🔀 Merge pull request #83 from DIVA-DIA/dev_resnet

Created resnet FCN header

README.md

@@ -94,3 +94,20 @@ python run.py trainer.max_epochs=20 datamodule.batch_size=64
 11. Go to the root folder of the framework and activate the environment (source .autoenv OR conda activate unsupervised_learning)
 12. Log into wandb. Execute `wandb login` and follow the instructions
 13. Now you should be able to run the basic experiment from PyCharm
+
+
+### Loading models
+You can load the different model parts `backbone` or `header` as well as the whole task.
+To load the `backbone` or the `header` you need to add to your experiment config the field `path_to_weights`.
+e.g.
+```
+model:
+    header:
+        path_to_weights: /my/path/to/the/pth/file
+```
+To load the whole task you need to provide the path to the whole task to the trainer. This is with the field `resume_from_checkpoint`.
+e.g.
+```
+trainer:
+    resume_from_checkpoint: /path/to/.ckpt/file
+```

configs/experiment/dev_rotnet_pt_resnet18_cb55_10_segmentation.yaml

@@ -0,0 +1,74 @@
+# @package _global_
+
+# to execute this experiment run:
+# python run.py +experiment=exp_example_full
+
+defaults:
+    - /plugins: default.yaml
+    - /task: semantic_segmentation_HisDB.yaml
+    - /loss: crossentropyloss.yaml
+    - /metric: hisdbiou.yaml
+    - /model/backbone: resnet18.yaml
+    - /model/header: resnet_segmentation.yaml
+    - /optimizer: adam.yaml
+    - /callbacks:
+          - check_compatibility.yaml
+          - model_checkpoint.yaml
+          - watch_model_wandb.yaml
+    - /logger:
+          - wandb.yaml # set logger here or use command line (e.g. `python run.py logger=wandb`)
+          - csv.yaml
+
+# we override default configurations with nulls to prevent them from loading at all
+# instead we define all modules and their paths directly in this config,
+# so everything is stored in one place for more readibility
+
+seed: 42
+
+train: True
+test: False
+
+trainer:
+    _target_: pytorch_lightning.Trainer
+    gpus: -1
+    accelerator: 'ddp'
+    min_epochs: 1
+    max_epochs: 3
+    weights_summary: full
+    precision: 16
+
+task:
+    confusion_matrix_log_every_n_epoch: 1
+    confusion_matrix_val: True
+    confusion_matrix_test: True
+
+datamodule:
+    _target_: src.datamodules.DivaHisDB.datamodule_cropped.DivaHisDBDataModuleCropped
+
+    data_dir: /netscratch/datasets/semantic_segmentation/datasets_cropped/CB55-10-segmentation
+    crop_size: 256
+    num_workers: 4
+    batch_size: 16
+    shuffle: True
+    drop_last: True
+    data_folder_name: data
+    gt_folder_name: gt
+
+model:
+    backbone:
+        path_to_weights: /netscratch/experiments_lars_paul/lars/2021-11-15/16-08-51/checkpoints/epoch=1/backbone.pth
+    header:
+        in_channels: 512
+
+callbacks:
+    model_checkpoint:
+        filename: ${checkpoint_folder_name}dev-rotnet-pt-resnet18-cb55-10-segmentation
+    watch_model:
+        log_freq: 1
+
+logger:
+    wandb:
+        name: 'dev-rotnet-pt-resnet18-cb55-10-segmetnation'
+        tags: [ "best_model", "USL" ]
+        group: 'dev-runs'
+        notes: "Testing"

configs/experiment/dev_rotnet_resnet18_cb55_10.yaml

@@ -9,7 +9,7 @@ defaults:
     - /loss: crossentropyloss.yaml
     - /metric: accuracy.yaml
     - /model/backbone: resnet18.yaml
-    - /model/header: null
+    - /model/header: resnet_classification.yaml
     - /optimizer: adam.yaml
     - /callbacks:
           - check_compatibility.yaml
@@ -55,21 +55,18 @@ datamodule:
 
 model:
     header:
-        _target_: src.models.headers.fully_connected.SingleLinear
-
-        num_classes: ${datamodule:num_classes}
         # needs to be calculated from the output of the last layer of the backbone (do not forget to flatten!)
-        input_size: 2048
+        in_channels: 32768
 
 callbacks:
     model_checkpoint:
-        filename: ${checkpoint_folder_name}dev-rotnet-basic-cnn-cb55-10
+        filename: ${checkpoint_folder_name}dev-rotnet-resnet18-cb55-10
     watch_model:
         log_freq: 1
 
 logger:
     wandb:
-        name: 'dev-rotnet-basic-cnn-cb55-10'
+        name: 'dev-rotnet-resnet18-cb55-10'
         tags: [ "best_model", "USL" ]
         group: 'dev-runs'
         notes: "Testing"

