It's been days that I've been struggling just to simply view layers' gradients in the debug mode of Keras2. Needless to say, I have already tried codes such as:
import Keras.backend as K
gradients = K.gradients(model.output, model.input)
sess = tf.compat.v1.keras.backend.get_session()
evaluated_gradients = sess.run(gradients, feed_dict={model.input:images})
or
evaluated_gradients = sess.run(gradients, feed_dict{model.input.experimantal_ref():images})
or
with tf.compat.v1.Session(graph=tf.compat.v1.keras.backend.get_default_graph())
or similar approaches using
tf.compat.v1
which all lead to the following error:
RuntimeError: The Session graph is empty. Add operations to the graph
before calling run().
I assume this should be the most basic tool any deep learning package could provide, it is strange why there seems no easy way to do so in Keras2. Any ideas?
You can try to do this on TF 2 with eager mode on.
Please notice that you need to use tf.keras for everything, including your model, layers, etc. For this to work you can never use keras alone, it must be tf.keras. This means, for instance, using tf.keras.layers.Dense, tf.keras.models.Sequential, etc..
input_images_tensor = tf.constant(input_images_numpy)
with tf.GradientTape() as g:
g.watch(input_images_tensor)
output_tensor = model(input_images_tensor)
gradients = g.gradient(output_tensor, input_images_tensor)
If you are going to calculate the gradients more than once with the same tape, you need the tape to be persistent=True and delete it manually after you get the gradients. (See details on the link below)
You can get the gradients regarding any "trainable" weight without needing watch. If you are going to get gradients with respect to non-trainable tensors (such as the input images), then you must call g.watch as above for each of these variables).
More details on GradientTape: https://www.tensorflow.org/api_docs/python/tf/GradientTape
Related
I'm trying to compute the gradient of the output layer with respect to the input layer. My neural network is relatively small (input layer composed of 9 activation units and the output layer of 1) and the training went fine as the test provided a very good accuracy. I made the NN model using Keras.
In order to solve my problem, I need to compute the gradient of the output with respect to the input. This is, I need to obtain the Jacobian which as dimension [1x9]. The gradients function in tensorflow should provide me with everything I need, but when I run the code below I obtain a different solution every time.
output_v = model.output
input_v = model.input
gradients = tf.gradients(output_v, input_v)
sess = tf.Session()
sess.run(tf.initialize_all_variables())
print(sess.run(model.input,feed_dict={model.input:x_test_N[0:1,:]}))
evaluated_gradients = sess.run(gradients,feed_dict{model.input:x_test_N[0:1,:]})
print(evaluated_gradients)
sess.close()
The first print command shows this value every time I run it (just to make sure that the input values are not modified):
[[-1.4306372 -0.1272892 0.7145787 1.338818 -1.2957293 -0.5402862-0.7771702 -0.5787912 -0.9157122]]
But the second print shows different ones:
[[ 0.00175761, -0.0490326 , -0.05413761, 0.09952173, 0.06112418, -0.04772799, 0.06557006, -0.02473242, 0.05542536]]
[[-0.00416433, 0.08235116, -0.00930298, 0.04440641, 0.03752216, 0.06378302, 0.03508484, -0.01903783, -0.0538374 ]]
Using finite differences, evaluated_gradients[0,0] = 0.03565103, which isn't close to any of the first values previously printed.
Thanks for your time!
Alberto
Solved by creating a specific session just before training my model:
sess = tf.Session()
sess.run(tf.global_variables_initializer())
K.set_session(sess)
history = model.fit(x_train_N, y_train_N, epochs=n_epochs,
validation_split=split, verbose=1, batch_size=n_batch_size,
shuffle='true', callbacks=[early_stop, tensorboard])
And evaluating the gradient after training, while tf.session is still open:
evaluated_gradients = sess.run(K.gradients(model.output, model.input), feed_dict={model.input: x_test_N})
Presumably your network is set up to initialize weights to random values. When you run sess.run(tf.initialize_all_variables()), you are initializing your variables to new random values. Therefore you get different values for output_v in every run, and hence different gradients. If you want to use a model you trained before, you should replace the initialization with initialize_all_variables() with a restore command. I am not familiar with how this is done in Keras since I usually work directly with tensorflow, but I would try this.
