log-probability plot for output - Google Groups

5 downloads 174 Views 1MB Size Report
Plot of log-probability vs iterations for output. 0. 200. 400. 600. 800. 1000. Iteration. 0.30. 0.25. 0.20. 0.15. 0.10.
1 FIG. 1: output

log-probability plot for output

Plot of log-probability vs iterations for

0.05

log-probability

0.10 0.15 0.20 0.25 0.30 0

200

400

600 Iteration

800

train exp/chain/tdnn7o_sp valid exp/chain/tdnn7o_sp

1000

2

r-dimension average-(value, derivative) and rms-oderivative percentiles for prefina

Oderivative-RectifiedLinear Derivative-RectifiedLinear

Value-RectifiedLinear

FIG. 2: Per-dimension average-(value, derivative) and rms-oderivative percentiles for prefinal-chain.relu

1 0

0

200

400

600

800

1000

0

200

400

600

800

1000

0

200

400

600 Iteration

800

1000

0.75 0.50 0.25

0.010 0.005

[1]/tdnn7o_sp

[1]:exp/chain 5th percentile 50th percentile

95th percentile

3

er-dimension average-(value, derivative) and rms-oderivative percentiles for prefina

Oderivative-RectifiedLinear

Derivative-RectifiedLinear

Value-RectifiedLinear

FIG. 3: Per-dimension average-(value, derivative) and rms-oderivative percentiles for prefinal-xent.relu

1 0

0

200

400

600

800

1000

0

200

400

600

800

1000

0

200

400

600 Iteration

800

1000

1.0 0.5

0.004 0.002

[1]/tdnn7o_sp

[1]:exp/chain 5th percentile 50th percentile

95th percentile

4

Per-dimension average-(value, derivative) and rms-oderivative percentiles for tdn

Oderivative-RectifiedLinear Derivative-RectifiedLinear

Value-RectifiedLinear

FIG. 4: Per-dimension average-(value, derivative) and rms-oderivative percentiles for tdnn1.relu

1 0

0

200

400

600

800

1000

0

200

400

600

800

1000

0

200

400

600 Iteration

800

1000

1.0 0.5 0.0

0.10 0.05

[1]/tdnn7o_sp

[1]:exp/chain 5th percentile 50th percentile

95th percentile

5

Per-dimension average-(value, derivative) and rms-oderivative percentiles for tdnn

Oderivative-RectifiedLinear Derivative-RectifiedLinear

Value-RectifiedLinear

FIG. 5: Per-dimension average-(value, derivative) and rms-oderivative percentiles for tdnn10.relu

1 0

0

200

400

600

800

1000

0

200

400

600

800

1000

0

200

400

600 Iteration

800

1000

0.75 0.50 0.25

0.010 0.005

[1]/tdnn7o_sp

[1]:exp/chain 5th percentile 50th percentile

95th percentile

6

Per-dimension average-(value, derivative) and rms-oderivative percentiles for tdnn

Oderivative-RectifiedLinear

Derivative-RectifiedLinear

Value-RectifiedLinear

FIG. 6: Per-dimension average-(value, derivative) and rms-oderivative percentiles for tdnn11.relu

1 0

0

200

400

600

800

1000

0

200

400

600

800

1000

0

200

400

600 Iteration

800

1000

0.75 0.50 0.25

0.0075 0.0050 0.0025

[1]/tdnn7o_sp

[1]:exp/chain 5th percentile 50th percentile

95th percentile

7

Per-dimension average-(value, derivative) and rms-oderivative percentiles for tdn

Oderivative-RectifiedLinear Derivative-RectifiedLinear

Value-RectifiedLinear

FIG. 7: Per-dimension average-(value, derivative) and rms-oderivative percentiles for tdnn2.relu

1 0

0

200

400

600

800

1000

0

200

400

600

800

1000

0

200

400

600 Iteration

800

1000

1.0 0.5

0.02 0.01 0.00

[1]/tdnn7o_sp

[1]:exp/chain 5th percentile 50th percentile

95th percentile

8

Per-dimension average-(value, derivative) and rms-oderivative percentiles for tdn

