Using the example training file bigdata.tr.txt and validation file bigdata.te.txt to perform the ffm test:

difacto local.conf data_in=data/bigdata.tr.txt data_val=data/bigdata.te.txt learner=ffmsgd  V_dim=4 max_num_epochs=10 batch_size=1000 has_aux=1 field_num=18

The program exits after only 2 epoch:

[12:23:58] /opt/codebase/dmlc/DiFacto2_ffm/src/sgd/sgd_learner.cc:80: Start epoch 0
    1        200    2e+02 |         0 | 0.6932  0.64692
[12:23:59] /opt/codebase/dmlc/DiFacto2_ffm/src/sgd/sgd_learner.cc:82: Epoch[0] Training: Rows = 200, loss = 0.693153, AUC = 0.646919
[12:24:01] /opt/codebase/dmlc/DiFacto2_ffm/src/sgd/sgd_learner.cc:87: Epoch[0] Validation: Rows = 200, loss = 0.693147, AUC = 0.619823
[12:24:01] /opt/codebase/dmlc/DiFacto2_ffm/src/sgd/sgd_learner.cc:80: Start epoch 1
    4        400    2e+02 |         0 | 0.6931  0.61127
[12:24:02] /opt/codebase/dmlc/DiFacto2_ffm/src/sgd/sgd_learner.cc:82: Epoch[1] Training: Rows = 200, loss = 0.693147, AUC = 0.611268
[12:24:04] /opt/codebase/dmlc/DiFacto2_ffm/src/sgd/sgd_learner.cc:87: Epoch[1] Validation: Rows = 200, loss = 0.693147, AUC = 0.619823
[12:24:04] /opt/codebase/dmlc/DiFacto2_ffm/src/sgd/sgd_learner.cc:94: Change of loss [8.80543e-06] < stop_rel_objv [1e-05]

It seems ffm can not converge during the training process

Hi，thanks for your work. Does this distributed FFM have any report?

How about the precision compared with libFFM when running on multiple machines and convergence situation？

I wondered how data is dispatched to different machines and make sure each machine reads different data.

Can you give me some suggestions when implement a distributed FFM on ps-lite from scratch? you can say in Chinese.

Thank you!

output:
[17:03:55] src/main.cc:40: - l1_shrk = 0
[17:03:55] src/main.cc:71: start run learner
[17:03:55] src/sgd/sgd_learner.cc:80: Start epoch 0
1 0 0 | 0 | -nan -nan
2 0 0 | 0 | -nan -nan
3 0 0 | 0 | -nan -nan
4 0 0 | 0 | -nan -nan
5 0 0 | 0 | -nan -nan
6 0 0 | 0 | -nan -nan
7 0 0 | 0 | -nan -nan
8 0 0 | 0 | -nan -nan
9 0 0 | 0 | -nan -nan
10 0 0 | 0 | -nan -nan
11 0 0 | 0 | -nan -nan
12 0 0 | 0 | -nan -nan
13 0 0 | 0 | -nan -nan
14 0 0 | 0 | -nan -nan
15 0 0 | 0 | -nan -nan
16 0 0 | 0 | -nan -nan
17 1e+04 1e+04 | 0 | 0.6931 0.55892
local.conf:

data

data_in = /data/data1/ffm/data/train.data
data_val = /data/data1/ffm/data/test.data

learner

learner = sgd
max_num_epochs = 2
batch_size = 10000
field_num=68

V_dim = 2
has_aux=1
l1_shrk=0

void SGDUpdater::UpdateV(real_t const* gV, SGDEntry* e) {
int nnz = e->nnz;
for (int i = 0; i < feat_dim; ++i) {
real_t sg = e->Z[i];
real_t vi = e->V[i];

// update sqrt_g
real_t gv = gV[i] + vi * param_.l2;
**e->Z[i] = sqrt(sg * sg + gv * gv);
e->V[i] -= param_.lr * e->Z[i] * gv;**