def test_valtest(sess, dcmf, epoch, step):
val_epochs = Misc.get_epoch_files(dcmf.mfp.val_epochs)
print 'Testing Perf: Val\t', epoch, step
val_file = val_epochs[0]
print 'Val file: ', val_file
v_val_review = FromDisk.FromDisk(val_file)
val_iter = v_val_review.BatchIter(dcmf.mfp.batch_size)
oth_mse_val, full_mse_val = test(sess, dcmf, val_iter)
print 'Testing MSE Other: Val\t', epoch, step, '\t', oth_mse_val
print 'Testing MSE Full: Val\t', epoch, step, '\t', full_mse_val
test_epochs = Misc.get_epoch_files(dcmf.mfp.test_epochs)
print 'Testing Perf: Test\t', epoch, step
test_file = test_epochs[0]
print 'Test file: ', test_file
v_test_review = FromDisk.FromDisk(test_file)
test_iter = v_test_review.BatchIter(dcmf.mfp.batch_size)
oth_mse_test, full_mse_test = test(sess, dcmf, test_iter)
print 'Testing MSE Other: Test\t', epoch, step, '\t', oth_mse_test
print 'Testing MSE Full: Test\t', epoch, step, '\t', full_mse_test
return full_mse_val
def test_valtest(sess, dcmf, epoch, step):
val_epochs = Misc.get_epoch_files(dcmf.mfp.val_epochs)
print('Testing Perf: Val\t', epoch, step)
val_file = val_epochs[0]
print('Val file: ', val_file)
v_val_review = FromDisk.FromDisk(val_file)
val_iter = v_val_review.BatchIter(dcmf.mfp.batch_size)
oth_mse_val, full_mse_val = test(sess, dcmf, val_iter)
print('Testing MSE Other: Val\t', epoch, step, '\t', oth_mse_val)
print('Testing MSE Full: Val\t', epoch, step, '\t', full_mse_val)
test_epochs = Misc.get_epoch_files(dcmf.mfp.test_epochs)
print('Testing Perf: Test\t', epoch, step)
test_file = test_epochs[0]
print('Test file: ', test_file)
v_test_review = FromDisk.FromDisk(test_file)
test_iter = v_test_review.BatchIter(dcmf.mfp.batch_size)
oth_mse_test, full_mse_test = test(sess, dcmf, test_iter)
print('Testing MSE Other: Test\t', epoch, step, '\t', oth_mse_test)
print('Testing MSE Full: Test\t', epoch, step, '\t', full_mse_test)
if full_mse_val < 1.7:
#TODO: need to find the saving criteria from the previous MSE
return True, full_mse_val #save this model
else:
return False, full_mse_val
def BatchIter(self, batch_size):
'''
batch size = number of training u,b,r examples in the batch
returns:
uList = uA useres
bList = iB items
rList: rAB (float)
user_revlist: the UText converted to int list
item_revlist: the BText converted to int list
'''
while True:
#one batch
start = time.time()
uList = []
bList = []
rList = []
uTextList = []
bTextList = []
for line in self.fin:
vals = line.split("\t")
if len(vals) == 0:
continue
u = vals[0]
b = vals[1]
r = float(vals[2])
uText = vals[3]
bText = vals[4]
uList.append(u)
bList.append(b)
rList.append(r)
uTextList.append(Misc.int_list(uText))
bTextList.append(Misc.int_list(bText))
if len(uList) >= batch_size:
break
if len(uList) == 0:
#end of data
self._close()
print('Total Batch gen time = ', (self.tot_batch / 60.0),
' min')
raise StopIteration
end = time.time()
bg = (end - start)
print('Batch gen time = ', bg, ' sec')
self.tot_batch += bg
yield uList, bList, rList, uTextList, bTextList
def train(dcmf, savedir):
cfg = tf.ConfigProto(allow_soft_placement=True)
cfg.gpu_options.allow_growth = True
sess = tf.Session(config=cfg)
min_MSE = float("inf")
#read the embedding
emb = pickle.load(open(dcmf.mfp.word_embedding_file, "rb"))
dcmf.run_init_all(sess, emb)
del emb
step = 0
#get the epoch files from the train dir
train_epochs = Misc.get_epoch_files(dcmf.mfp.train_epochs)
print 'Train Epochs: found ', len(train_epochs), ' files'
#get the epoch files from the val dir
val_epochs = Misc.get_epoch_files(dcmf.mfp.val_epochs)
print 'Val Epochs: found ', len(val_epochs), ' files'
#get the epoch files from the test dir
test_epochs = Misc.get_epoch_files(dcmf.mfp.test_epochs)
print 'Test Epochs: found ', len(test_epochs), ' files'
#load the revAB from the train file
dp_mgr = DataPairMgr.DataPairMgr(dcmf.mfp.train_data)
names = True
for epoch in range(dcmf.mfp.max_epoch):
print 'Epoch: ', epoch
train_time = 0
train_file = train_epochs[epoch % len(train_epochs)]
