This was one of my first projects in Tensorflow and therefore neither is very elegent, nor very efficiently implemented. <\br> The trained weights of the network are missing.

The task of the 2016 PlantCLEF challenge (https://www.imageclef.org/lifeclef/2016/plant) was to predict the class and probability of a flower displayed on an image.

The code is based upon the research presented in
Šulc, M., Mishkin, D. & Matas, J. (2016). Very deep residual networks with maxout for plant identifcation in the wild. Working notes of CLEF 2016 conference. (http://ceur-ws.org/Vol-1609/16090579.pdf)

A Deep Residual Neural Network to identify flowers is created. The network is based on a pre-trained model within the Tensorflow framework (https://github.com/tensorflow/models/tree/master/research/slim).

To fine-tune the existing weights, images of the PlantCLEF dataset were used.

• plantclef_download_and_convert_data This script downloads the given dataset of the PlantCLEF challenge 2016 and converts it into tfRecords. It saves the image, species_label, genus_label, family_label, organ_label (see changed_scripts/dataset_utils: image_to_tfexample()).
• changed_scripts/resnet_utils and changed_scripts/resnet_v2 Contains the architecture of the Residual Networks. Methods containing "intermediate" in their name are changed, so that they can be cut of at any level.
• resNetClassifier Example on how to train and evaluate a network. The network architecture is further edited (my_cnn()) and initialized (get_init_fn()). The code contains the load_batch() method, here the data is loaded and preprocessed (see my_resnet_preprocessing). Probably the method as to be adjusted, when it was created the tfRecords only contained one label \glqq label\grqq{*, not four! Just load the one you need, or all (see layerwise_generalization's load_batch method for more insights).
• train_my_network script calling resNetClassifier to train the models. Also contains the method final_evaluation() and final_evaluation_generic(). The first uses three networks to make a prediction, using the 6 time data augmentation. The latter uses n networks, but you have to specify which preprocessing method you want to use.
• my_resnet_preprocessing Here the preprocessing is done. One for training, one for evaluation and one for the final evaluation. You can call the first two by using preprocess_image(). There are deprecated methods included, the most recent ones are preprocess_for_train() and preprocess_for_eval(). Always change the dtype of the image to float (tf.convert_image_dtype), int type and float type preprocessings for train and eval don't work! You can find more preprocessing methods in the Tensorflow documentation under "Image".
• zeiler_fergus Contains code calling tf_cnnvis, this github repo contains code to visualize individual layers of your network.
• my_functions Contains additional functions, e.g. the MaxOut function used in the network introduced by the CMP group in the PlantCLEF challenge 2016.
• DataVisualisation Contains the splitting method used to get the tfRecords you want to use. Note: the whole script besides the show_example() method should be moved to my_functions or another script. (It is still in this script, because I didn't want to mess my code up in the last days). At least rename it :D
• calculate_optimal_stimuli Goes through the dataset and finds the image which has the highest softmax output for a class. Saves it into a text file.
• freezer Method used to freeze a graph. This means, save its weights and architecture together into a .pb file. This can then be used to be deployed e.g. on a smartphone.
• crop_acts zeiler_fergus creates visualizations for all channels of one layer, this method now crops out the specified channels and saves them into a grid (9 images needed, change the code for more or less)
• grid_plotter Plots 9 images into a grid, good for visualizations ;)
• kernel_visualization Contains methods from the DeepDream tutorial of Tensorflow. The outputs seem reasonable but did not really give me insights into what the network learned. The outputs for all classes almost looked the same. I'm not sure whether I messed up something, it does not work as well with the ResNet or if the outputs actually look like this, when finetuned on the plantclef-challenge!
• layerwise_generalization Similar to ResNetClassifier, but trains ONLY ONE LAYER: speciefied in train_op: variable_to_train !\ Used to train the fc-softmax-classifiers afther a cutoff of the network. The methods train_and_eval() as well as train_and_eval_slim() are training and evaluating the new layers and saving the results into a text-file.
• maximum_pathway Only the methods fin_max_activation_per_filter() and top_channel_per_layer() are interesting. The find the 9 images which excite each channel of each layer the most and save them as a text file. The latter now searches for the top activated channels in each layer and saves it into a text-file. The txt-files are not really readable (see max_act_worker). Need to be changed.
• max_act_worker This script is only necessary because the max_activation.txt is not readable (see maximum_pathway). Change the file, maybe into a csv or something similar.

The best network resulted in a Top-1 accuracy of 64.48% and a top-5 accuracy of 82.73%. This network only used scaled, ramdom cuts of the images to train. Further preprocessing did not improve the results.