cntk 0.0.3

Wrapper around Microsoft CNTK library
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294

#[macro_use]
extern crate cpp;

mod shape;
pub use shape::Shape;

mod variable_set;
pub use variable_set::VariableSet;

mod variable;
pub use variable::{Variable, ParameterInitializer};

pub mod ops;

mod function;
pub use function::Function;
pub use function::BackPropState;

mod value;
pub use value::Value;

mod device;
pub use device::{DeviceDescriptor, set_max_num_cpu_threads};

mod data_map;
pub use data_map::DataMap;

mod replacement_map;
pub use replacement_map::ReplacementMap;

mod learner;
pub use learner::{Learner, DoubleParameterSchedule};

mod trainer;
pub use trainer::Trainer;

#[cfg(test)]
mod tests {
    use super::*;

    use ops::*;
    #[test]
    fn simple_add() {
        let var = Variable::input_variable(Shape::from_slice(&vec!(5)));
        let var2 = Variable::input_variable(Shape::from_slice(&vec!(5)));
        let plus = plus(&var, &var2);

        let data: Vec<f32> = vec!(1.0, 2.0, 3.0, 4.0, 5.0, 6.0, 7.0, 8.0, 9.0, 10.0);
        let data2: Vec<f32> = vec!(11.0, 12.0, 13.0, 14.0, 15.0, 16.0, 17.0, 18.0, 19.0, 110.0);

        let val = Value::batch(&var.shape(), &data, DeviceDescriptor::cpu());
        let val2 = Value::batch(&var2.shape(), &data2, DeviceDescriptor::cpu());

        let mut datamap = DataMap::new();
        datamap.add(&var, &val);
        datamap.add(&var2, &val2);

        let emptyval = Value::empty();
        let mut outdatamap = DataMap::new();
        outdatamap.add(&plus, &emptyval);

        Function::from_variable(&plus).evaluate(&datamap, &mut outdatamap, DeviceDescriptor::cpu());

        let result = outdatamap.get(&plus).unwrap().to_vec();
        assert_eq!(result, vec!(12., 14., 16., 18., 20., 22., 24., 26., 28., 120.));
    }

    #[test]
    fn gradient() {
        let var = Variable::input_variable_with_gradient(Shape::scalar());
        let var2 = Variable::input_variable_with_gradient(Shape::scalar());
        let var3 = Variable::input_variable_with_gradient(Shape::scalar());
        let out = plus(&elem_times(&var, &var2), &var3);

        let data: Vec<f32> = vec!(4.0, 7.0);
        let data2: Vec<f32> = vec!(11.0, 12.0);
        let data3: Vec<f32> = vec!(11.0, 12.0);

        let val = Value::batch(&var.shape(), &data, DeviceDescriptor::cpu());
        let val2 = Value::batch(&var2.shape(), &data2, DeviceDescriptor::cpu());
        let val3 = Value::batch(&var3.shape(), &data3, DeviceDescriptor::cpu());

        let mut datamap = DataMap::new();
        datamap.add(&var, &val);
        datamap.add(&var2, &val2);
        datamap.add(&var3, &val3);

        let emptyval = Value::empty();
        let mut outdatamap = DataMap::new();
        outdatamap.add(&out, &emptyval);

        let mut retain_state = VariableSet::new();
        retain_state.add(&out);

        let of = Function::from_variable(&out);

        let bpstate = of.forward(&datamap, &mut outdatamap, DeviceDescriptor::cpu(), &retain_state, &VariableSet::new());
        let out_val = outdatamap.get(&out).unwrap();

        let mut result = DataMap::new();
        let emptyval2 = Value::empty();
        let emptyval3 = Value::empty();
        result.add(&var2, &emptyval2);
        result.add(&var3, &emptyval3);

        let mut rgvalues = DataMap::new();
        let rootgrad = Value::from_vec(&out_val.shape(), &(vec![1.; out_val.shape().total_size()]), DeviceDescriptor::cpu());
        rgvalues.add(&out, &rootgrad);

        of.backward(&bpstate, &rgvalues, &mut result);

        assert_eq!(result.get(&var2).unwrap().to_vec(), vec!(4., 7.));
        assert_eq!(result.get(&var3).unwrap().to_vec(), vec!(1., 1.));
    }

    fn rng_next(seed: &mut i32) -> f32 {
        let ret = (((*seed % 201)+201)%201 - 100) as f32 / 100.0;
        *seed = seed.wrapping_mul(23);
        ret
    }

    #[test]
    fn feedforward_net_training() {
        set_max_num_cpu_threads(1);
        let x = Variable::input_variable_with_name(Shape::from_slice(&vec!(3)), "X");
        let y = Variable::input_variable_with_name(Shape::from_slice(&vec!(1)), "Y");
        let w1 = Variable::parameter(Shape::from_slice(&vec!(20, 3)), ParameterInitializer::glorot_uniform(), DeviceDescriptor::cpu());
        let b1 = Variable::parameter(Shape::from_slice(&vec!(20)), ParameterInitializer::glorot_uniform(), DeviceDescriptor::cpu());
        let w2 = Variable::parameter(Shape::from_slice(&vec!(1, 20)), ParameterInitializer::glorot_uniform(), DeviceDescriptor::cpu());
        let b2 = Variable::parameter(Shape::from_slice(&vec!(1)), ParameterInitializer::glorot_uniform(), DeviceDescriptor::cpu());

        let hidden_value = tanh(&plus(&times(&w1, &x), &b1));
        let output_value = named_alias(&plus(&times(&w2, &hidden_value), &b2), "output");
        let error = reduce_sum_all(&squared_error(&output_value, &y));

        let output_func = Function::from_variable(&output_value);
        let error_func = Function::from_variable(&error);

        let mut rng_seed = 47;

        let learner = Learner::sgd(&vec!(&w1, &b1, &w2, &b2), &DoubleParameterSchedule::constant(0.01));
        let trainer = Trainer::new(&output_func, &error_func, &learner);
        let mut lastloss = 1000000.0;

        for iter in 0..5000 {
            let data = vec!(rng_next(&mut rng_seed), rng_next(&mut rng_seed), rng_next(&mut rng_seed),
                            rng_next(&mut rng_seed), rng_next(&mut rng_seed), rng_next(&mut rng_seed));
            let odata = vec!(data[0]*data[1] + data[2],
                             data[3]*data[4] + data[5]);
            let value = Value::batch(&x.shape(), &data, DeviceDescriptor::cpu());
            let ovalue = Value::batch(&y.shape(), &odata, DeviceDescriptor::cpu());
            let mut datamap = DataMap::new();
            datamap.add(&x, &value);
            datamap.add(&y, &ovalue);
            let mut outdatamap = DataMap::new();
            outdatamap.add_null(&output_value);
            outdatamap.add_null(&error);

            trainer.train_minibatch(&datamap, &mut outdatamap, DeviceDescriptor::cpu());
            let result = outdatamap.get(&output_value).unwrap().to_vec();
            let loss = outdatamap.get(&error).unwrap().to_vec();
            lastloss = loss[0];
        }
        assert!(lastloss < 0.1);

        output_func.save("test.dat");

        {
            let func_loaded = Function::load("test.dat", DeviceDescriptor::cpu());
            let inputs = func_loaded.inputs();

            let outputs = func_loaded.outputs();

            let loaded_input = inputs.iter().find(|x| x.name() == "X").unwrap();
            let loaded_output = outputs.iter().find(|x| x.name() == "output").unwrap();

            let data = vec!(rng_next(&mut rng_seed), rng_next(&mut rng_seed), rng_next(&mut rng_seed),
                           rng_next(&mut rng_seed), rng_next(&mut rng_seed), rng_next(&mut rng_seed));
            let odata = vec!(data[0]*data[1] + data[2],
                            data[3]*data[4] + data[5]);
            let value = Value::batch(&loaded_input.shape(), &data, DeviceDescriptor::cpu());
            let mut datamap = DataMap::new();
            datamap.add(&loaded_input, &value);
            let mut outdatamap = DataMap::new();
            outdatamap.add_null(&loaded_output);

            func_loaded.evaluate(&datamap, &mut outdatamap, DeviceDescriptor::cpu());

            let result = outdatamap.get(&loaded_output).unwrap().to_vec();
            assert!((result[0] - odata[0]).abs() < 0.05);
            assert!((result[1] - odata[1]).abs() < 0.05);
        }
    }

    #[test]
    fn classification_net_training() {
        set_max_num_cpu_threads(1);
        let x = Variable::input_variable(Shape::from_slice(&vec!(2)));
        let y = Variable::input_variable(Shape::from_slice(&vec!(3)));
        let w1 = Variable::parameter(Shape::from_slice(&vec!(20, 2)), ParameterInitializer::glorot_uniform(), DeviceDescriptor::cpu());
        let b1 = Variable::parameter(Shape::from_slice(&vec!(20)), ParameterInitializer::glorot_uniform(), DeviceDescriptor::cpu());
        let w2 = Variable::parameter(Shape::from_slice(&vec!(3, 20)), ParameterInitializer::glorot_uniform(), DeviceDescriptor::cpu());
        let b2 = Variable::parameter(Shape::from_slice(&vec!(3)), ParameterInitializer::glorot_uniform(), DeviceDescriptor::cpu());

        let hidden_value = tanh(&plus(&times(&w1, &x), &b1));
        let output_value = plus(&times(&w2, &hidden_value), &b2);
        let loss = reduce_sum_all(&cross_entropy_with_softmax(&output_value, &y));
        let wrong_labels = reduce_sum_all(&classification_error(&output_value, &y));


        let output_func = Function::from_variable(&output_value);
        let loss_func = Function::from_variable(&loss);
        let wrong_labels_func = Function::from_variable(&wrong_labels);

        let mut rng_seed = 47;

        let learner = Learner::sgd(&vec!(&w1, &b1, &w2, &b2), &DoubleParameterSchedule::constant(0.01));
        let trainer = Trainer::new_with_evalatuion(&output_func, &loss_func, &wrong_labels_func, &learner);
        let mut lastloss = 1000000.0;

        for iter in 0..50000 {
            let data = vec!(rng_next(&mut rng_seed), rng_next(&mut rng_seed), rng_next(&mut rng_seed),
                           rng_next(&mut rng_seed));
            let r1 = data[0]*data[0] + data[1]*data[1];
            let r2 = data[2]*data[2] + data[3]*data[3];

            let odata = vec!(if (r1 < 0.3) {1.0} else {0.0},
                             if (r1 >= 0.3 && r1 < 0.6) {1.0} else {0.0},
                             if (r1 >= 0.6) {1.0} else {0.0},
                             if (r2 < 0.3) {1.0} else {0.0},
                             if (r2 >= 0.3 && r2 < 0.6) {1.0} else {0.0},
                             if (r2 >= 0.6) {1.0} else {0.0}
                            );

            let value = Value::batch(&x.shape(), &data, DeviceDescriptor::cpu());
            let ovalue = Value::batch(&y.shape(), &odata, DeviceDescriptor::cpu());
            let mut datamap = DataMap::new();
            datamap.add(&x, &value);
            datamap.add(&y, &ovalue);
            let mut outdatamap = DataMap::new();
            outdatamap.add_null(&output_value);
            outdatamap.add_null(&loss);
            outdatamap.add_null(&wrong_labels);

            trainer.train_minibatch(&datamap, &mut outdatamap, DeviceDescriptor::cpu());
            let result = outdatamap.get(&output_value).unwrap().to_vec();
            let loss_val = outdatamap.get(&loss).unwrap().to_vec();
            let wrong_labels_val = outdatamap.get(&wrong_labels).unwrap().to_vec();
            lastloss = lastloss * 0.9 + 0.1*loss_val[0];
        }
        assert!(lastloss < 0.5);
    }

    #[test]
    fn simple_recurrence() {
        let x = Variable::input_variable(Shape::from_slice(&vec!(2)));
        let y = Variable::input_variable(Shape::from_slice(&vec!(2)));
        let placeholder = Variable::placeholder(Shape::from_slice(&vec!(2)));
        let output = plus(&placeholder, &elem_times(&x, &y));
        let placeholder_replacement = past_value(&output);

        let mut replacement_map = ReplacementMap::new();
        replacement_map.add(&placeholder, &placeholder_replacement);

        let output_function = Function::from_variable(&output).replace_placeholders(&replacement_map);

        let last_output = last(&output_function.outputs()[0]);
        let last_output_function = Function::from_variable(&last_output);

        let data: Vec<f32> = vec!(1.0, 2.0, 3.0, 4.0, 5.0, 6.0, 7.0, 8.0, 9.0, 10.0);
        let data2: Vec<f32> = vec!(1.0, 2.0, 3.0, 4.0, 5.0, 6.0, 7.0, 8.0, 9.0, 100.0);

        let val = Value::sequence(&x.shape(), &data, DeviceDescriptor::cpu());
        let val2 = Value::sequence(&y.shape(), &data2, DeviceDescriptor::cpu());

        let mut datamap = DataMap::new();
        datamap.add(&x, &val);
        datamap.add(&y, &val2);

        let mut outdatamap = DataMap::new();
        outdatamap.add_null(&output);

        output_function.evaluate(&datamap, &mut outdatamap, DeviceDescriptor::cpu());

        let result = outdatamap.get(&output).unwrap().to_vec();
        assert_eq!(result, vec!(1., 4., 10., 20., 35., 56., 84., 120., 165., 1120.));

        let mut outdatamap_last = DataMap::new();
        outdatamap_last.add_null(&last_output);
        last_output_function.evaluate(&datamap, &mut outdatamap_last, DeviceDescriptor::cpu());
        let result_last = outdatamap_last.get(&last_output).unwrap().to_vec();
        assert_eq!(result_last, vec!(165., 1120.));
    }
}