Variational Autoencoder

Variational Autoencoder#

Hide code cell source

 
import logging
from functools import partial

import pandas as pd
import sklearn
import torch
from fastai import learner
from fastai.basics import *
from fastai.callback.all import *
from fastai.callback.all import EarlyStoppingCallback
from fastai.learner import Learner
from fastai.torch_basics import *
from IPython.display import display
from sklearn.impute import SimpleImputer
from sklearn.preprocessing import StandardScaler
from torch.nn import Sigmoid

import pimmslearn
import pimmslearn.model
import pimmslearn.models as models
import pimmslearn.nb
from pimmslearn.analyzers import analyzers
from pimmslearn.io import datasplits
# overwriting Recorder callback with custom plot_loss
from pimmslearn.models import ae, plot_loss

learner.Recorder.plot_loss = plot_loss


logger = pimmslearn.logging.setup_logger(logging.getLogger('pimmslearn'))
logger.info(
    "Experiment 03 - Analysis of latent spaces and performance comparisions")

figures = {}  # collection of ax or figures

pimmslearn - INFO     Experiment 03 - Analysis of latent spaces and performance comparisions

Hide code cell source

 
# catch passed parameters
args = None
args = dict(globals()).keys()

Papermill script parameters:

# files and folders
# Datasplit folder with data for experiment
folder_experiment: str = 'runs/example'
folder_data: str = ''  # specify data directory if needed
file_format: str = 'csv'  # file format of create splits, default pickle (pkl)
# Machine parsed metadata from rawfile workflow
fn_rawfile_metadata: str = 'data/dev_datasets/HeLa_6070/files_selected_metadata_N50.csv'
# training
epochs_max: int = 50  # Maximum number of epochs
batch_size: int = 64  # Batch size for training (and evaluation)
cuda: bool = True  # Whether to use a GPU for training
# model
# Dimensionality of encoding dimension (latent space of model)
latent_dim: int = 25
# A underscore separated string of layers, '256_128' for the encoder, reverse will be use for decoder
hidden_layers: str = '256_128'
# force_train:bool = True # Force training when saved model could be used. Per default re-train model
patience: int = 50  # Patience for early stopping
sample_idx_position: int = 0  # position of index which is sample ID
model: str = 'VAE'  # model name
model_key: str = 'VAE'  # potentially alternative key for model (grid search)
save_pred_real_na: bool = True  # Save all predictions for missing values
# metadata -> defaults for metadata extracted from machine data
meta_date_col: str = None  # date column in meta data
meta_cat_col: str = None  # category column in meta data

# Parameters
model = "VAE"
latent_dim = 10
batch_size = 64
epochs_max = 300
hidden_layers = "64"
sample_idx_position = 0
cuda = False
save_pred_real_na = True
fn_rawfile_metadata = "https://raw.githubusercontent.com/RasmussenLab/njab/HEAD/docs/tutorial/data/alzheimer/meta.csv"
folder_experiment = "runs/alzheimer_study"
model_key = "VAE"

Some argument transformations

Hide code cell source

 
args = pimmslearn.nb.get_params(args, globals=globals())
args

{'folder_experiment': 'runs/alzheimer_study',
 'folder_data': '',
 'file_format': 'csv',
 'fn_rawfile_metadata': 'https://raw.githubusercontent.com/RasmussenLab/njab/HEAD/docs/tutorial/data/alzheimer/meta.csv',
 'epochs_max': 300,
 'batch_size': 64,
 'cuda': False,
 'latent_dim': 10,
 'hidden_layers': '64',
 'patience': 50,
 'sample_idx_position': 0,
 'model': 'VAE',
 'model_key': 'VAE',
 'save_pred_real_na': True,
 'meta_date_col': None,
 'meta_cat_col': None}

Hide code cell source

 
args = pimmslearn.nb.args_from_dict(args)

if isinstance(args.hidden_layers, str):
    args.overwrite_entry("hidden_layers", [int(x)
                         for x in args.hidden_layers.split('_')])
else:
    raise ValueError(
        f"hidden_layers is of unknown type {type(args.hidden_layers)}")
args

{'batch_size': 64,
 'cuda': False,
 'data': Path('runs/alzheimer_study/data'),
 'epochs_max': 300,
 'file_format': 'csv',
 'fn_rawfile_metadata': 'https://raw.githubusercontent.com/RasmussenLab/njab/HEAD/docs/tutorial/data/alzheimer/meta.csv',
 'folder_data': '',
 'folder_experiment': Path('runs/alzheimer_study'),
 'hidden_layers': [64],
 'latent_dim': 10,
 'meta_cat_col': None,
 'meta_date_col': None,
 'model': 'VAE',
 'model_key': 'VAE',
 'out_figures': Path('runs/alzheimer_study/figures'),
 'out_folder': Path('runs/alzheimer_study'),
 'out_metrics': Path('runs/alzheimer_study'),
 'out_models': Path('runs/alzheimer_study'),
 'out_preds': Path('runs/alzheimer_study/preds'),
 'patience': 50,
 'sample_idx_position': 0,
 'save_pred_real_na': True}

Some naming conventions

Hide code cell source

 
TEMPLATE_MODEL_PARAMS = 'model_params_{}.json'

Load data in long format#

Hide code cell source

 
data = datasplits.DataSplits.from_folder(
    args.data, file_format=args.file_format)

pimmslearn.io.datasplits - INFO     Loaded 'train_X' from file: runs/alzheimer_study/data/train_X.csv
pimmslearn.io.datasplits - INFO     Loaded 'val_y' from file: runs/alzheimer_study/data/val_y.csv
pimmslearn.io.datasplits - INFO     Loaded 'test_y' from file: runs/alzheimer_study/data/test_y.csv

data is loaded in long format

Hide code cell source

 
data.train_X.sample(5)

