No validation set detected with validation_split:0
Philip642 opened this issue · comments
Hi Jake,
I have now over 3h video footage from different animals doing different behaviors and I got the message No validation set detected with validation_split:0 so I changed it to 0.1.
Fit the model works with these settings:
model.fit(
batch_size=12, #less than n_train from data generator
validation_batch_size=1, # less than n_val from data generator
callbacks=callbacks,
#epochs=1000, # Increase the number of epochs to train the model longer
epochs=200,
n_workers=1,
steps_per_epoch=500,
)
and the output from the training generator looks like this:
{'n_train': 19,
'n_validation': 2,
'validation_split': 0.1,
'downsample_factor': 3,
'output_shape': (48, 48),
'n_output_channels': 88,
'shuffle': True,
'sigma': 5,
'output_sigma': 0.625,
'use_graph': True,
'graph_scale': 1,
'random_seed': 1,
'augmenter': True,
'datapath': '/home/einat/deepposekit-data/datasets/whipspider/annotation_data_release.h5',
'dataset': 'images',
'generator': 'DataGenerator',
'n_samples': 21,
'image_shape': (384, 384, 1),
'keypoints_shape': (43, 2)}
Output from fit the model
Epoch 1/200
13/500 [..............................] - ETA: 2:36:46 - loss: 178.4359 - output_0_loss: 71.6774 - output_1_loss: 56.4588 - output_2_loss: 50.2997
Epoch 1/200
405/500 [=======================>......] - ETA: 29:31 - loss: 52.7862 - output_0_loss: 19.2698 - output_1_loss: 17.0771 - output_2_loss: 16.4393
Does this looks good?
Cheers, Philip
Yes, if you set validation split to 0 then there will be no validation set. If the loss is decreasing then training is working. You should assess the quality of the results using the notebooks for step 4.