Feature extraction interface, model cleanup, 'real labels' validation support

.github/workflows/tests.yml

@@ -6,6 +6,10 @@ on:
  pull_request:
  branches: [ master ]
 
+env:
+ OMP_NUM_THREADS: 2
+ MKL_NUM_THREADS: 2
+
 jobs:
  test:
  name: Run tests on ${{ matrix.os }} with Python ${{ matrix.python }}

avg_checkpoints.py

@@ -9,7 +9,7 @@
 EMA (exponential moving average) of the model weights or performing SWA (stochastic
 weight averaging), but post-training.
 
-Hacked together by Ross Wightman (https://github.com/rwightman)
+Hacked together by / Copyright 2020 Ross Wightman (https://github.com/rwightman)
 """
 import torch
 import argparse

clean_checkpoint.py

@@ -5,7 +5,7 @@
 and outputs a CPU tensor checkpoint with only the `state_dict` along with SHA256
 calculation for model zoo compatibility.
 
-Hacked together by Ross Wightman (https://github.com/rwightman)
+Hacked together by / Copyright 2020 Ross Wightman (https://github.com/rwightman)
 """
 import torch
 import argparse

docs/archived_changes.md

@@ -0,0 +1,83 @@
+# Archived Changes
+
+### Feb 29, 2020
+* New MobileNet-V3 Large weights trained from stratch with this code to 75.77% top-1
+* IMPORTANT CHANGE - default weight init changed for all MobilenetV3 / EfficientNet / related models
+ * overall results similar to a bit better training from scratch on a few smaller models tried
+ * performance early in training seems consistently improved but less difference by end
+ * set `fix_group_fanout=False` in `_init_weight_goog` fn if you need to reproducte past behaviour
+* Experimental LR noise feature added applies a random perturbation to LR each epoch in specified range of training
+
+### Feb 18, 2020
+* Big refactor of model layers and addition of several attention mechanisms. Several additions motivated by 'Compounding the Performance Improvements...' (https://arxiv.org/abs/2001.06268):
+ * Move layer/module impl into `layers` subfolder/module of `models` and organize in a more granular fashion
+ * ResNet downsample paths now properly support dilation (output stride != 32) for avg_pool ('D' variant) and 3x3 (SENets) networks
+ * Add Selective Kernel Nets on top of ResNet base, pretrained weights
+ * skresnet18 - 73% top-1
+ * skresnet34 - 76.9% top-1 
+ * skresnext50_32x4d (equiv to SKNet50) - 80.2% top-1
+ * ECA and CECA (circular padding) attention layer contributed by [Chris Ha](https://github.com/VRandme)
+ * CBAM attention experiment (not the best results so far, may remove)
+ * Attention factory to allow dynamically selecting one of SE, ECA, CBAM in the `.se` position for all ResNets
+ * Add DropBlock and DropPath (formerly DropConnect for EfficientNet/MobileNetv3) support to all ResNet variants
+* Full dataset results updated that incl NoisyStudent weights and 2 of the 3 SK weights
+
+### Feb 12, 2020
+* Add EfficientNet-L2 and B0-B7 NoisyStudent weights ported from [Tensorflow TPU](https://github.com/tensorflow/tpu/tree/master/models/official/efficientnet)
+
+### Feb 6, 2020
+* Add RandAugment trained EfficientNet-ES (EdgeTPU-Small) weights with 78.1 top-1. Trained by [Andrew Lavin](https://github.com/andravin) (see Training section for hparams)
+
+### Feb 1/2, 2020
+* Port new EfficientNet-B8 (RandAugment) weights, these are different than the B8 AdvProp, different input normalization.
+* Update results csv files on all models for ImageNet validation and three other test sets
+* Push PyPi package update
+
+### Jan 31, 2020
