# @title Load some example data (BAIR).
batch_size = 16
# If unable to download the dataset automatically due to "not enough disk space", please download manually to Google Drive and
# load using tf.data.TFRecordDataset.
ds = builder.as_dataset(split="test")
test_videos = ds.batch(batch_size)
first_batch = next(iter(test_videos))
input_frames = first_batch['image_aux1'][:, ::15]
input_frames = tf.cast(input_frames, tf.float32)
2024-03-09 12:55:10.586839: E external/local_xla/xla/stream_executor/cuda/cuda_driver.cc:282] failed call to cuInit: CUDA_ERROR_NO_DEVICE: no CUDA-capable device is detected WARNING:absl:`FeatureConnector.dtype` is deprecated. Please change your code to use NumPy with the field `FeatureConnector.np_dtype` or use TensorFlow with the field `FeatureConnector.tf_dtype`. WARNING:absl:`FeatureConnector.dtype` is deprecated. Please change your code to use NumPy with the field `FeatureConnector.np_dtype` or use TensorFlow with the field `FeatureConnector.tf_dtype`.
# @title Visualize loaded videos start and end frames.
print('Test videos shape [batch_size, start/end frame, height, width, num_channels]: ', input_frames.shape)
sns.set_style('white')
plt.figure(figsize=(4, 2*batch_size))
for i in range(batch_size)[:4]:
plt.subplot(batch_size, 2, 1 + 2*i)
plt.imshow(input_frames[i, 0] / 255.0)
plt.title('Video {}: First frame'.format(i))
plt.axis('off')
plt.subplot(batch_size, 2, 2 + 2*i)
plt.imshow(input_frames[i, 1] / 255.0)
plt.title('Video {}: Last frame'.format(i))
plt.axis('off')
Test videos shape [batch_size, start/end frame, height, width, num_channels]: (16, 2, 64, 64, 3)
Load Hub Module
hub_handle = 'https://tfhub.dev/google/tweening_conv3d_bair/1'
module = hub.load(hub_handle).signatures['default']
WARNING:tensorflow:Unable to create a python object for variable <tf.Variable 'video_discriminator/conv_0/weights:0' shape=(4, 4, 4, 3, 64) dtype=float32_ref> because it is a reference variable. It may not be visible to training APIs. If this is a problem, consider rebuilding the SavedModel after running tf.compat.v1.enable_resource_variables(). WARNING:tensorflow:Unable to create a python object for variable <tf.Variable 'video_discriminator/conv_0/weights:0' shape=(4, 4, 4, 3, 64) dtype=float32_ref> because it is a reference variable. It may not be visible to training APIs. If this is a problem, consider rebuilding the SavedModel after running tf.compat.v1.enable_resource_variables(). WARNING:tensorflow:Unable to create a python object for variable <tf.Variable 'video_discriminator/conv_0/LayerNorm/beta:0' shape=(64,) dtype=float32_ref> because it is a reference variable. It may not be visible to training APIs. If this is a problem, consider rebuilding the SavedModel after running tf.compat.v1.enable_resource_variables(). WARNING:tensorflow:Unable to create a python object for variable <tf.Variable 'video_discriminator/conv_0/LayerNorm/beta:0' shape=(64,) dtype=float32_ref> because it is a reference variable. It may not be visible to training APIs. If this is a problem, consider rebuilding the SavedModel after running tf.compat.v1.enable_resource_variables(). WARNING:tensorflow:Unable to create a python object for variable <tf.Variable 'video_discriminator/conv_0/LayerNorm/gamma:0' shape=(64,) dtype=float32_ref> because it is a reference variable. It may not be visible to training APIs. If this is a problem, consider rebuilding the SavedModel after running tf.compat.v1.enable_resource_variables(). WARNING:tensorflow:Unable to create a python object for variable <tf.Variable 'video_discriminator/conv_0/LayerNorm/gamma:0' shape=(64,) dtype=float32_ref> because it is a reference variable. It may not be visible to training APIs. If this is a problem, consider rebuilding the SavedModel after running tf.compat.v1.enable_resource_variables(). WARNING:tensorflow:Unable to create a python object for variable <tf.Variable 'video_discriminator/conv_1/weights:0' shape=(4, 4, 4, 64, 128) dtype=float32_ref> because it is a reference variable. It may not be visible to