Instruction
stringlengths 0
86.4k
⌀ | Response
stringlengths 0
127k
⌀ |
|---|---|
Great explanation! By the way Can we calculate the number from the locations of each player? I want to fit the value into its NaN. | I see clearly. Thanks a lot ) |
High scoring public kernels were posted. They broke the leaderboard. A bad marketing tactic in my opinion. | I never understood this argument. If you work hard on your solution and you believe it is unfair that everyone can use ready-to-use model why you can't ensemble them? |
Why is so important to you direction in which players are facing? It does no tell much. They may have been rotating their heads just when the data were collected. Also, they don't have to move in the way they are looking. | For sure, I am not questing the importance of direction, but direction on contrary to orientation, can not change in a split of a second, because players have momentum. If we had more frames from a single play, then we could have been sure that a player was looking in a certain direction, but from a single frame, it is not so certain. Nevertheless, if it works, then it is great :) Also, the reason why the score lowers when the orientation is transformed, so that it is same every year, may be, that a model is using orientation just to distinguish 2017 plays from other plays. |
it looks like You haven't use all the folds you have used only one fold. msss = MultilabelStratifiedShuffleSplit(n_splits = 20, test_size = TEST_SIZE, random_state = SEED) X = train_df.index Y = train_df.Label.values Get train and test index msss_splits = next(msss.split(X, Y)) train_idx = msss_splits[0] valid_idx = msss_splits[1] But you have made 20 splits. Why?? I think it will improve the score if we have used all the 20 splits. | Hi . That's exactly correct :-) I use the 20 folds when running my code at home on my datascience desktop. Just copied it as is. Feel free to change it though. With the 9 hour time limit you won't be able to use them all 20 or you should decrease the steps_per_epoch even further (set it to 1/12 or even 1/13 or 1/14). Would be interresting to see if that would improve the score indeed. |
Validation - 1.18 LB - 1.29 What is the validation technique are you using? | K-fold, but I'm getting similar validation scores with random subset or time split. I don't know their correlation with LB however. |
What about the nn. What private score, can you share the code? | may I ask what is "kernel median" and "PrEp" in the code? |
Hi , I just tried running your exact code on the kaggle kernal but it produce this error on the Unet chunk: AttributeError: module 'segmentation_models.backbones' has no attribute 'get_preprocessing'. Any idea why this might occur? Thanks! | The first parameter on the model is "backbone_name" so you can pass it explicitly or just put "BACKBONE" first, it will work anyway, just pay attention to the log because it says what model is being loaded. |
Great charts! | Thank you! |
Those are some crazy plays! Sorry, Rob, but think about the happiness experienced by the Michigan fans! | Never! 😄 |
Nice approach 😄 | thank you |
Thanks for sharing! :) | you're always welcome ;) |
Nice one | thank you |
Is anyone or any team looking to seriously compete in this competition that I can join with? I can rent cloud GPUs for us to train on via potentially vast.ai. | Hello , I have added all the team members on Skype, kindly share your skype details too. Thanks, Ranjan Rishi |
Our apologies . Our team is working on this and we hope to have it resolved sometime on Monday. Thank you for your patience. | Thanks for replying . At least now we know the reason. |
Thanks! | thanks a lot! |
How can we know safely that our inference kernel will work (within 1 hour) on unseen data? Is it already what happen when we click on "Submit to competition"? Kernel might have bugs or memory issue due to new data. Update: It's still not clear for me if 1 hour is on commit or on submit: https://www.kaggle.com/c/severstal-steel-defect-detection/discussion/108038 Some discussions are reporting 3 hours on submit and 1 hour on commit. Submit would include full private data so if it runs then we know we don't have blocking bug. Could you confirm or not? 1 hour on commit or on submit. And submit already runs on all private data (but currently scoring only on public)? Thanks a lot. | You can have code as part of dataset. |
