Closed EderSantana closed 8 years ago
around1991
commented
8 years ago Yeah, when you call model_from_json, set the custom_objects kwarg to be a dictionary with the keys being the custom layer name and the value being the custom layer itself.
In my case, I have a FeedForwardAttention layer, so I have to do
model = model_from_json(config, custom_objects={'FeedForwardAttention': FeedForwardAttention})Hope this helps Kris
EderSantana
commented
8 years ago @around1991 it did work! Thank you very much!
visheshmistry
commented
7 years ago I did the exact thing as mentioned above but still got the following error:
(what I did) model=load_model('model.h5',custom_objects={"SpatialTransformer": SpatialTransformer()})
(error)
Traceback (most recent call last):
File "stn_test.py", line 119, in
(SpatialTransformer.py) class SpatialTransformer(Layer): def init(self, localization_net, output_size, kwargs): self.locnet = localization_net self.output_size = output_size super(SpatialTransformer, self).init(kwargs) ...
I even tried:
model=load_model('model.h5',custom_objects={"SpatialTransformer": SpatialTransformer(localization_net=locnet,output_size=(30,30))})
but got the same error again
ankitmishra262
commented
7 years ago Just send model=load_model('model.h5',custom_objects={"SpatialTransformer": SpatialTransformer})
visheshmistry
commented
7 years ago Thank you Ankit but I tried this too and it still did not work.
ankitmishra262
commented
7 years ago Okay, I was getting the same issue today. If you have not done it, try updating keras to 2.0. Then follow the syntax given in https://keras.io/layers/writing-your-own-keras-layers/ for your layer.
This mostly should solve your issue, if that doesn't work, check about adding get_config in your custom layer declaration.
A workaround that I can suggest is to save the model as JSON, save the weights separately by save_weights, then when required load both of them. If you aren't clear how to do it, look at http://machinelearningmastery.com/save-load-keras-deep-learning-models/
salarjf
commented
6 years ago Just send model=load_model('model.h5',custom_objects={"SpatialTransformer": SpatialTransformer})
Thanks Ankit! Your solution helped me.
Just to complete your hint, 'custom_objects' also works with other model elements like custom loss functions and so. Salar
MingleiLI
commented
6 years ago Okay, I was getting the same issue today. If you have not done it, try updating keras to 2.0. Then follow the syntax given in https://keras.io/layers/writing-your-own-keras-layers/ for your layer.
This mostly should solve your issue, if that doesn't work, check about adding
get_configin your custom layer declaration.A workaround that I can suggest is to save the model as JSON, save the weights separately by
save_weights, then when required load both of them. If you aren't clear how to do it, look at http://machinelearningmastery.com/save-load-keras-deep-learning-models/
It seems that there is nothing about how to write get_config in the links.
Monduiz
commented
5 years ago When using a custom layer, you will have to define a get_config function into the layer class. Example:
https://github.com/keras-team/keras/blob/master/keras/layers/convolutional.py#L214
This will show you how to adapt the get_config code to your custom layers.
With the example above:
`def get_config(self):
config = {
'localization_net': self.localization_net,
'output_size': self.output_size
}
base_config = super(SpatialTransformer, self).get_config()
return dict(list(base_config.items()) + list(config.items()))`You will need to retrain the model using the new class code.
offchan42
commented
5 years ago @Monduiz
I had a lot of problems with saving and loading Keras custom layers because keras forget my layer's fields.
Why don't they mentioned get_config thing in their custom layer tutorial?
offchan42
commented
5 years ago Also why keras prefer to use return dict(list(base_config.items()) + list(config.items()))
instead of
config.update(base_config)
return configBecause this is a simpler code. Is there any other reason you know?
AlkaSaliss
commented
5 years ago @Monduiz I had a lot of problems with saving and loading Keras custom layers because keras forget my layer's fields. Why don't they mentioned
get_configthing in their custom layer tutorial?
I confirm, I had tough time before finding the get_config trick, it would be nice if it is added in the doc that one might need to rewrite the get_config method in the custom layer.
AdnanRiaz107
commented
4 years ago Yeah, when you call model_from_json, set the custom_objects kwarg to be a dictionary with the keys being the custom layer name and the value being the custom layer itself.
