systemml-dev mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
Subject Re: Please reply ASAP : Regarding incubator systemml/breast_cancer project
Date Tue, 25 Apr 2017 18:27:37 GMT
Hi Aishwarya,

Unfortunately this mailing list removes all images, so I can't view your screenshot.  I'm
assuming that it is the same issue with the missing SparkContext `sc` object, but please let
me know if it is a different issue.  This sounds like it could be an issue with multiple kernels
installed in Jupyter.  When you start the notebook, can you see if there are multiple kernels
listed in the "Kernel" -> "Change Kernel" menu?  If so, please try one of the other kernels
to see if Jupyter is starting by default with a non-spark kernel.  Also, is it possible that
you have more than one instance of the Jupyter server running?  I.e. for this scenario, we
start Jupyter itself directly via pyspark using the command sent previously, whereas usually
Jupyter can just be started with `jupyter notebook`.  In the latter case, PySpark (and thus
`sc`) would *not* be available (unless you've set up special PySpark kernels separately).
 In summary, can you (1) check for other kernels via the menus, and (2) check for other running
Jupyter servers that are non-PySpark?

As for the other inquiry, great question!  When training models, it's quite useful to track
the loss and other metrics (i.e. accuracy) from *both* the training and validation sets. 
The reasoning is that it allows for a more holistic view of the overall learning process,
such as evaluating whether any overfitting or underfitting is occurring.  For example, say
that you train a model and achieve an accuracy of 80% on the validation set.  Is this good?
 Is this the best that can be done?  Without also tracking performance on the training set,
it can be difficult to make these decisions.  Say that you then measure the performance on
the training set and find that the model achieves 100% accuracy on that data.  That might
be a good indication that your model is overfitting the training set, and that a combination
of more data, regularization, and a smaller model may be helpful in raising the generalization
performance, i.e. the performance on the validation set and future real examples on which
you wish to make predictions.  If on the other hand, the model achieved an 82% on the training
set, this could be a good indication that the model is underfitting, and that a combination
of a more expressive model and better data could be helpful.  In summary, tracking performance
on both the training and validation datasets can be useful for determining ways in which to
improve the overall learning process.

- Mike


Mike Dusenberry

Sent from my iPhone.

> On Apr 25, 2017, at 8:47 AM, Aishwarya Chaurasia <> wrote:
> We had another query, sir. We read the entire MachineLearning.ipynb code.
> in it the training samples and the validation samples have both been
> evaluated separately and their respective losses and accuracies obtained.
> Why are the training samples being evaluated again if they were used to
> train the model in the first place? Shouldn't only the validation data
> frames be evaluated to find out the loss and accuracy?
> Thank you
> On 25-Apr-2017 4:00 PM, "Aishwarya Chaurasia" <>
> wrote:
>> Hello sir,
>> The NameError is occuring again sir. Why does it keep resurfacing?
>> Attaching the screenshot of the error.
>>> On 25-Apr-2017 2:50 AM, <> wrote:
>>> Hi Aishwarya,
>>> For the error message, that just means that the SystemML jar isn't being
>>> found.  Can you add a `--driver-class-path $SYSTEMML_HOME/target/SystemML.jar`
>>> to the invocation of Jupyter?  I.e. `PYSPARK_PYTHON=python3
>>> pyspark  --jars $SYSTEMML_HOME/target/SystemML.jar --driver-class-path
>>> $SYSTEMML_HOME/target/SystemML.jar`. There was a PySpark bug that was
>>> supposed to have been fixed in Spark 2.x, but it's possible that it is
>>> still an issue.
>>> As for the output, the notebook will create SystemML `Matrix` objects for
>>> all of the weights and biases of the trained models.  To save, please
>>> convert each one to a DataFrame, i.e. `Wc1.toDF()` and repeated for each
>>> matrix, and then simply save the DataFrames.  This could be done all at
>>> once like this for a SystemML Matrix object `Wc1`:
>>> `Wc1.toDf()"path/to/save/Wc1.parquet", format="parquet")`.
