def run(config):
print(config)
checkConfigParams(config)
np.random.seed(config["random_seed"])
random.seed(config["random_seed"])
dataset = Dataset(**config)
# preconditions: the dataset has generated a pool of data, and the folds have been generated
if config["kfcv"] == 1 and config["kfcv_serial"] == 1:
for i in range(0, config["folds"]):
neuralnet = ConvModel(dataset, **config)
#validate that the folds work
#dataset.kfcvPrintFoldInfo()
#xit()
gc.collect()
runNet(neuralnet, config, i)
# preconditions: the dataset has generated a pool of data, and the folds have been generated
elif config["kfcv"] == 1:
neuralnet = ConvModel(dataset, **config)
runNet(neuralnet, config)
# preconditions: the dataset has generated distinct training and testing set
else:
neuralnet = ConvModel(dataset, **config)
runNet(neuralnet, config)
def main():
args = parser.parse_args()
if os.path.isdir(args.output_folder):
raise Exception("Experiment name " + args.output_folder +
" already exists.")
os.mkdir(args.output_folder)
with open(args.output_folder + "/args.pckl", "wb") as f:
pickle.dump(args, f)
transform = None
if args.normalize:
transform = NormalizeFixedFactor(1280)
# utterance_dict = build_dataset_structure(args.utterance_folder)
# metadata_structure = [utterance_dict]
if "Text" in args.model:
dataset = FastTextPoseDataset(args.data,
args.max_frames,
transform,
use_rand_tokens=args.rand_tokens)
else:
dataset = FastPoseDataset(args.data, args.max_frames, transform)
loader = DataLoader(dataset, batch_size=1, collate_fn=collate_function)
if args.model == "Conv":
model = ConvModel(args.conv_channels,
activation="ReLU",
pos_emb=args.conv_pos_emb)
elif args.model == "TransformerEncoder":
model = TransformerEncoder(args, 100)
elif args.model == "ConvTransformerEncoder":
model = ConvTransformerEncoder(args, 21 * 2)
elif args.model == "TransformerEnc":
model = TransformerEnc(ninp=12 * 2,
nhead=4,
nhid=100,
nout=21 * 2,
nlayers=4,
dropout=0.0)
elif args.model == "TextPoseTransformer":
model = TextPoseTransformer(n_tokens=1000,
n_joints=12,
joints_dim=2,
nhead=4,
nhid=128,
nout=21 * 2,
n_enc_layers=4,
n_dec_layers=4,
dropout=args.transformer_dropout)
else:
raise ValueError()
model.load_state_dict(torch.load(args.model_checkpoint))
infer_utterance(model, loader, args)
def run(args, use_cuda, output_dir):
trial_list = list(range(args.n_trials))
np.random.shuffle(trial_list)
for trial_i in trial_list:
trial_dir = os.path.join(output_dir, 'trial_{}'.format(trial_i))
os.makedirs(trial_dir, exist_ok=True)
loaders, params = get_dataloaders(args.batch_size,
trial_i,
args.dataset,
args.augment_data,
early_stop=args.early_stop)
if args.network_type == 'fc':
model = DenseModel(input_dim=np.prod(params['input_shape']),
output_dim=params['output_dim'],
hidden_nodes=args.hidden_nodes,
num_modules=args.n_modules,
activation=args.activation)
elif args.network_type == 'conv':
model = ConvModel(input_shape=params['input_shape'],
output_dim=params['output_dim'],
num_filters=args.filters,
kernel_sizes=args.kernels,
strides=args.strides,
dilations=args.dilations,
num_modules=args.n_modules,
activation=args.activation,
final_layer=args.conv_final_layer)
elif args.network_type == 'densenet':
model = DenseNet(input_shape=params['input_shape'],
output_dim=params['output_dim'],
growth_rate=args.densenet_k,
depth=args.densenet_depth,
reduction=args.densenet_reduction,
bottleneck=args.densenet_bottleneck,
num_modules=args.n_modules)
logging.debug(args)
logging.debug('Parameters: {}'.format(model.n_parameters()))
device = torch.device("cuda" if use_cuda else "cpu")
model = model.to(device)
model.reset_parameters()
weight_path = os.path.join(trial_dir, 'initial_weights.pt')
torch.save(model.state_dict(), weight_path)
for lambda_i, (lambda_, learning_rate) in enumerate(
zip(args.lambda_values, args.learning_rates)):
model.load_state_dict(torch.load(weight_path))
lambda_dir = os.path.join(trial_dir, str(lambda_))
os.makedirs(lambda_dir, exist_ok=True)
do_lambda_value(model, lambda_, learning_rate, args, loaders,
params['distribution'], device, lambda_dir)
def main():
args = parser.parse_args()
if os.path.isfile(args.out_h5data):
raise Exception("Experiment name " + args.out_h5data + " already exists.")
