a
hK @ s d Z ddlZddlZddlZddlmZ ddlmZmZ ddlZ ddl
m Z ddl m
Z
mZmZmZ ddlmZ dd lmZ eeZeG d
d d ZG d d
d
ZeG dd deZG dd dZG dd deZ G dd deZ G dd deZ G dd deZ G dd deeZ dS ) zJ
Callbacks to use with the Trainer class and customize the training loop.
N) dataclass)OptionalUnion)tqdm )HPSearchBackendIntervalStrategy SaveStrategy
has_length)TrainingArguments)loggingc @ sv e Zd ZU dZdZee ed< dZe ed< dZ
e ed< dZ e ed< dZ e ed < dZ
e ed
< dZee ed < dZe ed < dZe ed
< dZeed< dZeeeef ed< dZee ed< dZee ed< dZee ed< dZeed< dZeed< dZ eed< dZ ee ed< dZ eee eee ef f ed< dZ ed ed< d d Z!ed d d Z"e#ed d!d"Z$d#d$ Z%d%d& Z&dS )' TrainerStatea
A class containing the [`Trainer`] inner state that will be saved along the model and optimizer when checkpointing
and passed to the [`TrainerCallback`].
In all this class, one step is to be understood as one update step. When using gradient accumulation, one update
step may require several forward and backward passes: if you use `gradient_accumulation_steps=n`, then one update
step requires going through *n* batches.
Args:
epoch (`float`, *optional*):
Only set during training, will represent the epoch the training is at (the decimal part being the
percentage of the current epoch completed).
global_step (`int`, *optional*, defaults to 0):
During training, represents the number of update steps completed.
max_steps (`int`, *optional*, defaults to 0):
The number of update steps to do during the current training.
logging_steps (`int`, *optional*, defaults to 500):
Log every X updates steps
eval_steps (`int`, *optional*):
Run an evaluation every X steps.
save_steps (`int`, *optional*, defaults to 500):
Save checkpoint every X updates steps.
train_batch_size (`int`, *optional*):
The batch size for the training dataloader. Only needed when
`auto_find_batch_size` has been used.
num_input_tokens_seen (`int`, *optional*, defaults to 0):
When tracking the inputs tokens, the number of tokens seen during training (number of input tokens, not the
number of prediction tokens).
total_flos (`float`, *optional*, defaults to 0):
The total number of floating operations done by the model since the beginning of training (stored as floats
to avoid overflow).
log_history (`list[dict[str, float]]`, *optional*):
The list of logs done since the beginning of training.
best_metric (`float`, *optional*):
When tracking the best model, the value of the best metric encountered so far.
best_global_step (`int`, *optional*):
When tracking the best model, the step at which the best metric was encountered.
Used for setting `best_model_checkpoint`.
best_model_checkpoint (`str`, *optional*):
When tracking the best model, the value of the name of the checkpoint for the best model encountered so
far.
is_local_process_zero (`bool`, *optional*, defaults to `True`):
Whether or not this process is the local (e.g., on one machine if training in a distributed fashion on
several machines) main process.
is_world_process_zero (`bool`, *optional*, defaults to `True`):
Whether or not this process is the global main process (when training in a distributed fashion on several
machines, this is only going to be `True` for one process).
is_hyper_param_search (`bool`, *optional*, defaults to `False`):
Whether we are in the process of a hyper parameter search using Trainer.hyperparameter_search. This will
impact the way data will be logged in TensorBoard.
stateful_callbacks (`list[StatefulTrainerCallback]`, *optional*):
Callbacks attached to the `Trainer` that should have their states be saved or restored.
Relevant callbacks should implement a `state` and `from_state` function.