>>> Just repeat for each matrix returned by the "Train" code for the
>>> algorithms.  At that point, you will have a set of saved DataFrames
>>> representing a trained SystemML model, and these can be used in downstream
>>> classification tasks in a similar manner to the "Eval" sections.
>>> -Mike
>>> --
>>> Mike Dusenberry
>>> GitHub:
>>> LinkedIn:
>>> Sent from my iPhone.
>>>> On Apr 24, 2017, at 3:07 AM, Aishwarya Chaurasia <
>>>> wrote:
>>>> Further more :
>>>> What is the output of MachineLearning.ipynb you're obtaining sir?
>>>> We are actually nearing our deadline for our problem.
>>>> Thanks a lot.
>>>> On 24-Apr-2017 2:58 PM, "Aishwarya Chaurasia" <>
>>>> wrote:
>>>> Hello sir,
>>>> Thanks a lot for replying sir. But unfortunately it did not work.
>>> Although
>>>> the NameError did not appear this time but another error came about :
>>>> 5M1UNdIGYhyRLivL9gydE=
>>>> This error was obtained after executing the second block of code of
>>>> in terminal. ( ml = MLContext(sc) )
>>>> We have installed the bleeding-edge version of systemml only and the
>>>> installation was done correctly. We are in a fix now. :/
>>>> Kindly look into the matter asap
>>>> On 24-Apr-2017 12:15 PM, "Mike Dusenberry" <>
>>> wrote:
>>>> Hi Aishwarya,
>>>> Glad to hear that the preprocessing stage was successful!  As for the
>>>> `MachineLearning.ipynb` notebook, here is a general guide:
>>>>  - The `MachineLearning.ipynb` notebook essentially (1) loads in the
>>>>  training and validation DataFrames from the preprocessing step, (2)
>>>>  converts them to normalized & one-hot encoded SystemML matrices for
>>>>  consumption by the ML algorithms, and (3) explores training a couple
>>> of
>>>>  models.
>>>>  - To run, you'll need to start Jupyter in the context of PySpark via
>>>>  PYSPARK_DRIVER_PYTHON_OPTS="notebook" pyspark  --jars
>>>>  $SYSTEMML_HOME/target/SystemML.jar`.  Note that if you have installed
>>>>  SystemML with pip from PyPy (`pip3 install systemml`), this will
>>> install
>>>>  our 0.13 release, and the `--jars $SYSTEMML_HOME/target/SystemML.jar`
>>>> will
>>>>  not be necessary.  If you instead have installed a bleeding-edge
>>> version
>>>> of
>>>>  SystemML locally (git clone locally, maven build, `pip3 install -e
>>>>  src/main/python` as listed in `projects/breast_cancer/`),
>>> the
>>>>  `--jars $SYSTEMML_HOME/target/SystemML.jar` part *is* necessary.  We
>>> are
>>>>  about to release 0.14, and for this project, I *would* recommend
>>> using a
>>>>  bleeding edge install.
>>>>  - Once Jupyter has been started in the context of PySpark, the `sc`
>>>>  SparkContext object should be available.  Please let me know if you
>>>>  continue to see this issue.
>>>>  - The "Read in train & val data" section simply reads in the training
>>>>  and validation data generated in the preprocessing stage.  Be sure
>>> that
>>>> the
>>>>  `size` setting is the same as the preprocessing size.  The percentage
>>> `p`
>>>>  setting determines whether the full or sampled DataFrames are
>>> loaded.  If
>>>>  you set `p = 1`, the full DataFrames will be used.  If you instead
>>> would
>>>>  prefer to use the smaller sampled DataFrames while getting started,
>>>> please
>>>>  set it to the same value as used in the preprocessing to generate the
>>>>  smaller sampled DataFrames.
>>>>  - The `Extract X & Y matrices` section splits each of the train and
>>>>  validation DataFrames into effectively X & Y matrices (still as
>>> DataFrame
>>>>  types), with X containing the images, and Y containing the labels.
>>>>  - The `Convert to SystemML Matrices` section passes the X & Y
>>> DataFrames
>>>>  into a SystemML script that performs some normalization of the images
>>> &
>>>>  one-hot encoding of the labels, and then returns SystemML `Matrix`
>>> types.
