Here is a great question on how to find the first occurence of Nan in a tensorflow graph:
Debugging nans in the backward pass
The answer is quite helpful, here is the code from it:
train_op = ...
check_op = tf.add_check_numerics_ops()
sess = tf.Session()
sess.run([train_op, check_op]) # Runs training and checks for NaNs
Apparently, running the training and the numerical check at the same time will result in an error report as soon as Nan is encountered for the first time.
How do I integrate this into Keras ?
In the documentation, I can't find anything that looks like this.
I checked the code, too.
The update step is executed here:
https://github.com/fchollet/keras/blob/master/keras/engine/training.py
There is a function called _make_train_function where an operation to compute the loss and apply updates is created. This is later called to train the network.
I could change the code like this (always assuming that we're running on a tf backend):
check_op = tf.add_check_numerics_ops()
self.train_function = K.function(inputs,
[self.total_loss] + self.metrics_tensors + [check_op],
updates=updates, name='train_function', **self._function_kwargs)
I'm currently trying to set this up properly and not sure whether the code above actually works.
Maybe there is an easier way ?
I've been running into the exact same problem, and found an alternative to the check_add_numerics_ops() function. Instead of going that route, I use the TensorFlow Debugger to walk through my model, following the example in https://www.tensorflow.org/guide/debugger to figure out exactly where my code produces nans. This snippet should work for replacing the TensorFlow Session that Keras is using with a debugging session, allowing you to use tfdbg.
from tensorflow.python import debug as tf_debug
sess = K.get_session()
sess = tf_debug.LocalCLIDebugWrapperSession(sess)
K.set_session(sess)
I have the following lines as part of a program:
tensor_gradients = optimizer.compute_gradients(cross_entropy)
with tf.Session() as session:
for step in range(20000):
batch = mnist.train.next_batch(train_batch_size)
feed = {input_x: batch[0], input_y: batch[1]}
gradients = session.run([tensor_gradients], feed)[0]
for i in range(len(gradients)):
gradients[i] = (gradients[i][0], tensor_gradients[i][1])
... computation on gradients ...
training_step = optimizer.apply_gradients(gradients)
training = session.run([training_step], feed)
The reason I'm doing this is because I want to modify the gradients using numpy. The above code runs out of memory around step 800. However, if you replace the optimizer.apply_gradients step by tensor_gradients, then the code does not run out of memory.
training_step = optimizer.apply_gradients(tensor_gradients)
Any ideas at what might be happening? The rest of the code remains the same except for the line above. Is it possible that the numpy arrays in gradients is not being garbage collected because they are being passed into the apply_gradients step? I have no idea where the memory leak could be or if I'm inadvertently adding to the tensorflow graph by passing modified gradients (in numpy array form) back into apply_gradients.
Any ideas at what might be happening?
OOM happens because you're constructing the graph inside the loop: This builds a graph with 20,000x nodes, and running it may need more memory than you have.
Move all TF operations that build the graph outside the loop, i.e. everything except feed_dict construction and sess.run calls.
Reply to comments
Apply gradients builds the graph?
Yes, if you look in the docs:
Returns:
An `Operation` that applies the specified gradients. If `global_step`
was not None, that operation also increments `global_step`.
Say I have access to a number of GPUs in a single machine (for the sake of argument assume 8GPUs each with max memory of 8GB each in one single machine with some amount of RAM and disk). I wanted to run in one single script and in one single machine a program that evaluates multiple models (say 50 or 200) in TensorFlow, each with a different hyper parameter setting (say, step-size, decay rate, batch size, epochs/iterations, etc). At the end of training assume we just record its accuracy and get rid of the model (if you want assume the model is being check pointed every so often, so its fine to just throw away the model and start training from scratch. You may also assume some other data may be recorded like the specific hyper params, train, validation, train errors are recorded as we train etc).
Currently I have a (pseudo-)script that looks as follow:
def train_multiple_modles_in_one_script_with_gpu(arg):
'''
trains multiple NN models in one session using GPUs correctly.
arg = some obj/struct with the params for trianing each of the models.
'''
#### try mutliple models
for mdl_id in range(100):
#### define/create graph
graph = tf.Graph()
with graph.as_default():
### get mdl
x = tf.placeholder(float_type, get_x_shape(arg), name='x-input')
y_ = tf.placeholder(float_type, get_y_shape(arg))
y = get_mdl(arg,x)
### get loss and accuracy
loss, accuracy = get_accuracy_loss(arg,x,y,y_)
### get optimizer variables
opt = get_optimizer(arg)
train_step = opt.minimize(loss, global_step=global_step)
#### run session
with tf.Session(graph=graph) as sess:
# train
for i in range(nb_iterations):
batch_xs, batch_ys = get_batch_feed(X_train, Y_train, batch_size)
sess.run(fetches=train_step, feed_dict={x: batch_xs, y_: batch_ys})
# check_point mdl
if i % report_error_freq == 0:
sess.run(step.assign(i))
#
train_error = sess.run(fetches=loss, feed_dict={x: X_train, y_: Y_train})
test_error = sess.run(fetches=loss, feed_dict={x: X_test, y_: Y_test})
print( 'step %d, train error: %s test_error %s'%(i,train_error,test_error) )
essentially it tries lots of models in one single run but it builds each model in a separate graph and runs each one in a separate session.
I guess my main worry is that its unclear to me how tensorflow under the hood allocates resources for the GPUs to be used. For example, does it load the (part of the) data set only when a session is ran? When I create a graph and a model, is it brought in the GPU immediately or when is it inserted in the GPU? Do I need to clear/free the GPU each time it tries a new model? I don't actually care too much if the models are ran in parallel in multiple GPU (which can be a nice addition), but I want it to first run everything serially without crashing. Is there anything special I need to do for this to work?
Currently I am getting an error that starts as follow:
I tensorflow/core/common_runtime/bfc_allocator.cc:702] Stats:
Limit: 340000768
InUse: 336114944
MaxInUse: 339954944
NumAllocs: 78
MaxAllocSize: 335665152
W tensorflow/core/common_runtime/bfc_allocator.cc:274] ***************************************************xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx
W tensorflow/core/common_runtime/bfc_allocator.cc:275] Ran out of memory trying to allocate 160.22MiB. See logs for memory state.
W tensorflow/core/framework/op_kernel.cc:975] Resource exhausted: OOM when allocating tensor with shape[60000,700]
and further down the line it says:
ResourceExhaustedError (see above for traceback): OOM when allocating tensor with shape[60000,700]
[[Node: standardNN/NNLayer1/Z1/add = Add[T=DT_FLOAT, _device="/job:localhost/replica:0/task:0/gpu:0"](standardNN/NNLayer1/Z1/MatMul, b1/read)]]
I tensorflow/core/common_runtime/gpu/gpu_device.cc:975] Creating TensorFlow device (/gpu:0) -> (device: 0, name: Tesla P100-SXM2-16GB, pci bus id: 0000:06:00.0)
however further down the output file (where it prints) it seems to print fine the errors/messages that should show as training proceeds. Does this mean that it didn't run out of resources? Or was it actually able to use the GPU? If it was able to use the CPU instead of the CPU, when why is this an error only happening when GPU are about to be used?
The weird thing is that the data set is really not that big (all 60K points are 24.5M) and when I run a single model locally in my own computer it seems that the process uses less than 5GB. The GPUs have at least 8GB and the computer with them has plenty of RAM and disk (at least 16GB). Thus, the errors that tensorflow is throwing at me are quite puzzling. What is it trying to do and why are they occurring? Any ideas?
