model save in .h5

[Lunatik00]

Hi, i have tried to save the model as shown in the tensorflow page https://www.tensorflow.org/tutorials/keras/save_and_load and i get the following error

tensorflow.python.framework.errors_impl.FailedPreconditionError: model.h5 is not a directory [Op:WriteFile]

I can save it in the other way, is the h5 format that is not working, and it would be better for what i need if i could save it in just one file.

The version are tensorflow==2.6.0 h5py==3.1.0 pyyaml==6.0 tensorflow_similarity==0.14.8

I think it could be a versions issue

[Lunatik00]

well, trial and error, it finally worked

model.save(os.path.dirname('model.h5'))

but i get the following warning

2021-11-10 00:53:53.232095: W tensorflow/python/util/util.cc:348] Sets are not currently considered sequences, but this may change in the future, so consider avoiding using them.
WARNING:absl:Found untraced functions such as dense_layer_call_fn, dense_layer_call_and_return_conditional_losses, dense_layer_call_fn, dense_layer_call_and_return_conditional_losses, dense_layer_call_and_return_conditional_losses while saving (showing 5 of 5). These functions will not be directly callable after loading.
/usr/local/lib/python3.7/dist-packages/keras/utils/generic_utils.py:497: CustomMaskWarning: Custom mask layers require a config and must override get_config. When loading, the custom mask layer must be passed to the custom_objects argument.
  category=CustomMaskWarning)

I am not sure if this affect in any way the usability of the network.

[kechan]

I ran into same issue. I tried your workaround os.path.dirname, some h5 seemed to be saved. but i also go this error:

serialize_concrete_function(concrete_function, node_ids, coder) 66 except KeyError: 67 raise KeyError( ---> 68 f"Failed to add concrete function '{concrete_function.name}' to object-" 69 f"based SavedModel as it captures tensor {capture!r} which is unsupported" 70 " or not reachable from root. "

KeyError: "Failed to add concrete function 'b'__inference_similarity_model_layer_call_fn_145923'' to object-based SavedModel as it captures tensor <tf.Tensor: shape=(), dtype=resource, value=<Resource Tensor>> which is unsupported or not reachable from root. One reason could be that a stateful object or a variable that the function depends on is not assigned to an attribute of the serialized trackable object (see SaveTest.test_captures_unreachable_variable)."

[kechan]

Update: My particular model has some data augmentation layers in them, once i remove them, the model is saved. There may be serialization related issue. One such layer is like:

tensorflow.keras.layers.experimental.preprocessing.RandomFlip

[kechan]

Actually, I discovered this is indeed a bug, that is distinct from inability to save in .h5 format (which may necessitate more code to save the _index?). The error message I got was for saving with the expected 'tf' format.

Trackable Python objects referring to this tensor (from gc.get_referrers, limited to two hops): <tf.Variable 'random_crop_1/cond/Variable:0' shape=(3,) dtype=int64>

This happens for EfficientNetSim (which is used in supervised_visualization.ipynb)

embedding_size = 128 #@param {type:"integer"}
model = EfficientNetSim(train_ds.example_shape, embedding_size)

What i found is a bit surprising. The default for augmentation is "basic" (not nothing), and it involves:

augmentation_layers = tf.keras.Sequential([
            layers.experimental.preprocessing.RandomCrop(img_size, img_size),
            layers.experimental.preprocessing.RandomFlip("horizontal")
        ])

So when i used this same model in work, I didn't know it has applied these augmentations. I am biased to think default should be None, since this is highly data-centric and task-centric. Eg. one may argue you shouldn't do this for x-ray if the disease doesn't have left/right symmetry. Model summary sort of gives a high level sequential layer and i miss this as a result.

I suspect if i don't use augmentation, the model will save correctly (avoid using RandomFlip). I will try this and update more if that doesn't resolve.

Finally, for those who really want to save and not redo expensive training, you can try model.save_weights(...) and load_weights(...) as a tmp workaround.

Let me know if this should be a separate issue.

[shibuiwilliam]

I found it is possible to save if I don't use augmentation in EfficientNetSim.

model = EfficientNetSim(train_ds.example_shape, embedding_size, augmentation=None)
tf.keras.models.save_model(model, "./sim_model.h5")
# OR 
tf.keras.models.save_model(model, "./sim_model/0")

[owenvallis]

We just pushed 0.15 to the main branch and removed the augmentation arg from all the tfsim.architectures. This is a breaking change but should resolve some of these issues.