Overview

This dataset contains features extracted from the Imagenet dataset using a pre-trained ResNet neural network. The network was configured with an input layer of (200, 200, 3). Feature extraction was performed using the Python package Py Image Feature Extractor.

Source

• Imagenet dataset from Kaggle Competition

Related datasets

• Imagenet extracted features with VGG-19
• Imagenet extracted features with Inception V3
• Imagenet extracted features with LBP

Abstract

In the last years, neural networks have evolved from laboratory environments to the state-of-the-art for many real-world problems. Our hypothesis is that neural network models (i.e., their weights and biases) evolve on unique, smooth trajectories in weight space during training. Following, a population of such neural network models (refereed to as “model zoo”) would form topological structures in weight space. We think that the geometry, curvature and smoothness of these structures contain information about the state of training and can be reveal latent properties of individual models. With such zoos, one could investigate novel approaches for (i) model analysis, (ii) discover unknown learning dynamics, (iii) learn rich representations of such populations, or (iv) exploit the model zoos for generative modelling of neural network weights and biases. Unfortunately, the lack of standardized model zoos and available benchmarks significantly increases the friction for further research about populations of neural networks. With this work, we publish a novel dataset of model zoos containing systematically generated and diverse populations of neural network models for further research. In total the proposed model zoo dataset is based on six image datasets, consist of 27 model zoos with varying hyperparameter combinations are generated and includes 50’360 unique neural network models resulting in over 2’585’360 collected model states. Additionally, to the model zoo data we provide an in-depth analysis of the zoos and provide benchmarks for multiple downstream tasks as mentioned before.

Dataset

This dataset is part of a larger collection of model zoos and contains the zoo of 1000 ResNet18 models trained on Tiny Imagenet. All zoos with extensive information and code can be found at www.modelzoos.cc.

The complete zoo is 2.6TB large. Due to the size, this repository contains the checkpoints of the first 115 models at their last epoch 60. For a link to the full dataset as well as more information on the zoos and code to access and use the zoos, please see www.modelzoos.cc.

LGV models used as surrogate in the original paper.

Those resnet50 models were collected along the SGD trajectory with a high learning rate. The zip file contains three random seeds in respective subfolders. Each one contains a subfolder with the original pretrained model from which the model collection started. These pretrained models were trained by Ashukha, A., et al. Pitfalls of In-Domain Uncertainty Estimation and Ensembling in Deep Learning (2020).

Performance on ImageNet dataset. We train the models following the standard training strategy with pre-trained teacher networks ResNet-34 and ResNet-50 provided by Torchvision [42].

One table and six figures. Table 1 shows the number of images for each label in the 1μ–2μ data set, adopting the same labelling used in [11, 12, 13], reported here for completeness: 0 = Porous sponges, 1 = Patterned surfaces, 2 = Particles, 3 = Films and coated surfaces, 4 = Powders, 5 = Tips, 6 = Nanowires, 7 = Biological, 8 = MEMS devices and electrodes, 9 = Fibres.Figure 1 shows test accuracy as a function of the number of training epochs obtained by training from scratch Inception-v3 (magenta), Inception-v4 (orange), Inception-Resnet (green), and AlexNet (black) on SEM data set. All the models were trained with the best combination of hyperparameters, according to the memory capability of the available hardware. In Figure 2, Main: Test accuracy as a function of the number of training epochs obtained when fine tuning on the SEM data set Inception-v3 (magenta) and Inception-v4 (orange) starting from the ImageNet checkpoint, and Inception-v3 (blue) from the SEM checkpoint that, as expected, converges very rapidly. Inset: Test accuracy as a function of the number of training epochs obtained when performing feature extraction of Inception-v3 (magenta), Inception-v4 (orange), and Inception-Resnet (green) on the SEM data set starting from the ImageNet checkpoint. All the models were trained with the best combination of hyperparameters, according to the memory capability of the hardware available. Figure 3 shows intrinsic Dimension of the 1μ–2μ_1001 data set, varying the sample size, computed before autoencoding (green lines) and after autoencoding (red lines). The three brightness levels for each color correspond to the percentage of points used in the linear fi t: 90%, 70%, and 50%. Figure 4 shows ddisc heatmap for a manually labelled subset of images. Figure 5 presents heatmaps of the distances obtained via Inception-v3. The image captions specify the methods used and indicate the correlation index with ddisc. Figure 6 shows NMI scores of the clustering obtained by the five hierarchical algorithms (solid lines) considered as a function of k, the number of clusters. The scores of the artificial scenarios are reported as orange (good case) and green (uniform case) dashed lines.