Sample ID   protein groups
Sample_140  Q96S96           19.588
Sample_127  O14498           20.153
Sample_060  P54802           17.842
Sample_143  Q16769           19.584
Sample_180  Q8IW52           15.998
Name: intensity, dtype: float64

Infer index names from long format

Hide code cell source

 
index_columns = list(data.train_X.index.names)
sample_id = index_columns.pop(args.sample_idx_position)
if len(index_columns) == 1:
    index_column = index_columns.pop()
    index_columns = None
    logger.info(f"{sample_id = }, single feature: {index_column = }")
else:
    logger.info(f"{sample_id = }, multiple features: {index_columns = }")

if not index_columns:
    index_columns = [sample_id, index_column]
else:
    raise NotImplementedError(
        "More than one feature: Needs to be implemented. see above logging output.")

pimmslearn - INFO     sample_id = 'Sample ID', single feature: index_column = 'protein groups'

load meta data for splits

Hide code cell source

 
if args.fn_rawfile_metadata:
    df_meta = pd.read_csv(args.fn_rawfile_metadata, index_col=0)
    display(df_meta.loc[data.train_X.index.levels[0]])
else:
    df_meta = None
_collection site _age at CSF collection _gender _t-tau [ng/L] _p-tau [ng/L] _Abeta-42 [ng/L] _Abeta-40 [ng/L] _Abeta-42/Abeta-40 ratio _primary biochemical AD classification _clinical AD diagnosis _MMSE score
Sample ID
Sample_000 Sweden 71.000 f 703.000 85.000 562.000 NaN NaN biochemical control NaN NaN
Sample_001 Sweden 77.000 m 518.000 91.000 334.000 NaN NaN biochemical AD NaN NaN
Sample_002 Sweden 75.000 m 974.000 87.000 515.000 NaN NaN biochemical AD NaN NaN
Sample_003 Sweden 72.000 f 950.000 109.000 394.000 NaN NaN biochemical AD NaN NaN
Sample_004 Sweden 63.000 f 873.000 88.000 234.000 NaN NaN biochemical AD NaN NaN
... ... ... ... ... ... ... ... ... ... ... ...
Sample_205 Berlin 69.000 f 1,945.000 NaN 699.000 12,140.000 0.058 biochemical AD AD 17.000
Sample_206 Berlin 73.000 m 299.000 NaN 1,420.000 16,571.000 0.086 biochemical control non-AD 28.000
Sample_207 Berlin 71.000 f 262.000 NaN 639.000 9,663.000 0.066 biochemical control non-AD 28.000
Sample_208 Berlin 83.000 m 289.000 NaN 1,436.000 11,285.000 0.127 biochemical control non-AD 24.000
Sample_209 Berlin 63.000 f 591.000 NaN 1,299.000 11,232.000 0.116 biochemical control non-AD 29.000

210 rows × 11 columns

Initialize Comparison#

  • replicates idea for truely missing values: Define truth as by using n=3 replicates to impute each sample

  • real test data:

    • Not used for predictions or early stopping.

    • [x] add some additional NAs based on distribution of data

Hide code cell source

 
freq_feat = pimmslearn.io.datasplits.load_freq(args.data)
freq_feat.head()  # training data

protein groups
A0A024QZX5;A0A087X1N8;P35237                                                     197
A0A024R0T9;K7ER74;P02655                                                         208
A0A024R3W6;A0A024R412;O60462;O60462-2;O60462-3;O60462-4;O60462-5;Q7LBX6;X5D2Q8   185
A0A024R644;A0A0A0MRU5;A0A1B0GWI2;O75503                                          208
A0A075B6H7                                                                        97
Name: freq, dtype: int64

Produce some addional simulated samples#

The validation simulated NA is used to by all models to evaluate training performance.

Hide code cell source

 
val_pred_simulated_na = data.val_y.to_frame(name='observed')
val_pred_simulated_na
observed
Sample ID protein groups
Sample_158 Q9UN70;Q9UN70-2 14.630
Sample_050 Q9Y287 15.755
Sample_107 Q8N475;Q8N475-2 15.029
Sample_199 P06307 19.376
Sample_067 Q5VUB5 15.309
... ... ...
Sample_111 F6SYF8;Q9UBP4 22.822
Sample_002 A0A0A0MT36 18.165
Sample_049 Q8WY21;Q8WY21-2;Q8WY21-3;Q8WY21-4 15.525
Sample_182 Q8NFT8 14.379
Sample_123 Q16853;Q16853-2 14.504

12600 rows × 1 columns

Hide code cell source

 
test_pred_simulated_na = data.test_y.to_frame(name='observed')
test_pred_simulated_na.describe()
observed
count 12,600.000
mean 16.339
std 2.741
min 7.209
25% 14.412
50% 15.935
75% 17.910
max 30.140

Data in wide format#

  • Autoencoder need data in wide format

Hide code cell source

 
data.to_wide_format()
args.M = data.train_X.shape[-1]
data.train_X.head()
protein groups A0A024QZX5;A0A087X1N8;P35237 A0A024R0T9;K7ER74;P02655 A0A024R3W6;A0A024R412;O60462;O60462-2;O60462-3;O60462-4;O60462-5;Q7LBX6;X5D2Q8 A0A024R644;A0A0A0MRU5;A0A1B0GWI2;O75503 A0A075B6H7 A0A075B6H9 A0A075B6I0 A0A075B6I1 A0A075B6I6 A0A075B6I9 ... Q9Y653;Q9Y653-2;Q9Y653-3 Q9Y696 Q9Y6C2 Q9Y6N6 Q9Y6N7;Q9Y6N7-2;Q9Y6N7-4 Q9Y6R7 Q9Y6X5 Q9Y6Y8;Q9Y6Y8-2 Q9Y6Y9 S4R3U6
Sample ID
Sample_000 15.912 16.852 15.570 16.481 17.301 20.246 16.764 17.584 16.988 20.054 ... 16.012 15.178 NaN 15.050 16.842 NaN NaN 19.563 NaN 12.805
Sample_001 NaN 16.874 15.519 16.387 NaN 19.941 18.786 17.144 NaN 19.067 ... 15.528 15.576 NaN 14.833 16.597 20.299 15.556 19.386 13.970 12.442
Sample_002 16.111 NaN 15.935 16.416 18.175 19.251 16.832 15.671 17.012 18.569 ... 15.229 14.728 13.757 15.118 17.440 19.598 15.735 20.447 12.636 12.505
Sample_003 16.107 17.032 15.802 16.979 15.963 19.628 17.852 18.877 14.182 18.985 ... 15.495 14.590 14.682 15.140 17.356 19.429 NaN 20.216 NaN 12.445
Sample_004 15.603 15.331 15.375 16.679 NaN 20.450 18.682 17.081 14.140 19.686 ... 14.757 NaN NaN 15.256 17.075 19.582 15.328 NaN 13.145 NaN

5 rows × 1421 columns

Add interpolation performance#

Fill Validation data with potentially missing features#

Hide code cell source

 
data.train_X
protein groups A0A024QZX5;A0A087X1N8;P35237 A0A024R0T9;K7ER74;P02655 A0A024R3W6;A0A024R412;O60462;O60462-2;O60462-3;O60462-4;O60462-5;Q7LBX6;X5D2Q8 A0A024R644;A0A0A0MRU5;A0A1B0GWI2;O75503 A0A075B6H7 A0A075B6H9 A0A075B6I0 A0A075B6I1 A0A075B6I6 A0A075B6I9 ... Q9Y653;Q9Y653-2;Q9Y653-3 Q9Y696 Q9Y6C2 Q9Y6N6 Q9Y6N7;Q9Y6N7-2;Q9Y6N7-4 Q9Y6R7 Q9Y6X5 Q9Y6Y8;Q9Y6Y8-2 Q9Y6Y9 S4R3U6
Sample ID
Sample_000 15.912 16.852 15.570 16.481 17.301 20.246 16.764 17.584 16.988 20.054 ... 16.012 15.178 NaN 15.050 16.842 NaN NaN 19.563 NaN 12.805
Sample_001 NaN 16.874 15.519 16.387 NaN 19.941 18.786 17.144 NaN 19.067 ... 15.528 15.576 NaN 14.833 16.597 20.299 15.556 19.386 13.970 12.442
Sample_002 16.111 NaN 15.935 16.416 18.175 19.251 16.832 15.671 17.012 18.569 ... 15.229 14.728 13.757 15.118 17.440 19.598 15.735 20.447 12.636 12.505
Sample_003 16.107 17.032 15.802 16.979 15.963 19.628 17.852 18.877 14.182 18.985 ... 15.495 14.590 14.682 15.140 17.356 19.429 NaN 20.216 NaN 12.445
Sample_004 15.603 15.331 15.375 16.679 NaN 20.450 18.682 17.081 14.140 19.686 ... 14.757 NaN NaN 15.256 17.075 19.582 15.328 NaN 13.145 NaN
... ... ... ... ... ... ... ... ... ... ... ... ... ... ... ... ... ... ... ... ... ...
Sample_205 15.682 16.886 14.910 16.482 NaN 17.705 17.039 NaN 16.413 19.102 ... NaN 15.684 14.236 15.415 17.551 17.922 16.340 19.928 12.929 NaN
Sample_206 15.798 17.554 15.600 15.938 NaN 18.154 18.152 16.503 16.860 18.538 ... 15.422 16.106 NaN 15.345 17.084 18.708 NaN 19.433 NaN NaN
Sample_207 15.739 NaN 15.469 16.898 NaN 18.636 17.950 16.321 16.401 18.849 ... 15.808 16.098 14.403 15.715 NaN 18.725 16.138 19.599 13.637 11.174
Sample_208 15.477 16.779 14.995 16.132 NaN 14.908 NaN NaN 16.119 18.368 ... 15.157 16.712 NaN 14.640 16.533 19.411 15.807 19.545 NaN NaN
Sample_209 NaN 17.261 15.175 16.235 NaN 17.893 17.744 16.371 15.780 18.806 ... 15.237 15.652 15.211 14.205 16.749 19.275 15.732 19.577 11.042 11.791

