paddleslim.common package

class paddleslim.common.EvolutionaryController

Bases: object

Abstract controller for all evolutionary searching method.

next_tokens()

Generate new tokens.

Returns:The next searched tokens.
Return type:list<list>
reset(range_table, constrain_func=None)

Reset the controller.

Parameters:
  • range_table (list<int>) – It is used to define the searching space of controller. The tokens[i] generated by controller should be in [0, range_table[i]).
  • constrain_func (function) – It is used to check whether tokens meet the constraint. None means there is no constraint. Default: None.
update(tokens, reward)

Update the status of controller according current tokens and reward.

Parameters:
  • tokens (list<int>) – A solution of searching task.
  • reward (list<int>) – The reward of tokens.
class paddleslim.common.SAController(range_table=None, reduce_rate=0.85, init_temperature=None, max_try_times=300, init_tokens=None, reward=-1, max_reward=-1, iters=0, best_tokens=None, constrain_func=None, checkpoints=None, searched=None)

Bases: paddleslim.common.controller.EvolutionaryController

Simulated annealing controller.

Parameters:
  • range_table (list<int>) – Range table.
  • reduce_rate (float) – The decay rate of temperature.
  • init_temperature (float) – Init temperature.
  • max_try_times (int) – max try times before get legal tokens. Default: 300.
  • init_tokens (list<int>) – The initial tokens. Default: None.
  • reward (float) – The reward of current tokens. Default: -1.
  • max_reward (float) – The max reward in the search of sanas, in general, best tokens get max reward. Default: -1.
  • iters (int) – The iteration of sa controller. Default: 0.
  • best_tokens (list<int>) – The best tokens in the search of sanas, in general, best tokens get max reward. Default: None.
  • constrain_func (function) – The callback function used to check whether the tokens meet constraint. None means there is no constraint. Default: None.
  • checkpoints (str) – if checkpoint is None, donnot save checkpoints, else save scene to checkpoints file.
  • searched (dict<list, float>) – remember tokens which are searched.
best_tokens

Get current best tokens.

Returns:The best tokens.
Return type:list<int>
current_tokens

Get tokens generated in current searching step.

Returns:The best tokens.
Return type:list<int>
max_reward
next_tokens(control_token=None)

Get next tokens.

Parameters:control_token – The tokens used to generate next tokens.
Returns:The next tokens.
Return type:list<int>
update(tokens, reward, iter, client_num)

Update the controller according to latest tokens and reward.

Parameters:
  • tokens (list<int>) – The tokens generated in current step.
  • reward (float) – The reward of tokens.
  • iter (int) – The current step of searching client.
  • client_num (int) – The total number of searching client.
paddleslim.common.get_logger(name, level, fmt='%(asctime)s-%(levelname)s: %(message)s')

Get logger from logging with given name, level and format without setting logging basicConfig. For setting basicConfig in paddle will disable basicConfig setting after import paddle.

Parameters:
  • name (str) – The logger name.
  • level (logging.LEVEL) – The base level of the logger
  • fmt (str) – Format of logger output
Returns:

logging logger with given setttings
Return type:

logging.Logger

Examples:

logger = log_helper.get_logger(__name__, logging.INFO,
                     fmt='%(asctime)s-%(levelname)s: %(message)s')
class paddleslim.common.ControllerServer(controller=None, address=('', 0), max_client_num=100, search_steps=None, key=None)

Bases: object

The controller wrapper with a socket server to handle the request of search agent. :param controller: The controller used to generate tokens. :type controller: slim.searcher.Controller :param address: The address of current server binding with format (ip, port). Default: (‘’, 0).

which means setting ip automatically
Parameters:
  • max_client_num (int) – The maximum number of clients connecting to current server simultaneously. Default: 100.
  • search_steps (int|None) – The total steps of searching. None means never stopping. Default: None
  • key (str|None) – Config information. Default: None.
close()

Close the server.

ip()

Get the ip.

port()

Get the port.

run()

Start the server.

start()
class paddleslim.common.ControllerClient(server_ip=None, server_port=None, key=None, client_name=None)

Bases: object

Controller client. :param server_ip: The ip that controller server listens on. None means getting the ip automatically. Default: None. :type server_ip: str :param server_port: The port that controller server listens on. 0 means getting usable port automatically. Default: 0. :type server_port: int :param key: The key used to identify legal agent for controller server. Default: “light-nas” :type key: str :param client_name: Current client name, random generate for counting client number. Default: None. :type client_name: str

next_tokens()

Get next tokens.

request_current_info()

Request for current information.

update(tokens, reward, iter)

Update the controller according to latest tokens and reward.

Parameters:
  • tokens (list<int>) – The tokens generated in last step.
  • reward (float) – The reward of tokens.
  • iter (int) – The iteration number of current client.
paddleslim.common.lock(file)

Lock the file in local file system.

paddleslim.common.unlock(file)

Unlock the file in local file system.

paddleslim.common.cached_reader(reader, sampled_rate, cache_path, cached_id)

Sample partial data from reader and cache them into local file system.

Parameters:
  • reader – Iterative data source.
  • sampled_rate (float) – The sampled rate used to sample partial data for evaluation. None means using all data in eval_reader. default: None.
  • cache_path (str) – The path to cache the sampled data.
  • cached_id (int) – The id of dataset sampled. Evaluations with same cached_id use the same sampled dataset. default: 0.

Submodules

paddleslim.common.cached_reader module

paddleslim.common.cached_reader.cached_reader(reader, sampled_rate, cache_path, cached_id)

Sample partial data from reader and cache them into local file system.

