stk = comstok.w2i['<s>']
comstart[0] = stk
outfn = outdir+"/predictions/predict-{}-cfw.txt".format(outfile.split('.')[0])
outf = open(outfn, 'w')
print("writing to file: " + outfn)
batch_sets = [allfids[i:i+batchsize] for i in range(0, len(allfids), batchsize)]
prep("computing predictions...\n")
for c, fid_set in enumerate(batch_sets):
st = timer()
for fid in fid_set:
refcoms[fid] = seqdata['c'+testval][fid]
seqdata['c'+testval][fid] = comstart #np.asarray([stk])
bg = batch_gen(seqdata, testval, config, training=False)
batch = bg.make_batch(fid_set)
if config['batch_maker'] == 'datsonly':
batch_results = gendescr_2inp(model, batch, comstok, comlen, batchsize, config, refcoms, c, strat='greedy')
elif config['batch_maker'] == 'ast':
batch_results = gendescr_3inp(model, batch, comstok, comlen, batchsize, config, refcoms, c, strat='greedy')
elif config['batch_maker'] == 'ast_threed':
batch_results = gendescr_4inp(model, batch, comstok, comlen, batchsize, config, refcoms, c, strat='greedy')
elif config['batch_maker'] == 'threed':
batch_results = gendescr_threed(model, batch, comstok, comlen, batchsize, config, refcoms, c, strat='greedy')
elif config['batch_maker'] == 'graphast':
batch_results = gendescr_graphast(model, batch, comstok, comlen, batchsize, config, refcoms, c, strat='greedy')
elif config['batch_maker'] == 'graphast_threed':
batch_results = gendescr_5inp(model, batch, comstok, comlen, batchsize, config, refcoms, c, strat='greedy')
elif config['batch_maker'] == 'pathast_threed':
config['tdatlen'] = len(list(seqdata['dttrain'].values())[0])
config['sdatlen'] = seqdata['config']['sdatlen']
config['comlen'] = len(list(seqdata['ctrain'].values())[0])
config['smllen'] = len(list(seqdata['strain'].values())[0])
config['multigpu'] = multigpu
config['batch_size'] = batch_size
prep('creating model... ')
config, model = create_model(modeltype, config)
drop()
print(model.summary())
gen = batch_gen(seqdata, 'train', modeltype, config)
checkpoint = ModelCheckpoint(outdir + '/models/' + modeltype +
'_E{epoch:02d}_' + str(timestart) + '.h5')
savehist = HistoryCallback()
savehist.setCatchExit(outdir, modeltype, timestart, config)
valgen = batch_gen(seqdata, 'val', modeltype, config)
callbacks = [checkpoint, savehist]
try:
history = model.fit_generator(gen,
steps_per_epoch=steps,
epochs=epochs,
verbose=1,
max_queue_size=3,
# cmd = "grep {} /nfs/projects/attn-to-fc/data/standard/output/sdats.test".format(fid)
# print('sdats for {}'.format(fid))
# os.system(cmd)
# print()
#print(modeltypewsdats, configwsdats['batch_maker'])
#print(modeltypewosdats, configwosdats['batch_maker'])
#sys.exit()
prep("computing predictions...\n")
for c, fid_set in enumerate(batch_sets):
st = timer()
for fid in fid_set:
seqdata['ctest'][fid] = comstart #np.asarray([stk])
bg = batch_gen(seqdata, 'test', configwsdats, training=False)
batch = bg.make_batch(fid_set)
if configwsdats['batch_maker'] == 'datsonly':
batch_resultswsdats = gendescr_2inp(modelwsdats, batch, comstok,
comlen, batchsize,
configwsdats, strat, beamwidth,
outfilewsdats, stopword)
elif configwsdats['batch_maker'] == 'ast':
batch_resultswsdats = gendescr_3inp(modelwsdats, batch, comstok,
comlen, batchsize,
configwsdats, strat, beamwidth,
outfilewsdats, stopword,
outdir)
elif configwsdats['batch_maker'] == 'ast_threed':
batch_resultswsdats = gendescr_4inp(modelwsdats, batch, comstok,
