![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F2294613%2Fac09d01c06c0e634f95454daaf99d94f%2Faldfly-0_aldfly-00-0313.jpg?generation=1598666499468588&alt=media\")
Hi, kaggler.
\nMany of the participants in this competition use Pytorch.
\nHowever, TPU is good for training a lot of data.
\nI'll propose to use a notebook with TPU.
birdcall_using_TPU_train
\nbirdcall_using_TPU_inference
Try training with TPU!
\nI will publish a dataset of images at the same time.
\nThe dataset contains more than 200K images such as the following
This image is a mel spectrogram of the pre-emphasis processed audio.
\nThe TPU notebooks train and infer this.
birdsongab
\nbirdsongcf
\nbirdsonggm
\nbirdsongnr
\nbirdsongsy
I hope I can contribute to your work.
\nI'm a newbie, so I may make a lot of mistakes.
\nor perhaps I should say, I feel like I'm making a big mistake.
\nThank you.
References:
\n[inference, PyTorch] Birdcall ResNet Baseline
\nGetting started with 100+ flowers on TPU
\nTriple Stratified KFold with TFRecords
\nResampled Train Audio on Kaggle Dataset
(8/29 8:28 added)
\nI'll publish preprocessing files which I executed.
\nPreprocessing Files
\nBrief Description
\npreprocess1.py
\nSilence removing and audio concatenation
\npreprocess2.py
\npre-emphasis and convert audio to image
\nJPGtoTFR.ipynb
\nConvert JPG to TFR
Hi, kaggler.
\nMany of the participants in this competition use Pytorch.
\nHowever, TPU is good for training a lot of data.
\nI'll propose to use a notebook with TPU.
birdcall_using_TPU_train
\nbirdcall_using_TPU_inference
Try training with TPU!
\nI will publish a dataset of images at the same time.
\nThe dataset contains more than 200K images such as the following
This image is a mel spectrogram of the pre-emphasis processed audio.
\nThe TPU notebooks train and infer this.
birdsongab
\nbirdsongcf
\nbirdsonggm
\nbirdsongnr
\nbirdsongsy
I hope I can contribute to your work.
\nI'm a newbie, so I may make a lot of mistakes.
\nor perhaps I should say, I feel like I'm making a big mistake.
\nThank you.
References:
\n[inference, PyTorch] Birdcall ResNet Baseline
\nGetting started with 100+ flowers on TPU
\nTriple Stratified KFold with TFRecords
\nResampled Train Audio on Kaggle Dataset
(8/29 8:28 added)
\nI'll publish preprocessing files which I executed.
\nPreprocessing Files
\nBrief Description
\npreprocess1.py
\nSilence removing and audio concatenation
\npreprocess2.py
\npre-emphasis and convert audio to image
\nJPGtoTFR.ipynb
\nConvert JPG to TFR
Thanks for the great post.
\nI have a question.
\nI'm talking about the process of creating a tfrecode.
\n1.preprocess1.py do
\n2.preprocess2.py do
\n3.JPGtoTFR.ipynb do
\nIs that correct?
\nAlso, where should I quote the speech_tools in preprocess1.py?
\nI tried to run preprocess1.py in the local environment, but I couldn't find speech_tools.
Please delete that line.
\nI forgot to delete it because I was diverting the code for speech processing.
\nclone again or import them
\n\nimport librosa\n
\n\nimport numpy as np\n
\n\nimport glob\n
\nThe order of execution of code is correct.
\nYou may need to rename the folder in these codes.
I understand!
\nThanks for the thoughtful answer!
", "rawMarkdown": "I understand!\n\nThanks for the thoughtful answer!", "votes": 1 } ] }, { "id": 989873, "postDate": "2020-08-29T07:00:33.653Z", "content": "hi, as per the rules, TPU is not allowed. Feel free to use for practice and training
\nbut submit a model using GPU or CPU in submission, else your notebook will be disqualified
@kmldas , I think the inference kernel is not TPU.
", "rawMarkdown": "@kmldas , I think the inference kernel is not TPU. ", "votes": 1 }, { "id": 989970, "postDate": "2020-08-29T08:42:04.107Z", "content": "@kmldas
\nsorry for confusing expression.
\nInference is executed with GPU.
thanks … as GPU and not TPU
\nit should be fine!
Thank you for sharing the code - I still have gone through it completely, but it looks good at first sight.
\nCan you share the code for sharing tfrecords please.
here.
\nPlease correct as appropriate.
\nThese may have defects.
\nyou can inprove it.
Thanks so much!! This will greatly reduce the training time and can experiment more within limited time!
", "rawMarkdown": "Thanks so much!! This will greatly reduce the training time and can experiment more within limited time!\n", "votes": 1 }, { "id": 993637, "postDate": "2020-09-01T04:56:22.620Z", "content": "The training phase runs for only 18-20 epochs but I gave around 30 epochs what could be the reason?
", "rawMarkdown": "The training phase runs for only 18-20 epochs but I gave around 30 epochs what could be the reason?", "replies": [ { "id": 993912, "postDate": "2020-09-01T08:16:09.740Z", "content": "Training is stopped by EarlyStopping.
\nI removed ES in my newest notebook.
\nIt seems many epochs can enhance score.
Thanks for the great post.
\nI have a question.
\nI'm talking about the process of creating a tfrecode.
\n1.preprocess1.py do
\n2.preprocess2.py do
\n3.JPGtoTFR.ipynb do
\nIs that correct?
\nAlso, where should I quote the speech_tools in preprocess1.py?
\nI tried to run preprocess1.py in the local environment, but I couldn't find speech_tools.
Please delete that line.
\nI forgot to delete it because I was diverting the code for speech processing.
\nclone again or import them
\n\nimport librosa\n
\n\nimport numpy as np\n
\n\nimport glob\n
\nThe order of execution of code is correct.
\nYou may need to rename the folder in these codes.
I understand!
\nThanks for the thoughtful answer!
", "votes": 1, "replies": [] } ] }, { "id": 989873, "author_name": "Kamal Das", "author_url": "", "post_date": "2020-08-29T07:00:33.653000", "content": "hi, as per the rules, TPU is not allowed. Feel free to use for practice and training
\nbut submit a model using GPU or CPU in submission, else your notebook will be disqualified
@kmldas , I think the inference kernel is not TPU.
", "votes": 1, "replies": [] }, { "id": 989970, "author_name": "ITK8191", "author_url": "", "post_date": "2020-08-29T08:42:04.107000", "content": "@kmldas
\nsorry for confusing expression.
\nInference is executed with GPU.
thanks … as GPU and not TPU
\nit should be fine!
Thank you for sharing the code - I still have gone through it completely, but it looks good at first sight.
\nCan you share the code for sharing tfrecords please.
here.
\nPlease correct as appropriate.
\nThese may have defects.
\nyou can inprove it.
Thanks so much!! This will greatly reduce the training time and can experiment more within limited time!
", "votes": 1, "replies": [] }, { "id": 993637, "author_name": "Convophile", "author_url": "", "post_date": "2020-09-01T04:56:22.620000", "content": "The training phase runs for only 18-20 epochs but I gave around 30 epochs what could be the reason?
", "votes": 0, "replies": [ { "id": 993912, "author_name": "ITK8191", "author_url": "", "post_date": "2020-09-01T08:16:09.740000", "content": "Training is stopped by EarlyStopping.
\nI removed ES in my newest notebook.
\nIt seems many epochs can enhance score.