After reading the answer that suggests to use the multiprocessing library I came up with the following script:
def train_mdl(args):
train(mdl,args)
if __name__ == '__main__':
for mdl_id in range(100):
# train one model with some specific hyperparms (assume they are chosen randomly inside the funciton bellow or read from a config file or they could just be passed or something)
p = Process(target=train_mdl, args=(args,))
p.start()
p.join()
print('Done training all models!')
honestly I am not sure why his answer suggests to use pool, or why there are weird tuple brackets but this is what would make sense for me. Would the resources for tensorflow be re-allocated every time a new process is created in the above loop?
I think that running all models in one single script can be bad practice in the long term (see my suggestion below for a better alternative). However, if you would like to do it, here is a solution: You can encapsulate your TF session into a process with the multiprocessing module, this will make sure TF releases the session memory once the process is done. Here is a code snippet:
from multiprocessing import Pool
import contextlib
def my_model((param1, param2, param3)): # Note the extra (), required by the pool syntax
< your code >
num_pool_worker=1 # can be bigger than 1, to enable parallel execution
with contextlib.closing(Pool(num_pool_workers)) as po: # This ensures that the processes get closed once they are done
pool_results = po.map_async(my_model,
((param1, param2, param3)
for param1, param2, param3 in params_list))
results_list = pool_results.get()
Note from OP: The random number generator seed does not reset automatically with the multi-processing library if you choose to use it. Details here: Using python multiprocessing with different random seed for each process
About TF resource allocation: Usually TF allocates much more resources than it needs. Many times you can restrict each process to use a fraction of the total GPU memory, and discover through trial and error the fraction your script requires.
You can do it with the following snippet
gpu_memory_fraction = 0.3 # Choose this number through trial and error
gpu_options = tf.GPUOptions(per_process_gpu_memory_fraction=gpu_memory_fraction,)
session_config = tf.ConfigProto(gpu_options=gpu_options)
sess = tf.Session(config=session_config, graph=graph)
Note that sometimes TF increases the memory usage in order to accelerate the execution. Therefore, reducing the memory usage might make your model run slower.
Answers to the new questions in your edit/comments:
Yes, Tensorflow will be re-allocated every time a new process is created, and cleared once a process ends.
The for-loop in your edit should also do the job. I suggest to use Pool instead, because it will enable you to run several models concurrently on a single GPU. See my notes about setting gpu_memory_fraction and "choosing the maximal number of processes". Also note that: (1) The Pool map runs the loop for you, so you don't need an outer for-loop once you use it. (2) In your example, you should have something like mdl=get_model(args) before calling train()
Weird tuple parenthesis: Pool only accepts a single argument, therefore we use a tuple to pass multiple arguments. See multiprocessing.pool.map and function with two arguments for more details. As suggested in one answer, you can make it more readable with
def train_mdl(params):
(x,y)=params
< your code >
As #Seven suggested, you can use CUDA_VISIBLE_DEVICES environment variable to choose which GPU to use for your process. You can do it from within your python script using the following on the beginning of the process function (train_mdl).
import os # the import can be on the top of the python script
os.environ["CUDA_VISIBLE_DEVICES"] = "{}".format(gpu_id)
A better practice for executing your experiments would be to isolate your training/evaluation code from the hyper parameters/ model search code.
E.g. have a script named train.py, which accepts a specific combination of hyper parameters and references to your data as arguments, and executes training for a single model.
Then, to iterate through the all the possible combinations of parameters you can use a simple task (jobs) queue, and submit all the possible combinations of hyper-parametrs as separate jobs. The task queue will feed your jobs one at a time to your machine. Usually, you can also set the queue to execute number of processes concurrently (see details below).
Specifically, I use task spooler, which is super easy to install and handful (doesn't requires admin privileges, details below).
Basic usage is (see notes below about task spooler usage):
ts <your-command>
In practice, I have a separate python script that manages my experiments, set all the arguments per specific experiment and send the jobs to the ts queue.