thanks for sharing the kernel Den. Could you please tell me what are you using pretrained efficient_net( model_eff in the code) for? As much as i can see, you are using pretrained resnet34 for transfer learning (via vision.cnn_learner), right? Just wanted to know if i am missing something. Thanks | Hello. Thanks for you question, you right. So, in efficient has one problem with bool type of size image. Now i fixing this bug and this bug exicted because we use FastAi in pytorch API all is good. Some about code, for using efficient net need to change vision.models.{model} to base_arch = {model}. If you have other questions, you can ask me. |
If your current LB is 0.898, I suggest that you focus on class 3 and 4 for now. With just class 3 and 4, you should be able to get above LB 0.910 with efficientnetB4 segmentation and classification. | , you mentioned "Please remember to do postprocesssing - remove noise masks from submission." Can you share a bit more how to identify noise? Thanks a lot. |
Is anyone or any team looking to seriously compete in this competition that I can join with? I can rent cloud GPUs for us to train on via potentially vast.ai. | Skype email: safi.zaidi@hotmail.com. I say lets all connect in one group so the conversation can be more simpler. |
Hi , I just tried running your exact code on the kaggle kernal but it produce this error on the Unet chunk: AttributeError: module 'segmentation_models.backbones' has no attribute 'get_preprocessing'. Any idea why this might occur? Thanks! | Hi , how do you switch from the Pytorch API to the Keras API? Thank you. |
I'm playing with inference kernel since a few days to find the some combinations that work for allowed infer time (1 hour). I can tell that if your commit time is around 25 minutes then it should work. Above 30 minutes leads to Notebook timeout on submit. As said by Kush, your infer time is around 3 times more than the one on commit (33% of data). | hi, my commit time is around 20 minutes, but i got "Notebook Exceeded Allowed Compute" on my all submits. Do u know what the problem is? thx a lot |
With raw target variables we can't use lovasz loss, but if we add one more 'undefined' class, we can use lovasz softmax. | That might actually work. Thanks for suggestion. |
Thanks for reporting . Our team is aware and are handling the issue, but may not be completed until Monday. Apologies; hang tight! | OKAY! 😑 |
Hi!) Hmm, your answer is correct, but for some reason the checker doesn't accept it. Have you tried restarting the session? I ran your query (with original formatting) and didn't get any errors. But I can't understand what this code segment does and how it influences the result. It limits the amount of data processed by your query to 10 GB (if this limit is exceeded, the query will fail) and converts the results to a pandas DataFrame. The query_to_pandas_safe() function you use in your solution does the same thing. I looked for it in the tutorial section but couldn't find any mention of any of these methods. The last section of the second tutorial briefly discusses possible ways of estimating query costs. For more info, read the docs. | Thank You! The cell runs perfectly after restarting the kernel. Thanks for the explanation as well. |
Hi , I uploaded the app icons as a zip file so people didn't have to download each one themselves, the link is in the description. Looking forward to seeing more of this kernel. | Thank you so much for your kindness. :) |
In the Rules tab https://www.kaggle.com/c/nfl-big-data-bowl-2020/rules I don't see any GPU restriction, am I missing something? | Found GPU disabled at https://www.kaggle.com/c/nfl-big-data-bowl-2020/overview/code-requirements but we can use GPU for training while use CPU for inference, right? |
I believe it's because when we use sigmoid as the final layer activation function, MSE and CRPS are the same thing.. | Exactly. |
It is not necessarily multilabel classification, you can also predict the number of gained yards (regression problem with MSE) and based on that set CRPS prediction. | If you treat this as a regression problem, then MAE should be used IMHO. See https://www.kaggle.com/c/nfl-big-data-bowl-2020/discussion/112621 |
This is very helpful to understand dataset. Thanks! | Thanks :) |
in main.py torch.backends.cudnn.deterministic = True This will more slow than set False. | The deterministic flag is to help reproduce results. Usually gpu cores work in semi chaotic order, producing slightly different results if you run the same training twice (even with the same random seeds everywhere else). The deterministic flag cause cudnn the produce the same result if you rerun it with the same inputs twice. in doing so, you lose a bit of speed as concurrency is more limited. At least that is my understanding.... |
