Code takes very long to compile on jupyter

[cgarciae]

This code takes very long to compile on Jupyter, I am running version 1.2.1 and also tested on coconut-develop. Since all dependencies are used inside methods you dont need the dependecies installed to test the compilation of this code on Jupyter.

class Model(object):
    
    def define_model(self, n_actions, n_states, y=0.98):
        
        with self.graph.as_default(), tf.device("cpu:0"):
            
            s = tf.placeholder(tf.int32, [None], name='s')
            a = tf.placeholder(tf.int32, [None], name='a')
            r = tf.placeholder(tf.float32, [None], name='r')
            max_Qs1 = tf.placeholder(tf.float32, [None], name='maxQs1')
            lr = tf.placeholder(tf.float32, [], name='lr')

            ops = dict(
                trainable=True, 
                kernel_initializer=tf.random_uniform_initializer(minval=0.0, maxval=0.01), 
                bias_initializer=None) #tf.random_uniform_initializer(minval=0, maxval=0.01))

            Qs = (
                s
                |> tf.one_hot$(?, n_states)
                |> tf.layers.dense$(?, n_actions, name='linear_layer', **ops)
            )

            Qsa = select_columns(Qs, a)

            max_Qs = tf.reduce_max(Qs, 1)

            error = r + y * max_Qs1 - Qsa
            loss = Pipe(error, tf.nn.l2_loss, tf.reduce_sum)
            update = tf.train.GradientDescentOptimizer(lr).minimize(loss)

        self.s = s; self.a = a; self.r = r; self.max_Qs1 = max_Qs1
        self.max_Qs = max_Qs; self.Qs = Qs; self.Qsa = Qsa; self.update = update
        self.lr = lr
                
    def choose_action(self, state, e=0.05):
        actions = self.sess.run(self.Qs, feed_dict={self.s: [state]})[0]
        n = len(actions)

        if random.random() < e:
            return random.randint(0, n-1)
        else:
            return np.argmax(actions)
        
    def get_trainer(self):
        return 1
    
    def reset(self):
        pass

    @staticmethod
    def learning_rate(t, b, k):
        return b * k / (k + t)
    
print(0)

[evhub]

@cgarciae It does seem like something is slightly off with compilation under Jupyter. The whole thing takes 1.5 seconds to compile in the standard interpreter, but almost a minute on Jupyter. I'll look into it and potentially try to put out a quick v1.2.2 release with a fix.

[evhub]

@cgarciae I've uploaded a coconut-develop release that should hopefully make the kernel a lot faster. Let me know if it helps.

[cgarciae]

Did a pip install coconut-develop --upgrade but its still taking more than 30s.

[evhub]

@cgarciae Try again now. I think I got it this time!

[evhub]

As far as I can tell, I believe this has been fixed. Marking as resolved and closing the issue. @cgarciae lmk if you're still experiencing any problems now on the latest coconut-develop (which will very soon be released).

[cgarciae]

@evhub Thanks a lot! It did reduce the time, it when from 40+ seconds to 20 seconds. However, the normal python interpreter takes around 1-2 seconds, I'd say coconut should be good around 2-5 seconds given the extra effort; the coconut compiler is really takes around 2 seconds on the same file.

I can't yet use coconut like this in Jupyter, hope this gets fixed soon :)

[evhub]

@cgarciae I'm going to put out v1.2.2 as is mostly to fix #219, but after that I'll try to keep working on Jupyter performance in develop.

[evhub]

@cgarciae As a temporary measure, if you delete the blank lines in your code (Jupyter thinks they might be where your code ends, and therefore tries to compile up to that point), you should be fine.

[evhub]

@cgarciae I just made a bunch of experimental modifications to develop which seem to have resolved the rest of the problem for me. Try it out with pip install coconut-develop and let me know if it's working for you now as well!

[cgarciae]

@evhub Excellent! Now taking around 5 secconds as expected :) I am going to close the issue.