سلام- من از تصاویر گوش استفاده میکنم. حدودا 6500 تصویر برای ترین دارم و 700 تصویر برای تست. که در واقع با data augmentation به دست آوردم. تعداد کلاس هام هم 50 تا هستش. موقعی که میخوام از شبکه های مختلف استفاده کنم ضمن رعایت همه موارد فاین تیون، از وزن کفه مدل شبکه ها هم استفاده میکنم. به عنوان مثال برای استفاده از گوگل نت، وزن آن را هم استفاده میکنم و فاین تیون میکنم. اما loss همیشه زیاد هست و دقت خیلی کم. چیزی حدود 0.01. بچ سایز روی یک تنظیم شده چون که بیشتر از اون بذارم رم کم میارم. از مین فایل و نرمال سازی هم استفاده می کنم ولی هیچ بهبودی حاصل نمیشه. روی شبکه های الکس نت، lenet و گوگل نت امتحان کردم. هیچ کدوم دقت بالایی نداشتن.
این سالور من هست:
test_iter: 700
test_interval: 6500
base_lr: 0.01
display: 40
max_iter: 10000
lr_policy: "poly"
power: 0.5
momentum: 0.9
weight_decay: 0.0002
snapshot: 40000
snapshot_prefix: "examples/bvlc_googlenet/bvlc_googlenet_quick"
solver_mode: GPU
device_id: 0
net: "examples/bvlc_googlenet/train_val.prototxt"
train_state {
level: 0
stage: ""
}
نمیدونم دیگه باید چیکار کنم؟ ایا تصاویر من کم هستند و یا اینکه مشکل از وزن ها هستش. اینم قسمتی از خروجی روی شبکه گوگل نت:
I0504 20:34:49.219352 6648 net.cpp:761] Ignoring source layer loss1/ave_pool
I0504 20:34:49.219352 6648 net.cpp:761] Ignoring source layer loss1/conv
I0504 20:34:49.220332 6648 net.cpp:761] Ignoring source layer loss1/relu_conv
I0504 20:34:49.220332 6648 net.cpp:761] Ignoring source layer loss1/fc
I0504 20:34:49.221334 6648 net.cpp:761] Ignoring source layer loss1/relu_fc
I0504 20:34:49.221334 6648 net.cpp:761] Ignoring source layer loss1/drop_fc
I0504 20:34:49.222334 6648 net.cpp:761] Ignoring source layer loss1/classifier
I0504 20:34:49.222334 6648 net.cpp:761] Ignoring source layer loss1/loss
I0504 20:34:49.224339 6648 net.cpp:761] Ignoring source layer loss2/ave_pool
I0504 20:34:49.224339 6648 net.cpp:761] Ignoring source layer loss2/conv
I0504 20:34:49.225338 6648 net.cpp:761] Ignoring source layer loss2/relu_conv
I0504 20:34:49.225338 6648 net.cpp:761] Ignoring source layer loss2/fc
I0504 20:34:49.225338 6648 net.cpp:761] Ignoring source layer loss2/relu_fc
I0504 20:34:49.226459 6648 net.cpp:761] Ignoring source layer loss2/drop_fc
I0504 20:34:49.226459 6648 net.cpp:761] Ignoring source layer loss2/classifier
I0504 20:34:49.226459 6648 net.cpp:761] Ignoring source layer loss2/loss
I0504 20:34:49.229475 6648 net.cpp:761] Ignoring source layer loss3/classifier
I0504 20:34:49.238483 6648 caffe.cpp:252] Starting Optimization
I0504 20:34:49.238483 6648 solver.cpp:279] Solving GoogleNet
I0504 20:34:49.239472 6648 solver.cpp:280] Learning Rate Policy: poly
I0504 20:34:49.400656 6648 solver.cpp:228] Iteration 0, loss = 4.93185
I0504 20:34:49.400656 6648 solver.cpp:244] Train net output #0: loss3/loss3 = 4.93185 (* 1 = 4.93185 loss)
I0504 20:34:49.400656 6648 sgd_solver.cpp:106] Iteration 0, lr = 0.01
I0504 20:34:55.188277 6648 solver.cpp:228] Iteration 40, loss = 87.3365
