def __init__(self, verbosity=True, latent_dim=100):
img_shape = (128, 128, 3)
the_disc = Discriminator()
the_gen = Generator()
self.discriminator = the_disc.define_discriminator(
verb=verbosity, sample_shape=img_shape)
self.generator = the_gen.define_generator(verb=verbosity,
sample_shape=img_shape,
latent_dim=latent_dim)
self.discriminator.trainable = False
optimizer = Adam(0.0002, 0.5)
self.discriminator.compile(
loss=['binary_crossentropy', 'categorical_crossentropy'],
loss_weights=[0.5, 0.5],
optimizer=optimizer,
metrics=['accuracy'])
noise = Input(shape=(latent_dim, ))
img = self.generator(noise)
valid, _ = self.discriminator(img)
self.combined = Model(noise, valid)
self.combined.compile(loss=['binary_crossentropy'],
optimizer=optimizer)
def __init__(self, _device):
self.device = _device
self.batch_size = 64
self.resolution = 28
self.d_criterion = None
self.d_optimizer = None
self.g_criterion = None
self.g_optimizer = None
self.discriminator = Discriminator(num_layers=5,
activations=["relu", "relu", "relu",
"sigmoid"],
device=_device,
num_nodes=[1, 64, 128, 64, 1],
kernels=[5, 5, 3],
strides=[2, 2, 2],
dropouts=[.25, .25, 0],
batch_size=64)
# pass one image through the network so as to initialize the output
# layer
self.discriminator(torch.rand(
size=[self.batch_size, 1, self.resolution, self.resolution]))
self.generator = Generator(num_layers=6,
activations=["relu", "relu", "relu", "relu",
"tanh"],
num_nodes=[1, 64, 128, 64, 64, 1],
kernels=[3, 3, 3, 3],
strides=[1, 1, 1, 1],
batch_norms=[1, 1, 1, 0],
upsamples=[1, 1, 0, 0],
dropouts=[.25, .25, 0])
def main():
# Loading Parameters
parser = init_parameters()
args, _ = parser.parse_known_args()
# Updating Parameters (cmd > yaml > default)
args = update_parameters(parser, args)
# Setting save_dir
save_dir = get_save_dir(args)
U.set_logging(save_dir)
with open('{}/config.yaml'.format(save_dir), 'w') as f:
yaml.dump(vars(args), f)
# Processing
if args.generate_data or args.generate_label:
g = Generator(args)
g.start()
elif args.extract or args.visualization:
if args.extract:
p = Processor(args, save_dir)
p.extract()
if args.visualization:
v = Visualizer(args)
v.start()
else:
p = Processor(args, save_dir)
p.start()
def main():
arg_parser = argparse.ArgumentParser()
arg_parser.add_argument('read_path')
arg_parser.add_argument('write_path')
args = arg_parser.parse_args()
with open(args.read_path, 'r') as source:
text = source.read()
lexer = Lexer(text)
tokens = lexer.lex()
parser = Parser(tokens)
ast = parser.parse()
symbolizer = Symbolizer(ast)
symbolizer.symbolize()
optimizer = Optimizer(ast)
optimizer.optimize()
grapher = Grapher(ast)
grapher.graph()
generator = Generator(ast)
generator.generate(args.write_path)
runner = Runner(ast)
runner.run()
def __init__(self, data, generator=None):
data_indicator = data.iloc[:, range(9, len(data.columns), 1)]
if generator is None:
self._generator = Generator(data_indicator)
else:
self._generator = generator
self._data = data
def test_generator(self):
for path in glob.glob("test/grader/*/src.pas"):
dir = os.path.dirname(path)
should_fail = not dir.endswith('16')
