Integrating Keras & TensorFlow: The Keras workflow, expanded (TensorFlow Dev Summit 2017)
ฝัง
- เผยแพร่เมื่อ 14 ก.พ. 2017
- Keras has the goal to make deep learning accessible to everyone, and it's one of the fastest growing machine learning frameworks. Join Francois Chollet, the primary author of Keras, as he demonstrates how Keras can be used in TensorFlow through a video QA example.
Visit the TensorFlow website for all session recordings: goo.gl/bsYmza
Subscribe to the Google Developers channel at goo.gl/mQyv5L
event: TensorFlow Dev Summit 2017; re_ty: Publish; - วิทยาศาสตร์และเทคโนโลยี
![มองหน้ากันไม่ติด (Awkward) - โอ๊ต ปราโมทย์ FEAT. MAIYARAP [Official MV]](http://i.ytimg.com/vi/4b02SY_r6Gw/mqdefault.jpg)
insane how terse keras can be describing something so complex
Great initiative!! I am an active Keras User and having it as part of TensorFlow will supercharge my work!! From DGX-1 and GPU Workstations to embedded platforms!! Keep up the great work!!
The volume is too low
gotta say, without surprisingly decent closed captioning i would not have been able to watch this vid past the first 2 minutes. glad i watched, it was very informative, so thanks for upload.
Thanks, Keras really impressive objective, bring deep learning techniques to mere mortals.
Good talk Francois, thanks for sharing!
anyone know how old fchollet is?
This is excellent Francois... so that I can try and duplicate this model myself is there a good public dataset available?
You can buy his book. I did. It's called Deep Learning with Python. It's published by MEAP. Available on Kindle and print.
Why is keras performance slow according to tensorflow low API
I wonder why he does not use sparse_softmax_cross_entropy_with_logits since it has only one valid output vs 1000
Do I need to pip install keras separately? Or can I just pip install tensorflow and use tf.keras?
Tf.keras
12:26 : tf.layers.LSTM cant be found in the official document.
GREAT INITIATIVE 🔝👏THANKS FOR SHARING
How to use tf.training.Experiment?
Any working code, where model definition is by Keras, but trained using "tf.training.Experiment" instead of "fit" ?
Same question. Am using TF 1.3. Have tried with tf.contrib.learn.Experiment but get "ValueError: `estimator` must implement `tf.contrib.learn.Evaluable` or `tf.estimator.Estimator`."
Here waiting for tf.contrib.keras by TF 1.1(planned in mid-March), now it's the end of March, TF1.1 is not coming yet... T_T
Good Luck Francois
Hey I got a question! I hope someone or Google Developers themselves answer me hehehe
I searched a lot about Keras and Tensorflow . I thought Tensorflow is for creating deep learning models. Where in fact I've come out that it was Keras who was doing that job.
Based on my research a lot of articles from various Deep learning and even machine learning sites tell that Tensorflow is the backend of Keras. MOST USED BACKEND OF KERAS AND POPULAR!
So am I right about what am i thinking? that Keras is for building or creating deep learning models and tensorflow is the one who is doing those various mathematical computations needed for that model. Is that it? cause I think it is but i would love to hear or learn other people's thoughts about it . Thank you so much
Working code using Keras 2.0.4
import keras
video = keras.layers.Input(shape=(None, 150, 150, 3))
cnn = keras.applications.InceptionV3(weights='imagenet',
include_top=False,
pooling='avg')
cnn.trainable = False
encoded_frames = keras.layers.wrappers.TimeDistributed(cnn)(video)
encoded_vid = keras.layers.LSTM(256)(encoded_frames)
question = keras.layers.Input(shape=[100], dtype='int32')
x = keras.layers.Embedding(10000, 256, mask_zero=True)(question)
encoded_q = keras.layers.LSTM(128)(x)
x = keras.layers.concatenate([encoded_vid, encoded_q])
x = keras.layers.Dense(128, activation=keras.activations.relu)(x)
outputs = keras.layers.Dense(1000)(x)
model = keras.models.Model([video, question], outputs)
model.compile(optimizer=keras.optimizers.Adam(),
loss=keras.losses.categorical_crossentropy)
i managed to run it but i got stuck with trying to load certain video & question and predict them /=
do you maybe have a complete code you can share? including the prediction?
tnx a lot
Cool
fantastic
OUI!
Please share the source code with sample data on Github!
They put a version of this model on the Keras docs. No sample data though. keras.io/getting-started/functional-api-guide/#visual-question-answering-model
Very generous on cloud computing... but why not bring it to Matlab so that students can use it.
I used do all my work in Matlab once upon a time, but in recent years Python is simply much better. Also, Matlab is proprietary and trying to bring such functionality is probably not worth the effort.
I fixed some problems (See github.com/hcchengithub/Toy-Video-QA-problem-on-Keras) with the code shown on the video but now I still don't know how to play the Q&A yet with the model compiled ... and then what?
cool
It would be nice if non-native English speakers would speak more slowly so as not to slur their phrases thereby making them undecipherable to those not familiar with the jargon of the topic. Software looks good though.
You should be thankful they are speaking english.
OK, some native English speakers do also slur their words, but most English speakers don't speak so quickly (since English is not spoken that way), so it is less severe without the accent of foreign speakers. Basically all speakers should enunciate clearly.
It was super clear to me. You should maybe learn to decipher non-native English accents instead of complaining
I wasn't trying to complain, but to put attention on speaking slowly, clearly and enunciating.
I understood pretty much everything! I'm french and I'm used to crappy french accent (including my own ^^)
Nice, but he should speak slower and stop walking. It looks like we are watching a tennis match