If your current LB is 0.898, I suggest that you focus on class 3 and 4 for now. With just class 3 and 4, you should be able to get above LB 0.910 with efficientnetB4 segmentation and classification. | You can calculate how many pixels are in mask and fill this mask with all zeros if number of pixels is small. This is a trick which boosts a score very much but I believe you use it already. Second trick is to change the threshold. Instead comparing output with 0.5 try other values. |
in main.py torch.backends.cudnn.deterministic = True This will more slow than set False. | Hi, What is the reason to set at True being slow instead of set False to be faster ? |
THANK YOU!!! | yes!!! have a nice day |
Judging by the very low importance of floor_count and year_built, do you think it is worth imputing the values or keep them NAs? | Depends on the model that you are using. For lightgbm, it is better to not impute anything most of the time. |
Thank you for posting this Konstantin! Your TPU code is very informative, I'd tried a few times to get pytorch_xla working before without success so seeing a pytorch model training on the tpu is most excellent. | Good point, large dataset size is another problem for Colab. With GCP it's not as issue (but be sure to use an SSD). |
How do you upload your files in colab? Or do you have access to the GCP? Thanks | I worked on GCP, yes. Unfortunately with pytorch dynamic loading approach, Colab VM is not powerful enough to fully utilize a TPU. There are some examples though https://github.com/pytorch/xla/tree/master/contrib/colab#get-started-with-our-colab-tutorials and you can also connect to Colab via ssh. |
I am confused from the beginning of the competition . But I interpreted the rules as : The Test dataset given to us and we see is both Private and Public data . Only the public score is calculated based on 33% of this dataset and everytime we submit the whole private score based on 100% of this test data also is getting calculated . There is no separate hidden dataset . But my interpretation could be wrong ? | My idea is same as (the reason is 3 times longer running time) So I think there is separate hidden dataset whose distribution might be different from public test dataset. |
Hi, it seems that the images are not in standard dicom format, either they are compressed or something else (the buffer size is not the full image size). What am I missing? thanks | So - I have found a workaround - one can read the images using SimpleITK |
Nice kernel! I tried running this in Colab, but I get the following error: ``` TypeError Traceback (most recent call last) in () 40 41 ---> 42 model.compile(optimizer = Adam(learning_rate = LR), 43 loss = 'binary_crossentropy', 44 metrics = ['acc', tf.keras.metrics.AUC()]) 1 frames /usr/local/lib/python3.6/dist-packages/keras/optimizers.py in init(self, lr, beta_1, beta_2, epsilon, decay, amsgrad, **kwargs) 455 def init(self, lr=0.001, beta_1=0.9, beta_2=0.999, 456 epsilon=None, decay=0., amsgrad=False, *kwargs): --> 457 super(Adam, self).init(*kwargs) 458 with K.name_scope(self.class._name_): 459 self.iterations = K.variable(0, dtype='int64', name='iterations') /usr/local/lib/python3.6/dist-packages/keras/optimizers.py in init(self, **kwargs) 77 if k not in allowed_kwargs: 78 raise TypeError('Unexpected keyword argument ' ---> 79 'passed to optimizer: ' + str(k)) 80 self.dict.update(kwargs) 81 self.updates = [] TypeError: Unexpected keyword argument passed to optimizer: learning_rate ``` | Hi thanks for the compliment. I have not yet tried the Colabs so I 'am not familiar what the differences between the Kaggle kernel and the Colabs are. If you have some more error details I'am happy to help you with some googling. |
If you are committing your kernel, yeah sure, you can close your browser, it only kills your process if the running time exceeds 6 hours. But if you are running the kernel normally (not committing), it will cancel your process when you close the browser | if we don't close the browse and commit ,how long will it run? |
interesting, i found using MAE objective alongside crps worsened my model's performance. i didn't use [1,0.01] loss weights though, i used [5,1] and [1,1]. Will have to retry. Thank You! | yes! my crps cannot improve my LB score too, if my crps wrong . i use from keras import backend as K func for metric def crps(y_true, y_pred): ans = 0 ground_t = y_true.argmax(1) for i, t in enumerate(ground_t): for n in range(-99, 100): h = n>=(t-99) ans+=(y_pred[i][n+99]-h)**2 return ans/(199*len(y_true)) |
having more ram helps no matter what you are using ;) | Sounds very nice. |
Is there still value in converting to KITTI as only the front facing camera data is retained ? What i am not able to grapple in this competition is where to even start ? | Thanks for the hint . Looking at PointPillars |
