A Complete Guide to Deep Learning CNN - TensorFlow Keras Version
From core theories of Deep Learning and CNN to implementation methods of various CNN models, and practical Deep Learning development know-how through real-world problems, If you want to become a Deep Learning CNN technology expert, join us in this lecture :)
2,115 learners
Level Basic
Course period Unlimited
Notice of uploading of modified practice code according to Tensorflow version update.
Hello, the Tensorflow version of Kaggle, the practice environment, has been upgraded a lot compared to when the lecture was created.
I have revised the practice code to reflect this. I am in the process of making a video to announce the changes, but it is taking longer than expected, so I will upload the revised practice code first. I will also make a video to announce the changes soon.
If an error occurs while practicing using the existing practice code, please download the new practice code and try performing it again.
Below is a summary of the changes that have been implemented. Thank you.
0. Common
- The initialization argument name of the Adam() optimizer was changed from lr to learning_rate.
- In the getitem () method of the Sequence Dataset class, the return result is separated into a dataset that includes or does not include the Target value depending on the learning and testing time.
- When inputting the training and validation data sets as Sequence Dataset arguments to model.fit (), exclude the steps_per_epoch and validation_steps arguments.
1. Fashion_Mnist_Practice
- When save_weights_only=True of ModelCheckpoint, the checkpoint model file name must end with .weights.h5. When False, it must end with .keras.
- period is changed to save_freq. save_freq can be "epoch" or an integer value. In the case of an integer value, it is not an epoch but the number of repetitions of the learning batch, and there is a risk of error when applying.
2. Inception_Practice
- With the Tensorflow version up, the string '/' cannot be entered in the name argument of the Conv and Maxpooling layers. Change it to '_'.
3. Learning_Rate_Scheduler
- model.optimizer.lr is converted to model.optimizer.learning_rate.
- Convert the y target value input from np.zeros(5) to 2D behavior with np.zeros(5).reshape(5, -1).
- CosineDecay is no longer experimental but is registered as a regular scheduler. Therefore, experimental.CosineDecay is changed to optimizers.schedules.CosineDecay.
- The bonus code for Cosine Decay and Cosine Decay Restart implemented as callbacks is no longer valid. tf.keras.backend.set_value is deprecated and no longer valid.
4. Plant_Pathology
-Rearranged source code to facilitate execution of source code when kernel is restarted to free GPU memory.