Here are some relevant snippets of python code from my experiments manager:
run_bash executes a bash command
def run_bash(cmd):
p = subprocess.Popen(cmd, shell=True, stdout=subprocess.PIPE, executable='/bin/bash')
out = p.stdout.read().strip()
return out # This is the stdout from the shell command
The next snippet sets the number of concurrent processes to be run (see note below about choosing the maximal number of processes):
max_job_num_per_gpu = 2
run_bash('ts -S %d'%max_job_num_per_gpu)
The next snippet iterates through a list of all combinations of hyper params / model params. Each element of the list is a dictionary, where the keys are the command line arguments for the train.py script
for combination_dict in combinations_list:
job_cmd = 'python train.py ' + ' '.join(
['--{}={}'.format(flag, value) for flag, value in combination_dict.iteritems()])
submit_cmd = "ts bash -c '%s'" % job_cmd
run_bash(submit_cmd)
A note about about choosing the maximal number of processes:
If you are short on GPUs, you can use gpu_memory_fraction you found, to set the number of processes as max_job_num_per_gpu=int(1/gpu_memory_fraction)
Notes about task spooler (ts):
You could set the number of concurrent processes to run ("slots") with:
ts -S <number-of-slots>
Installing ts doesn't requires admin privileges. You can download and compile it from source with a simple make, add it to your path and you're done.
You can set up multiple queues (I use it for multiple GPUs), with
TS_SOCKET=<path_to_queue_name> ts <your-command>
e.g.
TS_SOCKET=/tmp/socket-ts.gpu_queue_1 ts <your-command>
TS_SOCKET=/tmp/socket-ts.gpu_queue_2 ts <your-command>
See here for further usage example
A note about automatically setting the path names and file names:
Once you separate your main code from the experiment manager, you will need an efficient way to generate file names and directory names, given the hyper-params. I usually keep my important hyper params in a dictionary and use the following function to generate a single chained string from the dictionary key-value pairs.
Here are the functions I use for doing it:
def build_string_from_dict(d, sep='%'):
"""
Builds a string from a dictionary.
Mainly used for formatting hyper-params to file names.
Key-value pairs are sorted by the key name.
Args:
d: dictionary
Returns: string
:param d: input dictionary
:param sep: key-value separator
"""
return sep.join(['{}={}'.format(k, _value2str(d[k])) for k in sorted(d.keys())])
def _value2str(val):
if isinstance(val, float):
# %g means: "Floating point format.
# Uses lowercase exponential format if exponent is less than -4 or not less than precision,
# decimal format otherwise."
val = '%g' % val
else:
val = '{}'.format(val)
val = re.sub('\.', '_', val)
return val
As I understand, firstly tensorflow constructs a symbolic graph and infers the derivatives based on chain rule. Then allocates memory for all (necessary) tensors, including some inputs and outputs of layers for efficiency. When running a session, data will be loaded into the graph but in general, memory use will not change any more.
The error you met, I guess, may be caused by constructing several models in one GPU.
Isolating your training/evaluation code from the hyper parameters is a good choice, as #user2476373 proposed. But I am using bash script directly, not task spooler (may be it's more convenient), e.g.
CUDA_VISIBLE_DEVICES=0 python train.py --lrn_rate 0.01 --weight_decay_rate 0.001 --momentum 0.9 --batch_size 8 --max_iter 60000 --snapshot 5000
CUDA_VISIBLE_DEVICES=0 python eval.py
Or you can write a 'for' loop in the bash script, not necessarily in python script. Noting that I used CUDA_VISIBLE_DEVICES=0 at beginning of the script (the index could be 7 if you have 8 GPUs in one machine). Because based on my experience, I've found that tensorflow uses all GPUs in one machine if I didn't specify operations use which GPU with the code like this
with tf.device('/gpu:0'):
If you want to try multi-GPU implementation, there is some example.
Hope this could help you.
An easy solution: Give each model a unique session and graph.
It works for this platform: TensorFlow 1.12.0, Keras 2.1.6-tf, Python 3.6.7, Jupyter Notebook.