210 rows × 1421 columns

Hide code cell source

 
data.val_y  # potentially has less features
protein groups A0A024QZX5;A0A087X1N8;P35237 A0A024R0T9;K7ER74;P02655 A0A024R3W6;A0A024R412;O60462;O60462-2;O60462-3;O60462-4;O60462-5;Q7LBX6;X5D2Q8 A0A024R644;A0A0A0MRU5;A0A1B0GWI2;O75503 A0A075B6H7 A0A075B6H9 A0A075B6I0 A0A075B6I1 A0A075B6I6 A0A075B6I9 ... Q9Y653;Q9Y653-2;Q9Y653-3 Q9Y696 Q9Y6C2 Q9Y6N6 Q9Y6N7;Q9Y6N7-2;Q9Y6N7-4 Q9Y6R7 Q9Y6X5 Q9Y6Y8;Q9Y6Y8-2 Q9Y6Y9 S4R3U6
Sample ID
Sample_000 NaN NaN NaN NaN NaN NaN NaN NaN NaN NaN ... NaN NaN NaN NaN NaN 19.863 NaN NaN NaN NaN
Sample_001 NaN NaN NaN NaN NaN NaN NaN NaN NaN NaN ... NaN NaN NaN NaN NaN NaN NaN NaN NaN NaN
Sample_002 NaN 14.523 NaN NaN NaN NaN NaN NaN NaN NaN ... NaN NaN NaN NaN NaN NaN NaN NaN NaN NaN
Sample_003 NaN NaN NaN NaN NaN NaN NaN NaN NaN NaN ... NaN NaN NaN NaN NaN NaN NaN NaN NaN NaN
Sample_004 NaN NaN NaN NaN 15.473 NaN NaN NaN NaN NaN ... NaN NaN 14.048 NaN NaN NaN NaN 19.867 NaN 12.235
... ... ... ... ... ... ... ... ... ... ... ... ... ... ... ... ... ... ... ... ... ...
Sample_205 NaN NaN NaN NaN NaN NaN NaN NaN NaN NaN ... NaN NaN NaN NaN NaN NaN NaN NaN NaN 11.802
Sample_206 NaN NaN NaN NaN NaN NaN NaN NaN NaN NaN ... NaN NaN NaN NaN NaN NaN NaN NaN NaN NaN
Sample_207 NaN NaN NaN NaN NaN NaN NaN NaN NaN NaN ... NaN NaN NaN NaN NaN NaN NaN NaN NaN NaN
Sample_208 NaN NaN NaN NaN NaN NaN 17.530 NaN NaN NaN ... NaN NaN NaN NaN NaN NaN NaN NaN NaN NaN
Sample_209 15.727 NaN NaN NaN NaN NaN NaN NaN NaN NaN ... NaN NaN NaN NaN NaN NaN NaN NaN NaN NaN

210 rows × 1419 columns

Hide code cell source

 
data.val_y = pd.DataFrame(pd.NA, index=data.train_X.index,
                          columns=data.train_X.columns).fillna(data.val_y)
data.val_y
protein groups A0A024QZX5;A0A087X1N8;P35237 A0A024R0T9;K7ER74;P02655 A0A024R3W6;A0A024R412;O60462;O60462-2;O60462-3;O60462-4;O60462-5;Q7LBX6;X5D2Q8 A0A024R644;A0A0A0MRU5;A0A1B0GWI2;O75503 A0A075B6H7 A0A075B6H9 A0A075B6I0 A0A075B6I1 A0A075B6I6 A0A075B6I9 ... Q9Y653;Q9Y653-2;Q9Y653-3 Q9Y696 Q9Y6C2 Q9Y6N6 Q9Y6N7;Q9Y6N7-2;Q9Y6N7-4 Q9Y6R7 Q9Y6X5 Q9Y6Y8;Q9Y6Y8-2 Q9Y6Y9 S4R3U6
Sample ID
Sample_000 NaN NaN NaN NaN NaN NaN NaN NaN NaN NaN ... NaN NaN NaN NaN NaN 19.863 NaN NaN NaN NaN
Sample_001 NaN NaN NaN NaN NaN NaN NaN NaN NaN NaN ... NaN NaN NaN NaN NaN NaN NaN NaN NaN NaN
Sample_002 NaN 14.523 NaN NaN NaN NaN NaN NaN NaN NaN ... NaN NaN NaN NaN NaN NaN NaN NaN NaN NaN
Sample_003 NaN NaN NaN NaN NaN NaN NaN NaN NaN NaN ... NaN NaN NaN NaN NaN NaN NaN NaN NaN NaN
Sample_004 NaN NaN NaN NaN 15.473 NaN NaN NaN NaN NaN ... NaN NaN 14.048 NaN NaN NaN NaN 19.867 NaN 12.235
... ... ... ... ... ... ... ... ... ... ... ... ... ... ... ... ... ... ... ... ... ...
Sample_205 NaN NaN NaN NaN NaN NaN NaN NaN NaN NaN ... NaN NaN NaN NaN NaN NaN NaN NaN NaN 11.802
Sample_206 NaN NaN NaN NaN NaN NaN NaN NaN NaN NaN ... NaN NaN NaN NaN NaN NaN NaN NaN NaN NaN
Sample_207 NaN NaN NaN NaN NaN NaN NaN NaN NaN NaN ... NaN NaN NaN NaN NaN NaN NaN NaN NaN NaN
Sample_208 NaN NaN NaN NaN NaN NaN 17.530 NaN NaN NaN ... NaN NaN NaN NaN NaN NaN NaN NaN NaN NaN
Sample_209 15.727 NaN NaN NaN NaN NaN NaN NaN NaN NaN ... NaN NaN NaN NaN NaN NaN NaN NaN NaN NaN

210 rows × 1421 columns

Variational Autoencoder#

Analysis: DataLoaders, Model, transform#

Hide code cell source

 
default_pipeline = sklearn.pipeline.Pipeline(
    [
        ('normalize', StandardScaler()),
        ('impute', SimpleImputer(add_indicator=False))
    ])

Analysis: DataLoaders, Model#

Hide code cell source

 
analysis = ae.AutoEncoderAnalysis(  # datasplits=data,
    train_df=data.train_X,
    val_df=data.val_y,
    model=models.vae.VAE,
    model_kwargs=dict(n_features=data.train_X.shape[-1],
                      n_neurons=args.hidden_layers,
                      # last_encoder_activation=None,
                      last_decoder_activation=None,
                      dim_latent=args.latent_dim),
    transform=default_pipeline,
    decode=['normalize'],
    bs=args.batch_size)
args.n_params = analysis.n_params_ae
if args.cuda:
    analysis.model = analysis.model.cuda()
analysis.model

