I'm running a CNN for image classification. Every 100 steps, a file is created in a folder either as: model.ckpt-0.data-00000-of-00001, model.ckpt-0.index, model.ckpt-0.meta. There are also these files: graph.pbtxt and checkpoint.
Which of these files would I use to view the accuracy of my training model in Tensorboard?
None of these contain the accuracy values, they are the definition of the model (graph.pbtxt) and the model weights (checkpoint / ckpt files).
By default the fit method will output any losses or metrics (e.g. accuracy) you defined when you called compile on the model, e.g.
model.compile(optimizer="Adam", loss="mse", metrics=["mae", "acc"])
will compile the model with the mse loss and the mae and acc metrics. The values will be printed at the end of each epoch, or more often if you change the verbose argument when calling fit
Perhaps the best way to visualise these values is to use Tensorboard. To do this you create a tensorboard callback (a callback is a class with methods that are called at the start / end of training, epoch and batch) which will write the metrics and other info into the training directory.
Then you may run tensorboard from inside the training directory, e.g. tensorboard --logdir=/path/to/training/dir to get a nice web-based UI in which to monitor training.