Nepochr global_step max_stepsi
logging_steps
eval_steps
save_stepstrain_batch_sizenum_train_epochsnum_input_tokens_seen
total_flos log_history best_metricbest_global_stepbest_model_checkpointTis_local_process_zerois_world_process_zeroFis_hyper_param_search
trial_name trial_paramsTrainerCallbackstateful_callbacksc C s | j d u rg | _ | jd u r"i | _nt| jtr0n~i }| jD ]l}t|tsZtdt| |jj}||v rt|| t s|| g||< ||
| q:| ||< q:|| _d S )NzNAll callbacks passed to be saved must inherit `ExportableState`, but received ) r r"
isinstancedictExportableState TypeErrortype __class____name__listappendstate)selfr" callbackname r0 Y/var/www/html/assistant/venv/lib/python3.9/site-packages/transformers/trainer_callback.py
__post_init__u s&
zTrainerState.__post_init__) json_pathc C sV t jt| dddd }t|ddd}|| W d n1 sH0 Y dS ) zDSave the content of this instance in JSON format inside `json_path`. T)indent sort_keys
wutf-8encodingN)jsondumps dataclassesasdictopenwrite)r- r3 Z json_stringfr0 r0 r1 save_to_json s zTrainerState.save_to_jsonc C sH t |dd}| }W d n1 s*0 Y | f i t|S )z3Create an instance from the content of `json_path`.r9 r: N)r@ readr< loads)clsr3 rB textr0 r0 r1 load_from_json s &zTrainerState.load_from_jsonc C sN dD ]D}t || d}|dur|dk r6t|| }t| | d| qdS )z
Calculates and stores the absolute value for logging,
eval, and save steps based on if it was a proportion
or not.
)r evalsaveZ_stepsNr )getattrmathceilsetattr)r- argsr Z step_kindZ num_stepsr0 r0 r1
compute_steps s zTrainerState.compute_stepsc C s |j dur"|jdur"| |j| _d| _|dur\ddlm} |jtjkrN|j n|}||| _|| _
|| _ | | _ |
| _
dS )zI
Stores the initial training references needed in `self`
Nr ) hp_params)Zhp_nameZ_trialr r Ztransformers.integrationsrQ Zhp_search_backendr ZSIGOPT assignmentsr r r r )r- Ztrainerr r ZtrialrQ rR r0 r0 r1 init_training_references s
z%TrainerState.init_training_references)'r)
__module__ __qualname____doc__r r float__annotations__r intr r r r r r r r r r* r$ strr r r r boolr r r r r r" r2 rC classmethodrH rP rS r0 r0 r0 r1 r
# s6
;
r
c @ s* e Zd ZdZedddZedd ZdS )r% aj
A class for objects that include the ability to have its state
be saved during `Trainer._save_checkpoint` and loaded back in during
`Trainer._load_from_checkpoint`.
These must implement a `state` function that gets called during the respective
Trainer function call. It should only include parameters and attributes needed to
recreate the state at a particular time, to avoid utilizing pickle/maintain standard
file IO writing.
Example:
```python
class EarlyStoppingCallback(TrainerCallback, ExportableState):
def __init__(self, early_stopping_patience: int = 1, early_stopping_threshold: Optional[float] = 0.0):
self.early_stopping_patience = early_stopping_patience
self.early_stopping_threshold = early_stopping_threshold
# early_stopping_patience_counter denotes the number of times validation metrics failed to improve.
self.early_stopping_patience_counter = 0
def state(self) -> dict:
return {
"args": {
"early_stopping_patience": self.early_stopping_patience,
"early_stopping_threshold": self.early_stopping_threshold,
},
"attributes": {
"early_stopping_patience_counter": self.early_stopping_patience_counter,
}
}
```returnc C s t dd S )NzInternal class that just calls the list of callbacks in order.c C sf g | _ |D ]}| | q
|| _|| _|| _|| _d | _d | _tdd | j D sbt