>>>>  These are now ready to be passed into the subsequent algorithms.
>>>>  - The "Trigger Caching" and "Save Matrices" are experimental features,
>>>>  and not necessary to execute.
>>>>  - Next comes the two algorithms being explored in this notebook.  The
>>>>  "Softmax Classifier" is just a multi-class logistic regression model,
>>> and
>>>>  is simply there to serve as a baseline comparison with the subsequent
>>>>  convolutional neural net model.  You may wish to simply skip this
>>> softmax
>>>>  model and move to the latter convnet model further down in the
>>> notebook.
>>>>  - The actual softmax model is located at [
>>>> projects/breast_cancer/softmax_clf.dml],
>>>>  and the notebook calls functions from that file.
>>>>  - The softmax sanity check just ensures that the model is able to
>>>>  completely overfit when given a tiny sample size.  This should yield
>>>> ~100%
>>>>  training accuracy if the sample size in this section is small enough.
>>>> This
>>>>  is just a check to ensure that nothing else is wrong with the math or
>>> the
>>>>  data.
>>>>  - The softmax "Train" section will train a softmax model and return
>>> the
>>>>  weights (`W`) and biases (`b`) of the model as SystemML `Matrix`
>>> objects.
>>>>  Please adjust the hyperparameters in this section to your problem.
>>>>  - The softmax "Eval" section takes the trained weights and biases and
>>>>  evaluates the training and validation performance.
>>>>  - The next model is a LeNet-like convnet model.  The actual model is
>>>>  located at [
>>>> projects/breast_cancer/convnet.dml],
>>>>  and the notebook simply calls functions from that file.
>>>>  - Once again, there is an initial sanity check for the ability to
>>>>  overfit on a small amount of data.
>>>>  - The "Hyperparameter Search" contains a script to sample different
>>>>  hyperparams for the convnet, and save the hyperparams + validation
>>>> accuracy
>>>>  of each set after a single epoch of training.  These string files
>>> will be
>>>>  saved to HDFS.  Please feel free to adjust the range of the
>>>> hyperparameters
>>>>  for your problem.  Please also feel free to try using the `parfor`
>>>>  (parallel for-loop) instead of the while loop to speed up this
>>> section.
>>>>  Note that this is still a work in progress.  The hyperparameter
>>> tuning in
>>>>  this section makes use of random search (as opposed to grid search),
>>>> which
>>>>  has been promoted by Bengio et al. to speed up the search time.
>>>>  - The "Train" section trains the convnet and returns the weights and
>>>>  biases as SystemML `Matrix` types.  In this section, please replace
>>> the
>>>>  hyperparameters with the best ones from above, and please increase the
>>>>  number of epochs given your time constraints.
>>>>  - The "Eval" section evaluates the performance of the trained convnet.
>>>>  - Although it is not shown in the notebook yet, to save the weights
>>> and
>>>>  biases, please use the `toDF()` method on each weight and biases (i.e.
>>>>  `Wc1.toDF()`) to convert to a Spark DataFrame, and then simply save
>>> the
>>>>  DataFrame as desired.
>>>>  - Finally, please feel free to extend the model in `convnet.dml` for
>>>>  your particular problem!  The LeNet-like model just serves as a simple
>>>>  convnet, but there are much richer models currently, such as resnets,
>>>> that
>>>>  we are experimenting with.  To make larger models such as resnets
>>> easier
>>>> to
>>>>  define, we are also working on other tools for converting model
>>>> definitions
>>>>  + pretrained weights from other systems into SystemML.
>>>> Also, please keep in mind that the deep learning support in SystemML is
>>>> still a work in progress.  Therefore, if you run into issues, please
>>> let us
>>>> know and we'll do everything possible to help get things running!
>>>> Thanks!
>>>> - Mike
>>>> --
>>>> Michael W. Dusenberry
>>>> GitHub:
>>>> LinkedIn:
>>>> On Sat, Apr 22, 2017 at 4:49 AM, Aishwarya Chaurasia <
>>>>> wrote:
>>>>> Hey,
>>>>> Thank you so much for your help sir. We were finally able to run
>>>>> without any errors. And the results obtained were
>>>>> satisfactory i.e we got five set of data frames like you said we would.