VAE(
  (encoder): Sequential(
    (0): Linear(in_features=1421, out_features=64, bias=True)
    (1): Dropout(p=0.2, inplace=False)
    (2): BatchNorm1d(64, eps=1e-05, momentum=0.1, affine=True, bias=True, track_running_stats=True)
    (3): LeakyReLU(negative_slope=0.1)
    (4): Linear(in_features=64, out_features=20, bias=True)
  )
  (decoder): Sequential(
    (0): Linear(in_features=10, out_features=64, bias=True)
    (1): Dropout(p=0.2, inplace=False)
    (2): BatchNorm1d(64, eps=1e-05, momentum=0.1, affine=True, bias=True, track_running_stats=True)
    (3): LeakyReLU(negative_slope=0.1)
    (4): Linear(in_features=64, out_features=2842, bias=True)
  )
)

Training#

Hide code cell source

 
results = []
loss_fct = partial(models.vae.loss_fct, results=results)

Hide code cell source

 
analysis.learn = Learner(dls=analysis.dls,
                         model=analysis.model,
                         loss_func=loss_fct,
                         cbs=[ae.ModelAdapterVAE(),
                              EarlyStoppingCallback(patience=args.patience)
                              ])

analysis.learn.show_training_loop()

Start Fit
   - before_fit     : [TrainEvalCallback, Recorder, ProgressCallback, EarlyStoppingCallback]
  Start Epoch Loop
     - before_epoch   : [Recorder, ProgressCallback]
    Start Train
       - before_train   : [TrainEvalCallback, Recorder, ProgressCallback]
      Start Batch Loop
         - before_batch   : [ModelAdapterVAE, CastToTensor]
         - after_pred     : [ModelAdapterVAE]
         - after_loss     : [ModelAdapterVAE]
         - before_backward: []
         - before_step    : []
         - after_step     : []
         - after_cancel_batch: []
         - after_batch    : [TrainEvalCallback, Recorder, ProgressCallback]
      End Batch Loop
    End Train
     - after_cancel_train: [Recorder]
     - after_train    : [Recorder, ProgressCallback]
    Start Valid
       - before_validate: [TrainEvalCallback, Recorder, ProgressCallback]
      Start Batch Loop
         - **CBs same as train batch**: []
      End Batch Loop
    End Valid
     - after_cancel_validate: [Recorder]
     - after_validate : [Recorder, ProgressCallback]
  End Epoch Loop
   - after_cancel_epoch: []
   - after_epoch    : [Recorder, EarlyStoppingCallback]
End Fit
 - after_cancel_fit: []
 - after_fit      : [ProgressCallback, EarlyStoppingCallback]

Adding a EarlyStoppingCallback results in an error. Potential fix in PR3509 is not yet in current version. Try again later

Hide code cell source

 
# learn.summary()

Hide code cell source

 
suggested_lr = analysis.learn.lr_find()
analysis.params['suggested_inital_lr'] = suggested_lr.valley
suggested_lr

SuggestedLRs(valley=0.0030199517495930195)
_images/c9b509efe004090b8afe1e5b7ff66e6a04af8328e2c48dc72008b52702df36e9.png

Hide code cell source

 
results.clear()  # reset results

dump model config

Hide code cell source

 
# needs class as argument, not instance, but serialization needs instance
analysis.params['last_decoder_activation'] = Sigmoid()

pimmslearn.io.dump_json(
    pimmslearn.io.parse_dict(
        analysis.params, types=[
            (torch.nn.modules.module.Module, lambda m: str(m))
        ]),
    args.out_models / TEMPLATE_MODEL_PARAMS.format(args.model_key))

