I am running into ValueError: Tensor("conv2d_1/kernel:0", ...) must be from the same graph as Tensor("IteratorGetNext:0", ...). I am trying to reuse a keras model with Estimator class.
I tried enclosing everything possible into
g = tf.Graph()
with g.as_default():
import tensorflow as tf
g = tf.Graph()
with g.as_default():
MODEL = get_keras_model(...)
def model_fn(mode, features, labels, params):
logits = MODEL(features)
...
def parser(record):
...
def get_dataset_inp_fn(filenames, epochs=20):
def dataset_input_fn():
dataset = tf.contrib.data.TFRecordDataset(filenames)
dataset = dataset.map(parser)
...
with tf.Session(graph=g) as sess:
est = tf.estimator.Estimator(
model_fn,
model_dir=None,
config=None,
params={"optimizer": "AdamOptimizer",
"opt_params":{}}
)
est.train(get_dataset_inp_fn(["mydata.tfrecords"],epochs=20))
but that is not helpful.
Is there a way to list all graphs defined up to current point?
Here's a general debugging technique, put import pdb; pdb.set_trace() into tf.Graph constructor, and then use bt to figure out who is creating the Graph. My first guess would that Keras does not use the default graph and creates its own. You can do inspect.getsourcefile(tf.Graph) to find where Graph file is located locally
The function that checks the graphs and returns the error (wish they return the graph addresses as well) calls following function to check the graphs:
from tensorflow.python.framework.ops import _get_graph_from_inputs
_get_graph_from_inputs([x])
In this case the graph that keras has created is identical to graph g, but one that is created by get_dataset_inp_fn is different from g.
Related
I am trying to write a function that runs KMeans on a dataset and outputs the cluster centroids. My aim is to use this in a custom keras layer, so I am using TensorFlow's implementation of KMeans that takes a tensor as the input dataset.
My problem however is that I can't make it work even as a standalone function. The problem comes from the fact that KMeans accepts a generator function that provides mini-batches instead of a plain tensor, but when I am using closure to do that, I get a graph disconnected error:
import tensorflow as tf # version: 2.4.1
from tensorflow.compat.v1.estimator.experimental import KMeans
#tf.function
def KMeansCentroids(inputs, num_clusters, steps, use_mini_batch=False):
# `inputs` is a 2D tensor
def input_fn():
# Each one of the lines below results in the same "Graph Disconnected" error. Tuples don't really needed but just to be consistent with the documentation
return (inputs, None)
return (tf.data.Dataset.from_tensor_slices(inputs), None)
return (tf.convert_to_tensor(inputs), None)
kmeans = KMeans(
num_clusters=num_clusters,
use_mini_batch=use_mini_batch)
kmeans.train(input_fn, steps=steps) # This is where the error happens
return kmeans.cluster_centers()
>>> x = tf.random.uniform((100, 2))
>>> c = KMeansCentroids(x, 5, 10)
The exact error is:
ValueError:
Tensor("strided_slice:0", shape=(), dtype=int32)
must be from the same graph as
Tensor("Equal:0", shape=(), dtype=bool)
(graphs are FuncGraph(name=KMeansCentroids, id=..) and <tensorflow.python.framework.ops.Graph object at ...>).
If I were to use a numpy dataset and convert to tensor inside the function, the code would work just fine.
Also, making input_fn() return directly tf.random.uniform((100, 2)) (ignoring the inputs argument), would again work. That's why I am guessing that tensorflow doesn't support closures since it needs to build the computation graph at the beginning.
But I don't see how to work around that.
Could it be a version error due to KMeans being a compat.v1.experimental module?
Note that the documentation of KMeans states for the input_fn():
The function should construct and return one of the following:
A tf.data.Dataset object: Outputs of Dataset object must be a tuple (features, labels) with same constraints as below.
