Difference between revisions of "Minh Le (Work Log)"
Leminh.ams (talk | contribs) |
Leminh.ams (talk | contribs) |
||
Line 12: | Line 12: | ||
dropout_rate_firstlayer\tdropout_rate_secondlayer\trec_dropout_rate_firstlayer\trec_dropout_rate_secondlayer\tembedding_vector_length\tfirstlayer_units\tsecondlayer_units\t"dropout_rate_dropout_layer\tepochs\tbatch_size\tvalidation_split | dropout_rate_firstlayer\tdropout_rate_secondlayer\trec_dropout_rate_firstlayer\trec_dropout_rate_secondlayer\tembedding_vector_length\tfirstlayer_units\tsecondlayer_units\t"dropout_rate_dropout_layer\tepochs\tbatch_size\tvalidation_split | ||
*Talked to Ed about potentially just do a test run with the RandomForest model because we needed data soon. | *Talked to Ed about potentially just do a test run with the RandomForest model because we needed data soon. | ||
+ | *This tuning had been fruitless. I stopped the code. | ||
+ | *Looked into a word2vec representation of the input | ||
2018-07-19: | 2018-07-19: |
Revision as of 12:04, 23 July 2018
Summer 2018
2018-07-23:
- The tuning has not been completed yet. However, checking from the results, it seemed that the last 6 parameters did not significantly affect the result?
2018-07-20:
- Ran parameters tuning to tweak 11 different parameters:
dropout_rate_firstlayer\tdropout_rate_secondlayer\trec_dropout_rate_firstlayer\trec_dropout_rate_secondlayer\tembedding_vector_length\tfirstlayer_units\tsecondlayer_units\t"dropout_rate_dropout_layer\tepochs\tbatch_size\tvalidation_split
- Talked to Ed about potentially just do a test run with the RandomForest model because we needed data soon.
- This tuning had been fruitless. I stopped the code.
- Looked into a word2vec representation of the input
2018-07-19:
- Helped Grace with her Thicket project
- Helped Maxine with her classifier
- Delegated the data collecting task to Connor
- Continued optimizing the current Kera's LSTM. The accuracy is around 50% right now
2018-07-18:
- Edited the wiki page with more content and ideas.
- Tried an MLP with lbfgs solver, and got around 60% accuracy:
FINISHED classifying. Train accuracy score: 1.0 FINISHED classifying. Test accuracy score: 0.652542372881356
- Building a full fledge LSTM (not prototype) to see how things go
2018-07-17:
- try tuning the LSTM in keras but did not manage to increase the accuracy by much. Accuracy fluctuates around 50%
2018-07-16:
- Work to adapt the data to RNN
- Installed keras for BOTH python 2 and 3.
- For python2, installed using the command:
pip install keras
- For python3, installed by first downloading github repo:
git clone https://github.com/keras-team/keras.git
then run the following command
cd keras python3 setup.py install
Normally, having run the command for python 2 should be sufficient, but we have anaconda2 and anaconda3 both so for some reason, pip can't detect the ananconda 3 folder, hence we have to manually install it like that. Note that you can run:
python setup.py install
to install to python2 as well (and skip the pip installation). Source: https://keras.io/
- Prototyped a simple LSTM in keras, and the accuracy was 0.53. This is promising; after I complete the full model, the accuracy can be much higher.
2018-07-13:
- Finished installing tensorflow for all user. Create a new folder to work on the DBServer to use tensorflow. The folder can be found here:
Z:\AcceleratorDemoDay
or if accessed from PuTtY, use the following command:
cd \bulk\AcceleratorDemoDay
- The new RNN currently has words frequency as input features
2018-07-12:
- Followed this instruction here: https://www.tensorflow.org/install/install_linux#InstallingVirtualenv and install tensorflow with Wei. Specific is below.
- 1. Installed CUDA Toolkit 9.0 Base Installer. The toolkit is in
/usr/local/cuda-9.0
for the toolkit. Did NOT install NVDIA accelerated Graphics Driver for Linux-x86_64 384.81 (We believe we have a different graphic driver. we have a much Newer version(396.26)). Installed the CUDA 9.0 samples in
HOME/MCNAIR/CUDA-SAMPLES.
- 2. Installed Patch 1, 2 and 3. The command to install was
sudo sh cuda 9.0.176.2 linux.run # (9.0.176.1 for patch 1 and 9.0.176.3 for patch 3)
- 3. This was supposed to be what to do next:
""" Set up the environment variables: The PATH variable needs to include /usr/local/cuda-9.0/bin To add this path to the PATH variable:
$ export PATH=/usr/local/cuda-9.0/bin${PATH:+:${PATH}}
In addition, when using the runfile installation method, the LD_LIBRARY_PATH variable needs to contain /usr/local/cuda-9.0/lib64 on a 64-bit system To change the environment variables for 64-bit operating systems:
$ export LD_LIBRARY_PATH=/usr/local/cuda-9.2/lib64\${LD_LIBRARY_PATH:+:${LD_LIBRARY_PATH}}
Note that the above paths change when using a custom install path with the runfile installation method. """ But when we travel to /usr/local/ we saw cuda-9.2 which we did not install. So we are WAITING for Yang to get back to use so we can proceed.
- For now, I can't build anything without tensorflow, so I am going to continue classifying data.
- Helped Grace with Google Scholar Crawler's regex
- All installationote can be see here Installing TensorFlow
2018-07-11:
- With an extended dataset, the accuracy went down with the random forest model. Accuracy: 0.71 (+/- 0.15)
- Built codes for an RNN, running into problem of not having tensorflow installed
- Helped Grace with her Google Scholar Crawler.
- Asked Wei to help with installing tensorflow GPU version.