In my case, I have a FeedForwardAttention layer, so I have to do
model = model_from_json(config, custom_objects={'FeedForwardAttention': FeedForwardAttention})Hope this helps Kris
Hi Dear! I am trying to do the same but I got the error."ValueError: Unknown Layer: Attention"
AlkaSaliss
commented
4 years ago @AdnanRiaz107 do you have a Layer called Attention in your code ? (also a small code snippet might help us to get where the problem comes from)
AdnanRiaz107
commented
4 years ago @AdnanRiaz107 do you have a Layer called
Attentionin your code ? (also a small code snippet might help us to get where the problem comes from)
Yes , I have attention layer followed by a LSTM layer.
AlkaSaliss
commented
4 years ago And can u provide the code snippet you use for loading the model and that gives the error ?
AdnanRiaz107
commented
4 years ago And can u provide the code snippet you use for loading the model and that gives the error ?
I got this error.........
Traceback (most recent call last):
File "C:/Users/adnan/PycharmProjects/Code/rerun.py", line 268, in
Process finished with exit code 1
Here is my code class Attention(Layer): def init(self, step_dim, bias=True, W_regularizer=None, b_regularizer=None, W_constraint=None, b_constraint=None, **kwargs): self.supports_masking = True self.init = initializers.get('glorot_uniform') self.W_regularizer = regularizers.get(W_regularizer) self.b_regularizer = regularizers.get(b_regularizer)
self.W_constraint = constraints.get(W_constraint)
self.b_constraint = constraints.get(b_constraint)
self.bias = bias
self.step_dim = step_dim
self.features_dim = 0
super(Attention, self).__init__(**kwargs)
def build(self, input_shape):
assert len(input_shape) == 3
self.W = self.add_weight((input_shape[-1],),
initializer=self.init,
name='{}_W'.format(self.name),
regularizer=self.W_regularizer,
constraint=self.W_constraint)
self.features_dim = input_shape[-1]
if self.bias:
self.b = self.add_weight((input_shape[1],),
initializer='zero',
name='{}_b'.format(self.name),
regularizer=self.b_regularizer,
constraint=self.b_constraint)
else:
self.b = None
self.built = True
def compute_mask(self, input, input_mask=None):
return None
def call(self, x, mask=None):
features_dim = self.features_dim
step_dim = self.step_dim
eij = K.reshape(K.dot(K.reshape(x, (-1, features_dim)), K.reshape(self.W, (features_dim, 1))),
(-1, step_dim))
if self.bias:
eij += self.b
eij = K.tanh(eij)
a = K.exp(eij)
if mask is not None:
a *= K.cast(mask, K.floatx())
a /= K.cast(K.sum(a, axis=1, keepdims=True) + K.epsilon(), K.floatx())
a = K.expand_dims(a)
weighted_input = x * a
return K.sum(weighted_input, axis=1)
def compute_output_shape(self, input_shape):
return input_shape[0], self.features_dimdef LSTMS(X, Y, epochs=30, validation_split=0.2, patience=10): speed_input = Input(shape=(X.shape[1], X.shape[2]), name='speed') main_output = LSTM(input_shape = (X.shape[1], X.shape[2]), output_dim = X.shape[2], return_sequences=False)(speed_input) main_output = Attention(3)(mainoutput) NAME = "1LSTM-mse-rmse{}".format(int(time.time())) tensorboard= TensorBoard(log_dir='logs/{}'.format(NAME), histogram_freq=150, batch_size=32, write_graph=False, write_grads=True )
final_model = Model(input=[speed_input], output=[main_output])
final_model.summary()
final_model.compile(loss='mse', optimizer='rmsprop' )
history = LossHistory()
earlyStopping = EarlyStopping(monitor='val_loss', min_delta=0.00001, patience=patience, verbose=0, mode='auto')
final_model.fit([X], Y, validation_split=0.2, epochs=epochs, callbacks=[history, earlyStopping, tensorboard])
return final_model, historyI've copie-pasted your code but was unable to run it and train a model, I have shape errors. I can't figure out what the inputs/outputs look like. here the link to the colab notebook where I tried, can you check what is wrong with the code in the notebook and/or modify it to providing an example input to make it work so that I can try saving and loading ?
Please make sure that the boxes below are checked before you submit your issue. Thank you!
How do I use
model_from_jsonif I have custom layer. Here it says that we just have to insert it to globals: https://github.com/fchollet/keras/blob/master/keras/utils/layer_utils.py#L23-L24I did
importthe layer, but it didn't work for me. Am I missing something? Do they mean anything else with "insert to globals"?