with open(path, 'r') as source:
print(f"testing {path}")
text = source.read()
lexer = Lexer(text)
tokens = lexer.lex()
parser = Parser(tokens)
ast = parser.parse()
symbolizer = Symbolizer(ast)
symbolizer.symbolize()
grapher = Generator(ast, symbolizer)
grapher.generate()
sol = os.path.join(dir, 'src.c')
out = os.path.join(dir, 'out')
if os.path.exists(sol):
os.remove(sol)
if os.path.exists(out):
os.remove(out)
grapher.write(sol)
p = None
try:
p = sp.Popen(['gcc', sol, '-o', out], stdout=sp.PIPE)
retCode = p.wait()
self.assertTrue(retCode == 0)
p.stdout.close()
#s = str(p.stdout.read())
#self.assertTrue(s == '')
except Exception:
self.assertFalse(should_fail)
for i in range(1, 5):
inFile = os.path.join(dir, str(i) + '.in')
outFile = os.path.join(dir, str(i) + '.out')
with open(inFile, 'r') as inText:
with open(outFile, 'r') as outText:
inText = inText.read()
outText = outText.read()
try:
of = sp.Popen([out],
stdin=sp.PIPE,
stdout=sp.PIPE)
of.stdin.write(inText.encode('utf-8'))
of.stdin.close()
rc = of.wait()
self.assertTrue(rc == 0)
b = of.stdout.read()
s = b.decode('utf-8')
of.stdout.close()
if (not should_fail):
self.assertEqual(s, str(outText))
except Exception:
self.assertFalse(should_fail)
self.assertTrue(True)
#Tests().test_grapher()
def __collect_data(bca_inputs='bee_colony_inputs.txt', generator_inputs='settings.txt'):
settings = input_output.read(bca_inputs)
data = Generator(generator_inputs)
intelligence = Engine(data)
scouts = [Bee(intelligence) for _ in range(settings[0])]
workers = [Bee(intelligence) for _ in range(settings[1])]
steps = settings[2]
hive = []
return steps, scouts, workers, hive
def test_generator(self):
g = Generator()
w1 = Website('test1', num_pages=20)
w2 = Website('test2', num_pages=20)
g.generate(sites=[w1, w2], writeFile=True)
self.assertEqual(len(g.sites), 2)
self.assertEqual(len(g.sites[0].pages), 20)
for i, page in enumerate(g.sites[0].pages):
self.assertTrue(
f'<title>{w1.domain} - Page{i}</title>' in page.html)
def setUp(self):
self.batch_size = 32
self.generator = Generator(
csv_file=f"{gc['DATA_DIR']}/src/train.csv",
images_src_dir=f"{gc['DATA_DIR']}/src/train_images",
batch_size=self.batch_size,
target_image_size=(224, 224),
image_augmentation_options=default_image_augmenation_options,
cache_dir=gc["DATA_DIR"] + "/images_cache",
)
def createNew(self, species, length, gamutLength, cf):
if species == cf:
cf = ()
cf = tuple(cf)
key = (species, length, gamutLength, cf)
if key in self.generators:
return self.generators[key].createNew()
else:
if species == 'cf':
g = Generator(cantusSpec(length, gamutLength, 'cf'))
elif species == 's1':
g = Generator(firstSpeciesSpec([ConstPitch(x) for x in cf], gamutLength, 's1'))
elif species == 's2':
g = Generator(secondSpeciesSpec([ConstPitch(x) for x in cf], gamutLength, 's2'))
else:
g = Generator(thirdSpeciesSpec([ConstPitch(x) for x in cf], gamutLength, 's3'))
self.generators[key] = g
return g.createNew()
def main():
args = parse_args()
device = torch.device("cuda")
generator = Generator().to(device)
if args.freeze:
for name, param in generator.named_parameters():
