Fixed lbfgs for ray-cluster #180
Conversation
|
|
||
| def set_weights(new_weights): | ||
| sess.run(assignment_nodes, feed_dict={p: w for p, w in zip(assignment_placeholders, new_weights)}) | ||
| class LinearModel(object): |
There was a problem hiding this comment.
Can you move this class definition before the __main__ section of the file (after the imports)? Same for net_initialization and net_reinitialization, loss, grad, full_loss and full_grad!
| sess.run(assignment_nodes, feed_dict={p: w for p, w in zip(assignment_placeholders, new_weights)}) | ||
| class LinearModel(object): | ||
| def __init__(self, shape): | ||
| image_dimension = shape[0] |
There was a problem hiding this comment.
why don't you get rid of image_dimension and label_dimension? In general LinearModel can be an arbitrary linear model (not related to images) and these could be arbitrary shapes, and just substituting shape[0] and shape[1] is good!
| @@ -73,27 +70,29 @@ def set_weights(new_weights): | |||
| def net_reinitialization(net_vars): | |||
| return net_vars | |||
|
|
|||
| # Create a reusable variable for the network. | |||
| # Register the network with Ray and create a reusable variable for it. | |||
| ray.register_class(LinearModel) | |||
There was a problem hiding this comment.
register_class shouldn't be neccessary, we shouldn't have to serialize the network (and we shouldn't, if we have to there is a problem somewhere)
| @@ -73,27 +70,29 @@ def set_weights(new_weights): | |||
| def net_reinitialization(net_vars): | |||
| return net_vars | |||
|
|
|||
| # Create a reusable variable for the network. | |||
| # Register the network with Ray and create a reusable variable for it. | |||
| ray.register_class(LinearModel) | |||
| ray.reusables.net_vars = ray.Reusable(net_initialization, net_reinitialization) | |||
|
|
|||
| # Load the weights into the network. | |||
| def load_weights(theta): | |||
There was a problem hiding this comment.
can you define generic functions set_flat and get_flat in TensorFlowVariables for this and use them below?
No description provided.