Adding a python layer in caffe was fairly straightforward (creating a child class that inherits from caffe.layer and adding four basic methods, as described here and here. However, adding a custom python layer in caffe2 is not as straightforward to me. Can someone please explain the procedure for adding a python layer in caffe2?
First, you must implement your new layer as a Python class as shown in the example. In this case, it only outputs the input tensor in reverse order:
class ReverseOrderOp(object):
def forward(self, inputs, outputs):
blob_out = outputs[0]
blob_out.reshape(inputs[0].shape)
blob_out.data[...] = inputs[0].data[::-1]
Then, you can add your new layer to the model using model.net.Python:
model = ModelHelper(name="test")
l = np.asarray([0,1,2,3])
workspace.FeedBlob('l', l.astype(np.float32))
model.net.Python(ReverseOrderOp().forward)(
['l'], ['out'], name='ReverseOrder'
)
workspace.RunNetOnce(model.net)
print(workspace.FetchBlob('out'))
The output should be [ 3. 2. 1. 0.]
Related
I'm using Tensorboard to see the progress of the PettingZoo environment that my agents are playing. I can see the reward go up with time, which is good, but I'd like to add other metrics that are specific to my environment. i.e. I'd like TensorBoard to show me more charts with my metrics and how they improve over time.
The only way I could figure out how to do that was by inserting a few lines into the learn method of OnPolicyAlgorithm that's part of SB3. This works and I got the charts I wanted:
(The two bottom charts are the ones I added.)
But obviously editing library code isn't a good practice. I should make the modifications in my own code, not in the libraries. Is there currently a more elegant way to add a metric from my PettingZoo environment into TensorBoard?
You can add a callback to add your own logs. See the below example. In this case the call back is called every step. There are other callbacks that you case use depending on your use case.
import numpy as np
from stable_baselines3 import SAC
from stable_baselines3.common.callbacks import BaseCallback
model = SAC("MlpPolicy", "Pendulum-v1", tensorboard_log="/tmp/sac/", verbose=1)
class TensorboardCallback(BaseCallback):
"""
Custom callback for plotting additional values in tensorboard.
"""
def __init__(self, verbose=0):
super(TensorboardCallback, self).__init__(verbose)
def _on_step(self) -> bool:
# Log scalar value (here a random variable)
value = np.random.random()
self.logger.record('random_value', value)
return True
model.learn(50000, callback=TensorboardCallback())
So, I have a model with some parameters that I need to train. And also, there are some parameters in the class Dataset(torch.utils.data.Dataset) which do some preprocessing. I need to train them as well with the model’s parameters. So, can you please let me know if what I am doing is correct:
params = list(model.parameters())
params.extend(list(Dataset.parameters()))
opt = torch.optim.Adam(params,lr=1e-4)
One more question. As Dataset class will generate both the train_ds and val_ds, I only need to train the parameters of the Dataset class when getting train_ds. And to get val_ds, I need to use the trained parameters of the Dataset. So, how do I create the train_ds and val_ds? Should I initially create both of them and then train the model with tarin_ds, and then create them again and use the val_ds for testing?
Dataset class does not have .parameters - it is used only to handle data.
Your model class is derived from nn.Module class that has .parameters and can be trained.
It seems like you need to add the trainable preprocessing to your model, rather than have it as part of your dataset class.
Being a new guy and a beginner to deep learning and pytorch I am not sure what all inputs should I give you guys to answer my question. But I will try my best to make you guys understand my problem. I have loaded a model in pytorch using 'model= torch.load('model/resnet18-5c106cde.pth')'. But it is showing an AttributeError: 'collections.OrderedDict' object has no attribute 'predict', when I used the command 'prediction = model.predict(test_image)'. Hope you guys understood my problem and Thanks in advance...