if ("shared" not in name) and ("decoder.block.5" not in name):
param.requires_grad = False
train_dataset = DailyDialogueDataset(
path_join(args.dataset_path, "train/dialogues_train.txt"),
tokenizer=generator.tokenizer,
)
valid_dataset = DailyDialogueDataset(
path_join(args.dataset_path, "validation/dialogues_validation.txt"),
tokenizer=generator.tokenizer,
)
print(len(train_dataset), len(valid_dataset))
optimizer = AdamW(generator.parameters(), lr=args.lr)
best_loss = np.float("inf")
for epoch in tqdm(range(args.num_epochs)):
train_loss, valid_loss = [], []
generator.train()
for ind in np.random.permutation(len(train_dataset)):
optimizer.zero_grad()
context, reply = train_dataset.sample_dialouge(ind)
context, reply = context.to(device), reply.to(device)
loss = generator.get_loss(input_ids=context, labels=reply)
loss.backward()
optimizer.step()
train_loss.append(loss.item())
generator.eval()
for ind in range(len(valid_dataset)):
context, reply = valid_dataset[ind]
context, reply = context.to(device), reply.to(device)
with torch.no_grad():
loss = generator.get_loss(input_ids=context, labels=reply)
valid_loss.append(loss.item())
train_loss, valid_loss = np.mean(train_loss), np.mean(valid_loss)
print(
f"Epoch {epoch + 1}, Train Loss: {train_loss:.2f}, Valid Loss: {valid_loss:.2f}"
)
if valid_loss < best_loss:
best_loss = valid_loss
torch.save(generator.state_dict(), args.output_path)
def __init__(self, window_height, window_width, button_height, screen):
# Window parameters
self.window_height = window_height
self.window_width = window_width
self.button_height = button_height
self.screen = screen
self.buttons = None
self.clock = pg.time.Clock()
# Maze parameters
self.maze_size = 10
self.scale = (self.window_width / self.maze_size)
self.maze = Maze(self.maze_size, self.scale)
self.maze_list = [self.maze.grid]
self.build_maze = copy.deepcopy(self.maze.grid)
# Initialize searcher, generator, learner objects
self.generator = Generator() # Used to generate mazes
self.searcher = Searcher() # Used to search mazes
#self.learner = Learner() # Used to simulate learning on mazes
# Paths searched
self.paths = {}
self.shown_path = {}
# Number of cells visited in each search
self.visits = {"BFS": "-", "A*": "-", "DFS": "-"}
# Length of the path found from the start to the end
self.lengths = {"BFS": "-", "A*": "-", "DFS": "-"}
# Reinforcement Learning parameters
self.num_actions = 4
self.num_states = self.maze.size * self.maze.size
self.step_size = 0.5
self.epsilon = 0.1
self.discount = 0.9
self.num_episodes = 50
self.num_runs = 5
self.all_states_visited = {}
self.all_state_visits = {}
self.all_reward_sums = {}
# Colors
self.path_colors = {
"BFS": RED,
"A*": ORANGE,
"DFS": PINK,
"Explore_BFS": DARK_GREEN,
"Explore_A*": DARK_ORANGE,
"Explore_DFS": PURPLE
}
def main():
global window, ceilingtexture, floortexture, walltexture, map
glutInit(sys.argv)
glutInitDisplayMode(GLUT_RGBA | GLUT_DOUBLE | GLUT_DEPTH)
glutInitWindowSize(640, 480)
glutInitWindowPosition(200, 200)
window = glutCreateWindow('Experimental Maze')
# Generate map.
generator = Generator()
map = generator.generateMap(16)
