Skip to content

Latest commit

 

History

History
102 lines (93 loc) · 3.87 KB

itex_ops_override.md

File metadata and controls

102 lines (93 loc) · 3.87 KB

Operators Override

Python api itex.experimental_ops_override() is provided to automatically replace some TensorFlow operators by Custom Operators under itex.ops namespace, as well as to be compatible with existing trained parameters.

Usage

After itex.experimental_ops_override() is called, these TensorFlow APIs are automatically replaced by Customized Operators. For Keras layers, their call functions will be overloaded; layer names will be kept.

Layer Normalization

tf.keras.layers.LayerNormalization and keras.layers.LayerNormalization will be fused by Customized Operators of LayerNorm and LayerNormGrad. For example:

$ python
>>> import tensorflow as tf
>>> import intel_extension_for_tensorflow as itex
>>> itex.experimental_ops_override()
>>> tf.keras.layers.LayerNormalization(
      axis=-1, epsilon=0.001, center=True, scale=True,
      beta_initializer='zeros', gamma_initializer='ones',
      beta_regularizer=None, gamma_regularizer=None, beta_constraint=None,
      gamma_constraint=None, **kwargs)
>>> # it will run by op ITEXLayerNorm and ITEXLayerNormGrad

Dense Layer

tf.keras.layers.Dense and keras.layers.core.dense.Dense will be optimized by BatchMatMul, BiasAdd and Activation fusion for prediction, MatMul and BiasAdd fusion for training. For example:

$ python
>>> import tensorflow as tf
>>> import intel_extension_for_tensorflow as itex
>>> itex.experimental_ops_override()
>>> tf.keras.layers.Dense(32, activation='relu')

Gelu Activation

tf.nn.gelu will be replaced by itex.ops.gelu. For example:

$ python
>>> import tensorflow as tf
>>> import intel_extension_for_tensorflow as itex
>>> itex.experimental_ops_override()
>>> x = tf.constant([-3.0, -1.0, 0.0, 1.0, 3.0], dtype=tf.float32)
>>> y = tf.nn.gelu(x)
>>> # it will run by op ITEXGelu and ITEXGeluGrad

Instance Normalization

If TensorFlow Addons is installed, tfa.layers.InstanceNormalization will be replaced by custom implementation using Transpose and itex.ops.LayerNormalization. For example:

$ python
>>> import tensorflow as tf
>>> import intel_extension_for_tensorflow as itex
>>> itex.experimental_ops_override()
tfa.layers.InstanceNormalization(
      axis=-1,
      beta_initializer='zeros',
      gamma_initializer='ones',
      **kwargs)
>>> # it will run by op Transpose and ITEXLayerNorm

LSTM

If Intel® Extension for TensorFlow* backend is XPU, tf.keras.layers.LSTM will be replaced by itex.ops.ItexLSTM. For example:

$ python
>>> import tensorflow as tf
>>> import intel_extension_for_tensorflow as itex
>>> itex.experimental_ops_override()
>>> itex.ops.ItexLSTM(
    200, activation='tanh',
    recurrent_activation='sigmoid',
    use_bias=True,
    kernel_initializer='glorot_uniform',
    recurrent_initializer='orthogonal',
    bias_initializer='zeros', **kwargs
)
>>> # it will run by op ItexRnn

AdamW

If Intel® Extension for TensorFlow* backend is XPU, tf.keras.optimizers.AdamW and keras.optimizers.AdamW will be fused by itex.ops.AdamWithWeightDecayOptimizer. For examples:

$ python
>>> import tensorflow as tf
>>> import intel_extension_for_tensorflow as itex
>>> itex.experimental_ops_override()
>>> optimizer = tf.keras.optimizers.AdamW(
    learning_rate=0.02,
    weight_decay=0.01,
    beta_1=0.9,
    beta_2=0.999,
    epsilon=1e-6)
>>> # it will run by op ITEXResourceApplyAdamWithWeightDecay