00:34:11 Kerrie Geil: just a summary from earlier for the chat record since I neglected to hit record until after she said it:: comment out the "from mrcnn…." lines in the import block, there are instructions in section 2 about how to install mrcnn. these instructions are what worked on Laura's machine so there will likely be some troubleshooting required on your own system. 00:34:28 Kerrie Geil: We likely won't troubleshoot together today 00:46:27 Timothy: it took me about 30 min 00:47:59 ARS - Kossi Nouwakpo: It seems like many of these networks build off of other pre-existing networks (e.g. Yolo built off of VGG). Is transfer learning used in these instances? If so how do we ensure that these ML networks are better are solving new problems rather than solving the same problems faster? 00:51:56 ARS - Kossi Nouwakpo: Thanks 00:52:52 Scott Orr: In the paper, the kernels used varied from 7x7, 3x3, 1x1. How are they deciding on what size kernel to use for which convolutional layer? 01:16:30 Kerrie Geil: Our HPC users should probably be ok at this point with grabbing the zebra image, but just in case... File>New>Terminal cd to your working directory wget or curl -O and paste the link to the zebra image 01:27:06 Elizabeth Chin: can you do transfer learning with YOLO by fine-tuning the last layer(s) like we learned on previous days? 01:28:52 Elizabeth Chin: thanks! 01:34:41 Andrew.French: please let us know your thoughts on how to create libraries specific to our applications at ARS, is there already an effort to do this in the agency? then yolo would/could be useful 01:36:53 Jennifer Woodward-Greene: I would be interested in helping in that effort!! 01:37:32 Maria Laura Cangiano: In my lab we are starting to plan for a weeds repository 01:40:28 Andrew.French: would be great if ARS could create an (ad hoc?) working group so we don't have dozens of groups developing similar repositories, this would be timely for NP211 , water /nat resources planning 01:42:02 Kerrie Geil: Andy, Deb's been talking about an "ARS AI Center of Excellence". I'm not sure who all is involved in it, but your idea sounds like something that would be appropriate for that 01:47:03 Jennifer Woodward-Greene: I am getting a goat is a horse with .67 in one image, or a dog with .83.... I would have thought it may go with 'sheep' which is one of the labels! 01:47:41 Jennifer Woodward-Greene: Yes! 01:47:54 Jennifer Woodward-Greene: Sometimes people mix them up 01:48:11 Jennifer Woodward-Greene: Yes... some sheep have no wool... tricky 01:48:13 Maximilian Feldman: Yolo3 is 64.5% confident ‘latest_256_0193.jpg’ is a cake! :) 01:48:15 Jennifer Woodward-Greene: Beltsville 01:49:59 Andrew.French: cake is better than tick! 01:54:08 Jennifer Woodward-Greene: When I run the peppers.png I get a type error, expected str, .... not tuple 01:54:54 Lucas Heintzman: How sensitive is this network to orientation? If you turned the image 90 degrees would the classification change? 01:55:00 Gerardo A Armendariz: Monday’s exercise was able to detect a “brown bear” for me 01:56:02 Jennifer Woodward-Greene: the cameraman works fine 01:57:20 Suzy Stillman: I had an issue where load_image rotated one of my images and then it was not able to identify objects well 01:58:59 Jennifer Woodward-Greene: Reloading the image worked... although the original was still viewable; no obvious problems. 01:59:55 Jennifer Woodward-Greene: Good to remember this troubleshoot!! 02:00:26 Jennifer Woodward-Greene: Agree, I never use the tiff format.. to touchy. 02:00:55 Lucas Heintzman: And can you speak to the efficacy of ML system in general to differences in pixel cell SIZES? 