d| j d S )Nc s s | ]}t |tV qd S N)r# DefaultFlowCallback.0cbr0 r0 r1 z+CallbackHandler.__init__..zThe Trainer will not work properly if you don't have a `DefaultFlowCallback` in its callbacks. You
should add one before training with `trainer.add_callback(DefaultFlowCallback). The current list ofcallbacks is
:) callbacks add_callbackmodelprocessing_class optimizer lr_schedulertrain_dataloadereval_dataloaderanyloggerwarning
callback_list)r- r r r r r r r0 r0 r1 __init__ s zCallbackHandler.__init__c C sh t |tr| n|}t |tr"|n|j}|dd | jD v rXtd| dd | j | j| d S )Nc S s g | ]
}|j qS r0 )r( )r cr0 r0 r1
r z0CallbackHandler.add_callback..zYou are adding a zH to the callbacks of this Trainer, but there is already one. The currentzlist of callbacks is
:)r# r' r( r r r r r+ )r- r. r Zcb_classr0 r0 r1 r s
z CallbackHandler.add_callbackc C sb t |tr6| jD ]"}t ||r| j| | S qn(| jD ] }||kr<| j| | S q.)joinr r` r0 r0 r1 r s z CallbackHandler.callback_listrr c C s | d|||S )Nrv
call_eventr- rO r, rs r0 r0 r1 rv s zCallbackHandler.on_init_endc C s d|_ | d|||S )NFrw )rh r r r0 r0 r1 rw s z CallbackHandler.on_train_beginc C s | d|||S )Nrx r r r0 r0 r1 rx s z CallbackHandler.on_train_endc C s d|_ | d|||S )NFry )ri r r r0 r0 r1 ry s z CallbackHandler.on_epoch_beginc C s | d|||S )Nrz r r r0 r0 r1 rz s z CallbackHandler.on_epoch_endc C s" d|_ d|_d|_| d|||S )NFr{ )rl rk rj r r r0 r0 r1 r{ s z CallbackHandler.on_step_beginc C s | d|||S )Nr| r r r0 r0 r1 r| s z%CallbackHandler.on_pre_optimizer_stepc C s | d|||S )Nr} r r r0 r0 r1 r} s z!CallbackHandler.on_optimizer_stepc C s | d|||S )Nr~ r r r0 r0 r1 r~ s z CallbackHandler.on_substep_endc C s | d|||S )Nr r r r0 r0 r1 r s zCallbackHandler.on_step_endc C s d|_ | jd||||dS )NFr r )rk r r- rO r, rs r r0 r0 r1 r s zCallbackHandler.on_evaluatec C s | j d||||dS )Nr r r r r0 r0 r1 r s zCallbackHandler.on_predictc C s d|_ | d|||S )NFr )rj r r r0 r0 r1 r s zCallbackHandler.on_savec C s d|_ | jd||||dS )NFr )logs)rl r )r- rO r, rs r r0 r0 r1 r # s zCallbackHandler.on_logc C s | d|||S )Nr r r r0 r0 r1 r ' s z"CallbackHandler.on_prediction_stepc
K sP | j D ]D}t|||||f| j| j| j| j| j| jd|}|d ur|}q|S )N)r r r r r r )r rK r r r r r r )r- eventrO r, rs ru r. resultr0 r0 r1 r * s$
zCallbackHandler.call_eventN) r) rT rU rV r r r r propertyr r r
rg rv rw rx ry rz r{ r| r} r~ r r r r r r r r0 r0 r0 r1 r s.
r c @ s4 e Zd ZdZeeedddZeeedddZdS )r zx
A [`TrainerCallback`] that handles the default flow of the training loop for logs, evaluation and checkpoints.
rr c K s |j dkr|jrd|_|jtjkr8|j |j dkr8d|_|jtjkrf|j |j dkrf|j |j krfd|_
|j t jkr|j
dkr|j |j
dkrd|_|j |jkrd|_|j t jkrd|_|S )Nr Tr )r Zlogging_first_steprl logging_strategyr ZSTEPSr
eval_strategyr
eval_delayrk
save_strategyr r rj r rh rt r0 r0 r1 r C s.
z DefaultFlowCallback.on_step_endc K sF |j tjkrd|_|jtjkr0|j|jkr0d|_|jt jkrBd|_
|S )NT) r r EPOCHrl r r r rk r r rj rt r0 r0 r1 rz c s z DefaultFlowCallback.on_epoch_endN) r) rT rU rV r r
rg r rz r0 r0 r0 r1 r > s r c @ s\ e Zd ZdZdedddZdd Zdd Zdd d Zd
d Z dd Z
dddZ dd Z d
S )ProgressCallbackz
A [`TrainerCallback`] that displays the progress of training or evaluation.
You can modify `max_str_len` to control how long strings are truncated when logging.
d ) max_str_lenc C s d| _ d| _|| _dS )a!
Initialize the callback with optional max_str_len parameter to control string truncation length.
Args:
max_str_len (`int`):
Maximum length of strings to display in logs.
Longer strings will be truncated with a message.
