I'm trying to log my ML trials with mlflow.keras.autolog and mlflow.log_param simultaneously (mlflow v 1.22.0). However, the only things that are recorded are autolog's products, but not those of log_param.
experiment = mlf_client.get_experiment_by_name(experiment_name)
with mlflow.start_run(experiment_id=experiment.experiment_id):
mlflow.keras.autolog(log_input_examples=True)
mlflow.log_param('batch_size', self.batch_size)
mlflow.log_param('training_set_size', len(kwargs['training_ID_list']))
mlflow.log_param('testing_set_size', len(kwargs['testing_ID_list']))
history = self.train_NN_model(**kwargs)
I know I can use log_param with log_model to save the model itself, but then I lose some useful stuff that autolog can record for me automatically (e.g., model summary).
Is it possible to use autolog with custom parameters for logging?
Related
EDIT:
My question was horrifically put so I delete it and rephrase entirely here.
I'll give a tl;dr:
I'm trying to assign each computation to a designated worker that fits the computation type.
In long:
I'm trying to run a simulation, so I represent it using a class of the form:
Class Simulation:
def __init__(first_Client: Client, second_Client: Client)
self.first_client = first_client
self.second_client = second_client
def first_calculation(input):
with first_client.as_current():
return output
def second_calculation(input):
with second_client.as_current():
return output
def run(input):
return second_calculation(first_calculation(input))
This format has downsides like the fact that this simulation object is not pickleable.
I could edit the Simulation object to contain only addresses and not clients for example, but I feel as if there must be a better solution. For instance, I would like the simulation object to work the following way:
Class Simulation:
def first_calculation(input):
client = dask.distributed.get_client()
with client.as_current():
return output
...
Thing is, the dask workers best fit for the first calculation, are different than the dask workers best fit for the second calculation, which is the reason my Simulation object has two clients that connect to tow different schedulers to begin with. Is there any way to make it so there is only one client but two types of schedulers and to make it so the client knows to run the first_calculation to the first scheduler and the second_calculation to the second one?
Dask will chop up large computations in smaller tasks that can run in paralell. Those tasks will then be submitted by the client to the scheduler which in turn wil schedule those tasks on the available workers.
Sending the client object to a Dask scheduler will likely not work due to the serialization issue you mention.
You could try one of two approaches:
Depending on how you actually run those worker machines, you could specify different types of workers for different tasks. If you run on kubernetes for example you could try to leverage the node pool functionality to make different worker types available.
An easier approach using your existing infrastructure would be to return the results of your first computation back to the machine from which you are using the client using something like .compute(). And then use that data as input for the second computation. So in this case you're sending the actual data over the network instead of the client. If the size of that data becomes an issue you can always write the intermediary results to something like S3.
Dask does support giving specific tasks to specific workers with annotate. Here's an example snippet, where a delayed_sum task was passed to one worker and the doubled task was sent to the other worker. The assert statements check that those workers really were restricted to only those tasks. With annotate you shouldn't need separate clusters. You'll also need the most recent versions of Dask and Distributed for this to work because of a recent bug fix.
import distributed
import dask
from dask import delayed
local_cluster = distributed.LocalCluster(n_workers=2)
client = distributed.Client(local_cluster)
workers = list(client.scheduler_info()['workers'].keys())
with dask.annotate(workers=workers[0]):
delayed_sum = delayed(sum)([1, 2])
with dask.annotate(workers=workers[1]):
doubled = delayed_sum * 2
# use persist so scheduler doesn't clean up
# wrap in a distributed.wait to make sure they're there when we check the scheduler
distributed.wait([doubled.persist(), delayed_sum.persist()])
worker_restrictions = local_cluster.scheduler.worker_restrictions
assert worker_restrictions[delayed_sum.key] == {workers[0]}
assert worker_restrictions[doubled.key] == {workers[1]}
From the Best Practices Guide to using Sidekiq, I understand it's best to pass "string, integer, float, boolean, null(nil), array and hash" as arguments to the job.
I often just pass the id of a persisted object to my jobs, but due to latency constraints I need to save the object after running the job.