Oderivative-RectifiedLinear Derivative-RectifiedLinear

Value-RectifiedLinear

FIG. 8: Per-dimension average-(value, derivative) and rms-oderivative percentiles for tdnn3.relu

1.5 1.0 0.5 0

200

400

600

800

1000

0

200

400

600

800

1000

0

200

400

600 Iteration

800

1000

0.75 0.50 0.25

0.006 0.004 0.002

[1]/tdnn7o_sp

[1]:exp/chain 5th percentile 50th percentile

95th percentile

9

Per-dimension average-(value, derivative) and rms-oderivative percentiles for tdn

Oderivative-RectifiedLinear Derivative-RectifiedLinear

Value-RectifiedLinear

FIG. 9: Per-dimension average-(value, derivative) and rms-oderivative percentiles for tdnn4.relu

1 0

0

200

400

600

800

1000

0

200

400

600

800

1000

0

200

400

600 Iteration

800

1000

1.00 0.75 0.50 0.25

0.02 0.01 0.00

[1]/tdnn7o_sp

[1]:exp/chain 5th percentile 50th percentile

95th percentile

10

Per-dimension average-(value, derivative) and rms-oderivative percentiles for tdn

Oderivative-RectifiedLinear Derivative-RectifiedLinear

Value-RectifiedLinear

FIG. 10: Per-dimension average-(value, derivative) and rms-oderivative percentiles for tdnn5.relu

1.5 1.0 0.5 0.0

0

200

400

600

800

1000

0

200

400

600

800

1000

0

200

400

600 Iteration

800

1000

0.75 0.50 0.25

0.010 0.005

[1]/tdnn7o_sp

[1]:exp/chain 5th percentile 50th percentile

95th percentile

11

Per-dimension average-(value, derivative) and rms-oderivative percentiles for tdn

Oderivative-RectifiedLinear

Derivative-RectifiedLinear

Value-RectifiedLinear

FIG. 11: Per-dimension average-(value, derivative) and rms-oderivative percentiles for tdnn6.relu

1 0

0

200

400

600

800

1000

0

200

400

600

800

1000

0

200

400

600 Iteration

800

1000

0.75 0.50 0.25

0.0075 0.0050 0.0025

[1]/tdnn7o_sp

[1]:exp/chain 5th percentile 50th percentile

95th percentile

12

Per-dimension average-(value, derivative) and rms-oderivative percentiles for tdn

Oderivative-RectifiedLinear

Derivative-RectifiedLinear

Value-RectifiedLinear

FIG. 12: Per-dimension average-(value, derivative) and rms-oderivative percentiles for tdnn7.relu

1.5 1.0 0.5 0

200

400

600

800

1000

0

200

400

600

800

1000

0

200

400

600 Iteration

800

1000

0.75 0.50 0.25

0.0075 0.0050 0.0025

[1]/tdnn7o_sp

[1]:exp/chain 5th percentile 50th percentile

95th percentile

13

Per-dimension average-(value, derivative) and rms-oderivative percentiles for tdn

Oderivative-RectifiedLinear Derivative-RectifiedLinear

Value-RectifiedLinear

FIG. 13: Per-dimension average-(value, derivative) and rms-oderivative percentiles for tdnn8.relu

1 0

0

200

400

600

800

1000

0

200

400

600

800

1000

0

200

400

600 Iteration

800

1000

0.75 0.50 0.25

0.010 0.005

[1]/tdnn7o_sp

[1]:exp/chain 5th percentile 50th percentile

95th percentile

14

Per-dimension average-(value, derivative) and rms-oderivative percentiles for tdn

Oderivative-RectifiedLinear

Derivative-RectifiedLinear

Value-RectifiedLinear

FIG. 14: Per-dimension average-(value, derivative) and rms-oderivative percentiles for tdnn9.relu

1 0

0

200

400

600

800

1000

0

200

400

600

800

1000

0

200

400

600 Iteration

800

1000

0.75 0.50 0.25

0.0075 0.0050 0.0025

[1]/tdnn7o_sp

[1]:exp/chain 5th percentile 50th percentile

95th percentile

Parameter differences at output-xent.affine

Relative Parameter Differences Parameter Differences

FIG. 15: Parameter differences at output-xent.affine

15 10 5 0

200

400

600

800

1000

0

200

400

600 Iteration

800

1000

0.6 0.4 0.2 0.0

Parameter Differences exp/chain/tdnn7o_sp

15

16

Parameter differences at output.affine

Relative Parameter Differences

Parameter Differences

FIG. 16: Parameter differences at output.affine

8 6 4 2 0

200

400

600

800

1000

0

200

400

600 Iteration

800

1000

0.4 0.3 0.2 0.1

Parameter Differences exp/chain/tdnn7o_sp

17

Parameter differences at prefinal-chain-l

Relative Parameter Differences

Parameter Differences

FIG. 17: Parameter differences at prefinal-chain-l

4 3 2 1 0

200

400

600

800

1000

0

200

400

600 Iteration

800

1000

0.25 0.20 0.15 0.10 0.05

Parameter Differences exp/chain/tdnn7o_sp

18

Relative Parameter Differences Parameter Differences

FIG. 18: prefinal-chain.affine

Parameter differences at prefinal-chain.affine Parameter differences at

1.75 1.50 1.25 1.00 0.75

0

200

400

600

800

1000

0

200

400

600 Iteration

800

1000

0.06 0.05 0.04 0.03 0.02

Parameter Differences exp/chain/tdnn7o_sp

19

Parameter differences at prefinal-l

Relative Parameter Differences

Parameter Differences

FIG. 19: Parameter differences at prefinal-l

4 3 2 1

0.30 0.25 0.20 0.15 0.10 0.05

0

200

400

600

800

1000

0

200

400

600 Iteration

800

1000

Parameter Differences exp/chain/tdnn7o_sp

20

Parameter differences at prefinal-xent-l

Relative Parameter Differences

Parameter Differences

FIG. 20: Parameter differences at prefinal-xent-l

4 3 2 1 0

200

400

600

800

1000

0

200

400

600 Iteration

800

1000

0.25 0.20 0.15 0.10 0.05

Parameter Differences exp/chain/tdnn7o_sp

21

Relative Parameter Differences Parameter Differences

FIG. 21: prefinal-xent.affine

Parameter differences at prefinal-xent.affine Parameter differences at

1.5 1.0 0.5

0

200

400

600

800

1000

0

200

400

600 Iteration

800

1000

0.04 0.03 0.02

Parameter Differences exp/chain/tdnn7o_sp

22

Parameter differences at tdnn1.affine

Relative Parameter Differences Parameter Differences

FIG. 22: Parameter differences at tdnn1.affine

3.0 2.5 2.0 1.5 0

200

400

600

800

1000

0

200

400

600 Iteration

800

1000

0.08 0.06 0.04

Parameter Differences exp/chain/tdnn7o_sp

23

Parameter differences at tdnn10.affine

Relative Parameter Differences Parameter Differences

FIG. 23: Parameter differences at tdnn10.affine

1.6 1.4 1.2 1.0 0.8 0

200

400

600

800

1000

0

200

400

600 Iteration

800

1000

0.06 0.04 0.02

Parameter Differences exp/chain/tdnn7o_sp

24

Parameter differences at tdnn10l

Relative Parameter Differences

Parameter Differences

FIG. 24: Parameter differences at tdnn10l

6 4 2 0

200

400

600

800

1000

0

200

400

600 Iteration

800

1000

0.3 0.2 0.1

Parameter Differences exp/chain/tdnn7o_sp

25

Parameter differences at tdnn10l0

Relative Parameter Differences

Parameter Differences

FIG. 25: Parameter differences at tdnn10l0

4 3 2 1

0.25

0

200

400

600

800

1000

0

200

400

600 Iteration

800

1000

0.20 0.15 0.10 0.05

Parameter Differences exp/chain/tdnn7o_sp

26

Parameter differences at tdnn11.affine