>>>>> But alas! when we tried to run MachineLearning.ipynb the same NameError
>>>>> came :
>>>>> vnYEDTSTQH73l5M1UNdIGYhyRLivL9gydE=
>>>>> Could you guide us again as to how to proceed now?
>>>>> Also, could you please provide an overview of the process
>>>>> MachineLearning.ipynb is following to train the samples.
>>>>> Thanks a lot!
>>>>>> On 20-Apr-2017 12:16 AM, <> wrote:
>>>>>> Hi Aishwarya,
>>>>>> Looks like you've just encountered an out of memory error on one
>>> the
>>>>>> executors.  Therefore, you just need to adjust the
>>>>> `spark.executor.memory`
>>>>>> and `spark.driver.memory` settings with higher amounts of RAM.  What
>>> is
>>>>>> your current setup?  I.e. are you using a cluster of machines, or
>>>>> single
>>>>>> machine?  We generally use a large driver on one machine, and then
>>>>> single
>>>>>> large executor on each other machine.  I would give a sizable amount
>>> of
>>>>>> memory to the driver, and about half the possible memory on the
>>>> executors
>>>>>> so that the Python processes have enough memory as well.  PySpark
>>>> JVM
>>>>>> and Python components, and the Spark memory settings only pertain
>>> the
>>>>>> JVM side, thus the need to save about half the executor memory for
>>>>>> Python side.
>>>>>> Thanks!
>>>>>> - Mike
>>>>>> --
>>>>>> Mike Dusenberry
>>>>>> GitHub:
>>>>>> LinkedIn:
>>>>>> Sent from my iPhone.
>>>>>>> On Apr 19, 2017, at 5:53 AM, Aishwarya Chaurasia <
>>>>>>> wrote:
>>>>>>> Hello sir,
>>>>>>> We also wanted to ensure that the spark-submit command we're
using is
>>>>> the
>>>>>>> correct one for running ''.
>>>>>>> Command :  /home/new/sparks/bin/spark-submit
>>>>>>> Thank you.
>>>>>>> Aishwarya Chaurasia.
>>>>>>> On 19-Apr-2017 3:55 PM, "Aishwarya Chaurasia" <
>>>>>>> wrote:
>>>>>>> Hello sir,
>>>>>>> On running the file we are getting the following
>>> :
>>>>>>> YhyRLivL9gydE=
>>>>>>> Can you please help us by looking into the error and kindly tell
>>>> the
>>>>>>> solution for it.
>>>>>>> Thanks a lot.
>>>>>>> Aishwarya Chaurasia
>>>>>>>> On 19-Apr-2017 12:43 AM, <> wrote:
>>>>>>>> Hi Aishwarya,
>>>>>>>> Certainly, here is some more detailed information
>>>>> about``:
>>>>>>>> * The preprocessing Python script is located at
>>>>>>>> projects/breast_cancer/ Note that this is different
>>>>> than
>>>>>>>> the library module at
>>>>>>>> bator-systemml/blob/master/projects/breast_cancer/breastc
>>>>>>>> ancer/
>>>>>>>> * This script is used to preprocess a set of histology slide
>>>>>>>> which are `.svs` files in our case, and `.tiff` files in
your case.
>>>>>>>> * Lines 63-79 contain "settings" such as the output image
>>>>> folder
>>>>>>>> paths, etc.  Of particular interest, line 72 has the folder
path for
>>>>> the
>>>>>>>> original slide images that should be commonly accessible
from all
>>>>>> machines
>>>>>>>> being used, and lines 74-79 contain the names of the output
>>>> DataFrames
>>>>>> that
>>>>>>>> will be saved.
>>>>>>>> * Line 82 performs the actual preprocessing and creates a
>>>>>>>> DataFrame with the following columns: slide number, tumor
>>>>>> molecular
>>>>>>>> score, sample.  The "sample" in this case is the actual small,
>>>>>> chopped-up
>>>>>>>> section of the image that has been extracted and flattened
into a
>>> row
>>>>>>>> Vector.  For test images without labels (`training=false`),
only the
>>>>>> slide
>>>>>>>> number and sample will be contained in the DataFrame (i.e.
