This repository contains 1D variants of popular CNN models for classification like ResNets, DenseNets, VGG, etc. It also contains weights obtained by converting ImageNet weights from the same 2D models. It can be useful for classification of audio or some timeseries data.

This repository is based on great classification_models repo by @qubvel

• VGG [16, 19]
• ResNet [18, 34, 50, 101, 152]
• ResNeXt [50, 101]
• SE-ResNet [18, 34, 50, 101, 152]
• SE-ResNeXt [50, 101]
• SE-Net [154]
• DenseNet [121, 169, 201]
• Inception ResNet V2
• Inception V3
• MobileNet
• MobileNet v2
• EfficientNet
• EfficientNet v2

pip install classification-models-1D

from classification_models_1D.tfkeras import Classifiers

ResNet18, preprocess_input = Classifiers.get('resnet18')
model = ResNet18(input_shape=(224*224, 2), weights='imagenet')

All possible nets for Classifiers.get() method:

Based on Conv1D: 'resnet18', 'resnet34', 'resnet50', 'resnet101', 'resnet152', 'seresnet18', 'seresnet34', 'seresnet50', 'seresnet101', 'seresnet152', 'seresnext50','seresnext101', 'senet154', 'resnext50', 'resnext101', 'vgg16', 'vgg19', 'densenet121', 'densenet169', 'densenet201', 'mobilenet', 'mobilenetv2', 'inceptionresnetv2', 'inceptionv3', 'EfficientNetB0', 'EfficientNetB1', 'EfficientNetB2', 'EfficientNetB3', 'EfficientNetB4', 'EfficientNetB5', 'EfficientNetB6', 'EfficientNetB7', 'EfficientNetV2B0', 'EfficientNetV2B1', 'EfficientNetV2B2', 'EfficientNetV2B3', 'EfficientNetV2S', 'EfficientNetV2M', 'EfficientNetV2L'

Non-standard nets (Conv1D): resnet18_pool8

Based on spectrograms and Conv2D: 'EfficientNetB0_spectre', 'EfficientNetB1_spectre', 'EfficientNetB2_spectre', 'EfficientNetB3_spectre', 'EfficientNetB4_spectre', 'EfficientNetB5_spectre', 'EfficientNetB6_spectre', 'EfficientNetB7_spectre'

Code to convert 2D imagenet weights to 1D variant is available here: convert_imagenet_weights_to_1D_models.py.

If initial 2D model had shape (224, 224, 3) then you can use shape (W, 3) where W ~= 224*224, so something like (224*224, 2) will be ok.

• Default pooling/stride size for 1D models set equal to 4 to match (2, 2) pooling for 2D nets. Kernel size by default is 9 to match (3, 3) kernels. You can change it for your needs using parameters stride_size and kernel_size. Example:

from classification_models_1D.tfkeras import Classifiers

ResNet18, preprocess_input = Classifiers.get('resnet18')
model = ResNet18(
   input_shape=(224*224, 2),
   stride_size=6,
   kernel_size=3, 
   weights=None
)

• You can set different pooling for each pooling block. For example you don't need pooling at first convolution. You can do it using tuple as value for stride_size:

from classification_models_1D.tfkeras import Classifiers

ResNet18, preprocess_input = Classifiers.get('resnet34')
model = ResNet18(
   input_shape=(65536, 2),
   stride_size=(1, 4, 4, 8, 8),
   kernel_size=9,
   weights='imagenet'
)

• For some models like (resnet, resnext, senet, vgg16, vgg19, densenet) it's possible to change number of blocks/poolings. For example if you want to switch to pooling/stride = 2 but make more poolings overall. You can do it like that:

from classification_models_1D.tfkeras import Classifiers

ResNet18, preprocess_input = Classifiers.get('resnet34')
model = ResNet18(
   input_shape=(224*224, 2),
   include_top=False,
   weights=None,
   stride_size=(2, 4, 4, 4, 2, 2, 2, 2),
   kernel_size=3,
   repetitions=(2, 2, 2, 2, 2, 2, 2),
   init_filters=16,
)

Note: Since number of filters grows 2 times, you can set initial number of filters with init_filters parameter.

Imagenet weights available for all models except ('inceptionresnetv2', 'inceptionv3'). They available only for kernel_size == 3 or kernel_size == 9 and 2 channel input (e.g. stereo sound). Weights were converted from 2D models to 1D variant. Weights can be loaded with any pooling scheme.

AudioSet is large audio dataset. It's multilabel classifcation on 527 different classes. All available data was used for training. It's around 1.9 millions of audio tracks. Each track is around 10 seconds of length.

• AudioSet weights were obtained for default parameters kernel_size = 9, stride_size = (4, 4, 4, 4, 4).
• Random class sampling was used during training. To form batch first choose random class, then choose random sample, which contains this class.
• Validation data can be found here: AudioSet validation.

Quality table below:

• Note: Required memory is for input shape of (441000, 2) - it's for classification of 10 seconds stereo audio (like in AudioSet).

• https://github.com/qubvel/classification_models - original 2D repo
• https://github.com/ZFTurbo/classification_models_3D - 3D variant repo

• Create pretrained weights obtained on AudioSet