def train_npi(
steplists,
epochs: int = 100,
batch_size: int = 1,
pretrained_encoder_weights: str = f'{config.outdir}/weights/f_enc.weights'
):
state_encoder = StateEncoder().to(config.device)
warm_up: list = list(
filter(
lambda question: 0 <= question['question'][0] < 100 and 0 <=
question['question'][1] < 100, steplists))
if not os.path.exists(pretrained_encoder_weights):
print('start trainning f_enc model')
train_f_enc(warm_up, epochs=epochs, batch_size=batch_size)
else:
state_encoder.load_state_dict(torch.load(pretrained_encoder_weights))
state_encoder.trainable = False
npi = NPI(state_encoder).to(config.device)
optimizer = optim.Adam(npi.parameters(), lr=1e-4, weight_decay=1e-6)
# warm up with single digit addjj
for _ in range(10):
if train_with_plot(npi, optimizer, warm_up, epochs=100):
break
while True:
if train_with_plot(npi,
optimizer,
steplists,
epochs=100,
skip_correct=False):
break
def valid_npi(questions: list, pretrained_encoder_weights: str,
pretrained_npi_weights: str):
state_encoder = StateEncoder().to(config.device)
state_encoder.load_state_dict(torch.load(pretrained_encoder_weights))
npi = NPI(state_encoder, max_depth=20, max_steps=10000).to(config.device)
npi.load_state_dict(torch.load(pretrained_npi_weights))
env = AdditionEnv()
add_program: dict = {'pgid': 2, 'args': []}
wc: int = 0
correct: int = 0
npi.eval().to(config.device)
loop = tqdm(questions, postfix='correct: {correct} wrong: {wrong}')
for addend, augend in loop:
npi.reset()
with torch.no_grad():
env.setup(addend, augend)
# run npi algorithm
npi.step(env, add_program['pgid'], add_program['args'])
if env.result != (addend + augend):
wc += 1
loop.write('{:>5} + {:>5} = {:>5}'.format(addend, augend,
env.result))
else:
correct += 1
loop.set_postfix(correct=correct, wrong=wc)
return correct, wc
def test_question(question: list, npi: NPI) -> int:
env = AdditionEnv()
addend, augend = question
add_program: dict = {'pgid': 2, 'args': []}
npi.reset()
with torch.no_grad():
env.setup(addend, augend)
# run npi algorithm
npi.step(env, add_program['pgid'], add_program['args'])
# get environment observation
return env.result
def evaluate_addition():
"""
Load NPI Model from Checkpoint, and initialize REPL, for interactive carry-addition.
"""
with tf.Session() as sess:
# Load Data
with open(DATA_PATH_TEST, 'rb') as f:
data = pickle.load(f)
# Initialize Addition Core
core = AdditionCore()
# Initialize NPI Model
npi = NPI(core, CONFIG, LOG_PATH)
# Restore from Checkpoint
saver = tf.train.Saver()
saver.restore(sess, CKPT_PATH)
# with gfile.FastGFile("/tmp/tf/log/graph.pb", 'rb') as f:
# graph_def = tf.GraphDef()
# graph_def.ParseFromString(f.read())
# sess.graph.as_default()
# tf.import_graph_def(graph_def)
# print("map variables")
# Run REPL
for x in range(0, 20):
res = ""
# try:
repl(sess, npi, data, x)
# except:
print("--------------------------")
def train_with_plot(npi: NPI,
optimizer,
steplists: list,
epochs: int = 100,
skip_correct: bool = False):
arg_num, arg_depth = config.arg_shape
train_loss: list = []
valid_loss: list = []
# scheduler = optim.lr_scheduler.StepLR(optimizer, step_size=10, gamma=1e-1, last_epoch=-1)
# scheduler = optim.lr_scheduler.ReduceLROnPlateau(optimizer, mode='min', factor=1e-1, patience=2)
for epoch in range(epochs):
npi.train().to(config.device)
# initialize corrent / wrong count
losses: list = []
# np.random.shuffle(steplists)
loop = tqdm(steplists, ncols=100)
loop.write('epoch: {}/{}'.format(epoch + 1, epochs))
for idx, step in enumerate(loop):
question, trace = step['question'], step['trace']
npi.reset()
for env, (pgid, args), (pgid_out, args_out), term_out in trace:
optimizer.zero_grad()
weights = [1] + [1 if 0 <= pgid < 6 else 1e-10] + [
1e-10 if np.argmax(arg) == (arg_depth - 1) else 1
for arg in args_out
]
# get environment observation
term_pred, pgid_pred, args_pred = npi(
tensor(env).flatten().type(torch.FloatTensor).to(
config.device),
tensor(pgid).to(config.device),
tensor(args).flatten().type(torch.FloatTensor).to(
config.device))
total_loss = npi_criteria(
term_pred,
tensor(term_out).type(torch.FloatTensor).to(config.device),
pgid_pred,
tensor(pgid_out).to(config.device), args_pred,
tensor(args_out).type(torch.FloatTensor).to(config.device),
weights)
total_loss.backward()
optimizer.step()
losses.append(total_loss.item())
# total_loss
loop.set_postfix(loss=np.average(losses))
loop.close()
vloss, acc = validate(npi, steplists)
valid_loss.append(vloss)
train_loss.append(np.average(losses))
xlabel = np.array(range(len(train_loss))) + 1
plt.plot(xlabel, train_loss, 'b')
plt.plot(xlabel, valid_loss, 'g')
plt.ylabel('loss')
plt.xlabel('epochs')
plt.savefig(f'{config.outdir}/loss.png')
# scheduler.step()
loop.close()
npi.save()
if acc == 1.:
return True
def validate(npi: NPI, steplists: list, epochs: int = 100):
_, arg_depth = config.arg_shape
env = AdditionEnv()
valid_loss: list = []
correct = wc = 0
npi.eval().to(config.device)
for step in steplists:
question, trace = step['question'], step['trace']
res = test_question(question, npi)
if res == np.sum(question):
correct += 1
else:
wc += 1
npi.reset()
with torch.no_grad():
for env, (pgid, args), (pgid_out, args_out), term_out in trace:
weights = [1] + [1 if 0 <= pgid < 6 else 1e-10] + [
1e-10 if np.argmax(arg) == (arg_depth - 1) else 1
for arg in args_out
]
# get environment observation
term_pred, pgid_pred, args_pred = npi(
tensor(env).flatten().type(torch.FloatTensor).to(
config.device),
tensor(pgid).to(config.device),
tensor(args).flatten().type(torch.FloatTensor).to(
config.device))
total_loss = npi_criteria(
term_pred,
tensor(term_out).type(torch.FloatTensor).to(config.device),
pgid_pred,
tensor(pgid_out).to(config.device), args_pred,
tensor(args_out).type(torch.FloatTensor).to(config.device),
weights)
valid_loss.append(total_loss.item())
return np.average(valid_loss), correct / len(steplists)
def evaluate_addition():
"""
Load NPI Model from Checkpoint, and initialize REPL, for interactive carry-addition.
