Hello Everyone,\nI am very new to neural networks/ AI. My experience in this field is limited to completing a couple of initial modules of the fast.ai course (Jeremy Howard). I have a background in implementing statistical techniques using R.
\n\nMy goals: \nBuild a basic implementation from scratch that can yield me an accuracy of at-least 75%, more importantly - I want to make this work on RPi on a real-time basis (I use RPi for my hobby projects).
\n\nI have seen that a few people have already posted reference implementations, however I am keen on building my implementation from ground up.
\n\nSpecific Questions from the group:
\n\nThanks a lot for your time, and your help. I'll keep you updated on my progress.\nRama
", "messages": [ { "id": "246617", "postDate": "11/21/2017 12:47:30", "content": "Hello Everyone,\nI am very new to neural networks/ AI. My experience in this field is limited to completing a couple of initial modules of the fast.ai course (Jeremy Howard). I have a background in implementing statistical techniques using R.
\n\nMy goals: \nBuild a basic implementation from scratch that can yield me an accuracy of at-least 75%, more importantly - I want to make this work on RPi on a real-time basis (I use RPi for my hobby projects).
\n\nI have seen that a few people have already posted reference implementations, however I am keen on building my implementation from ground up.
\n\nSpecific Questions from the group:
\n\nThanks a lot for your time, and your help. I'll keep you updated on my progress.\nRama
", "rawMarkdown": "Hello Everyone,\nI am very new to neural networks/ AI. My experience in this field is limited to completing a couple of initial modules of the fast.ai course (Jeremy Howard). I have a background in implementing statistical techniques using R.\n\nMy goals: \nBuild a basic implementation from scratch that can yield me an accuracy of at-least 75%, more importantly - I want to make this work on RPi on a real-time basis (I use RPi for my hobby projects). \n\nI have seen that a few people have already posted reference implementations, however I am keen on building my implementation from ground up.\n\nSpecific Questions from the group:\n\n 1. I have read/reading posts by the organizers on the TensorFlow and on the dataset\n 2. I need help to understand a bit of the underlying theory relevant to solving this problem. Can someone point me to a textbook chapter(s), a youtube tutorial or an online course that can help me build a basic implementation\n 3. I had a question pertaining to the environment where people typically train these models. I dont have a GPU, however, I do have an AWS P2 instance set-up - should I use that, or do you have any other suggestions?\n\nThanks a lot for your time, and your help. I'll keep you updated on my progress.\nRama", "votes": null }, { "id": "246646", "postDate": "11/21/2017 14:08:14", "content": "Welcome, thanks for joining in the contest! Just to check, did you see the tutorial here?
\n\nhttps://www.tensorflow.org/versions/master/tutorials/audio_recognition
\n\nI'm hoping this should be a good way to get started, and it will give you a model you can run on a Pi. It will run without a GPU, though the training process will take longer.
", "rawMarkdown": "Welcome, thanks for joining in the contest! Just to check, did you see the tutorial here?\n\nhttps://www.tensorflow.org/versions/master/tutorials/audio_recognition\n\nI'm hoping this should be a good way to get started, and it will give you a model you can run on a Pi. It will run without a GPU, though the training process will take longer.", "votes": null }, { "id": "251311", "postDate": "12/01/2017 01:02:18", "content": "I ran train.py successfully on a Macbook Pro (i7 Version 7, 16GB, no GPU) - it took about 16 hours.
\n\nINFO:tensorflow:Step #18000: rate 0.000100, accuracy 91.0%, cross entropy 0.402003\nINFO:tensorflow:Confusion Matrix:\n [[258 0 0 0 0 0 0 0 0 0 0 0]\n [ 0 191 6 3 1 4 11 12 7 3 10 10]\n [ 3 3 248 2 0 0 2 2 0 0 0 1]\n [ 0 12 4 231 5 2 0 0 0 0 2 14]\n [ 3 6 0 0 243 0 0 0 0 3 4 1]\n [ 0 4 2 17 0 228 2 0 0 0 5 6]\n [ 1 3 10 2 0 0 229 2 0 0 0 0]\n [ 0 5 0 0 3 0 5 242 1 0 0 0]\n [ 4 4 0 0 3 2 2 0 239 2 0 1]\n [ 0 6 0 0 17 0 2 1 2 227 0 1]\n [ 2 3 0 0 13 1 2 0 0 1 224 0]\n [ 7 8 0 30 2 4 1 3 0 4 0 201]]\nINFO:tensorflow:Step 18000: Validation accuracy = 89.3% (N=3093)\nINFO:tensorflow:Saving to \"/tmp/speech_commands_train/conv.ckpt-18000\"\nINFO:tensorflow:set_size=3081\nINFO:tensorflow:Confusion Matrix:\n [[257 0 0 0 0 0 0 0 0 0 0 0]\n [ 1 195 4 2 3 8 5 11 7 2 5 14]\n [ 1 8 235 3 1 1 4 2 0 0 1 0]\n [ 1 3 1 216 1 10 1 3 1 0 2 13]\n [ 0 2 0 0 260 0 3 0 0 2 3 2]\n [ 2 4 0 16 2 216 3 0 0 0 0 10]\n [ 0 3 14 0 4 0 244 2 0 0 0 0]\n [ 1 9 0 0 5 0 1 241 0 2 0 0]\n [ 0 5 0 0 3 2 0 2 232 2 0 0]\n [ 1 5 0 0 13 1 2 1 7 228 4 0]\n [ 0 2 1 0 7 2 3 0 0 2 232 0]\n [ 0 10 0 31 4 5 3 2 0 0 3 193]]\nINFO:tensorflow:Final test accuracy = 89.2% (N=3081)\nBrandons-MacBook-Pro-2:speech_commands brandonpippin$
", "rawMarkdown": "I ran train.py successfully on a Macbook Pro (i7 Version 7, 16GB, no GPU) - it took about 16 hours.\n\nINFO:tensorflow:Step #18000: rate 0.000100, accuracy 91.0%, cross entropy 0.402003\nINFO:tensorflow:Confusion Matrix:\n [[258 0 0 0 0 0 0 0 0 0 0 0]\n [ 0 191 6 3 1 4 11 12 7 3 10 10]\n [ 3 3 248 2 0 0 2 2 0 0 0 1]\n [ 0 12 4 231 5 2 0 0 0 0 2 14]\n [ 3 6 0 0 243 0 0 0 0 3 4 1]\n [ 0 4 2 17 0 228 2 0 0 0 5 6]\n [ 1 3 10 2 0 0 229 2 0 0 0 0]\n [ 0 5 0 0 3 0 5 242 1 0 0 0]\n [ 4 4 0 0 3 2 2 0 239 2 0 1]\n [ 0 6 0 0 17 0 2 1 2 227 0 1]\n [ 2 3 0 0 13 1 2 0 0 1 224 0]\n [ 7 8 0 30 2 4 1 3 0 4 0 201]]\nINFO:tensorflow:Step 18000: Validation accuracy = 89.3% (N=3093)\nINFO:tensorflow:Saving to \"/tmp/speech_commands_train/conv.ckpt-18000\"\nINFO:tensorflow:set_size=3081\nINFO:tensorflow:Confusion Matrix:\n [[257 0 0 0 0 0 0 0 0 0 0 0]\n [ 1 195 4 2 3 8 5 11 7 2 5 14]\n [ 1 8 235 3 1 1 4 2 0 0 1 0]\n [ 1 3 1 216 1 10 1 3 1 0 2 13]\n [ 0 2 0 0 260 0 3 0 0 2 3 2]\n [ 2 4 0 16 2 216 3 0 0 0 0 10]\n [ 0 3 14 0 4 0 244 2 0 0 0 0]\n [ 1 9 0 0 5 0 1 241 0 2 0 0]\n [ 0 5 0 0 3 2 0 2 232 2 0 0]\n [ 1 5 0 0 13 1 2 1 7 228 4 0]\n [ 0 2 1 0 7 2 3 0 0 2 232 0]\n [ 0 10 0 31 4 5 3 2 0 0 3 193]]\nINFO:tensorflow:Final test accuracy = 89.2% (N=3081)\nBrandons-MacBook-Pro-2:speech_commands brandonpippin$", "votes": null }, { "id": "251313", "postDate": "12/01/2017 01:10:47", "content": "I'm using TensorFlow docker container. It works well under both Linux and Windows. I'm currently using CPU version and waiting for GCP credit so I can train faster.
", "rawMarkdown": "I'm using [TensorFlow docker container][1]. It works well under both Linux and Windows. I'm currently using CPU version and waiting for GCP credit so I can train faster.\n\n\n [1]: https://hub.docker.com/r/tensorflow/tensorflow/", "votes": null }, { "id": "251408", "postDate": "12/01/2017 06:01:23", "content": "I got an error when running tutorial code on the dataset download from kaggle. It's an encoding error. When I use the data downloaded from tutorial's script, it works. \nIs there any difference between the two dataset mentioned above?
