I want to create a network with two parallel layers (same input is given to two different layers and output of them is combined with some mathematical operations). Having said that, I am not sure the back-propagation will be done by Keras automatically. As a simple example of custom RNN cell,
class Example(keras.layers.Layer):
def __init__(self, units, **kwargs):
super(Example, self).__init__(**kwargs)
self.units = units
self.state_size = units
self.la = keras.layers.Dense(self.units)
self.lb = keras.layers.Dense(self.units)
def call(self, inputs, states):
prev_output = states[0]
# parallel layers
a = tf.sigmoid(self.la(inputs))
b = tf.sigmoid(self.lb(inputs))
# combined using mathematical operation
output = (-1 * prev_output * a) + (prev_output * b)
return output, [output]
Now, the loss gradient to `la` and `lb` layers are different (gradient of loss wrt `a`, should be `-output` but wrt `b` should be `output`), will this be taken care by Keras automatically or should we create custom gradient functions?
Any insights and suggestions are much appreciated :)
Check the answer by Daniel Möller
Back-propagation will be taken care by Keras as long as all calculation is linked by tensor object, i.e don't cast tensor to another type like array, so don't worry about it.
For example of gradient tape, you can check the gradient at each layer by:
gradients = grad_tape.gradient(total_loss, model.trainable_variables)
gradient_of_last_layer = tf.reduce_max(gradients[-1])
Related
I have a encoder model and a decoder model (RNN).
I want to compute the gradients and update the weights.
I'm somewhat confused by what I've seen so far on the web.
Which block is the best practice? Is there any difference between the two options? Gradients seems to converge faster in Block 1, I do not know why?
# BLOCK 1, in two operations
encoder_gradients,decoder_gradients = tape.gradient(loss,[encoder_model.trainable_variables,decoder_model.trainable_variables])
myoptimizer.apply_gradients(zip(encoder_gradients,encoder_model.trainable_variables))
myoptimizer.apply_gradients(zip(decoder_gradients,decoder_model.trainable_variables))
# BLOCK 2, in one operation
gradients = tape.gradient(loss,encoder_model.trainable_variables + decoder_model.trainable_variables)
myoptimizer.apply_gradients(zip(gradients,encoder_model.trainable_variables +
decoder_model.trainable_variables))
You can manually verify this.
First, let's simplify the model. Let the encoder and decoder both be a single dense layer. This is mostly for simplicity and you can print out the weights being applying the gradients, gradients and weights after applying the gradients.
import tensorflow as tf
import numpy as np
from copy import deepcopy
# create a simple model with one encoder and one decoder layer.
class custom_net(tf.keras.Model):
def __init__(self):
super().__init__()
self.encoder = tf.keras.layers.Dense(3, activation='relu')
self.decoder = tf.keras.layers.Dense(3, activation='relu')
def call(self, inp):
return self.decoder(self.encoder(inp))
net = model()
# create dummy input/output
inp = np.random.randn(1,1)
gt = np.random.randn(3,1)
# set persistent to true since we will be accessing the gradient 2 times
with tf.GradientTape(persistent=True) as tape:
out = custom_model(inp)
loss = tf.keras.losses.mean_squared_error(gt, out)
# get the gradients as mentioned in the question
enc_grad, dec_grad = tape.gradient(loss,
[net.encoder.trainable_variables,
net.decoder.trainable_variables])
gradients = tape.gradient(loss,
net.encoder.trainable_variables + net.decoder.trainable_variables)
First, let's use a stateless optimizer like SGD which updates the weights based on the following formula and compare it to the 2 approaches mentioned in the question.
new_weights = weights - learning_rate * gradients.
# Block 1
myoptimizer = tf.keras.optimizers.SGD(learning_rate=1)
# store weights before updating the weights based on the gradients
old_enc_weights = deepcopy(net.encoder.get_weights())
old_dec_weights = deepcopy(net.decoder.get_weights())
myoptimizer.apply_gradients(zip(enc_grad, net.encoder.trainable_variables))
myoptimizer.apply_gradients(zip(dec_grad, net.decoder.trainable_variables))
# manually calculate the weights after gradient update
# since the learning rate is 1, new_weights = weights - grad
cal_enc_weights = []
for weights, grad in zip(old_enc_weights, enc_grad):
cal_enc_weights.append(weights-grad)
cal_dec_weights = []
for weights, grad in zip(old_dec_weights, dec_grad):
cal_dec_weights.append(weights-grad)
for weights, man_calc_weight in zip(net.encoder.get_weights(), cal_enc_weights):
print(np.linalg.norm(weights-man_calc_weight))
for weights, man_calc_weight in zip(net.decoder.get_weights(), cal_dec_weights):
print(np.linalg.norm(weights-man_calc_weight))
# block 2
old_weights = deepcopy(net.encoder.trainable_variables + net.decoder.trainable_variables)
myoptimizer.apply_gradients(zip(gradients, net.encoder.trainable_variables + \
net.decoder.trainable_variables))
cal_weights = []
for weight, grad in zip(old_weights, gradients):
cal_weights.append(weight-grad)
for weight, man_calc_weight in zip(net.encoder.trainable_variables + net.decoder.trainable_variables, cal_weights):
print(np.linalg.norm(weight-man_calc_weight))
You will see that both the methods update the weights in the exact same way.