"""
with tf.Session() as sess:
# Load Data
with open(DATA_PATH_TEST, 'rb') as f:
data = pickle.load(f)
# Initialize Addition Core
core = AdditionCore()
# Initialize NPI Model
npi = NPI(core, CONFIG, LOG_PATH)
# Restore from Checkpoint
saver = tf.train.Saver()
saver.restore(sess, CKPT_PATH)
# with gfile.FastGFile("/tmp/tf/log/graph.pb", 'rb') as f:
# graph_def = tf.GraphDef()
# graph_def.ParseFromString(f.read())
# sess.graph.as_default()
# tf.import_graph_def(graph_def)
# print("map variables")
# Run REPL
predict = {};
predict["ncw"] = 0;
predict["ncr"] = 0;
predict["cw"] = 0;
predict["cr"] = 0;
count = 0;
#f = open('log/numbers.txt', 'r+')
#f.truncate()
limit = EVAL_LIMIT
r = list(range(1000))
# random.shuffle(r)
for x in r:
limit -= 1
if limit > 0:
res = ""
print(repl(sess, npi, data, x, predict))
print("predict_connect_right " + str(predict["cr"]) + " predict_connect_wrong " + str(predict["cw"]) + " predict_not_connect_right " + str(predict["ncr"]) + " predict_not_connect_wrong " + str(predict["ncw"]))
print (str(limit)+"--------------------------")
def __init__(self):
with open("/root/ContextToCode/predictor/log/1class/expect_to_prog", 'r') as handle:
self.sessions = json.load(handle)
for key, value in self.sessions.items():
value['session'] = tf.Session()
# Initialize Addition Core
core = AdditionCore(CONFIG)
# Initialize NPI Model
self.npi = NPI(core, CONFIG, LOG_PATH)
# Restore from Checkpoint
saver = tf.train.Saver()
for key, value in self.sessions.items():
saver.restore(value['session'], CKPT_PATH+value['dir']+"/models/model-0006.ckpt")
print(value)
def __init__(self):
self.sess1 = tf.Session()
self.sess2 = tf.Session()
self.sess3 = tf.Session()
self.sess4 = tf.Session()
self.sess5 = tf.Session()
# Initialize Addition Core
core = AdditionCore(CONFIG)
# Initialize NPI Model
self.npi = NPI(core, CONFIG, LOG_PATH)
# Restore from Checkpoint
saver = tf.train.Saver()
saver.restore(self.sess1, CKPT_PATH_CLASS1)
saver.restore(self.sess5, CKPT_PATH_CLASS5)
saver.restore(self.sess3, CKPT_PATH_CLASS3)
saver.restore(self.sess2, CKPT_PATH_CLASS2)
saver.restore(self.sess4, CKPT_PATH_CLASS4)
def evaluate_multiplication():
"""
Load NPI Model from Checkpoint, and initialize REPL, for interactive multiplication.
"""
with tf.Session() as sess:
# Load Data
with open(TEST_PATH, 'rb') as f:
data = pickle.load(f)
# Initialize Multiplication Core
core = MultiplicationCore()
# Initialize NPI Model
npi = NPI(core, CONFIG, LOG_PATH)
# Restore from Checkpoint
saver = tf.train.Saver()
saver.restore(sess, CKPT_PATH)
# Run REPL
repl(sess, npi, data)
def evaluate_card_pattern_matching():
"""
Load NPI Model from Checkpoint, and initialize REPL, for interactive carry-addition.
"""
# Load Data
with open(TEST_PATH, 'rb') as f:
data = pickle.load(f)
# Initialize Card Pattern Matching Core
print('Initializing Card Pattern Matching Core!')
core = CardPatternMatchingCore()
# Initialize NPI Model
npi = NPI(core, CONFIG, LOG_PATH)
with tf.Session() as sess:
# Restore from Checkpoint
saver = tf.train.Saver()
saver.restore(sess, CKPT_PATH)
# Run REPL
repl(sess, npi, data)
def multiclass_eval():
with open(DATA_PATH_TEST, 'rb') as f:
data = pickle.load(f)
sess1 = tf.Session()
sess2 = tf.Session()
sess3 = tf.Session()
# Initialize Addition Core
core = AdditionCore()
# Initialize NPI Model
npi = NPI(core, CONFIG, LOG_PATH)
# Restore from Checkpoint
saver = tf.train.Saver()
# saver.restore(sess1, CKPT_PATH_CLASS1)
# saver.restore(sess2, CKPT_PATH_CLASS2)
# saver.restore(sess3, CKPT_PATH_CLASS3)
predict = {};
predict["ncw"] = 0;
predict["ncr"] = 0;
predict["cw"] = 0;
predict["cr"] = 0;
f = open('log/prog_produced.txt', 'w+')
f.truncate()
limit = EVAL_LIMIT
r = list(range(1000))
# random.shuffle(r)
for x in r:
limit -= 1
if limit > 0:
with open("/root/ContextToCode/predictor/log/prog_produced.txt", "a") as myfile:
myfile.write(str(repl(sess1, npi, data, x, predict))+"\n")
myfile.write(str(repl(sess2, npi, data, x, predict))+"\n")
myfile.write(str(repl(sess3, npi, data, x, predict))+"\n")
print (str(limit)+"--------------------------")
def evaluate_vqa():
"""
Load NPI Model from Checkpoint, and initialize REPL, for interactive carry-addition.
"""
# Load Data
with open(TEST_PATH, 'rb') as f:
data = pickle.load(f)
data = data[50:100]
# Initialize Addition Core
core = VQAcore()
# Initialize NPI Model
npi = NPI(core, CONFIG, LOG_PATH)
with tf.Session() as sess:
# Restore from Checkpoint
saver = tf.train.Saver()
saver.restore(sess, CKPT_PATH)
# Run REPL
repl(sess, npi, data)
def evaluate_bubblesort():
"""
Load NPI Model from Checkpoint, and initialize REPL, for interactive bubblesort.
"""
with tf.Session() as sess:
# Load Data
with open(TEST_PATH, 'rb') as f:
data = pickle.load(f)
# Initialize Addition Core
print('initializing bubblesort core ...')
core = BubblesortCore()
# Initialize NPI Model
print('initializing npi model ...')
npi = NPI(core, CONFIG, LOG_PATH)
# Restore from Checkpoint
saver = tf.train.Saver()
saver.restore(sess, CKPT_PATH)
# Run REPL
repl(sess, npi, data)
def train_addition(epochs, verbose=0):
"""
Instantiates an Addition Core, NPI, then loads and fits model to data.
:param epochs: Number of epochs to train for.
"""
# Load Data
with open(DATA_PATH, 'r') as f:
data = pickle.load(f)
# Initialize Addition Core
print 'Initializing Addition Core!'
core = AdditionCore()
# Initialize NPI Model
print 'Initializing NPI Model!'
npi = NPI(core, CONFIG, LOG_PATH, verbose=verbose)
# Initialize TF Saver
saver = tf.train.Saver()
# Initialize TF Session
sess = tf.Session()
sess.run(tf.initialize_all_variables())
# Start Training
for ep in range(1, epochs + 1):
for i in range(len(data)):
# Reset NPI States
npi.reset_state()
# Setup Environment
in1, in2, steps = data[i]
scratch = ScratchPad(in1, in2)
x, y = steps[:-1], steps[1:]
# Run through steps, and fit!
step_def_loss, step_arg_loss, term_acc, prog_acc, = 0.0, 0.0, 0.0, 0.0
arg0_acc, arg1_acc, arg2_acc, num_args = 0.0, 0.0, 0.0, 0
for j in range(len(x)):
(prog_name, prog_in_id), arg, term = x[j]
(_, prog_out_id), arg_out, term_out = y[j]
# print(x[j], y[j])
# Update Environment if MOVE or WRITE
if prog_in_id == MOVE_PID or prog_in_id == WRITE_PID:
scratch.execute(prog_in_id, arg)
# Get Environment, Argument Vectors
env_in = [scratch.get_env()]
arg_in, arg_out = [get_args(arg, arg_in=True)
], get_args(arg_out, arg_in=False)
prog_in, prog_out = [[prog_in_id]], [prog_out_id]
term_out = [1] if term_out else [0]
