lib/models/resnet_video.py [159:254]: - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - if cfg.MODEL.DEPTH in [50, 101]: blob_in, dim_in = resnet_helper.res_stage_nonlocal( model, res_block, max_pool, 64, 256, stride=1, num_blocks=n1, prefix='res2', dim_inner=dim_inner, group=group, use_temp_convs=use_temp_convs_set[1], temp_strides=temp_strides_set[1]) layer_mod = cfg.NONLOCAL.LAYER_MOD if cfg.MODEL.DEPTH == 101: layer_mod = 2 if cfg.NONLOCAL.CONV3_NONLOCAL is False: layer_mod = 1000 blob_in = model.MaxPool(blob_in, 'pool2', kernels=[2, 1, 1], strides=[2, 1, 1], pads=[0, 0, 0] * 2) if cfg.MODEL.USE_AFFINE is False: blob_in, dim_in = resnet_helper.res_stage_nonlocal( model, res_block, blob_in, dim_in, 512, stride=2, num_blocks=n2, prefix='res3', dim_inner=dim_inner * 2, group=group, use_temp_convs=use_temp_convs_set[2], temp_strides=temp_strides_set[2], batch_size=batch_size, nonlocal_name='nonlocal_conv3', nonlocal_mod=layer_mod) else: crop_size = cfg.TRAIN.CROP_SIZE blob_in, dim_in = resnet_helper.res_stage_nonlocal_group( model, res_block, blob_in, dim_in, 512, stride=2, num_blocks=n2, prefix='res3', dim_inner=dim_inner * 2, group=group, use_temp_convs=use_temp_convs_set[2], temp_strides=temp_strides_set[2], batch_size=batch_size, pool_stride=pool_stride, spatial_dim=(int(crop_size / 8)), group_size=4, nonlocal_name='nonlocal_conv3', nonlocal_mod=layer_mod) layer_mod = cfg.NONLOCAL.LAYER_MOD if cfg.MODEL.DEPTH == 101: layer_mod = layer_mod * 4 - 1 if cfg.NONLOCAL.CONV4_NONLOCAL is False: layer_mod = 1000 blob_in, dim_in = resnet_helper.res_stage_nonlocal( model, res_block, blob_in, dim_in, 1024, stride=2, num_blocks=n3, prefix='res4', dim_inner=dim_inner * 4, group=group, use_temp_convs=use_temp_convs_set[3], temp_strides=temp_strides_set[3], batch_size=batch_size, nonlocal_name='nonlocal_conv4', nonlocal_mod=layer_mod) blob_in, dim_in = resnet_helper.res_stage_nonlocal( model, res_block, blob_in, dim_in, 2048, stride=2, num_blocks=n4, prefix='res5', dim_inner=dim_inner * 8, group=group, use_temp_convs=use_temp_convs_set[4], temp_strides=temp_strides_set[4]) else: raise Exception("Unsupported network settings.") blob_out = model.AveragePool(blob_in, 'pool5', kernels=[pool_stride, 7, 7], strides=[1, 1, 1], pads=[0, 0, 0] * 2) if cfg.TRAIN.DROPOUT_RATE > 0 and test_mode is False: blob_out = model.Dropout( blob_out, blob_out + '_dropout', ratio=cfg.TRAIN.DROPOUT_RATE, is_test=False) if split in ['train', 'val']: blob_out = model.FC( blob_out, 'pred', dim_in, cfg.MODEL.NUM_CLASSES, weight_init=('GaussianFill', {'std': cfg.MODEL.FC_INIT_STD}), bias_init=('ConstantFill', {'value': 0.}) ) elif split == 'test': blob_out = model.ConvNd( blob_out, 'pred', dim_in, cfg.MODEL.NUM_CLASSES, [1, 1, 1], strides=[1, 1, 1], pads=[0, 0, 0] * 2, ) if split == 'train': scale = 1. / cfg.NUM_GPUS softmax, loss = model.SoftmaxWithLoss( [blob_out, labels], ['softmax', 'loss'], scale=scale) elif split == 'val': #in ['test', 'val']: softmax = model.Softmax(blob_out, 'softmax', engine='CUDNN') loss = None elif split == 'test': # fully convolutional testing blob_out = model.Transpose(blob_out, 'pred_tr', axes=(0, 2, 3, 4, 1,)) blob_out, old_shape = model.Reshape( blob_out, ['pred_re', 'pred_shape5d'], shape=(-1, cfg.MODEL.NUM_CLASSES)) blob_out = model.Softmax(blob_out, 'softmax_conv', engine='CUDNN') blob_out = model.Reshape( [blob_out, 'pred_shape5d'], ['softmax_conv_re', 'pred_shape2d'])[0] blob_out = model.Transpose(blob_out, 'softmax_conv_tr', axes=(0, 4, 1, 2, 3)) blob_out = model.net.ReduceBackMean( [blob_out], ['softmax_ave_w']) blob_out = model.ReduceBackMean( [blob_out], ['softmax_ave_h']) softmax = model.ReduceBackMean( [blob_out], ['softmax']) loss = None return model, softmax, loss - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - lib/models/resnet_video_org.py [159:254]: - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - if cfg.MODEL.DEPTH in [50, 101]: blob_in, dim_in = resnet_helper.res_stage_nonlocal( model, res_block, max_pool, 64, 256, stride=1, num_blocks=n1, prefix='res2', dim_inner=dim_inner, group=group, use_temp_convs=use_temp_convs_set[1], temp_strides=temp_strides_set[1]) layer_mod = cfg.NONLOCAL.LAYER_MOD if cfg.MODEL.DEPTH == 101: layer_mod = 2 if cfg.NONLOCAL.CONV3_NONLOCAL is False: layer_mod = 1000 blob_in = model.MaxPool(blob_in, 'pool2', kernels=[2, 1, 1], strides=[2, 1, 1], pads=[0, 0, 0] * 2) if cfg.MODEL.USE_AFFINE is False: blob_in, dim_in = resnet_helper.res_stage_nonlocal( model, res_block, blob_in, dim_in, 512, stride=2, num_blocks=n2, prefix='res3', dim_inner=dim_inner * 2, group=group, use_temp_convs=use_temp_convs_set[2], temp_strides=temp_strides_set[2], batch_size=batch_size, nonlocal_name='nonlocal_conv3', nonlocal_mod=layer_mod) else: crop_size = cfg.TRAIN.CROP_SIZE blob_in, dim_in = resnet_helper.res_stage_nonlocal_group( model, res_block, blob_in, dim_in, 512, stride=2, num_blocks=n2, prefix='res3', dim_inner=dim_inner * 2, group=group, use_temp_convs=use_temp_convs_set[2], temp_strides=temp_strides_set[2], batch_size=batch_size, pool_stride=pool_stride, spatial_dim=(int(crop_size / 8)), group_size=4, nonlocal_name='nonlocal_conv3', nonlocal_mod=layer_mod) layer_mod = cfg.NONLOCAL.LAYER_MOD if cfg.MODEL.DEPTH == 101: layer_mod = layer_mod * 4 - 1 if cfg.NONLOCAL.CONV4_NONLOCAL is False: layer_mod = 1000 blob_in, dim_in = resnet_helper.res_stage_nonlocal( model, res_block, blob_in, dim_in, 1024, stride=2, num_blocks=n3, prefix='res4', dim_inner=dim_inner * 4, group=group, use_temp_convs=use_temp_convs_set[3], temp_strides=temp_strides_set[3], batch_size=batch_size, nonlocal_name='nonlocal_conv4', nonlocal_mod=layer_mod) blob_in, dim_in = resnet_helper.res_stage_nonlocal( model, res_block, blob_in, dim_in, 2048, stride=2, num_blocks=n4, prefix='res5', dim_inner=dim_inner * 8, group=group, use_temp_convs=use_temp_convs_set[4], temp_strides=temp_strides_set[4]) else: raise Exception("Unsupported network settings.") blob_out = model.AveragePool(blob_in, 'pool5', kernels=[pool_stride, 7, 7], strides=[1, 1, 1], pads=[0, 0, 0] * 2) if cfg.TRAIN.DROPOUT_RATE > 0 and test_mode is False: blob_out = model.Dropout( blob_out, blob_out + '_dropout', ratio=cfg.TRAIN.DROPOUT_RATE, is_test=False) if split in ['train', 'val']: blob_out = model.FC( blob_out, 'pred', dim_in, cfg.MODEL.NUM_CLASSES, weight_init=('GaussianFill', {'std': cfg.MODEL.FC_INIT_STD}), bias_init=('ConstantFill', {'value': 0.}) ) elif split == 'test': blob_out = model.ConvNd( blob_out, 'pred', dim_in, cfg.MODEL.NUM_CLASSES, [1, 1, 1], strides=[1, 1, 1], pads=[0, 0, 0] * 2, ) if split == 'train': scale = 1. / cfg.NUM_GPUS softmax, loss = model.SoftmaxWithLoss( [blob_out, labels], ['softmax', 'loss'], scale=scale) elif split == 'val': #in ['test', 'val']: softmax = model.Softmax(blob_out, 'softmax', engine='CUDNN') loss = None elif split == 'test': # fully convolutional testing blob_out = model.Transpose(blob_out, 'pred_tr', axes=(0, 2, 3, 4, 1,)) blob_out, old_shape = model.Reshape( blob_out, ['pred_re', 'pred_shape5d'], shape=(-1, cfg.MODEL.NUM_CLASSES)) blob_out = model.Softmax(blob_out, 'softmax_conv', engine='CUDNN') blob_out = model.Reshape( [blob_out, 'pred_shape5d'], ['softmax_conv_re', 'pred_shape2d'])[0] blob_out = model.Transpose(blob_out, 'softmax_conv_tr', axes=(0, 4, 1, 2, 3)) blob_out = model.net.ReduceBackMean( [blob_out], ['softmax_ave_w']) blob_out = model.ReduceBackMean( [blob_out], ['softmax_ave_h']) softmax = model.ReduceBackMean( [blob_out], ['softmax']) loss = None return model, softmax, loss - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -