Skip to content

Latest commit

 

History

History
372 lines (292 loc) · 11.5 KB

log.md

File metadata and controls

372 lines (292 loc) · 11.5 KB
Raw

Activity Log

Written by: Erick Cobos T. ([email protected])

Log with activities and questions arising during every week

Jun 2 - Jun 8

Activities

  • Write Experiment 3 in Ch 3
  • Write code for Experiment 3

To do

  • Run Experiment 3 Hyperparameter search
  • Run Experiment 3
  • Ask for institutional email

May 19 - May 25

Activities

  • Train best Experiment 2 model
  • Write results for Experiment 2 in Chapter 4

To do

  • Write code for model 3
  • Ask for institutional email

Questions

  1. Should I try using a Resnet-like model with dilated convolutions? Answer: Sure :)
  2. Should I try a simpler model with less parameters (maybe in exchange for more memory usage)?

May 12 - May 18

Activities

  • Write code to test models in test set
  • Run hyperparameter search for Experiment 2
  • Write hyperparameter search result for Exp 2

To do

  • Train best Experiment 2 model
  • Write results for Experiment 2 in Chapter 4
  • Ask for institutional email

May 5 - May 11

Activities

  • Write Chapter 3 in thesis
  • Write results from Experiment 1 in thesis.

To do

  • Change loss function for experiment 2
  • Run Experiment 2 hyperparameter search
  • Ask for institutional email

Apr 29 - May 4

Activities

  • Run refined hyperparameter search
  • Train best network

To do

  • Write hyperparameter results in thesis
  • Write results from the best network
  • Ask for institutional email

Apr 25 - Apr 28

Activities

  • Run hyperparameter search
  • Get hyperparameter figures

To do

  • Refine hyperparameter search
  • Write Model and Results
  • Train best network
  • Ask for institutional email

Apr 14 - Apr 22

Activities

  • Install CUDA and Tensorflow 0.8.0 in computers in A3-401
  • Modify model to run inside 2GB GPUs
  • Do the slides for the Research congress
  • Expose in research congress

To do

  • Run hyperparamter search
  • Write Section 3.2(Model) and 3.3(Training)
  • Ask for institutional email

Apr 07 - Apr 13

Activities

  • Install CUDA and Tensorflow in computers in A3-401
  • Modify model to run in TF 0.8.0

To do

  • Install Tensorflow in computers at A3-301
  • Run experiments
  • Write Section 3.2(Model) and 3.3(Training)
  • Ask for institutional email

Mar 31 - Apr 06

Activities

  • Writing Tensorflow implementation
  • Run sanity checks and initial tests

To do

  • Install Tensorflow in computers at A3-401
  • Run experiments
  • Write Section 3.2(Model) and 3.3(Training)
  • Ask for institutional email

Mar 17 - Mar 30

Activities

  • Spoke with Dr. Parra
  • Writing Tensorflow implementation

To do

  • Write network in Tensorflow
  • Install Tensorflow in computers at A3-401
  • Run experiments
  • Write Section 3.2(Model) and 3.3(Training)
  • Ask for institutional email

Mar 10 - Mar 16

Activities

  • Spoke with Dr. Garrido
  • Writing Tensorflow implementation

To do

  • Write network in Tensorflow
  • Ask for a computer in A3-401 (Dr. Parra) or the cluster of CPUs (Dr. Nolazco)
  • Write chapter 3 (Solution Model)
  • Ask for institutional email

Mar 04 - Mar 09

Activities

  • Reading Tensorflow tutorials

To do

  • Write network in Tensorflow
  • Ask for a computer in A3-401 (Dr. Parra) or the cluster of CPUs (Dr. Nolazco)
  • Write chapter 3 (Solution Model)
  • Ask for institutional email

Feb 25 - Mar 03

Activities

  • Preprocessing database
  • Writing Section 3.2
  • Writing Section 3.3
  • Watching cs231n lectures/ updating knowledge
  • Installing Tensorflow (Laptop/CTS)
  • Reading Tensorflow tutorials