>>>> labels).
>>>>>>>> This calls the `preprocess(...)` function located on line
371 of
>>>>>>>> projects/breast_cancer/breastcancer/, which
is a
>>>>>>>> different file.
>>>>>>>> * Line 87 simply saves the above DataFrame to HDFS with the
>>>> from
>>>>>>>> line 74.
>>>>>>>> * Line 93 splits the above DataFrame row-wise into separate
>>>>> "training"
>>>>>>>> and "validation" DataFrames, based on the split percentage
from line
>>>>> 70
>>>>>>>> (`train_frac`).  This is performed so that downstream machine
>>>> learning
>>>>>>>> tasks can learn from the training set, and validate performance
>>>>>>>> hyperparameter choices on the validation set.  These DataFrames
>>>>>> start
>>>>>>>> with the same columns as the above DataFrame.  If `add_row_indices`
>>>>> from
>>>>>>>> line 69 is true, then an additional row index column (`__INDEX`)
>>> will
>>>>> be
>>>>>>>> pretended.  This is useful for SystemML in downstream machine
>>>> learning
>>>>>>>> tasks as it gives the DataFrame row numbers like a real matrix
>>>>>> have,
>>>>>>>> and SystemML is built to operate on matrices.
>>>>>>>> * Lines 97 & 98 simply save the training and validation
>>>>>> using
>>>>>>>> the names defined on lines 76 & 78.
>>>>>>>> * Lines 103-137 create smaller train and validation DataFrames
>>>>>> taking
>>>>>>>> small row-wise samples of the full train and validation DataFrames.
>>>>> The
>>>>>>>> percentage of the sample is defined on line 111 (`p=0.01`
for a 1%
>>>>>>>> sample).  This is generally useful for quicker downstream
>>>>> without
>>>>>>>> having to load in the larger DataFrames, assuming you have
a large
>>>>>> amount
>>>>>>>> of data.  For us, we have ~7TB of data, so having 1% sampled
>>>>> DataFrames
>>>>>> is
>>>>>>>> useful for quicker downstream tests.  Once again, the same
>>>>> from
>>>>>> the
>>>>>>>> larger train and validation DataFrames will be used.
>>>>>>>> * Lines 146 & 147 simply save these sampled train and
>>>>>>>> DataFrames.
>>>>>>>> As a summary, after running ``, you will be
left with
>>>> the
>>>>>>>> following saved DataFrames in HDFS:
>>>>>>>> * Full DataFrame
>>>>>>>> * Training DataFrame
>>>>>>>> * Validation DataFrame
>>>>>>>> * Sampled training DataFrame
>>>>>>>> * Sampled validation DataFrame
>>>>>>>> As for visualization, you may visualize a "sample" (i.e.
>>>>>> chopped-up
>>>>>>>> section of original image) from a DataFrame by using the
>>>>>>>> breastcancer.visualization.visualize_sample(...)` function.
>>> will
>>>>>>>> need to do this after creating the DataFrames.  Here is a
snippet to
>>>>>>>> visualize the first row sample in a DataFrame, where `df`
is one of
>>>>> the
>>>>>>>> DataFrames from above:
>>>>>>>> ```
>>>>>>>> from breastcancer.visualization import visualize_sample
>>>>>>>> visualize_sample(df.first().sample)
>>>>>>>> ```
>>>>>>>> Please let me know if you have any additional questions.
>>>>>>>> Thanks!
>>>>>>>> - Mike
>>>>>>>> --
>>>>>>>> Mike Dusenberry
>>>>>>>> GitHub:
>>>>>>>> LinkedIn:
>>>>>>>> Sent from my iPhone.
>>>>>>>>> On Apr 15, 2017, at 4:38 AM, Aishwarya Chaurasia <
>>>>>>>>> wrote:
>>>>>>>>> Hello sir,
>>>>>>>>> Can you please elaborate more on what output we would
be getting
>>>>>> because
>>>>>>>> we
>>>>>>>>> tried executing the file using spark submit
it keeps
>>>> on
>>>>>>>>> adding the tiles in rdd and while running the
>>>>> it
>>>>>>>>> isn't showing any output. Can you please help us out
asap stating
>>>> the
>>>>>>>>> output we will be getting and the sequence of execution
of files.
