{ "id": 41335, "title": "[Python] Keras or Tensorflow? ", "url": "/competitions/dog-breed-identification/discussion/41335", "author_name": "Carmen J Su", "post_date": "2017-10-16T16:09:41.421000", "votes": 7, "comment_count": 21, "views": 0, "content": "

Hi, \nI just wanted to set up this little post here to find out which library people are using for this competition and why. \nStarting from me. I am planning to use Tensorflow image classifiers myself, as it seems quite robust and I want to learn it.
\nPlease enlighten me with your thought.

", "messages": [ { "id": 232814, "postDate": "2017-10-18T14:19:08.640Z", "content": "

Life is short, I chose Keras.

", "rawMarkdown": "Life is short, I chose Keras.", "votes": 9, "replies": [ { "id": 232844, "postDate": "2017-10-18T15:22:55.043Z", "content": "

I agree. Although time does seem to crawl when I'm coding in pure TensorFlow...

", "rawMarkdown": "I agree. Although time does seem to crawl when I'm coding in pure TensorFlow...", "votes": 1 } ] }, { "id": 232016, "postDate": "2017-10-16T16:09:41.423Z", "content": "

Hi, \nI just wanted to set up this little post here to find out which library people are using for this competition and why. \nStarting from me. I am planning to use Tensorflow image classifiers myself, as it seems quite robust and I want to learn it.
\nPlease enlighten me with your thought.

", "rawMarkdown": "Hi, \nI just wanted to set up this little post here to find out which library people are using for this competition and why. \nStarting from me. I am planning to use Tensorflow image classifiers myself, as it seems quite robust and I want to learn it. \nPlease enlighten me with your thought. ", "votes": 7 }, { "id": 255999, "postDate": "2017-12-10T22:06:13.377Z", "content": "

keras with the backend of tensorflow, \ni suggest you the Keras, it is really great. for fast implementation.

", "rawMarkdown": "keras with the backend of tensorflow, \ni suggest you the Keras, it is really great. for fast implementation.", "votes": 3 }, { "id": 232056, "postDate": "2017-10-16T17:56:30.867Z", "content": "

I have tested CNTK and caffe2. Problem with cntk and caffe2 is availability of pretrained models but they are both faster and more memory efficient than TF. I usually start with Keras 2-tensorflow for fast prototyping, then tensorlayer. Base model is InceptionResNetV2, method: transfer learning on imagenet with finetuning, custom: 4 layer dense top network and custom image augmentation. If you are new at this, Keras is very simple for fast prototyping, you can also check tensorlayer for more advanced (but still manageable) development.

", "rawMarkdown": "I have tested CNTK and caffe2. Problem with cntk and caffe2 is availability of pretrained models but they are both faster and more memory efficient than TF. I usually start with Keras 2-tensorflow for fast prototyping, then tensorlayer. Base model is InceptionResNetV2, method: transfer learning on imagenet with finetuning, custom: 4 layer dense top network and custom image augmentation. If you are new at this, Keras is very simple for fast prototyping, you can also check tensorlayer for more advanced (but still manageable) development.", "votes": 3 }, { "id": 232036, "postDate": "2017-10-16T17:12:36.007Z", "content": "

Keras 2.x with TensorFlow backend

", "rawMarkdown": "Keras 2.x with TensorFlow backend", "votes": 3, "replies": [ { "id": 237199, "postDate": "2017-10-29T19:53:04.740Z", "content": "

Me too :)

", "rawMarkdown": "Me too :)" } ] }, { "id": 268683, "postDate": "2018-01-15T07:43:27.143Z", "content": "", "rawMarkdown": "* [TensorLayer: Deep Learning and Reinforcement Learning Library for Researchers and Engineers.](https://github.com/tensorlayer/tensorlayer) or Keras, tflearn\n* pytorch\n* pure tensorflow is hard for researcher", "votes": 1 }, { "id": 238241, "postDate": "2017-10-31T23:30:40.777Z", "content": "

Keras and PyTorch. I find debugging in PyTorch to be much easier, but fine-tuning pretrained Keras networks is easier than in PyTorch since Keras lets you remove and add layers trivially.

