primal_lr = 1e-6
dual_lr = 1e-4
num_epochs = 1000
log_every = 100 # batches
test_every = 5 # epochs
optimizer_primal = torch.optim.Adam(model.primal.parameters(), lr=primal_lr)
optimizer_dual = torch.optim.Adam(model.dual.parameters(), lr=dual_lr)
history = SPDG(model,
optimizer_primal,
optimizer_dual,
sequence_loader,
data_loader,
test_loader,
num_epochs,
log_every,
test_every,
eval_predictions_on_data=True,
show_dual=False)
# %% DUAL TRAINING (CONTINUATION)
# history = SPDG(model, optimizer_primal, optimizer_dual, sequence_loader,
# data_loader, test_loader, num_epochs=10, log_every=100,
# test_every=1, eval_predictions_on_data=True, show_dual=True, history=history)
# # %% DUAL TRAINING (CONTINUATION VERY LONG)
# history = SPDG(model, optimizer_primal, optimizer_dual, sequence_loader,
# test_loader, num_epochs=1000, log_every=50, test_every=1, history=history)
history = History()
for idx in model.dual:
history['dual ' + str(idx)] = []
epochs_done = 0
while epochs_done < num_epochs:
history = SPDG(model,
optimizer_primal,
optimizer_dual,
sequence_loader,
data_loader,
test_loader,
save_every,
log_every,
test_every,
sequence_test_loader=sequence_test_loader,
predictions_on_data=predictions_on_data,
show_dual=show_dual,
predictions_on_sequences=predictions_on_sequences,
ngram_data_stats=ngram_data_stats,
ngram_test_stats=ngram_test_stats,
loss_on_test=loss_on_test,
history=history)
save(history, model, ngram, optimizer_primal, optimizer_dual, primal_lr,
dual_lr)
epochs_done += save_every
# %% PLOTTING
lr=primal_lr)
optimizer_dual = torch.optim.Adam(model.dual.parameters(), lr=dual_lr)
history = History()
for idx in model.dual:
history['dual ' + str(idx)] = []
epochs_done = 0
while epochs_done < num_epochs:
history = SPDG(model,
optimizer_primal,
optimizer_dual,
sequence_loader,
data_loader,
test_loader,
save_every,
log_every,
test_every,
sequence_test_loader=sequence_test_loader,
show_dual=False,
eval_predictions_on_sequences=True,
history=history)
save(history, model, ngram, optimizer_primal, optimizer_dual, primal_lr,
dual_lr)
epochs_done += save_every
# %% PLOTTING TEST
xs = np.arange(len(history['predictions'])) * test_every
ys = [[
100.0 - preds[i, i] / preds[i].sum() * 100
model = Model(ngram, output_size=2)
model.to(DEVICE)
model.init_weights()
primal_lr = 1e-6
dual_lr = 1e-4
optimizer_primal = torch.optim.Adam(model.primal.parameters(), lr=primal_lr)
optimizer_dual = torch.optim.Adam(model.dual.parameters(), lr=dual_lr)
history = SPDG(model,
optimizer_primal,
optimizer_dual,
sequence_loader,
sequence_loader,
test_loader,
num_epochs=1000,
log_every=100,
test_every=5,
eval_predictions_on_data=True,
show_dual=True)
# %% SAVE
fname = 't1-1-1'
comment = ''
np.save(fname + '_hist', history)
np.save(fname + '_model', model)
np.save(fname + '_ngram', ngram)
with open(fname + '_doc', "w+") as doc:
