wassname/ray
1
0
Fork 0
mirror of https://github.com/wassname/ray.git synced 2026-07-13 17:45:08 +08:00
Code Issues Packages Projects Releases Wiki Activity

[tune] Remove accidentally added files (#9835)

Browse Source
This commit is contained in:
Richard Liaw
2020-07-30 21:47:27 -07:00
committed by GitHub
parent 02fd950252
commit a47121476f
11 changed files with 0 additions and 544 deletions
Download Patch File Download Diff File Expand all files Collapse all files
python/ray/.anyscale.yaml
-1
View File
@@ -1 +0,0 @@
project_id: 643
python/ray/session-default.yaml
-77
View File
@@ -1,77 +0,0 @@
# An unique identifier for the head node and workers of this cluster.
cluster_name: sgd-pytorch
# The maximum number of workers nodes to launch in addition to the head
# node. This takes precedence over min_workers. min_workers default to 0.
min_workers: 2
initial_workers: 2
max_workers: 2
target_utilization_fraction: 0.9
# If a node is idle for this many minutes, it will be removed.
idle_timeout_minutes: 20
# docker:
# image: tensorflow/tensorflow:1.5.0-py3
# container_name: ray_docker
# Cloud-provider specific configuration.
provider:
type: aws
region: us-west-2
# How Ray will authenticate with newly launched nodes.
auth:
ssh_user: ubuntu
head_node:
InstanceType: p3.8xlarge
ImageId: latest_dlami
InstanceMarketOptions:
MarketType: spot
# SpotOptions:
# MaxPrice: "9.0"
worker_nodes:
InstanceType: p3.8xlarge
ImageId: latest_dlami
# Run workers on spot by default. Comment this out to use on-demand.
InstanceMarketOptions:
MarketType: spot
# SpotOptions:
# MaxPrice: "9.0"
setup_commands:
- ray || pip install -U https://s3-us-west-2.amazonaws.com/ray-wheels/latest/ray-0.9.0.dev0-cp36-cp36m-manylinux1_x86_64.whl
- pip install -U ipdb ray[rllib] torch torchvision
- cp -r ~/tune ~/anaconda3/lib/python3.6/site-packages/ray
- cp -r ~/torch_ ~/anaconda3/lib/python3.6/site-packages/ray/util/sgd
- cp -r ~/autoscaler ~/anaconda3/lib/python3.6/site-packages/ray/
# Install apex.
# - rm -rf apex || true
# - git clone https://github.com/NVIDIA/apex && cd apex && pip install -v --no-cache-dir ./ || true
file_mounts: {
~/tune: ./tune/,
~/torch_: ./util/sgd/torch/,
~/autoscaler: ./autoscaler/
}
# Custom commands that will be run on the head node after common setup.
head_setup_commands: []
# Custom commands that will be run on worker nodes after common setup.
worker_setup_commands: []
# # Command to start ray on the head node. You don't need to change this.
head_start_ray_commands:
- ray stop
- ray start --head --port=6379 --object-manager-port=8076 --autoscaling-config=~/ray_bootstrap_config.yaml --object-store-memory=1000000000
