DQN Double

Instead of using the actions from memory to train on, use the actions chosen by the current model.

Training DataPipes

DoubleQCalc

 DoubleQCalc (*args, **kwds)

Iterable-style DataPipe.

All DataPipes that represent an iterable of data samples should subclass this. This style of DataPipes is particularly useful when data come from a stream, or when the number of samples is too large to fit them all in memory. IterDataPipe is lazily initialized and its elements are computed only when next() is called on the iterator of an IterDataPipe.

All subclasses should overwrite :meth:__iter__, which would return an iterator of samples in this DataPipe. Calling __iter__ of an IterDataPipe automatically invokes its method reset(), which by default performs no operation. When writing a custom IterDataPipe, users should override reset() if necessary. The common usages include resetting buffers, pointers, and various state variables within the custom IterDataPipe.

Note: Only one iterator can be valid for each IterDataPipe at a time, and the creation a second iterator will invalidate the first one. This constraint is necessary because some IterDataPipe have internal buffers, whose states can become invalid if there are multiple iterators. The code example below presents details on how this constraint looks in practice. If you have any feedback related to this constraint, please see GitHub IterDataPipe Single Iterator Issue_.

These DataPipes can be invoked in two ways, using the class constructor or applying their functional form onto an existing IterDataPipe (recommended, available to most but not all DataPipes). You can chain multiple IterDataPipe together to form a pipeline that will perform multiple operations in succession.

.. _GitHub IterDataPipe Single Iterator Issue: https://github.com/pytorch/data/issues/45

Note: When a subclass is used with :class:~torch.utils.data.DataLoader, each item in the DataPipe will be yielded from the :class:~torch.utils.data.DataLoader iterator. When :attr:num_workers > 0, each worker process will have a different copy of the DataPipe object, so it is often desired to configure each copy independently to avoid having duplicate data returned from the workers. :func:~torch.utils.data.get_worker_info, when called in a worker process, returns information about the worker. It can be used in either the dataset’s :meth:__iter__ method or the :class:~torch.utils.data.DataLoader ’s :attr:worker_init_fn option to modify each copy’s behavior.

Examples: General Usage: >>> # xdoctest: +SKIP >>> from torchdata.datapipes.iter import IterableWrapper, Mapper >>> dp = IterableWrapper(range(10)) >>> map_dp_1 = Mapper(dp, lambda x: x + 1) # Using class constructor >>> map_dp_2 = dp.map(lambda x: x + 1) # Using functional form (recommended) >>> list(map_dp_1) [1, 2, 3, 4, 5, 6, 7, 8, 9, 10] >>> list(map_dp_2) [1, 2, 3, 4, 5, 6, 7, 8, 9, 10] >>> filter_dp = map_dp_1.filter(lambda x: x % 2 == 0) >>> list(filter_dp) [2, 4, 6, 8, 10] Single Iterator Constraint Example: >>> from torchdata.datapipes.iter import IterableWrapper, Mapper >>> dp = IterableWrapper(range(10)) >>> it1 = iter(source_dp) >>> list(it1) [0, 1, 2, 3, 4, 5, 6, 7, 8, 9] >>> it1 = iter(source_dp) >>> it2 = iter(source_dp) # The creation of a new iterator invalidates it1 >>> next(it2) 0 >>> next(it1) # Further usage of it1 will raise a RunTimeError

Try training with basic defaults…

from fastrl.envs.gym import *
from fastrl.loggers.vscode_visualizers import *

# Setup Loggers
logger_base = ProgressBarLogger(epoch_on_pipe=EpocherCollector,
                 batch_on_pipe=BatchCollector)

# Setup up the core NN
torch.manual_seed(0)
model = DQN(4,2)
# Setup the Agent
agent = DQNAgent(model,[logger_base],max_steps=4000)
# Setup the DataBlock
block = DataBlock(
    GymTransformBlock(agent=agent,nsteps=1,nskips=1,firstlast=False,n=1000,bs=1)
)
dls = L(block.dataloaders(['CartPole-v1']*1))
# Setup the Learner
learner = DQNLearner(model,dls,logger_bases=[logger_base],bs=128,max_sz=100_000,
                    dp_augmentation_fns=[
                        TargetModelUpdater.insert_dp(),
                        DoubleQCalc.replace_dp()
                    ])
learner.fit(3)
# learner.fit(25)

loss	episode	rolling_reward	epoch	batch	epsilon
0.017258495	46	21.644444	1	1000	0.749500
0.0485615	75	26.837838	2	1000	0.499250
0.03659891	94	31.946237	2	1000	0.249250

The DQN learners, but I wonder if we can get it to learn faster…

# Setup Loggers
logger_base = ProgressBarLogger(epoch_on_pipe=EpocherCollector,
                 batch_on_pipe=BatchCollector)

# Setup up the core NN
torch.manual_seed(0)
model = DQN(4,2)
# Setup the Agent
agent = DQNAgent(model,[logger_base],max_steps=10000)
# Setup the DataBlock
block = DataBlock(
    GymTransformBlock(agent=agent,nsteps=2,nskips=2,firstlast=True,n=1000,bs=1), # We basically merge 2 steps into 1 and skip. 
    (GymTransformBlock(agent=agent,nsteps=2,nskips=2,firstlast=True,n=100,include_images=True),VSCodeTransformBlock())
)
# pipes = L(block.datapipes(['CartPole-v1']*1,n=10))
dls = L(block.dataloaders(['CartPole-v1']*1))
# Setup the Learner
learner = DQNLearner(model,dls,logger_bases=[logger_base],bs=128,max_sz=20_000,nsteps=2,lr=0.001,
                    dp_augmentation_fns=[
                        TargetModelUpdater.insert_dp(),
                        DoubleQCalc.replace_dp()
                    ])
learner.fit(3)
# learner.fit(10)

loss	episode	rolling_reward	epoch	batch	epsilon
0.12804106	71	29.120000	1	1000	0.809700
0.9983938	118	41.120000	2	1000	0.615900
1.8011488	135	66.460000	2	1000	0.418300