# Fit!
if prog_out_id == MOVE_PID or prog_out_id == WRITE_PID:
loss, t_acc, p_acc, a_acc, _ = sess.run(
[
npi.arg_loss, npi.t_metric, npi.p_metric,
npi.a_metrics, npi.arg_train_op
],
feed_dict={
npi.env_in: env_in,
npi.arg_in: arg_in,
npi.prg_in: prog_in,
npi.y_prog: prog_out,
npi.y_term: term_out,
npi.y_args[0]: [arg_out[0]],
npi.y_args[1]: [arg_out[1]],
npi.y_args[2]: [arg_out[2]]
})
# print({npi.prg_in: prog_in, npi.y_prog: prog_out, npi.y_term: term_out})
# print({npi.y_args[0]: [arg_out[0]], npi.y_args[1]: [arg_out[1]], npi.y_args[2]: [arg_out[2]]})
step_arg_loss += loss
term_acc += t_acc
prog_acc += p_acc
arg0_acc += a_acc[0]
arg1_acc += a_acc[1]
arg2_acc += a_acc[2]
num_args += 1
else:
loss, t_acc, p_acc, _ = sess.run(
[
npi.default_loss, npi.t_metric, npi.p_metric,
npi.default_train_op
],
feed_dict={
npi.env_in: env_in,
npi.arg_in: arg_in,
npi.prg_in: prog_in,
npi.y_prog: prog_out,
npi.y_term: term_out
})
step_def_loss += loss
term_acc += t_acc
prog_acc += p_acc
print "Epoch {0:02d} Step {1:03d} Default Step Loss {2:05f}, " \
"Argument Step Loss {3:05f}, Term: {4:03f}, Prog: {5:03f}, A0: {6:03f}, " \
"A1: {7:03f}, A2: {8:03}"\
.format(ep, i, step_def_loss / len(x), step_arg_loss / len(x), term_acc / len(x),
prog_acc / len(x), arg0_acc / num_args, arg1_acc / num_args,
arg2_acc / num_args)
# Save Model
saver.save(sess, 'tasks/addition/log/model.ckpt')
TRAIN_DATA_PATH = 'tasks/reverse_polish/data/train_8.pik'
with open(TRAIN_DATA_PATH, 'rb', ) as f:
train_data = pickle.load(f)
EVAL_DATA_PATH = 'tasks/reverse_polish/data/eval_8.pik'
with open(EVAL_DATA_PATH, 'rb', ) as f:
eval_data = pickle.load(f)
TEST_DATA_PATH = 'tasks/reverse_polish/data/test_8.pik'
with open(TEST_DATA_PATH, 'rb', ) as f:
test_data = pickle.load(f)
func_core = RevPolishCore().to(device)
npi = NPI(func_core, CONFIG).to(device)
print_net(npi)
optimizer = torch.optim.Adam(npi.parameters(), lr=1e-4)
# optimizer = torch.optim.SGD(npi.parameters(), lr=1e-4)
lr_schedulers = torch.optim.lr_scheduler.CosineAnnealingLR(optimizer, T_max=max_num_epochs)
if not FLG_TEST:
for epoch in range(start_epoch, max_num_epochs+1):
run_epoch(npi, 'train', train_data, writer, os.path.join(exp_dir, 'npi_{}.pth'.format(epoch)))
lr_schedulers.step(epoch)
EVAL_DATA_PATH = 'tasks/reverse_polish/data/eval_8.pik'
with open(
EVAL_DATA_PATH,
'rb',
) as f:
eval_data = pickle.load(f)
TEST_DATA_PATH = 'tasks/reverse_polish/data/test_8.pik'
with open(
TEST_DATA_PATH,
'rb',
) as f:
test_data = pickle.load(f)
func_core = RevPolishCore().to(device)
npi = NPI(func_core, CONFIG).to(device)
print_net(npi)
# optimizer = torch.optim.Adam(npi.parameters(), lr=1e-4)
optimizer = torch.optim.SGD(npi.parameters(), lr=1e-4)
lr_schedulers = torch.optim.lr_scheduler.CosineAnnealingLR(
optimizer, T_max=max_num_epochs)
if not FLG_TEST:
for epoch in range(start_epoch, max_num_epochs + 1):
run_epoch(npi, 'train', train_data, writer,
os.path.join(exp_dir, 'npi_{}.pth'.format(epoch)))
def train_vqa(epochs, verbose=0):
"""
Instantiates a VQA Core, NPI, then loads and fits model to data.
:param epochs: Number of epochs to train for.
"""
# Load Data
with open(DATA_PATH, 'rb') as f:
dataT = pickle.load(f)
data = dataT[:80]
test = dataT[80:100]
with open(DATA_PATH2, 'rb') as f:
dataT2 = pickle.load(f)
test_out = dataT2[:20]
with open(DATA_PATH3, 'rb') as f:
dataT3 = pickle.load(f)
test_out2 = dataT3[:20]
# Initialize VQA Core
print('Initializing VQA Core!')
core = VQAcore()
# Initialize NPI Model
print('Initializing NPI Model!')
npi = NPI(core, CONFIG, LOG_PATH, verbose=verbose)
# Initialize TF Saver
saver = tf.train.Saver()
# Initialize TF Session
with tf.Session() as sess:
sess.run(tf.global_variables_initializer())
# Start Training
removed = {}
errors = {}
# for learning curve
count = 0
# tot_loss_def = []
# tot_loss_arg = []
# test_loss_def = []
# test_loss_arg = []
# test1_loss_def = []
# test1_loss_arg = []
# test2_loss_def = []
# test2_loss_arg = []
test_term_acct = []
test_prog_acct = []
test_arg_acct = []
train_term_acct = []
train_prog_acct = []
train_arg_acct = []
test1_term_acct = []
test1_prog_acct = []
test1_arg_acct = []
test2_term_acct = []
test2_prog_acct = []
test2_arg_acct = []
step = []
for ep in range(1, epochs + 1):
removed[ep] = 0
for i in range(len(data)):
# Reset NPI States
npi.reset_state()
# Setup Environment
_, imgid, qid, qtype, steps = data[i]
scene = Scene(imgid)
x, y = steps[:-1], steps[1:]
if len(x) == 0 or len(y) == 0:
removed[ep] += 1
continue
count += 1
# Run through steps, and fit!
step_def_loss, step_arg_loss, term_acc, prog_acc, = 0.0, 0.0, 0.0, 0.0
arg0_acc, arg1_acc, arg2_acc, num_args = 0.0, 0.0, 0.0, 0
for j in range(len(x)):
(prog_name, prog_in_id), arg, term = x[j]
(_, prog_out_id), arg_out, term_out = y[j]
# Update Environment if MOVE or WRITE
if prog_in_id in EX_PROG_PID:
scene.execute(prog_in_id, arg)
# Get Environment, Argument Vectors
env_in = [scene.get_env()]
# env_in = [np.asarray(list(env_in.values())).transpose().flatten()]
arg_in, arg_out = [get_args(arg, arg_in=True)
], get_args(arg_out, arg_in=False)
prog_in, prog_out = [[prog_in_id]], [prog_out_id]
term_out = [1] if term_out else [0]
