I am trying to train a classifier based on the InceptionV3 architecture in Keras.
For this I loaded the pre-trained InceptionV3 model, without top, and added a final fully connected layer for the classes of my classification problem. In the first training I froze the InceptionV3 base model and only trained the final fully connected layer.
In the second step I want to "fine tune" the network by unfreezing a part of the InceptionV3 model.
Now I know that the InceptionV3 model makes extensive use of BatchNorm layers. It is recommended (link to documentation), when BatchNorm layers are "unfrozen" for fine tuning when transfer learning, to keep the mean and variances as computed by the BatchNorm layers fixed. This should be done by setting the BatchNorm layers to inference mode instead of training mode.
Please also see: What's the difference between the training argument in call() and the trainable attribute?
Now my main question is: how to set ONLY the BatchNorm layers of the InceptionV3 model to inference mode?
Currently I set the whole InceptionV3 base model to inference mode by setting the "training" argument when assembling the network:
inputs = keras.Input(shape=input_shape)
# Scale the 0-255 RGB values to 0.0-1.0 RGB values
x = layers.experimental.preprocessing.Rescaling(1./255)(inputs)
# Set include_top to False so that the final fully connected (with pre-loaded weights) layer is not included.
# We will add our own fully connected layer for our own set of classes to the network.
base_model = keras.applications.InceptionV3(input_shape=input_shape, weights='imagenet', include_top=False)
x = base_model(x, training=False)
# Classification block
x = layers.GlobalAveragePooling2D(name='avg_pool')(x)
x = layers.Dense(num_classes, activation='softmax', name='predictions')(x)
model = keras.Model(inputs=inputs, outputs=x)
What I don't like about this, is that in this way I set the whole model to inference mode which may set some layers to inference mode which should not be.
Here is the part of the code that loads the weights from the initial training that I did and the code that freezes the first 150 layers and unfreezes the remaining layers of the InceptionV3 part:
model.load_weights(load_model_weight_file_name)
for layer in base_model.layers[: 150]:
layer.trainable = False
for layer in base_model.layers[ 150:]:
layer.trainable = True
The rest of my code (not shown here) are the usual compile and fit calls.
Running this code seems to result a network that doesn't really learn (loss and accuracy remain approximately the same). I tried different orders of magnitude for the optimization step size, but that doesn't seem to help.
Another thing that I observed it that when I make the whole InceptionV3 part trainable
base_model.trainable = True
that the training starts with an accuracy server orders of magnitude smaller than were my first training round finished (and of course a much higher loss). Can someone explain this to me? I would at least expect the training to continue were it left off in terms of accuracy and loss.
You could do something like:
for layer in base_model.layers:
if isinstance(layer ,tf.keras.layers.BatchNormalization):
layer.trainable=False
This will iterate over each layer and check the type, setting to inference mode if the layer is BatchNorm.
As for the low starting accuracy during transfer learning, you're only loading the weights and not the optimiser state (as would occur with a full model.load() which loads architecture, weights, optimiser state etc).
This doesn't mean there's an error, but if you must load weights only just let it train, the optimiser will configure eventually and you should see progress. Also as you're potentially over-writing the pre-trained weights in your second run, make sure you use a lower learning rate so the updates are small in comparison i.e. fine-tune the weights rather than blast them to pieces.
Related
I'm trying to solve a multilabel classification task of 10 classes with a relatively balanced training set consists of ~25K samples and an evaluation set consists of ~5K samples.
I'm using the huggingface:
model = transformers.BertForSequenceClassification.from_pretrained(...
and obtain quite nice results (ROC AUC = 0.98).
However, I'm witnessing some odd behavior which I don't seem to make sense of -
I add the following lines of code:
for param in model.bert.parameters():
param.requires_grad = False
while making sure that the other layers of the model are learned, that is:
[param[0] for param in model.named_parameters() if param[1].requires_grad == True]
gives
['classifier.weight', 'classifier.bias']
Training the model when configured like so, yields some embarrassingly poor results (ROC AUC = 0.59).