02:02:13 Timothy: Laura we can see your screen anymore 02:02:21 Gerardo A Armendariz: rotation worked for me using the bear image with slightly less confidence 02:03:45 Gerardo A Armendariz: Right 02:05:06 Kerrie Geil: How are we saving the output image in the jupyter notebook to a png or jpg file? 02:05:12 Jennifer Woodward-Greene: It's impressive too that they are facing the camera, not a side view which would seem to be easier to determine what they are! 02:05:28 Kerrie Geil: ah ok. not on the HPC 02:06:31 Jennifer Woodward-Greene: Is there a config file needed for the next step? 02:06:47 Jennifer Woodward-Greene: NameError: name 'Config' is not defined 02:06:50 Jennifer Woodward-Greene: my own 02:06:50 Gerardo A Armendariz: Interesting, with 180 degree rotation, the bear is not recognized. 02:08:07 Jennifer Woodward-Greene: what is the line? 02:08:11 Timothy: could not detect the zebra in that image 02:21:51 Jennifer Woodward-Greene: I am searching for the tf.xxx code lines to change, and not seeing them... 02:24:14 Maria Laura Cangiano: it tells me that setup.py doesn't exist 02:25:07 ARS - Kossi Nouwakpo: Maria you may need to cd to the mask-rcnn folder that has been downloaded 02:26:53 Jennifer Woodward-Greene: should we try it first, or check our version of tensorflow? 02:27:05 Jennifer Woodward-Greene: in case we have the right version to match? 02:28:27 Laura Boucheron: import tensorflow as tf print(tf.__version__) 02:29:51 Kerrie Geil: I would need a few more minutes to get it up and running personally 02:31:08 Maria Laura Cangiano: pip show mask-rcn not found 02:31:28 ARS - Kossi Nouwakpo: maria typo add another n 02:31:34 ARS - Kossi Nouwakpo: mask-rcnn 02:40:09 Elizabeth Chin: can you please switch the screen to the jupyter notebook? 02:40:28 Elizabeth Chin: (it was showing the terminal) 02:40:32 Elizabeth Chin: thanks! 02:48:18 ARS - Kossi Nouwakpo: Is mask-rcnn limited to 3-layered images? Can it be trained for more layers? 02:49:26 ARS - Kossi Nouwakpo: Thanks 02:51:13 Lucas Heintzman: Is the ML package "smart" enough to expand a Grayscale image? 02:51:17 Jennifer Woodward-Greene: yes 02:55:13 Jennifer Woodward-Greene: ++ 02:55:16 ARS - Kossi Nouwakpo: yes 02:56:43 Jennifer Woodward-Greene: What environment should we reopen the notebook... assume aiworkshop...? 02:58:09 Laura Boucheron: DeepVis, Yosinki 02:59:28 Laura Boucheron: http://yosinski.com/deepvis#toolbox 03:00:02 Gerardo A Armendariz: Identified the flipped bear 👍🏼 03:00:40 Elizabeth Chin: very cool. thank you! 03:03:55 Jennifer Woodward-Greene: sure play 03:04:20 Jennifer Woodward-Greene: yes 03:05:25 ARS - Kossi Nouwakpo: there is an echo I suggest muting your computer 03:08:03 Jennifer Woodward-Greene: everyone else needs to mute... someone else is listening in their own time... 03:12:31 Kerrie Geil: If you want to leave for lunch we'll start back up at 1:30pm Mountain 03:13:30 Jennifer Woodward-Greene: Am unable to get rid of tf.log error... I have changed it to tf.math.log in model.py 03:13:48 Jennifer Woodward-Greene: and have re-imported 03:14:29 Jennifer Woodward-Greene: trying restart kernel and rerun 03:16:57 Jennifer Woodward-Greene: Getting a None type error now... 03:17:54 Jennifer Woodward-Greene: I think it is all installed, but not got it working in the notebook... 03:18:05 Maria Laura Cangiano: Haven't been able to make it work yet 03:18:25 Elizabeth Chin: thank you! 03:18:31 Jennifer Woodward-Greene: c:\users\jennifer.woodward\appdata\local\conda\conda\envs\aiworkshop\lib\site-packages\tensorflow\python\framework\tensor_util.py in make_tensor_proto(values, dtype, shape, verify_shape, allow_broadcast) 443 if values is None: --> 444 raise ValueError("None values not supported.") 