# restore original value
analysis.params['last_decoder_activation'] = Sigmoid

Hide code cell source

 
# papermill_description=train
analysis.learn.fit_one_cycle(args.epochs_max, lr_max=suggested_lr.valley)
epoch train_loss valid_loss time
0 1687.690063 92.853897 00:00
1 1684.072021 94.253754 00:00
2 1682.008545 93.654343 00:00
3 1680.848022 94.276993 00:00
4 1677.843262 94.543213 00:00
5 1674.314209 94.483749 00:00
6 1671.795410 95.606239 00:00
7 1668.650391 94.714317 00:00
8 1664.342041 94.447678 00:00
9 1661.116821 94.811409 00:00
10 1658.250610 94.286163 00:00
11 1654.668091 94.680588 00:00
12 1652.127563 94.857422 00:00
13 1647.596558 94.615448 00:00
14 1643.616455 94.526794 00:00
15 1639.167236 94.321518 00:00
16 1635.758911 93.897591 00:00
17 1631.248657 94.201653 00:00
18 1626.358276 93.568756 00:00
19 1620.609131 93.316338 00:00
20 1615.250610 93.887321 00:00
21 1608.340210 93.022896 00:00
22 1602.393188 93.212524 00:00
23 1595.197998 92.748955 00:00
24 1587.932129 92.617218 00:00
25 1579.862427 92.455856 00:00
26 1571.290405 92.446465 00:00
27 1561.633789 92.426659 00:00
28 1553.348511 92.766273 00:00
29 1542.468750 92.917358 00:00
30 1531.960083 92.627068 00:00
31 1521.738159 92.251694 00:00
32 1510.727905 92.465820 00:00
33 1499.094727 91.890121 00:00
34 1487.274780 92.090309 00:00
35 1476.110229 92.557198 00:00
36 1464.784668 92.736641 00:00
37 1453.802734 92.863693 00:00
38 1444.145508 93.311028 00:00
39 1433.271362 92.853676 00:00
40 1422.980225 93.517776 00:00
41 1411.958130 93.790619 00:00
42 1401.539917 93.531166 00:00
43 1391.165527 93.017197 00:00
44 1381.529785 92.615097 00:00
45 1370.921143 92.814789 00:00
46 1361.922363 92.267792 00:00
47 1351.974976 92.003693 00:00
48 1342.457275 91.327927 00:00
49 1333.360107 91.926071 00:00
50 1323.971924 92.287209 00:00
51 1316.353394 92.177658 00:00
52 1309.750000 91.305267 00:00
53 1302.246460 91.404045 00:00
54 1295.872559 91.496574 00:00
55 1288.320068 91.928032 00:00
56 1281.898438 91.848923 00:00
57 1274.719849 91.374931 00:00
58 1268.132080 91.358017 00:00
59 1263.293213 91.763466 00:00
60 1257.992065 91.418343 00:00
61 1251.996948 91.021141 00:00
62 1247.211182 90.919334 00:00
63 1240.906372 91.148483 00:00
64 1236.417358 91.044113 00:00
65 1231.177490 90.869568 00:00
66 1226.341064 90.813431 00:00
67 1220.830688 90.359451 00:00
68 1215.408691 90.782143 00:00
69 1211.099487 90.637329 00:00
70 1208.188110 90.580917 00:00
71 1204.232544 90.758720 00:00
72 1199.649414 90.929260 00:00
73 1194.221069 90.767067 00:00
74 1189.766846 90.478180 00:00
75 1185.052490 90.395721 00:00
76 1181.417480 90.613792 00:00
77 1176.756226 90.217949 00:00
78 1172.523560 90.441696 00:00
79 1168.366699 90.840851 00:00
80 1166.126831 91.061554 00:00
81 1161.902954 90.962173 00:00
82 1158.528198 91.296844 00:00
83 1157.384033 91.697533 00:00
84 1154.165405 91.800049 00:00
85 1151.832886 91.743706 00:00
86 1148.914429 91.072235 00:00
87 1145.506592 91.132202 00:00
88 1142.072876 91.205467 00:00
89 1140.492188 90.984337 00:00
90 1135.829712 90.698921 00:00
91 1133.213867 90.819473 00:00
92 1131.597534 90.890831 00:00
93 1127.713623 90.612862 00:00
94 1124.663452 90.696167 00:00
95 1121.846924 90.566162 00:00
96 1119.002808 90.506332 00:00
97 1116.055176 90.827263 00:00
98 1112.790405 90.975975 00:00
99 1110.860229 91.092041 00:00
100 1106.851440 90.936577 00:00
101 1105.353882 91.020493 00:00
102 1103.032959 90.754257 00:00
103 1101.691650 91.482796 00:00
104 1099.664185 91.190804 00:00
105 1097.008423 91.477036 00:00
106 1095.709717 91.468277 00:00
107 1094.388428 91.769699 00:00
108 1091.720703 91.755775 00:00
109 1089.645386 91.468925 00:00
110 1087.975830 92.009521 00:00
111 1086.766724 91.988625 00:00
112 1083.847656 92.180069 00:00
113 1082.315308 91.400108 00:00
114 1081.112305 91.625328 00:00
115 1079.576416 91.901672 00:00
116 1077.492310 92.045853 00:00
117 1075.705933 91.871468 00:00
118 1073.776733 92.004204 00:00
119 1074.155518 92.318443 00:00
120 1072.533203 92.266335 00:00
121 1072.517700 92.367157 00:00
122 1071.231079 92.370911 00:00
123 1070.232178 92.277496 00:00
124 1068.897827 92.229790 00:00
125 1068.957642 91.922592 00:00
126 1067.020874 91.758888 00:00
127 1066.698242 91.621353 00:00 
No improvement since epoch 77: early stopping

Save number of actually trained epochs

Hide code cell source

 
args.epoch_trained = analysis.learn.epoch + 1
args.epoch_trained

128

Loss normalized by total number of measurements#

Hide code cell source

 
N_train_notna = data.train_X.notna().sum().sum()
N_val_notna = data.val_y.notna().sum().sum()
fig = models.plot_training_losses(analysis.learn, args.model_key,
                                  folder=args.out_figures,
                                  norm_factors=[N_train_notna, N_val_notna])

pimmslearn.plotting - INFO     Saved Figures to runs/alzheimer_study/figures/vae_training
_images/b521e0f00dd15e38f7935a24fbcee9771caa3f35dc5f7f5582da1ca3e63b24d2.png

Predictions#

create predictions and select validation data predictions

Hide code cell source

 
analysis.model.eval()
pred, target = res = ae.get_preds_from_df(df=data.train_X, learn=analysis.learn,
                                          position_pred_tuple=0,
                                          transformer=analysis.transform)
pred = pred.stack()
pred

Sample ID   protein groups                                                                
Sample_000  A0A024QZX5;A0A087X1N8;P35237                                                     15.944
            A0A024R0T9;K7ER74;P02655                                                         16.778
            A0A024R3W6;A0A024R412;O60462;O60462-2;O60462-3;O60462-4;O60462-5;Q7LBX6;X5D2Q8   15.807
            A0A024R644;A0A0A0MRU5;A0A1B0GWI2;O75503                                          16.743
            A0A075B6H7                                                                       17.329
                                                                                              ...  
Sample_209  Q9Y6R7                                                                           19.171
            Q9Y6X5                                                                           15.961
            Q9Y6Y8;Q9Y6Y8-2                                                                  19.710
            Q9Y6Y9                                                                           11.879
            S4R3U6                                                                           11.448
Length: 298410, dtype: float32

Hide code cell source

 
val_pred_simulated_na['VAE'] = pred  # 'model_key' ?
val_pred_simulated_na
observed VAE
Sample ID protein groups
Sample_158 Q9UN70;Q9UN70-2 14.630 15.598
Sample_050 Q9Y287 15.755 16.879
Sample_107 Q8N475;Q8N475-2 15.029 14.352
Sample_199 P06307 19.376 19.046
Sample_067 Q5VUB5 15.309 15.046
... ... ... ...
Sample_111 F6SYF8;Q9UBP4 22.822 22.925
Sample_002 A0A0A0MT36 18.165 16.001
Sample_049 Q8WY21;Q8WY21-2;Q8WY21-3;Q8WY21-4 15.525 15.407
Sample_182 Q8NFT8 14.379 13.615
Sample_123 Q16853;Q16853-2 14.504 14.488

