Datasets
Models
Downloads: 1
386 lines (342 with data), 18.7 kB
3
R└A]N Ń @ s\ d Z ddlmZ ddlmZ ddlmZ ddlm Z ddl
Z
ddl Z ddl m Z ddl
mZ dd l
mZ dd
lmZ ddlZddlZddlZddljZddl Z dd l T ddl Z dd l m!Z! ddl"Z"dd
l#m$Z$m%Z%m&Z&m'Z' e j(â ddl)m*Z*m+Z+ ddl,m-Z- dd
lmZ ddl.Z.G ddä dej/âZ0G ddä dej/âZ1d ddäZ2d ddäZ3G d d ä d âZ4dS ) zů
models.py
=======================
Houses all of the PyTorch models to access and the corresponding Scikit-Learn like model trainer.
Ú )┌UNet)┌
NestedUNet)┌ UNetSmall)┌
get_fast_scnnN)┌models)┌ segmentation)┌nn)┌
functional)┌*)┌Variable)┌ roc_curve┌confusion_matrix┌classification_report┌r2_score)┌GeneralizedDiceLoss┌ FocalLoss)┌ampc s" e Zd ZdZdç fddä Zç ZS )┌MLPa> Multi-layer perceptron model.
Parameters
----------
n_input:int
Number input dimensions.
hidden_topology:list
List of hidden topology
dropout_p:float
Amount dropout.
n_outputs:int
Number outputs.
binary:bool
Binary output with sigmoid transform.
softmax:bool
Whether to apply softmax on output.
Ú TFc
sţ t tłâjâ |g| |g ł_çfddättłjâd âD â}x |D ]}tjjj |j
â qFW ç fddä|D âł_ tj łjd łjd âł_
tjjj łj
j
â |rČtjâ } n|r║tjâ } n tjddŹ} łj jtjłj
| ââ tjłj Ä ł_d S )
Nc s( g | ] }t jł j| ł j|d âĹqS )r )r ┌Linear┌topology)┌.0┌i)┌selfę ˙B/Users/joshualevy/Documents/GitHub/PathFlowAI/pathflowai/models.py˙
<listcomp>4 s z MLP.__init__.<locals>.<listcomp>Ú c s& g | ] }t j|t jâ t jł d ŹâĹqS ))┌p)r ┌
Sequential┌ LeakyReLU┌Dropout)r ┌layer)┌ dropout_pr r r 7 s r g )r Ú■ Ú )┌superr ┌__init__r ┌range┌len┌torchr ┌init┌xavier_uniform_┌weight┌layersr Z output_layer┌Sigmoid┌Softmaxr! ┌appendr ┌mlp)
r Zn_inputZhidden_topologyr# ┌ n_outputs┌binary┌softmaxr. r" Zoutput_transform)┌ __class__)r# r r r' 1 s
z MLP.__init__)r TF)┌__name__┌
__module__┌ __qualname__┌__doc__r' ┌
__classcell__r r )r6 r r s r c s( e Zd ZdZç fddäZddä Zç ZS )┌FixedSegmentationModulez¬Special model modification for segmentation tasks. Gets output from some of the models' forward loops.
Parameters
----------
segnet:nn.Module
Segmentation network
c s t t| âjâ || _d S )N)r& r< r' ┌segnet)r r= )r6 r r r' K s z FixedSegmentationModule.__init__c C s | j |âd S )ztForward pass.
Parameters
----------
x:Tensor
Input
Returns
-------
Tensor
Output from model.
┌out)r= )r ┌xr r r ┌forwardO s z FixedSegmentationModule.forward)r7 r8 r9 r: r' r@ r; r r )r6 r r< C s r< TÚd Fc C s■ d}|dkr t d|dŹ}ÉnŮ|dkr2td|â}Én╚|dkrFt|â}Én┤|dkr\td|â}Én×|jdârČd d
lm} | rî|j|t|d Źd Ź}n|j |t|d Źd Ź}t
|â ÉnNd
dä t t âD â}ddä t t
âD â} ||krŔtt |â| dŹ}|Érp|| kÉr
tt
|â| dŹ}n
t ddŹ}|jdâÉrBtjd|d#d$dŹ|jd< t|â}n,|jdâÉr˙tjd|d%d&dŹ|jd< t|â}nŐ|jdâÉsł|jdâÉrČ|jj}
t|
dgd ||d d Źj|_nN|jd âÉsđ|jd âÉsđ|jd!âÉr˙|jd" j}
t|
dgd ||d d Źj|jd"< |S )'a Generate a nn.Module for use.
