DeepRecipe

step1: download images from here wget http://data.csail.mit.edu/im2recipe/recipe1M_images_val.tar. This thing is 20GB, so takes quite a while.

step2: upload dataset.tar.gz to DeepRecipe/ and then decompress UPDATE there is a new pickle available with parsed ingredient list!!

step3: modify default.json, point the val/ folder and the pickles correctly.

step4: run try_dataloader.ipynb

Using GloVE vectors

• download: wget http://nlp.stanford.edu/data/wordvecs/glove.6B.zip
• preprocess using process_glove_embedding.ipynb
• visualize things using visualize_ingredient_glove_vector.ipynb
• the dataloader takes glove.py to use the gloVE vectors

Models

Currently there are 2 models: naive (image -> ingredient categories); attention (image -> masked embedding res_attn.py) Training is by pretrain_resnet_tagger.py and pretrain_resnet_embed.py which is similar to experiment.py and model handler. They use slightly different dataloader options and collate function. It's not elegant :(

see try_dataloader to find explanations

With the new design:

• Whenver you create a new model, create a new class for that architecture.
• Now add this new architecture to the models folder and the corresponding config json to the json folder.
• Note that the config has a parameter called mode - which tells the model to be used to the modelHandler
• So you have to come up with a string name. Use this name as identifier in the config json for the mode key and make the necessary updation in the model factory to create your class.