tivaro / cavfefe

Cavfefe

TODOs

• If you download a checkpoint with a BasicLSTMCell (such as the language model and the sentence classifier) from the DAS, it will not be accepted by our local Tensorflow installations because the variable names were changed after Tensorflow 1.0. Therefore, run fix_basic_lstm.sh with the path to the checkpoint (e.g. pretrained/sc/SentenceClassifier) to replace the old names with the new ones.

Preparing the Dataset

Building a Vocabulary

For reasons of efficiency, we start by mapping words (literals) to indices (integers), and save the resulting vocabulary to disk so we can use it from several implementations (the sentence classifier and the language model).

python cavfefe.py --ds-build-vocabulary

Computer Vision Model

Preparing the Dataset

The vision model accepts only 299x299 images with 3 channels. In addition, we also need to split the training and the testset.

python cavfefe.py --cv-prepare

Training

Once we have prepared the dataset, we can train the model by running the command below. Note that this will only train the logits layer (the last layer), which means that even the features that will be extracted later on, remain unchanged.

python cavfefe.py --cv-train

Evaluation

After training, it is important to verify that the model is behaving correctly before proceeding to integrating the entire project. The accuracy of the CNN can be computed in this intermediate stage by running the command below.

python cavfefe.py --cv-evaluate

Feature Extraction

We want to extract the features from the vision model so we can concatenate those in the language model. The command below extracts both features of images in the training set as well as the test set and saves both to disk.

python cavfefe.py --cv-extract-features

Sentence Classifier

Preparing the Dataset

Again, we first need to preprocess the dataset. In this case, we need to pad the sentences to a fixed maximum length (default: 30). Running the command below requires that the vocabulary has already been built (step #1).

python cavfefe.py --sc-prepare

Training

The sentence classifier can be trained by running the following command. In our experience, the classifier takes a bit longer than usual before beginning to converge. There is also a significant variance between the accuracy after convergence: we suspect that it gets more easily stuck in local optima than other models.

python cavfefe.py --sc-train

Evaluation

Again, we need to evaluate the accuracy of the sentence classifier before proceeding with the rest of the project. The authors of the paper mention they achieved 22% classification error on their validation set. We managed to get 21.754% which is close enough. Note that our model uses 512 hidden units (instead of 1000).

python cavfefe.py --sc-evaluate

Embedding Extraction

In addition to the features we retrieve from the computer vision model, we also want to concatenate a class embedding to the input of the language model. The authors of the paper take the average over all activations of sentences in the training set per class. For each class, we have 30 examples with 10 sentences each, so for each class we take the average over 300 512-dimensional vectors and end up with 200 512-dimensional vectors, one for each class. This command writes the embeddings to disk.

python cavfefe.py --sc-extract-embeddings