Parameters
----------
pretrain:bool
Pretrain using ImageNet?
architecture:str
See model_training for list of all architectures you can train with.
num_classes:int
Number of classes to predict.
add_sigmoid:type
Add sigmoid non-linearity at end.
n_hidden:int
Number of hidden fully connected layers.
segmentation:bool
Whether segment task?
Returns
-------
nn.Module
Pytorch model.
N┌unetÚ )┌
n_channels┌ n_classes┌unet2┌ fast_scnn┌ nested_unetZ efficientnetr )┌ EfficientNet)┌ num_classes)Zoverride_paramsc S s g | ]}|j d âs|ĹqS )┌__)┌
startswith)r ┌mr r r r î s z"generate_model.<locals>.<listcomp>c S s g | ]}|j d âs|ĹqS )rK )rL )r rM r r r r Ź s )┌
pretrained)rD ┌deeplabÚ r )┌ kernel_size┌strideÚ ┌fcni ┌resnet┌ inceptioniŔ g F)r# r3 r4 r5 ┌alexnet┌vggZ densenetsÚ )r r )r r )r r )r r )r ┌UNet2r r rL Zefficientnet_pytorchrI ┌from_pretrained┌dict┌ from_name┌print┌dirr ┌ segmodels┌getattrr ┌Conv2d┌
classifierr< ┌fc┌ in_featuresr r2 ) ┌pretrain┌ architecturerJ ┌ add_sigmoid┌n_hiddenr ┌modelrI Z model_namesZsegmentation_model_namesZnum_ftrsr r r ┌generate_model_ sH
$ rk šH»╝ܲÎz>c C sL | j d }|dkrÂtj|d â|jdâ }|jddddâjâ }|ddůddůddůddůf }|ddůddůddůddůf }tj||gddŹ}tj| â}d| }tj||gddŹ} n6tj|â|jdâ }|jddddâjâ }tj | ddŹ} |j
| j
â â}dt t d|j
â ââ }
tj| | |
â} tj| | |
â} d| | | jâ }
d|
S ) u▀ https://github.com/kevinzakka/pytorch-goodies
Computes the S├ŞrensenÔÇôDice loss.
Note that PyTorch optimizers minimize a loss. In this
case, we would like to maximize the dice loss so we
return the negated dice loss.
Args:
true: a tensor of shape [B, 1, H, W].
logits: a tensor of shape [B, C, H, W]. Corresponds to
the raw output or logits of the model.
eps: added to the denominator for numerical stability.
Returns:
dice_loss: the S├ŞrensenÔÇôDice loss.
r r rC r N)┌dimg @)r )┌shaper* ┌eye┌squeeze┌permute┌float┌cat┌sigmoid┌Fr5 ┌type┌tupler( ┌
ndimension┌sum┌mean)┌logits┌true┌epsrJ Z
true_1_hotZ true_1_hot_fZ true_1_hot_sZpos_probZneg_probZprobas┌dims┌ intersectionZ cardinality┌ dice_lossr r r rÇ Ę s&
rÇ c
@ s e Zd ZdZddeddddŹedd d
d d d
ŹdddfddäZddä Zddä Zddä Zddä Z ddä Z
d d ä Z d d ä Z d0d!d"äZ
d#d$ä Zd1d&d'äZd2d(d)äZd*d+ä Zd,d-ä Zd.d/ä ZdS )3┌ ModelTrainera+ Trainer for the neural network model that wraps it into a scikit-learn like interface.
Parameters
----------
model:nn.Module
Deep learning pytorch model.
n_epoch:int
Number training epochs.
validation_dataloader:DataLoader
Dataloader of validation dataset.
optimizer_opts:dict
Options for optimizer.
scheduler_opts:dict
Options for learning rate scheduler.
loss_fn:str
String to call a particular loss function for model.
reduction:str
Mean or sum reduction of loss.
num_train_batches:int
Number of training batches for epoch.