# Represents a top-down view of the maze.
# map = [
# [1, 1, 0, 0, 0, 1, 1, 1, 1, 1, 1],
# [1, 1, 0, 0, 0, 1, 1, 0, 1, 0, 1],
# [1, 1, 0, 0, 0, 1, 1, 0, 0, 0, 1],
# [1, 1, 1, 0, 0, 1, 1, 0, 0, 0, 1],
# [1, 1, 1, 0, 0, 0, 0, 0, 0, 0, 1],
# [1, 1, 0, 0, 0, 1, 1, 0, 0, 1, 1],
# [1, 1, 0, 0, 0, 1, 1, 0, 0, 1, 1],
# [1, 1, 0, 0, 0, 1, 1, 0, 0, 0, 1],
# [1, 1, 1, 1, 1, 1, 1, 0, 0, 0, 1],
# [1, 1, 1, 1, 1, 1, 1, 0, 0, 0, 1],
# [1, 1, 0, 0, 0, 0, 0, 0, 1, 0, 1],
# [1, 1, 0, 0, 0, 0, 0, 0, 0, 0, 1],
# [1, 1, 0, 0, 0, 0, 0, 0, 0, 0, 1],
# [1, 1, 1, 0, 0, 0, 1, 1, 1, 1, 1],
# [1, 1, 1, 0, 0, 0, 1, 1, 1, 1, 1],
# [1, 1, 1, 0, 0, 0, 1, 1, 1, 1, 1]
# ]
# Load texture.
texture = Texture()
ceilingtexture = texture.loadImage('tex/ceiling.bmp')
floortexture = texture.loadImage('tex/floor.bmp')
walltexture = texture.loadImage('tex/wall.bmp')
glutIgnoreKeyRepeat(1)
glutKeyboardFunc(input.registerKeyDown)
glutKeyboardUpFunc(input.registerKeyUp)
glutDisplayFunc(drawScene)
glutIdleFunc(drawScene)
initGL(640, 480)
glutMainLoop()
def __init__(self):
self.logging = True
# Game logic
self.players = {}
self.bots = {}
self.botsInactive = {}
self.bullets = {}
self.bulletsIndex = 0
self.geysers = {}
self.geysersInactive = {}
self.geysersIndex = 0
self.doors = {}
self.doorsIndex = 0
self.fortified = {}
self.fortifyIndex = 0
# Game managment
self.titlesToUpdate = []
self.deadConnections = []
self.mapDimensions = MAPDIMENSIONS
gen = Generator(self.mapDimensions[0], self.mapDimensions[1], -200)
self.map = [gen.getUnderworld(), gen.getOverworld()]
self.bouldersMap = gen.getBoulders()
# Create initial boulders
for floor in [0, 1]:
for boulder in self.bouldersMap[floor]:
boulderx = boulder[0]
bouldery = boulder[1]
self.map[floor][bouldery][boulderx] = BOULDER_CHAR
# Create initial bots
botId = 0
for x in range(0, round(BOTAMOUNT)):
if x % 2 == 0: # 1/2 of bot amount in overworld
self.botsInactive[botId] = Bot(self, self.getSpawnPosition(),
1)
botId += 1
self.botsInactive[botId] = Bot(self,
self.getSpawnPosition(floor=0), 0)
botId += 1
def __init__(self, master_key, clipboard, keys_and_emojis={}):
"""
@param master_key: Veja a documentação na classe Listener.
@param clipboard: Veja a documentação na classe Generator.
@param keys_and_emojis: Dicionário com teclas e emojis.
"""
self.__master_key = master_key
self.__clipboard = clipboard
self.__keys_and_emojis = keys_and_emojis
self.__emoji_generator = Generator()
self.__listener = None
def test_run_generator():
g = Generator()
z_code = torch.randn(BATCH, D_Z)
sent_emb = torch.randn(BATCH, D_HIDDEN)
word_embs = torch.randn(BATCH, D_WORD, CAP_MAX_LEN)
gen, att, mu, logvar = g(z_code, sent_emb, word_embs, None)
print('Generated shape:')
for k in gen:
print(gen[k].size())
print('Attention shape:')
for a in att:
print(att[a].size())
print(f'Mu shape: {mu.size()} logvar shape: {logvar.size()}')
return gen, att, mu, logvar
def test_get_neighbours_0(param):
inp1, inp2, answer = param
cells = [
generator._cells[3][3], generator._cells[2][2], generator._cells[1][3]
]
for cell in cells:
cell.accepted = True
neighbours = generator._get_neighbours(
Generator.Cell(Position(inp1, inp2)), lambda c: c.accepted)
generator.debug()
assert len(neighbours['roads']) == answer
def main(args):
mode = args.mode
amount = args.number_amount
bits_number = args.bits
if mode not in range(3):
raise Exception('Mode ' + str(mode) + ' not allowed.')
if amount < 0:
raise Exception('Amount ' + str(amount) + ' not allowed.')
if bits_number < 0:
raise Exception('Number of bits ' + str(bits_number) + ' not allowed.')
qrangen = Generator(mode, amount, bits_number)
print(qrangen.generate_number())
def generate(self, event):
from src.generator import Generator
import os
schoolName = self.inputs["schoolName"].get()
examName = self.inputs["examName"].get()
authorityName = self.inputs["authorityName"].get()
if (not schoolName) and (not examName):
messagebox.showwarning("Required",
"School Name and Exam Name are required!!")
return
elif not schoolName:
messagebox.showwarning("Required", "School Name is required!!")
return
elif not examName:
messagebox.showwarning("Required", "Exam Name is required!!")
return
try:
self.waitLabel["text"] = "Wait..."
factory = Generator(
schoolName,
examName,
self.file,
self.signatureFile.name if self.signatureFile else None,
authorityName,
)
if not os.path.isdir("pdfs"):
os.mkdir("pdfs", 0o666)
print("Starting to generate PDFs")
for student in factory.getCSV():
factory.generatePDF(student["Roll Number"], log=True)
self.waitLabel["text"] = ""
messagebox.showinfo("Success!!",
"Generated PDFs!! Check the 'pdfs' folder")
except Exception as e:
print(e.with_traceback())
self.waitLabel["text"] = ""
messagebox.showerror("Error!!",
"Something went wrong!! Please try again!!")