training APIs. If this is a problem, consider rebuilding the SavedModel after running tf.compat.v1.enable_resource_variables(). WARNING:tensorflow:Unable to create a python object for variable <tf.Variable 'video_discriminator/conv_1/weights:0' shape=(4, 4, 4, 64, 128) dtype=float32_ref> because it is a reference variable. It may not be visible to training APIs. If this is a problem, consider rebuilding the SavedModel after running tf.compat.v1.enable_resource_variables(). WARNING:tensorflow:Unable to create a python object for variable <tf.Variable 'video_discriminator/conv_1/LayerNorm/beta:0' shape=(128,) dtype=float32_ref> because it is a reference variable. It may not be visible to training APIs. If this is a problem, consider rebuilding the SavedModel after running tf.compat.v1.enable_resource_variables(). WARNING:tensorflow:Unable to create a python object for variable <tf.Variable 'video_discriminator/conv_1/LayerNorm/beta:0' shape=(128,) dtype=float32_ref> because it is a reference variable. It may not be visible to training APIs. If this is a problem, consider rebuilding the SavedModel after running tf.compat.v1.enable_resource_variables(). WARNING:tensorflow:Unable to create a python object for variable <tf.Variable 'video_discriminator/conv_0/weights:0' shape=(4, 4, 4, 3, 64) dtype=float32_ref> because it is a reference variable. It may not be visible to training APIs. If this is a problem, consider rebuilding the SavedModel after running tf.compat.v1.enable_resource_variables(). WARNING:tensorflow:Unable to create a python object for variable <tf.Variable 'video_discriminator/conv_0/weights:0' shape=(4, 4, 4, 3, 64) dtype=float32_ref> because it is a reference variable. It may not be visible to training APIs. If this is a problem, consider rebuilding the SavedModel after running tf.compat.v1.enable_resource_variables(). WARNING:tensorflow:Unable to create a python object for variable <tf.Variable 'video_discriminator/conv_0/LayerNorm/beta:0' shape=(64,) dtype=float32_ref> because it is a reference variable. It may not be visible to training APIs. If this is a problem, consider rebuilding the SavedModel after running tf.compat.v1.enable_resource_variables(). WARNING:tensorflow:Unable to create a python object for variable <tf.Variable 'video_discriminator/conv_0/LayerNorm/beta:0' shape=(64,) dtype=float32_ref> because it is a reference variable. It may not be visible to training APIs. If this is a problem, consider rebuilding the SavedModel after running tf.compat.v1.enable_resource_variables(). WARNING:tensorflow:Unable to create a python object for variable <tf.Variable 'video_discriminator/conv_0/LayerNorm/gamma:0' shape=(64,) dtype=float32_ref> because it is a reference variable. It may not be visible to training APIs. If this is a problem, consider rebuilding the SavedModel after running tf.compat.v1.enable_resource_variables(). WARNING:tensorflow:Unable to create a python object for variable <tf.Variable 'video_discriminator/conv_0/LayerNorm/gamma:0' shape=(64,) dtype=float32_ref> because it is a reference variable. It may not be visible to training APIs. If this is a problem, consider rebuilding the SavedModel after running tf.compat.v1.enable_resource_variables(). WARNING:tensorflow:Unable to create a python object for variable <tf.Variable 'video_discriminator/conv_1/weights:0' shape=(4, 4, 4, 64, 128) dtype=float32_ref> because it is a reference variable. It may not be visible to training APIs. If this is a problem, consider rebuilding the SavedModel after running tf.compat.v1.enable_resource_variables(). WARNING:tensorflow:Unable to create a python object for variable <tf.Variable 'video_discriminator/conv_1/weights:0' shape=(4, 4, 4, 64, 128) dtype=float32_ref> because it is a reference variable. It may not be visible to training APIs. If this is a problem, consider rebuilding the SavedModel after running tf.compat.v1.enable_resource_variables(). WARNING:tensorflow:Unable to create a python object for variable <tf.Variable 