+* Update ResNet50 weights with a new 79.038 result from further JSD / AugMix experiments. Full command line for reproduction in training section below.
+
+### Jan 11/12, 2020
+* Master may be a bit unstable wrt to training, these changes have been tested but not all combos
+* Implementations of AugMix added to existing RA and AA. Including numerous supporting pieces like JSD loss (Jensen-Shannon divergence + CE), and AugMixDataset
+* SplitBatchNorm adaptation layer added for implementing Auxiliary BN as per AdvProp paper
+* ResNet-50 AugMix trained model w/ 79% top-1 added
+* `seresnext26tn_32x4d` - 77.99 top-1, 93.75 top-5 added to tiered experiment, higher img/s than 't' and 'd'
+
+### Jan 3, 2020
+* Add RandAugment trained EfficientNet-B0 weight with 77.7 top-1. Trained by [Michael Klachko](https://github.com/michaelklachko) with this code and recent hparams (see Training section)
+* Add `avg_checkpoints.py` script for post training weight averaging and update all scripts with header docstrings and shebangs.
+
+### Dec 30, 2019
+* Merge [Dushyant Mehta's](https://github.com/mehtadushy) PR for SelecSLS (Selective Short and Long Range Skip Connections) networks. Good GPU memory consumption and throughput. Original: https://github.com/mehtadushy/SelecSLS-Pytorch
+
+### Dec 28, 2019
+* Add new model weights and training hparams (see Training Hparams section)
+ * `efficientnet_b3` - 81.5 top-1, 95.7 top-5 at default res/crop, 81.9, 95.8 at 320x320 1.0 crop-pct
+ * trained with RandAugment, ended up with an interesting but less than perfect result (see training section)
+ * `seresnext26d_32x4d`- 77.6 top-1, 93.6 top-5
+ * deep stem (32, 32, 64), avgpool downsample
+ * stem/dowsample from bag-of-tricks paper
+ * `seresnext26t_32x4d`- 78.0 top-1, 93.7 top-5
+ * deep tiered stem (24, 48, 64), avgpool downsample (a modified 'D' variant)
+ * stem sizing mods from Jeremy Howard and fastai devs discussing ResNet architecture experiments
+
+### Dec 23, 2019
+* Add RandAugment trained MixNet-XL weights with 80.48 top-1.
+* `--dist-bn` argument added to train.py, will distribute BN stats between nodes after each train epoch, before eval
+
+### Dec 4, 2019
+* Added weights from the first training from scratch of an EfficientNet (B2) with my new RandAugment implementation. Much better than my previous B2 and very close to the official AdvProp ones (80.4 top-1, 95.08 top-5).
+
+### Nov 29, 2019
+* Brought EfficientNet and MobileNetV3 up to date with my https://github.com/rwightman/gen-efficientnet-pytorch code. Torchscript and ONNX export compat excluded.
+ * AdvProp weights added
+ * Official TF MobileNetv3 weights added
+* EfficientNet and MobileNetV3 hook based 'feature extraction' classes added. Will serve as basis for using models as backbones in obj detection/segmentation tasks. Lots more to be done here...
+* HRNet classification models and weights added from https://github.com/HRNet/HRNet-Image-Classification
+* Consistency in global pooling, `reset_classifer`, and `forward_features` across models
+ * `forward_features` always returns unpooled feature maps now
+* Reasonable chance I broke something... let me know
+
+### Nov 22, 2019
+* Add ImageNet training RandAugment implementation alongside AutoAugment. PyTorch Transform compatible format, using PIL. Currently training two EfficientNet models from scratch with promising results... will update.
+* `drop-connect` cmd line arg finally added to `train.py`, no need to hack model fns. Works for efficientnet/mobilenetv3 based models, ignored otherwise.