having more ram helps no matter what you are using ;) | I'm using R too. I have 48GB of ram and it makes things a lot easier. |
Hi all, to clarify: You are allowed to train offline and upload model weights back to Kaggle. However we'd caution you in your approach. 1) External datasets and pre-trained models are still prohibited (External datasets and pre-trained models do not include work you create on your own, whether locally or in a cloud environment). 2) Internet is turned off. 3) You are required to submit your model through the API, and the Stage 2 leaderboard will be determined on how your model performs on games yet to be played this season. Submitting only predictions would not apply, as the plays you will be predicting have not yet been determined. 4) Before the evaluation period begins, we will be updating the train.csv file to include current season games through Stage 1. Please take note should you want to retraining to be a part of your model submission. Thanks! | I agree. I also haven't had luck loading the model directly. I get this error because the model file is too large to paste as a string. If I upload it as a dataset it works, but won't let me submit. Error: The kernel source must be less than 1 megabyte in size - Is there any other way we could upload models as datasets? Otherwise I don't think training of many model types offline is feasible. |
Thanks for sharing! I have s silly question: in the training part, I tried using early stop and restore best weights, but only got 0.01386 on LB, is that some trick in not using early stop? | My guess is that the 0.01385 LB score is not that easily reproducible, scores may vary slightly. This is due to the random factors inside the NN. For instance, the seed is not defined at the Dropout layers. Which is nice, to stop users from Fork+Run+Submit without understanding what's going on! Hehe :) |
efficientnet-b2 410x410 single fold hflip tta publicLB: 0.064 | Cool. Could you please share the windowing method? |
thanks for sharing! | Thanks for support sir ! I've helped you ! 🙏 |
test dataset has 2-year data, train set has 1. | Got it. Thanks! |
If I may play devil's advocate . He/she did participate in discussion forums ,so not in hiding . The public kernel tweak can give you good score and some more if you have worked on image segmentation before . The person in question could genuinely be talented , not dumb . | Kaggle has good detectives, they will find out. |
I'm using Keras as well. One (of several) neat techniques I've learned over this very competition is how to implement early stopping. Once you do it, you can forget about number of epochs: one parameter less to worry about. Check this great notebook from ! | But then you get the fun of tuning patience! |
tokenizer.fit_on_texts(list(train_X)) ตรงนี้ถ้าเราใช้พวกตัดคำภาษาไทยแทนได้ไหมครับ เหมือนเวลาตัดออกมามันไม่เป็นคำ มันเป็นตัวอักษรเลย {'ท': 1, 'า': 2, 'ั': 3, 'ี': 4, 'ด': 5, 'ร': 6, 'บ': 7, 'ม': 8, 'ง': 9, 'อ': 10, 'ะ': 11, 'จ': 12, '้': 13, 'น': 14, '่': 15, 'ย': 16, 'ส': 17, 'ิ': 18, 'ธ': 19, 'ก' : 20, 'ป': 21, 'ใ': 22, 'ผ': 23, 'ู': 24, 'ถ': 25, 'แ': 26, 'ล': 27, 'พ': 28, 'ค': 29} | พอจะยกตัวอย่างการเขียน code ตัดคำภาษาไทยแล้วทำการแปลงข้อความให้มีเครื่องหมายตัดคำ เช่น '|' ได้ไหมคะ จะลองนำไปใช้กับข้อมูลภาษาไทยที่ตัวเองมีอยู่ค่า ขอบคุณค่ะ |
having more ram helps no matter what you are using ;) | True. True. But my laptop I'm using rn is not that great. I do encounter situations such as 'R session aborted', which is a nightmare for R users haha |
Nice kernel, thank you | Thanks, 👊 |
ถ้ามีคำแนะนำ คำถามเพิ่มเติม หรืออยากได้ workshop ประเภทไหนในอนาคต สามารถเม้นตรงนี้ได้เลยครับ | อยากได้ workshop การทำ text classification ของภาษาไทยแบบละเอียดเหมือน workshop นี้ค่ะ เพราะอ่านเข้าใจง่าย แต่ตัวเองเพิ่งเริ่มศึกษา NLP จึงยังติดขัดกับการประยุกต์ใช้กับภาษาไทยค่ะ ขอบคุณล่วงหน้านะคะ |
That is the one of the best kernels that I've ever read about this topic! Thank you so much for sharing it! ;-) | Thank you for kind words 😃 |
I get this error while trying your code AttributeError: ‘AUROC’ object has no attribute ‘output’ def on_epoch_begin(self, **kwargs): self.output, self.target = [], [] def on_batch_end(self, last_target, last_output, train, **kwargs): if not train: self.output.append(last_output) self.target.append(last_target) Is it a version issue? Thanks in advance. | Hi, what's your solution? Tanks! |
this is a great kerenl that help us to learn how to Create baseline on our own, I have do some time with this kernel,BUT THER IS ONE PIECE OF CODE I DON'T UNDERSTAND, train.loc[train.VisitorTeamAbbr == "ARI", 'VisitorTeamAbbr'] = "ARZ" train.loc[train.HomeTeamAbbr == "ARI", 'HomeTeamAbbr'] = "ARZ" what is logical within the code, I see It change Visitor with Home abbr here, BUT I DON'T UNDERSATND the purpose of the code, could you please kindly help me out with it? thanks a lot | thanks~ |
