def add_visual_embedding()

in lib/modeling/VGG16_rel_softmaxed_triplet.py [0:0]
72 lines of code
9 McCabe index (conditional complexity)

def add_visual_embedding(model,
                         blob_sbj, dim_sbj,
                         blob_obj, dim_obj,
                         blob_rel_prd, dim_rel_prd,
                         blob_rel_sbj, dim_rel_sbj,
                         blob_rel_obj, dim_rel_obj):

    model.add_FC_layer_with_weight_name(
        'x_sbj_and_obj',
        blob_sbj, 'x_sbj_raw', dim_sbj, cfg.OUTPUT_EMBEDDING_DIM,
        weight_init=('GaussianFill', {'std': 0.01}),
        bias_init=('ConstantFill', {'value': 0.}))
    model.add_FC_layer_with_weight_name(
        'x_sbj_and_obj',
        blob_obj, 'x_obj_raw', dim_obj, cfg.OUTPUT_EMBEDDING_DIM,
        weight_init=('GaussianFill', {'std': 0.01}),
        bias_init=('ConstantFill', {'value': 0.}))
    if cfg.MODEL.SUBTYPE.find('w_ishans') == 0:
        model.FC(
            blob_rel_prd, 'x_rel_prd_raw_1',
            dim_rel_prd, 4 * cfg.OUTPUT_EMBEDDING_DIM,
            weight_init=('GaussianFill', {'std': 0.01}),
            bias_init=('ConstantFill', {'value': 0.}))
        model.LeakyRelu('x_rel_prd_raw_1', 'x_rel_prd_raw_1', alpha=0.1)
        model.FC(
            'x_rel_prd_raw_1', 'x_rel_prd_raw_2',
            4 * cfg.OUTPUT_EMBEDDING_DIM, 2 * cfg.OUTPUT_EMBEDDING_DIM,
            weight_init=('GaussianFill', {'std': 0.01}),
            bias_init=('ConstantFill', {'value': 0.}))
        model.LeakyRelu('x_rel_prd_raw_2', 'x_rel_prd_raw_2', alpha=0.1)
        model.FC(
            'x_rel_prd_raw_2', 'x_rel_prd_raw_3',
            2 * cfg.OUTPUT_EMBEDDING_DIM, cfg.OUTPUT_EMBEDDING_DIM,
            weight_init=('GaussianFill', {'std': 0.01}),
            bias_init=('ConstantFill', {'value': 0.}))
        model.net.Alias('x_rel_prd_raw_3', 'x_rel_prd_raw')
    else:
        model.FC(
            blob_rel_prd, 'x_rel_prd_raw_1',
            dim_rel_prd, cfg.OUTPUT_EMBEDDING_DIM,
            weight_init=('GaussianFill', {'std': 0.01}),
            bias_init=('ConstantFill', {'value': 0.}))
        if cfg.MODEL.SUBTYPE.find('w_ishans') == 0:
            model.Relu('x_rel_prd_raw_1', 'x_rel_prd_raw_1')
            model.FC(
                'x_rel_prd_raw_1', 'x_rel_prd_raw_2',
                cfg.OUTPUT_EMBEDDING_DIM, cfg.OUTPUT_EMBEDDING_DIM,
                weight_init=('GaussianFill', {'std': 0.01}),
                bias_init=('ConstantFill', {'value': 0.}))
            model.net.Alias('x_rel_prd_raw_2', 'x_rel_prd_raw')
        else:
            model.net.Alias('x_rel_prd_raw_1', 'x_rel_prd_raw')
    model.net.Normalize('x_sbj_raw', 'x_sbj')
    model.net.Normalize('x_obj_raw', 'x_obj')
    model.net.Normalize('x_rel_prd_raw', 'x_rel_prd')

    # get x_rel_sbj and x_rel_obj for the p branch
    if model.train and cfg.MODEL.SUBTYPE.find('embd_fusion') >= 0:
        model.add_FC_layer_with_weight_name(
            'x_sbj_and_obj',
            blob_rel_sbj, 'x_rel_sbj_raw',
            dim_rel_sbj, cfg.OUTPUT_EMBEDDING_DIM,
            weight_init=('GaussianFill', {'std': 0.01}),
            bias_init=('ConstantFill', {'value': 0.}))
        model.add_FC_layer_with_weight_name(
            'x_sbj_and_obj',
            blob_rel_obj, 'x_rel_obj_raw',
            dim_rel_obj, cfg.OUTPUT_EMBEDDING_DIM,
            weight_init=('GaussianFill', {'std': 0.01}),
            bias_init=('ConstantFill', {'value': 0.}))
        x_rel_sbj = model.net.Normalize('x_rel_sbj_raw', 'x_rel_sbj')
        x_rel_obj = model.net.Normalize('x_rel_obj_raw', 'x_rel_obj')
        if cfg.MODEL.SPECS.find('not_stop_gradient') < 0:
            # this is to stop gradients from x_rel_sbj and x_rel_obj
            model.StopGradient(x_rel_sbj, x_rel_sbj)
            model.StopGradient(x_rel_obj, x_rel_obj)