I was working under the assumption that an out-of-the-box pre-trained BERT model (without any fine-tuning) should serve as a relatively good feature extractor for the classification layers. So, where do I got it wrong?
From my experience, you are going wrong in your assumption
an out-of-the-box pre-trained BERT model (without any fine-tuning) should serve as a relatively good feature extractor for the classification layers.
I have noticed similar experiences when trying to use BERT's output layer as a word embedding value with little-to-no fine-tuning, which also gave very poor results; and this also makes sense, since you effectively have 768*num_classes connections in the simplest form of output layer. Compared to the millions of parameters of BERT, this gives you an almost negligible amount of control over intense model complexity. However, I also want to cautiously point to overfitted results when training your full model, although I'm sure you are aware of that.
The entire idea of BERT is that it is very cheap to fine-tune your model, so to get ideal results, I would advise against freezing any of the layers. The one instance in which it can be helpful to disable at least partial layers would be the embedding component, depending on the model's vocabulary size (~30k for BERT-base).
I think the following will help in demystifying the odd behavior I reported here earlier –
First, as it turned out, when freezing the BERT layers (and using an out-of-the-box pre-trained BERT model without any fine-tuning), the number of training epochs required for the classification layer is far greater than that needed when allowing all layers to be learned.
For example,
Without freezing the BERT layers, I’ve reached:
ROC AUC = 0.98, train loss = 0.0988, validation loss = 0.0501 # end of epoch 1
ROC AUC = 0.99, train loss = 0.0484, validation loss = 0.0433 # end of epoch 2
Overfitting, train loss = 0.0270, validation loss = 0.0423 # end of epoch 3
Whereas, when freezing the BERT layers, I’ve reached:
ROC AUC = 0.77, train loss = 0.2509, validation loss = 0.2491 # end of epoch 10
ROC AUC = 0.89, train loss = 0.1743, validation loss = 0.1722 # end of epoch 100
ROC AUC = 0.93, train loss = 0.1452, validation loss = 0.1363 # end of epoch 1000
The (probable) conclusion that arises from these results is that working with an out-of-the-box pre-trained BERT model as a feature extractor (that is, freezing its layers) while learning only the classification layer suffers from underfitting.
This is demonstrated in two ways:
First, after running 1000 epochs, the model still hasn’t finished learning (the training loss is still higher than the validation loss).
Second, after running 1000 epochs, the loss values are still higher than the values achieved with the non-freeze version as early as the 1’st epoch.
To sum it up, #dennlinger, I think I completely agree with you on this:
The entire idea of BERT is that it is very cheap to fine-tune your model, so to get ideal results, I would advise against freezing any of the layers.
A model should be set in the evaluation mode for inference by calling model.eval().
Do we need to also do this during training before getting the model outputs? Like within a training epoch if the network contains one or more dropout and/or batch-normalization layers.
If this is not done then the output of the forward pass in the training epoch might be affected by the randomness in the dropout?
Many example codes do not do this and something along these lines is the common approach:
for t in range(num_epochs):
# forward pass
yhat = model(x)
# get the loss
loss = criterion(yhat , y)
# backward pass, optimizer step
optimizer.zero_grad()
loss.backward()
optimizer.step()
For example here is an example code to look at : convolutional_neural_network/main.py
Should this instead be?
for t in range(num_epochs):
# forward pass
model.eval() # disable dropout etc
yhat = model(x)
# get the loss
loss = criterion(yhat , y)
# backward pass, optimizer step
model.train()
optimizer.zero_grad()
loss.backward()
optimizer.step()
TLDR:
Should this instead be?
No!
Why?
More explanation:
Different Modules behave differently depending on whether they are in training or evaluation/test mode.
BatchNorm and Dropout are only two examples of such modules, basically any module that has a training phase follows this rule.
When you do .eval(), you are signaling all modules in the model to shift operations accordingly.
Update
The answer is during training you should not use eval mode and yes, as long as you have not set the eval mode, the dropout will be active and act randomly in each forward passes. Similarly all other modules that have two phases, will perform accordingly. That is BN will always update the mean/var for each pass, and also if you use batch_size of 1, it will error out as it can not do BN with batch of 1
As it was pointed out in comments, it should be noted that during training, you should not do eval() before the forward pass, as it effectively disables all modules that has different phases for train/test mode such as BN and Dropout (basically any module that has updateable/learnable parameters, or impacts network topology like dropout) will be disabled and you will not see them contributing to your network learning. So don't code like that!
Let me explain a bit what happens during training:
When you are in training mode, all of your modules that make up your model may have two modes, training and test mode. These modules either have learnable parameters that need to be updated during training, like BN, or affect network topology in a sense like Dropout (by disabling some features during forward pass). some modules such as ReLU() only operate in one mode and thus do not have any change when modes change.
When you are in training mode, you feed an image, it passes trough layers until it faces a dropout and here, some features are disabled, thus theri responses to the next layer is omitted, the output goes to other layers until it reaches the end of the network and you get a prediction.
the network may have correct or wrong predictions, which will accordingly update the weights. if the answer was right, the features/combinations of features that resulted in the correct answer will be positively affected and vice versa.
So during training you do not need and should not disable dropout, as it affects the output and should be affecting it so that the model learns a better set of features.
I hope this makes it a bit more clear for you. if you still feel you need more, say so in the comments.
I was building a neural network model and my question is that by any chance the ordering of the dropout and batch normalization layers actually affect the model?
Will putting the dropout layer before batch-normalization layer (or vice-versa) actually make any difference to the output of the model if I am using ROC-AUC score as my metric of measurement.
I expect the output to have a large (ROC-AUC) score and want to know that will it be affected in any way by the ordering of the layers.
The order of the layers effects the convergence of your model and hence your results. Based on the Batch Normalization paper, the author suggests that the Batch Normalization should be implemented before the activation function. Since Dropout is applied after computing the activations. Then the right order of layers are:
Dense or Conv
Batch Normalization
Activation
Droptout.
In code using keras, here is how you write it sequentially:
model = Sequential()
model.add(Dense(n_neurons, input_shape=your_input_shape, use_bias=False)) # it is important to disable bias when using Batch Normalization
model.add(BatchNormalization())
model.add(Activation('relu')) # for example
model.add(Dropout(rate=0.25))
Batch Normalization helps to avoid Vanishing/Exploding Gradients when training your model. Therefore, it is specially important if you have many layers. You can read the provided paper for more details.
I have implemented an autoencoder using Keras. I understand that I can add accuracy performance metric as follows:
autoencoder.compile(optimizer='adam',
loss='mean_squared_error',
metrics=['accuracy'])
My question is:
Is the accuracy metric applied on the last layer of the decoder by default? If so, how can I set it so that it would get the representations from middle (hidden) layer to compute accuracy performance? Do I need to define a custom metric? How would that work?
It seems that what you really want is a multiple output network.
So on top of your middle layer that defines your embedding, add a layer (or more) to do your classification.
Then have a look at Multiple outputs in Keras to create your global cost.
You may also want to start by training the autoendoder only, then the classifier additional layers only to see the performance, you can also balance the accuracy of the encoder vs the accuracy of the classifier as a loss, training "both" networks at the same time.
I have 10 class dataset with this I got 85% accuracy, got the same accuracy on a saved model.
now I want to add a new class, how to add a new class To the saved model.
I tried by deleting the last layer and train but model get overfit and in prediction every Images show same result (newly added class).
This is what I did
model.pop()
base_model_layers = model.output
pred = Dense(11, activation='softmax')(base_model_layers)
model = Model(inputs=model.input, outputs=pred)
# compile and fit step
I have trained model with 10 class I want to load the model train with class 11 data and give predictions.
Using the model.pop() method and then the Keras Model() API will lead you to an error. The Model() API does not have the .pop() method, so if you want to re-train your model more than once you will have this error.
But the error only occurs if you, after the re-training, save the model and use the new saved model in the next re-training.
Another very wrong and used approach is to use the model.layers.pop(). This time the problem is that function only removes the last layer in the copy it returns. So, the model still has the layer, and just the method's return does not have the layer.