", "rawMarkdown": "Keras and PyTorch. I find debugging in PyTorch to be much easier, but fine-tuning pretrained Keras networks is easier than in PyTorch since Keras lets you remove and add layers trivially. ", "votes": 1 }, { "id": 233750, "postDate": "2017-10-21T02:59:44.403Z", "content": "

I'm using Keras + Theano for now. I only recently started entering Kaggle competitions as I started going through the fast.ai Deep Learning Part 1 course, which uses Keras + Theano. I believe the nest course will use Karas + TensorFlow and they are starting to move both courses to Pytorch.

", "rawMarkdown": "I'm using Keras + Theano for now. I only recently started entering Kaggle competitions as I started going through the fast.ai Deep Learning Part 1 course, which uses Keras + Theano. I believe the nest course will use Karas + TensorFlow and they are starting to move both courses to Pytorch.", "votes": 1, "replies": [ { "id": 237316, "postDate": "2017-10-30T03:44:57.903Z", "content": "

Today I updated Keras (1.2.2 -> 2.0.8) and switched the backend to TensorFlow. I was using the older Keras since that's what fast.ai part 1 used, but I upgraded so I could run more Kernels locally. I made the switch to TF to have access to the Xception model.

", "rawMarkdown": "Today I updated Keras (1.2.2 -> 2.0.8) and switched the backend to TensorFlow. I was using the older Keras since that's what fast.ai part 1 used, but I upgraded so I could run more Kernels locally. I made the switch to TF to have access to the Xception model." } ] }, { "id": 233103, "postDate": "2017-10-19T08:07:07.123Z", "content": "

Thanks for all your insights, I would challenge myself to keras with tensorlow backend for this exercise, wish myself luck XD

\n\n

I went to a tensorflow event yesterday, One of the speakers use pure tensorflow without keras on top to do a product demand prediction at Argos. He put on his code on the screen, and I was like O.o

", "rawMarkdown": "Thanks for all your insights, I would challenge myself to keras with tensorlow backend for this exercise, wish myself luck XD\n\nI went to a tensorflow event yesterday, One of the speakers use pure tensorflow without keras on top to do a product demand prediction at Argos. He put on his code on the screen, and I was like O.o\n", "votes": 1 }, { "id": 232621, "postDate": "2017-10-17T22:52:36.967Z", "content": "

I've had a lot of luck with Keras + TensorFlow. I used to use TensorFlow by itself, but I don't recommend that unless you are a sucker for punishment, heh.

", "rawMarkdown": "I've had a lot of luck with Keras + TensorFlow. I used to use TensorFlow by itself, but I don't recommend that unless you are a sucker for punishment, heh.", "votes": 1 }, { "id": 232866, "postDate": "2017-10-18T16:25:09.010Z", "content": "

Both of them are developed by folks in Google. TensorFlow came out first and it quickly became one of dominant deep learning choices. Before this, most frameworks (MXNet, Caffe) were developed by university with relative slim support. However, people realized TensorFlow is still a relatively bottom level package, where you can twist code and have lots of control over things but not very friendly for folks without much advanced knowledge with deep learning. That's why Keras was built out, which can wrap TensorFlow in the backend. I guess given its well adoption rate by now, it accomplishes its purpose.

", "rawMarkdown": "Both of them are developed by folks in Google. TensorFlow came out first and it quickly became one of dominant deep learning choices. Before this, most frameworks (MXNet, Caffe) were developed by university with relative slim support. However, people realized TensorFlow is still a relatively bottom level package, where you can twist code and have lots of control over things but not very friendly for folks without much advanced knowledge with deep learning. That's why Keras was built out, which can wrap TensorFlow in the backend. I guess given its well adoption rate by now, it accomplishes its purpose.", "votes": 2 }, { "id": 232599, "postDate": "2017-10-17T20:18:05.940Z", "content": "

Keras with TF backend. Keras is easier and faster to build up a model for this kind of problem. While tensorflow gives more flexibility so that you can build more complicate models.

", "rawMarkdown": "Keras with TF backend. Keras is easier and faster to build up a model for this kind of problem. While tensorflow gives more flexibility so that you can build more complicate models.", "votes": 2 }, { "id": 276961, "postDate": "2018-02-01T22:33:22.690Z", "content": "

The new fast.ai course now recommends pytorch: http://www.fast.ai/2018/01/26/v2-launch/

", "rawMarkdown": "The new fast.ai course now recommends pytorch: http://www.fast.ai/2018/01/26/v2-launch/ " }, { "id": 262258, "postDate": "2017-12-26T01:37:56.717Z", "content": "

I'm using Keras + Tensorflow backend now. It is really convenient and it can satisfy most of my requirements.

", "rawMarkdown": "I'm using Keras + Tensorflow backend now. It is really convenient and it can satisfy most of my requirements. " }, { "id": 252553, "postDate": "2017-12-03T08:14:41.440Z", "content": "

Pytorch maybe a good choice.

", "rawMarkdown": "Pytorch maybe a good choice." }, { "id": 241568, "postDate": "2017-11-09T08:29:01.313Z", "content": "

I recommend pytorch. As for keras and tensorflow, in this test I guess the differences between them are not that important. For faster speed I prefer tensorflow.

", "rawMarkdown": "I recommend pytorch. As for keras and tensorflow, in this test I guess the differences between them are not that important. For faster speed I prefer tensorflow." }, { "id": 240651, "postDate": "2017-11-07T04:48:56.963Z", "content": "

fastai

", "rawMarkdown": "fastai" }, { "id": 239116, "postDate": "2017-11-02T18:10:22.030Z", "content": "

I am using keras with a tensorflow backend. Depending on how smoothly that goes I may play with pytorch to give dynamic neural nets a shot.

", "rawMarkdown": "I am using keras with a tensorflow backend. Depending on how smoothly that goes I may play with pytorch to give dynamic neural nets a shot." }, { "id": 233061, "postDate": "2017-10-19T04:39:40.757Z", "rawMarkdown": "", "votes": 1, "isDeleted": true } ], "comments": [ { "id": 232814, "author_name": "杨培文 (Yang Peiwen)", "author_url": "", "post_date": "2017-10-18T14:19:08.640000", "content": "

Life is short, I chose Keras.

", "votes": 9, "replies": [ { "id": 232844, "author_name": "Jason King", "author_url": "", "post_date": "2017-10-18T15:22:55.043000", "content": "

I agree. Although time does seem to crawl when I'm coding in pure TensorFlow...

", "votes": 1, "replies": [] } ] }, { "id": 255999, "author_name": "Aydin Ayanzadeh", "author_url": "", "post_date": "2017-12-10T22:06:13.377000", "content": "

keras with the backend of tensorflow, \ni suggest you the Keras, it is really great. for fast implementation.

", "votes": 3, "replies": [] }, { "id": 232056, "author_name": "Bostjan Mrak", "author_url": "", "post_date": "2017-10-16T17:56:30.867000", "content": "

I have tested CNTK and caffe2. Problem with cntk and caffe2 is availability of pretrained models but they are both faster and more memory efficient than TF. I usually start with Keras 2-tensorflow for fast prototyping, then tensorlayer. Base model is InceptionResNetV2, method: transfer learning on imagenet with finetuning, custom: 4 layer dense top network and custom image augmentation. If you are new at this, Keras is very simple for fast prototyping, you can also check tensorlayer for more advanced (but still manageable) development.

", "votes": 3, "replies": [] }, { "id": 232036, "author_name": "SM", "author_url": "", "post_date": "2017-10-16T17:12:36.007000", "content": "

Keras 2.x with TensorFlow backend

", "votes": 3, "replies": [ { "id": 237199, "author_name": "amsr", "author_url": "", "post_date": "2017-10-29T19:53:04.740000", "content": "

Me too :)

", "votes": 0, "replies": [] } ] }, { "id": 268683, "author_name": "云不懂风吹", "author_url": "", "post_date": "2018-01-15T07:43:27.143000", "content": "", "votes": 1, "replies": [] }, { "id": 238241, "author_name": "Rohan", "author_url": "", "post_date": "2017-10-31T23:30:40.777000", "content": "

Keras and PyTorch. I find debugging in PyTorch to be much easier, but fine-tuning pretrained Keras networks is easier than in PyTorch since Keras lets you remove and add layers trivially.

