Losses come out to be None, as if the model isn't training on the data

[iHasna]

Environment: Google Colab with GPU

This is more of an issue than a bug. I'm running everything with no errors thankfully. But its as if the model is not running or finding anything in my datasets. I added my images and XML files to the same folder 'Images' like it was mentioned to do in the medium article. I do want to note that some images don't have XML files cause there was no object in the image but I believe that to be fine because the 300+ images do have the object 'table'. I even ran the utils.xml_to_csv to make sure the xml files were being read and it was. This is the result:

I then run the next few lines as shown in the tutorial:

`dataset = core.Dataset('/content/train_labels.csv', '/content/Images/')
loader = core.DataLoader(dataset, batch_size=3, shuffle=True)

model = core.Model(sni)
losses = model.fit(loader, epochs=7,
                   learning_rate=0.001, verbose=True)`

And everything runs:

I then get this error:

The losses are none? As if its not training on anything? I can't seem to see where the problem is.

[alankbi]

You have to pass in a validation dataset/loader as the second argument to model.fit (or the keyword argument val_dataset) in order for it to return a list of losses. So something like this:

dataset = core.Dataset('/content/train_labels.csv', '/content/Images/')
loader = core.DataLoader(dataset, batch_size=3, shuffle=True)

validation = ...  # initialize your validation dataset (either core.Dataset or core.DataLoader object)

model = core.Model(sni)
losses = model.fit(loader, validation, epochs=7, learning_rate=0.001, verbose=True)

[iHasna]

Ah okay thanks for letting me know. However, when I try to predict on an image I used for training I get empty lists. Code below:

(idk why it turns it yellow, its originally white.)

[alankbi]

That likely means the model wasn't trained well enough in order to detect any objects. You could try printing or plotting the returned list of losses (if you have a validation dataset specified) to see if the loss is continuously decreasing (maybe you need more epochs), or potentially add additional images or augmentations to strengthen your dataset.