To do

  • Read Tensorflow tutorials
  • Write network in Tensorflow
  • Ask for a computer in A3-401 or somewhere else (maybe ask Dr. Garrido)
  • Write chapter 3 (Solution Model)
  • Ask for institutional email

Questions

  1. Should I add code as appendix (prepareDB.py or tensorflow.py) or link to the github project or not do any? Answer: Put a link to Github

Feb 18 - Feb 24

Activities

  • Writing Section 2.6.2, adding to 2.5
  • Taking more desing decisions
  • Writing Section 3.2

To do

  • Ready database
  • Write chapter 3 (Solution Model)
  • Install TensorFlow (CTS/Laptop)
  • Write network in Tensorflow/Keras
  • Ask for a computer in A3-401 or somewhere else (maybe ask Dr. Garrido)
  • Ask for institutional email

Questions

  1. Should I leave on the database part in Breast cancer or better just delete it and degrade mammograms and maybe put BCDR info in Model/Data set/Database? Answer: Leave it there.
  2. Should I try to fit a simple model and a more advanced model (ADAM, batchnorm, leaky relus) or should I go directly for the best model? Answer: Best model.

Feb 12 - Feb 17

Activities

  • Writing section 2.6 and 2.7
  • Writing Section 3.1

To do

  • Write chapter 3 (Solution Model)
  • Ready database
  • Install TensorFlow (CTS/Laptop)
  • Write network in Tensorflow/Keras
  • Ask for a computer in A3-401 or somewhere else (maybe ask Dr. Garrido)
  • Ask for institutional email

Questions

  1. Is IOU fine for unbalanced data sets? Answer: For model selection IOU is going to try to maximize the intersection and minimize the union as the union is waaay bigger (because objects are small), it will probably try to minimize the union more prediciting less positive labels and it may lose sensitivity (for the sake of specificity). Not sure about this, though, seems like F-1 is gonna do the same.
  2. Should I leave Section 2.7 citations as "[23] trained ..." or write "Ge et al. trained"? Answer: Say names
  3. For background, should I cite all articles where a netwrok appear, if for example they reported something twice. Answer: No
  4. In the solution model, should I write all alternatives, say what I chose, and explain why I choose the one i chose or just say what I chose? Answer: Just what you chose and justify why. Maybe an alternative but only slightly.

Jan 25 - Feb 11

Activities

  • Writing final draft of Chapter 2
  • Taking last important design decisions

To do

  • Write chapter 3 (Solution Model)
  • Ready database
  • Install TensorFlow (CTS/Laptop)
  • Write network in Tensorflow/Keras
  • Ask for a computer in A3-401 or somewhere else (maybe ask Dr. Garrido)
  • Ask for institutional email

Questions

  1. Which post-processing should I use? Gaussian smoothing, cluster-based enhancement, fully connected CRFs or a combination?
  2. Which evaluation metric should I use? Accuracy, F1-score, PRAUC, ROC, IOU, Dice? Answer: IOU. F1-score in second place.
  3. Should I cite Agarwal2015 (unpublished Stanford report)? Answer: No.

Jul 2 - Jul 8

Activities

  • Write emails looking for more digital mammograms
  • Document decisions on how to crop big images.

To do

  • Ask for institutional email.

Jun 24 - Jul 1

Activities

  • Look for databases and its features
  • Decide how to crop the images from the entire mammograms

To do

  • Write the features of the database and its labelling.
  • Decide how to obtain the small crops from the big mammograms.
  • Write code to automatically obtain the small crops from the mammogram.
  • Update LaTex template to thesis template
  • Choose software (probably Caffe)

Questions

  1. Should I try to get more digital mammograms or just go with film?. Answer: Enough digital mammograms, work with wath you have. If needed, ask Dr. Tamez.