", "votes": 1, "replies": [] }, { "id": 233750, "author_name": "Ed Miller", "author_url": "", "post_date": "2017-10-21T02:59:44.403000", "content": "

I'm using Keras + Theano for now. I only recently started entering Kaggle competitions as I started going through the fast.ai Deep Learning Part 1 course, which uses Keras + Theano. I believe the nest course will use Karas + TensorFlow and they are starting to move both courses to Pytorch.

", "votes": 1, "replies": [ { "id": 237316, "author_name": "Ed Miller", "author_url": "", "post_date": "2017-10-30T03:44:57.903000", "content": "

Today I updated Keras (1.2.2 -> 2.0.8) and switched the backend to TensorFlow. I was using the older Keras since that's what fast.ai part 1 used, but I upgraded so I could run more Kernels locally. I made the switch to TF to have access to the Xception model.

", "votes": 0, "replies": [] } ] }, { "id": 233103, "author_name": "Carmen J Su", "author_url": "", "post_date": "2017-10-19T08:07:07.123000", "content": "

Thanks for all your insights, I would challenge myself to keras with tensorlow backend for this exercise, wish myself luck XD

\n\n

I went to a tensorflow event yesterday, One of the speakers use pure tensorflow without keras on top to do a product demand prediction at Argos. He put on his code on the screen, and I was like O.o

", "votes": 1, "replies": [] }, { "id": 232621, "author_name": "Jason King", "author_url": "", "post_date": "2017-10-17T22:52:36.967000", "content": "

I've had a lot of luck with Keras + TensorFlow. I used to use TensorFlow by itself, but I don't recommend that unless you are a sucker for punishment, heh.

", "votes": 1, "replies": [] }, { "id": 232866, "author_name": "DSEverything", "author_url": "", "post_date": "2017-10-18T16:25:09.010000", "content": "

Both of them are developed by folks in Google. TensorFlow came out first and it quickly became one of dominant deep learning choices. Before this, most frameworks (MXNet, Caffe) were developed by university with relative slim support. However, people realized TensorFlow is still a relatively bottom level package, where you can twist code and have lots of control over things but not very friendly for folks without much advanced knowledge with deep learning. That's why Keras was built out, which can wrap TensorFlow in the backend. I guess given its well adoption rate by now, it accomplishes its purpose.

", "votes": 2, "replies": [] }, { "id": 232599, "author_name": "JiaqiZhu", "author_url": "", "post_date": "2017-10-17T20:18:05.940000", "content": "

Keras with TF backend. Keras is easier and faster to build up a model for this kind of problem. While tensorflow gives more flexibility so that you can build more complicate models.

", "votes": 2, "replies": [] }, { "id": 276961, "author_name": "Emily Lawton", "author_url": "", "post_date": "2018-02-01T22:33:22.690000", "content": "

The new fast.ai course now recommends pytorch: http://www.fast.ai/2018/01/26/v2-launch/

", "votes": 0, "replies": [] }, { "id": 262258, "author_name": "Xiao Wang", "author_url": "", "post_date": "2017-12-26T01:37:56.717000", "content": "

I'm using Keras + Tensorflow backend now. It is really convenient and it can satisfy most of my requirements.

", "votes": 0, "replies": [] }, { "id": 252553, "author_name": "Mr.Zhao", "author_url": "", "post_date": "2017-12-03T08:14:41.440000", "content": "

Pytorch maybe a good choice.

", "votes": 0, "replies": [] }, { "id": 241568, "author_name": "SystemErrorWang", "author_url": "", "post_date": "2017-11-09T08:29:01.313000", "content": "

I recommend pytorch. As for keras and tensorflow, in this test I guess the differences between them are not that important. For faster speed I prefer tensorflow.