Wonderful kernel. Thanks for sharing!! | Thanks you for a good comment :) |
efficientnet-b2 410x410 single fold hflip tta publicLB: 0.064 | I used Appian's windowing and didn't use CQ500. |
nice job :) may I ask how many features you are training on roughly for each tree based and nn based models? | I think I didn't quiet get to you. I can see that the data consists of 23171 plays and each play consists of 22 player-based rows.. hope I'm correct. then how did you preprocess the data to make (509762, 49) to (23171, 120) shape? are you just picking one player row out of 22 rows in each play? sorry for the newbie question :) |
Hello, good job for this classification work. Could you explain to me why you choosed to make data augmentation with autoencoder and not with "traditional" tools like image rotation? If i understand well, you'r autoencoder add noise on your reconstrcuted image, why you don't directly add noise in original image? Also, why do you use augmented image as test set? I though that was dangerous to do this | it looks to me that using augmented image as test set results in such high accuracy in the test set, because that it is divided on such a large number. In addition, there is not much change before and after data augmentation. i doubt how meaning it is. |
nice job :) may I ask how many features you are training on roughly for each tree based and nn based models? | Yes I am, since the beginning. Actually I don't see how not to: the way the submission procedure is setup, it requires us to submit one play at a time, by feeding us one play at a time (22 rows of data from the 22 players who participate in the play. To be clear, in my pre-processing routine I take in train.csv in the shape (509762, 49) and output a pre-processed matrix ready to train in the shape (23171, 120). (just added 2 more features :)) There are 23171 different plays in the training set, which is exactly 509762 divided by 22. |
nice job :) may I ask how many features you are training on roughly for each tree based and nn based models? | so you are not treating the data as single rows? mentioned in here |
nice job :) may I ask how many features you are training on roughly for each tree based and nn based models? | so far, 118 |
can you share your notebook link. try to understand approach of ml problem from this kernel. https://www.kaggle.com/startupsci/titanic-data-science-solutions/comments#653066 | https://www.kaggle.com/abilp13/flower-classification-model-tensorflow Hello! There is the link to my dataset, I will check out the link you sent me. Thank you so much! |
Look here (for R) or here (for Python) on how to map Yardline info to line of scrimmage. | Thank you. |
Very cool and thorough. Like your detailed thoughts and conclusions. 👍 Have you by any chance already checked if the outlying buildings actually degrade predictions? | I should. I'll add this to my backlog. |
Great ! I upvoted ! 💪 Have you ever verify if each buildings (low multivariate series) have outliers? And later using approximate harmonic series for identity and removing positive outliers? | I have planned to do it. This notebook is for outlier whole timeseries, I've carried out one specific notebook for zeros, I'm making ready another for gaps. And finally I'll do one for values. |
Glad to see you here This competition has a very large data set. Can you focus on some memory optimization and resampling topics this time? | Thanks about this revelation . I always used to think that everyone else is using kaggle kernel efficiently and I am the only one who find it difficult to use the kaggle kernel because memory/CPU limitation and random failures; May be I process is not efficient enough! |
Great work Güneş. I have many questions, I will be very happy if you respond to all, so here they come: You said "it is convenient to work on concatanated training and test set however this approach causes "data leak" and therefore avoided. Assuming you know what data leak is, why did you carry out imputing it in concatanated version? Have you tried imputing Age using Regression? Was this current imputing of yours your only approach/try? Well done finding about actual Embarked value of two passengers by Googling! How come you accept higher than 0.1 is a high correlation? Over 0.50 is considered high correlation as a thumb of rule. About encoding, how did you decide which feature to encode as One-Hot and which feature to encode as especially target encoding? Can you please explain furthermore about the decision of Target Encoding where you said "Target