Does my linear model with variable hidden layers look ok?

1

model_help
model_help817×564 26 KB

I am trying to build a model for my final year project but I get very high loss values does this model look okay or am I doing something wrong? Just wanted to make sure before I start looking at the input data more.
Thanks for any help!

2

Hey,

can you describe in more detail your problem? What is your loss-function and what does your labels look like? High-Loss values can be the reason of selecting the wrong loss-function…

As far as I understand your code (and there is nothing more) the model looks fine.

Best regards,

~ Linus

1 Like
3

Hi thanks for the input, so essentially my problem is to take lip movements and label the word that has been said for each sequence of lip movements. The lip movements have been extracted using a gabor transform and I have combined each sequence per word with a label at the end such as ‘blue’. So say the first 7 columns are centroid x movement’s then the next 7 are centroid y and so on with all the other features. so there are around 53 different words to be classified, this is the training method I am using:

training
training1216×651 37.7 KB

4

Hi thanks for the input, so essentially my problem is to take lip movements and label the word that has been said for each sequence of lip movements. The lip movements have been extracted using a gabor transform and I have combined each sequence per word with a label at the end such as ‘blue’. So say the first 7 columns are centroid x movement’s then the next 7 are centroid y and so on with all the other features. so there are around 53 different words to be classified, this is the training method I am using:

training
training1216×651 37.7 KB

5

Hey!

Thank you for your fast answer. Regarding to the PyTorch docs:
PyTorch - Training a Classifier

You have to change your train-loop a bit: You have to zero the parameter gradients before you call your model(input). This is the first thing.

The second thing is: How does your input-data looks like (dataset, dataloader). Do you normalize your data? What is the “High-Loss” for you? If the problem is to complex and the model to simple, the model maybe cannot approximate very well given your data.

Best regards,

~ Linus