Also note that initialize_all_variables is deprecated and you should use global_variables_initializer instead.
How do I update moving mean and moving variance in keras BatchNormalization?
I found this in tensorflow documentation, but I don't know where to put train_op or how to work it with keras models:
update_ops = tf.get_collection(tf.GraphKeys.UPDATE_OPS)
with tf.control_dependencies(update_ops):
train_op = optimizer.minimize( loss )
No posts I found say what to do with train_op and whether you can use it in model.compile.
You do not need to manually update the moving mean and variances if you are using the BatchNormalization layer. Keras takes care of updating these parameters during training, and to keep them fixed during testing (by using the model.predict and model.evaluate functions, same as with model.fit_generator and friends).
Keras also keeps track of the learning phase so different codepaths run during training and validation/testing.
If you need just update the weights for existing model with some new values then you can do the following:
w = model.get_layer('batchnorm_layer_name').get_weights()
# Order: [gamma, beta, mean, std]
for j in range(len(w[0])):
gamma = w[0][j]
beta = w[1][j]
run_mean = w[2][j]
run_std = w[3][j]
w[2][j] = new_run_mean_value1
w[3][j] = new_run_std_value2
model.get_layer('batchnorm_layer_name').set_weights(w)
There are two interpretations of the question: the first is assuming that the goal is to use high level training api and this question was answered by Matias Valdenegro.
The second - as discussed in the comments - is whether it is possible to use batch normalization with the standard tensorflow optimizer as discussed here keras a simplified tensorflow interface and the section "Collecting trainable weights and state updates". As mentioned there the update ops are accessible in layer.updates and not in tf.GraphKeys.UPDATE_OPS, in fact if you have a keras model in tensorflow you can optimize with a standard tensorflow optimizer and batch normalization like this
update_ops = model.updates
with tf.control_dependencies(update_ops):
train_op = optimizer.minimize( loss )
and then use a tensorflow session to fetch the train_op. To distinguish training and evaluation modes of the batch normalization layer you need to feed the
learning phase state of the keras engine (see "Different behaviors during training and testing" on the same tutorial page as given above). This would work for example like this
...
# train
lo, _ = tf_sess.run(fetches=[loss, train_step],
feed_dict={tf_batch_data: bd,
tf_batch_labels: bl,
tensorflow.keras.backend.learning_phase(): 1})
...
# eval
lo = tf_sess.run(fetches=[loss],
feed_dict={tf_batch_data: bd,
tf_batch_labels: bl,
tensorflow.keras.backend.learning_phase(): 0})
I tried this in tensorflow 1.12 and it works with models containing batch normalization. Given my existing tensorflow code and in the light of approaching tensorflow version 2.0 I was tempted to use this approach myself, but given that this approach is not being mentioned in the tensorflow documentation I am not sure this will be supported in the long term and I finally have decided to not use it and to invest a little bit more to change the code to use the high level api.
This is my third attempt to get a deep learning project off the ground. I'm working with protein sequences. First I tried TFLearn, then raw TensorFlow, and now I'm trying Keras.
The previous two attempts taught me a lot, and gave me some code and concepts that I can re-use. However there has always been an obstacle, and I've asked questions that the developers can't answer (in the case of TFLearn), or I've simply gotten bogged down (TensorFlow object introspection is tedious).
I have written this TensorFlow loss function, and I know it works:
def l2_angle_distance(pred, tgt):
with tf.name_scope("L2AngleDistance"):
# Scaling factor
count = tgt[...,0,0]
scale = tf.to_float(tf.count_nonzero(tf.is_finite(count)))
# Mask NaN in tgt
tgt = tf.where(tf.is_nan(tgt), pred, tgt)
# Calculate L1 losses
losses = tf.losses.cosine_distance(pred, tgt, -1, reduction=tf.losses.Reduction.NONE)
# Square the losses, then sum, to get L2 scalar loss.
# Divide the loss result by the scaling factor.
return tf.reduce_sum(losses * losses) / scale
My target values (tgt) can include NaN, because my protein sequences are passed in a 4D Tensor, despite the fact that the individual sequences differ in length. Before you ask, the data can't be resampled like an image. So I use NaN in the tgt Tensor to indicate "no prediction needed here." Before I calculate the L2 cosine loss, I replace every NaN with the matching values in the prediction (pred) so the loss for every NaN is always zero.
