failed precondition: error while reading resource variable block8_sepconv3_bn/gamma

[8rax]

Hello,

I have found very few documentation concerning the error : failed precondition: error while reading resource variable block8_sepconv3_bn/gamma from Container: localhost. This could mean that the variable was uninitialized. Not found : Resource localhost/block8_speconv3_bn/gamma/class tensorflow::Var does not exist

I guess the problem must be in the train function before the model.predict and has something to do with the sessions and threads but I would be happy to get any help I can get.

I use :
Python 3.7.8
Cuda 10.1
CUDNN v7.6.5
Tensorflow 2.3.0
keras 2.4.3

You can find my code in attachment or here : https://github.com/8rax/Carla/blob/master/JDO_Tutorial_5.py
Carla_Env_DQN.zip

[bhack]

Can you check #28287 (comment)

[8rax]

Hello, thank you very much for the suggestion. I indeed had a look at that thread as well and tried the solution. But I might have done it wrong. My understanding about sessions, graphs and threads might be a little bit lackluster as well, apologies for that.

I created on line 70 my gpu_options and my sess at the begining of my script after the regular imports:

gpu_options=tf.compat.v1.GPUOptions(per_process_gpu_memory_fraction=MEMORY_FRACTION) #TENSORFLOW DON'T USE ALL MY GPU PLZ
sess=set_session(tf.compat.v1.Session(config=tf.compat.v1.ConfigProto(gpu_options=gpu_options)))

I understand I should also declare here the graph globally instead of declaring it in the init function of my DQNAgent and should therefore also put :

graph = tf.compat.v1.get_default_graph()

then, if my understanding of the problem is correct, I should then set the session every time I am calling the .predict method in the train function of my DQNAgent class.

I suppose that I should then use the following syntax:

global sess
global graph
with self.graph.as_default():
  set_session(sess)
  current_qs_list = self.model.predict(current_states, PREDICTION_BATCH_SIZE)
  new_current_states = np.array([transition[3] for transition in minibatch])/255 

global sess
global graph
with self.graph.as_default():
  set_session(sess)
  future_qs_list = self.target_model.predict(new_current_states, PREDICTION_BATCH_SIZE)

Not sure I should add it here though as there is no graph involved :

def get_qs(self, state):
        return self.model.predict(np.array(state).reshape(-1, * state.shape)/255)[0]

would that make sense?

[bhack]

I think that it is not a bug so you can close this.
But if you need a specific support on your use case I suggest you to use our channel at:
https://stackoverflow.com/questions/tagged/tensorflow

[8rax]

Hi,
Alright thank you for the suggestion, posted it there : https://stackoverflow.com/q/64004994/14319427

[google-ml-butler]

Are you satisfied with the resolution of your issue?
Yes
No