", "rawMarkdown": "I got an error when running tutorial code on the dataset download from kaggle. It's an encoding error. When I use the data downloaded from tutorial's script, it works. \nIs there any difference between the two dataset mentioned above?", "votes": null }, { "id": "251754", "postDate": "12/01/2017 16:47:10", "content": "Is it valid to run the tutorial model on the kaggle competition data?
", "rawMarkdown": "Is it valid to run the tutorial model on the kaggle competition data?", "votes": null }, { "id": "251963", "postDate": "12/02/2017 01:01:48", "content": "No, it isn't.
", "rawMarkdown": "No, it isn't.", "votes": null } ], "comments": [ { "id": 246646, "author_name": "petewarden", "author_url": "", "post_date": "11/21/2017 14:08:14", "content": "Welcome, thanks for joining in the contest! Just to check, did you see the tutorial here?
\n\nhttps://www.tensorflow.org/versions/master/tutorials/audio_recognition
\n\nI'm hoping this should be a good way to get started, and it will give you a model you can run on a Pi. It will run without a GPU, though the training process will take longer.
", "votes": null, "replies": [ { "id": 251408, "author_name": "jihaoliu", "author_url": "", "post_date": "12/01/2017 06:01:23", "content": "I got an error when running tutorial code on the dataset download from kaggle. It's an encoding error. When I use the data downloaded from tutorial's script, it works. \nIs there any difference between the two dataset mentioned above?
", "votes": null, "replies": [] }, { "id": 251754, "author_name": "ksumlin", "author_url": "", "post_date": "12/01/2017 16:47:10", "content": "Is it valid to run the tutorial model on the kaggle competition data?
", "votes": null, "replies": [] }, { "id": 251963, "author_name": "jihaoliu", "author_url": "", "post_date": "12/02/2017 01:01:48", "content": "No, it isn't.
", "votes": null, "replies": [] } ] }, { "id": 251311, "author_name": "flagshipdynamics", "author_url": "", "post_date": "12/01/2017 01:02:18", "content": "I ran train.py successfully on a Macbook Pro (i7 Version 7, 16GB, no GPU) - it took about 16 hours.
\n\nINFO:tensorflow:Step #18000: rate 0.000100, accuracy 91.0%, cross entropy 0.402003\nINFO:tensorflow:Confusion Matrix:\n [[258 0 0 0 0 0 0 0 0 0 0 0]\n [ 0 191 6 3 1 4 11 12 7 3 10 10]\n [ 3 3 248 2 0 0 2 2 0 0 0 1]\n [ 0 12 4 231 5 2 0 0 0 0 2 14]\n [ 3 6 0 0 243 0 0 0 0 3 4 1]\n [ 0 4 2 17 0 228 2 0 0 0 5 6]\n [ 1 3 10 2 0 0 229 2 0 0 0 0]\n [ 0 5 0 0 3 0 5 242 1 0 0 0]\n [ 4 4 0 0 3 2 2 0 239 2 0 1]\n [ 0 6 0 0 17 0 2 1 2 227 0 1]\n [ 2 3 0 0 13 1 2 0 0 1 224 0]\n [ 7 8 0 30 2 4 1 3 0 4 0 201]]\nINFO:tensorflow:Step 18000: Validation accuracy = 89.3% (N=3093)\nINFO:tensorflow:Saving to \"/tmp/speech_commands_train/conv.ckpt-18000\"\nINFO:tensorflow:set_size=3081\nINFO:tensorflow:Confusion Matrix:\n [[257 0 0 0 0 0 0 0 0 0 0 0]\n [ 1 195 4 2 3 8 5 11 7 2 5 14]\n [ 1 8 235 3 1 1 4 2 0 0 1 0]\n [ 1 3 1 216 1 10 1 3 1 0 2 13]\n [ 0 2 0 0 260 0 3 0 0 2 3 2]\n [ 2 4 0 16 2 216 3 0 0 0 0 10]\n [ 0 3 14 0 4 0 244 2 0 0 0 0]\n [ 1 9 0 0 5 0 1 241 0 2 0 0]\n [ 0 5 0 0 3 2 0 2 232 2 0 0]\n [ 1 5 0 0 13 1 2 1 7 228 4 0]\n [ 0 2 1 0 7 2 3 0 0 2 232 0]\n [ 0 10 0 31 4 5 3 2 0 0 3 193]]\nINFO:tensorflow:Final test accuracy = 89.2% (N=3081)\nBrandons-MacBook-Pro-2:speech_commands brandonpippin$
", "votes": null, "replies": [] }, { "id": 251313, "author_name": "bsp2020", "author_url": "", "post_date": "12/01/2017 01:10:47", "content": "I'm using TensorFlow docker container. It works well under both Linux and Windows. I'm currently using CPU version and waiting for GCP credit so I can train faster.
", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "246617": "Hello Everyone,\nI am very new to neural networks/ AI. My experience in this field is limited to completing a couple of initial modules of the fast.ai course (Jeremy Howard). I have a background in implementing statistical techniques using R.\n\nMy goals: \nBuild a basic implementation from scratch that can yield me an accuracy of at-least 75%, more importantly - I want to make this work on RPi on a real-time basis (I use RPi for my hobby projects). \n\nI have seen that a few people have already posted reference implementations, however I am keen on building my implementation from ground up.\n\nSpecific Questions from the group:\n\n 1. I have read/reading posts by the organizers on the TensorFlow and on the dataset\n 2. I need help to understand a bit of the underlying theory relevant to solving this problem. Can someone point me to a textbook chapter(s), a youtube tutorial or an online course that can help me build a basic implementation\n 3. I had a question pertaining to the environment where people typically train these models. I dont have a GPU, however, I do have an AWS P2 instance set-up - should I use that, or do you have any other suggestions?\n\nThanks a lot for your time, and your help. I'll keep you updated on my progress.\nRama", "246646": "Welcome, thanks for joining in the contest! Just to check, did you see the tutorial here?\n\nhttps://www.tensorflow.org/versions/master/tutorials/audio_recognition\n\nI'm hoping this should be a good way to get started, and it will give you a model you can run on a Pi. It will run without a GPU, though the training process will take longer.", "251311": "I ran train.py successfully on a Macbook Pro (i7 Version 7, 16GB, no GPU) - it took about 16 hours.\n\nINFO:tensorflow:Step #18000: rate 0.000100, accuracy 91.0%, cross entropy 0.402003\nINFO:tensorflow:Confusion Matrix:\n [[258 0 0 0 0 0 0 0 0 0 0 0]\n [ 0 191 6 3 1 4 11 12 7 3 10 10]\n [ 3 3 248 2 0 0 2 2 0 0 0 1]\n [ 0 12 4 231 5 2 0 0 0 0 2 14]\n [ 3 6 0 0 243 0 0 0 0 3 4 1]\n [ 0 4 2 17 0 228 2 0 0 0 5 6]\n [ 1 3 10 2 0 0 229 2 0 0 0 0]\n [ 0 5 0 0 3 0 5 242 1 0 0 0]\n [ 4 4 0 0 3 2 2 0 239 2 0 1]\n [ 0 6 0 0 17 0 2 1 2 227 0 1]\n [ 2 3 0 0 13 1 2 0 0 1 224 0]\n [ 7 8 0 30 2 4 1 3 0 4 0 201]]\nINFO:tensorflow:Step 18000: Validation accuracy = 89.3% (N=3093)\nINFO:tensorflow:Saving to \"/tmp/speech_commands_train/conv.ckpt-18000\"\nINFO:tensorflow:set_size=3081\nINFO:tensorflow:Confusion Matrix:\n [[257 0 0 0 0 0 0 0 0 0 0 0]\n [ 1 195 4 2 3 8 5 11 7 2 5 14]\n [ 1 8 235 3 1 1 4 2 0 0 1 0]\n [ 1 3 1 216 1 10 1 3 1 0 2 13]\n [ 0 2 0 0 260 0 3 0 0 2 3 2]\n [ 2 4 0 16 2 216 3 0 0 0 0 10]\n [ 0 3 14 0 4 0 244 2 0 0 0 0]\n [ 1 9 0 0 5 0 1 241 0 2 0 0]\n [ 0 5 0 0 3 2 0 2 232 2 0 0]\n [ 1 5 0 0 13 1 2 1 7 228 4 0]\n [ 0 2 1 0 7 2 3 0 0 2 232 0]\n [ 0 10 0 31 4 5 3 2 0 0 3 193]]\nINFO:tensorflow:Final test accuracy = 89.2% (N=3081)\nBrandons-MacBook-Pro-2:speech_commands brandonpippin$", "251313": "I'm using [TensorFlow docker container][1]. It works well under both Linux and Windows. I'm currently using CPU version and waiting for GCP credit so I can train faster.\n\n\n [1]: https://hub.docker.com/r/tensorflow/tensorflow/", "251408": "I got an error when running tutorial code on the dataset download from kaggle. It's an encoding error. When I use the data downloaded from tutorial's script, it works. \nIs there any difference between the two dataset mentioned above?", "251754": "Is it valid to run the tutorial model on the kaggle competition data?", "251963": "No, it isn't." }, "source": "meta" }