def encode_lowshot_trainset()

in matching_network.py [0:0]

• 30 lines of code
• 2 McCabe index (conditional complexity)

def encode_lowshot_trainset(model, base_classes, train_file_handle, novel_idx, lowshotn, num_base=100):
    all_labels = train_file_handle['all_labels'][...]
    all_feats = train_file_handle['all_feats']

    feats = []
    Y = []
    #for each base class, randomly pick 100 examples
    for i, k in enumerate(base_classes):
        idx = np.where(all_labels==k)[0]
        idx = np.sort(np.random.choice(idx, num_base, replace=False))
        feats.append(all_feats[list(idx)])
        Y_this = np.zeros((num_base,1000))
        Y_this[:,k] = 1
        Y.append(Y_this)

    #next get the novel classes
    sorted_novel_idx = np.sort(novel_idx.reshape(-1))
    novel_feats = all_feats[list(sorted_novel_idx)]
    novel_labels = all_labels[sorted_novel_idx]
    Y_novel = np.zeros((novel_feats.shape[0],1000))
    Y_novel[np.arange(novel_feats.shape[0]), novel_labels] = 1

    num_repeats = int(np.ceil(float(num_base)/float(lowshotn)))
    novel_feats = np.tile(novel_feats, (num_repeats,1))
    Y_novel = np.tile(Y_novel, (num_repeats,1))


    feats.append(novel_feats)
    Y.append(Y_novel)

    feats = np.concatenate(feats, axis=0)
    Y = np.concatenate(Y, axis=0)

    model = model.cuda()
    feats = Variable(torch.Tensor(feats).cuda())
    Y = Variable(torch.Tensor(Y).cuda())


    G, G_norm = model.encode_training_set(feats)
    print(novel_feats.shape, len(base_classes))
    return G, G_norm, Y