def __init__(self, damsm, device=DEVICE):
self.gen = Generator(device)
self.disc = Discriminator(device)
self.damsm = damsm.to(device)
self.damsm.txt_enc.eval(), self.damsm.img_enc.eval()
freeze_params_(self.damsm.txt_enc), freeze_params_(self.damsm.img_enc)
self.device = device
self.gen.apply(init_weights), self.disc.apply(init_weights)
self.gen_optimizer = torch.optim.Adam(self.gen.parameters(),
lr=GENERATOR_LR,
betas=(0.5, 0.999))
self.discriminators = [self.disc.d64, self.disc.d128, self.disc.d256]
self.disc_optimizers = [
torch.optim.Adam(d.parameters(),
lr=DISCRIMINATOR_LR,
betas=(0.5, 0.999)) for d in self.discriminators
]
def setUp(self):
self.batch_size = 32
self.df = pd.read_csv(f"{c['WORK_DIR']}/work.csv")
self.generator = Generator(df=self.df)
parser.add_argument(
"--size",
type=int,
default=[224, 224],
nargs="+",
help="Target image size (WxH)",
)
args = parser.parse_args()
print(f"* Arguments:\n{pformat(vars(args))}")
# endregion
g = Generator(
csv_file=f"{gc['DATA_DIR']}/src/{args.set}.csv",
images_src_dir=f"{gc['DATA_DIR']}/src/{args.set}_images",
target_image_size=tuple(args.size),
image_augmentation_options=None,
cache_dir=gc["DATA_DIR"] + "/images_cache",
)
def _mapping(x):
x, y = g.get_one(x,
use_cached=False,
write_cache=True,
normalize=False,
augment=False)
with Pool(cpu_count()) as pool:
list(tqdm(
def __init__(self, description, path_to_settings):
self.description = description
self.path_to_generator = path_to_settings
data = Generator(self.path_to_generator)
self.intelligence = Engine(data)
def main():
parser = argparse.ArgumentParser()
parser.add_argument('--cuda',
default=False,
action='store_true',
help='Enable CUDA')
args = parser.parse_args()
use_cuda = True if args.cuda and torch.cuda.is_available() else False
random.seed(SEED)
np.random.seed(SEED)
netG = Generator(VOCAB_SIZE, G_EMB_SIZE, G_HIDDEN_SIZE, G_LR, use_cuda)
netD = Discriminator(VOCAB_SIZE, D_EMB_SIZE, D_NUM_CLASSES, D_FILTER_SIZES,
D_NUM_FILTERS, DROPOUT, D_LR, D_L2_REG, use_cuda)
oracle = Oracle(VOCAB_SIZE, G_EMB_SIZE, G_HIDDEN_SIZE, use_cuda)
# generating synthetic data
# print('Generating data...')
# generate_samples(oracle, BATCH_SIZE, GENERATED_NUM, REAL_FILE)
# pretrain generator
gen_set = GeneratorDataset(REAL_FILE)
genloader = DataLoader(dataset=gen_set,
batch_size=BATCH_SIZE,
shuffle=True)
print('\nPretraining generator...\n')
for epoch in range(PRE_G_EPOCHS):
loss = netG.pretrain(genloader)
print('Epoch {} pretrain generator training loss: {}'.format(
epoch, loss))
generate_samples(netG, BATCH_SIZE, GENERATED_NUM, EVAL_FILE)
val_set = GeneratorDataset(EVAL_FILE)
valloader = DataLoader(dataset=val_set,
batch_size=BATCH_SIZE,
shuffle=True)
loss = oracle.val(valloader)
print('Epoch {} pretrain generator val loss: {}'.format(
epoch + 1, loss))
# pretrain discriminator
print('\nPretraining discriminator...\n')
for epoch in range(D_STEPS):
generate_samples(netG, BATCH_SIZE, GENERATED_NUM, FAKE_FILE)
dis_set = DiscriminatorDataset(REAL_FILE, FAKE_FILE)
disloader = DataLoader(dataset=dis_set,
batch_size=BATCH_SIZE,
shuffle=True)
for _ in range(K_STEPS):
loss = netD.dtrain(disloader)
print('Epoch {} pretrain discriminator training loss: {}'.format(
epoch + 1, loss))
# adversarial training
rollout = Rollout(netG,
update_rate=ROLLOUT_UPDATE_RATE,
rollout_num=ROLLOUT_NUM)
print('\n#####################################################')
print('Adversarial training...\n')
for epoch in range(TOTAL_EPOCHS):
for _ in range(G_STEPS):
netG.pgtrain(BATCH_SIZE, SEQUENCE_LEN, rollout, netD)
for d_step in range(D_STEPS):
# train discriminator
generate_samples(netG, BATCH_SIZE, GENERATED_NUM, FAKE_FILE)
dis_set = DiscriminatorDataset(REAL_FILE, FAKE_FILE)
disloader = DataLoader(dataset=dis_set,
batch_size=BATCH_SIZE,
shuffle=True)
for k_step in range(K_STEPS):
loss = netD.dtrain(disloader)
print(
'D_step {}, K-step {} adversarial discriminator training loss: {}'
.format(d_step + 1, k_step + 1, loss))
rollout.update_params()
generate_samples(netG, BATCH_SIZE, GENERATED_NUM, EVAL_FILE)
val_set = GeneratorDataset(EVAL_FILE)
valloader = DataLoader(dataset=val_set,
batch_size=BATCH_SIZE,
shuffle=True)
loss = oracle.val(valloader)
print('Epoch {} adversarial generator val loss: {}'.format(
epoch + 1, loss))
############################################
pred = Predictor(weights_path, meta_path)
type(pred)
pred.predict('This is incredible! I love it, this is the best chicken I have ever had.')