A tuple (features, labels): Where features is a tf.Tensor or a dictionary of string feature name to Tensor and labels is a Tensor or a dictionary of string label name to Tensor. Both features and labels are consumed by model_fn. They should satisfy the expectation of model_fn from inputs.
The problem you're facing is more about invoking tensor outside the created graph. Basically, when you called the .train function, a new graph will be created and that is with the graph defined in that input_fn and the graph defined in the model_fn.
kmeans.train(input_fn, steps=steps)
And, after that all the tensors those coming outside these functions will be treated as outsiders and won't part of this new graph. That's why you're getting a graph disconnected error for trying to use outsider tensor. To resolve this, you need to create the necessary tensors within these graphs.
import tensorflow as tf
from tensorflow.compat.v1.estimator.experimental import KMeans
#tf.function
def KMeansCentroids(num_clusters, steps, use_mini_batch=False):
def input_fn(batch_size):
pinputs = tf.random.uniform((100, 2))
dataset = tf.data.Dataset.from_tensor_slices((pinputs))
dataset = dataset.shuffle(1000).repeat()
return dataset.batch(batch_size)
kmeans = KMeans(
num_clusters=num_clusters,
use_mini_batch=use_mini_batch)
kmeans.train(input_fn = lambda: input_fn(5),
steps=steps)
return kmeans.cluster_centers()
c = KMeansCentroids(5, 10)
Here is some more info for reading, 1. FYI, I tested your code with a few versions of tf > 2, and I don't think it's related to version error or something.
Re-mentioning here for future readers. An alternative of using KMeans within Keras layers:
tf_kmeans.py
ClusteringLayer
I want to implement a custom optimization algorithm for TF models.
I have read the following sources
tf documentation on custom optimizers
tf SGD implementation
keras documentation on custom models
towardsdatascience guide on custom optimizers
However lot of questions remain.
It seems like it is not possible to evaluate the loss function multiple times (for different weight settings) before applying a gradient step, when using the custom optimizer API. For example in a line-search type of algorithm this is necessary.
I tried to do all steps manually.
Assume I have setup my model and my optimization problem like this
from tensorflow.keras import layers
from tensorflow.keras import losses
from tensorflow.keras import models
model = models.Sequential()
model.add(layers.Dense(15, input_dim=10))
model.add(layers.Dense(20))
model.add(layers.Dense(1))
x_train, y_train = get_train_data()
loss = losses.MeanSquaredError()
def val_and_grads(weights):
model.set_weights(weights)
with tf.GradientTape() as tape:
val = loss(y_train, model(x_train))
grads = tape.gradient(val, model.trainable_variables)
return val, grads
initial_weights = model.get_weights()
optimal_weigths = my_fancy_optimization_algorithm(val_and_grads, initial_weights)
However my function val_and_grads needs a list of weights and returns a list of gradients from my_fancy_optimization_algorithms point of view that seems unnatural.
I could warp val_and_grads to "stack" the returned gradients and "split" the passed weights like this
def wrapped_val_and_grad(weights):
grads = val_and_grads(split_weights(weights))
return stack_grads(grads)
however that seems very inefficient.
Anyway, I do not like this approach since it seems that I would loose out out on a lot of the surrounding tensorflow infrastructure (printing of current loss function values and metrics during learning, tensorboard stuff, ...).
I could also pack the above in a custom model with a tailored train_step like this
def CustomModel(keras.Model):
def train_step(self, data):
x_train, y_train = data
def val_and_grads(weights):
self.set_weights(weights)
with tf.GradientTape() as tape:
val = loss(y_train, self(x_train))
grads = tape.gradient(val, self.trainable_variables)
return val, grads
trainable_vars = self.trainable_variables
old_weights = self.get_weights()
update = my_fancy_update_finding_algorithm(val_and_grads, self.get_weights()) # this can do multiple evaluations of the model
self.set_weights(old_weights) # restore the weights
self.optimizer.apply_gradients(zip(update, trainable_vars))
Here I would need a accompanying custom optimizer that does nothing else than updating the current weights by adding the update (new_weigths = current_weights + update).