N) training_barprediction_barr )r- r r0 r0 r1 r y s zProgressCallback.__init__c K s |j rt|jdd| _d| _d S )NT)total
dynamic_ncolsr )r r r r current_steprt r0 r0 r1 rw s z ProgressCallback.on_train_beginc K s& |j r"| j|j| j |j| _d S r )r r updater r rt r0 r0 r1 r s z ProgressCallback.on_step_endNc K sB |j r>t|r>| jd u r2tt|| jd u dd| _| jd d S )NT)r Zleaver r )r r
r r lenr r )r- rO r, rs r ru r0 r0 r1 r s
z#ProgressCallback.on_prediction_stepc K s$ |j r | jd ur| j d | _d S r r r closert r0 r0 r1 r s
z ProgressCallback.on_evaluatec K s$ |j r | jd ur| j d | _d S r r rt r0 r0 r1 r s
zProgressCallback.on_predictc
K s |j r| jd uri }| D ]F\}}t|trZt|| jkrZdt| d| j d||< q |||< q |dd } d|v rt|d d|d< | j t| d S )Nz%[String too long to display, length: z > z/. Consider increasing `max_str_len` if needed.]r r r4 )
r r rb r# rZ r r poproundrA )
r- rO r, rs r ru Z shallow_logsrd re _r0 r0 r1 r s
zProgressCallback.on_logc K s |j r| j d | _d S r )r r r rt r0 r0 r1 rx s
z ProgressCallback.on_train_end)r )N)N)
r) rT rU rV rY r rw r r r r r rx r0 r0 r0 r1 r s s
r c @ s e Zd ZdZdddZdS )PrinterCallbackz?
A bare [`TrainerCallback`] that just prints the logs.
Nc K s | dd }|jrt| d S )Nr )r r print)r- rO r, rs r ru r r0 r0 r1 r s zPrinterCallback.on_log)N)r) rT rU rV r r0 r0 r0 r1 r s r c @ sL e Zd ZdZdeee dddZdd Zd d
Z d d Z
e d
ddZ dS )EarlyStoppingCallbacka1
A [`TrainerCallback`] that handles early stopping.
Args:
early_stopping_patience (`int`):
Use with `metric_for_best_model` to stop training when the specified metric worsens for
`early_stopping_patience` evaluation calls.
early_stopping_threshold(`float`, *optional*):
Use with TrainingArguments `metric_for_best_model` and `early_stopping_patience` to denote how much the
specified metric must improve to satisfy early stopping conditions. `
This callback depends on [`TrainingArguments`] argument *load_best_model_at_end* functionality to set best_metric
in [`TrainerState`]. Note that if the [`TrainingArguments`] argument *save_steps* differs from *eval_steps*, the
early stopping will not occur until the next save step.
r early_stopping_patienceearly_stopping_thresholdc C s || _ || _d| _d S )Nr r r early_stopping_patience_counter)r- r r r0 r0 r1 r s z EarlyStoppingCallback.__init__c C sV |j r tjntj}|jd u s<|||jrDt||j | jkrDd| _n| jd7 _d S )Nr r )Zgreater_is_betternpZgreaterlessr absr r )r- rO r, rs metric_valueoperatorr0 r0 r1 check_metric_value s
z(EarlyStoppingCallback.check_metric_valuec K s: |j std |jd us"J d|jtjks6J dd S )NzUsing EarlyStoppingCallback without load_best_model_at_end=True. Once training is finished, the best model will not be loaded automatically.zBEarlyStoppingCallback requires metric_for_best_model to be definedzAEarlyStoppingCallback requires IntervalStrategy of steps or epoch)Zload_best_model_at_endr r metric_for_best_modelr r NOrt r0 r0 r1 rw s z$EarlyStoppingCallback.on_train_beginc K sh |j }|dsd| }||}|d u rBtd| d d S | |||| | j| jkrdd|_d S )NZeval_z@early stopping required metric_for_best_model, but did not find z so early stopping is disabledT) r
startswithgetr r r r r rh )r- rO r, rs r ru Zmetric_to_checkr r0 r0 r1 r s
z!EarlyStoppingCallback.on_evaluater] c C s | j | jdd| jidS )Nr r rq r r` r0 r0 r1 r, s zEarlyStoppingCallback.stateN)r r )
r) rT rU rV rY r rW r r rw r r$ r, r0 r0 r0 r1 r s
r )!rV r> r< rL r typingr r numpyr Z tqdm.autor Z
trainer_utilsr r r r
Z
training_argsr utilsr Z
get_loggerr) r r
r% rg r! r r r r r r0 r0 r0 r1 s2
,= 5J