diff --git a/benchmark/alexnet/train.net b/benchmark/alexnet/train.net index 87c1d2da9b8899ffa2264e244288de0cda3b09bc..98749ac8e5ae46f0da2c72653b54108d64e36eba 100644 --- a/benchmark/alexnet/train.net +++ b/benchmark/alexnet/train.net @@ -6,12 +6,18 @@ op { name: "encoded" data_type: kFloat shape { - dim: 224 - dim: 224 + dim: 227 + dim: 227 dim: 3 } encode_case { jpeg { + preprocess { + resize { + width: 227 + height: 227 + } + } } } } @@ -45,7 +51,7 @@ op { in: "transpose/out" out: "out" filters: 64 - padding: "SAME" + padding: "VALID" data_format: "channels_first" kernel_size: 11 kernel_size: 11 @@ -63,23 +69,23 @@ op { } } -op { - name: "lrn1" - local_response_normalization_conf { - in: "conv1/out" - out: "out" - depth_radius: 2 - bias: 2 - alpha: 0.0001 - beta: 0.75 - data_format: "channels_first" - } -} +#op { +# name: "lrn1" +# local_response_normalization_conf { +# in: "conv1/out" +# out: "out" +# depth_radius: 2 +# bias: 2 +# alpha: 0.0001 +# beta: 0.75 +# data_format: "channels_first" +# } +#} op { name: "pool1" max_pooling_2d_conf { - in: "lrn1/out" + in: "conv1/out" out: "out" padding: "VALID" data_format: "channels_first" @@ -114,23 +120,23 @@ op { } } -op { - name: "lrn2" - local_response_normalization_conf { - in: "conv2/out" - out: "out" - depth_radius: 2 - bias: 2 - alpha: 0.0001 - beta: 0.75 - data_format: "channels_first" - } -} +#op { +# name: "lrn2" +# local_response_normalization_conf { +# in: "conv2/out" +# out: "out" +# depth_radius: 2 +# bias: 2 +# alpha: 0.0001 +# beta: 0.75 +# data_format: "channels_first" +# } +#} op { name: "pool2" max_pooling_2d_conf { - in: "lrn2/out" + in: "conv2/out" out: "out" padding: "VALID" data_format: "channels_first" diff --git a/benchmark/alexnet/train.other b/benchmark/alexnet/train.other index d8b0639b2347d50e6db379d3743c4d820dbc37ac..d1454639aca0b1d35e7a3cc3b29089b53cd0e9c3 100644 --- a/benchmark/alexnet/train.other +++ b/benchmark/alexnet/train.other @@ -4,7 +4,7 @@ globalfs_conf { } } data_part_num: 30 -piece_size: 240 +piece_size: 480 model_load_snapshot_path: "" max_data_id_length: 0 train_conf { @@ -18,12 +18,11 @@ train_conf { std: 0.01 } } - batch_size: 240 + batch_size: 480 l2: 0.0005 model_update_conf { learning_rate: 0.01 - momentum_conf { - beta: 0.9 + naive_conf { } learning_rate_decay { exponential_conf { diff --git a/benchmark/alexnet/train.placement b/benchmark/alexnet/train.placement index 903bdf3c7a77f93b92caa51baeecfa0fb4a3818c..980726d53532fcdc15d499df790bab1d9cd1cb1a 100644 --- a/benchmark/alexnet/train.placement +++ b/benchmark/alexnet/train.placement @@ -1,6 +1,7 @@ placement_group { op_set { op_name: "decode" + op_name: "transpose" } parallel_conf { policy: kDataParallel @@ -10,12 +11,11 @@ placement_group { placement_group { op_set { - op_name: "transpose" op_name: "conv1" - op_name: "lrn1" +# op_name: "lrn1" op_name: "pool1" op_name: "conv2" - op_name: "lrn2" +# op_name: "lrn2" op_name: "pool2" op_name: "conv3" op_name: "conv4"