I recommend the following solution:
Admitting you have your already trained model saved in the model variable, something like:
model = load_my_trained_model_function()
# creating a new model
model_2 = Sequential()
# getting all the layers except the output one
for layer in model.layers[:-1]: # just exclude last layer from copying
model_2.add(layer)
# prevent the already trained layers from being trained again
# (you can use layers[:-n] to only freeze the model layers until the nth layer)
for layer in model_2.layers:
layer.trainable = False
# adding the new output layer, the name parameter is important
# otherwise, you will add a Dense_1 named layer, that normally already exists, leading to an error
model_2.add(Dense(num_neurons_you_want, name='new_Dense', activation='softmax'))
Now you should specify the compile and fit methods to train your model and it's done:
model_2.compile(loss='categorical_crossentropy',
optimizer='adam',
metrics=['accuracy'])
# model.fit trains the model
model_history = model_2.fit(x_train, y_train,
batch_size=batch_size,
epochs=epochs,
verbose=1,
validation_split=0.1)
EDIT:
Note that by adding a new output layer we do not have the weights and biases adjusted in the last training.
Thereby we lost pretty much everything from the previous training.
We need to save the weights and biases of the output layer of the previous training, and then we must add them to the new output layer.
We also must think if we should let all the layers train or not, or even if we should allow the training of only some intercalated layers.
To get the weights and biases from the output layer using Keras we can use the following method:
# weights_training[0] = layer weights
# weights_training[1] = layer biases
weights_training = model.layers[-1].get_weights()
Now you should specify the weights for the new output layer. You can use, for example, the mean of the weights for the weights of the new classes. It's up to you.
To set the weights and biases of the new output layer using Keras we can use the following method:
model_2.layers[-1].set_weights(weights_re_training)
model.pop()
base_model_layers = model.output
pred = Dense(11, activation='softmax')(base_model_layers)
model = Model(inputs=model.input, outputs=pred)
Freeze the first layers, before train it
for layer in model.layers[:-2]:
layer.trainable = False
I am assuming that the problem is singlelabel-multiclass classification i.e. a sample will belong to only 1 of the 11 classes.
This answer will be completely based on implementing the way humans learn into machines. Hence, this will not provide you with a proper code of how to do that but it will tell you what to do and you will be able to easily implement it in keras.
How does a human child learn when you teach him new things? At first, we ask him to forget the old and learn the new. This does not actually mean that the old learning is useless but it means that for the time while he is learning the new, the old knowledge should not interfere as it will confuse the brain. So, the child will only learn the new for sometime.
But the problem here is, things are related. Suppose, the child learned C programming language and then learned compilers. There is a relation between compilers and programming language. The child cannot master computer science if he learns these subjects separately, right? At this point we introduce the term 'intelligence'.
The kid who understands that there is a relation between the things he learned before and the things he learned now is 'intelligent'. And the kid who finds the actual relation between the two things is 'smart'. (Going deep into this is off-topic)
What I am trying to say is:
Make the model learn the new class separately.
And then, make the model find a relation between the previously learned classes and the new class.
To do this, you need to train two different models:
The model which learns to classify on the new class: this model will be a binary classifier. It predicts a 1 if the sample belongs to class 11 and 0 if it doesn't. Now, you already have the training data for samples belonging to class 11 but you might not have data for the samples which doesn't belong to class 11. For this, you can randomly select samples which belong to classes 1 to 10. But note that the ratio of samples belonging to class 11 to that not belonging to class 11 must be 1:1 in order to train the model properly. That means, 50% of the samples must belong to class 11.
Now, you have two separate models: the one which predicts class 1-10 and one which predicts class 11. Now, concatenate the outputs of (the 2nd last layers) these two models with a newly created Dense layer with 11 nodes and let the whole model retrain itself adjusting the weights of pretrained two models and learning new weights of the dense layer. Keep the learning rate low.
The final model is the third model which is a combination of two models (without last Dense layer) + a new Dense layer.
Thank you..