loading trained Spektral GeneralGNN model

[senemaktas]

Hi,

I use the GeneralGNN model from Spektral with my own dataset. Training and evaluation works fine. When I try to load the trained model, I get different errors for different loading approaches for example, weights,SavedModel, model.to_json etc..
So, my question is how to save and load the trained GeneralGNN model, is there any way to do it?. Note: I do not make any changes to this model.

thanks.

[danielegrattarola]

Sorry for the late reply but for some reason I stopped receiving notifications.

Can you provide more details on the error you are getting?

[senemaktas]

Hello again,

For example, in one of my approach; I use the below code:

from keras.models import model_from_config
model = model_from_config(open("model.json", "r").read())

The json file consist of those informations: "{"class_name": "GeneralGNN", "config": {"output": 2, "activation": "softmax", "hidden": 256, "message_passing": 4, "pre_process": 2, "post_process": 2, "connectivity": "cat", "batch_norm": true, "dropout": 0.0, "aggregate": "sum", "hidden_activation": "prelu", "pool": "sum"}, "keras_version": "2.8.0", "backend": "tensorflow"}"

The error is in the below.
ValueError: Unknown layer: GeneralGNN. Please ensure this object is passed to the custom_objects argument. See https://www.tensorflow.org/guide/keras/save_and_serialize#registering_the_custom_object for details.

Same thing for

from tensorflow.keras.models import model_from_json
loaded_model = model_from_json(open("model.json", "r").read())
# load weights into new model
loaded_model.load_weights("model.h5")
print("Loaded model from disk")

For the another approach:

# Retrieve the config
config = GeneralGNN(data.n_labels, activation="softmax").get_config()
print("config", config)

# At loading time, register the custom objects with a `custom_object_scope`:
custom_objects = {"GeneralGNN": GeneralGNN, "custom_activation": "softmax"}
with tf.keras.utils.custom_object_scope(custom_objects):
    new_model = keras.Model.from_config(config)

Error:
new_model = keras.Model.from_config(config)
File "C:\Users...\lib\site-packages\keras\engine\training.py", line 2641, in from_config functional.reconstruct_from_config(config, custom_objects))
File "C:\Users....\lib\site-packages\keras\engine\functional.py", line 1325, in reconstruct_from_config for layer_data in config['layers']:
KeyError: 'layers'

thank you.

[danielegrattarola]

Let me look into it.

In the meantime, does it work if you reinitialize a GeneralGNN model from scratch and then simply restore the weights?

[senemaktas]

I tried loading the weights as shown in the below. I got config not found error in 3 models (model1,model2, model3). I have try them seperately, one by one.

from spektral.models import GeneralGNN
# from general_gnn import GeneralGNN
model = GeneralGNN(data.n_labels, activation="softmax")
model1 = load_model('model.h5')
model2 = keras.models.load_model("model.h5", custom_objects={'GeneralGNN': model})
model3 = keras.models.load_model("model.h5", custom_objects={'GeneralGNN': GeneralGNN, "custom_activation": "softmax"})
# model3 = keras.models.load_model("model.h5", custom_objects={'GeneralGNN': GeneralGNN(data.n_labels, activation="softmax")})

raise ValueError(f'No model config found in the file at {filepath}.')
ValueError: No model config found in the file at <tensorflow.python.platform.gfile.GFile object at 0x000002EA83502820>.

---

Also I have tried another approach

# Retrieve the config
# from general_gnn import GeneralGNN
from spektral.models import GeneralGNN
config = GeneralGNN(data.n_labels, activation="softmax").get_config()
# config = {'output': 2, 'activation': 'softmax', 'hidden': 256, 'message_passing': 4, 'pre_process': 2, 'post_process': 2,
# 'connectivity': 'cat', 'batch_norm': True, 'dropout': 0.0, 'aggregate': 'sum', 'hidden_activation': 'prelu', 'pool': 'sum'}

# At loading time, register the custom objects with a `custom_object_scope`:
custom_objects = {"GeneralGNN": GeneralGNN, "custom_activation": "softmax"}
with tf.keras.utils.custom_object_scope(custom_objects):
    new_model = keras.Model.from_config(config)

File "C:\Users.....\lib\site-packages\keras\engine\training.py", line 2641, in from_config
functional.reconstruct_from_config(config, custom_objects))
File "C:\Users......\lib\site-packages\keras\engine\functional.py", line 1325, in reconstruct_from_config
for layer_data in config['layers']:
KeyError: 'layers'

When i change the last line of code -> new_model = keras.Sequential.from_config(config)
It gives "ValueError: Unknown layer: output. Please ensure this object is passed to the custom_objects argument."
(The "output" layer in the config.)

---

For the last option:

When I load from json and get the weights

from tensorflow.keras.models import model_from_json
loaded_model = model_from_json(open("model.json", "r").read())
# load weights into new model
loaded_model.load_weights("model.h5")
print("Loaded model from disk")

• Error: Unknown layer: GeneralGNN. Please ensure this object is passed to the custom_objects argument.
• İnside json: {"class_name": "GeneralGNN", "config": {"output": 2, "activation": "softmax", "hidden": 256, "message_passing": 4, "pre_process": 2, "post_process": 2, "connectivity": "cat", "batch_norm": true, "dropout": 0.0, "aggregate": "sum", "hidden_activation": "prelu", "pool": "sum"}, "keras_version": "2.8.0", "backend": "tensorflow"}

It wants right config files, but the files which i store does not save whole things related to GNN model (GeneralGNN). It only saves some basic or last layers. I am not sure how to do this. Training is working perfectly.

Thank you.

[danielegrattarola]

The problem is in recreating a model from the config, so the best you can do is instantiate a new model using the GeneralGNN constructor and then simply calling load_weights. This is only a temporary workaround but it should work

[senemaktas]

Thank you, now i am able to upload my trained model. Solution for other people:

model.save_weights('./model_path/')
loaded_model = GeneralGNN(data.n_labels, activation="softmax")
loaded_model.load_weights("./model_path/")

note: model.save('my_model', include_optimizer=True) and model.save_weights("model.h5") does not work