Parameters:
  • reader – Iterative data source.
  • sampled_rate (float) – The sampled rate used to sample partial data for evaluation. None means using all data in eval_reader. default: None.
  • cache_path (str) – The path to cache the sampled data.
  • cached_id (int) – The id of dataset sampled. Evaluations with same cached_id use the same sampled dataset. default: 0.

paddleslim.common.controller module

The controller used to search hyperparameters or neural architecture

class paddleslim.common.controller.EvolutionaryController

Bases: object

Abstract controller for all evolutionary searching method.

next_tokens()

Generate new tokens.

Returns:The next searched tokens.
Return type:list<list>
reset(range_table, constrain_func=None)

Reset the controller.

Parameters:
  • range_table (list<int>) – It is used to define the searching space of controller. The tokens[i] generated by controller should be in [0, range_table[i]).
  • constrain_func (function) – It is used to check whether tokens meet the constraint. None means there is no constraint. Default: None.
update(tokens, reward)

Update the status of controller according current tokens and reward.

Parameters:
  • tokens (list<int>) – A solution of searching task.
  • reward (list<int>) – The reward of tokens.

paddleslim.common.controller_client module

class paddleslim.common.controller_client.ControllerClient(server_ip=None, server_port=None, key=None, client_name=None)

Bases: object

Controller client. :param server_ip: The ip that controller server listens on. None means getting the ip automatically. Default: None. :type server_ip: str :param server_port: The port that controller server listens on. 0 means getting usable port automatically. Default: 0. :type server_port: int :param key: The key used to identify legal agent for controller server. Default: “light-nas” :type key: str :param client_name: Current client name, random generate for counting client number. Default: None. :type client_name: str

next_tokens()

Get next tokens.

request_current_info()

Request for current information.

update(tokens, reward, iter)

Update the controller according to latest tokens and reward.

Parameters:
  • tokens (list<int>) – The tokens generated in last step.
  • reward (float) – The reward of tokens.
  • iter (int) – The iteration number of current client.

paddleslim.common.controller_server module

class paddleslim.common.controller_server.ControllerServer(controller=None, address=('', 0), max_client_num=100, search_steps=None, key=None)

Bases: object

The controller wrapper with a socket server to handle the request of search agent. :param controller: The controller used to generate tokens. :type controller: slim.searcher.Controller :param address: The address of current server binding with format (ip, port). Default: (‘’, 0).

which means setting ip automatically
Parameters:
  • max_client_num (int) – The maximum number of clients connecting to current server simultaneously. Default: 100.
  • search_steps (int|None) – The total steps of searching. None means never stopping. Default: None
  • key (str|None) – Config information. Default: None.
close()

Close the server.

ip()

Get the ip.

port()

Get the port.

run()

Start the server.

start()

paddleslim.common.lock module

paddleslim.common.lock.lock(file)

Lock the file in local file system.

paddleslim.common.lock.unlock(file)

Unlock the file in local file system.

paddleslim.common.log_helper module

paddleslim.common.log_helper.get_logger(name, level, fmt='%(asctime)s-%(levelname)s: %(message)s')

Get logger from logging with given name, level and format without setting logging basicConfig. For setting basicConfig in paddle will disable basicConfig setting after import paddle.

Parameters:
  • name (str) – The logger name.
  • level (logging.LEVEL) – The base level of the logger
  • fmt (str) – Format of logger output
Returns:

logging logger with given setttings
Return type:

logging.Logger

Examples:

logger = log_helper.get_logger(__name__, logging.INFO,
                     fmt='%(asctime)s-%(levelname)s: %(message)s')

paddleslim.common.sa_controller module

The controller used to search hyperparameters or neural architecture

class paddleslim.common.sa_controller.SAController(range_table=None, reduce_rate=0.85, init_temperature=None, max_try_times=300, init_tokens=None, reward=-1, max_reward=-1, iters=0, best_tokens=None, constrain_func=None, checkpoints=None, searched=None)

Bases: paddleslim.common.controller.EvolutionaryController

Simulated annealing controller.

Parameters:
  • range_table (list<int>) – Range table.
  • reduce_rate (float) – The decay rate of temperature.
  • init_temperature (float) – Init temperature.
  • max_try_times (int) – max try times before get legal tokens. Default: 300.
  • init_tokens (list<int>) – The initial tokens. Default: None.
  • reward (float) – The reward of current tokens. Default: -1.
  • max_reward (float) – The max reward in the search of sanas, in general, best tokens get max reward. Default: -1.
  • iters (int) – The iteration of sa controller. Default: 0.
  • best_tokens (list<int>) – The best tokens in the search of sanas, in general, best tokens get max reward. Default: None.
  • constrain_func (function) – The callback function used to check whether the tokens meet constraint. None means there is no constraint. Default: None.
  • checkpoints (str) – if checkpoint is None, donnot save checkpoints, else save scene to checkpoints file.
  • searched (dict<list, float>) – remember tokens which are searched.
best_tokens

Get current best tokens.

Returns:The best tokens.
Return type:list<int>
current_tokens

Get tokens generated in current searching step.

Returns:The best tokens.
Return type:list<int>
max_reward
next_tokens(control_token=None)

Get next tokens.

Parameters:control_token – The tokens used to generate next tokens.
Returns:The next tokens.
Return type:list<int>
update(tokens, reward, iter, client_num)

Update the controller according to latest tokens and reward.

Parameters:
  • tokens (list<int>) – The tokens generated in current step.
  • reward (float) – The reward of tokens.
  • iter (int) – The current step of searching client.
  • client_num (int) – The total number of searching client.