Now, how can I re-use this function in Keras? It appears that the Keras Lambda core layer is not a good choice, because a Lambda only takes a single argument, and a loss function needs two arguments.
Alternately, can I rewrite this function in Keras? I shouldn't ever need to use the Theano or CNTK backend, so it isn't necessary for me to rewrite my function in Keras. I'll use whatever works.
I just looked at the Keras losses.py file to get some clues. I imported keras.backend and had a look around. I also found https://keras.io/backend/. I don't seem to find wrappers for ANY of the TensorFlow function calls I happen to use: to_float(), count_nonzero(), is_finite(), where(), is_nan(), cosine_distance(), or reduce_sum().
Thanks for your suggestions!
I answered my own question. I'm posting the solution for anyone who may come across this same problem.
I tried using my TF loss function directly in Keras, as was independently suggested by Matias Valdenegro. I did not provoke any errors from Keras by doing so, however, the loss value went immediately to NaN.
Eventually I identified the problem. The calling convention for a Keras loss function is first y_true (which I called tgt), then y_pred (my pred). But the calling convention for a TensorFlow loss function is pred first, then tgt. So if you want to keep a Tensorflow-native version of the loss function around, this fix works:
def keras_l2_angle_distance(tgt, pred):
return l2_angle_distance(pred, tgt)
<snip>
model.compile(loss = keras_l2_angle_distance, optimizer = "something")
Maybe Theano or CNTK uses the same parameter order as Keras, I don't know. But I'm back in business.
You don't need to use keras.backend, as your loss is directly written in TensorFlow, then you can use it directly in Keras. The backend functions are an abstraction layer so you can code a loss/layer that will work with the multiple available backends in Keras.
You just have to put your loss in the model.compile call:
model.compile(loss = l2_angle_distance, optimizer = "something")
I have a Keras model that I would like to convert to a Tensorflow protobuf (e.g. saved_model.pb).
This model comes from transfer learning on the vgg-19 network in which and the head was cut-off and trained with fully-connected+softmax layers while the rest of the vgg-19 network was frozen
I can load the model in Keras, and then use keras.backend.get_session() to run the model in tensorflow, generating the correct predictions:
frame = preprocess(cv2.imread("path/to/img.jpg")
keras_model = keras.models.load_model("path/to/keras/model.h5")
keras_prediction = keras_model.predict(frame)
print(keras_prediction)
with keras.backend.get_session() as sess:
tvars = tf.trainable_variables()
output = sess.graph.get_tensor_by_name('Softmax:0')
input_tensor = sess.graph.get_tensor_by_name('input_1:0')
tf_prediction = sess.run(output, {input_tensor: frame})
print(tf_prediction) # this matches keras_prediction exactly
If I don't include the line tvars = tf.trainable_variables(), then the tf_prediction variable is completely wrong and doesn't match the output from keras_prediction at all. In fact all the values in the output (single array with 4 probability values) are exactly the same (~0.25, all adding to 1). This made me suspect that weights for the head are just initialized to 0 if tf.trainable_variables() is not called first, which was confirmed after inspecting the model variables. In any case, calling tf.trainable_variables() causes the tensorflow prediction to be correct.
The problem is that when I try to save this model, the variables from tf.trainable_variables() don't actually get saved to the .pb file:
with keras.backend.get_session() as sess:
tvars = tf.trainable_variables()
constant_graph = graph_util.convert_variables_to_constants(sess, sess.graph.as_graph_def(), ['Softmax'])
graph_io.write_graph(constant_graph, './', 'saved_model.pb', as_text=False)
What I am asking is, how can I save a Keras model as a Tensorflow protobuf with the tf.training_variables() intact?
Thanks so much!
