VideoPrismBackbone model

[source]

VideoPrismBackbone class

keras_hub.models.VideoPrismBackbone(
    num_frames,
    patch_size,
    hidden_dim,
    intermediate_dim,
    num_heads,
    num_spatial_layers,
    num_temporal_layers,
    num_auxiliary_layers,
    vocabulary_size=0,
    num_text_layers=0,
    dropout_rate=0.0,
    attention_dropout_rate=0.0,
    attention_logit_soft_cap=None,
    layer_norm_epsilon=1e-06,
    image_shape=(288, 288, 3),
    data_format=None,
    dtype=None,
    **kwargs
)

VideoPrism backbone for video and multimodal understanding.

This backbone implements the VideoPrism architecture, a powerful video understanding model that uses a factorized encoder design. The model can operate in two modes:

1. Video-only mode (num_text_layers=0): Contains only a video encoder that processes videos through spatial and temporal factorized encoding, outputting frame-level video features.
2. Multimodal mode (num_text_layers>0): Includes both a video encoder and a CoCa-style text encoder, producing aligned video and text embeddings suitable for contrastive learning and vision-language tasks.

The video encoder uses a factorized design that separately processes spatial and temporal information for efficiency and scalability.

Arguments

• num_frames: int. The number of frames in the input video sequence.
• patch_size: int. The size of each square patch in the input image.
• hidden_dim: int. The dimensionality of the hidden representations throughout the model.
• intermediate_dim: int. The dimensionality of the intermediate layer in the feedforward MLP blocks.
• num_heads: int. The number of attention heads for each transformer.
• num_spatial_layers: int. Number of transformer layers in the spatial encoder that processes within-frame information.
• num_temporal_layers: int. Number of transformer layers in the temporal encoder that processes across-frame information.
• num_auxiliary_layers: int. Number of additional transformer layers applied after the factorized video encoder. Only used when num_text_layers > 0.
• vocabulary_size: int. The size of the token vocabulary. Only required when num_text_layers > 0. Defaults to 0.
• num_text_layers: int. The number of transformer encoder layers for the text encoder. Set to 0 for video-only mode, or a positive value for multimodal mode with both video and text encoders. Defaults to 0.
• dropout_rate: float. Dropout probability for the Transformer encoder. Defaults to 0.0.
• attention_dropout: float. Dropout probability applied to the attention weights. Defaults to 0.0.
• attention_logit_soft_cap: None or float. Soft cap for the attention logits. Defaults to None.
• image_shape: tuple of ints. The shape of each input frame as (height, width, channels). For example, (288, 288, 3).
• layer_norm_epsilon: float. The epsilon for the layer normalization. Defaults to 1e-6.
• data_format: None or str. If specified, either "channels_last" or "channels_first". The ordering of the dimensions in the inputs. "channels_last" corresponds to inputs with shape (batch_size, height, width, channels) while "channels_first" corresponds to inputs with shape (batch_size, channels, height, width). It defaults to the image_data_format value found in your Keras config file at ~/.keras/keras.json. If you never set it, then it will be "channels_last".
• dtype: string or keras.mixed_precision.DTypePolicy. The dtype to use for the model's computations and weights. Note that some operations, such as softmax and layer normalization, will always be performed in float32 precision regardless of dtype.

Returns

• When num_text_layers=0 (video-only mode): A tensor of shape (batch_size, num_frames, num_patches, hidden_dim) containing the video features for each patch in each frame.
• __ When num_text_layers>0 (multimodal mode)__: A dictionary with two keys:
  • "vision_embeddings": A tensor of shape (batch_size, hidden_dim) containing the pooled and normalized video embeddings.
  • "text_embeddings": A tensor of shape (batch_size, hidden_dim) containing the normalized text embeddings from the final token.

Example

# Video-only mode
backbone = keras_hub.models.VideoPrismBackbone.from_preset(
    "videoprism_public_v1_base"
)
# (batch_size, frames, H, W, C)
pixel_values = np.random.rand(2, 16, 288, 288, 3)
# (batch_size, num_frames, num_patches, hidden_dim)
features = backbone.predict(pixel_values)

# Multimodal mode with text encoder
token_ids = np.ones((2, 64), dtype="int32")  # (batch_size, seq_len)
padding_mask = np.ones((2, 64), dtype="int32")  # (batch_size, seq_len)
backbone = keras_hub.models.VideoPrismBackbone.from_preset(
    "videoprism_lvt_public_v1_base"
)
inputs = {
    "pixel_values": pixel_values,
    "token_ids": token_ids,
    "padding_mask": padding_mask,
}
outputs = backbone.predict(inputs)
outputs["vision_embeddings"]  # (batch_size, hidden_dim)
outputs["text_embeddings"]  # (batch_size, hidden_dim)

[source]

from_preset method

VideoPrismBackbone.from_preset(preset, load_weights=True, **kwargs)

Instantiate a keras_hub.models.Backbone from a model preset.

A preset is a directory of configs, weights and other file assets used to save and load a pre-trained model. The preset can be passed as a one of:

1. a built-in preset identifier like 'bert_base_en'
2. a Kaggle Models handle like 'kaggle://user/bert/keras/bert_base_en'
3. a Hugging Face handle like 'hf://user/bert_base_en'
4. a ModelScope handle like 'modelscope://user/bert_base_en'
5. a path to a local preset directory like './bert_base_en'

This constructor can be called in one of two ways. Either from the base class like keras_hub.models.Backbone.from_preset(), or from a model class like keras_hub.models.GemmaBackbone.from_preset(). If calling from the base class, the subclass of the returning object will be inferred from the config in the preset directory.

For any Backbone subclass, you can run cls.presets.keys() to list all built-in presets available on the class.

Arguments

• preset: string. A built-in preset identifier, a Kaggle Models handle, a Hugging Face handle, or a path to a local directory.
• load_weights: bool. If True, the weights will be loaded into the model architecture. If False, the weights will be randomly initialized.

Examples

# Load a Gemma backbone with pre-trained weights.
model = keras_hub.models.Backbone.from_preset(
    "gemma_2b_en",
)

# Load a Bert backbone with a pre-trained config and random weights.
model = keras_hub.models.Backbone.from_preset(
    "bert_base_en",
    load_weights=False,
)