I'm trying to use a placeholder in my graph as a Variable (so I can later optimise something with respect to it), but I don't know the best way to do that. I've tried this:
x = tf.placeholder(tf.float32, shape = [None,1])
x_as_variable = tf.Variable(x, validate_shape = False)
but everytime I build my graph, I get an error when I try to add my loss function:
train = tf.train.AdamOptimizer().minimize(MSEloss)
The error is:
ValueError: as_list() is not defined on an unknown TensorShape.
Even if you're not familiar with the error entirely, I'd really appreciate it if you could guide me on how to build a replicate Variable which takes on the value of my placeholder.
Thanks!
As you've noticed, TensorFlow optimizers (i.e. subclasses of tf.train.Optimizer) operate on tf.Variable objects because they need to be able to assign new values to those objects, and in TensorFlow only variables support an assign operation. If you use a tf.placeholder(), there's nothing to update, because the value of a placeholder is immutable within each step.
So how do you optimize with respect to a fed-in value? I can think of two options:
Instead of feeding a tf.placeholder(), you could first assign a fed-in value to a variable and then optimize with respect to it:
var = tf.Variable(...)
set_var_placeholder = tf.placeholder(tf.float32, ...)
set_var_op = var.assign(set_var_placeholder)
# ...
train_op = tf.train.AdamOptimizer(...).minimize(mse_loss, var_list=[var, ...])
# ...
initial_val = ... # A NumPy array.
sess.run(set_var_op, feed_dict={set_var_placeholder: initial_val})
sess.run(train_op)
updated_val = sess.run(var)
You could use the lower-level tf.gradients() function to get the gradient of the loss with respect to a placeholder in a single step. You could then use that gradient in Python:
var = tf.placeholder(tf.float32, ...)
# ...
mse_loss = ...
var_grad, = tf.gradients(loss, [var])
var_grad_val = sess.run(var_grad, feed_dict={var: ...})
PS. The code in your question, where you define a tf.Variable(tf.placeholder(...), ...) is just defining a variable whose initial value is fed by the placeholder. This probably isn't what you want, because the training op that the optimizer creates will only use the value assigned to the variable, and ignore whatever you feed to the placeholder (after the initialization step).
Related
I am trying to create a physics-informed neural network (PINN) in JAX. I want to differentiate the defined model (neural network) by the input (x). If I set model to jax.grad(params), I get an error.
If I set model to jax.grad(model), I don't get an error, but I don't know if I am able to differentiate the model of the neural network by x.
class MLP(fnn.Module):
#fnn.compact
def __call__(self, x):
x = fnn.Dense(128)(x)
x = fnn.relu(x)
x = fnn.Dense(256)(x)
x = fnn.relu(x)
x = fnn.Dense(10)(x)
return x
model = MLP()
params = model.init(jax.random.PRNGKey(0), jnp.ones([1]))['params']
tx = optax.adam(0.001)
state = TrainState.create(apply_fn=model.apply, params=params, tx=tx)
You can differentiate a model in JAX by (1) defining a function that you want to differentiate, (2) transforming it with jax.grad, jax.jacrev, jax.jacfwd, etc. as appropriate for your application, and (3) passing data to the transformed function.
It's not entirely clear from your question what operation you're hoping to differentiate, but here is an example of computing a forward-mode jacobian of the training state creation with respect to the params:
def f(params):
return TrainState.create(apply_fn=model.apply, params=params, tx=tx)
result = jax.jacfwd(f)(params)
If that doesn't help, I'd suggest editing your question to make clear what operation you're interested in differentiating.
So my goal is basically implementing global top-k subsampling. Gradient sparsification is quite simple and I have already done this building on stateful clients example, but now I would like to use encoders as you have recommended here at page 28. Additionally I would like to average only the non-zero gradients, so say we have 10 clients but only 4 have nonzero gradients at a given position for a communication round then I would like to divide the sum of these gradients to 4, not 10. I am hoping to achieve this by summing gradients at numerator and masks, 1s and 0s, at denominator. Also moving forward I will add randomness to gradient selection so it is imperative that I create those masks concurrently with gradient selection. The code I have right now is
import tensorflow as tf
from tensorflow_model_optimization.python.core.internal import tensor_encoding as te
#te.core.tf_style_adaptive_encoding_stage
class GrandienrSparsificationEncodingStage(te.core.AdaptiveEncodingStageInterface):
"""An example custom implementation of an `EncodingStageInterface`.