configs/experiment/dev_rotnet_resnet50_cb55_10.yaml

@@ -0,0 +1,72 @@
+# @package _global_
+
+# to execute this experiment run:
+# python run.py +experiment=exp_example_full
+
+defaults:
+    - /plugins: default.yaml
+    - /task: classification.yaml
+    - /loss: crossentropyloss.yaml
+    - /metric: accuracy.yaml
+    - /model/backbone: resnet50.yaml
+    - /model/header: resnet_classification.yaml
+    - /optimizer: adam.yaml
+    - /callbacks:
+          - check_compatibility.yaml
+          - model_checkpoint.yaml
+          - watch_model_wandb.yaml
+    - /logger:
+          - wandb.yaml # set logger here or use command line (e.g. `python run.py logger=wandb`)
+          - csv.yaml
+
+# we override default configurations with nulls to prevent them from loading at all
+# instead we define all modules and their paths directly in this config,
+# so everything is stored in one place for more readibility
+
+seed: 42
+
+train: True
+test: False
+
+trainer:
+    _target_: pytorch_lightning.Trainer
+    gpus: -1
+    accelerator: 'ddp'
+    min_epochs: 1
+    max_epochs: 3
+    weights_summary: full
+    precision: 16
+
+task:
+    confusion_matrix_log_every_n_epoch: 1
+    confusion_matrix_val: False
+    confusion_matrix_test: False
+
+datamodule:
+    _target_: src.datamodules.RotNet.datamodule_cropped.RotNetDivaHisDBDataModuleCropped
+
+    data_dir: /netscratch/datasets/semantic_segmentation/datasets_cropped/CB55-10-segmentation
+    crop_size: 256
+    num_workers: 4
+    batch_size: 16
+    shuffle: True
+    drop_last: True
+    data_folder_name: data
+
+model:
+    header:
+        # needs to be calculated from the output of the last layer of the backbone (do not forget to flatten!)
+        in_channels: 131072
+
+callbacks:
+    model_checkpoint:
+        filename: ${checkpoint_folder_name}dev-rotnet-resnet50-cb55-10
+    watch_model:
+        log_freq: 1
+
+logger:
+    wandb:
+        name: 'dev-rotnet-resnet50-cb55-10'
+        tags: [ "best_model", "USL" ]
+        group: 'dev-runs'
+        notes: "Testing"

configs/experiment/rotnet_resnet18_cb55_full.yaml

@@ -9,7 +9,7 @@ defaults:
     - /loss: crossentropyloss.yaml
     - /metric: accuracy.yaml
     - /model/backbone: resnet18.yaml
-    - /model/header: single_layer.yaml
+    - /model/header: resnet_classification.yaml
     - /optimizer: adam.yaml
     - /callbacks:
           - check_compatibility.yaml

configs/experiment/rotnet_resnet18_cb55_train10_last.yaml

@@ -9,7 +9,7 @@ defaults:
     - /loss: crossentropyloss.yaml
     - /metric: accuracy.yaml
     - /model/backbone: resnet18.yaml
-    - /model/header: single_layer.yaml
+    - /model/header: resnet_classification.yaml
     - /optimizer: adam.yaml
     - /callbacks:
           - check_compatibility.yaml

configs/experiment/rotnet_resnet18_cb55_train19_last.yaml

@@ -9,7 +9,7 @@ defaults:
     - /loss: crossentropyloss.yaml
     - /metric: accuracy.yaml
     - /model/backbone: resnet18.yaml
-    - /model/header: single_layer.yaml
+    - /model/header: resnet_classification.yaml
     - /optimizer: adam.yaml
     - /callbacks:
           - check_compatibility.yaml

configs/model/header/single_layer.yaml → configs/model/header/resnet_classification.yaml

@@ -1,5 +1,5 @@
-_target_: src.models.headers.fully_connected.SingleLinear
+_target_: src.models.headers.fully_connected.ResNetHeader
 
 num_classes: ${datamodule:num_classes}
 # needs to be calculated from the output of the last layer of the backbone (do not forget to flatten!)
-input_size: 109512
+in_channels: 109512

configs/model/header/resnet_segmentation.yaml

@@ -0,0 +1,8 @@
+_target_: src.models.headers.fully_convolution.ResNetFCNHead
+
+#FCN header for resnets. The in_channels are fixed for the different resnet architectures:
+#resnet18, 34 = 512
+#resnet50, 101, 152 = 2048
+in_channels: 512
+num_classes: ${datamodule:num_classes}
+output_dims: ${datamodule:dims}

requirements.txt

@@ -1,24 +1,23 @@
 # --------- pytorch --------- #
 torch==1.8.1
-torchvision>=0.9.1
-pytorch-lightning>=1.4.4
-lightning-bolts>=0.3.2
-torchmetrics>=0.5.0
+torchvision==0.9.1
+pytorch-lightning==1.4.8
+lightning-bolts==0.4.0
+torchmetrics==0.5.1
 
 # --------- hydra --------- #
 hydra-core==1.1.0
 hydra-colorlog==1.1.0
 hydra-optuna-sweeper==1.1.0
 
 # --------- loggers --------- #
-wandb>=0.10.31
+wandb==0.12.6
 
 # --------- others --------- #
-rich
-python-dotenv
-pre-commit
-scikit-learn>=0.23.2
-pandas
-matplotlib
-seaborn
-pytest
+rich==10.1.0
+python-dotenv==0.17.0
+scikit-learn==0.24.1
+pandas==1.2.4
+matplotlib==3.4.1
+seaborn==0.11.1
+pytest==6.2.3