The non-persisted object I'm working with contains a mixture of data types:
#MyObject<00x000>{
id: nil
start_time: Fri, 11 Dec 2020 08:45:00 PST -08:00 (*this is a TimeWithZone object)
rate: 18.0 (*this is a BigDecimal object)
...
}
I plan to pass this object to my job by converting it to a hash first:
MyJob.perform_async(my_object.attributes)
and then later persist the object like so:
MyObject.new(my_object_hash).save
My question is, is this safe? Even though I am passing a 'simple' datatype to Sidekiq, it actually contains complex objects. Am I going to lose precision?
Thank you!
This sounds like a "potayto, potahto" solution. You are not not using the serialisation of Sidekiq, but instead serialize it yourself.
Let's have a look at why sidekiq has this rule:
Even if they did serialize correctly, what happens if your queue backs up and that quote object changes in the meantime? [...]
Don't pass symbols, named parameters, keyword arguments or complex Ruby objects (like Date or Time!) as those will not survive the dump/load round trip correctly.
I like to add a third:
Serializing state makes it impossible to distinguish between persisted and ethereal (in-memory, memoized, lazy-loaded etc) data. E.g. a def sent_mails; #sent_mails ||= Mail.for(user_id: id); end now gets serialized: do you want that?
The solution is also provided by sidekiq:
Don't save state to Sidekiq, save simple identifiers. Look up the objects once you actually need them in your perform method.
The XY problem here
Your real problem is not where or how to serialize state. Because sidekiq warns against serializing state regardless of where and how you do this.
The problem you need to solve is either how to store state somewhere where it can be stored properly. Or to avoid storing the state at all: not in redis/sidekiq, nor in the storage that is giving you problems.
Latency
Is your storage slow? Is it not a validation, a serialisation, some side-effect of storage that is slow?
Can you improve this by making it a two-step: insert the state and update/enrich/validate it async later? If you are using Rails, it won't help you here, or might even work against you, but a common model is to store objects in a special "queue" table or events queue; e.g. kafka is famous for this.
When e.g. storage happens over a slow network to a slow API, this is probably unsolvable, but when storage happens in a local database, there are decades of solutions to improve write performance here that you can use. Both inside your database, or with some specialised queue for state-storage (sidekiq is not such a specialised storage queue) depending on the tech used to store. E.g. Linux will allow you to store through memory, making writes to disk really quick, but removing the guarantee that it was really written to disk.
E.g. In a bookkeeping api, we would store the validated object in PostgreSQL and then have async jobs add expensive attributes to this later (e.g. state that had to be retrieved from legacy APIs or through complex calculations).
E.g. in a write-heavy GIS system, we would store objects into a "to_process_places" table, that was monitored by tooling which processes the Places. It all really depends on your domain, and requirements.
Not using state.
A common solution is not to make objects, but use the actual payload by the customer. Just send the HTTP payload (in rails, the params) along and leave it at that. Maybe merge in a header (like the Request Date) or filter out some data (header tokens or cookies).
If your controller can operate with this data, so can a delayed job. Instead of building objects in the controller, leave that to the delayed job. This can even result in really neat and lean controllers: all they do is (some authentication and authorization and then) call the proper job and pass it a sanitized params.
Obviously this requires trade-offs like not being able to validate in-sync, but to give such info over email, push-notification, or delayed response instead, depending on your requirements (e.g. a large CSV import could just email any validation issues, but a login request might need to get immediate response if the login is invalid).
It also requires some thought: you probably don't want to send the Base64 encoded CSV along to sidekiq, but instead write the file to a (temp) storage and pass the filename/url along instead. This might sound obvious, because it is: file uploads are essentially an implementation of the earlier mentioned "temporary state storage": you don't pass the entire PDF/high-res-header-image/CSV along to sidekiq, but store it somewhere so sidekiq can pick it up later to process it. Why should the other attributes not employ the same pattern if passing them along to sidekiq is problematic?
The most important part from the best practices you linked is
Complex Ruby objects do not convert to JSON
Therefore you're not supposed to pass instances of a model to a worker.
If you're using Sidekiq workers, you should comply with this statement and the hash you're passing should be just fine. I am not exactly sure about the TimeWithZone object, but you could try converting this to a JSON or to a string as they do in the best practices guide.