# Command to start ray on worker nodes. You don't need to change this.
worker_start_ray_commands:
- ray stop
- ray start --address=$RAY_HEAD_IP:6379 --object-manager-port=8076 --object-store-memory=1000000000
python/ray/tune/.anyscale.yaml
-1
View File
@@ -1 +0,0 @@
project_id: 578
python/ray/tune/_test.py
-43
View File
@@ -1,43 +0,0 @@
class Trial:
hypers: dict = {} # static
config: dict = {} # static
status: str = None
trace: List[Dict] = []
checkpoints: List[str] = []
space = {}
trials = []
trial_checkpoints = {}
while not Optimizer.is_finished():
while Optimizer.has_next(space, trials, state):
trials += [Optimizer.next(space, trials, state)]
trial = Optimizer.choose(trials, state)
if Optimizer.should_stop(trial, trials, state):
Executor.stop(trial)
elif Optimizer.should_pause(trial, state):
Executor.pause(trial)
elif Optimizer.should_restore(trial, state):
restore(trial, trial.checkpoints[-1])
elif Optimizer.should_save(trial, state):
checkpoint = save(trial)
elif Optimizer.should_continue(trial, state):
step(trial)
exp = Experiment(logdir, name, restore=True)
failed_trials = exp.get_failed_trials()
run(failed_trials)
exp = Experiment(logdir, name, restore=True)
trials = exp.trials_finished()
trials.reset_status()
run(trials)
optimizer = Optimizer(sweep, metric, *parameters)
sweep.configure_server()
sweep.add_logger(Logging)
sweep.set_executor(executor)
sweep.run(func, verbose=verbose)
python/ray/tune/_test2.py
-207
View File
@@ -1,207 +0,0 @@
storage = TrialStorage(location)
trials = storage.get_trials()
failed_trials = trials.filter(status=Failed)
parameters = [t.hypers for t in failed_trials]
# Builder Pattern
factory = TrialFactory()
factory.queue(grid)
run(func, factory)
factory = TrialFactory()
factory.queue(distribution, num_samples=3, repeat=5)
run(func, factory)
factory = TrialFactory(optimizer)
factory.queue(distribution, delay_feedback=3, num_samples=20, max_concurrent=3)
run(func, factory)
optimizer.restore(storage)
factory = TrialFactory(optimizer)
factory.queue(parameter_list)
factory.queue(distribution, num_samples=3)
# single process
trials = []
while factory.has_next():
x = factory.next()
trial = build(func, x)
trials.put(trial)
storage.save(factory, trials)
while not trial.done():
result = get_next_result(trial)
log(result)
storage.update_checkpoint(trial)
factory.update(result)
storage.save(factory, trials)
# concurrent
trials = []
class Actor:
def __init__(self):
pass
def configure():
pass
def step():
pass
def save():
pass
def restore():
pass
factory, trials = storage.recover()
optimizer = factory.optimizer
result_streams = []
while factory.has_next() or not trials.not_done():
while factory.has_next():
x = factory.next()
trial = build(func, x)
trials.put(trial)
storage.save(factory, trials)
while Cluster.has_space(trials.live()) and trials.has_pending():
trial = trials.pop_pending()
handle = Actor.configure(trial)
result_streams.add(handle)
trial, handle, payload = process_next(result_streams)
if payload.type == "SAVE":
trial.update(payload.checkpoint)
storage.save(trials)
elif payload.type == "STEP":
trial.track(payload.result)
log(payload.result)
else:
pass
if should_checkpoint(trial):
Executor.save(handle)
elif not is_finished(trial):
action = Scheduler(trial, trials)
Executor.execute_action(action, trial)
elif is_finished(trial):
factory.update(trial, result)
storage.save(factory, trials)
# concurrent with checkpointing
# concurrent with pbt
while factory.has_next() or not trials.not_done():
# ...
trial, handle, payload = process_next(result_streams)
elif not is_finished(trial):
action = pbt(trial, trials)
factory.queue(new_hps, trial3.checkpoint)
Executor.execute_action(action, trial)
# Restore last experiment
exp = Experiment.restore(storage=X)
trials = exp.get_trial(filter=failed)
run(func, manual_list)
run(func, space, searcher)
run(func, grid)
run(func, manual_list, checkpoints)
run(func, manual_list)
run(func, exp)
# Core concepts:
# Result: Dict[str, value]
# t_state: Any
# Trial: hps[Dict], static_config[Dict]
# TrialTrace: List[Result], t_state, Trial
# Trainable: t_state, Trial -> t_state, Result
# Optimizer: o_state, List[TrialTrace], Trainable -> (
# o_state, List[TrialTrace])
# SearchAlg: state, Dict[hps, Result] -> state, hps
# Execution concepts
# Checkpoint
# LiveTrial: TrialTrace, location, status, is_idle
# Status: PENDING, SAVING, RESTORING, TRAINING, SETUP, STOP, ERROR
# Trainer: Trainable, location, t_state, Trial -> t_state, Result
step(o_state, LiveTrial, List[LiveTrial]) -> LiveTrial, *args
Server(List[LiveTrial]) -> List[LiveTrial]
checkpointer(LiveTrial, manager_state) -> TrialTrace
Logger(TrialTrace)
Optimizer(o_trace, ...)