transforms = []
# TODO Encode body points also differentially to some joint not only hand wrt wrist
if args.dif_encoding:
transforms.append(WristDifference())
transforms.append(ChestDifference())
# TODO Change Normalization scheme to fixed bone dist
if args.normalize:
transforms.append(NormalizeFixedFactor(1280))
if args.predict == "right_index":
n_input = 12 + 17
n_output = 4
transforms.append(BuildIndexItem())
elif args.predict == "right_3fingers":
n_input = 12 + 9
n_output = 12
transforms.append(Build3fingerItem())
elif args.predict == "right_hand":
# n_input = 12
n_input = 8
n_output = 21
transforms.append(BuildRightHandItem())
else:
raise ValueError()
transforms = torchvision.transforms.Compose(transforms)
if "Text" in args.model:
dataset = TextPoseH5Dataset(args.valid_h5data, args.valid_textdata, args.max_frames, transforms, selection=args.frames_selection,
use_rand_tokens=args.rand_tokens)
else:
dataset = FastPoseDataset(args.data, args.max_frames, transforms)
loader = DataLoader(dataset, batch_size=128, collate_fn=collate_function_h5)
if args.model == "Conv":
model = ConvModel(args.conv_channels, "ReLU", pos_emb=args.conv_pos_emb)
elif args.model == "ConvTransformerEncoder":
model = ConvTransformerEncoder(args, 21 * 2)
elif args.model == "TransformerEnc":
model = TransformerEnc(ninp=12*2, nhead=4, nhid=128, nout=21*2,
nlayers=4, dropout=args.transformer_dropout)
elif args.model == "TextPoseTransformer":
model = TextPoseTransformer(n_tokens=1000, n_joints=n_input, joints_dim=2, nhead=4,
nhid=128, nout=n_output*2, n_enc_layers=4, n_dec_layers=4,
dropout=args.transformer_dropout)
else:
raise ValueError()
model.load_state_dict(torch.load(args.model_checkpoint))
infer_utterance_h5(model, loader, args)
def create_model_and_optimizer(args, dataset, params=None):
"""creates model from the dataset and load/initializes its parameters if necessary."""
vocab_sizes = [
len(dataset.vocab[k]["w2i"]) for k in ["word", "ent", "num"]
]
embed_sizes = [
args.word_embed_size,
args.entdist_embed_size,
args.numdist_embed_size,
]
pads = [dataset.vocab[k]["w2i"]["PAD"] for k in ["word", "ent", "num"]]
if args.model == "LSTM":
model = BLSTMModel(
vocab_sizes,
embed_sizes,
sum(embed_sizes),
args.blstm_fc_hidden_dim,
len(dataset.vocab["label"]["w2i"]),
pads,
args.dropout,
)
elif args.model == "CNN":
model = ConvModel(
vocab_sizes,
embed_sizes,
sum(embed_sizes),
args.conv_fc_hidden_dim,
len(dataset.vocab["label"]["w2i"]),
pads,
args.num_filters,
args.dropout,
)
if params is not None:
model.load_state_dict(params)
else:
# initialize all the model weights
for p in model.parameters():
torch.nn.init.uniform_(p, -args.uniform_init, args.uniform_init)
# Make sure that pad vectors are zero
model.embed.pad_init()
if args.cuda:
model = model.to("cuda")
optimizer = torch.optim.SGD(model.parameters(), lr=args.initial_lr)
return model, optimizer
train_dataset = FastPoseDataset(args.train_data, args.max_frames,
transform)
valid_dataset = FastPoseDataset(args.valid_data, args.max_frames,
transform)
train_dataloader = DataLoader(train_dataset,
batch_size=args.batch_size,
collate_fn=collate_function)
valid_dataloader = DataLoader(valid_dataset,
batch_size=args.batch_size,
collate_fn=collate_function)
if args.model == "Conv":
model = ConvModel(args.conv_channels,
"ReLU",
pos_emb=args.conv_pos_emb)
elif args.model == "TransformerEncoder":
model = TransformerEncoder(args, 100)
elif args.model == "ConvTransformerEncoder":
model = ConvTransformerEncoder(args, 21 * 2)
elif args.model == "TransformerEnc":
model = TransformerEnc(ninp=12 * 2,
nhead=4,
nhid=100,
nout=21 * 2,
nlayers=4,
dropout=0.0)
else:
raise ValueError()
print(args.resume)
def create_model(self):
return ConvModel(self.train_set.in_channels(),
self.train_set.out_channels(),
dropout=self.helper.opt.cnn.dropout)