'video_discriminator/conv_1/LayerNorm/beta:0' shape=(128,) dtype=float32_ref> because it is a reference variable. It may not be visible to training APIs. If this is a problem, consider rebuilding the SavedModel after running tf.compat.v1.enable_resource_variables(). WARNING:tensorflow:Unable to create a python object for variable <tf.Variable 'video_discriminator/conv_1/LayerNorm/beta:0' shape=(128,) dtype=float32_ref> because it is a reference variable. It may not be visible to training APIs. If this is a problem, consider rebuilding the SavedModel after running tf.compat.v1.enable_resource_variables(). WARNING:tensorflow:Unable to create a python object for variable <tf.Variable 'video_discriminator/conv_0/weights:0' shape=(4, 4, 4, 3, 64) dtype=float32_ref> because it is a reference variable. It may not be visible to training APIs. If this is a problem, consider rebuilding the SavedModel after running tf.compat.v1.enable_resource_variables(). WARNING:tensorflow:Unable to create a python object for variable <tf.Variable 'video_discriminator/conv_0/weights:0' shape=(4, 4, 4, 3, 64) dtype=float32_ref> because it is a reference variable. It may not be visible to training APIs. If this is a problem, consider rebuilding the SavedModel after running tf.compat.v1.enable_resource_variables(). WARNING:tensorflow:Unable to create a python object for variable <tf.Variable 'video_discriminator/conv_0/LayerNorm/beta:0' shape=(64,) dtype=float32_ref> because it is a reference variable. It may not be visible to training APIs. If this is a problem, consider rebuilding the SavedModel after running tf.compat.v1.enable_resource_variables(). WARNING:tensorflow:Unable to create a python object for variable <tf.Variable 'video_discriminator/conv_0/LayerNorm/beta:0' shape=(64,) dtype=float32_ref> because it is a reference variable. It may not be visible to training APIs. If this is a problem, consider rebuilding the SavedModel after running tf.compat.v1.enable_resource_variables(). WARNING:tensorflow:Unable to create a python object for variable <tf.Variable 'video_discriminator/conv_0/LayerNorm/gamma:0' shape=(64,) dtype=float32_ref> because it is a reference variable. It may not be visible to training APIs. If this is a problem, consider rebuilding the SavedModel after running tf.compat.v1.enable_resource_variables(). WARNING:tensorflow:Unable to create a python object for variable <tf.Variable 'video_discriminator/conv_0/LayerNorm/gamma:0' shape=(64,) dtype=float32_ref> because it is a reference variable. It may not be visible to training APIs. If this is a problem, consider rebuilding the SavedModel after running tf.compat.v1.enable_resource_variables(). WARNING:tensorflow:Unable to create a python object for variable <tf.Variable 'video_discriminator/conv_1/weights:0' shape=(4, 4, 4, 64, 128) dtype=float32_ref> because it is a reference variable. It may not be visible to training APIs. If this is a problem, consider rebuilding the SavedModel after running tf.compat.v1.enable_resource_variables(). WARNING:tensorflow:Unable to create a python object for variable <tf.Variable 'video_discriminator/conv_1/weights:0' shape=(4, 4, 4, 64, 128) dtype=float32_ref> because it is a reference variable. It may not be visible to training APIs. If this is a problem, consider rebuilding the SavedModel after running tf.compat.v1.enable_resource_variables(). WARNING:tensorflow:Unable to create a python object for variable <tf.Variable 'video_discriminator/conv_1/LayerNorm/beta:0' shape=(128,) dtype=float32_ref> because it is a reference variable. It may not be visible to training APIs. If this is a problem, consider rebuilding the SavedModel after running tf.compat.v1.enable_resource_variables(). WARNING:tensorflow:Unable to create a python object for variable <tf.Variable 'video_discriminator/conv_1/LayerNorm/beta:0' shape=(128,) dtype=float32_ref> because it is a reference variable. It may not be visible to training APIs. If this is a problem, consider rebuilding the SavedModel after running tf.compat.v1.enable_resource_variables(). WARNING:tensorflow:Unable