However, if you're using ActiveJob instead of Sidekiq workers (does your Job inherit from ApplicationJob or does it include Sidekiq::Worker ?), then you don't have that problem because ActiveJob uses Global ID to convert objects into a String. And then before performing the job is deserializing the object again. Meaning you can pass an object to your job.
my_object = MyObject.find(1)
my_object.to_global_id #=> #<GlobalID:0x000045432da2344 [...] gid://your_app_name/MyObject/1>>
serialized_my_object = my_object.to_global_id.to_s
my_object = GlobalID.find(serialized_my_object)
You can find more information here
https://github.com/toptal/active-job-style-guide#active-record-models-as-arguments
After doing some experimentation on the Time objects in my job, I found that I am losing nanosecond precision at the other end of the job.
my_object.start_time
=> Mon, 21 Dec 2020 11:35:50 PST -08:00
my_object.strftime('%Y-%m-%d %H:%M:%S.%N')
=> "2020-12-21 11:35:50.151893000"
You can see here, we have precision including 6 digits after the decimal.
(see this answer for more about 'strftime')
Once we call JSON methods on the object:
generated = JSON.generate(my_object.attributes))
=> \"start_time\":\"2020-12-21T11:35:50.151-08:00\"
You can see here we are down to 3 digits of precision after the decimal. The remaining 3 digits are lost at this point.
parsed = JSON.parse(generated)
parsed[‘start_time’] = "2020-12-21T11:35:50.151-08:00"
It appears at the most basic level, the JSON library recursively calls as_json on each of the key-value pairs in the hash. So really it depends on how your particular object implements as_json.
This issue caused test failures that involved querying our db for persisted objects (initialized with something like, start_time = Time.zone.now (!)) that are meant to overlap in time exactly with our MyObject class. Once the half-baked my_object blueprints made it through Sidekiq, they lost a sliver of precision, causing a slight misalignment.
One way to hack away at this issue is by monkey patching the Time class.
In our case, a better solution was to go in the opposite direction and to not use so much precision in our tests. The my_object in the example is something that a human user will have on their calendar; in production we never receive so much precision from clients. So instead we fixed our tests by instructing some of our test objects to use something like Time.zone.now.beginning_of_minute, rather than Time.zone.now. We intentionally removed precision to fix the issue, as well as more closely mirror reality.
Firestore recently started to print a warning in the console that I'd have to use areTimestampsInSnapshotsEnabled = true on FIRFirestoreSettings.
At the same time I have to review and update the use of dates by replacing them with timestamps.
For my purpose the higher precision of FIRTimestamps is not required. Furthermore I would like to keep using (NS)Date in my models in order to avoid the firebase dependency within the models.
Is there a way I can keep using (NS)Dates in my Codable models without having to implement a custom decoder/ encoder when areTimestampsInSnapshotsEnabled is enabled?
I have an assignment to make an AI Agent that will learn to play a video game using ML. I want to create a new environment using OpenAI Gym because I don't want to use an existing environment. How can I create a new, custom Environment?
Also, is there any other way I can start to develop making AI Agent to play a specific video game without the help of OpenAI Gym?
See my banana-gym for an extremely small environment.
Create new environments
See the main page of the repository:
https://github.com/openai/gym/blob/master/docs/creating_environments.md
The steps are:
Create a new repository with a PIP-package structure
It should look like this
gym-foo/
README.md
setup.py
gym_foo/
__init__.py
envs/
__init__.py
foo_env.py
foo_extrahard_env.py
For the contents of it, follow the link above. Details which are not mentioned there are especially how some functions in foo_env.py should look like. Looking at examples and at gym.openai.com/docs/ helps. Here is an example:
class FooEnv(gym.Env):
metadata = {'render.modes': ['human']}
def __init__(self):
pass
def _step(self, action):
"""
Parameters
----------
action :
Returns
-------
ob, reward, episode_over, info : tuple
ob (object) :
an environment-specific object representing your observation of
the environment.
reward (float) :
amount of reward achieved by the previous action. The scale
varies between environments, but the goal is always to increase
your total reward.
episode_over (bool) :
whether it's time to reset the environment again. Most (but not
all) tasks are divided up into well-defined episodes, and done
being True indicates the episode has terminated. (For example,
perhaps the pole tipped too far, or you lost your last life.)
info (dict) :
diagnostic information useful for debugging. It can sometimes
be useful for learning (for example, it might contain the raw
probabilities behind the environment's last state change).