I think you used an optimizer like Adam/RMSProp which is stateful. For such optimizers invoking apply_gradients will update the optimizer parameters based on the gradient value and sign. In the first case, the optimizer parameters are updated twice and in the second case only once.
I would stick to the second option if I were you, since you are performing just one step of optimization here.
i want to create a custom layer with weights that update only in training phase.
from the official documentation this is the way:
from keras import backend as K
from keras.layers import Layer
class MyLayer(Layer):
def __init__(self, output_dim, **kwargs):
self.output_dim = output_dim
super(MyLayer, self).__init__(**kwargs)
def build(self, input_shape):
# Create a trainable weight variable for this layer.
self.kernel = self.add_weight(name='kernel',
shape=(input_shape[1], self.output_dim),
initializer='uniform',
trainable=True)
super(MyLayer, self).build(input_shape) # Be sure to call this at the end
def call(self, x):
return K.dot(x, self.kernel)
def compute_output_shape(self, input_shape):
return (input_shape[0], self.output_dim)
in this github repo
the author added
new_centers = self.centers - self.alpha * delta_centers
self.add_update((self.centers, new_centers), x)
where self.centers are the weights.
I cant understand why self.add_update is useful in that situation.
Weights are not updated if i dont add self.add_update? If not, why new_centers must be in the updates list and not in the inputs list?And why x is a requirement?
from the source code,
self.add_update(updates, inputs)
updates: update op or list of update ops to add to the layer.
inputs: input tensor or list of inputs tensors to mark the updates as conditional on these inputs.If None is passed, the updates are assumed unconditional.
There are two types of weights:
Trainable = Updated automatically by the optimizer with backpropagation
Untrainable = Not updated by backpropagation
For the trainable weights, it's really not recommended to use updates, you will be mixing the optimizer's updates with your own updates and that could cause many issues
For the untrainable weights, you can do whatever you want. Sometimes you want constants and you will do nothing, sometimes, you want these weights to change (but not via backpropagation)
Notice how in that example the weights updated by the user are untrainable:
self.centers = self.add_weight(name='centers',
shape=(10, 2),
initializer='uniform',
#UNTRAINABLE
trainable=False)
But the user wants these weights to be updated following some rules. I don't know what they are doing there (didn't analyse the code), but I assume that they are calculating, for instance, something similar to the center point of a group of images, and each batch will have this center in a different position. They want to update this position.
A classical example is the BatchNormalization layer. Besides having trainable scale and bias weights used to rescale the outputs, they have the mean and variance weights. These are statistical properties of the data that need to be updated with every batch.
You are not training the "mean" or the "variance", but each batch of data updates these values.
How does it work?
This is obscure and lies deep down in Keras code.
We need the update operation so we make sure self.centers will have new values for every batch, otherwise it won't.
We use self.add_update in a layer to register that this variable should be updated. (We do similar things in custom optimizers as well, the optimizers contain the updates to the weights made via backpropagation)
Later in the source code for training the model, Keras will collect all these registered updates and make a train function. Somewhere inside this, these updates will be applied to the vars:
#inside a training function from keras
with K.name_scope('training'):
with K.name_scope(self.optimizer.__class__.__name__):
training_updates = self.optimizer.get_updates(
params=self._collected_trainable_weights,
loss=self.total_loss)
updates = (self.updates + #probably the updates registered in layers
training_updates + #the updates registered in optimizers
self.metrics_updates) #don't know....
# Gets loss and metrics. Updates weights at each call.
self.train_function = K.function(
inputs,
[self.total_loss] + self.metrics_tensors,
updates=updates,
name='train_function',
**self._function_kwargs)
Here is what I'm trying to implement:
We calculate loss based on F(X), as usual. But we also define "adversarial loss" which is a loss based on F(X + e). e is defined as dF(X)/dX multiplied by some constant. Both loss and adversarial loss are backpropagated for the total loss.
In tensorflow, this part (getting dF(X)/dX) can be coded like below:
grad, = tf.gradients( loss, X )
grad = tf.stop_gradient(grad)
e = constant * grad
Below is my pytorch code:
class DocReaderModel(object):
def __init__(self, embedding=None, state_dict=None):
self.train_loss = AverageMeter()
self.embedding = embedding
self.network = DNetwork(opt, embedding)
self.optimizer = optim.SGD(parameters)
def adversarial_loss(self, batch, loss, embedding, y):
self.optimizer.zero_grad()
loss.backward(retain_graph=True)
grad = embedding.grad
grad.detach_()
perturb = F.normalize(grad, p=2)* 0.5
self.optimizer.zero_grad()
adv_embedding = embedding + perturb
network_temp = DNetwork(self.opt, adv_embedding) # This is how to get F(X)
network_temp.training = False
network_temp.cuda()
start, end, _ = network_temp(batch) # This is how to get F(X)
del network_temp # I even deleted this instance.
return F.cross_entropy(start, y[0]) + F.cross_entropy(end, y[1])
def update(self, batch):
self.network.train()
start, end, pred = self.network(batch)
loss = F.cross_entropy(start, y[0]) + F.cross_entropy(end, y[1])
loss_adv = self.adversarial_loss(batch, loss, self.network.lexicon_encoder.embedding.weight, y)
loss_total = loss + loss_adv
self.optimizer.zero_grad()
loss_total.backward()
self.optimizer.step()
I have few questions:
1) I substituted tf.stop_gradient with grad.detach_(). Is this correct?
2) I was getting "RuntimeError: Trying to backward through the graph a second time, but the buffers have already been freed. Specify retain_graph=True when calling backward the first time." so I added retain_graph=True at the loss.backward. That specific error went away.
However now I'm getting a memory error after few epochs (RuntimeError: cuda runtime error (2) : out of memory at /opt/conda/conda-bld/pytorch_1525909934016/work/aten/src/THC/generic/THCStorage.cu:58
). I suspect I'm unnecessarily retaining graph.
Can someone let me know pytorch's best practice on this? Any hint / even short comment will be highly appreciated.
I think you are trying to implement generative adversarial network (GAN), but from the code, I don't understand and can't follow to what you are trying to achieve as there are a few missing pieces for a GAN to works. I can see there's a discriminator network module, DNetwork but missing the generator network module.
If to guess, when you say 'loss function twice', I assumed you mean you have one loss function for the discriminator net and another for the generator net. If that's the case, let me share how I would implement a basic GAN model.
As an example, let's take a look at this Wasserstein GAN Jupyter notebook
I'll skip the less important bits and zoom into the important ones here:
First, import PyTorch libraries and set up
# Set up batch size, image size, and size of noise vector:
bs, sz, nz = 64, 64, 100 # nz is the size of the latent z vector for creating some random noise later
Build a discriminator module
class DCGAN_D(nn.Module):
def __init__(self):
... truncated, the usual neural nets stuffs, layers, etc ...
def forward(self, input):
... truncated, the usual neural nets stuffs, layers, etc ...
Build a generator module
class DCGAN_G(nn.Module):
def __init__(self):
... truncated, the usual neural nets stuffs, layers, etc ...
def forward(self, input):
... truncated, the usual neural nets stuffs, layers, etc ...
Put them all together
netG = DCGAN_G().cuda()
netD = DCGAN_D().cuda()
Optimizer needs to be told what variables to optimize. A module automatically keeps track of its variables.
optimizerD = optim.RMSprop(netD.parameters(), lr = 1e-4)
optimizerG = optim.RMSprop(netG.parameters(), lr = 1e-4)
One forward step and one backward step for Discriminator
Here, the network can calculate gradient during the backward pass, depends on the input to this function. So, in my case, I have 3 type of losses; generator loss, dicriminator real image loss, dicriminator fake image loss. I can get gradient of loss function three times for 3 different net passes.
def step_D(input, init_grad):
# input can be from generator's generated image data or input image from dataset
err = netD(input)
err.backward(init_grad) # backward pass net to calculate gradient
return err # loss
Control trainable parameters [IMPORTANT]
Trainable parameters in the model are those that require gradients.
def make_trainable(net, val):
for p in net.parameters():
p.requires_grad = val # note, i.e, this is later set to False below in netG update in the train loop.