Syncer()
TrialExecutor(reuse_actors, queue_trials)
ServerConfig(server_port)
Optimizer(stop, search_alg, scheduler)
Experiment(resume, local_dir)
CheckpointManager(
sync_on_checkpoint,
keep_checkpoints_num,
global_checkpoint_period,
export_formats,
checkpoint_score_attr
)
### Tune commands
tune.set_log_config(
upload_dir,
sync_to_cloud,
trial_name_creator,
sync_to_driver,
progress_reporter,
loggers,
verbose
)
tune.set_server(ServerConfig)
tune.run(
experiment,
trainable_fn,
raise_on_failed_trial, # where can this go?
max_failures: int or "fail-fast",
trial_executor,
restore_from, # checkpoint path to restore from
resources_per_trial,
num_samples,
search_space, # I'm not a big fan of this because Search Algs have their own search_space too
Optimizer,
CheckpointManager)
python/ray/tune/_test3.py
-42
View File
@@ -1,42 +0,0 @@
checkpoint_manager = State(location)
checkpoint_manager.optimizer_state
checkpoint_manager.generator_state
checkpoint_manager.trial_state
# How much have we learned
optimizer = Optimizer.from_checkpoint(checkpoint_manager)
optimizer = Optimizer(space, checkpoint=checkpoint_manager)
for x, y in warm_start:
optimizer.report(x, y)
samples = [optimizer.sample(random=True) for i in range(50)]
spec = TrialSpec(func, local_dir, checkpoint)
generator = TrialGenerator.from_checkpoint(checkpoint, optimizer)
generator = TrialGenerator.from_trials(trials)
generator = TrialGenerator.from_spec(spec, optimizer)
generator.configure(checkpoint_callback)
generator.queue(samples)
generator.queue(num_samples=50, repeat=3, max_concurrent=4)
generator.next()
generator = TrialGenerator.from_multi_spec(spec)
run(generator)
###################################################
# Exploration process
trial_list = get_trials(checkpoint_manager)
failed_trials = [t.reset() for t in trial_list if t.status == "FAILED"]
generator = TrialGenerator.from_trials(failed_trials)
tune.run(generator)
builder = Builder()
for params in samples:
yield builder.build(params)
python/ray/tune/examples/cluster-p3.yaml
-71
View File
@@ -1,71 +0,0 @@
# An unique identifier for the head node and workers of this cluster.
cluster_name: sgd-pytorch
# The maximum number of workers nodes to launch in addition to the head
# node. This takes precedence over min_workers. min_workers default to 0.
min_workers: 0
initial_workers: 0
max_workers: 0
target_utilization_fraction: 0.9
# If a node is idle for this many minutes, it will be removed.
idle_timeout_minutes: 20
# docker:
# image: tensorflow/tensorflow:1.5.0-py3
# container_name: ray_docker
# Cloud-provider specific configuration.
provider:
type: aws
region: us-west-2
# How Ray will authenticate with newly launched nodes.
auth:
ssh_user: ubuntu
head_node:
InstanceType: p3.8xlarge
ImageId: latest_dlami
InstanceMarketOptions:
MarketType: spot
# SpotOptions:
# MaxPrice: "9.0"
worker_nodes:
InstanceType: p3.8xlarge
ImageId: latest_dlami
# Run workers on spot by default. Comment this out to use on-demand.
InstanceMarketOptions:
MarketType: spot
# SpotOptions:
# MaxPrice: "9.0"
setup_commands:
- ray || pip install -U https://s3-us-west-2.amazonaws.com/ray-wheels/latest/ray-0.9.0.dev0-cp36-cp36m-manylinux1_x86_64.whl
