00:44:45 ARS - Kossi Nouwakpo: How do you handle situations where you have a sequence of images over time and we want to classify the image based on changes in the images. 00:44:50 Jerry M: Squares are easy to rotate 00:45:38 Jennifer Woodward-Greene: Looking at changes over time... packages or programs developed for video would be a good choice? 00:47:47 Maximilian Feldman: Is the process for inputting images that are large (500 x 500 pixels) the same as for small images (28 x 28 pixels)? 00:54:45 Elizabeth Chin: when I plotted the preds of model 4, it was, for example, predicting a 4 when the ground truth was 4, but it thought it was incorrect. 00:55:00 Lucas Heintzman: Any idea why model4 would have an accuracy of .13 via the CNN and a .98 via personal evaluation? 00:57:25 Jennifer Woodward-Greene: model3 and model4 myacc is the same as the evaluate functions 01:33:52 Elizabeth Chin: I got a higher test score than training- isn’t that kind of weird? 01:34:06 Andrew.French: c hanging batch size doesn't seem to change results 01:34:07 Timothy: epochs=5, accuracy =0.7185. A little lower than the train test 01:34:21 Kerrie Geil: if I allow both fully connected layers to be trainable and do 3 epochs I can get about 90% accuracy on the test data 01:35:13 Elizabeth Chin: can you apply cross validation when re-training the last layer? 01:36:56 Elizabeth Chin: ok thanks! 01:37:36 Andrew.French: my results all maxed out around 52% 01:37:53 Jennifer Woodward-Greene: the accuracy prediction was really bad! everything worked tho 01:43:48 Andrew.French: weird, took out 10 layers and went up in accuracy a lot, now over 90%?? 01:45:20 Andrew.French: that for loop minus valu8e 01:45:20 Timothy: accuracy increased to 94% with freezing 3 layers 01:46:21 ARS - Kossi Nouwakpo: So the information in the first few layers is not really that dependent on the training data? 01:46:34 Timothy: oh yes I did for 5 epochs 01:47:10 ARS - Kossi Nouwakpo: thanks 01:48:41 Jerry M: Is using transfer learning similar to building a cascade classifier? 01:50:08 Jennifer Woodward-Greene: With all but last two layers, 4 epoch, I got .92 accuracy. Prediction was .9034 accuracy, and the same as myacc. 966 incorrect classifications. 01:53:31 Lucas Heintzman: This is a general question: but in terms of image identification/classification, are there "significance" thresholds that are recommended for different CNNs? 01:55:17 Lucas Heintzman: Yes 02:01:12 Jerry M: Comment- found a link about the difference between binary and categorical loss functions. I’m thinking if binary is used it means a digit is non-exclusive. (A single digit can be 1 and 2 simultaneously or maybe not a digit at all). If categorical is used a digit is exclusive. (A digit can be 1 or 2 but not both and it has to be a digit). https://stats.stackexchange.com/questions/260505/should-i-use-a-categorical-cross-entropy-or-binary-cross-entropy-loss-for-binary 02:09:08 Timothy: Thank you so much 02:09:09 ARS - Kossi Nouwakpo: Thank you 02:09:12 Elizabeth Chin: is the h5 filetype exclusive to keras? could you load these models using pytorch? 02:09:13 Dylan Burruss: Thanks! 02:09:20 Matthew.McEntire: thank you ! 02:09:21 Jerry M: thanks! 02:09:21 Amy Hudson (she/her): thank you! 02:09:22 Kale Harbick: Laura great workshop thanks! 02:09:30 Jennifer Woodward-Greene: Thank you, agree AWESOME job! 02:09:31 Gerardo A Armendariz: Thanks for the great workshop! 02:09:31 Maximilian Feldman: Thanks!