Using different learning rates for different variables in TensorFlow - tensorflow

Is it possible to set different learning rates for different variables in the same layer in TensorFlow?
For example, in a dense layer, how can you set a learning rate of 0.001 for the kernel while setting the learning rate for the bias to be 0.005?
One solution is to divide the layer into 2 layers. In one layer you only train the kernel (with a non-trainable 0 bias) and in the other one you only train the bias (with a non-trainable identity kernel). This way one can use tfa.optimizers.MultiOptimzer to set different learning rates for the two layers. But this slightly slows down the training, because now the training of the bias and the kernel is not parallelised. So, I'm wondering if there is an standard way of setting different learning rates for different variables in the same layer in TF?

Related

How to freeze batch-norm layers during Transfer-learning

I am following the Transfer learning and fine-tuning guide on the official TensorFlow website. It points out that during fine-tuning, batch normalization layers should be in inference mode:
Important notes about BatchNormalization layer
Many image models contain BatchNormalization layers. That layer is a
special case on every imaginable count. Here are a few things to keep
in mind.
BatchNormalization contains 2 non-trainable weights that get updated during training. These are the variables tracking the mean and variance of the inputs.
When you set bn_layer.trainable = False, the BatchNormalization layer will run in inference mode, and will not update its mean & variance statistics. This is not the case for other layers in general, as weight trainability & inference/training modes are two orthogonal concepts. But the two are tied in the case of the BatchNormalization layer.
When you unfreeze a model that contains BatchNormalization layers in order to do fine-tuning, you should keep the BatchNormalization layers in inference mode by passing training=False when calling the base model. Otherwise the updates applied to the non-trainable weights will suddenly destroy what the model has learned.
You'll see this pattern in action in the end-to-end example at the end
of this guide.
Even tho, some other sources, for example this article (titled Transfer Learning with ResNet), says something completely different:
for layer in resnet_model.layers:
if isinstance(layer, BatchNormalization):
layer.trainable = True
else:
layer.trainable = False
ANYWAY, I know that there is a difference between training and trainable parameters in TensorFlow.
I am loading my model from file, as so:
model = tf.keras.models.load_model(path)
And I am unfreezing (or actually freezing the rest) some of the top layers in this way:
model.trainable = True
for layer in model.layers:
if layer not in model.layers[idx:]:
layer.trainable = False
NOW about batch normalization layers: I can either do:
for layer in model.layers:
if isinstance(layer, keras.layers.BatchNormalization):
layer.trainable = False
or
for layer in model.layers:
if layer.name.startswith('bn'):
layer.call(layer.input, training=False)
Which one should I do? And whether finally it is better to freeze batch norm layer or not?
Not sure about the training vs trainable difference, but personally I've gotten good results settings trainable = False.
Now as to whether to freeze them in the first place: I've had good results with not freezing them. The reasoning is simple, the batch norm layer learns the moving average of the initial training data. This may be cats, dogs, humans, cars e.t.c. But when you're transfer learning, you could be moving to a completely different domain. The moving averages of this new domain of images are far different from the prior dataset.
By unfreezing those layers and freezing the CNN layers, my model saw a 6-7% increase in accuracy (82 -> 89% ish). My dataset was far different from the inital Imagenet dataset that efficientnet was trained on.
P.S. Depending on how you plan on running the mode post training, I would advise you to freeze the batch norm layers once the model is trained. For some reason, if you ran the model online (1 image at a time), the batch norm would get all funky and give irregular results. Freezing them post training fixed the issue for me.
Use the code below to see whether the batch norm layer are being freezed or not. It will not only print the layer names but whether they are trainable or not.
def print_layer_trainable(conv_model):
for layer in conv_model.layers:
print("{0}:\t{1}".format(layer.trainable, layer.name))
In this case i have tested your method but did not freezed my model's batch norm layers.
for layer in model.layers:
if isinstance(layer, keras.layers.BatchNormalization):
layer.trainable = False
The code below worked nice for me. In my case the model is a ResNetV2 and the batch norm layers are named with the suffix "preact_bn". By using the code above for printing layers you can see how the batch norm layers are named and configure as you want.
for layer in new_model.layers[:]:
if ('preact_bn' in layer.name):
trainable = False
else:
trainable = True
layer.trainable = trainable
Just to add to #luciano-dourado answer;
In my case, I started by following the Transfer Learning guide as is, that is, freezing BN layers throughout the entire training (classifier + fine-tuning).
What I saw is that training the classifier worked without problems but as soon as I started fine-tuning, the loss went to NaN after a few batches.
After running the usual checks: input data without NaNs, loss functions yielding correct values, etc. I checked if BN layers were running in inference mode (trainable = False).
But in my case, the dataset was so different to ImageNet that I needed to do the contrary, set all trainable BN attributes to True. I found this empirically just as #zwang commented. Just remember to freeze them after training, before you deploy the model for inference.
By the way, just as an informative note, ResNet50V2, for example, has a total 49 BN layers of which only 16 are pre-activations BNs. This means that the remaining 33 layers were updating their mean and variance values.
Yet another case where one has to run several empirical tests to find out why the "standard" approach does not work in his/her case. I guess this further reinforces the importance of data in Deep Learning :)