pred.predict('god is love')
top_labels = pred.predict_top_k('god is love', top_k_n=5)
############################################
# Get generated text similar to as above
############################################
gen = Generator(weights_path, meta_path)
gen.generate(
cond_text="In the beginning",
bag_of_words='data/bow_confucianism.tsv',
class_label=None,
)
############################################
# Recommendations
############################################
attr_labels = {
"None": None,
def generateOtherLines(powersOfTwo, specMaker, name):
lineLengths = powersOfTwo.copy()
gamutSizes = powersOfTwo.copy()
tasks = list(itertools.product(lineLengths, gamutSizes))
results = [] # (lineLength, GamutSize, time)
causes = [] #(lineLength, gamutSize, result)
lines = []
print(tasks)
for (length, gamut) in tasks:
print(length, gamut)
try:
cf = Generator(cantusSpec(length, gamut, '')).createNew()
if cf == []:
causes.append((length, gamut, 'Invalid cf'))
except Exception:
causes.append((length, gamut, 'Invalid cf'))
continue
try:
g = Generator(specMaker([ConstPitch(x) for x in cf], gamut, ''))
except Exception:
causes.append((length, gamut, 'Invalid species'))
continue
i = 0
while True:
if i == 5:
causes.append((length, gamut, 'over5'))
break
i += 1
res, runtime = timeWithTimeout(lambda: g.createNew(), 10)
# No more unique output for task
if res == []:
if i == 1:
causes.append((length, gamut, 'No possible species'))
else:
causes.append((length, gamut, 'All outputs found'))
break
#timeout
elif res is None:
causes.append((length, gamut, 'timeout'))
break
else:
out = (length, gamut, i, runtime)
lines.append((length, gamut, res))
results.append(out)
success = pd.DataFrame(columns=['length', 'gamut', 'iteration', 'times'],
data=results)
fails = pd.DataFrame(columns=['length', 'gamut', 'result'], data=causes)
linesDF = pd.DataFrame(columns=['length', 'gamut', 'line'], data=lines)
print(success)
print(fails)
print(lines)
dt = str(datetime.datetime.now())
success.to_csv(
'output/generator/' + name + '/generate' + name + 'Success_Max:' +
str(powersOfTwo[-1]) + "_Time:" + dt + '.csv',
index=False,
)
fails.to_csv(
'output/generator/' + name + '/generate' + name + 'Fails_Max:' +
str(powersOfTwo[-1]) + "_Time:" + dt + '.csv',
index=False,
)
linesDF.to_csv(
'output/generator/' + name + '/generate' + name + 'Lines_Max:' +
str(powersOfTwo[-1]) + '.csv',
index=False,
)
def main():
args = parse_args()
device = torch.device("cuda")
generator = Generator.from_file(args.generator_path).to(device)
if args.freeze:
for name, param in generator.named_parameters():
if ("shared" not in name) and ("decoder.block.5" not in name):
param.requires_grad = False
discriminator = Discriminator.from_file(
args.discriminator_path, tokenizer=generator.tokenizer
).to(device)
if args.freeze:
for name, param in discriminator.named_parameters():
if ("shared" not in name) and ("decoder.block.5" not in name):
param.requires_grad = False
train_dataset = DailyDialogueDataset(
path_join(args.dataset_path, "train/dialogues_train.txt"),
tokenizer=generator.tokenizer,
debug=args.debug,
)
valid_dataset = DailyDialogueDataset(
path_join(args.dataset_path, "validation/dialogues_validation.txt"),
tokenizer=generator.tokenizer,
debug=args.debug,
)
print(len(train_dataset), len(valid_dataset))
generator_optimizer = AdamW(generator.parameters(), lr=args.lr)
discriminator_optimizer = AdamW(discriminator.parameters(), lr=args.lr)
rewards = deque([], maxlen=args.log_every * args.generator_steps)
rewards_real = deque([], maxlen=args.log_every * args.generator_steps)
generator_loss = deque([], maxlen=args.log_every * args.generator_steps)