", "votes": 0, "replies": [] }, { "id": 240651, "author_name": "Meliodas", "author_url": "", "post_date": "2017-11-07T04:48:56.963000", "content": "

fastai

", "votes": 0, "replies": [] }, { "id": 239116, "author_name": "TylerCosner", "author_url": "", "post_date": "2017-11-02T18:10:22.030000", "content": "

I am using keras with a tensorflow backend. Depending on how smoothly that goes I may play with pytorch to give dynamic neural nets a shot.

", "votes": 0, "replies": [] }, { "id": 233061, "author_name": "", "author_url": "", "post_date": "2017-10-19T04:39:40.757000", "content": "", "votes": 1, "replies": [] } ], "raw_markdown_by_id": { "232814": "Life is short, I chose Keras.", "232016": "Hi, \nI just wanted to set up this little post here to find out which library people are using for this competition and why. \nStarting from me. I am planning to use Tensorflow image classifiers myself, as it seems quite robust and I want to learn it. \nPlease enlighten me with your thought. ", "255999": "keras with the backend of tensorflow, \ni suggest you the Keras, it is really great. for fast implementation.", "232056": "I have tested CNTK and caffe2. Problem with cntk and caffe2 is availability of pretrained models but they are both faster and more memory efficient than TF. I usually start with Keras 2-tensorflow for fast prototyping, then tensorlayer. Base model is InceptionResNetV2, method: transfer learning on imagenet with finetuning, custom: 4 layer dense top network and custom image augmentation. If you are new at this, Keras is very simple for fast prototyping, you can also check tensorlayer for more advanced (but still manageable) development.", "232036": "Keras 2.x with TensorFlow backend", "268683": "* [TensorLayer: Deep Learning and Reinforcement Learning Library for Researchers and Engineers.](https://github.com/tensorlayer/tensorlayer) or Keras, tflearn\n* pytorch\n* pure tensorflow is hard for researcher", "238241": "Keras and PyTorch. I find debugging in PyTorch to be much easier, but fine-tuning pretrained Keras networks is easier than in PyTorch since Keras lets you remove and add layers trivially. ", "233750": "I'm using Keras + Theano for now. I only recently started entering Kaggle competitions as I started going through the fast.ai Deep Learning Part 1 course, which uses Keras + Theano. I believe the nest course will use Karas + TensorFlow and they are starting to move both courses to Pytorch.", "233103": "Thanks for all your insights, I would challenge myself to keras with tensorlow backend for this exercise, wish myself luck XD\n\nI went to a tensorflow event yesterday, One of the speakers use pure tensorflow without keras on top to do a product demand prediction at Argos. He put on his code on the screen, and I was like O.o\n", "232621": "I've had a lot of luck with Keras + TensorFlow. I used to use TensorFlow by itself, but I don't recommend that unless you are a sucker for punishment, heh.", "232866": "Both of them are developed by folks in Google. TensorFlow came out first and it quickly became one of dominant deep learning choices. Before this, most frameworks (MXNet, Caffe) were developed by university with relative slim support. However, people realized TensorFlow is still a relatively bottom level package, where you can twist code and have lots of control over things but not very friendly for folks without much advanced knowledge with deep learning. That's why Keras was built out, which can wrap TensorFlow in the backend. I guess given its well adoption rate by now, it accomplishes its purpose.", "232599": "Keras with TF backend. Keras is easier and faster to build up a model for this kind of problem. While tensorflow gives more flexibility so that you can build more complicate models.", "276961": "The new fast.ai course now recommends pytorch: http://www.fast.ai/2018/01/26/v2-launch/ ", "262258": "I'm using Keras + Tensorflow backend now. It is really convenient and it can satisfy most of my requirements. ", "252553": "Pytorch maybe a good choice.", "241568": "I recommend pytorch. As for keras and tensorflow, in this test I guess the differences between them are not that important. For faster speed I prefer tensorflow.", "240651": "fastai", "239116": "I am using keras with a tensorflow backend. Depending on how smoothly that goes I may play with pytorch to give dynamic neural nets a shot.", "233061": "" } }