2018-07-10:
- Doing further research into how RNN can be used to classify
- Reorganize the code under a new folder "Experiment" to prepare for testing with a new RNN
- Ran the reorganized code to make sure there is no problem. I kept running into this error: "TypeError: ufunc 'isfinite' not supported for the input types, and the inputs could not be safely coerced to any supported types according to the casting rule safe"
- Apparently this was caused by random question marks I have in the column (??) Removed it and it seems to run fine.
2018-07-09:
- Continued studying machine learning models.
- Helped Grace with her LinkedIn Crawler.
- Cleaned up working folder.
- Populate the project page with some information.
2018-07-06:
- Review Augi's classified training data to make sure it meets the correct requirement.
- Continued studying machine learning models and neural nets
2018-07-05:
- Studied different machine learning models and different classifier algorithms to prepare to build the RNN.
- Worked on classifying more training data.
2018-07-03:
- Ran a 0.84 classifier on the newly crawled data from the Chrome driver. From observation, the data still was not good enough. I will started building the RNN
- Still waiting Augi to release lock on the new excel data so i can work on it.
2018-07-02:
- Why did the code not run while I logged out of RDP omg these codes were running for so 3 hours last time I logged off :(
- The accuracy got to 0.875 today with just the new improved word list, which I thought might have overfitted the data. This was also rare because I never got it again
- Ran the improved crawler again to see how it went. (The ran start at 10AM ~It has been 5 hours-ish and it only processes the 50% of the list)
- After painfully seeing firefox crawling (literally) through webpages, I had installed the chromedriver in the working folder and changed the DemoDayCrawler.py back to Chrome Webdriver
- It seems like Firefox has a tendancy to pause randomly when i don't log into rdp and keep an eye on it. Chrome resolves this problem
2018-06-29:
- Delegated Augi to work on building the training data.
- Started to work on the classifier by studying machine learning models
- Edited words.txt with new words and remove words that i don't think help with the classification. Removed: march/ Added: rundown, list, mentors, overview, graduating, company, founders, autumn.
- The new words.txt had increased the accuracy from 0.76 to 0.83 in the first run
- The accuracy really fluctuated. Got as low as 0.74 but the highest run has been 0.866
- Note: testing inside of KyranGoogleClassifier instead of the main folder because the main folder was testing out the new improved crawler.
- It also seemed that rundown and autumn is the least important with 0.0 score so I removed them
2018-06-28:
- Continued to find more ways to optimize the crawler: adding several constraints as well as blacklist websites like Eventbrite, LinkedIn and Twitter. Needed to figure out a way to bypass Eventbrite's time expire script. LinkedIn required login before seeing details. Twitter's post was too short and frankly distracting.
- Ran improved results on the classifier.
- Classified some training data.
- Helped Grace debug the LinkedIn Crawler.
2018-06-27:
- Worked on optimizing and fixing issues with the crawler.
- It was observed that we may not need to change our criteria for the demo day pages. The page containing cohort list often includes dates (which is a data we now need to find). I might add more words to the words bag to improve it further but it seems unnecessary for now
2018-06-26:
- Finished running the Analysis code (for some reasons the shell didn't run after i logged off of RDP
- Talked to Ed about where to head with the code
- Connected the 2 projects together: got rid of Kyran's crawler and Peter's analysis script for now (we might want the analysis code later on to see how good the crawler was)
- Ran on the list of accelerators Connor gave me. Got mixed results (probably because the 80% is low) and we had to deal with website with expire timestamp like Eventbrite (the html showed the list, but displaying the html in the web browser doesn't). Found a problem that the crawler only get the number of results of the first page so if we want to gather large numbers of result, it would not work.
2018-06-25:
- Fixed Peter's Parser's compatibility issue with Python3. All code can now be used with Python 3
- Ran through everything in the Parser on a small test set.
- Completed moving all the files.
- Ran the Parser on the entire list.
- The run took 3h45m to execute the crawling (not counting the other steps) with 5 results per accelerators
- Update @6:00PM The Analysis has been taking an hour and 30m to run and only 80% done. I need to go home now but these steps are taking a lot of time
2018-06-22:
- Moved Peter's Parser into my project folder. Details can be read under the folder "E:\McNair\Projects\Accelerator Demo Day\Notes. READ THIS FIRST\movelog".
- The current Selenium version and Chrome seem to hate each other on the RDP (throwing a bunch of errors on registry key), so I had to switch to a Firefox webdriver. Adjusting code and inserting a bunch of sleep statements.
- For some reason (yet to be understood) if I save HTML pages with the utf-8 encoding, it will get mad at me. So commented that out for now.
- The code seemed slow compared to those existed in Kyran's project. Might attempt to optimize and parallelize it?
- it seems that python 3 does not support write(stuff).encoding('utf-8')?
2018-06-21:
- Continued reading through past projects (it's so disorganized...)
- Moved Kyran's Google Classifier to my project folder. Details can be read under the folder "Notes. READ THIS FIRST\movelog".
- Tried running the Classifier from a new folder. The Shell crashed once on the web_demo_feature.py
- Ran through everything in the Classfier. Things seemed to be functioning with occasional error messages
- Talked to Kyran about the project and clarified some confusions up
- Made a to-do list in the general note file ("Notes. READ THIS FIRST\NotesAndTasks.txt")
2018-06-20:
- Set up Work Log page.
- Edited Profile page with more information.
- Created project page: Accelerator Demo Day.
- Made new project folder at E:\McNair\Projects\Accelerator Demo Day.
- Read through old projects and started copying scripts over as well as cleaned things up.
- Created movelog.txt to track these moving details.
- Talked to Ed more about the project goals and purposes
2018-06-19: More SQL. Talked to Ed and received my project (Demo Day Crawler).
2018-06-18: Set up RDPs, Slacks, Profile page and learned about SQL.