encoding could be very useful as well because of the high correlations with Survived feature." I suppose you did not even use Target Encoding in the end? How did you choose the number to bin, specifically 13 for Fare and 10 for Age? You even mentioned that people below 15 are highly likely to survive but you chose 10 to bin Age so it's a bit confusing here. You didn't actually fit to training set first, then transform it, and then using that fit to transform test, rather you fit them separately, isn't that wrong in terms of methodology? As far as I know both should have been transformed into same space (as if it's PCA). What do you think about feature importances? Different algorithms in sci-kit learn provides different feature importances for the same dataset? I see you used RandomForest here but what algorithms in scikit learn are the most consistent and true ones to return feature importance? | Thanks It is convenient to work on concatenated data because train and test could be from different samples. When I combine them, we see the whole distribution. Imputing age with regression could work but it is overkill. I didn't think on different approaches for age imputation so far. There is no general rule of high and low correlation. It is totally dependent on the data set. I was experimenting with One-Hot encoding back then when I was making this kernel and I tried it on many of the categorical features. I wasn't expecting it to work, but it did and I kept it. :D I used target encoding on Family and Ticket, then averaged the rates. Actually, that is the leak. Bin numbers doesn't have any meaning. I tried all of them and kept the best number for me. I should have fitted them together as you said. This way they are scaled differently. There are different ways to evaluate feature importance rather than the model's importance. I suggest you to use an external package like shap or lofo-importance for feature selection. They are more reliable. |
I'm using Keras as well. One (of several) neat techniques I've learned over this very competition is how to implement early stopping. Once you do it, you can forget about number of epochs: one parameter less to worry about. Check this great notebook from ! | I'm getting a very similar chart... with early stop, step size 32, it rarely gets to 20 epochs Usually converges between 10-15 |
Great Sayani. Upvoted. | Thanks Marilia |
Excellent work is done!!👍 | Thanks |
Minification part added: pickle is not the unique option btw. You can try use feather DataFrame.to_feather() or Google BigQuery table DataFrame.to_gbq() -> it would be amazing if someone can prepare BigQuery integration notebook for with competition or parquet DataFrame.to_parquet() | Here is a blog post which compares feather, md5, parquet, pickle etc. And it shows that feather is the most optimized one. I am using feather for sometime (both at work and projects) and found it to be good (Didn't do any comparison though). https://towardsdatascience.com/the-best-format-to-save-pandas-data-414dca023e0d |
I'm using Keras as well. One (of several) neat techniques I've learned over this very competition is how to implement early stopping. Once you do it, you can forget about number of epochs: one parameter less to worry about. Check this great notebook from ! | May I ask how many epochs until the validation loss don`t drop in your model ? About 6-10 epochs in my model..png?generation=1571718349214672&alt=media) |
efficientnet-b2 410x410 single fold hflip tta publicLB: 0.064 | Are you using the new CQ500 dataset? |
Great visualizations and commentary, thanks for sharing. Couple of questions for you regarding Grid Search: - First, how did you decide on your param_grid? In other words, how did you choose which hyperparameters to tune and over what range of values to iterate? - Second, just out of curiosity, did you perform Grid Search within a Kaggle kernel or on a local machine/cloud service provider? In my experience, Kaggle kernels are noticeably throttled. I was just wondering how you managed to get something like that to finish and how long it took. I personally used RandomizedSearchCV instead, just for efficiency. Thanks again. | Thansk for your comment!! I got param grid values with the help of grid search , in above kernel section 7 talks about the same *Below is the relevant code for your reference * `from sklearn.model_selection import GridSearchCV param_grid = { 'bootstrap': [True], 'max_depth': [80, 90, 100, 110], 'max_features': [2, 3], 'min_samples_leaf': [3, 4, 5], 'min_samples_split': [8, 10, 12], 'n_estimators': [100, 300, 500, 1000] } grid_search = GridSearchCV(estimator = classifier_rf, param_grid = param_grid,cv = 3, n_jobs = -1) grid_search = grid_search.fit(X, y)` I got list of parameters from Grid search official documentation and it's corresponding values by hit and try method . In addition to this I refer other people's blogs and kernels for values. Regarding performing Grid Search within a Kaggle kernel , you are correct kernel will throttle that's why I have commented that particular section after first time run because it took more than 40 minutes to complete the execution. I am still learning on Grid search and hyper parameter tuning , will update you if I came across a better approach . I hope this answer your questions 😊 |
