Open groot-1313 opened 6 years ago
Thanks :)
The training is done in the demo ipython notebook. It looks like this:
for i in range(20):
history=detection_model.fit_generator(train_gen, 6400//batch_size,
epochs=5,
callbacks=[nan_terminator],
validation_data = val_gen,
validation_steps = 1600//batch_size,
#use_multiprocessing=False)
workers =4,
max_queue_size=24)
histories.append(history)
times.append(time.time())
I'm training 20 times on 5 epochs with 6400 samples = 640000 samples. The weird outer loop allows me to break out of the training (ctrl + c) and still keep the history of most of it.
Please note: This is just to demonstrate that my loss function works. I'm only retraining the last few layers. The feature extraction is pre-trained with transfer learning.
In general, it is very hard to make predictions about the number of training samples necessary. They will vary wildly depending on your data. I'm training on original VOC data, so this is pretty much what this architecture was made for. With other datasets, the training times go up significantly, orders of magnitude. This is due to the fact that previous layers have to be trained, too.
The final expected loss is almost impossible to predict. It depends very much on the network architecture: The yolo loss looks at all the cells defined in the output and sums the loss for every one of them. So a different resolution will change the loss. Depending on your data, you will have to tune the different weights for the loss (objectness, class, location, etc) this will also change the loss. I'm afraid, for this I can't give any number at all.
In my experience, even if you thought that the curve has flattened and no more improvement can be made, it pays to continue training. Just train 2 or 3 times as long as you would think.
Hi, (YOLOv2) Great work here! I was wondering, since the loss starts from a very high value how many epochs did it take for you achieve the results you've displayed? And, what is the final expected loss to stop the training process?