Hi I am having some serious problems saving and loading a tensorflow model which is combination of hugging face transformers + some custom layers to do classfication. I am using the latest Huggingface transformers tensorflow keras version. The idea is to extract features using distilbert and then run the features through CNN to do classification and extraction. I have got everything to work as far as getting the correct classifications.
The problem is in saving the model once trained and then loading the model again.
I am using tensorflow keras and tensorflow version 2.2
Following is the code to design the model, train it, evaluate it and then save and load it
bert_config = DistilBertConfig(dropout=0.2, attention_dropout=0.2, output_hidden_states=False)
bert_config.output_hidden_states = False
transformer_model = TFDistilBertModel.from_pretrained(DISTIL_BERT, config=bert_config)
input_ids_in = tf.keras.layers.Input(shape=(BERT_LENGTH,), name='input_token', dtype='int32')
input_masks_in = tf.keras.layers.Input(shape=(BERT_LENGTH,), name='masked_token', dtype='int32')
embedding_layer = transformer_model(input_ids_in, attention_mask=input_masks_in)[0]
x = tf.keras.layers.Bidirectional(
tf.keras.layers.LSTM(50, return_sequences=True, dropout=0.1,
recurrent_dropout=0, recurrent_activation="sigmoid",
unroll=False, use_bias=True, activation="tanh"))(embedding_layer)
x = tf.keras.layers.GlobalMaxPool1D()(x)
outputs = []
# lots of code here to define the dense layers to generate the outputs
# .....
# .....
model = Model(inputs=[input_ids_in, input_masks_in], outputs=outputs)
for model_layer in model.layers[:3]:
logger.info(f"Setting layer {model_layer.name} to not trainable")
model_layer.trainable = False
rms_optimizer = RMSprop(learning_rate=0.001)
model.compile(loss=SigmoidFocalCrossEntropy(), optimizer=rms_optimizer)
# the code to fit the model (which works)
# then code to evaluate the model (which also works)
# finally saving the model. This too works.
tf.keras.models.save_model(model, save_url, overwrite=True, include_optimizer=True, save_format="tf")
However, when I try to load the saved model using the following
tf.keras.models.load_model(
path, custom_objects={"Addons>SigmoidFocalCrossEntropy": SigmoidFocalCrossEntropy})
I get the following load error
ValueError: The two structures don't have the same nested structure.
First structure: type=TensorSpec str=TensorSpec(shape=(None, 128), dtype=tf.int32, name='inputs')
Second structure: type=dict str={'input_ids': TensorSpec(shape=(None, 5), dtype=tf.int32, name='inputs/input_ids')}
More specifically: Substructure "type=dict str={'input_ids': TensorSpec(shape=(None, 5), dtype=tf.int32, name='inputs/input_ids')}" is a sequence, while substructure "type=TensorSpec str=TensorSpec(shape=(None, 128), dtype=tf.int32, name='inputs')" is not
Entire first structure:
.
Entire second structure:
{'input_ids': .}
I believe the issue is because TFDistilBertModel layer can be called using a dictionary input from DistilBertTokenizer.encode() and that happens to be the first layer. So the model compiler on load expects that to be the input signature to the call model. However, the inputs defined to the model are two tensors of shape (None, 128)
So how do I tell the load function or the save function to assume the correct signatures?
I solved the issue.
The issue was the object transformer_model in the above code is itself not a layer. So if we want to embed it inside another keras layer we should use the internal keras layer that is wrapped in the model
So changing the line
embedding_layer = transformer_model(input_ids_in, attention_mask=input_masks_in[0]
to
embedding_layer = transformer_model.distilbert([input_ids_in, input_masks_in])[0]
makes everything work. Hope this helps someone else. Took a long time to debug through tf.keras code to figure this one out although in hindsight it is obvious. :)
I suffered the same problem, casually, yesterday. My solution is very similar to yours, I supposed that the problem was due to how tensorflow keras processes custom models so, the idea was to use the layers of the custom model inside my model. This has the advantage of not calling explicitly the layer by its name (in my case, it is useful for easy building more generic models using different pretrained encoders):
sent_encoder = getattr(transformers, self.model_name).from_pretrained(self.shortcut_weights).layers[0]
I don't explored all the models of HuggingFace, but a few that I tested seem to be a custom model with only one custom layer.
Your solution also works like a charm, in fact, both solutions are the same if "distilbert" references to ".layers[0]".
Related
I'm trying to build a model with DenseVariational layer so that it can report epistemic uncertainties. Something like https://www.tensorflow.org/probability/examples/Probabilistic_Layers_Regression#figure_3_epistemic_uncertainty
The model training works just fine and now I would like to save the model and load it in a production environment. However, when I tried model.save('path/model.h5'), I got
Layer DenseVariational has arguments in `__init__` and therefore must override `get_config`.