12600 rows × 2 columns

Hide code cell source

 
test_pred_simulated_na['VAE'] = pred  # model_key?
test_pred_simulated_na
observed VAE
Sample ID protein groups
Sample_000 A0A075B6P5;P01615 17.016 17.316
A0A087X089;Q16627;Q16627-2 18.280 18.197
A0A0B4J2B5;S4R460 21.735 22.292
A0A140T971;O95865;Q5SRR8;Q5SSV3 14.603 15.275
A0A140TA33;A0A140TA41;A0A140TA52;P22105;P22105-3;P22105-4 16.143 16.844
... ... ... ...
Sample_209 Q96ID5 16.074 16.154
Q9H492;Q9H492-2 13.173 13.393
Q9HC57 14.207 13.975
Q9NPH3;Q9NPH3-2;Q9NPH3-5 14.962 15.077
Q9UGM5;Q9UGM5-2 16.871 16.471

12600 rows × 2 columns

save missing values predictions

Hide code cell source

 
if args.save_pred_real_na:
    pred_real_na = ae.get_missing_values(df_train_wide=data.train_X,
                                         val_idx=val_pred_simulated_na.index,
                                         test_idx=test_pred_simulated_na.index,
                                         pred=pred)
    display(pred_real_na)
    pred_real_na.to_csv(args.out_preds / f"pred_real_na_{args.model_key}.csv")

Sample ID   protein groups          
Sample_000  A0A075B6J9                 15.720
            A0A075B6Q5                 15.832
            A0A075B6R2                 16.975
            A0A075B6S5                 16.327
            A0A087WSY4                 16.405
                                        ...  
Sample_209  Q9P1W8;Q9P1W8-2;Q9P1W8-4   16.208
            Q9UI40;Q9UI40-2            16.392
            Q9UIW2                     16.629
            Q9UMX0;Q9UMX0-2;Q9UMX0-4   13.372
            Q9UP79                     15.961
Name: intensity, Length: 46401, dtype: float32

Plots#

  • validation data

Hide code cell source

 
analysis.model = analysis.model.cpu()
# underlying data is train_X for both
# assert analysis.dls.valid.data.equals(analysis.dls.train.data)
# Reconstruct DataLoader for case that during training singleton batches were dropped
_dl = torch.utils.data.DataLoader(
    pimmslearn.io.datasets.DatasetWithTarget(
        analysis.dls.valid.data),
    batch_size=args.batch_size,
    shuffle=False)
df_latent = pimmslearn.model.get_latent_space(analysis.model.get_mu_and_logvar,
                                              dl=_dl,
                                              dl_index=analysis.dls.valid.data.index)
df_latent
latent dimension 1 latent dimension 2 latent dimension 3 latent dimension 4 latent dimension 5 latent dimension 6 latent dimension 7 latent dimension 8 latent dimension 9 latent dimension 10
Sample ID
Sample_000 -0.402 -2.382 2.161 -0.388 -0.181 0.384 -0.172 -0.856 -0.324 1.218
Sample_001 -0.083 -2.159 1.263 -0.132 0.568 -0.523 0.816 0.617 0.151 1.028
Sample_002 0.370 -2.081 0.907 -0.741 -1.677 0.112 -0.894 0.577 1.805 1.780
Sample_003 0.374 -2.146 1.483 -1.190 -0.191 0.250 -1.295 0.729 0.417 1.498
Sample_004 -0.022 -2.219 1.783 -0.520 -0.401 0.196 -0.493 0.796 -0.723 0.524
... ... ... ... ... ... ... ... ... ... ...
Sample_205 0.256 -2.932 -1.082 -0.217 -1.455 1.739 -0.546 0.346 -0.071 -0.383
Sample_206 -0.799 -0.074 -1.286 1.353 -1.551 -1.475 0.949 -1.029 0.688 0.525
Sample_207 -1.423 -2.442 0.636 2.176 -1.026 1.148 -0.439 -1.272 -0.742 -1.433
Sample_208 -0.121 -1.993 -0.673 1.808 -0.939 -1.504 0.555 0.224 1.270 -1.597
Sample_209 -0.293 -1.581 -1.173 0.956 -0.700 -1.652 -1.576 0.970 1.368 -0.781

210 rows × 10 columns

Hide code cell source

 
ana_latent = analyzers.LatentAnalysis(df_latent,
                                      df_meta,
                                      args.model_key,
                                      folder=args.out_figures)
if args.meta_date_col and df_meta is not None:
    figures[f'latent_{args.model_key}_by_date'], ax = ana_latent.plot_by_date(
        args.meta_date_col)

Hide code cell source

 
if args.meta_cat_col and df_meta is not None:
    figures[f'latent_{args.model_key}_by_{"_".join(args.meta_cat_col.split())}'], ax = ana_latent.plot_by_category(
        args.meta_cat_col)

Hide code cell source

 
feat_freq_val = val_pred_simulated_na['observed'].groupby(level=-1).count()
feat_freq_val.name = 'freq_val'
ax = feat_freq_val.plot.box()
_images/cf9e9c20e966b08686dc7f1335a5340e530ffd0a521cae060530e7d2d1604f67.png

Hide code cell source

 
feat_freq_val.value_counts().sort_index().head()  # require more than one feat?