445 # if dtype is provided, forces numpy array to be the type ValueError: None values not supported. 03:19:01 Jennifer Woodward-Greene: I forced it to install 1.2.1 03:27:09 Jennifer Woodward-Greene: Attempting uninstall: tensorflow Found existing installation: tensorflow 2.3.1 Uninstalling tensorflow-2.3.1: Successfully uninstalled tensorflow-2.3.1 03:27:13 Jennifer Woodward-Greene: and more like that.. 03:29:05 Jennifer Woodward-Greene: Successfully installed astor-0.8.1 gast-0.2.2 keras-applications-1.0.8 tensorboard-1.15.0 tensorflow-1.15.3 tensorflow-estimator-1.15.1 (aiworkshop) D:\ImgWorkshop> (aiworkshop) D:\ImgWorkshop>python Python 3.7.9 (default, Aug 31 2020, 17:10:11) [MSC v.1916 64 bit (AMD64)] :: Anaconda, Inc. on win32 Type "help", "copyright", "credits" or "license" for more information. >>> import tensorflow as tf >>> tf.__version__ '1.15.3' >>> 03:29:29 Jennifer Woodward-Greene: try the notebook! 03:31:18 Jennifer Woodward-Greene: LOL 03:31:20 Jennifer Woodward-Greene: ImportError: Keras requires TensorFlow 2.2 or higher. Install TensorFlow via `pip install tensorflow` 03:31:36 Maria Laura Cangiano: module 'tensorflow_core._api.v2.sets' has no attribute 'tf' 03:32:11 Jennifer Woodward-Greene: I will go back to the higher version of tensorflow.... thanks! 05:05:15 Lucas Heintzman: Is this akin to a modified 'Monte Carlo' design? 05:06:10 Lucas Heintzman: Not important...just an idea. 05:06:28 Kerrie Geil: If your airlines data downloaded as .csv and not .txt, make sure you change the file name to .csv in the pandas read_csv function 05:07:17 ARS - Kossi Nouwakpo: i got a .csv.txt 05:14:20 ARS - Kossi Nouwakpo: So are X and Y shifted by 1 value? 05:16:54 ARS - Kossi Nouwakpo: the size of X changes as the lookback changes. Why is that? 05:17:43 ARS - Kossi Nouwakpo: The column number increases as well 05:18:18 ARS - Kossi Nouwakpo: Thanks 05:35:48 Jennifer Woodward-Greene: fast 05:36:49 Jennifer Woodward-Greene: is there a setting to have it stop itself? 05:36:57 Jennifer Woodward-Greene: when the loss stabilizes? 05:47:11 Jennifer Woodward-Greene: I have done that with SAS bootstrap I think.... 05:49:05 Jennifer Woodward-Greene: even # look back fails 05:49:15 Jennifer Woodward-Greene: 2 05:49:33 Jennifer Woodward-Greene: dimensions 05:49:53 Jennifer Woodward-Greene: I could run it but not the subsequent steps that had dimension error 05:50:17 Jennifer Woodward-Greene: my luck?? 05:51:08 Jennifer Woodward-Greene: i can't repeat it either... 05:51:17 Jennifer Woodward-Greene: my luck!!! 05:52:33 Lucas Heintzman: Anyone getting memories of Lokta-Voltera models while looking at this data? 05:55:56 Kerrie Geil: I changed the test data to 20% and used a look back of 13 but still getting a test score of 38.86 RMSE. I guess a much longer timeseries or the addition of more "features" might improve this? 05:58:36 Andrew.French: you can pre-process with a box convolution filter to trim the extremes, then use the filtered time series as the update input 06:10:18 Suzy Stillman: I think this is what Maria did, but if look_back is 12, results are pretty good because we remove the seasonality (predicting January from the previous January, etc.) is that correct? 06:12:59 Kerrie Geil: where is the doc page that shows all the options to put in for loss in .compile? 06:13:14 Elizabeth Chin: I set the n epochs to 50 and the look back to 12 as Maria did but get different RMSE values even thoug hthe same seed was set (42). Is this expected behavior and the seed not controlling inner workings of the model but some other stuff?