Hello, I'm also seeing the same newsfeed for 15 days. Any trick to solve this? Thank you Jesús | After my post the problem was fixed. |
I enjoyed your visualizations and mapping! Something I noticed is that there are 3 states in Brazil that begin with "Rio" (similar case with Mato Grosso), but they all get the same labels in the data. I go more into this and made an analysis of the same data if you're interested in taking a look. https://www.kaggle.com/etsc9287/a-quick-analysis-of-forest-fires-in-brazil | Hey! Thank you - yes, I will take a look. Yes, the Rio breakdown was already well discussed in Discussion section; hopefully those better breakdowns could be made by the owner of the dataset :) |
Great work. I am working on a Data Science project for my University and this is very helpfull. I would like to know if there are any others questions/key challanges that can be inspired from this Dataset. | Thank you! I believe this dataset could be an extremely good example for practicing your geographical analysis - for example, breaking down each state on a heat map by years provided that would be a good amount of work and good amount of powerful visuals :) In any case, it is always good to go with some sort of the idea on what problem you want to approach and what you want to achieve! |
Hey, another amazing kernel, could you do a Kernel intensive on Loops? They are very useful but not so intuitive at the beginning. Thanks | Hey! Yup, in my next kernel that I'll be doing I can provide extended explanations for Loops; there I'm going to use them to parse an address string, so it would be a good example for many! |
Nice work bro! Thanks for sharing. Commenting would be useful though. | Thanks ! ✔️ Thanks for insights sir! And help me! I'll descript all outputs and sections. |
Hi all, to clarify: You are allowed to train offline and upload model weights back to Kaggle. However we'd caution you in your approach. 1) External datasets and pre-trained models are still prohibited (External datasets and pre-trained models do not include work you create on your own, whether locally or in a cloud environment). 2) Internet is turned off. 3) You are required to submit your model through the API, and the Stage 2 leaderboard will be determined on how your model performs on games yet to be played this season. Submitting only predictions would not apply, as the plays you will be predicting have not yet been determined. 4) Before the evaluation period begins, we will be updating the train.csv file to include current season games through Stage 1. Please take note should you want to retraining to be a part of your model submission. Thanks! | isn't there a way of freezing an external dataset so that there is no internet leakage possible during the competition? Even though it might be possible to copy paste the weights of a model (not sure it's very convenient), it's really hard to copy an entire library with dependencies... I managed to install my library offline by creating datasets containing wheel files but I can't submit from this notebook. Freezing dataset would ensure that this is just external code that can't leak anything in the future. Any help? |
Interesting! Quite a good result already. I wonder what a linear regression with only log_square_feet as feature would get. | This was my first quick attempt to do a LR benchmark. There are many additional ideas I’d like to explore, and might give the log squared feet a try. |
Nice, thanks for sharing. This is what makes Kaggle competitions worthy to participate even when you're a beginner :) Just a few words on this: be careful, features might be more important for GBMs but not for NNs, and vice-versa! | Obrigado! |
Big thanks for sharing!! This post came just when I was running out of ideas and was sort of stuck. Guess now got some new features to implement and play around with. :D PS: I focused more on features such as weather playing conditions (i.e. what conditions each team is used to practice/play in vs the conditions they play the match), voronoi area of runner, but a lot of the features I engineered don't seem to contribute much... | I played with some of these features as well. For me, features like stadium type, weather, etc add very little to my models as well... |