Honey bees are essential pollinators of ecosystems and agriculture worldwide. With an estimated 50-80% of crops pollinated by honey bees, they generate approximately $20 billion in market value in the U.S. alone. However, commercial beekeepers often face an uphill battle, losing anywhere from 40-90% of their hives annually, significantly impacted by brood diseases caused by bacterial, viral, and fungal pathogens. Accurate diagnosis of these brood diseases, especially distinguishing bacterial diseases like European Foulbrood (EFB) from viral infections with a superficial resemblance to EFB (EFB-like disease), remains challenging. Incorrect diagnoses often lead to prophylactic antibiotic treatment across entire apiaries, exacerbating antibiotic resistance, disrupting native gut microbiota, and increasing susceptibility to opportunistic pathogens. Correct field diagnosis of brood disease is challenging and requires years of experience to identify and differentiate various disease states according to subtle differences in larval symptomology. To explore the feasibility of an image-based AI diagnosis tool, we collaborated with apiary inspectors and researchers to generate a dataset of 2,759 honey bee larvae images from Michigan apiaries, molecularly verified through 16S rRNA microbiome sequencing and qPCR viral screening. Our dataset included EFB cases and viral infections (ABPV, DWVA, and DWVB), which were augmented to 8,430 and 8,124 images respectively. We leveraged transfer learning techniques, fine-tuning deep convolutional neural networks (ResNet-50v2, ResNet-101v2, InceptionResNet-v2) pre-trained on ImageNet to discriminate between EFB and viral infections. Our proof-of-concept models achieved 73-88% accuracy on the training/validation sets. When tested on an independent dataset from Illinois containing additional viral pathogens not present in training data, the models showed higher accuracy for EFB (72-88%) than viral infections (28-68%), highlighting both the promise and current limitations of this approach. Implementing AI-based diagnostic tools could reduce unnecessary antibiotic treatments and help maintain the microbiome integrity critical to colony health. However, expanding training datasets to include all major pathogens, healthy larvae, and diverse geographic regions will be essential for developing field-ready diagnostic tools.

DeepWeeds: A Multiclass Weed Species Image Dataset for Deep Learning

This repository makes available the source code and public dataset for the work, "DeepWeeds: A Multiclass Weed Species Image Dataset for Deep Learning", published with open access by Scientific Reports: https://www.nature.com/articles/s41598-018-38343-3. The DeepWeeds dataset consists of 17,509 images capturing eight different weed species native to Australia in situ with neighbouring flora. In our work, the dataset was classified to an average accuracy of 95.7% with the ResNet50 deep convolutional neural network.

The source code, images and annotations are licensed under CC BY 4.0 license. The contents of this repository are released under an Apache 2 license.

Download the dataset images and our trained models

images.zip (468 MB)

models.zip (477 MB)

Due to the size of the images and models they are hosted outside of the Github repository. The images and models must be downloaded into directories named "images" and "models", respectively, at the root of the repository. If you execute the python script (deepweeds.py), as instructed below, this step will be performed for you automatically.

TensorFlow Datasets

Alternatively, you can access the DeepWeeds dataset with TensorFlow Datasets, TensorFlow's official collection of ready-to-use datasets. DeepWeeds was officially added to the TensorFlow Datasets catalog in August 2019.