where does class_weights or weighted loss penalize the network?

I am working on a Semantic segmentation project where I have to work on multiclass data which is highly imbalanced. I searched for optimizing it during training using the model.fit parameter and in that to use class_weights or sample_weights.
I can implement a following using a class_weight dictionary as
{ 0:1, 1:10,2:15 }
I also saw a method of updating weights in loss function
But at what point do these weights get updated?
If class_weights are used where will it get penalized? I already have a kernel_regularizer for each layer so if my classes have to be penalized based on my class weights then will it penalize the output of each layer y=Wx+b or only at the final layer?
Same if I use a weighted loss function will it get penalized only on the final layer before loss calculation or on each layer and then the final loss is calculated?
Any explanation on this would be very useful.
The class_weights you mentioned in your dictionary are there to account for your imbalanced data. They will never change, they are only there to increase the penalty for misclassified instances of minority classes (that way your network pays more attention to them and the gradients returned treat one 'Class2' instance as if it was 15 times more important than one 'Class0' instance).
The kernel_regularizer you mention resides at your loss function and penalizes large weight norms for weight matrices throughout the network (if you use kernel_regularizer = tf.keras.regularizers.l1(0.01) in a Dense layer, it only affects that layer). So that is a different weight that has nothing to do with classes, only with weights inside your network. Your eventual loss will be something like loss = Cross_entropy + a * norm(Weight_matrix) and that way the network will have as an additional task assigned to it to minimize the classification loss (cross entropy) while the weight norms remain low.

train keras applications from scratch in tensorflow 2

For benchmarking different Frameworks, I want to train a inception v3 network from scratch.
Here the code snippet to build the model:
IMAGE_RES = 229
NUM_CLASSES = 102
model = tf.keras.applications.InceptionV3(include_top=True,weights=None,classes=NUM_CLASSES)
model.build(input_shape=(None, IMAGE_RES , IMAGE_RES , channels))
according to the official keras website, the argument weight=None , means a random initialization. Does this mean that I am training my network from scratch? If not, how is it possible to train the nerwork from scratch?
Yes it means that you are training your model from scratch.
Weight and biases in deep learning models are randomly initialized following some specific shemes. (See the Xavier Glorot scheme for example) Those schemes generally helps the network converge faster and achieves better results, by preventing the gradient to either vanish or explode, and by maintaining a low variance in the gradient across all layers.

