It's an old version of dnn fo golang. Please see new updated version https://github.com/go-ml-dev/nn
import (
"github.com/sudachen/go-dnn/data/mnist"
"github.com/sudachen/go-dnn/fu"
"github.com/sudachen/go-dnn/mx"
"github.com/sudachen/go-dnn/ng"
"github.com/sudachen/go-dnn/nn"
"gotest.tools/assert"
"testing"
"time"
)
var mnistConv0 = nn.Connect(
&nn.Convolution{Channels: 24, Kernel: mx.Dim(3, 3), Activation: nn.ReLU},
&nn.MaxPool{Kernel: mx.Dim(2, 2), Stride: mx.Dim(2, 2)},
&nn.Convolution{Channels: 32, Kernel: mx.Dim(5, 5), Activation: nn.ReLU, BatchNorm: true},
&nn.MaxPool{Kernel: mx.Dim(2, 2), Stride: mx.Dim(2, 2)},
&nn.FullyConnected{Size: 32, Activation: nn.Swish, BatchNorm: true},
&nn.FullyConnected{Size: 10, Activation: nn.Softmax})
func Test_mnistConv0(t *testing.T) {
gym := &ng.Gym{
Optimizer: &nn.Adam{Lr: .001},
Loss: &nn.LabelCrossEntropyLoss{},
Input: mx.Dim(32, 1, 28, 28),
Epochs: 5,
Verbose: ng.Printing,
Every: 1 * time.Second,
Dataset: &mnist.Dataset{},
Metric: &ng.Classification{Accuracy: 0.98},
Seed: 42,
}
acc, params, err := gym.Train(mx.CPU, mnistConv0)
assert.NilError(t, err)
assert.Assert(t, acc >= 0.98)
err = params.Save(fu.CacheFile("tests/mnistConv0.params"))
assert.NilError(t, err)
net, err := nn.Bind(mx.CPU, mnistConv0, mx.Dim(10, 1, 28, 28), nil)
assert.NilError(t, err)
err = net.LoadParamsFile(fu.CacheFile("tests/mnistConv0.params"), false)
assert.NilError(t, err)
_ = net.PrintSummary(false)
ok, err := ng.Measure(net, &mnist.Dataset{}, &ng.Classification{Accuracy: 0.98}, ng.Printing)
assert.Assert(t, ok)
}Network Identity: 158cf5bd604e12e7bd438084e135703bd89dc10f
Symbol | Operation | Output | Params #
----------------------------------------------------------------------
_input | null | (32,1,28,28) | 0
Convolution01 | Convolution((3,3)//) | (32,24,26,26) | 240
Convolution01$A | Activation(relu) | (32,24,26,26) | 0
MaxPool02 | Pooling(max) | (32,24,13,13) | 0
Convolution03 | Convolution((5,5)//) | (32,32,9,9) | 19232
Convolution03$BN | BatchNorm | (32,32,9,9) | 128
Convolution03$A | Activation(relu) | (32,32,9,9) | 0
MaxPool04 | Pooling(max) | (32,32,4,4) | 0
FullyConnected05 | FullyConnected | (32,32) | 16416
FullyConnected05$BN | BatchNorm | (32,32) | 128
sigmoid@sym07 | sigmoid | (32,32) | 0
FullyConnected05$A | elemwise_mul | (32,32) | 0
FullyConnected06 | FullyConnected | (32,10) | 330
FullyConnected06$A | SoftmaxActivation() | (32,10) | 0
BlockGrad@sym08 | BlockGrad | (32,10) | 0
make_loss@sym09 | make_loss | (32,10) | 0
pick@sym10 | pick | (32,1) | 0
log@sym11 | log | (32,1) | 0
_mul_scalar@sym12 | _mul_scalar | (32,1) | 0
mean@sym13 | mean | (1) | 0
make_loss@sym14 | make_loss | (1) | 0
----------------------------------------------------------------------
Total params: 36474
[000] batch: 389, loss: 0.09991227
[000] batch: 1074, loss: 0.055281825
[000] batch: 1855, loss: 0.0760978
[000] metric: 0.988, final loss: 0.0515
Achieved reqired metric
Symbol | Operation | Output | Params #
----------------------------------------------------------------------
_input | null | (10,1,28,28) | 0
Convolution01 | Convolution((3,3)//) | (10,24,26,26) | 240
Convolution01$A | Activation(relu) | (10,24,26,26) | 0
MaxPool02 | Pooling(max) | (10,24,13,13) | 0
Convolution03 | Convolution((5,5)//) | (10,32,9,9) | 19232
Convolution03$BN | BatchNorm | (10,32,9,9) | 128
Convolution03$A | Activation(relu) | (10,32,9,9) | 0
MaxPool04 | Pooling(max) | (10,32,4,4) | 0
FullyConnected05 | FullyConnected | (10,32) | 16416
FullyConnected05$BN | BatchNorm | (10,32) | 128
sigmoid@sym07 | sigmoid | (10,32) | 0
FullyConnected05$A | elemwise_mul | (10,32) | 0
FullyConnected06 | FullyConnected | (10,10) | 330
FullyConnected06$A | SoftmaxActivation() | (10,10) | 0
----------------------------------------------------------------------
Total params: 36474
Accuracy over 1000*10 batchs: 0.988
--- PASS: Test_mnistConv0 (6.51s)