docs/changes.md

@@ -1,3 +1,20 @@
+# Recent Changes
+
+### Aug 1, 2020
+Universal feature extraction, new models, new weights, new test sets.
+* All models support the `features_only=True` argument for `create_model` call to return a network that extracts features from the deepest layer at each stride.
+* New models
+ * CSPResNet, CSPResNeXt, CSPDarkNet, DarkNet
+ * ReXNet
+ * (Aligned) Xception41/65/71 (a proper port of TF models)
+* New trained weights
+ * SEResNet50 - 80.3
+ * CSPDarkNet53 - 80.1 top-1
+ * CSPResNeXt50 - 80.0 to-1
+ * DPN68b - 79.2 top-1
+ * EfficientNet-Lite0 (non-TF ver) - 75.5 (submitted by @hal-314)
+* Add 'real' labels for ImageNet and ImageNet-Renditions test set, see [`results/README.md`](results/README.md)
+
 ### June 11, 2020
 Bunch of changes:
 
@@ -35,28 +52,3 @@ Bunch of changes:
 ### March 18, 2020
 * Add EfficientNet-Lite models w/ weights ported from [Tensorflow TPU](https://github.com/tensorflow/tpu/tree/master/models/official/efficientnet/lite)
 * Add RandAugment trained ResNeXt-50 32x4d weights with 79.8 top-1. Trained by [Andrew Lavin](https://github.com/andravin) (see Training section for hparams)
-
-### Feb 29, 2020
-* New MobileNet-V3 Large weights trained from stratch with this code to 75.77% top-1
-* IMPORTANT CHANGE - default weight init changed for all MobilenetV3 / EfficientNet / related models
- * overall results similar to a bit better training from scratch on a few smaller models tried
- * performance early in training seems consistently improved but less difference by end
- * set `fix_group_fanout=False` in `_init_weight_goog` fn if you need to reproducte past behaviour
-* Experimental LR noise feature added applies a random perturbation to LR each epoch in specified range of training
-
-### Feb 18, 2020
-* Big refactor of model layers and addition of several attention mechanisms. Several additions motivated by 'Compounding the Performance Improvements...' (https://arxiv.org/abs/2001.06268):
- * Move layer/module impl into `layers` subfolder/module of `models` and organize in a more granular fashion
- * ResNet downsample paths now properly support dilation (output stride != 32) for avg_pool ('D' variant) and 3x3 (SENets) networks
- * Add Selective Kernel Nets on top of ResNet base, pretrained weights
- * skresnet18 - 73% top-1
- * skresnet34 - 76.9% top-1 
- * skresnext50_32x4d (equiv to SKNet50) - 80.2% top-1
- * ECA and CECA (circular padding) attention layer contributed by [Chris Ha](https://github.com/VRandme)
- * CBAM attention experiment (not the best results so far, may remove)
- * Attention factory to allow dynamically selecting one of SE, ECA, CBAM in the `.se` position for all ResNets
- * Add DropBlock and DropPath (formerly DropConnect for EfficientNet/MobileNetv3) support to all ResNet variants
-* Full dataset results updated that incl NoisyStudent weights and 2 of the 3 SK weights
-
-### Feb 12, 2020
-* Add EfficientNet-L2 and B0-B7 NoisyStudent weights ported from [Tensorflow TPU](https://github.com/tensorflow/tpu/tree/master/models/official/efficientnet)