Note: This is likely not what one would want to use in practice. Rather, this
serves as an illustration of how a custom compression algorithm can be
provided to `tff`.
This encoding stage is expected to be run in an iterative manner, and
alternatively zeroes out values corresponding to odd and even indices. Given
the determinism of the non-zero indices selection, the encoded structure does
not need to be represented as a sparse vector, but only the non-zero values
are necessary. In the decode mehtod, the state (i.e., params derived from the
state) is used to reconstruct the corresponding indices.
Thus, this example encoding stage can realize representation saving of 2x.
"""
ENCODED_VALUES_KEY = 'stateful_topk_values'
INDICES_KEY = 'indices'
SHAPES_KEY = 'shapes'
ERROR_COMPENSATION_KEY = 'error_compensation'
def encode(self, x, encode_params):
shapes_list = [tf.shape(y) for y in x]
flattened = tf.nest.map_structure(lambda y: tf.reshape(y, [-1]), x)
gradients = tf.concat(flattened, axis=0)
error_compensation = encode_params[self.ERROR_COMPENSATION_KEY]
gradients_and_error_compensation = tf.math.add(gradients, error_compensation)
percentage = tf.constant(0.1, dtype=tf.float32)
k_float = tf.multiply(percentage, tf.cast(tf.size(gradients_and_error_compensation), tf.float32))
k_int = tf.cast(tf.math.round(k_float), dtype=tf.int32)
values, indices = tf.math.top_k(tf.math.abs(gradients_and_error_compensation), k = k_int, sorted = False)
indices = tf.expand_dims(indices, 1)
sparse_gradients_and_error_compensation = tf.scatter_nd(indices, values, tf.shape(gradients_and_error_compensation))
new_error_compensation = tf.math.subtract(gradients_and_error_compensation, sparse_gradients_and_error_compensation)
state_update_tensors = {self.ERROR_COMPENSATION_KEY: new_error_compensation}
encoded_x = {self.ENCODED_VALUES_KEY: values,
self.INDICES_KEY: indices,
self.SHAPES_KEY: shapes_list}
return encoded_x, state_update_tensors
def decode(self,
encoded_tensors,
decode_params,
num_summands=None,
shape=None):
del num_summands, decode_params, shape # Unused.
flat_shape = tf.math.reduce_sum([tf.math.reduce_prod(shape) for shape in encoded_tensors[self.SHAPES_KEY]])
sizes_list = [tf.math.reduce_prod(shape) for shape in encoded_tensors[self.SHAPES_KEY]]
scatter_tensor = tf.scatter_nd(
indices=encoded_tensors[self.INDICES_KEY],
updates=encoded_tensors[self.ENCODED_VALUES_KEY],
shape=[flat_shape])
nonzero_locations = tf.nest.map_structure(lambda x: tf.cast(tf.where(tf.math.greater(x, 0), 1, 0), tf.float32) , scatter_tensor)
reshaped_tensor = [tf.reshape(flat_tensor, shape=shape) for flat_tensor, shape in
zip(tf.split(scatter_tensor, sizes_list), encoded_tensors[self.SHAPES_KEY])]
reshaped_nonzero = [tf.reshape(flat_tensor, shape=shape) for flat_tensor, shape in
zip(tf.split(nonzero_locations, sizes_list), encoded_tensors[self.SHAPES_KEY])]
return reshaped_tensor, reshaped_nonzero
def initial_state(self):
return {self.ERROR_COMPENSATION_KEY: tf.constant(0, dtype=tf.float32)}
def update_state(self, state, state_update_tensors):
return {self.ERROR_COMPENSATION_KEY: state_update_tensors[self.ERROR_COMPENSATION_KEY]}
def get_params(self, state):
encode_params = {self.ERROR_COMPENSATION_KEY: state[self.ERROR_COMPENSATION_KEY]}
decode_params = {}
return encode_params, decode_params
#property
def name(self):
return 'gradient_sparsification_encoding_stage'
#property
def compressible_tensors_keys(self):
return False
#property
def commutes_with_sum(self):
return False
#property
def decode_needs_input_shape(self):
return False
#property
def state_update_aggregation_modes(self):
return {}
I have run some simple tests manually following the steps you outlined here at page 45. It works but I have some questions/problems.