I'd guess that the checkpoint you are loading stores a model state dict (the model's parameters) rather than a model (the structure of the model plus its parameters). Try:
model = resnet18(*args, **kwargs)
model.load_state_dict(torch.load(PATH))
model.eval()
where PATH is the path to the model checkpoint. You need to declare model as an instance of the object class (declare the model structure) so that you can load the checkpoint (parameters only, no structure). So you'll need to find the appropriate class to import for the resnet18, probably something along the lines of:
from torchvision.models import resnet18
We are working on a Top-Down-RPG-like Multiplayer game for learning purposes (and fun!) with some friends. We already have some Entities in the Game and Inputs are working, but the network implementation gives us headache :D
The Issues
When trying to convert with dict some values will still contain the pygame.Surface, which I dont want to transfer and it causes errors when trying to jsonfy them. Other objects I would like to transfer in a simplyfied way like Rectangle cannot be converted automatically.
Already functional
Client-Server connection
Transfering JSON objects in both directions
Async networking and synchronized putting into a Queue
Situation
A new player connects to the server and wants to get the current game state with all objects.
Data-Structure
We use a "Entity-Component" based architecture, so we separated the game logic very strictly into "systems", while the data is stored in the "components" of each Entity. The Entity is a very simple container and has nothing more than a ID and a list of components. Example Entity (shorten for better readability):
Entity
|-- Component (Moveable)
|-- Component (Graphic)
| |- complex datatypes like pygame.SURFACE
| `- (...)
`- Component (Inventory)
We tried different approaches, but all seems not to fit very well or feel "hacky".
pickle
Very Python near, so not easy to implement other clients in future. And I´ve read about some security risks when creating items from network in this dynamic way how pickle it offers. It does not even solve the Surface/Rectangle issue.
__dict__
Still contains the reference to the old objects, so a "cleanup" or "filter" for unwanted datatypes happens also in the origin. A deepcopy throws Exception.
...\Python\Python36\lib\copy.py", line 169, in deepcopy
rv = reductor(4)
TypeError: can't pickle pygame.Surface objects
Show some code
The method of the "EnitityManager" Class which should generate the Snapshot of all Entities, including their components. This Snapshot should be converted to JSON without any errors - and if possible without much configuration in this core-class.
class EnitityManager:
def generate_world_snapshot(self):
""" Returns a dictionary with all Entities and their components to send
this to the client. This function will probably generate a lot of data,
but, its to send the whole current game state when a new player
connects or when a complete refresh is required """
# It should be possible to add more objects to the snapshot, so we
# create our own Snapshot-Datastructure
result = {'entities': {}}
entities = self.get_all_entities()
for e in entities:
result['entities'][e.id] = deepcopy(e.__dict__)
# Components are Objects, but dictionary is required for transfer
cmp_obj_list = result['entities'][e.id]['components']
# Empty the current list of components, its going to be filled with
# dictionaries of each cmp which are cleaned for the dump, because
# of the errors directly coverting the whole datastructure to JSON
result['entities'][e.id]['components'] = {}
for cmp in cmp_obj_list:
cmp_copy = deepcopy(cmp)
cmp_dict = cmp_copy.__dict__
# Only list, dict, int, str, float and None will stay, while
# other Types are being simply deleted including their key
# Lists and directories will be cleaned ob recursive as well
cmp_dict = self.clean_complex_recursive(cmp_dict)
result['entities'][e.id]['components'][type(cmp_copy).__name__] \
= cmp_dict
logging.debug("EntityMgr: Entity#3: %s" % result['entities'][3])
return result
Expectation and actual results
We can find a way to manually override elements which we dont want. But as the list of components will increase we have to put all the filter logic into this core class, which should not contain any components specializations.
Do we really have to put all the logic into the EntityManager for filtering the right objects? This does not feel good, as I would like to have all convertion to JSON done without any hardcoded configuration.
How to convert all this complex data in a most generic approach?
Thanks for reading so far and thank you very much for your help in advance!
Interesting articles which we were already working threw and maybe helpful for others with similar issues
https://gafferongames.com/post/what_every_programmer_needs_to_know_about_game_networking/
http://code.activestate.com/recipes/408859/
https://docs.python.org/3/library/pickle.html
UPDATE: Solution - thx 2 sloth
We used a combination of the following architecture, which works really great so far and is also good to maintain!