docs/feature_extraction.md

@@ -0,0 +1,173 @@
+# Feature Extraction
+
+All of the models in `timm` have consistent mechanisms for obtaining various types of features from the model for tasks besides classification.
+
+## Penultimate Layer Features (Pre-Classifier Features)
+
+The features from the penultimate model layer can be obtained in severay ways without requiring model surgery (although feel free to do surgery). One must first decide if they want pooled or un-pooled features.
+
+### Unpooled
+
+There are three ways to obtain unpooled features.
+
+Without modifying the network, one can call `model.forward_features(input)` on any model instead of the usual `model(input)`. This will bypass the head classifier and global pooling for networks.
+
+If one wants to explicitly modify the network to return unpooled features, they can either create the model without a classifier and pooling, or remove it later. Both paths remove the parameters associated with the classifier from the network.
+
+#### forward_features()
+```python hl_lines="3 6"
+import torch
+import timm
+m = timm.create_model('xception41', pretrained=True)
+o = m(torch.randn(2, 3, 299, 299))
+print(f'Original shape: {o.shape}')
+o = m.forward_features(torch.randn(2, 3, 299, 299))
+print(f'Unpooled shape: {o.shape}')
+```
+Output:
+```text
+Original shape: torch.Size([2, 1000])
+Unpooled shape: torch.Size([2, 2048, 10, 10])
+```
+
+#### Create with no classifier and pooling
+```python hl_lines="3"
+import torch
+import timm
+m = timm.create_model('resnet50', pretrained=True, num_classes=0, global_pool='')
+o = m(torch.randn(2, 3, 224, 224))
+print(f'Unpooled shape: {o.shape}')
+```
+Output:
+```text
+Unpooled shape: torch.Size([2, 2048, 7, 7])
+```
+
+#### Remove it later
+```python hl_lines="3 6"
+import torch
+import timm
+m = timm.create_model('densenet121', pretrained=True)
+o = m(torch.randn(2, 3, 224, 224))
+print(f'Original shape: {o.shape}')
+m.reset_classifier(0, '')
+o = m(torch.randn(2, 3, 224, 224))
+print(f'Unpooled shape: {o.shape}')
+```
+Output:
+```text
+Original shape: torch.Size([2, 1000])
+Unpooled shape: torch.Size([2, 1024, 7, 7])
+```
+
+### Pooled
+
+To modify the network to return pooled features, one can use `forward_features()` and pool/flatten the result themselves, or modify the network like above but keep pooling intact. 
+
+#### Create with no classifier
+```python hl_lines="3"
+import torch
+import timm
+m = timm.create_model('resnet50', pretrained=True, num_classes=0)
+o = m(torch.randn(2, 3, 224, 224))
+print(f'Pooled shape: {o.shape}')
+```
+Output:
+```text
+Pooled shape: torch.Size([2, 2048])
+```
+
+#### Remove it later
+```python hl_lines="3 6"
+import torch
+import timm
+m = timm.create_model('ese_vovnet19b_dw', pretrained=True)
+o = m(torch.randn(2, 3, 224, 224))
+print(f'Original shape: {o.shape}')
+m.reset_classifier(0)
+o = m(torch.randn(2, 3, 224, 224))
+print(f'Pooled shape: {o.shape}')
+```
+Output:
+```text
+Pooled shape: torch.Size([2, 1024])
+```
+
+
+## Multi-scale Feature Maps (Feature Pyramid)
+
+Object detection, segmentation, keypoint, and a variety of dense pixel tasks require access to feature maps from the backbone network at multiple scales. This is often done by modifying the original classification network. Since each network varies quite a bit in structure, it's not uncommon to see only a few backbones supported in any given obj detection or segmentation library.
+
+`timm` allows a consistent interface for creating any of the included models as feature backbones that output feature maps for selected levels. 
+
+A feature backbone can be created by adding the argument `features_only=True` to any `create_model` call. By default 5 strides will be output from most models (not all have that many), with the first starting at 2 (some start at 1 or 4).
+
+### Create a feature map extraction model
+```python hl_lines="3"
+import torch
+import timm
+m = timm.create_model('resnest26d', features_only=True, pretrained=True)
+o = m(torch.randn(2, 3, 224, 224))
+for x in o:
+ print(x.shape)
+```
+Output:
+```text
+torch.Size([2, 64, 112, 112])
+torch.Size([2, 256, 56, 56])
+torch.Size([2, 512, 28, 28])
+torch.Size([2, 1024, 14, 14])
+torch.Size([2, 2048, 7, 7])
+```
+
+### Query the feature information
+
+After a feature backbone has been created, it can be queried to provide channel or resolution reduction information to the downstream heads without requiring static config or hardcoded constants. The `.feature_info` attribute is a class encapsulating the information about the feature extraction points.
+
+```python hl_lines="3 4"
+import torch
+import timm
+m = timm.create_model('regnety_032', features_only=True, pretrained=True)
+print(f'Feature channels: {m.feature_info.channels()}')
+o = m(torch.randn(2, 3, 224, 224))
+for x in o:
+ print(x.shape)
+```
+Output:
+```text
+Feature channels: [32, 72, 216, 576, 1512]
+torch.Size([2, 32, 112, 112])
+torch.Size([2, 72, 56, 56])
+torch.Size([2, 216, 28, 28])
+torch.Size([2, 576, 14, 14])
+torch.Size([2, 1512, 7, 7])
+```
+
+### Select specific feature levels or limit the stride
+
+There are to additional creation arguments impacting the output features. 
+
+* `out_indices` selects which indices to output
+* `output_stride` limits the feature output stride of the network (also works in classification mode BTW)
+
+`out_indices` is supported by all models, but not all models have the same index to feature stride mapping. Look at the code or check feature_info to compare. The out indices generally correspond to the `C(i+1)th` feature level (a `2^(i+1)` reduction). For most models, index 0 is the stride 2 features, and index 4 is stride 32.
+
+`output_stride` is achieved by converting layers to use dilated convolutions. Doing so is not always straightforward, some networks only support `output_stride=32`.
+
+```python hl_lines="3 4 5"
+import torch
+import timm
+m = timm.create_model('ecaresnet101d', features_only=True, output_stride=8, out_indices=(2, 4), pretrained=True)
+print(f'Feature channels: {m.feature_info.channels()}')
+print(f'Feature reduction: {m.feature_info.reduction()}')
+o = m(torch.randn(2, 3, 320, 320))
+for x in o:
+ print(x.shape)
+```
+Output:
+```text
+Feature channels: [512, 2048]
+Feature reduction: [8, 8]
+torch.Size([2, 512, 40, 40])
+torch.Size([2, 2048, 40, 40])
+```