>>>>>>>>> Thank you.
>>>>>>>>>> On 07-Apr-2017 5:54 AM, <>
>>>>>>>>>> Hi Aishwarya,
>>>>>>>>>> Thanks for sharing more info on the issue!
>>>>>>>>>> To facilitate easier usage, I've updated the preprocessing
code by
>>>>>>>> pulling
>>>>>>>>>> out most of the logic into a `breastcancer/`
>>>>> module,
>>>>>>>>>> leaving just the execution in the `Preprocessing.ipynb`
>>>>>>>> There is
>>>>>>>>>> also a `` script with the same contents
as the
>>>> notebook
>>>>>> for
>>>>>>>>>> use with `spark-submit`.  The choice of the notebook
or the script
>>>>> is
>>>>>>>> just
>>>>>>>>>> a matter of convenience, as they both import from
the same
>>>>>>>>>> `breastcancer/` package.
>>>>>>>>>> As part of the updates, I've added an explicit SparkSession
>>>>> parameter
>>>>>>>>>> (`spark`) to the `preprocess(...)` function, and
updated the body
>>>> to
>>>>>> use
>>>>>>>>>> this SparkSession object rather than the older SparkContext
>>>>>> object.
>>>>>>>>>> Previously, the `preprocess(...)` function accessed
the `sc`
>>> object
>>>>>> that
>>>>>>>>>> was pulled in from the enclosing scope, which would
work while all
>>>>> of
>>>>>>>> the
>>>>>>>>>> code was colocated within the notebook, but not if
the code was
>>>>>>>> extracted
>>>>>>>>>> and imported.  The explicit parameter now allows
for the code to
>>> be
>>>>>>>>>> imported.
>>>>>>>>>> Can you please try again with the latest updates?
 We are
>>> currently
>>>>>>>> using
>>>>>>>>>> Spark 2.x with Python 3.  If you use the notebook,
the pyspark
>>>>> kernel
>>>>>>>>>> should have a `spark` object available that can be
supplied to the
>>>>>>>>>> functions (as is done now in the notebook), and if
you use the
>>>>>>>>>> `` script with `spark-submit`, the `spark`
>>> will
>>>>> be
>>>>>>>>>> created explicitly by the script.
>>>>>>>>>> For a bit of context to others, Aishwarya initially
reached out to
>>>>>> find
>>>>>>>>>> out if our breast cancer project could be applied
to TIFF images,
>>>>>> rather
>>>>>>>>>> than the SVS images we are currently using (the answer
is "yes" so
>>>>>> long
>>>>>>>> as
>>>>>>>>>> they are "generic tiled TIFF images, according to
the OpenSlide
>>>>>>>>>> documentation), and then followed up with Spark issues
related to
>>>>> the
>>>>>>>>>> preprocessing code.  This conversation has been promptly
moved to
>>>>> the
>>>>>>>>>> mailing list so that others in the community can
>>>>>>>>>> Thanks!
>>>>>>>>>> -Mike
>>>>>>>>>> --
>>>>>>>>>> Mike Dusenberry
>>>>>>>>>> GitHub:
>>>>>>>>>> LinkedIn:
>>>>>>>>>> Sent from my iPhone.
>>>>>>>>>>> On Apr 6, 2017, at 5:09 AM, Aishwarya Chaurasia
>>>>>>>>>> wrote:
>>>>>>>>>>> Hey,
>>>>>>>>>>> The object sc is already defined in pyspark and
yet this name
>>>> error
>>>>>>>> keeps
>>>>>>>>>>> occurring. We are using spark 2.*
>>>>>>>>>>> Here is the link to error that we are getting
>>>>> 89iQODxzpNZVbSfgwocH8l5M1UNdIG
>>>>>>>>>> YhyRLivL9gydE=

  • Unnamed multipart/alternative (inline, 7-Bit, 0 bytes)
View raw message