to create a python object for variable <tf.Variable 'video_discriminator/conv_0/weights:0' shape=(4, 4, 4, 3, 64) dtype=float32_ref> because it is a reference variable. It may not be visible to training APIs. If this is a problem, consider rebuilding the SavedModel after running tf.compat.v1.enable_resource_variables(). WARNING:tensorflow:Unable to create a python object for variable <tf.Variable 'video_discriminator/conv_0/weights:0' shape=(4, 4, 4, 3, 64) dtype=float32_ref> because it is a reference variable. It may not be visible to training APIs. If this is a problem, consider rebuilding the SavedModel after running tf.compat.v1.enable_resource_variables(). WARNING:tensorflow:Unable to create a python object for variable <tf.Variable 'video_discriminator/conv_0/LayerNorm/beta:0' shape=(64,) dtype=float32_ref> because it is a reference variable. It may not be visible to training APIs. If this is a problem, consider rebuilding the SavedModel after running tf.compat.v1.enable_resource_variables(). WARNING:tensorflow:Unable to create a python object for variable <tf.Variable 'video_discriminator/conv_0/LayerNorm/beta:0' shape=(64,) dtype=float32_ref> because it is a reference variable. It may not be visible to training APIs. If this is a problem, consider rebuilding the SavedModel after running tf.compat.v1.enable_resource_variables(). WARNING:tensorflow:Unable to create a python object for variable <tf.Variable 'video_discriminator/conv_0/LayerNorm/gamma:0' shape=(64,) dtype=float32_ref> because it is a reference variable. It may not be visible to training APIs. If this is a problem, consider rebuilding the SavedModel after running tf.compat.v1.enable_resource_variables(). WARNING:tensorflow:Unable to create a python object for variable <tf.Variable 'video_discriminator/conv_0/LayerNorm/gamma:0' shape=(64,) dtype=float32_ref> because it is a reference variable. It may not be visible to training APIs. If this is a problem, consider rebuilding the SavedModel after running tf.compat.v1.enable_resource_variables(). WARNING:tensorflow:Unable to create a python object for variable <tf.Variable 'video_discriminator/conv_1/weights:0' shape=(4, 4, 4, 64, 128) dtype=float32_ref> because it is a reference variable. It may not be visible to training APIs. If this is a problem, consider rebuilding the SavedModel after running tf.compat.v1.enable_resource_variables(). WARNING:tensorflow:Unable to create a python object for variable <tf.Variable 'video_discriminator/conv_1/weights:0' shape=(4, 4, 4, 64, 128) dtype=float32_ref> because it is a reference variable. It may not be visible to training APIs. If this is a problem, consider rebuilding the SavedModel after running tf.compat.v1.enable_resource_variables(). WARNING:tensorflow:Unable to create a python object for variable <tf.Variable 'video_discriminator/conv_1/LayerNorm/beta:0' shape=(128,) dtype=float32_ref> because it is a reference variable. It may not be visible to training APIs. If this is a problem, consider rebuilding the SavedModel after running tf.compat.v1.enable_resource_variables(). WARNING:tensorflow:Unable to create a python object for variable <tf.Variable 'video_discriminator/conv_1/LayerNorm/beta:0' shape=(128,) dtype=float32_ref> because it is a reference variable. It may not be visible to training APIs. If this is a problem, consider rebuilding the SavedModel after running tf.compat.v1.enable_resource_variables().
Generate and show the videos
filled_frames = module(input_frames)['default'] / 255.0
# Show sequences of generated video frames.
# Concatenate start/end frames and the generated filled frames for the new videos.
generated_videos = np.concatenate([input_frames[:, :1] / 255.0, filled_frames, input_frames[:, 1:] / 255.0], axis=1)
for video_id in range(4):
fig = plt.figure(figsize=(10 * 2, 2))
for frame_id in range(1, 16):
ax = fig.add_axes([frame_id * 1 / 16., 0, (frame_id + 1) * 1 / 16., 1],
xmargin=0, ymargin=0)
ax.imshow(generated_videos[video_id, frame_id])
ax.axis('off')