- pip install -U ipdb ray[rllib] torch torchvision
# Install apex.
# - rm -rf apex || true
# - git clone https://github.com/NVIDIA/apex && cd apex && pip install -v --no-cache-dir ./ || true
file_mounts: {
}
# Custom commands that will be run on the head node after common setup.
head_setup_commands: []
# Custom commands that will be run on worker nodes after common setup.
worker_setup_commands: []
# # Command to start ray on the head node. You don't need to change this.
head_start_ray_commands:
- ray stop
- ray start --head --redis-port=6379 --object-manager-port=8076 --autoscaling-config=~/ray_bootstrap_config.yaml --object-store-memory=1000000000
# Command to start ray on worker nodes. You don't need to change this.
worker_start_ray_commands:
- ray stop
- ray start --address=$RAY_HEAD_IP:6379 --object-manager-port=8076 --object-store-memory=1000000000
python/ray/tune/session-default.yaml
-73
View File
@@ -1,73 +0,0 @@
# An unique identifier for the head node and workers of this cluster.
cluster_name: sgd-pytorch
# The maximum number of workers nodes to launch in addition to the head
# node. This takes precedence over min_workers. min_workers default to 0.
min_workers: 0
initial_workers: 0
max_workers: 0
target_utilization_fraction: 0.9
# If a node is idle for this many minutes, it will be removed.
idle_timeout_minutes: 20
# docker:
# image: tensorflow/tensorflow:1.5.0-py3
# container_name: ray_docker
# Cloud-provider specific configuration.
provider:
type: aws
region: us-west-2
# How Ray will authenticate with newly launched nodes.
auth:
ssh_user: ubuntu
head_node:
InstanceType: g3.8xlarge
ImageId: latest_dlami
InstanceMarketOptions:
MarketType: spot
# SpotOptions:
# MaxPrice: "9.0"
worker_nodes:
InstanceType: g3.8xlarge
ImageId: latest_dlami
# Run workers on spot by default. Comment this out to use on-demand.
InstanceMarketOptions:
MarketType: spot
# SpotOptions:
# MaxPrice: "9.0"
setup_commands:
- ray || pip install -U https://s3-us-west-2.amazonaws.com/ray-wheels/latest/ray-0.9.0.dev0-cp36-cp36m-manylinux1_x86_64.whl
- pip install -U ipdb ray[rllib] torch torchvision
# Install apex.
# - rm -rf apex || true
# - git clone https://github.com/NVIDIA/apex && cd apex && pip install -v --no-cache-dir ./ || true
file_mounts: {
~/anaconda3/lib/python3.6/site-packages/ray/tune: ./tune/,
~/anaconda3/lib/python3.6/site-packages/ray/util/sgd/torch: ./util/sgd/torch/
}
# Custom commands that will be run on the head node after common setup.
head_setup_commands: []
# Custom commands that will be run on worker nodes after common setup.
worker_setup_commands: []
# # Command to start ray on the head node. You don't need to change this.
head_start_ray_commands:
- ray stop
- ray start --head --redis-port=6379 --object-manager-port=8076 --autoscaling-config=~/ray_bootstrap_config.yaml --object-store-memory=1000000000
# Command to start ray on worker nodes. You don't need to change this.
worker_start_ray_commands:
- ray stop
- ray start --address=$RAY_HEAD_IP:6379 --object-manager-port=8076 --object-store-memory=1000000000
python/ray/tune/test.th
BIN
View File
Binary file not shown.
python/ray/tune/tests/test_serial.py
-28
View File
@@ -1,28 +0,0 @@
import pickle
import ray
import torch
import torch.nn.functional as F
import torch.nn as nn
class ConvNet(nn.Module):
def __init__(self):
super(ConvNet, self).__init__()
self.conv1 = nn.Conv2d(1, 3, kernel_size=3)
self.fc = nn.Linear(192, 10)
def forward(self, x):
x = F.relu(F.max_pool2d(self.conv1(x), 3))
x = x.view(-1, 192)
x = self.fc(x)
return F.log_softmax(x, dim=1)
def save_me():
model = ConvNet()
torch.save(model, "./test.th")
return 1
ray_func = ray.remote(save_me)
ray.init()
ray.get(ray_func.remote())
python/ray/tune/trialv2.py
-1
View File
@@ -1 +0,0 @@
trialv2.py