Entity Manager now calls the get_state() function of the entity.
class EntitiyManager:
def generate_world_snapshot(self):
""" Returns a dictionary with all Entities and their components to send
this to the client. This function will probably generate a lot of data,
but, its to send the whole current game state when a new player
connects or when a complete refresh is required """
# It should be possible to add more objects to the snapshot, so we
# create our own Snapshot-Datastructure
result = {'entities': {}}
entities = self.get_all_entities()
for e in entities:
result['entities'][e.id] = e.get_state()
return result
The Entity has only some basic attributes to add to the state and forwards the get_state() call to all the Components:
class Entity:
def get_state(self):
state = {'name': self.name, 'id': self.id, 'components': {}}
for cmp in self.components:
state['components'][type(cmp).__name__] = cmp.get_state()
return state
The components itself now inherit their get_state() method from their new superclass components, which simply cares about all simple datatypes:
class Component:
def __init__(self):
logging.debug('generic component created')
def get_state(self):
state = {}
for attr, value in self.__dict__.items():
if value is None or isinstance(value, (str, int, float, bool)):
state[attr] = value
elif isinstance(value, (list, dict)):
# logging.warn("Generating state: not supporting lists yet")
pass
return state
class GraphicComponent(Component):
# (...)
Now every developer has the opportunity to overlay this function to create a more detailed get_state() function for complex types directly in the Component Classes (like Graphic, Movement, Inventory, etc.) if it is required to safe the state in a more accurate way - which is a huge thing for maintaining the code in future, to have these code pieces in one Class.
Next step is to implement the static method for creating the items from the state in the same Class. This makes this working really smooth.
Thank you so much sloth for your help.
Do we really have to put all the logic into the EntityManager for filtering the right objects?
No, you should use polymorphism.
You need a way to represent your game state in a form that can be shared between different systems; so maybe give your components a method that will return all of their state, and a factory method that allows you create the component instances out of that very state.
(Python already has the __repr__ magic method, but you don't have to use it)
So instead of doing all the filtering in the entity manager, just let him call this new method on all components and let each component decide that the result will look like.
Something like this:
...
result = {'entities': {}}
entities = self.get_all_entities()
for e in entities:
result['entities'][e.id] = {'components': {}}
for cmp in e.components:
result['entities'][e.id]['components'][type(cmp).__name__] = cmp.get_state()
...
And a component could implement it like this:
class GraphicComponent:
def __init__(self, pos=...):
self.image = ...
self.rect = ...
self.whatever = ...
def get_state(self):
return { 'pos_x': self.rect.x, 'pos_y': self.rect.y, 'image': 'name_of_image.jpg' }
#staticmethod
def from_state(state):
return GraphicComponent(pos=(state.pos_x, state.pos_y), ...)
And a client's EntityManager that recieves the state from the server would iterate for the component list of each entity and call from_state to create the instances.
I am new to theano, can anyone help me defining a theano function like this:
Basically, I have a network model looks like this:
y_hat, cost, mu, output_hiddens, cells = nn_f(x, y, in_size, out_size, hidden_size, layer_models, 'MDN', training=False)
here the input x is a tensor:
x = tensor.tensor3('features', dtype=theano.config.floatX)
I want to define two theano functions for later use:
f_x_hidden = theano.function([x], [output_hiddens])
f_hidden_mu = theano.function([output_hiddens], [mu], on_unused_input = 'warn')
the first one is fine. for the second one, the problem is both the input and the output are output of the original function. it gives me error:
theano.gof.fg.MissingInputError: An input of the graph, used to compute Elemwise{identity}(features), was not provided and not given a value.
my understanding is, both of [output_hiddens] and [mu] are related to the input [x], there should be an relation between them. I tried define another theano function from [x] to [mu] like:
f_x_mu = theano.function([x], [mu]),
then
f_hidden_mu = theano.function(f_x_hidden, f_x_mu),
but it still does not work. Does anyone can help me? Thanks.
The simple answer is NO WAY. In here
Because in Theano you first express everything symbolically and afterwards compile this expression to get functions, ...
You can't use the output of theano.function as input/output for another theano.function since they are already a compiled graph/function.
You should pass the symbolic variables, such as x in your example code for f_x_hidden, to build the model.