I0504 20:34:55.189512 6648 solver.cpp:244] Train net output #0: loss3/loss3 = 87.3365 (* 1 = 87.3365 loss)
I0504 20:34:55.190076 6648 sgd_solver.cpp:106] Iteration 40, lr = 0.00997998
I0504 20:35:00.982188 6648 solver.cpp:228] Iteration 80, loss = 87.3365
I0504 20:35:00.982322 6648 solver.cpp:244] Train net output #0: loss3/loss3 = 87.3365 (* 1 = 87.3365 loss)
I0504 20:35:00.983325 6648 sgd_solver.cpp:106] Iteration 80, lr = 0.00995992
I0504 20:35:06.763850 6648 solver.cpp:228] Iteration 120, loss = 87.3365
I0504 20:35:06.764673 6648 solver.cpp:244] Train net output #0: loss3/loss3 = 87.3365 (* 1 = 87.3365 loss)
I0504 20:35:06.765547 6648 sgd_solver.cpp:106] Iteration 120, lr = 0.00993982
I0504 20:35:12.561303 6648 solver.cpp:228] Iteration 160, loss = 87.3365
I0504 20:35:18.875185 6648 solver.cpp:244] Train net output #0: loss3/loss3 = 87.3365 (* 1 = 87.3365 loss)
I0504 20:35:18.876188 6648 sgd_solver.cpp:106] Iteration 160, lr = 0.00991968
I0504 20:35:24.607239 6648 solver.cpp:228] Iteration 200, loss = 87.3365
I0504 20:35:26.163414 6648 solver.cpp:244] Train net output #0: loss3/loss3 = 87.3365 (* 1 = 87.3365 loss)
I0504 20:35:26.164357 6648 sgd_solver.cpp:106] Iteration 200, lr = 0.0098995
I0504 20:35:31.888242 6648 solver.cpp:228] Iteration 240, loss = 87.3365
I0504 20:35:31.888242 6648 solver.cpp:244] Train net output #0: loss3/loss3 = 87.3365 (* 1 = 87.3365 loss)
I0504 20:35:31.888242 6648 sgd_solver.cpp:106] Iteration 240, lr = 0.00987927
I0504 20:35:37.685355 6648 solver.cpp:228] Iteration 280, loss = 87.3365
I0504 20:35:37.686475 6648 solver.cpp:244] Train net output #0: loss3/loss3 = 87.3365 (* 1 = 87.3365 loss)
I0504 20:35:37.686882 6648 sgd_solver.cpp:106] Iteration 280, lr = 0.00985901
I0504 20:35:43.489472 6648 solver.cpp:228] Iteration 320, loss = 87.3365
I0504 20:35:43.489472 6648 solver.cpp:244] Train net output #0: loss3/loss3 = 87.3365 (* 1 = 87.3365 loss)
I0504 20:35:43.490864 6648 sgd_solver.cpp:106] Iteration 320, lr = 0.0098387
I0504 20:35:49.283823 6648 solver.cpp:228] Iteration 360, loss = 87.3365
I0504 20:35:49.284389 6648 solver.cpp:244] Train net output #0: loss3/loss3 = 87.3365 (* 1 = 87.3365 loss)
I0504 20:35:49.284389 6648 sgd_solver.cpp:106] Iteration 360, lr = 0.00981835
I0504 20:35:55.083504 6648 solver.cpp:228] Iteration 400, loss = 87.3365
I0504 20:35:55.084228 6648 solver.cpp:244] Train net output #0: loss3/loss3 = 87.3365 (* 1 = 87.3365 loss)
I0504 20:35:55.084854 6648 sgd_solver.cpp:106] Iteration 400, lr = 0.00979796
I0504 20:36:00.871918 6648 solver.cpp:228] Iteration 440, loss = 87.3365
I0504 20:36:00.871918 6648 solver.cpp:244] Train net output #0: loss3/loss3 = 87.3365 (* 1 = 87.3365 loss)
I0504 20:36:00.872941 6648 sgd_solver.cpp:106] Iteration 440, lr = 0.00977752
I0504 20:36:06.654637 6648 solver.cpp:228] Iteration 480, loss = 87.3365
I0504 20:36:06.655269 6648 solver.cpp:244] Train net output #0: loss3/loss3 = 87.3365 (* 1 = 87.3365 loss)