i, N┌adamgŘę˝ĎMbP?g-C ŰÔ6?)┌name┌lr┌ weight_decay┌
warm_restartsg Ó?Ú
gH»╝ܲÎj>r )┌ scheduler┌lr_scheduler_decay┌T_max┌eta_min┌T_mult┌cerz c
s || _ tjjtjjdť} tj|dŹtj|dŹtj|dŹtj |dŹt
t ddŹt ddŹdťë ç fdd äł d
< d t
|jâ âkrÇd |d < | |jd â | j jâ f|Ä| _tj| j | jd
dŹ\| _ | _t| j|dŹ| _|| _|| _ł | | _|| _| jdkp°| jjj| _ tj â | _ t j ł | â| _!|| _"d S )N)ré Zsgd)┌ reductionr )┌ num_classT)┌ add_softmax)┌bcerŹ ┌mse┌nll┌dice┌focal┌gdlc s t | |âł d | |â S )NrŹ )rÇ )┌y_pred┌y_true)┌loss_functionsr r ┌<lambda>Ű s z'ModelTrainer.__init__.<locals>.<lambda>zdice+cerâ ré ZO2)┌ opt_level)┌ optimizer┌optsrĹ )#rj r* ┌optim┌Adam┌SGDr ┌BCEWithLogitsLoss┌CrossEntropyLoss┌MSELoss┌NLLLossrÇ r r ┌list┌keys┌pop┌
parametersrť r ┌
initializeZ Schedulerrł ┌n_epoch┌validation_dataloader┌loss_fn┌ loss_fn_name┌dataset┌mt_bcerĹ r/ rt ┌copy┌deepcopy┌original_loss_fn┌num_train_batches)
r rj r¬ rź ┌optimizer_opts┌scheduler_optsrČ rÄ r│ Z
optimizersr )rÖ r r' Š s" @
zModelTrainer.__init__c C s | j ||âS )z┬Calculates loss supplied in init statement and modified by reweighting.
Parameters
----------
y_pred:tensor
Predictions.
y_true:tensor
True values.
Returns
-------
loss
)rČ )r rŚ rś r r r ┌ calc_loss˙ s zModelTrainer.calc_lossc C s | j ||âS )z╝Calculates loss supplied in init statement on validation set.
Parameters
----------
y_pred:tensor
Predictions.
y_true:tensor
True values.
Returns
-------
val_loss
)r▓ )r rŚ rś r r r ┌
calc_val_loss s zModelTrainer.calc_val_lossc C s | j | _dS )z'Resets loss to original specified loss.N)r▓ rČ )r r r r ┌
reset_loss_fn s zModelTrainer.reset_loss_fnc sé |j â ł _tjł jâł _tjł jtjdŹ}tj j
â r<|j â }ł j dkrVt j
|dŹł _n(ł j dkrpt j|dŹł _nç fddäł _dS )z░Updates loss function to handle class imbalance by weighting inverse to class appearance.
Parameters
----------
dataset:DynamicImageDataset
Dataset to balance by.
)┌dtyperŹ )r- rô c s t çç çfddätdâD ââS )Nc s> g | ]6}t ł|kârł j| ł jłł|k łł|k â ĹqS r )ry ┌
class_weightsr▓ )r r )r rŚ rś r r r 5 s zIModelTrainer.add_class_balance_loss.<locals>.<lambda>.<locals>.<listcomp>r )ry r( )rŚ rś )r )rŚ rś r rÜ 5 s z5ModelTrainer.add_class_balance_loss.<locals>.<lambda>N)┌get_class_weightsr║ r░ r▒ rČ r▓ r* ┌tensorrr ┌cuda┌ is_availablerş r ró rĄ )r r« r- r )r r ┌add_class_balance_loss" s
z#ModelTrainer.add_class_balance_lossc C sĺ t ||â\}}}|tjtjtjddgâtj||fâj d ddŹd â }||kjtâ}|t j
t ||âddgdd gd
Źj d d d ůd d d
ůf jfS )a= Calculate confusion matrix on validation set for classification/segmentation tasks, optimize threshold where positive.
Parameters
----------
y_pred:array
Predictions.
y_true:array
Ground truth.
Returns
-------
float
Optimized threshold to use on test set.
dataframe
Confusion matrix.
r r r )┌axisg Ó?ru ┌T˙-˙+)┌index┌columnsNr% r% )
r ┌np┌argminry ┌array┌vstackr┴ ┌astype┌int┌pd┌ DataFramer
┌iloc)r rŚ rś ┌fpr┌tpr┌
thresholds┌ thresholdr r r ┌calc_best_confusion7 s :z ModelTrainer.calc_best_confusionc C s& t j|| jâĆ}|jâ W dQ R X dS )zzBackprop using mixed precision for added speed boost.