# Fit!
if prog_out_id in PAR_PROG_PID:
loss, t_acc, p_acc, a_acc, _ = sess.run(
[
npi.arg_loss, npi.t_metric, npi.p_metric,
npi.a_metrics, npi.arg_train_op
],
feed_dict={
npi.env_in: env_in,
npi.arg_in: arg_in,
npi.prg_in: prog_in,
npi.y_prog: prog_out,
npi.y_term: term_out,
npi.y_args[0]: [arg_out[0]],
npi.y_args[1]: [arg_out[1]],
npi.y_args[2]: [arg_out[2]]
})
step_arg_loss += loss
term_acc += t_acc
prog_acc += p_acc
arg0_acc += a_acc[0]
arg1_acc += a_acc[1]
arg2_acc += a_acc[2]
num_args += 1
else:
loss, t_acc, p_acc, _ = sess.run(
[
npi.default_loss, npi.t_metric, npi.p_metric,
npi.default_train_op
],
feed_dict={
npi.env_in: env_in,
npi.arg_in: arg_in,
npi.prg_in: prog_in,
npi.y_prog: prog_out,
npi.y_term: term_out
})
step_def_loss += loss
term_acc += t_acc
prog_acc += p_acc
try:
print ("Epoch {0:02d} Step {1:03d} Default Step Loss {2:05f}, " \
"Argument Step Loss {3:05f}, Term: {4:03f}, Prog: {5:03f}, A0: {6:03f}, " \
"A1: {7:03f}, A2: {8:03}".format(ep, i, step_def_loss / len(x), step_arg_loss / len(x), term_acc / len(x),
prog_acc / len(x), arg0_acc / num_args, arg1_acc / num_args,
arg2_acc / num_args))
if count % 10 == 0:
# Save Model
tmp = stat.mean([
arg0_acc / num_args, arg1_acc / num_args,
arg2_acc / num_args
])
saver.save(sess, 'tasks/vqa/log/model.ckpt')
train_arg_acct.append(tmp / len(x))
train_prog_acct.append(prog_acc / len(x))
train_term_acct.append(term_acc / len(x))
step.append(count)
a, b, c = test_vqa(test, npi, core, sess)
test_arg_acct.append(c)
test_prog_acct.append(b)
test_term_acct.append(a)
a, b, c = test_vqa(test_out, npi, core, sess)
test1_arg_acct.append(c)
test1_prog_acct.append(b)
test1_term_acct.append(a)
a, b, c = test_vqa(test_out2, npi, core, sess)
test2_arg_acct.append(c)
test2_prog_acct.append(b)
test2_term_acct.append(a)
except:
print('main print failed')
# Save Model
saver.save(sess, 'tasks/vqa/log/model.ckpt')
# print learning curve
print('train term,prog,arg: ', test_term_acct[-1], test_prog_acct[-1],
test_arg_acct[-1])
print('test_inside term,prog,arg: ', test_term_acct[-1],
test_prog_acct[-1], test_arg_acct[-1])
print('test_out term,prog,arg: ', test1_term_acct[-1],
test1_prog_acct[-1], test1_arg_acct[-1])
print('test_out2 term,prog,arg: ', test2_term_acct[-1],
test2_prog_acct[-1], test2_arg_acct[-1])
plt.figure(figsize=(20, 5))
plt.plot(step, train_term_acct, 'b', label='train_query_term')
plt.plot(step, test_term_acct, 'm', label='test_query_term')
plt.plot(step, test1_term_acct, 'c', label='test_count_term')
plt.plot(step, test2_term_acct, 'k', label='test_exist_term')
plt.legend()
plt.xticks(step)
plt.xlabel('step')
plt.ylabel('acc')
plt.title('learning curve for termination')
plt.savefig(SAVE_PATH + 'acc_query_term')
plt.close()
plt.figure(figsize=(20, 5))
plt.plot(step, train_prog_acct, 'b', label='train_query_prog')
plt.plot(step, test_prog_acct, 'm', label='test_query_prog')
plt.plot(step, test1_prog_acct, 'c', label='test_count_prog')
plt.plot(step, test2_prog_acct, 'k', label='test_exist_prog')
plt.legend()
plt.xticks(step)
plt.xlabel('step')
plt.ylabel('acc')
plt.title('learning curve for program')
plt.savefig(SAVE_PATH + 'acc_query_prog')
plt.close()
plt.figure(figsize=(20, 5))
plt.plot(step, train_arg_acct, 'b', label='train_query_arg')
plt.plot(step, test_arg_acct, 'm', label='test_query_arg')
plt.plot(step, test1_arg_acct, 'c', label='test_count_arg')
plt.plot(step, test2_arg_acct, 'k', label='test_exist_arg')
plt.legend()
plt.xticks(step)
plt.xlabel('step')
plt.ylabel('acc')
plt.title('learning curve for arguments')
plt.savefig(SAVE_PATH + 'acc_query_arg')
plt.close()
def train_addition(epochs, verbose=0):
"""
Instantiates an Addition Core, NPI, then loads and fits model to data.
:param epochs: Number of epochs to train for.
