summary | shortlog | log | commit | commitdiff | tree
raw | patch | inline | side by side (parent: b6d40b8)
raw | patch | inline | side by side (parent: b6d40b8)
author | borisgin <boris.ginsburg@gmail.com> | |
Mon, 25 Sep 2017 22:12:36 +0000 (15:12 -0700) | ||
committer | borisgin <boris.ginsburg@gmail.com> | |
Mon, 25 Sep 2017 22:12:36 +0000 (15:12 -0700) |
models/bvlc_alexnet/solver_augment.prototxt | [new file with mode: 0644] | patch | blob |
models/bvlc_alexnet/train_alexnet_augment.sh | [new file with mode: 0755] | patch | blob |
models/bvlc_alexnet/train_val_augment.prototxt | [moved from models/bvlc_alexnet/train_val_transform.prototxt with 92% similarity] | patch | blob | history |
models/resnet50/train_val_augment.prototxt | patch | blob | history |
diff --git a/models/bvlc_alexnet/solver_augment.prototxt b/models/bvlc_alexnet/solver_augment.prototxt
--- /dev/null
@@ -0,0 +1,43 @@
+net: "models/bvlc_alexnet/train_val_augment.prototxt"
+
+test_iter: 196 #1562 = 50000/32
+test_interval: 1250
+test_initialization: false
+
+display: 250
+
+max_iter: 125000
+
+lr_policy: "poly"
+power: 2
+
+base_lr: 2 # B=1024
+
+#rampup_lr: 0.1
+#rampup_interval: 2500
+
+local_lr_auto: true
+local_gw_ratio: 0.001
+
+momentum: 0.9
+weight_decay: 0.0005
+
+snapshot: 500000
+snapshot_prefix: "models/bvlc_alexnet/snapshots/alexnet_fp32"
+snapshot_after_train: false
+
+solver_mode: GPU
+random_seed: 1
+
+# Train dataset size = 1,281,167
+# Test dataset size = 50,000
+
+# batch 64 --> epoch = 20,000
+# batch 96 --> epoch = 15,000
+# batch 128 --> epoch = 10,000
+# batch 256 --> epoch = 5,000
+# batch 512 --> epoch = 2,500
+# batch 1024--> epoch = 1,250
+# batch 2048--> epoch = 625
+# batch 4096--> epoch = 312
+# batch 8192--> epoch = 156
diff --git a/models/bvlc_alexnet/train_alexnet_augment.sh b/models/bvlc_alexnet/train_alexnet_augment.sh
--- /dev/null
@@ -0,0 +1,5 @@
+#!/usr/bin/env sh
+
+./build/tools/caffe train \
+ --solver=models/bvlc_alexnet/solver_augment.prototxt -gpu=all \
+ 2>&1 | tee models/bvlc_alexnet/logs/alexnet_augment_B1024_lr2.0.log
diff --git a/models/bvlc_alexnet/train_val_transform.prototxt b/models/bvlc_alexnet/train_val_augment.prototxt
similarity index 92%
rename from models/bvlc_alexnet/train_val_transform.prototxt
rename to models/bvlc_alexnet/train_val_augment.prototxt
index 8e724a5e1a676387e59141fa790165d9f4ebff44..d48fd5412d18a3da5a37f6d0276c37b8fcfbab8a 100644 (file)
rename from models/bvlc_alexnet/train_val_transform.prototxt
rename to models/bvlc_alexnet/train_val_augment.prototxt
index 8e724a5e1a676387e59141fa790165d9f4ebff44..d48fd5412d18a3da5a37f6d0276c37b8fcfbab8a 100644 (file)
top: "label"
transform_param {
# mean_file: "data/ilsvrc12/imagenet_mean.binaryproto"
+ mean_file: "data/imagenet/imagenet_mean.binaryproto"
+ img_rand_resize_lower: 240
+ img_rand_resize_upper: 272
crop_size: 227
- mean_value: 123.68
- mean_value: 116.78
- mean_value: 103.94
- img_rand_resize_lower: 250
- img_rand_resize_upper: 260
- var_sz_img_enabled: true
mirror: true
}
data_param {
source: "examples/imagenet/ilsvrc12_train_lmdb"
- batch_size: 256
-# batch_size: 1024 # DGX1
+ batch_size: 1024 # DGX1
backend: LMDB
- cache: true
+ cache: true
shuffle: true
}
include { phase: TRAIN }
top: "label"
transform_param {
# mean_file: "data/ilsvrc12/imagenet_mean.binaryproto"
- crop_size: 227
- mean_value: 123.68
- mean_value: 116.78
- mean_value: 103.94
+ mean_file: "data/imagenet/imagenet_mean.binaryproto"
img_rand_resize_lower: 256
img_rand_resize_upper: 256
- var_sz_img_enabled: true
+ crop_size: 227
mirror: false
}
data_param {
source: "examples/imagenet/ilsvrc12_val_lmdb"
# batch_size: 32
- batch_size: 256 # DGX1
+ batch_size: 256
backend: LMDB
}
include { phase: TEST }
diff --git a/models/resnet50/train_val_augment.prototxt b/models/resnet50/train_val_augment.prototxt
index 7e082006555615422ef8daeffe3c9e8c48dac981..7032c0c177c379a0f1b0784810a448d6eeb535c0 100644 (file)
shuffle: true
}
transform_param {
- var_sz_img_enabled: true
- img_rand_resize_lower: 256
- img_rand_resize_upper: 288
- mean_value: 104
- mean_value: 117
- mean_value: 123
+ img_rand_resize_lower: 224
+ img_rand_resize_upper: 272
+ mean_file: "data/imagenet/imagenet_mean.binaryproto"
crop_size: 224
mirror: true
scale: 0.00390625
batch_size: 32
}
transform_param {
- mean_value: 104
- mean_value: 117
- mean_value: 123
+ mean_file: "data/imagenet/imagenet_mean.binaryproto"
+ img_rand_resize_lower: 240
+ img_rand_resize_upper: 240
crop_size: 224
mirror: false
scale: 0.00390625