Parameters
----------
loss:loss
Torch loss calculated.
N)r Z
scale_lossrť ┌backward)r ┌lossZ scaled_lossr r r ┌
loss_backwardN s zModelTrainer.loss_backwardc C s | j jdâ d}| jdkr*t|jâ|j n| j}x╚t|âD ]╝\}}||krLP t|d ddŹ}t|d â}|jjrä| j dkrä|j
dâ}t j j
â r×|j â }|j â }| j |â} | j| |â}
|
jâ } || 7 }| jjâ | j|
â | jjâ tdj|||| ââ q:W | jjâ || }|S ) zńOne training epoch, calculate predictions, loss, backpropagate.
Parameters
----------
epoch:int
Current epoch.
train_dataloader:DataLoader
Training data.
Returns
-------
float
Training loss for epoch
Tg Nr )┌
requires_gradr rö z Epoch {}[{}/{}] Train Loss:{})rj ┌trainr│ r) r« ┌
batch_size┌ enumerater r rş rp r* rŻ rż r ┌itemrť ┌ zero_gradrÍ ┌stepr^ ┌formatrł ) r ┌epoch┌train_dataloader┌ running_loss┌n_batchr ┌batch┌Xrś rŚ rŇ ┌
train_lossr r r ┌
train_loop[ s0
zModelTrainer.train_loopTc C s@ | j jdâ t|jâ|j }d}g g dť}tjâ ÉĆ┤ Éx¬t|âD É]ť\}} t| d ddŹ}
t| d â} |jj rä| j
dkrä| j dâ} tj j
â r×|
j â }
| j â } | j |
â} |Ér░|jj Ér |d jtj|jjsđ| n| âjâ jâ jâ jtâjâ â |d j| jâ jâ jâ jdd
Źjtâjâ â nö|d j| jâ jâ jâ jtâjâ â | jÉsL| n| j| âjâ jâ jâ jtâ}
t|
âdkÉr×|
jd dkÉr×|jj Ér×|
jdd
Ź}
|d j|
jâ â | j | | â}|j â }||7 }t d j ||||ââ qBW W d Q R X |Ér4|Ér4t j!|d ât j!|d â } } |jj És&| j
dkÉr||jj Ér|| j"| | â\}}t dj ||ââ t |â | jtâ} | |kjtâ} nĘ|jjÉr4t|jj#â}| | dk } | | dk } | t j$| âdk } | t j$| âdk } dÉr
|dkÉr
t| â| }| j%t|â|â} | j%t|â|â} t dj |t&t'| | ââââ nt t(| | ââ || }|S )aE Calculate loss over validation set.
Parameters
----------
epoch:int
Current epoch.
val_dataloader:DataLoader
Validation iterator.
print_val_confusion:bool
Calculate confusion matrix and plot.
save_predictions:int
Print validation results.
Returns
-------
float
Validation loss for epoch.
Fg )┌predr| r )rÎ r rö r| rš )r└ zEpoch {}[{}/{}] Val Loss:{}NrĹ rĺ z%Epoch {} Val Confusion, Threshold {}:z$Epoch {} Val Regression, R2 Score {})rĹ rĺ ))rj rě r) r« r┘ r* ┌no_gradr┌ r r rş rp rŻ rż r1 ┌flattenrľ ┌detach┌cpu┌numpyr╩ r╦ ┌argmaxrĹ rt rr rn r» rĚ r█ r^ rŮ rĂ ┌hstackrË ┌targets┌isnan┌reshape┌strr r )r r▀ Zval_dataloader┌print_val_confusion┌save_predictionsrÔ rß ┌Yr rŃ rń rś rŚ Z y_pred_numpyrŇ ┌val_lossrĎ Zbest_confusionZ n_targets┌n_rowr r r ┌val_loopů sd
6.$(* "
zModelTrainer.val_loopc C s g }d}t jâ Ć╚ x└t|âD ]┤\}\}}t jjâ r:|jâ }|jjrx| j|âjâ j â j
â j ddŹ}|j }|j
|jtââ q | j|â}t|jjâd sś| jrĄ| j|â}n|jjr║tj|ddŹ}|j
|jâ j â j
â â q W W dQ R X tj|ddŹ}|S )zźCalculate final predictions on loss.