However, official evaluations of your agent are not allowed to
use this for learning.
"""
self._take_action(action)
self.status = self.env.step()
reward = self._get_reward()
ob = self.env.getState()
episode_over = self.status != hfo_py.IN_GAME
return ob, reward, episode_over, {}
def _reset(self):
pass
def _render(self, mode='human', close=False):
pass
def _take_action(self, action):
pass
def _get_reward(self):
""" Reward is given for XY. """
if self.status == FOOBAR:
return 1
elif self.status == ABC:
return self.somestate ** 2
else:
return 0
Use your environment
import gym
import gym_foo
env = gym.make('MyEnv-v0')
Examples
https://github.com/openai/gym-soccer
https://github.com/openai/gym-wikinav
https://github.com/alibaba/gym-starcraft
https://github.com/endgameinc/gym-malware
https://github.com/hackthemarket/gym-trading
https://github.com/tambetm/gym-minecraft
https://github.com/ppaquette/gym-doom
https://github.com/ppaquette/gym-super-mario
https://github.com/tuzzer/gym-maze
Its definitely possible. They say so in the Documentation page, close to the end.
https://gym.openai.com/docs
As to how to do it, you should look at the source code of the existing environments for inspiration. Its available in github:
https://github.com/openai/gym#installation
Most of their environments they did not implement from scratch, but rather created a wrapper around existing environments and gave it all an interface that is convenient for reinforcement learning.
If you want to make your own, you should probably go in this direction and try to adapt something that already exists to the gym interface. Although there is a good chance that this is very time consuming.
There is another option that may be interesting for your purpose. It's OpenAI's Universe
https://universe.openai.com/
It can integrate with websites so that you train your models on kongregate games, for example. But Universe is not as easy to use as Gym.
If you are a beginner, my recommendation is that you start with a vanilla implementation on a standard environment. After you get passed the problems with the basics, go on to increment...
I have a class method (placed in /app/lib/) which performs some heavy calculations and sub-http requests until a result is received.
The result isn't too dynamic, and requested by multiple users accessing a specific view in the app.
So, I want to schedule a periodic run of the method (using cron and Whenever gem), store the results somewhere in the server using JSON format and, by demand, read the results alone to the view.
How can this be achieved? what would be the correct way of doing that?
What I currently have:
def heavyMethod
response = {}
# some calculations, eventually building the response
File.open(File.expand_path('../../../tmp/cache/tests_queue.json', __FILE__), "w") do |f|
f.write(response.to_json)
end
end
and also a corresponding method to read this file.
I searched but couldn't find an example of achieving this using Rails cache convention (and not some private code that I wrote), on data which isn't related with ActiveRecord.
Thanks!
Your solution should work fine, but using Rails.cache should be cleaner and a bit faster. Rails guides provides enough information about Rails.cache and how to get it to work with memcached, let me summarize how I would use it in your case
Heavy method
def heavyMethod
response = {}
# some calculations, eventually building the response
Rails.cache.write("heavy_method_response", response)
end
Request
response = Rails.cache.fetch("heavy_method_response")
The only problem here is that when ur server starts for the first time, the cache will be empty. Also if/when memcache restarts.
One advantage is that somewhere on the flow, the data u pass in is marshalled into storage, and then unmartialled on the way out. Meaning u can pass in complex datastructures, and dont need to serialize to json manually.
Edit: memcached will clear your item if it runs out of memory. Will be very rare since its using a LRU (i think) algoritm to expire things, and I presume you will use this often.
To prevent this,
set expires_in larger than your cron period,
change your fetch code to call the heavy_method if ur fetch fails (like Rails.cache.fetch("heavy_method_response") {heavy_method}, and change heavy_method to just return the object.
Use something like redis which will not delete items.