When I use list of tensors of same shape (ex:2 2x25 tensors) as input,x, of encode it works without any issues but when I try to use list of tensors of different shapes (2x20 and 6x10) it gives and error saying
InvalidArgumentError: Shapes of all inputs must match: values[0].shape = [2,20] != values1.shape = [6,10] [Op:Pack] name: packed
How can I resolve this issue? As i said I want to use global top-k so it is essential I encode entire trainable model weights at once. Take the cnn model used here, all the tensors have different shapes.
How can I do the averaging I described at the beginning? For example here you have done
mean_factory = tff.aggregators.MeanFactory(
tff.aggregators.EncodedSumFactory(mean_encoder_fn), # numerator
tff.aggregators.EncodedSumFactory(mean_encoder_fn), # denominator )
Is there a way to repeat this with one output of decode going to numerator and other going to denominator? How can I handle dividing 0 by 0? tensorflow has divide_no_nan function, can I use it somehow or do I need to add eps to each?
How is partition handled when I use encoders? Does each client get a unique encoder holding a unique state for it? As you have discussed here at page 6 client states are used in cross-silo settings yet what happens if client ordering changes?
Here you have recommended using stateful clients example. Can you explain this a bit further? I mean in the run_one_round where exactly encoders go and how are they used/combined with client update and aggregation?
I have some additional information such as sparsity I want to pass to encode. What is the suggested method for doing that?
Here are some answers, hope it helps:
If you want to treat all of the aggregated structure just as a single tensor, use concat_factory as the outermost aggregator. That will concatenate entire structure to a rank-1 Tensor at clients, and then unpack back to the original structure at the end. Example use: tff.aggregators.concat_factory(tff.aggregators.MeanFactory(...))
Note the encoding stage objects are meant to work with a single tensor, so what you describe with identical tensors probably works only accidentally.
There are two options.
a. Modify the client training code such that the weights being passed to the weighted aggregator are already what you want it to be (zero/one
mask). In the stateful clients example you link, that would be here. You will then get what you need by default (by summing the numerator).
b. Modify UnweightedMeanFactory to do exactly the variant of averaging you describe and use that. Start would be modifying this
(and 4.) I think that is what you would need to implement. The same way existing client states are initialized in the example here, you would need extend it to contain the aggregator states, and make sure those are sampled together with the clients, as done here. Then, to integrate the aggregators in the example you would need to replace this hard-coded tff.federated_mean. An example of such integration is in the implementation of tff.learning.build_federated_averaging_process, primarily here
I am not sure what the question is. Perhaps get the previous working (seems like a prerequisite to me), and then clarify and ask in a new post?
I am training a Naive Bayes model using the mlr package.
I would like to tune the threshold (and only the threshold) for the classification. The tutorial provides an example for doing this while also doing additional hyperparameter tuning in a nested CV-setting. I actually do not want to tune any other (hyper)parameter while finding the optimal threshold value.
Based on the discussion here I set up a makeTuneWrapper() object and set another parameter (laplace) to a fixed value (1) and subsequently run resample() in a nested CV-setting.
nbayes.lrn <- makeLearner("classif.naiveBayes", predict.type = "prob")
nbayes.lrn
nbayes.pst <- makeParamSet(makeDiscreteParam("laplace", value = 1))
nbayes.tcg <- makeTuneControlGrid(tune.threshold = TRUE)
# Inner
rsmp.cv5.desc<-makeResampleDesc("CV", iters=5, stratify=TRUE)
nbayes.lrn<- makeTuneWrapper(nbayes.lrn, par.set=nbayes.pst, control=nbayes.tcg, resampling=rsmp.cv5.desc, measures=tpr)
# Outer
rsmp.cv10.desc<-makeResampleDesc("CV", iters=10, stratify=TRUE)
nbayes.res<-resample(nbayes.lrn, beispiel3.tsk, resampling= rsmp.cv10.desc, measures=list(tpr,ppv), extract=getTuneResult)
print(nbayes.res$extract)
Setting up a resampling scheme for the inner loop in the nested CV seems superfluous. The internal call to tuneThreshold() apparently does a more thorough optimization anyhow. However, calling makeTuneWrapper() without a resampling scheme leads to an error message.
I have two specific questions:
1.) Is there a simpler way of tuning a threshold (and only the threshold)?
2.) Given the setup used above: how can I access the threshold values that were actually tested?
EDIT:
This would be a code example for tuning the threshold for different measures (accuraccy, sensitivity, precision) based on the answer by #Lars Kotthoff.
### Create fake data
y<-c(rep(0,500), rep(1,500))
x<-c(rep(0, 300), rep(1,200), rep(0,100), rep(1,400))
balanced.df<-data.frame(y=y, x=x)
balanced.df$y<-as.factor(balanced.df$y)
balanced.df$x<-as.factor(balanced.df$x)
balanced.tsk<-makeClassifTask(data=balanced.df, target="y", positive="1")
summarizeColumns(balanced.tsk)
### TuneThreshold
logreg.lrn<-makeLearner("classif.logreg", predict.type="prob")
logreg.mod<-train(logreg.lrn, balanced.tsk)
logreg.preds<-predict(logreg.mod, balanced.tsk)
threshold_tpr<-tuneThreshold(logreg.preds, measure=list(tpr))
threshold_tpr
threshold_acc<-tuneThreshold(logreg.preds, measure=list(acc))
threshold_acc
threshold_ppv<-tuneThreshold(logreg.preds, measure=list(ppv))
threshold_ppv
You can use tuneThreshold() directly:
require(mlr)
iris.model = train(makeLearner("classif.naiveBayes", predict.type = "prob"), iris.task)
iris.preds = predict(iris.model, iris.task)
res = tuneThreshold(iris.preds)
Unfortunately, you can't access the threshold values that were tested when using tuneThreshold(). You could however treat the threshold value as a "normal" hyperparameter and use any of the tuning methods in mlr. This would allow you to get the values and corresponding performance.
Is there a way to plot both the training losses and validation losses on the same graph?
It's easy to have two separate scalar summaries for each of them individually, but this puts them on separate graphs. If both are displayed in the same graph it's much easier to see the gap between them and whether or not they have begin to diverge due to overfitting.
Is there a built in way to do this? If not, a work around way? Thank you much!
The work-around I have been doing is to use two SummaryWriter with different log dir for training set and cross-validation set respectively. And you will see something like this:
Rather than displaying the two lines separately, you can instead plot the difference between validation and training losses as its own scalar summary to track the divergence.
This doesn't give as much information on a single plot (compared with adding two summaries), but it helps with being able to compare multiple runs (and not adding multiple summaries per run).
Just for anyone coming accross this via a search: The current best practice to achieve this goal is to just use the SummaryWriter.add_scalars method from torch.utils.tensorboard. From the docs:
from torch.utils.tensorboard import SummaryWriter
writer = SummaryWriter()
r = 5
for i in range(100):
writer.add_scalars('run_14h', {'xsinx':i*np.sin(i/r),
'xcosx':i*np.cos(i/r),
'tanx': np.tan(i/r)}, i)
writer.close()
# This call adds three values to the same scalar plot with the tag
# 'run_14h' in TensorBoard's scalar section.
Expected result:
Many thanks to niko for the tip on Custom Scalars.
I was confused by the official custom_scalar_demo.py because there's so much going on, and I had to study it for quite a while before I figured out how it worked.