Weeds and locations

The selected weed species are local to pastoral grasslands across the state of Queensland. They include: "Chinee apple", "Snake weed", "Lantana", "Prickly acacia", "Siam weed", "Parthenium", "Rubber vine" and "Parkinsonia". The images were collected from weed infestations at the following sites across Queensland: "Black River", "Charters Towers", "Cluden", "Douglas", "Hervey Range", "Kelso", "McKinlay" and "Paluma". The table and figure below break down the dataset by weed, location and geographical distribution.

Data organization

Images are assigned unique filenames that include the date/time the image was photographed and an ID number for the instrument which produced the image. The format is like so: YYYYMMDD-HHMMSS-ID, where the ID is simply an integer from 0 to 3. The unique filenames are strings of 17 characters, such as 20170320-093423-1.

labels

The labels.csv file assigns species labels to each image. It is a comma separated text file in the format:

Filename,Label,Species ... 20170207-154924-0,jpg,7,Snake weed 20170610-123859-1.jpg,1,Lantana 20180119-105722-1.jpg,8,Negative ...

Note: The specific label subsets of training (60%), validation (20%) and testing (20%) for the five-fold cross validation used in the paper are also provided here as CSV files in the same format as "labels.csv".

models

We provide the most successful ResNet50 and InceptionV3 models saved in Keras' hdf5 model format. The ResNet50 model, which provided the best results, has also been converted to UFF format in order to construct a TensorRT inference engine.

resnet.hdf5 inception.hdf5 resnet.uff

deepweeds.py

This python script trains and evaluates Keras' base implementation of ResNet50 and InceptionV3 on the DeepWeeds dataset, pre-trained with ImageNet weights. The performance of the networks are cross validated for 5 folds. The final classification accuracy is taken to be the average across the five folds. Similarly, the final confusion matrix from the associated paper aggregates across the five independent folds. The script also provides the ability to measure the inference speeds within the TensorFlow environment.

The script can be executed to carry out these computations using the following commands.

To train and evaluate the ResNet50 model with five-fold cross validation, use python3 deepweeds.py cross_validate --model resnet.

To train and evaluate the InceptionV3 model with five-fold cross validation, use python3 deepweeds.py cross_validate --model inception.

To measure inference times for the ResNet50 model, use python3 deepweeds.py inference --model models/resnet.hdf5.

To measure inference times for the InceptionV3 model, use python3 deepweeds.py inference --model models/inception.hdf5.

Dependencies

The required Python packages to execute deepweeds.py are listed in requirements.txt.

tensorrt

This folder includes C++ source code for creating and executing a ResNet50 TensorRT inference engine on an NVIDIA Jetson TX2 platform. To build and run on your Jetson TX2, execute the following commands:

cd tensorrt/src make -j4 cd ../bin ./resnet_inference

Citations

If you use the DeepWeeds dataset in your work, please cite it as:

IEEE style citation: “A. Olsen, D. A. Konovalov, B. Philippa, P. Ridd, J. C. Wood, J. Johns, W. Banks, B. Girgenti, O. Kenny, J. Whinney, B. Calvert, M. Rahimi Azghadi, and R. D. White, “DeepWeeds: A Multiclass Weed Species Image Dataset for Deep Learning,” Scientific Reports, vol. 9, no. 2058, 2 2019. [Online]. Available: https://doi.org/10.1038/s41598-018-38343-3 ”

BibTeX

@article{DeepWeeds2019, author = {Alex Olsen and Dmitry A. Konovalov and Bronson Philippa and Peter Ridd and Jake C. Wood and Jamie Johns and Wesley Banks and Benjamin Girgenti and Owen Kenny and James Whinney and Brendan Calvert and Mostafa {Rahimi Azghadi} and Ronald D. White}, title = {{DeepWeeds: A Multiclass Weed Species Image Dataset for Deep Learning}}, journal = {Scientific Reports}, year = 2019, number = 2058, month = 2, volume = 9, issue = 1, day = 14, url = "https://doi.org/10.1038/s41598-018-38343-3", doi = "10.1038/s41598-018-38343-3" }

