Source code for invertedai.api.initialize

import time
from pydantic import BaseModel, validate_call
from typing import List, Optional, Dict, Tuple
import asyncio

import invertedai as iai
from invertedai.api.config import TIMEOUT, should_use_mock_api
from invertedai.error import TryAgain, InvalidInputType, InvalidInput
from invertedai.api.mock import (
    get_mock_agent_attributes,
    get_mock_agent_properties,
    get_mock_agent_state,
    get_mock_recurrent_state,
    get_mock_birdview,
    get_mock_infractions,
    get_mock_traffic_light_states,
    get_mock_light_recurrent_states
)
from invertedai.common import (
    RecurrentState,
    AgentState,
    AgentAttributes,
    AgentProperties,
    TrafficLightStatesDict,
    Image,
    InfractionIndicators,
    LightRecurrentStates,
    LightRecurrentState,
)


[docs]class InitializeResponse(BaseModel): """ Response returned from an API call to :func:`iai.initialize`. """ agent_states: List[AgentState] #: Initial states of all initialized agents. recurrent_states: List[Optional[RecurrentState]] #: To pass to :func:`iai.drive` at the first time step. agent_attributes: List[Optional[AgentAttributes]] #: Static attributes of all initialized agents. agent_properties: List[AgentProperties] #: Static agent properties of all initialized agents. birdview: Optional[Image] #: If `get_birdview` was set, this contains the resulting image. infractions: Optional[List[InfractionIndicators]] #: If `get_infractions` was set, they are returned here. traffic_lights_states: Optional[TrafficLightStatesDict] #: Traffic light states for the full map, each key-value pair corresponds to one particular traffic light. light_recurrent_states: Optional[LightRecurrentStates] #: Light recurrent states for the full map. Pass this to :func:`iai.drive` at the first time step to let the server generate a realistic continuation of the traffic light state sequence. This does not work correctly if any specific light states were specified as input to `initialize`. api_model_version: str #: Model version used for this API call
[docs]@validate_call def initialize( location: str, agent_attributes: Optional[List[AgentAttributes]] = None, agent_properties: Optional[List[AgentProperties]] = None, states_history: Optional[List[List[AgentState]]] = None, traffic_light_state_history: Optional[List[TrafficLightStatesDict]] = None, get_birdview: bool = False, location_of_interest: Optional[Tuple[float, float]] = None, get_infractions: bool = False, agent_count: Optional[int] = None, random_seed: Optional[int] = None, api_model_version: Optional[str] = None # Model version used for this API call ) -> InitializeResponse: """ Initializes a simulation in a given location, using a combination of **user-defined** and **sampled** agents. **User-defined** agents are placed in a scene first, after which a number of agents are sampled conditionally inferred from the `agent_count` argument. If **user-defined** agents are desired, `states_history` must contain a list of `AgentState's` of all **user-defined** agents per historical time step. Any **user-defined** agent must have a corresponding fully specified static `AgentAttribute` in `agent_attributes`. Any **sampled** agents not specified in `agent_attributes` will be generated with default static attribute values however **sampled** agents may be defined by specifying `agent_type` only. Agents are identified by their list index, so ensure the indices of each agent match in `states_history` and `agent_attributes` when applicable. If traffic lights are present in the scene, their states history can be specified with a list of `TrafficLightStatesDict`, each represent light states for one timestep, with the last element representing the current time step. It is legal to omit the traffic light state specification, and the scene will be initialized with a light state configuration consistent with agent states. Every simulation must start with a call to this function in order to obtain correct recurrent states for :func:`drive`. Parameters ---------- location: Location name in IAI format. agent_attributes: Static attributes for all agents. The pre-defined agents should be specified first, followed by the sampled agents. The optional waypoint passed will be ignored for Initialize. agent_properties: Agent properties for all agents, replacing soon to be deprecated `agent_attributes`. The pre-defined agents should be specified first, followed by the sampled agents. The optional waypoint passed will be ignored for Initialize. max_speed: optional [float], the desired maximum speed of the agent in m/s. states_history: History of pre-defined agent states - the outer list is over time and the inner over agents, in chronological order, i.e., index 0 is the oldest state and index -1 is the current state. The order of agents should be the same as in `agent_attributes`. For best results, provide at least 10 historical states for each agent. traffic_light_state_history: History of traffic light states - the list is over time, in chronological order, i.e. the last element is the current state. If there are traffic lights in the map, not specifying traffic light state is equivalent to using server generated light states. location_of_interest: Optional coordinates for spawning agents with the given location as center instead of the default map center get_birdview: If True, a birdview image will be returned representing the current world. Note this will significantly impact on the latency. get_infractions: If True, infraction metrics will be returned for each agent. agent_count: Deprecated. Equivalent to padding the `agent_attributes` list to this length with default `AgentAttributes`. random_seed: Controls the stochastic aspects of initialization for reproducibility. api_model_version: Optionally specify the version of the model. If None is passed which is by default, the best model will be used. See Also -------- :func:`drive` :func:`location_info` :func:`light` :func:`blame` """ if should_use_mock_api(): assert agent_properties is not None or agent_attributes is not None or agent_count is not None if agent_properties is not None: agent_count = len(agent_properties) elif agent_attributes is not None: agent_count = len(agent_attributes) agent_properties = [get_mock_agent_properties() for _ in range(agent_count)] agent_attributes = [get_mock_agent_attributes() for _ in range(agent_count)] if agent_attributes is None: agent_states = [get_mock_agent_state() for _ in range(agent_count)] else: agent_states = states_history[-1] if states_history is not None else [] recurrent_states = [get_mock_recurrent_state() for _ in range(agent_count)] birdview = get_mock_birdview() infractions = get_mock_infractions(len(agent_states)) response = InitializeResponse( agent_states=agent_states, agent_attributes=agent_attributes, agent_properties=agent_properties, recurrent_states=recurrent_states, birdview=birdview, infractions=infractions, api_model_version=api_model_version if api_model_version is not None else "best", traffic_lights_states=traffic_light_state_history[-1] if traffic_light_state_history is not None else None, light_recurrent_states=get_mock_light_recurrent_states(len(traffic_light_state_history[0])) if traffic_light_state_history is not None else None ) return response model_inputs = dict( location=location, num_agents_to_spawn=agent_count, states_history=states_history if states_history is None else [[st.tolist() for st in states] for states in states_history], agent_attributes=agent_attributes if agent_attributes is None else [state.tolist() for state in agent_attributes], agent_properties=agent_properties if agent_properties is None else [ap.serialize() if ap else None for ap in agent_properties] , traffic_light_state_history=traffic_light_state_history, get_birdview=get_birdview, location_of_interest=location_of_interest, get_infractions=get_infractions, random_seed=random_seed, model_version=api_model_version if api_model_version is not None else "best" ) start = time.time() timeout = TIMEOUT while True: try: response = iai.session.request(model="initialize", data=model_inputs) response = InitializeResponse( agent_states=[ AgentState.fromlist(state) for state in response["agent_states"] ], agent_attributes=[ AgentAttributes.fromlist(attr) for attr in response["agent_attributes"] ] if response["agent_attributes"] is not None else [], agent_properties=[ AgentProperties.deserialize(ap) for ap in response["agent_properties"] ], recurrent_states=[ RecurrentState.fromval(r) for r in response["recurrent_states"] ], birdview=Image.fromval(response["birdview"]) if response["birdview"] is not None else None, infractions=[ InfractionIndicators.fromlist(infractions) for infractions in response["infraction_indicators"] ] if response["infraction_indicators"] else [], api_model_version=response["model_version"], traffic_lights_states=response["traffic_lights_states"] if response["traffic_lights_states"] is not None else None, light_recurrent_states=[ LightRecurrentState(state=state_arr[0], time_remaining=state_arr[1]) for state_arr in response["light_recurrent_states"] ] if response["light_recurrent_states"] is not None else None ) return response except TryAgain as e: if timeout is not None and time.time() > start + timeout: raise e iai.logger.info(iai.logger.logfmt("Waiting for model to warm up", error=e))
@validate_call async def async_initialize( location: str, agent_attributes: Optional[List[AgentAttributes]] = None, agent_properties: Optional[List[AgentProperties]] = None, states_history: Optional[List[List[AgentState]]] = None, traffic_light_state_history: Optional[List[TrafficLightStatesDict]] = None, get_birdview: bool = False, location_of_interest: Optional[Tuple[float, float]] = None, get_infractions: bool = False, agent_count: Optional[int] = None, random_seed: Optional[int] = None, api_model_version: Optional[str] = None ) -> InitializeResponse: """ The async version of :func:`initialize` """ model_inputs = dict( location=location, num_agents_to_spawn=agent_count, states_history=states_history if states_history is None else [[st.tolist() for st in states] for states in states_history], agent_attributes=agent_attributes if agent_attributes is None else [state.tolist() for state in agent_attributes], agent_properties=agent_properties if agent_properties is None else [ap.serialize() if ap else None for ap in agent_properties] , traffic_light_state_history=traffic_light_state_history, get_birdview=get_birdview, location_of_interest=location_of_interest, get_infractions=get_infractions, random_seed=random_seed, model_version=api_model_version ) response = await iai.session.async_request(model="initialize", data=model_inputs) agents_spawned = len(response["agent_states"]) if agents_spawned != agent_count: iai.logger.warning( f"Unable to spawn a scenario for {agent_count} agents, {agents_spawned} spawned instead." ) response = InitializeResponse( agent_states=[ AgentState.fromlist(state) for state in response["agent_states"] ], agent_attributes=[ AgentAttributes.fromlist(attr) for attr in response["agent_attributes"] ], agent_properties=[ AgentProperties.deserialize(ap) for ap in response["agent_properties"] ], recurrent_states=[ RecurrentState.fromval(r) for r in response["recurrent_states"] ], birdview=Image.fromval(response["birdview"]) if response["birdview"] is not None else None, infractions=[ InfractionIndicators.fromlist(infractions) for infractions in response["infraction_indicators"] ] if response["infraction_indicators"] else [], api_model_version=response["model_version"], traffic_lights_states=response["traffic_lights_states"] if response["traffic_lights_states"] is not None else None, light_recurrent_states=[ LightRecurrentState(state=state_arr[0], time_remaining=state_arr[1]) for state_arr in response["light_recurrent_states"] ] if response["light_recurrent_states"] is not None else None ) return response