a
d @ s d dl Z d dlmZmZmZ d dlZd dlmZ d dlm Z zd dl
m Z W n e yb dZ Y n0 d dl
mZ ddlmZmZmZ g dZG d d
d
ZdS ) N)CallableDictOptional)
DataLoader) Accelerator) AverageMeter )
Configuration StatsTracker TrainerState)
preprocess
augmentationsevaluatefit fit_epochon_epoch_starton_before_modelon_after_model
on_checkpoint on_epoch_endc @ s< e Zd ZdZi fejeeeej ej j
ej j j e
eeef dd ddZeejdddZejdd d
d Zedd d
dZddddZe edddZdd Z eedddZ eedddZ ejeee f ddd Z ejejdd d Z!eedd d Z"ed!d"d#Z#ejed$d%d&Z$d'd( Z%d)d* Z&dS )+Trainera Base class to train the different models in kornia.
.. warning::
The API is experimental and subject to be modified based on the needs of kornia models.
Args:
model: the nn.Module to be optimized.
train_dataloader: the data loader used in the training loop.
valid_dataloader: the data loader used in the validation loop.
criterion: the nn.Module with the function that computes the loss.
optimizer: the torch optimizer object to be used during the optimization.
scheduler: the torch scheduler object with defiing the scheduling strategy.
accelerator: the Accelerator object to distribute the training.
config: a TrainerConfiguration structure containing the experiment hyper parameters.
callbacks: a dictionary containing the pointers to the functions to overrides. The
main supported hooks are ``evaluate``, ``preprocess``, ``augmentations`` and ``fit``.
.. important::
The API heavily relies on `accelerate `_.
In order to use it, you must: ``pip install kornia[x]``
.. seealso::
Learn how to use the API in our documentation
`here `_.
N) modeltrain_dataloadervalid_dataloader criterion optimizer schedulerconfig callbacksreturnc C s t d u rtdt | _| j|| _| j|| _| j|| _|d u rNd n
|| j| _ | j|| _
|| _ || _ |
D ],\} }
| tvrtd| dtt| |
q~|j| _tj| _td| _d S )Nz s" zTrainer.__init__)r c C s | j jS N)r# r& )r0 r3 r3 r4 r& d s zTrainer.device)lossr c C s | j | d S r6 )r# backward)r0 r7 r3 r3 r4 r8 h s zTrainer.backward)epochr c C s | j t }t| jD ]\}}|d |d d}| j | |}| |}| |}|
| j |}| || | ||d }|
| | j || t|d |d dkr| jd|d d| j d |d dt| j d
|jd d |jd qd S )
Nr r inputtargetr< r; 2 zTrain: /z
Sample: z Loss: z.3f )r r! r enumerater r Z zero_gradr r
r on_modelr compute_lossr8 stepupdateitemlenr/ infor+ valavg)r0 r9 losses sample_idsampleoutputr7 r3 r3 r4 r k s4
zTrainer.fit_epochc C sh t | jD ]X}tj| _| | tj| _| }| | j || |
| jtj krX qd| j
q
d S r6 )ranger+ r ZTRAININGr, r ZVALIDATEr r r r Z TERMINATEr rC )r0 r9 Z valid_statsr3 r3 r4 r s
z Trainer.fitc C s | j t }t| jD ]\}}|d |d d}| |}| |}| | j |}| || t |d }| j
d ur| ||d }| d|
| || ||d | |d dkr| jd| d t | j d
| q| S ) Nr r r: r; r< rJ
zTest: r> r? )r evalr
r@ r r r rA r rF r rB rD rE Zupdate_from_dictcompute_metricsr/ rG Zas_dict)r0 statsrK rL outZ
batch_sizeZval_lossr3 r3 r4 r s
&zTrainer.evaluatec O s d S r6 r3 r0 argskwargsr3 r3 r4 r s zTrainer.on_epoch_start)xr c C s |S r6 r3 r0 rW r3 r3 r4 r s zTrainer.preprocessc C s |S r6 r3 rX r3 r3 r4 r
s zTrainer.augmentations)rU r c G s i S )z&Compute metrics during the evaluation.r3 r0 rU r3 r3 r4 rQ s zTrainer.compute_metricsc G s | j d u rtd| j | S )Nz `criterion` should not be None.)r RuntimeErrorrY r3 r3 r4 rB s
zTrainer.compute_lossc C s |S r6 r3 rX r3 r3 r4 r s zTrainer.on_before_model)rL c C s ||d S )Nr; r3 )r0 r rL r3 r3 r4 rA s zTrainer.on_model)rM rL c C s d S r6 r3 )r0 rM rL r3 r3 r4 r s zTrainer.on_after_modelc O s d S r6 r3 rT r3 r3 r4 r s zTrainer.on_checkpointc O s d S r6 r3 rT r3 r3 r4 r s zTrainer.on_epoch_end)'__name__
__module__ __qualname____doc__nnModuler r torchZoptimZ OptimizerZ lr_schedulerZCosineAnnealingLRr r strr r5 propertyr& ZTensorr8 intr r Zno_graddictr r r r
floatrQ rB r rA r r r r3 r3 r3 r4 r # s< #
&r )r- typingr r r ra Ztorch.nnr_ Ztorch.utils.datar Z
accelerater ImportErrorZkornia.metricsr utilsr r
r r( r r3 r3 r3 r4 s