refactor: Use old timm pretrained weights for resnets

* add: Old weights for resnet18 and wideresnet in anomaly models' configs

* build: Bump version 1.4.0 -> 1.4.1

* docs: Update changelog

* docs: Fix changelog

* add: Update weights in generic anomaly experiments

* add: breaking_changes file

* fix: Better text in breaking changes file

* docs: Specify torchvision weights also in docs

BREAKING_CHANGES.md

@@ -0,0 +1,12 @@
+# Breaking Changes
+All the breaking changes will be documented in this file.
+
+### [1.4.0]
+
+#### Changed
+
+- In Quadra 1.4.0, we upgraded timm to version 0.9.12, resulting in potential variations in default weights for timm backbones compared to previous versions. To continue utilizing the previous weights for resnet18 and wide_resnet50, which are the default backbones for quadra anomaly and classification fine-tuning tasks, we have introduced ".tv_in1k" to the model_name inside Quadra configuration files.
+
+Although the timm upgrade might have very likely adjusted default weights also for other backbones, we have reinstated the old weights only for these two (some internal tests showed better performance of old weights, especially for classification fine-tuning).
+
+If you are updating quadra to a version >= 1.4.0 and you want to keep consistent results, it is recommended to verify whether your timm's backbone is sill using the same weights.

CHANGELOG.md

@@ -2,6 +2,16 @@
 # Changelog
 All notable changes to this project will be documented in this file.
 
+### [1.4.1]
+
+#### Changed
+
+- Change weights of resnet18 and wideresnet50 to old ones in anomaly model configs
+
+#### Updated
+
+- Update anomalib to [v0.7.0+obx.1.2.7] (added default padim n_features for resnets' old weights)
+
 ### [1.4.0]
 
 #### Added

docs/tutorials/examples/anomaly_detection.md

@@ -169,7 +169,7 @@ dataset:
 
 model:
   input_size: [224, 224]
-  backbone: resnet18
+  backbone: resnet18.tv_in1k
   layers:
     - layer1
     - layer2
@@ -190,7 +190,10 @@ metrics:
     manual_pixel: null
 ```
 What we are mostly interested about is the `model` section. In this section we can specify the backbone of the model
-(mainly resnet18 and wide_resnet50_2), which layers are used for feature extraction and the number of features used for dimensionality reduction (there are some default values for resnet18 and wide_resnet50_2).
+(mainly resnet18.tv_in1k and wide_resnet50_2.tv_in1k), which layers are used for feature extraction and the number of features used for dimensionality reduction (there are some default values for resnet18 and wide_resnet50_2).
+```
+Notice: ".tv_in1k" is an extension for timm backbones' model_name which refers to torchvision pretrained weights.
+```
 Generally we always compute an adaptive threshold based on the validation data, but it is possible to specify a manual threshold for both image and pixel as we may want a different tradeoff between false 
 positives and false negatives. The threshold specified must be the unnormalized one.
 

pyproject.toml

@@ -4,7 +4,7 @@ build-backend = "setuptools.build_meta"
 
 [project]
 name = "quadra"
-version = "1.4.0"
+version = "1.4.1"
 description = "Deep Learning experiment orchestration library"
 authors = [
   { name = "Alessandro Polidori", email = "[email protected]" },
@@ -67,7 +67,7 @@ dependencies = [
   "timm==0.9.12",
   # Currently the only version of smp supporting timm 0.9.12 is the following
   "segmentation-models-pytorch@git+https://github.com/qubvel/segmentation_models.pytorch@7b381f899ed472a477a89d381689caf535b5d0a6",
-  "anomalib@git+https://github.com/orobix/[email protected]+obx.1.2.9",
+  "anomalib@git+https://github.com/orobix/[email protected]+obx.1.2.10",
   "xxhash==3.2.*",
   "torchinfo==1.8.*",
 ]
@@ -123,7 +123,7 @@ repository = "https://github.com/orobix/quadra"
 
 # Adapted from https://realpython.com/pypi-publish-python-package/#version-your-package
 [tool.bumpver]
-current_version = "1.4.0"
+current_version = "1.4.1"
 version_pattern = "MAJOR.MINOR.PATCH"
 commit_message = "build: Bump version {old_version} -> {new_version}"
 commit = true

quadra/__init__.py

@@ -1,4 +1,4 @@
-__version__ = "1.4.0"
+__version__ = "1.4.1"
 
 
 def get_version():

quadra/configs/experiment/generic/mvtec/anomaly/padim.yaml

@@ -6,7 +6,7 @@ defaults:
 model:
   model:
     input_size: [224, 224]
-    backbone: resnet18
+    backbone: resnet18.tv_in1k
 
 datamodule:
   num_workers: 12

quadra/configs/experiment/generic/mvtec/anomaly/patchcore.yaml

@@ -6,7 +6,7 @@ defaults:
 model:
   model:
     input_size: [224, 224]
-    backbone: resnet18
+    backbone: resnet18.tv_in1k
 
 datamodule:
   num_workers: 12

quadra/configs/model/anomalib/cfa.yaml

@@ -3,7 +3,7 @@ dataset:
 
 model:
   input_size: [224, 224]
-  backbone: resnet18
+  backbone: resnet18.tv_in1k
   gamma_c: 1
   gamma_d: 1
   num_nearest_neighbors: 3

quadra/configs/model/anomalib/cflow.yaml

@@ -3,7 +3,7 @@ dataset:
 
 model:
   name: cflow
-  backbone: resnet18
+  backbone: resnet18.tv_in1k
   input_size: [256, 256]
   layers:
     - layer2

quadra/configs/model/anomalib/dfm.yaml

@@ -2,7 +2,7 @@ dataset:
   task: classification
 
 model:
-  backbone: wide_resnet50_2
+  backbone: wide_resnet50_2.tv_in1k
   pca_level: 0.97
   score_type: fre # nll: for Gaussian modeling, fre: pca feature reconstruction error
   threshold:

quadra/configs/model/anomalib/fastflow.yaml

@@ -11,7 +11,7 @@ dataset:
 model:
   name: fastflow
   input_size: [224, 224]
-  backbone: resnet18 # options: [resnet18, wide_resnet50_2, cait_m48_448, deit_base_distilled_patch16_384]
+  backbone: resnet18.tv_in1k # options: [resnet18, wide_resnet50_2, cait_m48_448, deit_base_distilled_patch16_384]
   pre_trained: true
   flow_steps: 8 # options: [8, 8, 20, 20] - for each supported backbone
   hidden_ratio: 1.0 # options: [1.0, 1.0, 0.16, 0.16] - for each supported backbone