Relative Parameter Differences Parameter Differences

FIG. 26: Parameter differences at tdnn11.affine

3.5 3.0 2.5 2.0 1.5 1.0

0.07 0.06 0.05 0.04 0.03

0

200

400

600

800

1000

0

200

400

600 Iteration

800

1000

Parameter Differences exp/chain/tdnn7o_sp

27

Parameter differences at tdnn11l

Relative Parameter Differences

Parameter Differences

FIG. 27: Parameter differences at tdnn11l

6 4 2 0

200

400

600

800

1000

0

200

400

600 Iteration

800

1000

0.3 0.2 0.1

Parameter Differences exp/chain/tdnn7o_sp

28

Parameter differences at tdnn11l0

Relative Parameter Differences

Parameter Differences

FIG. 28: Parameter differences at tdnn11l0

3 2 1 0

200

400

600

800

1000

0

200

400

600 Iteration

800

1000

0.20 0.15 0.10 0.05

Parameter Differences exp/chain/tdnn7o_sp

29

Parameter differences at tdnn2.affine

Relative Parameter Differences Parameter Differences

FIG. 29: Parameter differences at tdnn2.affine

3.0 2.5 2.0 1.5 1.0

0

200

400

600

800

1000

0

200

400

600 Iteration

800

1000

0.07 0.06 0.05 0.04 0.03

Parameter Differences exp/chain/tdnn7o_sp

30

Parameter differences at tdnn2l

Relative Parameter Differences

Parameter Differences

FIG. 30: Parameter differences at tdnn2l

6 4 2 0

200

400

600

800

1000

0

200

400

600 Iteration

800

1000

0.4 0.3 0.2 0.1

Parameter Differences exp/chain/tdnn7o_sp

31

Parameter differences at tdnn2l0

Relative Parameter Differences

Parameter Differences

FIG. 31: Parameter differences at tdnn2l0

4 3 2 0

200

400

600

800

1000

0

200

400

600 Iteration

800

1000

0.25 0.20 0.15 0.10 0.05

Parameter Differences exp/chain/tdnn7o_sp

32

Parameter differences at tdnn3.affine

Relative Parameter Differences Parameter Differences

FIG. 32: Parameter differences at tdnn3.affine

2.0 1.5 1.0 0

200

400

600

800

1000

0

200

400

600 Iteration

800

1000

0.08 0.06 0.04 0.02

Parameter Differences exp/chain/tdnn7o_sp

33

Parameter differences at tdnn3l

Relative Parameter Differences

Parameter Differences

FIG. 33: Parameter differences at tdnn3l

4 3 2 1

0.25 0.20 0.15 0.10 0.05

0

200

400

600

800

1000

0

200

400

600 Iteration

800

1000

Parameter Differences exp/chain/tdnn7o_sp

34

Parameter differences at tdnn4.affine

Relative Parameter Differences Parameter Differences

FIG. 34: Parameter differences at tdnn4.affine

2.00 1.75 1.50 1.25 1.00 0.75 0

200

400

600

800

1000

0

200

400

600 Iteration

800

1000

0.08 0.06 0.04 0.02

Parameter Differences exp/chain/tdnn7o_sp

35

Parameter differences at tdnn4l

Relative Parameter Differences

Parameter Differences

FIG. 35: Parameter differences at tdnn4l

6 4 2

0.4

0

200

400

600

800

1000

0

200

400

600 Iteration

800

1000

0.3 0.2 0.1

Parameter Differences exp/chain/tdnn7o_sp

36

Parameter differences at tdnn4l0

Relative Parameter Differences

Parameter Differences

FIG. 36: Parameter differences at tdnn4l0

4 3 2 1 0

200

400

600

800

1000

0

200

400

600 Iteration

800

1000

0.25 0.20 0.15 0.10 0.05

Parameter Differences exp/chain/tdnn7o_sp

37

Parameter differences at tdnn5.affine

Relative Parameter Differences Parameter Differences

FIG. 37: Parameter differences at tdnn5.affine

2.25 2.00 1.75 1.50 1.25 1.00 0

200

400

600

800

1000

0

200

400

600 Iteration

800

1000

0.08 0.06 0.04

Parameter Differences exp/chain/tdnn7o_sp

38

Parameter differences at tdnn5l

Relative Parameter Differences

Parameter Differences

FIG. 38: Parameter differences at tdnn5l

4 3 2 1 0