"""
# Load Data
with open(DATA_PATH_TRAIN, 'rb') as f:
data = pickle.load(f)
# Initialize Addition Core
print ('Initializing Addition Core!')
core = AdditionCore()
# Initialize NPI Model
print ('Initializing NPI Model!')
npi = NPI(core, CONFIG, LOG_PATH, verbose=verbose)
# Initialize TF Saver
saver = tf.train.Saver()
# Initialize TF Session
sess = tf.Session()
sess.run(tf.global_variables_initializer())
# Start Training
for ep in range(1, epochs + 1):
for i in range(len(data)):
# Reset NPI States
npi.reset_state()
# Setup Environment
steps = data[i]
# print(data[i])
x, y = steps[:-1], steps[1:]
# Run through steps, and fit!
step_def_loss, step_arg_loss, term_acc, prog_acc, = 0.0, 0.0, 0.0, 0.0
arg0_acc, arg1_acc, arg2_acc, num_args = 0.0, 0.0, 0.0, 0
# dsl = DSL([], [])
for j in range(len(x)):
# {'program': {'program': 'check'}, 'environment': {'terminate': False, 'answer': 1, 'is_redirect': 2},'args': {'id': 0}}
# print(y[j])
prog_name, prog_in_id, arg, term = x[j]["program"]["program"], x[j]["program"]["id"], x[j]["args"]["id"], x[j]["environment"]["terminate"]
prog_name_out, prog_out_id, arg_out, term_out = y[j]["program"]["program"], y[j]["program"]["id"], y[j]["args"]["id"], y[j]["environment"]["terminate"]
# Get Environment, Argument Vectors
env_in = [get_env(x[j]["environment"])]
arg_in, arg_out = [get_args(arg, arg_in=True)], get_args(arg_out, arg_in=False)
term_out = [1] if term_out else [0]
# if prog_name_out=="WRITE":
# prog_out_id = dsl.get_code(y[j]["prog"]["arg"][1])
# os._exit()
prog_in, prog_out = [[prog_in_id]], [prog_out_id]
# Fit!
if True:
t_acc, p_acc, _, loss = sess.run(
[npi.t_metric, npi.p_metric, npi.default_train_op, npi.default_loss],
feed_dict={npi.env_in: env_in, npi.arg_in: arg_in, npi.prg_in: prog_in,
npi.y_prog: prog_out, npi.y_term: term_out})
# print({npi.env_in: env_in, npi.arg_in: arg_in, npi.prg_in: prog_in, npi.y_prog: prog_out, npi.y_term: term_out})
# print({npi.y_args[0]: [arg_out[0]], npi.y_args[1]: [arg_out[1]], npi.y_args[2]: [arg_out[2]]})
# step_arg_loss += loss
term_acc += t_acc
prog_acc += p_acc
step_def_loss += loss
# arg0_acc += a_acc[0]
# arg1_acc += a_acc[1]
# arg2_acc += a_acc[2]
# num_args += 1
# else:
# loss, t_acc, p_acc, _ = sess.run(
# [npi.default_loss, npi.t_metric, npi.p_metric, npi.default_train_op],
# feed_dict={npi.env_in: env_in, npi.arg_in: arg_in, npi.prg_in: prog_in,
# npi.y_prog: prog_out, npi.y_term: term_out})
# step_def_loss += loss
# term_acc += t_acc
# prog_acc += p_acc
print ("Epoch {0:02d} Step {1:03d} Loss: {2:03f} Term: {3:03f}, Prog: {4:03f}" \
.format(ep, i, step_def_loss / len(x), term_acc / len(x), prog_acc / len(x)))
# Save Model
saver.save(sess, CKPT_PATH)
# !!!!
tf.train.write_graph(sess.graph_def, '/tmp/tf/log', 'graph.pb', as_text=False)
# tf.train.write_graph(my_graph, path_to_model_pb, 'saved_model.pb', as_text=False)
# !!!!
def train_card_pattern_matching(epochs, verbose=0):
# Load Data
with open(DATA_PATH, 'rb') as f:
data = pickle.load(f)
# Initialize Card Pattern Matching Core
print('Initializing Card Pattern Matching Core!')
core = CardPatternMatchingCore()
# Initialize NPI Model
print('Initializing NPI Model!')
npi = NPI(core, CONFIG, LOG_PATH, verbose=verbose)
# Initialize TF Saver
saver = tf.train.Saver()
# Initialize TF Session
with tf.Session() as sess:
sess.run(tf.global_variables_initializer())
# Start Training
for ep in range(1, epochs + 1):
for i in range(len(data)):
# Reset NPI States
npi.reset_state()
# Setup Environment
card1, card2, steps = data[i]
scratch = ScratchPad(card1, card2)
x, y = steps[:-1], steps[1:]
# Run through steps, and fit!
step_def_loss, step_arg_loss, term_acc, prog_acc, = 0.0, 0.0, 0.0, 0.0
arg0_acc, arg1_acc, arg2_acc, num_args = 0.0, 0.0, 0.0, 0
for j in range(len(x)):
(prog_name, prog_in_id), arg, term = x[j]
(_, prog_out_id), arg_out, term_out = y[j]
# Update Environment if MOVE or WRITE
if prog_in_id == MOVE_PID or prog_in_id == WRITE_PID:
scratch.execute(prog_in_id, arg)
# Get Environment, Argument Vectors
env_in = [scratch.get_env()]
arg_in, arg_out = [get_args(arg, arg_in=True)
], get_args(arg_out, arg_in=False)
prog_in, prog_out = [[prog_in_id]], [prog_out_id]
term_out = [1] if term_out else [0]
