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')
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_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()
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!')