Tensorboard is a visualization tool with very, very many functions, and it is very very simple to implement, which can help us to view
- The progress of model training, such as loss and acc (used in the classification network, only look at loss in target detection)
- Weight histogram
- Gradient histogram
- The structure of the entire model
specific methods:
1. Import tensorboard in the library of the callback function
from keras.callbacks import TensorBoard
2. Create a TensorBoard before training the model. These TensorBoards need to specify some parameters. You can specify the callbacks parameters in fit during training:
Tensorboard= TensorBoard(log_dir="./model", histogram_freq=1,write_grads=True)
history=model.fit·····
There are 7 commonly used parameters:
- 1. log_dir: The location used to save Tensorboard log files and other content
- 2. histogram_freq: Calculate the frequency of activation value and model weight histogram for each layer in the model.
- 3. write_graph: Whether to visualize the image in TensorBoard (whether to draw the model's picture).
- 4. write_grads: Whether to visualize the histogram of gradient values in TensorBoard.
- 5. batch_size: The input batch size of the afferent neuron network used for histogram calculation.
- 6. write_images: Whether to visualize the model weights as pictures in TensorBoard.
- 7. update_freq: The three commonly used values are'batch','epoch' or integers. When
using'batch', write the loss and evaluation value into TensorBoard after each batch . 'epoch' is similar. If an integer is used, the loss and evaluation value will be written into
TensorBoard after every certain number of samples .
Show off
-
Navigate to the location where the model file is saved
-
Open the cmd terminal and cd to the upper level directory of the model file
-
Execute the following command in the upper level directory of the model folder: tensorboard --logdir=./model, press Enter to automatically load the URL
-
Copy the URL and open the visual interface in Firefox or Google browser