Skip to content

Latest commit

 

History

History
275 lines (238 loc) · 7.7 KB

FEATURES.md

File metadata and controls

275 lines (238 loc) · 7.7 KB

A Short Summary of New Features in Synopsys Caffe

Synopsys Caffe Version: 2023.09
New added features are compared with the original BVLC Caffe 1.0.0

evconvert (TensorFlow/ONNX/... to Caffe Converter) related

atan_layer
attention_layer
batch_to_space_nd_layer
broadcast_to_layer
ceil_layer
conv_depthwise_layer
count_nonzero_layer
crop_and_resize_layer
depth_to_space_layer
div_layer
embedding_lookup_layer
embedding_lookup_sparse_layer
expand_dims_nd_layer
farthest_point_sample_layer
floor_div_layer
floor_layer
floor_mod_layer
gather_layer
gather_v2_layer
gather_nd_layer
gemm_layer
group_point_layer
gru_layer
hard_sigmoid_layer
hard_swish_layer
hard_tanh_layer
icnet_subgraph_layer
layer_norm_layer
log_softmax_layer
lp_normalization_layer
luong_attention_layer
nms_gather_layer
matmul_layer
maximum_layer
minimum_layer
mirror_pad_layer
mish_layer
mul_layer
nms_layer
non_max_suppression_layer
not_equal_layer
one_hot_layer
pad_layer
peephole_lstm_layer
piece_layer
pooling3d_layer
pow_layer
query_ball_point_layer
range_layer
reduce_all_layer
reduce_any_layer
reduce_l1_layer
reduce_l2_layer
reduce_logsumexp_layer
reduce_max_layer
reduce_mean_layer
reduce_min_layer
reduce_prod_layer
reduce_sum_layer
resize_bilinear_layer
resize_nearest_neighbor_layer
reverse_layer
reverse_sequence_layer
rnn_v2_layer
round_layer
scaled_tanh_layer
scatter_nd_layer
shape_layer
shuffle_channel_layer
simple_rnn_layer
sin_layer
softplus_layer
softsign_layer
spatial_batching_pooling_layer
space_to_batch_nd_layer
space_to_depth_layer
sparse_to_dense_layer
squeeze_layer
stack_layer
strided_slice_layer
sub_layer
tensor2box_layer
three_interpolate_layer
three_NN_layer
thresholded_relu_layer
tile_nd_layer
topk_gather_layer
unstack_layer
where4_gathernd_crop_layer
where4_gathernd_layer
where4_layer

DIV and MIN in EltwiseOp
axis in EltwiseParameter (broadcasting support for 2nd bottom blob in eltwise_layer)
min_first in ArgMaxParameter
submanifold_sparse in ConvolutionParameter
pad_type (deprecated, "SAME" style padding) in ConvolutionParameter and PoolingParameter
pad_l, pad_r, pad_t and pad_b (arbitrary 2D padding) in ConvolutionParameter and PoolingParameter
AVE_EXC_PAD (average pooling excluding the paddings), AVE_TF (deprecated, alias for AVE_EXC_PAD) in PoolingParameter
ceil_mode in PoolingParameter
faceboxes, box_width, box_height, keras, tf and yx_order in PriorBoxParameter
relu6, maximum and minimum in ReLUParameter

deform_conv2d.py (customized Python layer)
eltwise.py (deprecated, customized Python layer, realize the broadcasting and add support for divide and minimum for eltwise layer)
matrix_inverse.py (customized Python layer, implementation of tf.matrix_inverse)
pad.py and pads.py (deprecated, customized Python layer, implementation of tf.pad)
range.py (deprecated, customized Python layer, implementation of tf.range)
rank.py (customized Python layer, implementation of tf.rank)
reshape.py (customized Python layer, implementation of tf.reshape with two inputs)
shape.py (deprecated, customized Python layer, implementation of tf.shape)
slice.py (deprecated, customized Python layer, implementation of tf.slice and tf.strided_slice)
stack.py (deprecated, customized Python layer, implementation of tf.stack)
statistics.py (deprecated, customized Python layer, implementation of tf.reduce_mean, tf.reduce_prod, tf.reduce_sum, tf.reduce_max, tf.reduce_min)
stridedslice.py (deprecated, customized Python layer)