To show exactly what needs to be done to create a custom scalar graph for an existing model, I put together the following complete example:
# + <
# We need these to make a custom protocol buffer to display custom scalars.
# See https://developers.google.com/protocol-buffers/
from tensorboard.plugins.custom_scalar import layout_pb2
from tensorboard.summary.v1 import custom_scalar_pb
# >
import tensorflow as tf
from time import time
import re
# Initial values
(x0, y0) = (-1, 1)
# This is useful only when re-running code (e.g. Jupyter).
tf.reset_default_graph()
# Set up variables.
x = tf.Variable(x0, name="X", dtype=tf.float64)
y = tf.Variable(y0, name="Y", dtype=tf.float64)
# Define loss function and give it a name.
loss = tf.square(x - 3*y) + tf.square(x+y)
loss = tf.identity(loss, name='my_loss')
# Define the op for performing gradient descent.
minimize_step_op = tf.train.GradientDescentOptimizer(0.092).minimize(loss)
# List quantities to summarize in a dictionary
# with (key, value) = (name, Tensor).
to_summarize = dict(
X = x,
Y_plus_2 = y + 2,
)
# Build scalar summaries corresponding to to_summarize.
# This should be done in a separate name scope to avoid name collisions
# between summaries and their respective tensors. The name scope also
# gives a title to a group of scalars in TensorBoard.
with tf.name_scope('scalar_summaries'):
my_var_summary_op = tf.summary.merge(
[tf.summary.scalar(name, var)
for name, var in to_summarize.items()
]
)
# + <
# This constructs the layout for the custom scalar, and specifies
# which scalars to plot.
layout_summary = custom_scalar_pb(
layout_pb2.Layout(category=[
layout_pb2.Category(
title='Custom scalar summary group',
chart=[
layout_pb2.Chart(
title='Custom scalar summary chart',
multiline=layout_pb2.MultilineChartContent(
# regex to select only summaries which
# are in "scalar_summaries" name scope:
tag=[r'^scalar_summaries\/']
)
)
])
])
)
# >
# Create session.
with tf.Session() as sess:
# Initialize session.
sess.run(tf.global_variables_initializer())
# Create writer.
with tf.summary.FileWriter(f'./logs/session_{int(time())}') as writer:
# Write the session graph.
writer.add_graph(sess.graph) # (not necessary for scalars)
# + <
# Define the layout for creating custom scalars in terms
# of the scalars.
writer.add_summary(layout_summary)
# >
# Main iteration loop.
for i in range(50):
current_summary = sess.run(my_var_summary_op)
writer.add_summary(current_summary, global_step=i)
writer.flush()
sess.run(minimize_step_op)
The above consists of an "original model" augmented by three blocks of code indicated by
# + <
[code to add custom scalars goes here]
# >
My "original model" has these scalars:
and this graph:
My modified model has the same scalars and graph, together with the following custom scalar:
This custom scalar chart is simply a layout which combines the original two scalar charts.
Unfortunately the resulting graph is hard to read because both values have the same color. (They are distinguished only by marker.) This is however consistent with TensorBoard's convention of having one color per log.
Explanation
The idea is as follows. You have some group of variables which you want to plot inside a single chart. As a prerequisite, TensorBoard should be plotting each variable individually under the "SCALARS" heading. (This is accomplished by creating a scalar summary for each variable, and then writing those summaries to the log. Nothing new here.)
To plot multiple variables in the same chart, we tell TensorBoard which of these summaries to group together. The specified summaries are then combined into a single chart under the "CUSTOM SCALARS" heading. We accomplish this by writing a "Layout" once at the beginning of the log. Once TensorBoard receives the layout, it automatically produces a combined chart under "CUSTOM SCALARS" as the ordinary "SCALARS" are updated.
Assuming that your "original model" is already sending your variables (as scalar summaries) to TensorBoard, the only modification necessary is to inject the layout before your main iteration loop starts. Each custom scalar chart selects which summaries to plot by means of a regular expression. Thus for each group of variables to be plotted together, it can be useful to place the variables' respective summaries into a separate name scope. (That way your regex can simply select all summaries under that name scope.)
Important Note: The op which generates the summary of a variable is distinct from the variable itself. For example, if I have a variable ns1/my_var, I can create a summary ns2/summary_op_for_myvar. The custom scalars chart layout cares only about the summary op, not the name or scope of the original variable.
Here is an example, creating two tf.summary.FileWriters which share the same root directory. Creating a tf.summary.scalar shared by the two tf.summary.FileWriters. At every time step, get the summary and update each tf.summary.FileWriter.
import os
import tqdm
import tensorflow as tf
def tb_test():
sess = tf.Session()
x = tf.placeholder(dtype=tf.float32)
summary = tf.summary.scalar('Values', x)
merged = tf.summary.merge_all()
sess.run(tf.global_variables_initializer())
writer_1 = tf.summary.FileWriter(os.path.join('tb_summary', 'train'))
writer_2 = tf.summary.FileWriter(os.path.join('tb_summary', 'eval'))
for i in tqdm.tqdm(range(200)):
# train
summary_1 = sess.run(merged, feed_dict={x: i-10})
writer_1.add_summary(summary_1, i)
# eval
summary_2 = sess.run(merged, feed_dict={x: i+10})
writer_2.add_summary(summary_2, i)
writer_1.close()
writer_2.close()
if __name__ == '__main__':
tb_test()
Here is the result:
The orange line shows the result of the evaluation stage, and correspondingly, the blue line illustrates the data of the training stage.
Also, there is a very useful post by TF team to which you can refer.
For completeness, since tensorboard 1.5.0 this is now possible.
You can use the custom scalars plugin. For this, you need to first make tensorboard layout configuration and write it to the event file. From the tensorboard example:
import tensorflow as tf
from tensorboard import summary
from tensorboard.plugins.custom_scalar import layout_pb2
# The layout has to be specified and written only once, not at every step
layout_summary = summary.custom_scalar_pb(layout_pb2.Layout(
category=[
layout_pb2.Category(
title='losses',
chart=[
layout_pb2.Chart(
title='losses',
multiline=layout_pb2.MultilineChartContent(
tag=[r'loss.*'],
)),
layout_pb2.Chart(
title='baz',
margin=layout_pb2.MarginChartContent(
series=[
layout_pb2.MarginChartContent.Series(
value='loss/baz/scalar_summary',
lower='baz_lower/baz/scalar_summary',
upper='baz_upper/baz/scalar_summary'),
],
)),
]),
layout_pb2.Category(
title='trig functions',
chart=[
layout_pb2.Chart(
title='wave trig functions',
multiline=layout_pb2.MultilineChartContent(
tag=[r'trigFunctions/cosine', r'trigFunctions/sine'],
)),
# The range of tangent is different. Let's give it its own chart.
layout_pb2.Chart(
title='tan',
multiline=layout_pb2.MultilineChartContent(
tag=[r'trigFunctions/tangent'],
)),
],
# This category we care less about. Let's make it initially closed.
closed=True),
]))
writer = tf.summary.FileWriter(".")
writer.add_summary(layout_summary)
# ...
# Add any summary data you want to the file
# ...
writer.close()
A Category is group of Charts. Each Chart corresponds to a single plot which displays several scalars together. The Chart can plot simple scalars (MultilineChartContent) or filled areas (MarginChartContent, e.g. when you want to plot the deviation of some value). The tag member of MultilineChartContent must be a list of regex-es which match the tags of the scalars that you want to group in the Chart. For more details check the proto definitions of the objects in https://github.com/tensorflow/tensorboard/blob/master/tensorboard/plugins/custom_scalar/layout.proto. Note that if you have several FileWriters writing to the same directory, you need to write the layout in only one of the files. Writing it to a separate file also works.
To view the data in TensorBoard, you need to open the Custom Scalars tab. Here is an example image of what to expect https://user-images.githubusercontent.com/4221553/32865784-840edf52-ca19-11e7-88bc-1806b1243e0d.png
The solution in PyTorch 1.5 with the approach of two writers:
import os
from torch.utils.tensorboard import SummaryWriter
LOG_DIR = "experiment_dir"
train_writer = SummaryWriter(os.path.join(LOG_DIR, "train"))
val_writer = SummaryWriter(os.path.join(LOG_DIR, "val"))
# while in the training loop
for k, v in train_losses.items()
train_writer.add_scalar(k, v, global_step)
# in the validation loop
for k, v in val_losses.items()
val_writer.add_scalar(k, v, global_step)
# at the end
train_writer.close()
val_writer.close()
Keys in the train_losses dict have to match those in the val_losses to be grouped on the same graph.
Tensorboard is really nice tool but by its declarative nature can make it difficult to get it to do exactly what you want.
