38 add data to train model

src/training_classification.py

@@ -142,17 +142,11 @@
 # %% [markdown]
 # ## Before starting
 # If you failed to obtain the preprocessing using the `t1-linear` pipeline,
-# please uncomment the next cell. You can extract tensors from this CAPS, but
-# for the training part you will need a bigger dataset.
+# please uncomment the next cell.
 # %%
 # !curl -k https://aramislab.paris.inria.fr/clinicadl/files/handbook_2023/data_oasis/CAPS_example.tar.gz -o oasisCaps.tar.gz
 # !tar xf oasisCaps.tar.gz
 
-# %% [markdown]
-# If you have already downloaded the full dataset and converted it to
-# CAPS, you can give the path to the dataset directory by changing
-# the CAPS path. If not, just run it as written but the results will 
-# not be relevant.
 # %% [markdown]
 # To perform the feature extraction for our dataset, run the following cell:     
 # %%
@@ -198,6 +192,17 @@
 # based on the labels and splits obtained in the [previous section](./label_extraction.ipynb).  
 # Of course, you can use another dataset, on which you will also have to perform
 # labels extraction and data splitting.
+#
+# The purpose of this notebook is not to fully train a network, but rather to understand
+# how ClinicaDL works. Therefore, we will keep working with a subset of OASIS-1. This new
+# subset contains 10 T1w images, pre-processed with the pipeline `t1-linear`of Clinia. The
+# `prepare-data` pipeline has already been performed on the dataset.
+#
+# You can remove your old `data_oasis` folder and download the new one:
+
+# %%
+# !curl -k https://aramislab.paris.inria.fr/clinicadl/files/handbook_2023/data_oasis/CAPS_example_train.tar.gz -o oasisCaps.tar.gz
+# !tar xf oasisCaps.tar.gz
 
 # %% [markdown]
 # ## `train classification` 
@@ -234,9 +239,9 @@
 # %% [markdown]
 # ### Prerequisites
 #
-# You need to execute `clinicadl tsvtools get-labels` and `clinicadl tsvtools
-# {split|kfold}` commands prior to running this task to have the correct TSV file
-# organization.  Moreover, there should be a CAPS, obtained running the
+# If you use your own dataset, you need to execute `clinicadl tsvtools get-labels` 
+# and `clinicadl tsvtools {split|kfold}` commands prior to running this task to have 
+# the correct TSV file organization. Moreover, there should be a CAPS, obtained running the
 # preprocessing pipeline wanted.
 # %% [markdown]
 # ### Running the task
@@ -252,7 +257,7 @@
 # [CAPS](https://aramislab.paris.inria.fr/clinica/docs/public/latest/CAPS/Introduction/)
 # hierarchy. In case of multi-cohort training, must be a path to a TSV file.
 # - `PREPROCESSING_JSON` (str) is the name of the preprocessing json file stored
-# in the `CAPS_DIRECTORY` that corresponds to the `clinicadl extract` output.
+# in the `CAPS_DIRECTORY` that corresponds to the `clinicadl prepare-data` output.
 # This will be used to load the correct tensor inputs with the wanted
 # preprocessing.
 # - `TSV_DIRECTORY` (Path) is the input folder of a TSV file tree generated by
@@ -299,21 +304,26 @@
 # The default label for the classification task is `diagnosis` but as long as it
 # is a categorical variable, it can be of any type.
 # %% [markdown]
-# The next cell train a `resnet18` to classify 2D slices of t1-linear MRI by
+# The next cells train `resnet18` networks to classify 2D slices of t1-linear MRI by
 # diagnosis (AD or CN). 
-# Please note that the purpose of this notebook is not to fully train a network
-# because we don't have enough data. The objective is to understand how ClinicaDL 
-# works and make inferences using pretrained models in the next section.
+# Please note once again that we don't expect any interesting results with a
+# network trained on only 10 MRI images. That's why we will train the networks for
+# only few epochs.
+#
+# Let's first train a **single-CNN** on all slice locations (actually 4 networks are
+# trained, one for each split):
 
 
 # %% 
 # 2D-slice single-CNN training
-#!clinicadl train classification -h
-!clinicadl train classification data_oasis/CAPS_example slice_classification_t1 data_oasis/split/4_fold/ data_oasis/maps_classification_2D_slice_resnet18 --n_splits 4 --architecture resnet18 
-
+!clinicadl train classification data_oasis/CAPS_example slice_classification_t1 data_oasis/split/4_fold/ data_oasis/maps_classification_2D_slice_resnet18 --n_splits 4 --architecture resnet18 --batch_size 8 --epochs 5
+# %% [markdown]
+# Then, let's train a **multi-CNN** (i.e.one CNN is trained per slice location).
+# 168 models will be trained, so this command may take a while. If you don't want 
+# to run it, the results can be downloaded a few lines further on.
 # %%
 # 2D-slice multi-CNN training
-!clinicadl train classification data_oasis/CAPS_example slice_classification_t1 data_oasis/split/4_fold/ data_oasis/maps_classification_2D_slice_multi --n_splits 4 --architecture resnet18 --multi_network
+!clinicadl train classification data_oasis/CAPS_example slice_classification_t1 data_oasis/split/4_fold/ data_oasis/maps_classification_2D_slice_multi --n_splits 4 --split 0 --architecture resnet18 --batch_size 2 --epochs 1 --multi_network
 
 # %% [markdown]
 # The `clinicadl train` command outputs a MAPS structure in which there are only
@@ -363,12 +373,12 @@
 #```
 
 # You can find more information about MAPS structure on our
-# [documentation](https://clinicadl.readthedocs.io/en/latest/Introduction/#maps-definition)
+# [documentation](https://clinicadl.readthedocs.io/en/latest/Introduction/#maps-definition).
 
