search

matroid / dlwithtf

TensorFlow for Deep Learning Book
http://shop.oreilly.com/product/0636920065869.do
321 stars 160 forks source link

Bug in ch3 example of linear regression #14

Open chobing97 opened 6 years ago

chobing97 commented 6 years ago

Hi. Thanks for this great deep learning examples!

There appears a bug in ch3 example linear_regression_tf.py At line 42, y becomes shape(100,) but I think it should be (100,1) because y-y_pred becomes (100,100) given y_pred is (100,1). So the loss function is very largely overestimated. Simply changing the 42 line as following makes the training converge to global minimum. (W=5, b=2)

y = tf.placeholder(tf.float32, (N,)) --> tf.placeholder(tf.float32, (N,1))

Of course, np.reshape should be removed from line 28 and some additional code change is necessary to make the script runnable.

So I don't think this example is a proper proof of gradient descent not converging to global minimum.

But I still deeply appreciate the great examples of tensorflow and it helps me with studying deep learning so much. Thanks,

rbharath commented 6 years ago

Yes, I believe you're correct (#9 points this out as well). This is a bug on our part that slipped through. We'll fix this in a future release for sure! Apologies for the error in the meanwhile.