I recommend you checkout Losswise (https://losswise.com) for plotting and keeping track of loss functions as an alternative to Tensorboard. With Losswise you specify exactly what should be graphed together:
import losswise
losswise.set_api_key("project api key")
session = losswise.Session(tag='my_special_lstm', max_iter=10)
loss_graph = session.graph('loss', kind='min')
# train an iteration of your model...
loss_graph.append(x, {'train_loss': train_loss, 'validation_loss': validation_loss})
# keep training model...
session.done()
And then you get something that looks like:
Notice how the data is fed to a particular graph explicitly via the loss_graph.append call, the data for which then appears in your project's dashboard.
In addition, for the above example Losswise would automatically generate a table with columns for min(training_loss) and min(validation_loss) so you can easily compare summary statistics across your experiments. Very useful for comparing results across a large number of experiments.
Please let me contribute with some code sample in the answer given by #Lifu Huang. First download the loger.py from here and then:
from logger import Logger
def train_model(parameters...):
N_EPOCHS = 15
# Set the logger
train_logger = Logger('./summaries/train_logs')
test_logger = Logger('./summaries/test_logs')
for epoch in range(N_EPOCHS):
# Code to get train_loss and test_loss
# ============ TensorBoard logging ============#
# Log the scalar values
train_info = {
'loss': train_loss,
}
test_info = {
'loss': test_loss,
}
for tag, value in train_info.items():
train_logger.scalar_summary(tag, value, step=epoch)
for tag, value in test_info.items():
test_logger.scalar_summary(tag, value, step=epoch)
Finally you run tensorboard --logdir=summaries/ --port=6006and you get:
I have a graph as follows, where the input x has two paths to reach y. They are combined with a gModule that uses cMulTable. Now if I do gModule:backward(x,y), I get a table of two values. Do they correspond to the error derivative derived from the two paths?
But since path2 contains other nn layers, I suppose I need to derive the derivates in this path in a stepwise fashion. But why did I get a table of two values for dy/dx?
To make things clearer, code to test this is as follows:
input1 = nn.Identity()()
input2 = nn.Identity()()
score = nn.CAddTable()({nn.Linear(3, 5)(input1),nn.Linear(3, 5)(input2)})
g = nn.gModule({input1, input2}, {score}) #gModule
mlp = nn.Linear(3,3) #path2 layer
x = torch.rand(3,3)
x_p = mlp:forward(x)
result = g:forward({x,x_p})
error = torch.rand(result:size())
gradient1 = g:backward(x, error) #this is a table of 2 tensors
gradient2 = g:backward(x_p, error) #this is also a table of 2 tensors
So what is wrong with my steps?
P.S, perhaps I have found out the reason because g:backward({x,x_p}, error) results in the same table. So I guess the two values stand for dy/dx and dy/dx_p respectively.
I think you simply made a mistake constructing your gModule. gradInput of every nn.Module has to have exactly the same structure as its input - that is the way backprop works.
Here's an example how to create a module like yours using nngraph:
require 'torch'
require 'nn'
require 'nngraph'
function CreateModule(input_size)
local input = nn.Identity()() -- network input
local nn_module_1 = nn.Linear(input_size, 100)(input)
local nn_module_2 = nn.Linear(100, input_size)(nn_module_1)
local output = nn.CMulTable()({input, nn_module_2})
-- pack a graph into a convenient module with standard API (:forward(), :backward())
return nn.gModule({input}, {output})
end
input = torch.rand(30)
my_module = CreateModule(input:size(1))
output = my_module:forward(input)
criterion_err = torch.rand(output:size())
gradInput = my_module:backward(input, criterion_err)
print(gradInput)
UPDATE
As I said, gradInput of every nn.Module has to have exactly the same structure as its input. So, if you define your module as nn.gModule({input1, input2}, {score}), your gradOutput (the result of the backward pass) will be a table of gradients w.r.t. input1 and input2 which in your case are x and x_p.
The only question remains: why on Earth don't you get an error when call:
gradient1 = g:backward(x, error)
gradient2 = g:backward(x_p, error)
An exception must be raised because the first argument must be not a tensor but a table of two tensors. Well, most (perhaps all) of torch modules during calculating :backward(input, gradOutput) don't use input argument (they usually store a copy of input from the last :forward(input) call). In fact, this argument is so useless that modules don't even bother themselves to verify it.