resnet-34 single model, single fold, tta LB 0.90283 can't reproduce on local GPU | 3 folds out of 5 LB 0.91928 I was 16th and now I am 18th kernel time is 866.6s limit is 1200s 3 days to go I can do 4 submissions on tuesday, 5 on wednesday and 5 on thursday GPU Total Usage: 12 hours, 19 minutes, so I have almost 18 hours to use, that's OK |
Weighted average and simple average, I prefer weighted average | Hi , I don't mean the technique, but how you handle loading all models and making predictions, when I try to do it I get memory error and my kernel dies. |
Very useful and productive. Thanks | Thanks, 👊 |
There is another option. Take the network trained on KITTI or nuscences datasets and perform inference on the Lyft dataset :) | I was also curious to see how this approach performs. I tried the PointRCNN trained on Kitti for car detection and used it for inference. It got 0.02 :-D Perhaps finetuning it will give better results, but i don't have sufficient gpu resources to give it at try. At times like this, i miss unlimited kaggle gpus of the old days. |
Try using the python library: OpenCv 😉 | Thanks, I used cv2 to convert the mp4 to frames. |
Just like I wrote in another thread, my 5/5 folds solution scored lower than 3/5 folds solution on public LB so I expect shakeup. I can finish with gold I can finish with bronze. Or without medal? | Others are fighting, now I can't enter gold zone without 0.920 |
You seem t understand YOLO pretty well. One question. as I understand the last detection layer of YOLO outputs co-ordinates of one of the 5 predefined anchor boxes. If we need train the model with our own data, how can we prepare the training data to be annotated as per size of anchor boxes. | Andrew Ng explains intuition behind anchors. I found this very useful. https://www.youtube.com/watch?v=RTlwl2bv0Tg |
Kaggle is currently experiencing a bug. It looks like you can't add GPU to any kernel that wasn't previously committed with GPU before 22:00 UTC, October 21. To see if you have a GPU installed, run the following command: ! nvidia-smi | Yes, that would have worked better. Good suggestion. Kaggle has just fixed the bug. Everything is working correctly again. |
Thank you for the warning. I have resolved this problem, meaning that future commits and kernels will no longer encounter a problem. If you have any currently running commits or kernels, it is possible that they are still affected and would need to be restarted for the correct changes to take place. | Thank you for a quick resolution. My notebooks are working correctly now. |
That's not good. 😬 The kernels team is looking into this now. | Thank you for a quick resolution. My notebooks are working correctly now. |
Kaggle is currently experiencing a bug. It looks like you can't add GPU to any kernel that wasn't previously committed with GPU before 22:00 UTC, October 21. To see if you have a GPU installed, run the following command: ! nvidia-smi | , , you can also download the new notebook that had the error and upload it to the template of your old working GPU kernel. This way you can avoid the copy paste step mentioned by Chris above. |
That's not good. 😬 The kernels team is looking into this now. | Thank you for bringing this issue up. I believe we have just now resolved the GPU issue some people were hitting. If you run into this problem again today, please don't hesitate to let us know. |
Hi Jeremy. I'm studying this in depth right now, and I was wondering was the rationale is for keeping ALL the rows with a label, and then only HALF of the resulting number of rows without a label. Doesn't that result in oversampling the positive class when, in reality, this class is less represented than "not any" ? Are there any reason why we would want our model to believe that most of the time there is hemorrhage when most of the time there is not ? Also, I'm trying to read the output files. I'm using: def png16read(self:Path): return array(Image.open(self), dtype=np.uint16) as found here But then, I can't convert uint16 to a pytorch tensor. Any type of signed int is supported, but only uint8 is on the unsigned side... | Good questions Ben. I'll be explaining more in a future notebook, but in short, we're creating a prototyping dataset, which we'll be fine-tuning with the full dataset later. I want something <20GB so I can put on Kaggle datasets, and I definitely don't want to delete any data with positive labels! You can convert uint16 to tensors by converting to float32 arrays first. |