print 'Train file: ', train_file
trainIter = FromDisk.FromDisk(train_file)
batch_iter = trainIter.BatchIter(dcmf.mfp.batch_size)
while True:
step += 1
rList, retUTextInt, retBTextInt, revABList = ([] for i in range(4))
try:
#read the values
uList, bList, rList, retUTextInt, retBTextInt = batch_iter.next(
)
#get the revAB
revABList = [
dp_mgr.get_int_review(u, b) for u, b in zip(uList, bList)
]
except StopIteration:
#end of this data epoch
break
start = time.time()
act_rmse, oth_rmse, full_rmse = dcmf.run_train_step(
sess, rList, retUTextInt, retBTextInt, revABList,
dcmf.mfp.dropout_keep_prob)
end = time.time()
if names:
names = False
names_all, names_act, names_oth, names_full = dcmf.get_params()
print 'Variables - all trainable: ', names_all
print 'Variables trained in act: ', names_act
print 'Variables trained in oth: ', names_oth
print 'Variables trained in full: ', names_full
tt = end - start
print 'Train time ', (tt), ' sec'
train_time += tt
print 'Step ', step, ' act: ', act_rmse
print 'Step ', step, ' oth: ', oth_rmse
print 'Step ', step, ' full: ', full_rmse
if step % 500 == 0:
mse = test_valtest(sess, dcmf, epoch, step)
if mse < min_MSE:
#save the current model
min_MSE = mse
save_model(sess, dcmf, mse, savedir)
print 'End of Epoch Testing'
mse = test_valtest(sess, dcmf, epoch, step)
if mse < min_MSE:
#save the current model
save_model(sess, dcmf, mse, savedir)
sys.stdout.flush()
def train(dcmf, savedir):
cfg = tf.ConfigProto(allow_soft_placement=True )
cfg.gpu_options.allow_growth = True
sess = tf.Session(config=cfg)
#read the embedding
emb = pickle.load( open( dcmf.mfp.word_embedding_file, "rb" ),encoding="bytes" )
dcmf.run_init_all(sess, emb)
del emb
step = 0
#get the epoch files from the train dir
train_epochs = Misc.get_epoch_files(dcmf.mfp.train_epochs)
print ('Train Epochs: found ', len(train_epochs), ' files')
#get the epoch files from the val dir
val_epochs = Misc.get_epoch_files(dcmf.mfp.val_epochs)
print ('Val Epochs: found ', len(val_epochs), ' files')
#get the epoch files from the test dir
test_epochs = Misc.get_epoch_files(dcmf.mfp.test_epochs)
print ('Test Epochs: found ', len(test_epochs), ' files')
#load the revAB from the train file NO DATA PAIr MANAGER
# dp_mgr = DataPairMgr.DataPairMgr(dcmf.mfp.train_data)
names = True
for epoch in range(dcmf.mfp.max_epoch):
print ('Epoch: ', epoch)
train_time = 0
train_file = train_epochs[epoch % len(train_epochs)]
print ('Train file: ', train_file)
trainIter = FromDisk.FromDisk(train_file,dcmf.mfp.review_delim,dcmf.mfp.review_emb)
batch_iter = trainIter.BatchIter(dcmf.mfp.batch_size)
while True:
step += 1
rList, retUTextInt, retBTextInt, revABList, revABList = ([] for i in range(5))
try:
#read the values
uList, bList, rList, retUTextInt, retBTextInt, revABList = next(batch_iter)
#get the revAB
# revABList = [ dp_mgr.get_int_review(u, b) for u,b in zip(uList, bList) ]
except StopIteration:
#end of this data epoch
break
start = time.time()
act_rmse, oth_rmse, full_rmse = dcmf.run_train_step(sess, rList, retUTextInt, retBTextInt, revABList, dcmf.mfp.dropout_keep_prob)
end = time.time()
if names:
names = False
names_all, names_act, names_oth, names_full = dcmf.get_params()
print ('Variables - all trainable: ', names_all)
print ('Variables trained in act: ', names_act)
print ('Variables trained in oth: ', names_oth)
print ('Variables trained in full: ', names_full)
tt= end - start
#print ('Train time ', (tt), ' sec')
train_time += tt
#print ('Step ', step, ' act: ', act_rmse)
#print ('Step ', step, ' oth: ', oth_rmse)
print ('Step ', step, ' full: ', full_rmse)
if step % 1000 == 0:
print('SAVING MODEL:')
save_model(sess, dcmf, full_rmse, savedir)
save, mse = test_valtest(sess, dcmf, epoch, step)
#if save:
#print('SAVING MODEL:',mse)
#save the current model
#save_model(sess, dcmf, mse, savedir)
print ('End of Epoch Testing')
save, mse = test_valtest(sess, dcmf, epoch, step)
#if save:
#save the current model
save_model(sess, dcmf, mse, savedir)
sys.stdout.flush()