In TensorFlow, this part can be coded like below:
grad = tf.gradients(loss, X)
grad = tf.stop_gradient(grad)
So, I think this will answer your first question, "I substituted tf.stop_gradient with grad.detach_(). Is this correct?"
Train loop
You can see here how's the 3 different loss functions are being called here.
def train(niter, first=True):
for epoch in range(niter):
# Make iterable from PyTorch DataLoader
data_iter = iter(dataloader)
i = 0
while i < n:
###########################
# (1) Update D network
###########################
make_trainable(netD, True)
# train the discriminator d_iters times
d_iters = 100
j = 0
while j < d_iters and i < n:
j += 1
i += 1
# clamp parameters to a cube
for p in netD.parameters():
p.data.clamp_(-0.01, 0.01)
data = next(data_iter)
##### train with real #####
real_cpu, _ = data
real_cpu = real_cpu.cuda()
real = Variable( data[0].cuda() )
netD.zero_grad()
# Real image discriminator loss
errD_real = step_D(real, one)
##### train with fake #####
fake = netG(create_noise(real.size()[0]))
input.data.resize_(real.size()).copy_(fake.data)
# Fake image discriminator loss
errD_fake = step_D(input, mone)
# Discriminator loss
errD = errD_real - errD_fake
optimizerD.step()
###########################
# (2) Update G network
###########################
make_trainable(netD, False)
netG.zero_grad()
# Generator loss
errG = step_D(netG(create_noise(bs)), one)
optimizerG.step()
print('[%d/%d][%d/%d] Loss_D: %f Loss_G: %f Loss_D_real: %f Loss_D_fake %f'
% (epoch, niter, i, n,
errD.data[0], errG.data[0], errD_real.data[0], errD_fake.data[0]))
"I was getting "RuntimeError: Trying to backward through the graph a second time..."
PyTorch has this behaviour; to reduce GPU memory usage, during the .backward() call, all the intermediary results (if you have like saved activations, etc.) are deleted when they are not needed anymore. Therefore, if you try to call .backward() again, the intermediary results don't exist and the backward pass cannot be performed (and you get the error you see).
It depends on what you are trying to do. You can call .backward(retain_graph=True) to make a backward pass that will not delete intermediary results, and so you will be able to call .backward() again. All but the last call to backward should have the retain_graph=True option.
Can someone let me know pytorch's best practice on this
As you can see from the PyTorch code above and from the way things are being done in PyTorch which is trying to stay Pythonic, you can get a sense of PyTorch's best practice there.
If you want to work with higher-order derivatives (i.e. a derivative of a derivative) take a look at the create_graph option of backward.
For example:
loss = get_loss()
loss.backward(create_graph=True)
loss_grad_penalty = loss + loss.grad
loss_grad_penalty.backward()
I'm training a language model in Keras and would like to speed up training by using sampled softmax as the final activation function in my network. From the TF docs, it looks like I need to supply arguments for weights and biases, but I'm unsure of what is expected as input for these. It seems like I could write a custom function in Keras as follows:
import keras.backend as K
def sampled_softmax(weights, biases, y_true, y_pred, num_sampled, num_classes):
return K.sampled_softmax(weights, biases, y_true, y_pred, num_sampled, num_classes)
However, I'm unsure of how to "plug this in" to my existing network. The architecture for the LM is pretty dead-simple:
model = Sequential()
model.add(Embedding(input_dim=len(vocab), output_dim=256))
model.add(LSTM(1024, return_sequence=True))
model.add(Dense(output_dim=len(vocab), activation='softmax'))
model.compile(loss='categorical_crossentropy', optimizer='adam')
Given this architecture, could I pass the sampled_softmax function as the loss argument when calling the compile method on the model? Or do this need to be written as a layer that comes after the final fully-connected layer. Any guidance here would be greatly appreciated. Thanks.
The key observation here is that the TensorFlow sampled softmax function returns actual losses, not a set of predictions over the set of possible labels to compare with the ground truth data to then compute losses as a separate step. This makes the model setup a little bit weird.
First, we add a second input layer to the model that encodes the target (training) data a second time as an input, in addition to being the target output. This is used for the labels argument of the sampled_softmax_loss function. It needs to be a Keras input, because it's treated as an input when we go to instantiate and set up the model.
Second, we construct a new custom Keras layer that calls the sampled_softmax_loss function with two Keras layers as its inputs: the output of the dense layer that predicts our classes, and then the second input that contains a copy of the training data. Note that we're doing some serious hackery accessing the _keras_history instance variable to fetch the weight and bias tensors from the output tensor of the original fully-connected layer.