Then I added
class CustomVariational(tfp.layers.DenseVariational):
def get_config(self):
config = super().get_config().copy()
config.update({
'units': self.units,
'make_posterior_fn': self._make_posterior_fn,
'make_prior_fn': self._make_prior_fn
})
return config
but it failed with a new error
Unable to create link (name already exists)
Is DenseVariational layer for research only?
I think we can circumvent this problem by using the save_weights method.
When you add with tf.name_scope(...) to prior & posterior functions, it should be resolved, otherwise they end up with the same name for both tensors.
We're also fixing the example tutorial colab, should be online soon, thanks.
Update:
Instead of fixing it at the applications, we fixed it in the library instead: https://github.com/tensorflow/probability/commit/0ca065fb526b50ce38b68f7d5b803f02c78c8f16. Once it is updated, the duplicate tensor names should be resolved. Thanks.
It's been almost 2 years, and the problem is still going on.
A workaround is to store only the weights:
tf.keras.Model.save_weights(filepath, overwrite=True)
Then, you can use the same model structure and load the weights.
For example:
# model
model = tf.keras.Sequential([
tf.keras.layers.InputLayer(input_shape=(input_shape,), name="input"),
tfp.layers.DenseVariational(32, posterior_mean_field, prior_trainable), # trainable
tf.keras.layers.Dense(32, activation="relu"),
tf.keras.layers.Dense(1)
])
# save weights after compiling and training your model
model.save_weights('model_weights.h5')
Initialize a new model with the same structure:
# different model, same weights
model2 = tf.keras.Sequential([
tf.keras.layers.InputLayer(input_shape=(input_shape,), name="input"),
tfp.layers.DenseVariational(32, posterior_mean_field, prior_trainable),
tf.keras.layers.Dense(32, activation="relu"),
tf.keras.layers.Dense(1)
])
# load weights
model2.load_weights('model_weights.h5')
I hope this helps!
I want to use ResNet50 with Imagenet weights.
The last layer of ResNet50 is (from here)
x = layers.Dense(1000, activation='softmax', name='fc1000')(x)
I need to keep the weights of this layer but remove the softmax function.
I want to manually change it so my last layer looks like this
x = layers.Dense(1000, name='fc1000')(x)
but the weights stay the same.
Currently I call my net like this
resnet = Sequential([
Input(shape(224,224,3)),
ResNet50(weights='imagenet', input_shape(224,224,3))
])
I need the Input layer because otherwise the model.compile says that placeholders aren't filled.
Generally there are two ways of achievieng this:
Quick way - supported functions:
To change the final layer's activation function, you can pass an argument classifier_activation.
So in order to get rid of activation all together, your module can be called like:
import tensorflow as tf
resnet = tf.keras.Sequential([
tf.keras.layers.Input(shape=(224,224,3)),
tf.keras.applications.ResNet50(
weights='imagenet',
input_shape=(224,224,3),
pooling="avg",
classifier_activation=None
)
])
This however, is not going to work if the you want a different function, that is not supported by Keras classifer_activation parameter (e. g. custom activation function).
To achieve this you can use the workaround solution:
Long way - copy the model's weights
This solution proposes copying the original model's weights onto your custom one. This approach works because apart from the activation function you are not chaning the model's architecture.
You need to:
1. Download original model.
2. Save it's weights.
3. Declare your modified version of the model (in your case, without the activation function).
4. Set the weights of the new model.
Below snippet explains this concept in more detail:
import tensorflow as tf
# 1. Download original resnet
resnet = tf.keras.Sequential([
tf.keras.layers.Input(shape=(224,224,3)),
tf.keras.applications.ResNet50(
weights='imagenet',
input_shape=(224,224,3),
pooling="avg"
)
])
# 2. Hold weights in memory:
imagenet_weights = resnet.get_weights()
# 3. Declare the model, but without softmax
resnet_no_softmax = tf.keras.Sequential([
tf.keras.layers.Input(shape=(224,224,3)),
tf.keras.applications.ResNet50(
include_top=False,
weights='imagenet',
input_shape=(224,224,3),
pooling="avg"
),
tf.keras.layers.Dense(1000, name='fc1000')
])
# 4. Pass the imagenet weights onto the second resnet
resnet_no_softmax.set_weights(imagenet_weights)
Hope this helps!