What are the uses of layers in keras/Tensorflow

So I am new to computer vision, and I do not really know what the layers do in keras. What is the use of adding layers (dense, Conv2D, etc) in keras? What do they add to it?
Convolution neural network has 4 main steps: Convolution, Pooling, Flatten, and Full connection.
Conv2D(), Conv3D(), etc. is for Feature extraction (It's a Convolution Layer).
Pooling layers (MaxPool2D(), AvgPool2D(), etc) is for Feature extraction as well (It has different operation though).
Flattening layers (Flatten() ) are to convert the extracted feature map into Vector before being fed into the Fully connection layers (The Dense layers).
Dense layers are for Fully connected step in Computer vision that acts as Classifier (The Neural network classify each extracted features from the Convolution layers.)
There are also optimization layers such as Dropout(), BatchNormalization(), etc.
For more information, just open the keras documentation.
If you want to start learning Convolution neural network, this article may help.
A layer in an Artificial Neural Network is a bunch of nodes bound together at a specific depth in a Neural Network. Keras is a high level API used over NN modules like TensorFlow or CNTK in order to simplify tasks. A Keras layer comprises 3 main parts:
Input Layer - Which contains the raw data
Hidden layer - Where the nodes of a layer learn some aspects about
the raw data which is input. It's similar to levels of abstraction
to form a Neural network.
Output Layer - Consists of a single output which is mostly a single
node and can be subjected to classification.
Keras, as a whole consists of many different types of layers. A Convolutional layer creates a kernel which is convoluted with the input over a single temporal space to derive a group of outputs. Pooling layers provide sampling of the feature maps by simplifying features in a map into patches. Max Pooling and Average Pooling are commonly used methods in a Pool layer.
Other commonly used layers in Keras are Embedding layers, Noise layers and Core layers. A single NN layer can represent only a Linearly seperable method. Most prediction problems are complicated and more than just one layer is required. This is where Multi Layer concept is required.
I think i clear your doubts and for any other queries you can see on https://www.tensorflow.org/api_docs/python/tf/keras
Neural networks are a great tool nowadays to automate classification problems. However when it comes to computer vision the amount of input data is too great to be handled efficiently by simple neural networks.
To reduce the network workload, your data needs to be preprocessed and certain features need to be identified. To find features in images we can use certain filters (like sobel edge detection), which will highlight the essential features needed for classification.
Again the amount of filters required to classify one image is too great, and thus the selection of those filters needs to be automated.
That's where the convolutional layer comes in.
We use a convolutional layer to generate multiple random (at first) filters that will highlight certain features in an image. While the network is training those filters are optimized to do a better job at highlighting features.
In Tensorflow we use Conv2D() to add one of those layers. An example of parameters is : Conv2D(64, 3, activation='relu'). 64 denotes the number of filters used, 3 denotes the size of the filters (in this case 3x3) and activation='relu' denotes the activation function
After the convolutional layer we use a pooling layer to further highlight the features produced by the previous convolutional layer. In Tensorflow this is usually done with MaxPooling2D() which takes the filtered image and applies a 2x2 (by default) layer every 2 pixels. The filter applied by MaxPooling is basically looking for the maximum value in that 2x2 area and adds it in a new image.
We can use this set of convolutional layer and pooling layers multiple times to make the image easier for the network to work with.
After we are done with those layers, we need to pass the output to a conventional (Dense) neural network.
To do that, we first need to flatten the image data from a 2D Tensor(Matrix) to a 1D Tensor(Vector). This is done by calling the Flatten() method.
Finally we need to add our Dense layers which are used to train on the flattened data. We do this by calling Dense(). An example of parameters is Dense(64, activation='relu')
where 64 is the number of nodes we are using.
Here is an example CNN structure I used recently:
# Build model
model = tf.keras.models.Sequential()
# Convolution and pooling layers
model.add(tf.keras.layers.Conv2D(64, 3, activation='relu', input_shape=(IMG_SIZE, IMG_SIZE, 1))) # Input layer
model.add(tf.keras.layers.MaxPooling2D())
model.add(tf.keras.layers.Conv2D(64, 3, activation='relu'))
model.add(tf.keras.layers.MaxPooling2D())
# Flattened layers
model.add(tf.keras.layers.Flatten())
# Dense layers
model.add(tf.keras.layers.Dense(64, activation='relu'))
model.add(tf.keras.layers.Dense(2, activation='softmax')) # Output layer
Of course this worked for a certain classification problem and the number of layers and method parameters differ depending on the problem.
The Youtube channel The Coding Train has a very helpful video explaining the Convolutional and Pooling layer.

tensorflow - how to use variational recurrent dropout correctly

The tensorflow config dropout wrapper has three different dropout probabilities that can be set: input_keep_prob, output_keep_prob, state_keep_prob.
I want to use variational dropout for my LSTM units, by setting the variational_recurrent argument to true. However, I don't know which of the three dropout probabilities I have to use for variational dropout to function correctly.
Can someone provide help?
According to this paper https://arxiv.org/abs/1512.05287 that is used for implementation of the variational_recurrent dropouts, you can think about as follows,
input_keep_prob - probability that dropping out input connections.
output_keep_prob - probability that dropping out output connections.
state_keep_prob - Probability that droping out recurrent connections.
See the diagram below,
If you set the variational_recurrent to be true you will get an RNN that's similar to the model in right and otherwise in left.
The basic differences in above two models are,
Variational RNN repeats the same dropout mask at each time
step for both inputs, outputs, and recurrent layers (drop
the same network units at each time step).
Native RNN uses different dropout masks at each time step for the
inputs and outputs alone (no dropout is used with the recurrent
connections since the use of different masks with these connections
leads to deteriorated performance).
In the above diagram, coloured connections represent the dropped-out connections, with different colours corresponding to different dropout masks. Dashed lines correspond to standard connections with no dropout.
Therefore, if you use a variational RNN you can set all three probability parameters according to your requirement.
Hope this helps.