Deep Learning Hyperparameter Tuning in Python, TensorFlow & Keras

[@GregHogg]

Take my courses at mlnow.ai/!

[@masteronepiece6559]

Next time summarize the results in a table in the last of the video. We're busy to watch the whole video.

[@billybobandboshow]

Thank you for this video! I have been learning about deep learning algorithms over the holiday break! Hope we see more videos from you! I love your channel and content! Keep up the awesome work, happy holidays and happy new year! :)

[@GregHogg]

You're very welcome and thanks so much for the kind words! Awesome work, happy new year!!

[@iftekharanam8980]

That was excellent. Need more videos on DL

[@tomaszzielonka9808]

@GreggHogg Hi, I got stuck with keras tuner. It seems that code below will only only create the function 'model_builder' once. If I change anything like add Dropout layer and rerun the function it keeps displaying the comment (see below the code), like it was consistenly reaching to the first version of function. Any clues on how to fix that? I would like to experiment with the 'model_builder' function (add/remove layers, dropouts, etc) and then observe what parameters tuner generates. def model_builder(hp) : model = Sequential() hp_activation = hp.Choice('activation', values = ['relu', 'tanh']) hp_layer_1 = hp.Int('layer_1', min_value = 2, max_value = 32, step = 2) hp_layer_2 = hp.Int('layer_2', min_value = 2, max_value = 32, step = 2) hp_learning_rate = hp.Choice('learning_rate', values = [1e-2, 1e-3, 1e-4]) model.add(Dense(units = hp_layer_1, activation = hp_activation)) model.add(Dense(units = hp_layer_2, activation = hp_activation)) model.add(Dense(units = 1, activation = 'sigmoid')) model.compile(optimizer = tf.keras.optimizers.Adam(learning_rate = hp_learning_rate), loss = 'binary_crossentropy', metrics = [tf.keras.metrics.Recall()]) return model tuner = kt.Hyperband(model_builder, objective = kt.Objective("val_recall", direction = "max"), max_epochs = 50, factor = 3, seed = 42) Comment : Reloading Tuner from .\untitled_project\tuner0.json

[@tigjuli]

Simple explanation, awesome video!

[@GregHogg]

Thank you!

[@rudrathakkar56]

Thank you . I am learning deep learning .This helped me much

[@GregHogg]

Perfect - Really glad to hear it!

[@arsheyajain7055]

Awesome video!!

[@GregHogg]

Thanks a bunch Arsheya! Hope you're having a great holiday break :)

[@dakshbhatnagar]

Great Video Man but tbh I was actually expecting some sort of automation of the hyperparameter tuning.

[@GregHogg]

kzbin.info/www/bejne/bH_JYqttprmbiJo

[@dakshbhatnagar]

@@GregHogg thanks

[@prabinbasyal1049]

Can you suggest data science course? I already read numpy,pandas and matplotlib.

[@GregHogg]

Awesome! IBM Data science is a great intro. Big big fan of Andrew Ng's deep learning as well.

[@haneulkim4902]

Thanks for an amazing video! Is there way to tune hyperparameters like in sklearn w/o using keras-tuner?

[@GregHogg]

You're very welcome! I'm sure there is, although I don't believe I've done it before

[@tigjuli]

yes, there is. you have to define a model as a function and use KerasClassifier from keras as a wrapper to work with sklearn's GridSearch or Ramdomized search. I'm sure there are videos on youtube

[@luisalbertoburbano9295]

good afternoon, I have a task and I have not been able to create the keras tuner for 5000 rows with 4 columns, in each column the numbers are random from 0 to 9 and I need an output of only 4 numbers this is the code # Initialising the RNN model = Sequential() # Adding the input layer and the LSTM layer model.add(Bidirectional(LSTM(neurons1, input_shape=(window_length, number_of_features), return_sequences=True))) # Adding a first Dropout layer model.add(Dropout(0.2)) # Adding a second LSTM layer model.add(Bidirectional(LSTM(neurons2, input_shape=(window_length, number_of_features), return_sequences=True))) # Adding a second Dropout layer model.add(Dropout(0.2)) # Adding a third LSTM layer model.add(Bidirectional(LSTM(neurons3, input_shape=(window_length, number_of_features), return_sequences=True))) # Adding a fourth LSTM layer model.add(Bidirectional(LSTM(neurons4, input_shape=(window_length, number_of_features), return_sequences=False))) # Adding a fourth Dropout layer model.add(Dropout(0.2)) # Adding the first output layer with ReLU activation function model.add(Dense(output_neurons, activation='relu')) # Adding the last output layer with softmax activation function model.add(Dense(number_of_features, activation='softmax')) Thank you very much

[@AllanAlmeidaOficial]

GPT, Google, Stack Overflow...

[@BB-2383]

Side comment - we divide x by 255, because the image is grayscale. An RGB of white is (255,255,255), so we are converting the values upon dividing to (1,1,1), then leaving black as (0,0,0). So, an important note when training images is first convert the images to grayscale.

[@GregHogg]

Yes thank you ☺️

[@no-name168]

it has nothing to do with rgb. rgb is 3 channels, grayscale is 1. you scale both to get "normal" value range because apparently model's learning process works better on scaled values. also you do not always convert to grayscale