200

400

600

800

1000

0

200

400

600 Iteration

800

1000

0.25 0.20 0.15 0.10 0.05

Parameter Differences exp/chain/tdnn7o_sp

39

Parameter differences at tdnn6.affine

Relative Parameter Differences Parameter Differences

FIG. 39: Parameter differences at tdnn6.affine

1.8 1.6 1.4 1.2 1.0 0.8

0

200

400

600

800

1000

0

200

400

600 Iteration

800

1000

0.07 0.06 0.05 0.04 0.03 0.02

Parameter Differences exp/chain/tdnn7o_sp

40

Parameter differences at tdnn6l

Relative Parameter Differences

Parameter Differences

FIG. 40: Parameter differences at tdnn6l

6 4 2 0

200

400

600

800

1000

0

200

400

600 Iteration

800

1000

0.3 0.2 0.1

Parameter Differences exp/chain/tdnn7o_sp

41

Parameter differences at tdnn6l0

Relative Parameter Differences

Parameter Differences

FIG. 41: Parameter differences at tdnn6l0

4 3 2 1 0

200

400

600

800

1000

0

200

400

600 Iteration

800

1000

0.25 0.20 0.15 0.10 0.05

Parameter Differences exp/chain/tdnn7o_sp

42

Parameter differences at tdnn7.affine

Relative Parameter Differences Parameter Differences

FIG. 42: Parameter differences at tdnn7.affine

2.5 2.0 1.5 0

200

400

600

800

1000

0

200

400

600 Iteration

800

1000

0.07 0.06 0.05 0.04 0.03

Parameter Differences exp/chain/tdnn7o_sp

43

Parameter differences at tdnn7l

Relative Parameter Differences

Parameter Differences

FIG. 43: Parameter differences at tdnn7l

6 4 2 0

200

400

600

800

1000

0

200

400

600 Iteration

800

1000

0.3 0.2 0.1

Parameter Differences exp/chain/tdnn7o_sp

44

Parameter differences at tdnn7l0

Relative Parameter Differences

Parameter Differences

FIG. 44: Parameter differences at tdnn7l0

3 2 1 0

200

400

600

800

1000

0

200

400

600 Iteration

800

1000

0.20 0.15 0.10 0.05

Parameter Differences exp/chain/tdnn7o_sp

45

Parameter differences at tdnn8.affine

Relative Parameter Differences Parameter Differences

FIG. 45: Parameter differences at tdnn8.affine

1.6 1.4 1.2 1.0 0.8

0.07 0.06 0.05 0.04 0.03 0.02

0

200

400

600

800

1000

0

200

400

600 Iteration

800

1000

Parameter Differences exp/chain/tdnn7o_sp

46

Parameter differences at tdnn8l

Relative Parameter Differences

Parameter Differences

FIG. 46: Parameter differences at tdnn8l

6 4 2 0

200

400

600

800

1000

0

200

400

600 Iteration

800

1000

0.3 0.2 0.1

Parameter Differences exp/chain/tdnn7o_sp

47

Parameter differences at tdnn8l0

Relative Parameter Differences

Parameter Differences

FIG. 47: Parameter differences at tdnn8l0

4 3 2 1

0.25

0

200

400

600

800

1000

0

200

400

600 Iteration

800

1000

0.20 0.15 0.10 0.05

Parameter Differences exp/chain/tdnn7o_sp

48

Parameter differences at tdnn9.affine

Relative Parameter Differences Parameter Differences

FIG. 48: Parameter differences at tdnn9.affine

2.5 2.0 1.5 1.0

0

200

400

600

800

1000

0

200

400

600 Iteration

800

1000

0.08 0.06 0.04

Parameter Differences exp/chain/tdnn7o_sp

49

Parameter differences at tdnn9l

Relative Parameter Differences

Parameter Differences

FIG. 49: Parameter differences at tdnn9l

6 4 2 0

200

400

600

800

1000

0

200

400

600 Iteration

800

1000

0.3 0.2 0.1

Parameter Differences exp/chain/tdnn7o_sp

50

Parameter differences at tdnn9l0

Relative Parameter Differences

Parameter Differences

FIG. 50: Parameter differences at tdnn9l0

3 2 1 0

200

400

600

800

1000

0

200

400

600 Iteration

800

1000

0.20 0.15 0.10 0.05

Parameter Differences exp/chain/tdnn7o_sp