Parameters
----------
test_dataloader:DataLoader
Test dataset.
Returns
-------
array
Predictions or embeddings.
g r )r└ )rm Nr )r* rŔ r┌ rŻ rż r« r rj rŕ rŰ rý rÝ rn r1 r╩ r╦ r) r´ rĹ rt ┌classify_annotationsru r5 rĂ ┌ concatenate) r ┌test_dataloaderrŚ rß r rń ┌y_test┌
predictionZ pred_sizer r r ┌ test_loop¤ s$
$zModelTrainer.test_loopFc C sŔ g | _ g | _xĂt| jâD ]Ş} tjâ }
| j| |â} tjâ } | |
}
| j j| â | j| | j||dŹ}tjâ | }| jj|â |r¬| | r¬|rö| j |â t
dj | | ||
|ââ |t | jâkr|r|}| }t
j| jâ}qW |rŮ|| _| ||fS )aŃ Fits the segmentation or classification model to the patches, saving the model with the lowest validation score.
Parameters
----------
train_dataloader:DataLoader
Training dataset.
verbose:bool
Print training and validation loss?
print_every:int
Number of epochs until print?
save_model:bool
Whether to save model when reaching lowest validation loss.
plot_training_curves:bool
Plot training curves over epochs.
plot_save_file:str
File to save training curves.
print_val_confusion:bool
Print validation confusion matrix.
save_val_predictions:bool
Print validation results.
Returns
-------
self
Trainer.
float
Minimum val loss.
int
Best validation epoch with lowest loss.
)rˇ r˘ z@Epoch {}: Train Loss {}, Val Loss {}, Train Time {}, Val Time {})┌ train_losses┌
val_lossesr( r¬ ┌timerŠ r1 r° rź ┌plot_train_val_curvesr^ rŮ ┌minr░ r▒ rj )r rÓ ┌verbose┌ print_everyZ
save_model┌plot_training_curves┌plot_save_filerˇ ┌save_val_predictionsr▀ ┌
start_timerň ┌ current_timeZ
train_timer÷ Zval_timeZ min_val_lossZ
best_epoch┌
best_modelr r r ┌fit˘ s, !
zModelTrainer.fitc C sr t jâ tjdddtjtjtjt | j
ââ| j
| j fâj dddgdŹj
dgddgdŹdŹ |d k rnt j|d
d Ź d S ) zhPlots training and validation curves.
Parameters
----------
save_file:str
File to save to.
r▀ ┌value┌variablerě ┌val)r┼ )┌id_vars┌
value_vars)┌hue┌dataNi, )┌dpi)┌plt┌figure┌sns┌lineplotr╠ r═ rĂ r╔ ┌aranger) r r r┴ ┌melt┌savefig)r ┌ save_filer r r r , s
$"z"ModelTrainer.plot_train_val_curvesc C s | j |â}|S )z┤Make classification segmentation predictions on testing data.
Parameters
----------
test_dataloader:DataLoader
Test data.
Returns
-------
array
Predictions.
)r■ )r rű rŚ r r r ┌predict< s
zModelTrainer.predictc C s | j |âd j|âS )a Fit model to training data and make classification segmentation predictions on testing data.
Parameters
----------
train_dataloader:DataLoader
Train data.
test_dataloader:DataLoader
Test data.
Returns
-------
array
Predictions.
r )r r )r rÓ rű r r r ┌ fit_predictM s zModelTrainer.fit_predictc C s | j S )zReturns pytorch model.
)rj )r r r r ┌ return_model_ s zModelTrainer.return_model)TT)Frç TFNTT)N)r7 r8 r9 r: r\ r' r rĚ rŞ r┐ rË rÍ rŠ r° r■ r r r r r r r r r rü đ s 0
*
J%
8
rü )TrA F)rl )5r: Zpathflowai.unetr Zpathflowai.unet2r Zpathflowai.unet4r rZ Zpathflowai.fast_scnnr r* ┌ torchvisionr Ztorchvision.modelsr r` r ┌torch.nnr ru ┌pandasr╠ rý rĂ ┌
matplotlib┌matplotlib.pyplot┌pyplotr ┌seabornr ┌pathflowai.schedulers┌ pysnooperZtorch.autogradr r░ Zsklearn.metricsr r
r r ┌set┌pathflowai.lossesr r Zapexr r ┌Moduler r< rk rÇ rü r r r r ┌<module> s8 %
I
(