 # %% [markdown]
 # # Inference using pretrained models
 #
-# (If you failed to train the model please uncomment the next cell)
+# If you failed to train the model please uncomment the next cells:
 # %%
 !curl -k https://aramislab.paris.inria.fr/clinicadl/files/handbook_2023/data_oasis/maps_classification_2D_slice_multi.tar.gz -o maps_classification_2D_slice_multi.tar.gz
 !tar xf maps_classification_2D_slice_multi.tar.gz
@@ -377,23 +387,12 @@
 !curl -k https://aramislab.paris.inria.fr/clinicadl/files/handbook_2023/data_oasis/maps_classification_2D_slice_resnet.tar.gz -o maps_classification_2D_slice_resnet.tar.gz
 !tar xf maps_classification_2D_slice_resnet.tar.gz
 
-# %% [markdown]
-# If you failed to train the model, you also need to download the TSV files with 
-# the list of participants for each split used for the training because `clinicadl 
-# tsvtools split` and `clinicadl tsvtools kfold` commands randomly split data so 
-# you can have data leakage error (see previous [notebook](notebooks/labels_extraction.ipynb) 
-# for more information about data leakage).
-
-# %% 
-!curl -k https://aramislab.paris.inria.fr/clinicadl/files/handbook_2023/data_oasis/split.tar.gz -o training_split.tar.gz
-!tar xf training_split.tar.gz
-
 # %% [markdown]
 # The `predict` functionality performs individual prediction and metrics
 # computation on a set of data using models trained with `clinicadl train` or
 # `clinicadl random-search` tasks. 
 # It can also use any pretrained models if they are structured like a
-# [MAPS](https://clinicadl.readthedocs.io/en/latest/Introduction/#maps-definition)
+# [MAPS](https://clinicadl.readthedocs.io/en/latest/Introduction/#maps-definition).
 
 # %% [markdown]
 # ### Running the task 
@@ -403,33 +402,32 @@
 #   clinicadl predict [OPTIONS] INPUT_MAPS_DIRECTORY DATA_GROUP
 #```
 # where:
-# - INPUT_MAPS_DIRECTORY (Path) is the path to the MAPS of the pretrained model.
-# - DATA_GROUP (str) is the name of the data group used for the prediction.
+# - `INPUT_MAPS_DIRECTORY` (Path) is the path to the MAPS of the pretrained model.
+# - `DATA_GROUP` (str) is the name of the data group used for the prediction.
 
 # ```{warning}
 # For ClinicaDL, a data group is linked to a list of participants / sessions and
 # a CAPS directory. When performing a prediction, interpretation or tensor
-# serialization the user must give a data group. If this data group does not
-# exist, the user MUST give a caps_directory and a participants_tsv. If this
-# data group already exists, the user MUST not give any caps_directory or
-# participants_tsv, or set overwrite to True.
+# serialization, the user must give a data group. If this data group does not
+# exist (in the MAPS), the user MUST give a `caps_directory` and a `participants_tsv`. If this
+# data group already exists, the user MUST not give any `caps_directory` or
+# `participants_tsv`, or set overwrite to True.
 # ```
 
 # If you want to add optional argument you can check the
 # [documentation](https://clinicadl.readthedocs.io/en/latest/Predict/).
 
 # %%
-# !clinicadl predict -h
-!clinicadl predict data_oasis/maps_classification_2D_slice_resnet18 'test-Oasis2' --participants_tsv ./data_oasis/split/test_baseline.tsv --caps_directory data_oasis/CAPS_example
+!clinicadl predict data_oasis/maps_classification_2D_slice_resnet18 'test-Oasis' --participants_tsv ./data_oasis/split/test_baseline.tsv --caps_directory data_oasis/CAPS_example
 
 # %%
 !clinicadl predict data_oasis/maps_classification_2D_slice_multi 'test-Oasis' --participants_tsv ./data_oasis/split/test_baseline.tsv --caps_directory data_oasis/CAPS_example
 
 # %% [markdown]
-# Results are stored in the MAPS of path `model_path`, according to the
+# Results are stored in the MAPS, according to the
 # following file system:
 # ```text
-# model_path>
+# <maps_directory>
 #     ├── split-0  
 #     ├── ...  
 #     └── split-<i>
@@ -447,6 +445,6 @@
 # running the next cell:
 # %%
 import pandas as pd
-metrics = pd.read_csv("data_oasis/maps_classification_2D_slice_resnet18/split-0/best-loss/test-Oasis/test-OASIS_slice_level_metrics.tsv", sep="\t")
+metrics = pd.read_csv("data_oasis/maps_classification_2D_slice_resnet18/split-0/best-loss/test-Oasis/test-Oasis_slice_level_metrics.tsv", sep="\t")
 metrics.head()
 # %%