evprune (Network Pruning Tool) related

squeeze_conv_layer
squeeze_inner_product_layer
squeeze_deconv_layer

Customized Quantization related

  • input_scale, input_zero_point, output_scale, output_zero_point, weight_scale, weight_zero_point, per_channel_scale_weight, per_channel_scale_output, saturate in
    ConvolutionParameter
  • input_scale, input_zero_point, output_scale, output_zero_point, weight_scale, weight_zero_point, saturate in
    InnerProductParameter
  • input_scale, input_zero_point, output_scale, output_zero_point, bias_scale, bias_zero_point, saturate in
    BiasParameter
  • input_scale, input_zero_point, output_scale, output_zero_point, saturate in
    ConcatParameter
    EltwiseParameter
    PoolingParameter
    ReLUParameter
    ResizeBilinearParameter
    SigmoidParameter
    SoftmaxParameter
    SpatialBatchingPoolingParameter
  • output_scale, output_zero_point in
    InputParameter
  • saturate in
    PowerParameter

evquantize related (only valid for CUDA version forwards implementation)

input_scale and output_scale in ConvolutionParameter, SoftmaxParameter and LRNParameter
output_scale in EltwiseParameter and InnerProductParameter
output_shift_instead_division in PoolingParameter
saturate in ConvolutionParameter, EltwiseParameter, ReLUParameter and PoolingParameter

Mask RCNN related

maskrcnn_detection_layer
maskrcnn_proposal_layer
pyramid_roi_align_layer
roi_align_layer

apply_box_deltas.py (customized Python layer)
generate_pyramid_anchors.py (customized Python layer)
maskrcnn_detection.py (customized Python layer)
maskrcnn_proposal.py (customized Python layer)
pre_roi_align.py (customized Python layer)

SNNs related

selu_dropout_layer
selu_layer

YOLO related

reorg_layer
upsample_darknet_layer
yolo_v2_loss_layer
yolo_v3_loss_layer

add_eps_before_sqrt in BatchNormParameter, MVNParameter and NormalizeParameter
caffe_yolo in TransformationParameter
jitter in ResizeParameter
exposure_lower, exposure_upper in DistortionParameter
side and random in DataParameter

yolov2.py (customized Python layer, implementation of darknet_reorg)
darknet.py

ICNet (PSPNet) related

adaptive_bias_channel_layer
bias_channel_layer
cudnn_bn_layer
densecrf_layer
domain_transform_forward_only_layer
domain_transform_layer
image_seg_data_layer
interp_layer
mat_read_layer
mat_write_layer
seg_accuracy_layer
spatial_product_layer
unique_label_layer
bn_layer (slope_filler, bias_filler, momentum and icnet in BNParameter)

update_global_stats and icnet in BatchNormParameter
scale_factors, crop_width and crop_height in TransformationParameter

SRGAN related

gan_loss_layer

gan_solver in SolverParameter
pixelshuffler in ReshapeParameter
dis_mode and gen_mode in ConvolutionParameter, InnerProductParameter and ScaleParameter
weight_fixed in ConvolutionParameter and InnerProductParameter

FlowNet2 related

accum_layer
augmentation_layer
black_augmentation_layer
channel_norm_layer
correlation_1d_layer
correlation_layer
custom_data_layer
data_augmentation_layer
downsample_layer
flo_writer_layer
float_reader_layer
float_writer_layer
flow_augmentation_layer
flow_warp_layer
generate_augmentation_parameters_layer
l1_loss_layer
lpq_loss_layer
mean_layer
resample_layer

SSD related

annotated_data_layer
detection_evaluate_layer
detection_output_layer
multibox_loss_layer
normalize_layer
permute_layer
prior_box_layer
smooth_L1_loss_layer
ssd_decoder_layer
ssd_sort_layer
video_data_layer

Faster RCNN related

proposal_layer
roi_pooling_layer
smooth_L1_loss_layer (sigma and abssum in SmoothL1LossParameter)

scale_train in DropoutParameter

SegNet related

bn_layer
dense_image_data_layer
upsample_layer

sample_weights_test in DropoutParameter
weight_by_label_freqs in LossParameter

For more details, please refer to caffe.proto and the corresponding source code.