I am still unsure if this is the best way to go.
If someone can comment on the snippets and ideas above, guide me to any other resource that I should consider or provide new approaches and other feedback I would be very glad.
Thanks all.
Franz
EDIT:
Sadly I did not get any response here so far. Maybe my question is not concrete enough. As a first smaller question:
Given the model and val_and_grads in the first listing. How would I efficiently calculate the norm of the WHOLE gradient? What I do so far is
import numpy as np
_, grads = val_and_grad(model.get_weights())
norm_grads = np.linalg.norm(np.concatenate([grad.numpy().flatten() for grad in grad]))
This surely cannot be the "right" way.
I have trained a Keras model based on this repo.
After the training I save the model as checkpoint files like this:
sess=tf.keras.backend.get_session()
saver = tf.train.Saver()
saver.save(sess, current_run_path + '/checkpoint_files/model_{}.ckpt'.format(date))
Then I restore the graph from the checkpoint files and freeze it using the standard tf freeze_graph script. When I want to restore the frozen graph I get the following error:
Input 0 of node Conv_BN_1/cond/ReadVariableOp/Switch was passed float from Conv_BN_1/gamma:0 incompatible with expected resource
How can I fix this issue?
Edit: My problem is related to this question. Unfortunately, I can't use the workaround.
Edit 2:
I have opened an issue on github and created a gist to reproduce the error.
https://github.com/keras-team/keras/issues/11032
Just resolved the same issue. I connected this few answers: 1, 2, 3 and realized that issue originated from batchnorm layer working state: training or learning. So, in order to resolve that issue you just need to place one line before loading your model:
keras.backend.set_learning_phase(0)
Complete example, to export model
import tensorflow as tf
from tensorflow.python.framework import graph_io
from tensorflow.keras.applications.inception_v3 import InceptionV3
def freeze_graph(graph, session, output):
with graph.as_default():
graphdef_inf = tf.graph_util.remove_training_nodes(graph.as_graph_def())
graphdef_frozen = tf.graph_util.convert_variables_to_constants(session, graphdef_inf, output)
graph_io.write_graph(graphdef_frozen, ".", "frozen_model.pb", as_text=False)
tf.keras.backend.set_learning_phase(0) # this line most important
base_model = InceptionV3()
session = tf.keras.backend.get_session()
INPUT_NODE = base_model.inputs[0].op.name
OUTPUT_NODE = base_model.outputs[0].op.name
freeze_graph(session.graph, session, [out.op.name for out in base_model.outputs])
to load *.pb model:
from PIL import Image
import numpy as np
import tensorflow as tf
# https://i.imgur.com/tvOB18o.jpg
im = Image.open("/home/chichivica/Pictures/eagle.jpg").resize((299, 299), Image.BICUBIC)
im = np.array(im) / 255.0
im = im[None, ...]
graph_def = tf.GraphDef()
with tf.gfile.GFile("frozen_model.pb", "rb") as f:
graph_def.ParseFromString(f.read())
graph = tf.Graph()
with graph.as_default():
net_inp, net_out = tf.import_graph_def(
graph_def, return_elements=["input_1", "predictions/Softmax"]
)
with tf.Session(graph=graph) as sess:
out = sess.run(net_out.outputs[0], feed_dict={net_inp.outputs[0]: im})
print(np.argmax(out))
This is bug with Tensorflow 1.1x and as another answer stated, it is because of the internal batch norm learning vs inference state. In TF 1.14.0 you actually get a cryptic error when trying to freeze a batch norm layer.
Using set_learning_phase(0) will put the batch norm layer (and probably others like dropout) into inference mode and thus the batch norm layer will not work during training, leading to reduced accuracy.
My solution is this:
Create the model using a function (do not use K.set_learning_phase(0)):
def create_model():
inputs = Input(...)
...
return model
model = create_model()
Train model
Save weights:
model.save_weights("weights.h5")
Clear session (important so layer names are the same) and set learning phase to 0:
K.clear_session()
K.set_learning_phase(0)
Recreate model and load weights:
model = create_model()
model.load_weights("weights.h5")
Freeze as before
Thanks for pointing the main issue! I found that keras.backend.set_learning_phase(0) to be not working sometimes, at least in my case.
Another approach might be: for l in keras_model.layers: l.trainable = False
Following the upgrade to Keras 2.0.9, I have been using the multi_gpu_model utility but I can't save my models or best weights using
model.save('path')
The error I get is
TypeError: can’t pickle module objects
I suspect there is some problem gaining access to the model object. Is there a work around this issue?
To be honest, the easiest approach to this is to actually examine the multi gpu parallel model using
parallel_model.summary()
(The parallel model is simply the model after applying the multi_gpu function). This clearly highlights the actual model (in I think the penultimate layer - I am not at my computer right now). Then you can use the name of this layer to save the model.
model = parallel_model.get_layer('sequential_1)
Often its called sequential_1 but if you are using a published architecture, it may be 'googlenet' or 'alexnet'. You will see the name of the layer from the summary.
Then its simple to just save
model.save()
Maxims approach works, but its overkill I think.
Rem: you will need to compile both the model, and the parallel model.
Workaround
Here's a patched version that doesn't fail while saving:
from keras.layers import Lambda, concatenate
from keras import Model
import tensorflow as tf
def multi_gpu_model(model, gpus):
if isinstance(gpus, (list, tuple)):
num_gpus = len(gpus)
target_gpu_ids = gpus
else:
num_gpus = gpus
target_gpu_ids = range(num_gpus)
def get_slice(data, i, parts):
shape = tf.shape(data)
batch_size = shape[:1]
input_shape = shape[1:]
step = batch_size // parts
if i == num_gpus - 1:
size = batch_size - step * i
else:
size = step
size = tf.concat([size, input_shape], axis=0)
stride = tf.concat([step, input_shape * 0], axis=0)
start = stride * i
return tf.slice(data, start, size)
all_outputs = []
for i in range(len(model.outputs)):
all_outputs.append([])
# Place a copy of the model on each GPU,
# each getting a slice of the inputs.
for i, gpu_id in enumerate(target_gpu_ids):
with tf.device('/gpu:%d' % gpu_id):
with tf.name_scope('replica_%d' % gpu_id):
inputs = []
# Retrieve a slice of the input.
for x in model.inputs:
input_shape = tuple(x.get_shape().as_list())[1:]
slice_i = Lambda(get_slice,
output_shape=input_shape,
arguments={'i': i,
'parts': num_gpus})(x)
inputs.append(slice_i)
# Apply model on slice
# (creating a model replica on the target device).
outputs = model(inputs)
if not isinstance(outputs, list):
outputs = [outputs]
# Save the outputs for merging back together later.
for o in range(len(outputs)):
all_outputs[o].append(outputs[o])
# Merge outputs on CPU.
with tf.device('/cpu:0'):
merged = []
for name, outputs in zip(model.output_names, all_outputs):
merged.append(concatenate(outputs,
axis=0, name=name))
return Model(model.inputs, merged)
You can use this multi_gpu_model function, until the bug is fixed in keras. Also, when loading the model, it's important to provide the tensorflow module object:
model = load_model('multi_gpu_model.h5', {'tf': tf})
How it works
The problem is with import tensorflow line in the middle of multi_gpu_model:
def multi_gpu_model(model, gpus):
...
import tensorflow as tf
...
This creates a closure for the get_slice lambda function, which includes the number of gpus (that's ok) and tensorflow module (not ok). Model save tries to serialize all layers, including the ones that call get_slice and fails exactly because tf is in the closure.
The solution is to move import out of multi_gpu_model, so that tf becomes a global object, though still needed for get_slice to work. This fixes the problem of saving, but in loading one has to provide tf explicitly.
