Try Kaggle, they offer 30 hours of TPU per week which might help if you are working with tensorflow or torch (TPUs have 128GB of memory). What you should try when running into memory issues is to use half precision floats which reduces memory requirements.
Here are a few options:
Commercial cloud provider: AWS is the go-to cloud provider for virtually any need, they have options for everything. Check out their options for Deep Learning.
You could also buy your own hardware
Finally from a data science perspective there is also the option of modifying the parameters of the model and/or data so that the training ...
So, I understand the question is asking what data should you use, when given new data.
You might want to provide more context to your question as this might seem like a loose answer. The way you go forward with this depends on a couple of factors:
Is the initial dataset ($D_1$) have the same features/columns as the new dataset ($D_2$), such that you can ...
It can assume also other meaning but the learning rate schedule process. For example in YOLOv3, during warm up epochs the ground truth bounding boxes are forced to be the same size of the anchors.
At the end of the day, the warm up procedure aim to soften the impact of the first epochs of learning that can mislead the entire training process. This is not ...
It seems like both models are on par in terms of performance. It will be interesting to see what would happen if you combine predictions from both models via a simple average. In many cases an ensemble of models has shown to yield better performance than any single model.
So maybe it turns out you want to keep the best of both worlds by averaging.
The questions I ask myself when I see learning graphs are the following ones:
Is the loss decreasing and the accuracy increasing ? if yes, your network is learning and everything works fine, which is already a good new.
Have we reached a kind of plateau ? (in accuracy especially), which means learning is over. (Here maybe you could train your network a bit ...
The loop for the validation data would look very similar to your training loop, but for your validation data you only have to calculate the loss and not backpropagate the error. It would look something like this:
BATCH_SIZE = 64
EPOCHS = 10
for epoch in range(EPOCHS):
# training loop
With Keras, you could use the functional API, to estimate a model with two outputs („multioutput“). Simply train the model on two outputs like:
out1 = Dense(1)(x)
out2 = Dense(1)(x)
# Compile/fit the model
model = Model(inputs=Input_1, outputs=[out1,out2])
model.compile(optimizer = "rmsprop", loss = 'mse')
# Add actual data here in the ...