I was finally able to fine-tune a model and get decent results. I’ve used a new dataset from kaggle. This has more and simple images of water surfaces so that helped a lot. Results The model was able to learn…
I trained L-RASPP again to detect cracks with 110 images.The results are the above. Comparing to Deeplabv3(Image Below),trained with the same dataset, you can see there are a lot of noise in the results from the L-RASPP model. > Reference:…
I retrained a L-RASPP(Light-Reduced Astrous Spatial Pyramid Pooling) model to detect water level using Pytorch. I used the WSOC dataset containing 496 images for training and 278 images for testing. The ROC score was close to 1 by epoch 25.…
You can use click when you want to add arguments when calling your python file. Let’s say we want to specify the following when running your python code which will train your model. – Data Directory – Number of Epochs…
For a Tensorflow model, you can use .summary() to see the structure of your model. In pytorch, there is no such thing. So instead, you can use torchinfo. This was really useful when I wanted to do transfer-learning but didn’t…
Another useful regularization technique is called Max-Norm Regularization. Implementation layer = keras.layers.Dense(100, activation=”selu”, kernel_initializer=”lecun_normal”, kernel_constraint=keras.constraints.max_norm(1.)) By setting a hyper-parameter r you can set a max value for weights to prevent over-fitting. References: Hands-On Machine Learning with Scikit-Learn, Keras, and TensorFlow,…
Dropout is one of the most popular regularization techniques for deep neural networks. Monte Carlo Dropout may help boost the dropout model even more. Full Implementation ys = np.stack([model(X_test, training=True) for sample in range(100)]) y = ys.mean(axis=0) Predict() method returns…
Transposed Convolutional Layer is doing the following: Stretch Image by inserting rows and columns with zeros. Perform regular convolution You can think of it as a fractional stride, for example, 1 This is often used when doing semantic segmentation, where…
Load model model_A = keras.models.load_model(“model_A.h5″) Clone Architecture model_A_clone = keras.models.clone_model(model_A) Clone Weights model_A_clone.set_weights(model_A.get_weights()) Delete Last Layer model_B = keras.models.Sequential(model_A_clone.layers[:-1]) Add Final layer => Change to Binary classifier model_B.add(keras.layers.Dense(1, activation=”sigmoid”)) You can prevent copied layers to be affected when training for…
One way to tune your keras model is using random search. Here is one example with sklearn. from scipy.stats import reciprocal from sklearn.model_selection import RandomizedSearchCV #Specify range for random initialization param_distribs = { “n_hidden”: [0, 1, 2, 3], “n_neurons”: np.arange(1,…