Key code:
with session.as_default():
with session.graph.as_default():
# do something about an ANN model
Full code:
import tensorflow as tf
from tensorflow import keras
import gc
def limit_memory():
""" Release unused memory resources. Force garbage collection """
keras.backend.clear_session()
keras.backend.get_session().close()
tf.reset_default_graph()
gc.collect()
#cfg = tf.ConfigProto()
#cfg.gpu_options.allow_growth = True
#keras.backend.set_session(tf.Session(config=cfg))
keras.backend.set_session(tf.Session())
gc.collect()
def create_and_train_ANN_model(hyper_parameter):
print('create and train my ANN model')
info = { 'result about this ANN model' }
return info
for i in range(10):
limit_memory()
session = tf.Session()
keras.backend.set_session(session)
with session.as_default():
with session.graph.as_default():
hyper_parameter = { 'A set of hyper-parameters' }
info = create_and_train_ANN_model(hyper_parameter)
limit_memory()
Inspired by this link: Keras (Tensorflow backend) Error - Tensor input_1:0, specified in either feed_devices or fetch_devices was not found in the Graph
I have the same issue. My solution is to run from another script doing the following as many times and in as many hyperparameter configurations as you want.
cmd = "python3 ./model_train.py hyperparameters"
os.system(cmd)
You probably don't want to do this.
If you run thousands and thousands of models on your data, and pick the one that evaluates best, you are not doing machine learning; instead you are memorizing your data set, and there is no guarantee that the model you pick will perform at all outside that data set.
In other words, that approach is similar to having a single model, which has thousands of degrees of liberty. Having a model with such high order of complexity is problematic, since it will be able to fit your data better than is actually warranted; such a model is annoyingly able to memorize any noise (outliers, measurement errors, and such) in your training data, which causes the model to perform poorly when the noise is even slightly different.
(Apologies for posting this as an answer, the site wouldn't let me add a comment.)
I'd like to write one decoder for both training (should pass gradient down to the encoder) and beam-search mode (single steps from python, sadly, so not linked to the encoder directly).
Ideally, something like this would work:
decoder(beamSearchFlag_boolPlaceholder, initalState_fromEncoder, initialState_placeholder, input):
initialState = tf.cond(beamSearchFlag_boolPlaceholder,
lambda: initialState_placeholder,
lambda: initalState_fromEncoder)
... = cell(input, initialState)
But with cond() TF still needs to resolve the dependencies of both branches. The _fromEncoder branch is executed when beamSearchFlag==False, even without effect, and that's a big part of unnecessary graph. Is there a way around this?
Today, I got a really weird thing.
I load a caffe model, feed input, net.forward, check the output data, perfect.
Then, I feed labels to the bottom layer blobs.diff, net.backward, then check the gradients (params.diff) with the result from same model caffe c++ program. They were different.
Further, when I continued to run net.backward several times at python, each time I got different gradients. This is not the case for C++ programs, they keep the same no matter how many time you run net.backward, as long as you did not change the bottom diff.
I check the bottom layer's blobs and diff, they kept unchanged both in python and C++ programs, and weights were also unchanged. This was really weird.
Anyone can provide some hints? I can provide codes if it is necessary.
Here is part of the codes :
def train_one_step(X, y, lr) :
net.blobs['data'].data[...] = X
#Forward, to get the softmax output
output = net.forward()
prob = output['prob']
#Calculate the loss of cross entropy loss function
net.blobs['prob'].diff[:] = y[:] - prob[:]
#Calculate the gradients of net parameter
net.backward()
#Renew weights based on gradients and learning rate
for key in net.params:
net.params[key][0].data[:] += lr * net.params[key][0].diff[:]
net.params[key][1].data[:] += lr * net.params[key][1].diff[:]
return loss, prob
I just want to dig out my own step function (the step of solver), so I can make some trick on the loss before it backwards, and something else. I know this is quite low efficient, data between GPU, CPU exchanged a lot.
In order to test it, I kept input the same sample(same X, y), you get different diff data. That means this function cannot work.