So your approach of freezing the variables in the graph (converting to constants), should work, but isn't necessary and is trickier than the other approaches. (more on this below). If your want graph freezing for some reason (e.g. exporting to a mobile device), I'd need more details to help debug, as I'm not sure what implicit stuff Keras is doing behind the scenes with your graph. However, if you want to just save and load a graph later, I can explain how to do that, (though no guarantees that whatever Keras is doing won't screw it up..., happy to help debug that).
So there are actually two formats at play here. One is the GraphDef, which is used for Checkpointing, as it does not contain metadata about inputs and outputs. The other is a MetaGraphDef which contains metadata and a graph def, the metadata being useful for prediction and running a ModelServer (from tensorflow/serving).
In either case you need to do more than just call graph_io.write_graph because the variables are usually stored outside the graphdef.
There are wrapper libraries for both these use cases. tf.train.Saver is primarily used for saving and restoring checkpoints.
However, since you want prediction, I would suggest using a tf.saved_model.builder.SavedModelBuilder to build a SavedModel binary. I've provided some boiler plate for this below:
from tensorflow.python.saved_model.signature_constants import DEFAULT_SERVING_SIGNATURE_DEF_KEY as DEFAULT_SIG_DEF
builder = tf.saved_model.builder.SavedModelBuilder('./mymodel')
with keras.backend.get_session() as sess:
output = sess.graph.get_tensor_by_name('Softmax:0')
input_tensor = sess.graph.get_tensor_by_name('input_1:0')
sig_def = tf.saved_model.signature_def_utils.predict_signature_def(
{'input': input_tensor},
{'output': output}
)
builder.add_meta_graph_and_variables(
sess, tf.saved_model.tag_constants.SERVING,
signature_def_map={
DEFAULT_SIG_DEF: sig_def
}
)
builder.save()
After running this code you should have a mymodel/saved_model.pb file as well as a directory mymodel/variables/ with protobufs corresponding to the variable values.
Then to load the model again, simply use tf.saved_model.loader:
# Does Keras give you the ability to start with a fresh graph?
# If not you'll need to do this in a separate program to avoid
# conflicts with the old default graph
with tf.Session(graph=tf.Graph()):
meta_graph_def = tf.saved_model.loader.load(
sess,
tf.saved_model.tag_constants.SERVING,
'./mymodel'
)
# From this point variables and graph structure are restored
sig_def = meta_graph_def.signature_def[DEFAULT_SIG_DEF]
print(sess.run(sig_def.outputs['output'], feed_dict={sig_def.inputs['input']: frame}))
Obviously there's a more efficient prediction available with this code through tensorflow/serving, or Cloud ML Engine, but this should work.
It's possible that Keras is doing something under the hood which will interfere with this process as well, and if so we'd like to hear about it (and I'd like to make sure that Keras users are able to freeze graphs as well, so if you want to send me a gist with your full code or something maybe I can find someone who knows Keras well to help me debug.)
EDIT: You can find an end to end example of this here: https://github.com/GoogleCloudPlatform/cloudml-samples/blob/master/census/keras/trainer/model.py#L85
I'm using bidirectional_rnn with GRUCell but this is a general question regarding the RNN in Tensorflow.
I couldn't find how to initialize the weight matrices (input to hidden, hidden to hidden). Are they initialized randomly? to zeros? are they initialized differently for each LSTM I create?
EDIT: Another motivation for this question is in pre-training some LSTMs and using their weights in a subsequent model. I don't currently know how to do that currently without saving all the states and restoring the entire model.
Thanks.
How to initialize weight matrices for RNN?
I believe people are using random normal initialization for weight matrices for RNN. Check out the example in TensorFlow GitHub Repo. As the notebook is a bit long, they have a simple LSTM model where they use tf.truncated_normal to initialize weights and tf.zeros to initialize biases (although I have tried using tf.ones to initialize biases before, seem to also work). I believe that the standard deviation is a hyperparameter you could tune yourself. Sometimes weights initialization is important to the gradient flow. Although as far as I know, LSTM itself is designed to handle gradient vanishing problem (and gradient clipping is for helping gradient exploding problem), so perhaps you don't need to be super careful with the setup of std_dev in LSTM? I've read papers recommending Xavier initialization (TF API doc for Xavier initializer) in Convolution Neural Network context. I don't know if people use that in RNN, but I imagine you can even try those in RNN if you want to see if it helps.