Brain tumors pose significant global health concerns due to their high mortality rates and limited treatment options. These tumors, arising from abnormal cell growth within the brain, exhibits various sizes and shapes, making their manual detection from magnetic resonance imaging (MRI) scans a subjective and challenging task for healthcare professionals, hence necessitating automated solutions. This study investigates the potential of deep learning, specifically the DenseNet architecture, to automate brain tumor classification, aiming to enhance accuracy and generalizability for clinical applications. We utilized the Figshare brain tumor dataset, comprising 3,064 T1-weighted contrast-enhanced MRI images from 233 patients with three prevalent tumor types: meningioma, glioma, and pituitary tumor. Four pre-trained deep learning models—ResNet, EfficientNet, MobileNet, and DenseNet—were evaluated using transfer learning from ImageNet. DenseNet achieved the highest test set accuracy of 96%, outperforming ResNet (91%), EfficientNet (91%), and MobileNet (93%). Therefore, we focused on improving the performance of the DenseNet, while considering it as base model. To enhance the generalizability of the base DenseNet model, we implemented a fine-tuning approach with regularization techniques, including data augmentation, dropout, batch normalization, and global average pooling, coupled with hyperparameter optimization. This enhanced DenseNet model achieved an accuracy of 97.1%. Our findings demonstrate the effectiveness of DenseNet with transfer learning and fine-tuning for brain tumor classification, highlighting its potential to improve diagnostic accuracy and reliability in clinical settings.

Comparison of identification accuracy and training time between different training models on the PLAID dataset.

The F1 for different data balance algorithms using transferred CBAM-ResNet34.

The PLIAD dataset of appliance types and instance statistics.

This repository introduces the open-source project dubbed Tencent ML-Images, which publishes ML-Images: the largest open-source multi-label image database, including 17,609,752 training and 88,739 validation image URLs, which are annotated with up to 11,166 categories Resnet-101 model: it is pre-trained on ML-Images, and achieves the top-1 accuracy 80.73% on ImageNet via transfer learning

The comparison of the proposed method and other power fingerprint identification methods.

Margin enlargement of training data has been an important strategy for perceptrons in machine learning for the purpose of boosting the confidence of training toward a good generalization ability. Yet Breiman (1999) shows a dilemma: a uniform improvement on margin distribution does not necessarily reduce generalization errors. In this paper, we revisit Breiman's dilemma in deep neural networks with recently proposed spectrally normalized margins from a novel perspective based on phase transitions of normalized margin distributions in training dynamics. Normalized margin distribution of a classifier of the data can be divided into two parts: low/small margins such as some negative margins for misclassified samples vs. high/large margins for high confident correctly classified samples, which often behave differently during the training process. Low margins for training and test datasets are often effectively reduced in training, along with reductions of training and test errors, whereas high margins may exhibit different dynamics, reflecting the trade-off between the expressive power of models and the complexity of data. When data complexity is comparable to the model expressiveness, high margin distributions for both training and test data undergo similar decrease-increase phase transitions during training. In such cases, one can predict the trend of generalization or test error through margin-based generalization bounds with restricted Rademacher complexities, shown in two ways in this paper with early stopping time exploiting such phase transitions. On the other hand, over-expressive models may have both low and high training margins undergoing uniform improvements with a distinct phase transition in test margin dynamics. This reconfirms the Breiman's dilemma associated with over-parameterized neural networks where margins fail to predict overfitting. Experiments are conducted with some basic convolutional networks, AlexNet, VGG-16, and ResNet-18, on several datasets, including Cifar10/100 and mini-ImageNet.