# Fit!
if prog_out_id == MOVE_PID or prog_out_id == WRITE_PID:
loss, t_acc, p_acc, a_acc, _ = sess.run(
[
npi.arg_loss, npi.t_metric, npi.p_metric,
npi.a_metrics, npi.arg_train_op
],
feed_dict={
npi.env_in: env_in,
npi.arg_in: arg_in,
npi.prg_in: prog_in,
npi.y_prog: prog_out,
npi.y_term: term_out,
npi.y_args[0]: [arg_out[0]],
npi.y_args[1]: [arg_out[1]],
npi.y_args[2]: [arg_out[2]]
})
step_arg_loss += loss
term_acc += t_acc
prog_acc += p_acc
arg0_acc += a_acc[0]
arg1_acc += a_acc[1]
arg2_acc += a_acc[2]
num_args += 1
else:
loss, t_acc, p_acc, _ = sess.run(
[
npi.default_loss, npi.t_metric, npi.p_metric,
npi.default_train_op
],
feed_dict={
npi.env_in: env_in,
npi.arg_in: arg_in,
npi.prg_in: prog_in,
npi.y_prog: prog_out,
npi.y_term: term_out
})
step_def_loss += loss
term_acc += t_acc
prog_acc += p_acc
print ("Epoch {0:02d} Step {1:03d} Default Step Loss {2:05f}, " \
"Argument Step Loss {3:05f}, Term: {4:03f}, Prog: {5:03f}, A0: {6:03f}, " \
"A1: {7:03f}, A2: {8:03}".format(ep, i, step_def_loss / len(x), step_arg_loss / len(x), term_acc / len(x),
prog_acc / len(x), arg0_acc / num_args, arg1_acc / num_args,
arg2_acc / num_args))
# Save Model
saver.save(
sess,
'tasks/card_pattern_matching/log/card_pattern_matching_model.ckpt'
)
print('Model generation complete!')
def train_addition(epochs, start_epoch, start_step, sub, verbose=0):
"""
Instantiates an Addition Core, NPI, then loads and fits model to data.
:param epochs: Number of epochs to train for.
"""
if not os.path.exists("log/" + sub + "/models/"):
os.mkdir("log/" + sub + "/models/")
# Load Data
with open(DATA_PATH_TRAIN, 'rb') as f:
data = pickle.load(f)
# f = open('log/log_train.txt', 'r+')
#f.truncate()
# Initialize Addition Core
print('Initializing Addition Core!')
core = AdditionCore()
# Initialize NPI Model
print('Initializing NPI Model!')
npi = NPI(core, CONFIG, LOG_PATH, verbose=verbose)
# Initialize TF Saver
saver = tf.train.Saver()
# Initialize TF Session
sess = tf.Session()
if start_epoch > 0 or start_step > 0:
saver = tf.train.Saver()
if start_step > 0:
saver.restore(
sess,
"log/" + sub + "/models/model-{0:04d}-{1:06d}.ckpt".format(
start_epoch + 1, start_step))
print("log/" + sub + "/models/model-{0:04d}-{1:06d}.ckpt".format(
start_epoch + 1, start_step))
else:
saver.restore(
sess, "log/" + sub +
"/models/model-{0:04d}.ckpt".format(start_epoch))
print("log/" + sub +
"/models/model-{0:04d}.ckpt".format(start_epoch))
else:
sess.run(tf.global_variables_initializer())
# Reset NPI States
npi.reset_state()
tf.get_default_graph().finalize()
# Start Training
for ep in range(start_epoch + 1, epochs + 1):
sum = 0
for i in range(len(data)):
if i > start_step:
# Setup Environment
steps = data[i]
# print(data[i])
x, y = steps[:-1], steps[1:]
# Run through steps, and fit!
step_def_loss, step_arg_loss, term_acc, prog_acc, = 0.0, 0.0, 0.0, 0.0
arg0_acc, arg1_acc, arg2_acc, num_args = 0.0, 0.0, 0.0, 0
# dsl = DSL([], [])
if len(x) > 0:
for j in range(len(x)):
# {'program': {'program': 'check'}, 'environment': {'terminate': False, 'answer': 1, 'is_redirect': 2},'args': {'id': 0}}
#print(x[j]['addinfo'])
prog_name, prog_in_id, arg, term = x[j]["program"][
"program"], x[j]["program"]["id"], x[j]["args"][
"id"], x[j]["environment"]["terminate"]
prog_name_out, prog_out_id, arg_out, term_out = y[j][
"program"]["program"], y[j]["program"]["id"], y[j][
"args"]["id"], y[j]["environment"]["terminate"]
# Get Environment, Argument Vectors
env_in = [get_env(x[j]["environment"])]
arg_in, arg_out = [get_args(arg, arg_in=True)
], get_args(arg_out, arg_in=False)
term_out = [1] if term_out else [0]
# if prog_name_out=="WRITE":
# prog_out_id = dsl.get_code(y[j]["prog"]["arg"][1])
# os._exit()
prog_in, prog_out = [[prog_in_id]], [prog_out_id]
# Fit!
if True:
t_acc, p_acc, _, loss = sess.run(
[
npi.t_metric, npi.p_metric,
npi.default_train_op, npi.default_loss
],
feed_dict={
npi.env_in: env_in,
npi.arg_in: arg_in,
npi.prg_in: prog_in,
npi.y_prog: prog_out,
npi.y_term: term_out
})
# print({npi.env_in: env_in, npi.arg_in: arg_in, npi.prg_in: prog_in, npi.y_prog: prog_out, npi.y_term: term_out})
# print({npi.y_args[0]: [arg_out[0]], npi.y_args[1]: [arg_out[1]], npi.y_args[2]: [arg_out[2]]})
# step_arg_loss += loss
term_acc += t_acc
prog_acc += p_acc
step_def_loss += loss
# arg0_acc += a_acc[0]
# arg1_acc += a_acc[1]
# arg2_acc += a_acc[2]
# num_args += 1
# else:
# loss, t_acc, p_acc, _ = sess.run(
# [npi.default_loss, npi.t_metric, npi.p_metric, npi.default_train_op],
# feed_dict={npi.env_in: env_in, npi.arg_in: arg_in, npi.prg_in: prog_in,
# npi.y_prog: prog_out, npi.y_term: term_out})
# step_def_loss += loss
# term_acc += t_acc
# prog_acc += p_acc
sum += prog_acc / len(x)
#with open('log/log_train.txt', "a") as myfile:
if i % 1000 == 0:
message = "Epoch "+sub+" {0:02d} Step {1:03d} Loss: {2:03f} Term: {3:03f}, Prog: {4:03f} AVG: {5:03f}" \
.format(ep, i, step_def_loss / len(x), term_acc / len(x), prog_acc / len(x), sum / (i - start_step))
print(message)
with open("log/" + sub + "/info", "a") as myfile:
myfile.write(message + "\n")
if i % chunk == 0:
saver.save(
sess, "log/" + sub +
"/models/model-{0:04d}-{1:06d}.ckpt".format(ep, i))
#if os.path.exists("log/model-{0:04d}-{1:06d}.ckpt.meta".format(ep, i-chunk)):
# os.remove("log/model-{0:04d}-{1:06d}.ckpt.meta".format(ep, i-chunk))
# os.remove("log/model-{0:04d}-{1:06d}.ckpt.index".format(ep, i-chunk))
# os.remove("log/model-{0:04d}-{1:06d}.ckpt.data-00000-of-00001".format(ep, i-chunk))
print ("Epoch {0:02d} Step {1:03d} AVG: {2:03f}" \
.format(ep, len(data), sum / len(data)))
# Save Model
saver.save(sess,
"log/" + sub + "/models/model-{0:04d}.ckpt".format(ep))