Now to follow up with #Allen's answer and your follow up question left in the comments.
How to control initialization with variable scope?
Using the simple LSTM model in the TensorFlow GitHub python notebook that I linked to as an example.
Specifically, if I want to re-factorize the LSTM part of the code in above picture using variable scope control, I may code something as following...
import tensorflow as tf
def initialize_LSTMcell(vocabulary_size, num_nodes, initializer):
'''initialize LSTMcell weights and biases, set variables to reuse mode'''
gates = ['input_gate', 'forget_gate', 'memory_cell', 'output_gate']
with tf.variable_scope('LSTMcell') as scope:
for gate in gates:
with tf.variable_scope(gate) as gate_scope:
wx = tf.get_variable("wx", [vocabulary_size, num_nodes], initializer)
wt = tf.get_variable("wt", [num_nodes, num_nodes], initializer)
bi = tf.get_variable("bi", [1, num_nodes, tf.constant_initializer(0.0)])
gate_scope.reuse_variables() #this line can probably be omitted, b.z. by setting 'LSTMcell' scope variables to 'reuse' as the next line, it'll turn on the reuse mode for all its child scope variables
scope.reuse_variables()
def get_scope_variables(scope_name, variable_names):
'''a helper function to fetch variable based on scope_name and variable_name'''
vars = {}
with tf.variable_scope(scope_name, reuse=True):
for var_name in variable_names
var = tf.get_variable(var_name)
vars[var_name] = var
return vars
def LSTMcell(i, o, state):
'''a function for performing LSTMcell computation'''
gates = ['input_gate', 'forget_gate', 'memory_cell', 'output_gate']
var_names = ['wx', 'wt', 'bi']
gate_comp = {}
with tf.variable_scope('LSTMcell', reuse=True):
for gate in gates:
vars = get_scope_variables(gate, var_names)
gate_comp[gate] = tf.matmul(i, vars['wx']) + tf.matmul(o, vars['wt']) + vars['bi']
state = tf.sigmoid(gate_comp['forget_gate']) * state + tf.sigmoid(gate_comp['input_gate']) * tf.tanh(gate_comp['memory_cell'])
output = tf.sigmoid(gate_comp['output_gate']) * tf.tanh(state)
return output, state
The usage of the re-factorized code would be something like following...
initialize_LSTMcell(volcabulary_size, num_nodes, tf.truncated_normal_initializer(mean=-0.1, stddev=.01))
#...Doing some computation...
LSTMcell(input_tensor, output_tensor, state)
Even though the refactorized code may look less straightforward, but using scope variable control ensures scope encapsulation and allows flexible variable controls (in my opinion at least).
In pre-training some LSTMs and using their weights in a subsequent model. How to do that without saving all the states and restoring the entire model.
Assuming you have a pre-trained model froze and loaded in, if you wanna use their frozen 'wx', 'wt' and 'bi', you can simply find their parent scope names and variable names, then fetch the variables using similar structure in get_scope_variables func.
with tf.variable_scope(scope_name, reuse=True):
var = tf.get_variable(var_name)
Here is a link to understanding variable scope and sharing variables. I hope this is helpful.
The RNN models will create their variables with get_variable, and you can control the initialization by wrapping the code which creates those variables with a variable_scope and passing a default initializer to it. Unless the RNN specifies one explicitly (looking at the code, it doesn't), uniform_unit_scaling_initializer is used.
You should also be able to share model weights by declaring the second model and passing reuse=True to its variable_scope. As long as the namespaces match up, the new model will get the same variables as the first model.
A simple way to initialize all kernel weights with certain initializer is to leave the initializer in tf.variable_scope(). For example:
with tf.variable_scope('rnn', initializer=tf.variance_scaling_initializer()):
basic_cell= tf.contrib.rnn.BasicRNNCell(num_units=n_neurons)
outputs, state= tf.nn.dynamic_rnn(basic_cell, X, dtype=tf.float32)