It's something that need a little work around by loading the multi_gpu_model weight to the regular model weight.
e.g.
#1, instantiate your base model on a cpu
with tf.device("/cpu:0"):
model = create_model()
#2, put your model to multiple gpus, say 2
multi_model = multi_gpu_model(model, 2)
#3, compile both models
model.compile(loss=your_loss, optimizer=your_optimizer(lr))
multi_model.compile(loss=your_loss, optimizer=your_optimizer(lr))
#4, train the multi gpu model
# multi_model.fit() or multi_model.fit_generator()
#5, save weights
model.set_weights(multi_model.get_weights())
model.save(filepath=filepath)
`
refrence: https://github.com/fchollet/keras/issues/8123
I can't achieve to make summaries work with the Estimator API of Tensorflow.
The Estimator class is very useful for many reasons: I have already implemented my own classes which are really similar but I am trying to switch to this one.
Here is the code sample:
import tensorflow as tf
import tensorflow.contrib.layers as layers
import tensorflow.contrib.learn as learn
import numpy as np
# To reproduce the error: docker run --rm -w /algo -v $(pwd):/algo tensorflow/tensorflow bash -c "python sample.py"
def model_fn(x, y, mode):
logits = layers.fully_connected(x, 12, scope="dense-1")
logits = layers.fully_connected(logits, 56, scope="dense-2")
logits = layers.fully_connected(logits, 4, scope="dense-3")
loss = tf.reduce_mean(tf.nn.sparse_softmax_cross_entropy_with_logits(logits=logits, labels=y), name="xentropy")
return {"predictions":logits}, loss, tf.train.AdamOptimizer(0.001).minimize(loss)
def input_fun():
""" To be completed for a 4 classes classification problem """
feature = tf.constant(np.random.rand(100,10))
labels = tf.constant(np.random.random_integers(0,3, size=(100,)))
return feature, labels
estimator = learn.Estimator(model_fn=model_fn, )
trainingConfig = tf.contrib.learn.RunConfig(save_checkpoints_secs=60)
estimator = learn.Estimator(model_fn=model_fn, model_dir="./tmp", config=trainingConfig)
# Works
estimator.fit(input_fn=input_fun, steps=2)
# The following code does not work
# Can't initialize saver
# saver = tf.train.Saver(max_to_keep=10) # Error: No variables to save
# The following fails because I am missing a saver... :(
hooks=[
tf.train.LoggingTensorHook(["xentropy"], every_n_iter=100),
tf.train.CheckpointSaverHook("./tmp", save_steps=1000, checkpoint_basename='model.ckpt'),
tf.train.StepCounterHook(every_n_steps=100, output_dir="./tmp"),
tf.train.SummarySaverHook(save_steps=100, output_dir="./tmp"),
]
estimator.fit(input_fn=input_fun, steps=2, monitors=hooks)
As you can see, I can create an Estimator and use it but I can achieve to add hooks to the fitting process.
The logging hooks works just fine but the others require both tensors and a saver which I can't provide.
The tensors are defined in the model function, thus I can't pass them to the SummaryHook and the Saver can't be initialized because there is no tensor to save...
Is there a solution to my problem? (I am guessing yes but there is a lack of documentation of this part in the tensorflow documentation)
How can I initialized my saver? Or should I use other objects such as Scaffold?
How can I pass summaries to the SummaryHook since they are defined in my model function?
Thanks in advance.
PS: I have seen the DNNClassifier API but I want to use the estimator API for Convolutional Nets and others. I need to create summaries for any estimator.
The intended use case is that you let the Estimator save summaries for you. There are options in RunConfig for configuring summary writing. RunConfigs get passed when constructing the Estimator.
Just have tf.summary.scalar("loss", loss) in the model_fn, and run the code without summary_hook. The loss is recorded and shown in the tensorboard.
See also:
Tensorflow - Using tf.summary with 1.2 Estimator API