freq_val
1    12
2    18
3    50
4    82
5   108
Name: count, dtype: int64

Hide code cell source

 
errors_val = val_pred_simulated_na.drop('observed', axis=1).sub(
    val_pred_simulated_na['observed'], axis=0)
errors_val = errors_val.abs().groupby(level=-1).mean()
errors_val = errors_val.join(freq_feat).sort_values(by='freq', ascending=True)


errors_val_smoothed = errors_val.copy()  # .loc[feat_freq_val > 1]
errors_val_smoothed[errors_val.columns[:-1]] = errors_val[errors_val.columns[:-1]
                                                          ].rolling(window=200, min_periods=1).mean()
ax = errors_val_smoothed.plot(x='freq', figsize=(15, 10))
# errors_val_smoothed
_images/760bd681d9947c273929aff9106645a73e5835a041c9ba775668eebac64878ea.png

Hide code cell source

 
errors_val = val_pred_simulated_na.drop('observed', axis=1).sub(
    val_pred_simulated_na['observed'], axis=0)
errors_val.abs().groupby(level=-1).agg(['mean', 'count'])
VAE
mean count
protein groups
A0A024QZX5;A0A087X1N8;P35237 0.122 7
A0A024R0T9;K7ER74;P02655 1.329 4
A0A024R3W6;A0A024R412;O60462;O60462-2;O60462-3;O60462-4;O60462-5;Q7LBX6;X5D2Q8 0.280 9
A0A024R644;A0A0A0MRU5;A0A1B0GWI2;O75503 0.242 6
A0A075B6H7 0.595 6
... ... ...
Q9Y6R7 0.353 10
Q9Y6X5 0.288 7
Q9Y6Y8;Q9Y6Y8-2 0.285 9
Q9Y6Y9 0.372 15
S4R3U6 0.482 24

1419 rows × 2 columns

Hide code cell source

 
errors_val
VAE
Sample ID protein groups
Sample_158 Q9UN70;Q9UN70-2 0.968
Sample_050 Q9Y287 1.124
Sample_107 Q8N475;Q8N475-2 -0.677
Sample_199 P06307 -0.330
Sample_067 Q5VUB5 -0.263
... ... ...
Sample_111 F6SYF8;Q9UBP4 0.103
Sample_002 A0A0A0MT36 -2.164
Sample_049 Q8WY21;Q8WY21-2;Q8WY21-3;Q8WY21-4 -0.118
Sample_182 Q8NFT8 -0.764
Sample_123 Q16853;Q16853-2 -0.016

12600 rows × 1 columns

Comparisons#

Simulated NAs : Artificially created NAs. Some data was sampled and set explicitly to misssing before it was fed to the model for reconstruction.

Validation data#

  • all measured (identified, observed) peptides in validation data

Hide code cell source

 
# papermill_description=metrics
# d_metrics = models.Metrics(no_na_key='NA interpolated', with_na_key='NA not interpolated')
d_metrics = models.Metrics()

The simulated NA for the validation step are real test data (not used for training nor early stopping)

Hide code cell source

 
added_metrics = d_metrics.add_metrics(val_pred_simulated_na, 'valid_simulated_na')
added_metrics

Selected as truth to compare to: observed

{'VAE': {'MSE': 0.45593273844805987,
  'MAE': 0.43189727793148736,
  'N': 12600,
  'prop': 1.0}}

Test Datasplit#

Hide code cell source

 
added_metrics = d_metrics.add_metrics(test_pred_simulated_na, 'test_simulated_na')
added_metrics

Selected as truth to compare to: observed

{'VAE': {'MSE': 0.48091149712742703,
  'MAE': 0.436850015541966,
  'N': 12600,
  'prop': 1.0}}

Save all metrics as json

Hide code cell source

 
pimmslearn.io.dump_json(d_metrics.metrics, args.out_metrics /
                        f'metrics_{args.model_key}.json')
d_metrics

{ 'test_simulated_na': { 'VAE': { 'MAE': 0.436850015541966,
                                  'MSE': 0.48091149712742703,
                                  'N': 12600,
                                  'prop': 1.0}},
  'valid_simulated_na': { 'VAE': { 'MAE': 0.43189727793148736,
                                   'MSE': 0.45593273844805987,
                                   'N': 12600,
                                   'prop': 1.0}}}

Hide code cell source

 
metrics_df = models.get_df_from_nested_dict(
    d_metrics.metrics, column_levels=['model', 'metric_name']).T
metrics_df
subset valid_simulated_na test_simulated_na
model metric_name
VAE MSE 0.456 0.481
MAE 0.432 0.437
N 12,600.000 12,600.000
prop 1.000 1.000

Save predictions#

Hide code cell source

 
# save simulated missing values for both splits
val_pred_simulated_na.to_csv(args.out_preds / f"pred_val_{args.model_key}.csv")
test_pred_simulated_na.to_csv(args.out_preds / f"pred_test_{args.model_key}.csv")

Config#

Hide code cell source

 
figures  # switch to fnames?

{}

Hide code cell source

 
args.dump(fname=args.out_models / f"model_config_{args.model_key}.yaml")
args

{'M': 1421,
 'batch_size': 64,
 'cuda': False,
 'data': Path('runs/alzheimer_study/data'),
 'epoch_trained': 128,
 'epochs_max': 300,
 'file_format': 'csv',
 'fn_rawfile_metadata': 'https://raw.githubusercontent.com/RasmussenLab/njab/HEAD/docs/tutorial/data/alzheimer/meta.csv',
 'folder_data': '',
 'folder_experiment': Path('runs/alzheimer_study'),
 'hidden_layers': [64],
 'latent_dim': 10,
 'meta_cat_col': None,
 'meta_date_col': None,
 'model': 'VAE',
 'model_key': 'VAE',
 'n_params': 277998,
 'out_figures': Path('runs/alzheimer_study/figures'),
 'out_folder': Path('runs/alzheimer_study'),
 'out_metrics': Path('runs/alzheimer_study'),
 'out_models': Path('runs/alzheimer_study'),
 'out_preds': Path('runs/alzheimer_study/preds'),
 'patience': 50,
 'sample_idx_position': 0,
 'save_pred_real_na': True}