Jun 18 - Jun 23

Activities

  • Ended review
  • Write ConvNet for Breast Cancer
  • Rewrite some parts of proposal (introduction, objectives, methodology)

To do

  • Investigate and write the features of the database and its labelling.
  • Decide how to obtain the small crops from the big mammograms
  • Write software to automatically obtain the small crops from the mammogram
  • Update LaTex template to thesis template
  • Choose software (probably Caffe)

Questions

  1. Should i put mass vs nonmass, microcalc vs nonmicrocalc, or put every lession together (mass, microcalc, distortions, etc.) vs nonlession?. Thus, only train one network that differentiates all lessions vs no lession?. Answer: Segmentation (mass(benign or malign) vs non-mass)
  2. Should I use a single network with multiple outputs to classify every kind of lession?. Answer: No.
  3. Should I use data augmentation only on the minority classes (lessions)? Answer: Use everywhere. No oversampling

Jun 10 - Jun 17

Activities

  • Write ConvNet for Breast Cancer
  • Some preprocessing experiments

To do

  • Rewrite Methodology with new experiments
  • Update LaTex template to thesis template
  • Investigate and write the features of the database and its labelling.
  • Choose software (probably Caffe)

Questions

  1. Is the unbalanced data thing needed or does the network learns by its own?. May i be overkilling it? Answer: Train normally, cross-validate the threshold

Jun 3 - Jun 10

Activities

  • Write PracticalDL section

To do

  • Write ConvNet for Breast Cancer
  • Rewrite Methodology with new experiments
  • Update LaTex template to thesis template
  • Investigate and write the features of the database and its labelling.
  • Choose software (probably Caffe)

Questions

  1. Use NAG or SGD+Momentum? Answer: NAG
  2. Use Bioinformatics account or create another one?. Answer: Bioinformatics
  3. Is there a standard way to report convolutional network architectures (Krizhevsky style or Karpathy style or a table as in Striving for simplicity)?. Answer: No. Image if small, Table if big.

May 27- Jun 2

Activities

  • Read CS231n
  • Write Convnet section in thesis

To do

  • End writing Background
    • Practical Deep Learning
    • ConvNet for Breast Cancer
  • Rewrite Methodology with new experiments
  • Update LaTex template to thesis template
  • Investigate and write the features of the database and its labelling.
  • Choose software (probably Caffe)
  • Select exactly what experiments will be run and what hyperparameters be crossvalidated

Questions

  1. Naming: Should I use loss or cost function?. Answer: Loss
  2. How to obtain the small training images from the big images. Random sampling, crop without overlapping, with overlapping.? How to measure performance?. What are the labels?. Answer: Patches not needed. Labels are 1 in lesion, 0 in no lesion.

May 25 - May 26

Activities

  • Installed in CTS 5
  • Read cs231n Stanford Convnet course (cs231n.github.io).

To do

  • End writing Background
    • ConvNet
    • ConvNet for Breast Cancer
    • Practical Deep Learning
    • Database specifics
  • Update LaTex template to thesis template
  • Choose software (probably Caffe)
  • Select which forms of preprocessing to try

Questions

  1. Should I preserve a test set just for the final step (in December) or is it ok to use all data for the preprocessing choosing and then all data for the small vs big and all that?. Answer: Separate a test set right at the beginning. Treat preprocessing as a hyperparameter to fit. For transfer leraning and big vs small you can use the entire dataset but shuffle the test set to be different.
  2. Validation or 5-fold crossvalidation? Answer: Validation. If validation set is too small, then 5-fold.
  3. When checking for different preprocessings, fit all hyperparameters or only a subset or none at all? Answer: Fit learning rate and regularization. All other hyperparameters would be set to standard (including the network architecture).
  4. mxn o nxd for the name of dimensions? Answer: mxn. m examples of n dimensions.
  5. Is it a binary classification(cancer/no cancer) or 3 classes (micro/mass/nothing) or something else(detection). Answer: Image segmentation. Lesion vs. background.
  6. Which forms of image enhancement should I use?. No preprocessing or global contrast stretching?