discriminator_loss = deque([], maxlen=args.log_every * args.discriminator_steps)
best_reward = 0
generator.train()
discriminator.train()
for iter in tqdm(range(args.num_iterations)):
for _ in range(args.discriminator_steps):
discriminator_optimizer.zero_grad()
context, real_reply = train_dataset.sample()
context, real_reply = (
context.to(device),
real_reply.to(device),
)
fake_reply = generator.generate(context, do_sample=True)
if args.regs:
split_real = random.randint(1, real_reply.size(1))
real_reply = real_reply[:, :split_real]
split_fake = random.randint(1, fake_reply.size(1))
fake_reply = fake_reply[:, :split_fake]
loss, _, _ = discriminator.get_loss(context, real_reply, fake_reply)
loss.backward()
discriminator_optimizer.step()
discriminator_loss.append(loss.item())
for _ in range(args.generator_steps):
generator_optimizer.zero_grad()
context, real_reply = train_dataset.sample()
context, real_reply = (
context.to(device),
real_reply.to(device),
)
fake_reply = generator.generate(context, do_sample=True)
logprob_fake = generator.get_logprob(context, fake_reply)
reward_fake = discriminator.get_reward(context, fake_reply)
baseline = 0 if len(rewards) == 0 else np.mean(list(rewards))
if args.regs:
partial_rewards = torch.tensor(
[
discriminator.get_reward(context, fake_reply[:, :t])
for t in range(1, fake_reply.size(1) + 1)
]
).to(device)
loss = -torch.mean(partial_rewards * logprob_fake)
else:
loss = -(reward_fake - baseline) * torch.mean(logprob_fake)
if args.teacher_forcing:
logprob_real = generator.get_logprob(context, real_reply)
reward_real = discriminator.get_reward(context, real_reply)
loss -= torch.mean(logprob_real)
rewards_real.append(reward_real)
loss.backward()
generator_optimizer.step()
generator_loss.append(loss.item())
rewards.append(reward_fake)
if iter % args.log_every == 0:
mean_reward = np.mean(list(rewards))
mean_reward_real = np.mean(list(rewards_real))
if args.discriminator_steps > 0:
print(f"Discriminator Loss {np.mean(list(discriminator_loss))}")
if args.generator_steps > 0:
print(f"Generator Loss {np.mean(list(generator_loss))}")
if args.teacher_forcing:
print(f"Mean real reward: {mean_reward_real}")
print(f"Mean fake reward: {mean_reward}\n")
context, real_reply = valid_dataset.sample()
context, real_reply = (
context.to(device),
real_reply.to(device),
)
fake_reply = generator.generate(context, do_sample=True)
reward_fake = discriminator.get_reward(context, fake_reply)
print_dialogue(
context=context,
real_reply=real_reply,
fake_reply=fake_reply,
tokenizer=generator.tokenizer,
)
print(f"Reward: {reward_fake}\n")
if mean_reward > best_reward:
best_reward = mean_reward
torch.save(discriminator.state_dict(), args.discriminator_output_path)
torch.save(generator.state_dict(), args.generator_output_path)
torch.save(
discriminator.state_dict(), "all_" + args.discriminator_output_path
)
torch.save(generator.state_dict(), "all_" + args.generator_output_path)
def main():
args = parse_args()
device = torch.device("cuda")
generator = Generator.from_file(args.generator_path).to(device)
generator.eval()
discriminator = Discriminator(tokenizer=generator.tokenizer).to(device)
train_dataset = DailyDialogueDataset(
path_join(args.dataset_path, "train/dialogues_train.txt"),
tokenizer=generator.tokenizer,
)
valid_dataset = DailyDialogueDataset(
path_join(args.dataset_path, "validation/dialogues_validation.txt"),
tokenizer=generator.tokenizer,
)
print(len(train_dataset), len(valid_dataset))
optimizer = AdamW(discriminator.parameters(), lr=args.lr)
for epoch in tqdm(range(args.num_epochs)):
train_loss, valid_loss = [], []
rewards_real, rewards_fake, accuracy = [], [], []
discriminator.train()
