Model compatibility for TF1/TF2

TF Hub model formats

TF Hub offers reusable model pieces that can be loaded back, built upon, and possibly be retrained in a TensorFlow program. These come in two different formats:

• The custom TF1 Hub format . Its main intended use is in TF1 (or TF1 compatibility mode in TF2) via its hub.Module API. Full compatibility details below.
• The native TF2 SavedModel format. Its main intended use is in TF2 via the hub.load and hub.KerasLayer APIs. Full compatibility details below.

The model format can be found on the model page on tfhub.dev. Model loading/inference, fine-tuning or creation might not be supported in TF1/2 based on the model formats.

Compatibility of the TF1 Hub format {:#compatibility_of_hubmodule}

<table style="width: 100%;"> <tr style="text-align: center"> <col style="width: 20%" /> <col style="width: 40%" /> <col style="width: 40%" /> <td style="text-align: center; background-color: #D0D0D0">Operation</td> <td style="text-align: center; background-color: #D0D0D0">TF1/ TF1 compat mode in TF2 <a href="#compatfootnote">[1]</a></td> <td style="text-align: center; background-color: #D0D0D0">TF2</td> </tr> <tr> <td>Loading / Inference</td> <td> Fully supported (<a href="https://www.tensorflow.org/hub/tf1_hub_module#using_a_module">complete TF1 Hub format loading guide</a>) <pre style="font-size: 12px;" lang="python">m = hub.Module(handle) outputs = m(inputs)</pre> </td> <td> It's recommended to use either hub.load <pre style="font-size: 12px;" lang="python">m = hub.load(handle) outputs = m.signatures["sig"](inputs)</pre> or hub.KerasLayer <pre style="font-size: 12px;" lang="python">m = hub.KerasLayer(handle, signature="sig") outputs = m(inputs)</pre> </td> </tr> <tr> <td>Fine-tuning</td> <td> Fully supported (<a href="https://www.tensorflow.org/hub/tf1_hub_module#for_consumers">complete TF1 Hub format fine-tuning guide</a>) <pre style="font-size: 12px;" lang="python">m = hub.Module(handle, trainable=True, tags=["train"]*is_training) outputs = m(inputs)</pre> <div style="font-style: italic; font-size: 14px"> Note: modules that don't need a separate train graph don't have a train tag. </div> </td> <td style="text-align: center"> Not supported </td> </tr> <tr> <td>Creation</td> <td> Fully supported (see <a href="https://www.tensorflow.org/hub/tf1_hub_module#general_approach">complete TF1 Hub format creation guide</a>) <div style="font-style: italic; font-size: 14px"> Note: The TF1 Hub format is geared towards TF1 and is only partially supported in TF2. Consider creating a TF2 SavedModel. </div></td> <td style="text-align: center">Not supported</td> </tr> </table>

Compatibility of TF2 SavedModel {:#compatibility_of_tf2_savedmodel}

Not supported before TF1.15.

<table style="width: 100%;"> <tr style="text-align: center"> <col style="width: 20%" /> <col style="width: 40%" /> <col style="width: 40%" /> <td style="text-align: center; background-color: #D0D0D0">Operation</td> <td style="text-align: center; background-color: #D0D0D0">TF1.15/ TF1 compat mode in TF2 <a href="#compatfootnote">[1]</a></td> <td style="text-align: center; background-color: #D0D0D0">TF2</td> </tr> <tr> <td>Loading / Inference</td> <td> Use either hub.load <pre style="font-size: 12px;" lang="python">m = hub.load(handle) outputs = m(inputs)</pre> or hub.KerasLayer <pre style="font-size: 12px;" lang="python">m = hub.KerasLayer(handle) outputs = m(inputs)</pre> </td> <td> Fully supported (<a href="https://www.tensorflow.org/hub/tf2_saved_model#using_savedmodels_from_tf_hub">complete TF2 SavedModel loading guide</a>). Use either hub.load <pre style="font-size: 12px;" lang="python">m = hub.load(handle) outputs = m(inputs)</pre> or hub.KerasLayer <pre style="font-size: 12px;" lang="python">m = hub.KerasLayer(handle) outputs = m(inputs)</pre> </td> </tr> <tr> <td>Fine-tuning</td> <td> Supported for a hub.KerasLayer used in tf.keras.Model when trained with Model.fit() or trained in an Estimator whose model_fn wraps the Model per the <a href="https://www.tensorflow.org/guide/migrate#using_a_custom_model_fn">custom model_fn guide</a>. <div style="font-style: italic; font-size: 14px;"> Note: hub.KerasLayer <span style="font-weight: bold;">does not</span> fill in graph collections like the old tf.compat.v1.layers or hub.Module APIs did. </div> </td> <td> Fully supported (<a href="https://www.tensorflow.org/hub/tf2_saved_model#for_savedmodel_consumers">complete TF2 SavedModel fine-tuning guide</a>). Use either hub.load: <pre style="font-size: 12px;" lang="python">m = hub.load(handle) outputs = m(inputs, training=is_training)</pre> or hub.KerasLayer: <pre style="font-size: 12px;" lang="python">m = hub.KerasLayer(handle, trainable=True) outputs = m(inputs)</pre> </td> </tr> <tr> <td>Creation</td> <td> The TF2 API <a href="https://www.tensorflow.org/api_docs/python/tf/saved_model/save"> tf.saved_model.save()</a> can be called from within compat mode. </td> <td>Fully supported (see <a href="https://www.tensorflow.org/hub/tf2_saved_model#creating_savedmodels_for_tf_hub">complete TF2 SavedModel creation guide</a>) </td> </tr> </table> <p id="compatfootnote">[1] "TF1 compat mode in TF2" refers to the combined effect of importing TF2 with <code style="font-size: 12px;" lang="python">import tensorflow.compat.v1 as tf</code> and running <code style="font-size: 12px;" lang="python">tf.disable_v2_behavior()</code> as described in the <a href="https://www.tensorflow.org/guide/migrate">TensorFlow Migration guide </a>.</p>