config['tdatlen'] = len(list(seqdata['dttrain'].values())[0])
config['sdatlen'] = seqdata['config']['sdatlen']
config['smllen'] = len(list(seqdata['strain'].values())[0])
except KeyError:
pass # some configurations do not have all data, which is fine
config['multigpu'] = multigpu
config['batch_size'] = batch_size
prep('creating model... ')
config, model = create_model(modeltype, config)
drop()
print(model.summary())
gen = batch_gen(seqdata, 'train', config)
#checkpoint = ModelCheckpoint(outdir+'/'+modeltype+'_E{epoch:02d}_TA{acc:.2f}_VA{val_acc:.2f}_VB{val_bleu:}.h5', monitor='val_loss')
checkpoint = ModelCheckpoint(outdir + '/models/' + modeltype +
'_E{epoch:02d}_' + str(timestart) + '.h5')
savehist = HistoryCallback()
savehist.setCatchExit(outdir, modeltype, timestart, config)
valgen = batch_gen(seqdata, 'val', config)
# If you want it to calculate BLEU Score after each epoch use callback_valgen and test_cb
#####
#callback_valgen = batch_gen_train_bleu(seqdata, comvocabsize, 'val', modeltype, batch_size=batch_size)
#test_cb = mycallback(callback_valgen, steps)
#####
callbacks = [checkpoint, savehist]
prep('loading model... ')
model = keras.models.load_model(modelfile, custom_objects={"tf":tf, "keras":keras, "OurCustomGraphLayer":OurCustomGraphLayer, "SeqSelfAttention":SeqSelfAttention})
print(model.summary())
drop()
batch_sets = [allfids[i:i+batchsize] for i in range(0, len(allfids), batchsize)]
refs = list()
preds = list()
predf = open('{}/predictions/{}_{}_{}.tsv'.format(outdir, modeltype, mid, timestart), 'w')
prep("computing predictions...\n")
for c, fid_set in enumerate(batch_sets):
st = timer()
bg = batch_gen(seqdata, testval, config, training=False, firstwords=firstwords)
batch = bg.make_batch(fid_set)
if config['batch_maker'] == 'fw_datsonly':
batch_results = gendescr_datsonly(model, batch, batchsize, config)
elif config['batch_maker'] == 'fw_datsfc':
batch_results = gendescr_datsfc(model, batch, batchsize, config)
elif config['batch_maker'] == 'fw_datspc':
batch_results = gendescr_datspc(model, batch, batchsize, config)
elif config['batch_maker'] == 'fw_datsast':
batch_results = gendescr_datsast(model, batch, batchsize, config)
elif config['batch_maker'] == 'fw_datsastfc':
batch_results = gendescr_datsastfc(model, batch, batchsize, config)
elif config['batch_maker'] == 'fw_datsastpc':
batch_results = gendescr_datsastpc(model, batch, batchsize, config)
elif config['batch_maker'] == 'fw_graphast':
config['multigpu'] = multigpu
config['batch_size'] = batch_size
prep('creating model... ')
config, model = create_model(modeltype, config)
drop()
print(model.summary())
fn = outdir + '/histories/' + modeltype + '_conf_' + datrim + str(
timestart) + '.pkl' #start with saved config
confoutfd = open(fn, 'wb')
pickle.dump(config, confoutfd)
print('saved config to: ' + fn)
if fwfile is not None:
gen = batch_gen(seqdata, 'train', config, firstwords=firstwords)
else:
gen = batch_gen(seqdata, 'train', config)
checkpoint = ModelCheckpoint(outdir + '/models/' + modeltype +
'_E{epoch:02d}_' + datrim + str(timestart) +
'.h5')
if fwfile is not None:
valgen = batch_gen(seqdata, 'val', config, firstwords=firstwords)
else:
valgen = batch_gen(seqdata, 'val', config)
callbacks = [checkpoint]
model.fit_generator(gen,
steps_per_epoch=steps,
epochs=epochs,
verbose=1,