Finally, we have to construct a new "dumb" loss function that ignores the training data and just uses the loss reported by the sampled_softmax_loss function.
Note that because the sampled softmax function returns losses, not class predictions, you can't use this model specification for validation or inference. You'll need to re-use the trained layers from this "training version" in a new specification that applies a standard softmax function to the original dense layer which has the default activation function applied.
There is definitely a more elegant way to do this, but I believe this works, so I figured I'd post it here now as-is rather than wait until I have something that's a little bit neater. For example, you'd probably want to make the number of classes an argument of the SampledSoftmax layer, or better yet, condense this all into the loss function as in the original question and avoid passing in the training data twice.
from keras.models import Model
from keras.layers import Input, Dense, Layer
from keras import backend as K
class SampledSoftmax(Layer):
def __init__(self, **kwargs):
super(SampledSoftmax, self).__init__(**kwargs)
def call(self, inputs):
"""
The first input should be the model as it were, and the second the
target (i.e., a repeat of the training data) to compute the labels
argument
"""
# the labels input to this function is batch size by 1, where the
# value at position (i, 1) is the index that is true (not zero)
# e.g., (0, 0, 1) => (2) or (0, 1, 0, 0) => (1)
return K.tf.nn.sampled_softmax_loss(weights=inputs[0]._keras_history[0].weights[0],
biases=inputs[0]._keras_history[0].bias,
inputs=inputs[0],
labels=K.tf.reshape(K.tf.argmax(inputs[1], 1), [-1, 1]),
num_sampled=1000,
num_classes=200000)
def custom_loss(y_true, y_pred):
return K.tf.reduce_mean(y_pred)
num_classes = 200000
input = Input(shape=(300,))
target_input = Input(shape=(num_classes,))
dense = Dense(num_classes)
outputs = dense(input)
outputs = SampledSoftmax()([outputs, target_input])
model = Model([input, target_input], outputs)
model.compile(optimizer=u'adam', loss=custom_loss)
# train as desired
I have a TensorFlow model (a convolutional neural network) which I successfully trained using gradient descent (GD) on some input data.
Now, in a second step, I would like to provide an input image as initialization then and optimize over this input image with fixed network parameters using GD. The loss function will be a different one, but this a detail.
So, my main question is how to tell the gradient descent algorithm to
stop optimizing the network parameters
to optimize over the input image
The first can probably done with this
Holding variables constant during optimizer
Do you guys have ideas about the second point?
I guess I can recode the gradient descent algorithm myself using the TF gradient function, but my gut feeling tells me that there should be an easier way, which also allows me to benefit from more complex GD variants (Adam etc.).
No need for your SDG own implementation. TensorFlow provides all functions:
import tensorflow as tf
import numpy as np
# some input
data_pldhr = tf.placeholder(tf.float32)
img_op = tf.get_variable('input_image', [1, 4, 4, 1], dtype=tf.float32, trainable=True)
img_assign = img_op.assign(data_pldhr)
# your starting image
start_value = (np.ones((4, 4), dtype=np.float32) + np.eye(4))[None, :, :, None]
# override variable_getter
def nontrainable_getter(getter, *args, **kwargs):
kwargs['trainable'] = False
return getter(*args, **kwargs)
# all variables in this scope are not trainable
with tf.variable_scope('myscope', custom_getter=nontrainable_getter):
x = tf.layers.dense(img_op, 10)
y = tf.layers.dense(x, 10)
# the usual stuff
cost_op = tf.losses.mean_squared_error(x, y)
train_op = tf.train.AdamOptimizer(0.1).minimize(cost_op)
# fire up the training process
with tf.Session() as sess:
sess.run(tf.global_variables_initializer())
sess.run(img_assign, {data_pldhr: start_value})
print(sess.run(img_op))
for i in range(10):
_, c = sess.run([train_op, cost_op])
print(c)
print(sess.run(img_op))
represent an image as tf.Variable with trainable=True
initialise this variable with the starting image (initial guess)
recreate the NN graph using TF variables with trainable=False and copy the weights from the trained NN graph using tf.assign
calculate the loss function
plug the loss into any TF optimiser algorithm you want
Another alternative is to use ScipyOptimizerInterface, which allows to use scipy's minimizer. This supports constrained minimization.
I'm looking for a solution to the same problem, but my model is not an easy one as I have an LSTM network with cells created with MultiRNNCell, I don't think it is possible to get the weight and clone the network. Is there any workaround so that I can compute the gradient wrt the input?