Week 2 - Training with more raw images

1 minute read

Meeting summary

  • The item established for this week of generating modeled samples and raw samples has been fulfilled.
  • With this new type of data, the range of possibilities is widened so that we can play with the degree of complexity of the movement, the type of network used and the two types of images to be used both as input and output.
  • To focus the goal of the week we decided to focus on linear movement with three degrees of freedom (speed, slope and initial position), using the LSTM and ConvLSTM networks, and using raw images as input and output.
  • We will increase the number of samples in training based on the results obtained previously since the number of parameters is very high due to the use of raw images.
  • Analyzing what we know so far, it is expected that the network will not have a good performance with this type of data, so in the next steps we would change the input and output for the modeled images, thus reducing the number of parameters that manage the network.

To Do

The tasks proposed for this week are

  • Train LSTM networks with more samples in linear movement dataset with 3 degrees of freedom with more samples (100000).
  • Train ConvLSTM networks with more samples in linear movement dataset with 3 degrees of freedom with more samples(100000).
  • Network training code refactoring for new data structure.
  • [X] Re-train the networks with the 10000 sample dataset for consistency in results.

Small dataset trainings

After examining the dataset we had, the trained networks and the different results obtained, I detected that there was some lack of coherence between them, so I decided to retrain the different types of networks that we have assessed.

Although there was an image that did not match the network that was being evaluated, after repeating the training I have reached the same conclusion as months ago and all I have done is change those images in the respective part of the blog so that everything is in order and matches.

Big dataset trainings

When I tried to train the networks with the code I was using so far I have obtained a memory error so I focus on modifying the code to read the dataset by batches and be able to use it to train.