for ind in np.random.permutation(len(train_dataset)):
optimizer.zero_grad()
context, real_reply = train_dataset.sample_dialouge(ind)
context, real_reply = (
context.to(device),
real_reply.to(device),
)
fake_reply = generator.generate(context, do_sample=True)
loss, _, _ = discriminator.get_loss(context, real_reply,
fake_reply)
loss.backward()
optimizer.step()
train_loss.append(loss.item())
discriminator.eval()
real_replies, fake_replies = [], []
for ind in range(len(valid_dataset)):
context, real_reply = valid_dataset[ind]
context, real_reply = (
context.to(device),
real_reply.to(device),
)
fake_reply = generator.generate(context, do_sample=True)
with torch.no_grad():
loss, reward_real, reward_fake = discriminator.get_loss(
context, real_reply, fake_reply)
valid_loss.append(loss.item())
rewards_real.append(reward_real)
rewards_fake.append(reward_fake)
accuracy.extend([reward_real > 0.5, reward_fake < 0.5])
real_reply, fake_reply = (
generator.tokenizer.decode(real_reply[0]),
generator.tokenizer.decode(fake_reply[0]),
)
real_replies.append(real_reply)
fake_replies.append(fake_reply)
train_loss, valid_loss = np.mean(train_loss), np.mean(valid_loss)
print(
f"Epoch {epoch + 1}, Train Loss: {train_loss:.2f}, Valid Loss: {valid_loss:.2f}, Reward real: {np.mean(rewards_real):.2f}, Reward fake: {np.mean(rewards_fake):.2f}, Accuracy: {np.mean(accuracy):.2f}"
)
print(f"Adversarial accuracy, {np.mean(accuracy):.2f}")
for order in range(1, 5):
print(
f"BLEU-{order}: {bleuscore(real_replies, fake_replies, order=order)}"
)
print(f"DIST-1: {dist1(fake_replies)}")
print(f"DIST-2: {dist2unbiased(fake_replies)}")
def generateCanti(powersOfTwo):
lineLengths = powersOfTwo.copy()
gamutSizes = powersOfTwo.copy()
tasks = list(itertools.product(lineLengths, gamutSizes))
results = [] # (lineLength, GamutSize, time)
causes = [] #(lineLength, gamutSize, result)
cfs = [] #(lineLength, gamutSize, result)
print(tasks)
for (length, gamut) in tasks:
print((length, gamut))
try:
g = Generator(cantusSpec(length, gamut, ''))
except Exception:
causes.append((length, gamut, 'Invalid'))
continue
i = 0
while True:
if i == 5:
causes.append((length, gamut, 'over5'))
break
i += 1
res, runtime = timeWithTimeout(lambda: g.createNew(), 10)
# No more unique output for task
if res == []:
if i == 1:
causes.append((length, gamut, 'No possible lines'))
else:
causes.append((length, gamut, 'All outputs found'))
break
#timeout
elif res is None:
causes.append((length, gamut, 'timeout'))
break
else:
out = (length, gamut, i, runtime)
cfs.append((length, gamut, res))
results.append(out)
success = pd.DataFrame(columns=['length', 'gamut', 'iteration', 'times'],
data=results)
fails = pd.DataFrame(columns=['length', 'gamut', 'result'], data=causes)
lines = pd.DataFrame(columns=['length', 'gamut', 'line'], data=cfs)
print(success)
print(fails)
print(cfs)
dt = str(datetime.datetime.now())
success.to_csv(
'output/generator/cf/generateCFSuccess_Max:' + str(powersOfTwo[-1]) +
"_Time:" + dt + '.csv',
index=False,
)
fails.to_csv(
'output/generator/cf/generateCFFails_Max:' + str(powersOfTwo[-1]) +
"_Time:" + dt + '.csv',
index=False,
)
lines.to_csv(
'output/generator/cf/generateCFCFs_Max:' + str(powersOfTwo[-1]) +
'.csv',
index=False,
)
type=float,
default=1,
help="Zoom factor",
)
args = parser.parse_args()
print(f"* Arguments:\n{pformat(vars(args))}")
# endregion
os.chdir(c["WORK_DIR"])
df = pd.read_csv(args.in_csv)
g = Generator(
df=df,
batch_size=1,
shuffle=False,
zoom=args.zoom,
augmentation_options=None,
image_output_size=tuple(args.size),
)
def _mapping(i):
g.__getitem__(i)
with Pool(cpu_count()) as pool:
list(
tqdm(
pool.imap(
_mapping,
range(df.shape[0]),