# !!!!
tf.train.write_graph(sess.graph_def,
'/tmp/tf/log',
'graph.pb',
as_text=False)
def test_vqa(verbose=0):
"""
Instantiates a VQA Core, NPI, then loads and fits model to data.
:param epochs: Number of epochs to train for.
"""
# Load Data
with open(DATA_PATH, 'rb') as f:
data = pickle.load(f)
data = data[80:300] # the default model saved has been trained on same data but [:80]
# Initialize VQA Core
print ('Initializing VQA Core!')
core = VQAcore()
# Initialize NPI Model
print ('Initializing NPI Model!')
npi = NPI(core, CONFIG, LOG_PATH, verbose=verbose)
with tf.Session() as sess:
# Restore from Checkpoint
saver = tf.train.Saver()
saver.restore(sess, CKPT_PATH)
term_acct = []
prog_acct = []
arg_acct = []
step = []
count = 0
# Start Testing
for i in range(len(data)):
# Reset NPI States
npi.reset_state()
# Setup Environment
_,imgid, qid, qtype, steps = data[i]
scene = Scene(imgid)
x, y = steps[:-1], steps[1:]
if len(x) == 0 or len(y) == 0:
continue
count += 1
# Run through steps, and fit!
step_def_loss, step_arg_loss, term_acc, prog_acc, = 0.0, 0.0, 0.0, 0.0
arg0_acc, arg1_acc, arg2_acc, num_args = 0.0, 0.0, 0.0, 0
for j in range(len(x)):
if random.uniform(0,1) > ABAL_THR:
(prog_name, prog_in_id), arg, term = y[j]
(_, prog_out_id), arg_out, term_out = x[j]
else:
(prog_name, prog_in_id), arg, term = x[j]
(_, prog_out_id), arg_out, term_out = y[j]
# Update Environment if MOVE or WRITE
if prog_in_id in EX_PROG_PID:
scene.execute(prog_in_id, arg)
# Get Environment, Argument Vectors
env_in = [scene.get_env()]
arg_in, arg_out = [get_args(arg, arg_in=True)], get_args(arg_out, arg_in=False)
prog_in, prog_out = [[prog_in_id]], [prog_out_id]
term_out = [1] if term_out else [0]
# Fit!
if prog_out_id in PAR_PROG_PID:
loss, t_acc, p_acc, a_acc= sess.run(
[npi.arg_loss, npi.t_metric, npi.p_metric, npi.a_metrics],
feed_dict={npi.env_in: env_in, npi.arg_in: arg_in, npi.prg_in: prog_in,
npi.y_prog: prog_out, npi.y_term: term_out,
npi.y_args[0]: [arg_out[0]], npi.y_args[1]: [arg_out[1]],
npi.y_args[2]: [arg_out[2]]})
step_arg_loss += loss
term_acc += t_acc
prog_acc += p_acc
arg0_acc += a_acc[0]
arg1_acc += a_acc[1]
arg2_acc += a_acc[2]
num_args += 1
else:
loss, t_acc, p_acc= sess.run(
[npi.default_loss, npi.t_metric, npi.p_metric],
feed_dict={npi.env_in: env_in, npi.arg_in: arg_in, npi.prg_in: prog_in,
npi.y_prog: prog_out, npi.y_term: term_out})
step_def_loss += loss
term_acc += t_acc
prog_acc += p_acc
try:
print ("Step {} Default Step Loss {}, " \
"Argument Step Loss {}, Term: {}, Prog: {}, A0: {}, " \
"A1: {}, A2: {}".format(i, step_def_loss / len(x), step_arg_loss / len(x), term_acc / len(x),
prog_acc / len(x), arg0_acc / num_args, arg1_acc / num_args,
arg2_acc / num_args))
tmp = stat.mean([arg0_acc / num_args, arg1_acc / num_args, arg2_acc / num_args])
term_acct.append(term_acc / len(x))
prog_acct.append(prog_acc / len(x))
arg_acct.append(tmp)
step.append(count)
except:
print('main print failed')
plt.figure(figsize=(20, 5))
plt.plot(step, term_acct, 'b', label='term')
plt.plot(step, prog_acct, 'm', label='prog')
plt.plot(step, arg_acct, 'c', label='arg')
plt.legend()
plt.xticks(step)
plt.xlabel('step')
plt.ylabel('acc')
plt.title('Ablation Study')
plt.savefig(SAVE_PATH + 'acc_ablation')
plt.close()