I'm trying to build a Keras model using a DenseFeatures layer as the input; the input comes as a dict of Tensors. TF is insisting that I use model.build() to build the model before optimization, but I can't build it due to DenseFeatures not having an input shape. I get the error
AttributeError: 'DenseFeatures' object has no attribute 'shape'
How can I resolve this? Here's my code:
input_layer = tf.keras.layers.DenseFeatures(params.columns)
predictions = Dense(1, input_dim=len(params.columns), activation='softmax')(input_layer)
model = Sequential([input_layer, predictions])
model.build()
ETA: some further information for insight: I'm not actually fitting the model with this code; rather, I'm creating an EstimatorSpec to use with a Sagemaker model (thus it seems like this will probably require a bit of weird footwork between two different paradigms.)
I'm building image processing network in tensorflow and I want to make use of texture loss. Texture loss seems simple to implement if you have pretrained model loaded.
I'm using TF to build the computational graph for my model and I want to incorporate Keras.application.VGG19 model to get output from layer 'block4_conv4'.
The problem is: I have two TF tensors target and result from my main model, how to feed them into keras VGG19 in the same session to compute their diff and use it in main loss for my model?
It seems following code does the trick
with tf.variable_scope("") as scope:
phi_func = VGG19(include_top=False, weights=None, input_shape=(128, 128, 3))
text_1 = phi_func(predicted)
scope.reuse_variables()
text_2 = phi_func(x)
text_loss = tf.reduce_mean((text_1 - text_2)**2)
right after session created I call phi_func.load_weights(path) to initiate weights
I wanna draw the weights of tf.layers.dense in tensorboard histogram, but it not show in the parameter, how could I do that?
The weights are added as a variable named kernel, so you could use
x = tf.dense(...)
weights = tf.get_default_graph().get_tensor_by_name(
os.path.split(x.name)[0] + '/kernel:0')
You can obviously replace tf.get_default_graph() by any other graph you are working in.
I came across this problem and just solved it. tf.layers.dense 's name is not necessary to be the same with the kernel's name's prefix. My tensor is "dense_2/xxx" but it's kernel is "dense_1/kernel:0". To ensure that tf.get_variable works, you'd better set the name=xxx in the tf.layers.dense function to make two names owning same prefix. It works as the demo below:
l=tf.layers.dense(input_tf_xxx,300,name='ip1')
with tf.variable_scope('ip1', reuse=True):
w = tf.get_variable('kernel')
By the way, my tf version is 1.3.
The latest tensorflow layers api creates all the variables using the tf.get_variable call. This ensures that if you wish to use the variable again, you can just use the tf.get_variable function and provide the name of the variable that you wish to obtain.
In the case of a tf.layers.dense, the variable is created as: layer_name/kernel. So, you can obtain the variable by saying:
with tf.variable_scope("layer_name", reuse=True):
weights = tf.get_variable("kernel") # do not specify
# the shape here or it will confuse tensorflow into creating a new one.
[Edit]: The new version of Tensorflow now has both Functional and Object-Oriented interfaces to the layers api. If you need the layers only for computational purposes, then using the functional api is a good choice. The function names start with small letters for instance -> tf.layers.dense(...). The Layer Objects can be created using capital first letters e.g. -> tf.layers.Dense(...). Once you have a handle to this layer object, you can use all of its functionality. For obtaining the weights, just use obj.trainable_weights this returns a list of all the trainable variables found in that layer's scope.
I am going crazy with tensorflow.
I run this:
sess.run(x.kernel)
after training, and I get the weights.
Comes from the properties described here.
I am saying that I am going crazy because it seems that there are a million slightly different ways to do something in tf, and that fragments the tutorials around.
Is there anything wrong with
model.get_weights()
After I create a model, compile it and run fit, this function returns a numpy array of the weights for me.
In TF 2 if you're inside a #tf.function (graph mode):
weights = optimizer.weights
If you're in eager mode (default in TF2 except in #tf.function decorated functions):
weights = optimizer.get_weights()
in TF2 weights will output a list in length 2
weights_out[0] = kernel weight
weights_out[1] = bias weight
the second layer weight (layer[0] is the input layer with no weights) in a model in size: 50 with input size: 784
inputs = keras.Input(shape=(784,), name="digits")
x = layers.Dense(50, activation="relu", name="dense_1")(inputs)
x = layers.Dense(50, activation="relu", name="dense_2")(x)
outputs = layers.Dense(10, activation="softmax", name="predictions")(x)
model = keras.Model(inputs=inputs, outputs=outputs)
model.compile(...)
model.fit(...)
kernel_weight = model.layers[1].weights[0]
bias_weight = model.layers[1].weights[1]
all_weight = model.layers[1].weights
print(len(all_weight)) # 2
print(kernel_weight.shape) # (784,50)
print(bias_weight.shape) # (50,)
Try to make a loop for getting the weight of each layer in your sequential network by printing the name of the layer first which you can get from:
model.summary()
Then u can get the weight of each layer running this code:
for layer in model.layers:
print(layer.name)
print(layer.get_weights())