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
| Operation | TF1/ TF1 compat mode in TF2 [1] | TF2 |
| Loading / Inference |
Fully supported (complete TF1 Hub format loading guide)
m = hub.Module(handle)
outputs = m(inputs) |
It's recommended to use either hub.load
m = hub.load(handle)
outputs = m.signatures["sig"](inputs)m = hub.KerasLayer(handle, signature="sig")
outputs = m(inputs) |
| Fine-tuning |
Fully supported (complete TF1 Hub format fine-tuning guide)
m = hub.Module(handle,
trainable=True,
tags=["train"]*is_training)
outputs = m(inputs)
Note: modules that don't need a separate train graph don't have a train
tag.
|
Not supported |
| Creation | Fully supported (see complete TF1 Hub format creation guide)
Note: The TF1 Hub format is geared towards TF1 and is only partially supported in TF2. Consider creating a TF2 SavedModel.
|
Not supported |
Compatibility of TF2 SavedModel
Not supported before TF1.15.
| Operation | TF1.15/ TF1 compat mode in TF2 [1] | TF2 |
| Loading / Inference |
Use either hub.load
m = hub.load(handle)
outputs = m(inputs)m = hub.KerasLayer(handle)
outputs = m(inputs) |
Fully supported (complete TF2 SavedModel loading guide). Use either hub.load
m = hub.load(handle)
outputs = m(inputs)m = hub.KerasLayer(handle)
outputs = m(inputs) |
| Fine-tuning |
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 custom model_fn guide.
Note: hub.KerasLayer does not
fill in graph collections like the old tf.compat.v1.layers or hub.Module
APIs did.
|
Fully supported (complete TF2 SavedModel fine-tuning guide).
Use either hub.load:
m = hub.load(handle)
outputs = m(inputs, training=is_training)m = hub.KerasLayer(handle, trainable=True)
outputs = m(inputs) |
| Creation | The TF2 API tf.saved_model.save() can be called from within compat mode. | Fully supported (see complete TF2 SavedModel creation guide) |
[1] "TF1 compat mode in TF2" refers to the combined
effect of importing TF2 with
import tensorflow.compat.v1 as tf
and running
tf.disable_v2_behavior()
as described in the
TensorFlow Migration guide
.