import sys
import pandas as pd
import torch
from torch.utils.data import DataLoader
from datasets import KPIDataset
from models import ConvModel
from util import print_progress_bar
dataset = KPIDataset('../data/test_preprocessed.csv',
seq_length=1001,
step_width=1,
evaluate=True)
model = ConvModel(1001)
model.load_state_dict(torch.load('./state'))
model = model.cuda()
loader = DataLoader(dataset, 256, False)
iter_per_epoch = len(loader)
result = []
with torch.no_grad():
for i, x in enumerate(loader):
x = x.cuda()
out = model(x).data.cpu().numpy()
result.extend(list(out.argmax(1)))
print_progress_bar(i, iter_per_epoch)
df = pd.read_csv('../data/test_preprocessed.csv')
def __init__(self, gpu, mode, args):
self.rank = gpu
self.mode = mode
self.args = args
self.lr = getattr(self.args, 'lr', 1e-5)
self.l2 = getattr(self.args, 'l2', 0.0001)
self.model_type = getattr(self.args, 'model_type', 'ConvRNN')
if self.args.on_cpu:
self.device = torch.device('cpu')
else:
self.device = torch.device('cuda:{}'.format(self.rank))
dist.init_process_group('nccl',
world_size=len(self.args.gpus),
rank=self.rank)
torch.manual_seed(self.args.seed)
if self.rank == 0:
print('Creating model...')
if self.model_type == 'CNN':
self.model = ConvModel(self.args, self.device).to(self.device)
else:
self.model = ConvRNNModel(self.args, self.device).to(self.device)
if not self.args.on_cpu:
self.model = torch.nn.parallel.DistributedDataParallel(self.model, device_ids=[self.rank],
find_unused_parameters=False)
ckpt_file = None
if self.mode != 'train':
# ckpt_file = self.args.out_dir
if self.args.train:
ckpt_epoch_offset = 1
ckpt_file = os.path.join(self.args.model_save_dir,
self.args.ckpt_file_tmplt.format(self.args.epochs - ckpt_epoch_offset))
while not os.path.exists(ckpt_file) and self.args.epochs - ckpt_epoch_offset >= 0:
ckpt_epoch_offset += 1
ckpt_file = os.path.join(self.args.model_save_dir,
self.args.ckpt_file_tmplt.format(self.args.epochs - ckpt_epoch_offset))
else:
ckpt_file = self.args.ckpt_file
if ckpt_file is not None:
if self.rank == 0:
print('Loading model from checkpoint...')
map_location = {'cuda:{}'.format(0): 'cuda:{}'.format(gpu_id) for gpu_id in args.gpus}
state_dict = torch.load(ckpt_file, map_location=map_location)
self.model.load_state_dict(state_dict) # , strict=False
if self.rank == 0:
print('Loading dataset...')
dataset_start_time = time.time()
if self.model_type == 'CNN':
frame_undersample_p = getattr(self.args, 'frame_undersample_p', 1.0)
frame_undersample_p = None if frame_undersample_p == 1.0 else frame_undersample_p
self.dataset = PhysicsFramesDataset(self.mode, self.args.data, self.args, undersample=frame_undersample_p)
else:
self.dataset = PhysicsDataset(self.mode, self.args.data, self.args)
if self.rank == 0:
print('Time to read dataset: {0:.2f}s'.format(time.time() - dataset_start_time))
if getattr(self.args, 'weight_xent_loss', False):
self.label_weights = self.dataset.label_weights
else:
self.label_weights = [1.0, 1.0]
if self.rank == 0:
print('Setting label weights to {}...'.format(self.label_weights))
self.model.module.set_label_weights(self.label_weights)
if self.args.on_cpu:
data_sampler = None
else:
data_sampler = torch.utils.data.distributed.DistributedSampler(self.dataset,
num_replicas=args.world_size,
rank=self.rank,
shuffle=True if self.mode == 'train'
else False)
self.data_loader = DataLoader(self.dataset, batch_size=self.args.batch_size, shuffle=False,
num_workers=self.args.n_data_workers, pin_memory=True, sampler=data_sampler)
self.n_iters = int(math.ceil(len(self.dataset) / (self.args.batch_size * len(self.args.gpus))))
self.aux_dataset = None
self.aux_data_loader = None
self.aux_n_iters = None
if self.mode == 'train' and (self.args.eval or self.args.eval_every > 0):
if self.rank == 0:
print('Loading aux dataset...')
if self.model_type == 'CNN':
frame_undersample_p = getattr(self.args, 'frame_undersample_p', 1.0)
frame_undersample_p = None if frame_undersample_p == 1.0 else frame_undersample_p
self.aux_dataset = PhysicsFramesDataset('dev', self.args.data, self.args,
force_unbalanced=True,
undersample=frame_undersample_p)
else:
self.aux_dataset = PhysicsDataset('dev', self.args.data, self.args, force_unbalanced=True,
disallow_supersample=True)
if self.args.on_cpu:
aux_data_sampler = None
else:
aux_data_sampler = torch.utils.data.distributed.DistributedSampler(self.aux_dataset,
num_replicas=args.world_size,
rank=self.rank)
self.aux_data_loader = DataLoader(self.aux_dataset, batch_size=self.args.batch_size, shuffle=False,
num_workers=self.args.n_data_workers, pin_memory=True,
sampler=aux_data_sampler)
self.aux_n_iters = int(math.ceil(len(self.aux_dataset) / (self.args.batch_size * len(self.args.gpus))))
self.summary_writer = None
if self.rank == 0 and self.mode == 'train':
self.summary_writer = SummaryWriter(log_dir=self.args.tb_dir)
self.n_epochs = 1
if self.mode == 'train':
self.n_epochs = self.args.epochs
opt_parms = filter(lambda p: p.requires_grad, self.model.parameters())
self.optimizer = optim.Adam(opt_parms, lr=self.lr, weight_decay=self.l2)
n_total_iters = self.n_iters * self.n_epochs
n_warmup_steps = self.args.warmup_proportion * n_total_iters
if n_warmup_steps > 0: # get_linear_schedule_with_warmup
self.scheduler = get_cosine_schedule_with_warmup(self.optimizer,
num_warmup_steps=n_warmup_steps,
num_training_steps=n_total_iters)
# self.scheduler = get_linear_schedule_with_warmup(self.optimizer,
# num_warmup_steps=n_warmup_steps,
# num_training_steps=n_total_iters)
else:
self.scheduler = None
self.run()
def main(name, test=False, chkpt=None, device="cuda"):
if not test:
wandb.init(project="dqn-tutorial", name=name)
do_boltzman_exploration = False
memory_size = 1000000
min_rb_size = 50000
sample_size = 32
lr = 0.0001
# eps_max = 1.0
eps_min = 0.1
eps_decay = 0.999999
env_steps_before_train = 16
tgt_model_update = 5000
epochs_before_test = 1500
env = gym.make("Breakout-v0")
env = FrameStackingAndResizingEnv(env, 84, 84, 4)
test_env = gym.make("Breakout-v0")
test_env = FrameStackingAndResizingEnv(test_env, 84, 84, 4)
last_observation = env.reset()
m = ConvModel(env.observation_space.shape, env.action_space.n, lr=lr).to(device)
if chkpt is not None:
m.load_state_dict(torch.load(chkpt))
tgt = ConvModel(env.observation_space.shape, env.action_space.n).to(device)
update_tgt_model(m, tgt)
rb = ReplayBuffer()
steps_since_train = 0
epochs_since_tgt = 0
epochs_since_test = 0
step_num = -1 * min_rb_size
episode_rewards = []
rolling_reward = 0
tq = tqdm()
try:
while True:
if test:
env.render()
time.sleep(0.05)
tq.update(1)
eps = eps_decay ** (step_num)
if test:
eps = 0
if do_boltzman_exploration:
logits = m(torch.Tensor(last_observation).unsqueeze(0).to(device))[0]
action = torch.distributions.Categorical(logits=logits).sample().item()
else:
if random() < eps:
action = (
env.action_space.sample()
) # your agent here (this takes random actions)
else:
action = m(torch.Tensor(last_observation).unsqueeze(0).to(device)).max(-1)[-1].item()
observation, reward, done, info = env.step(action)
rolling_reward += reward
rb.insert(Sarsd(last_observation, action, reward, observation, done))
last_observation = observation
if done:
episode_rewards.append(rolling_reward)
if test:
print(rolling_reward)
rolling_reward = 0
observation = env.reset()
steps_since_train += 1
step_num += 1
if (
(not test)
and rb.idx > min_rb_size
and steps_since_train > env_steps_before_train
):
loss = train_step(
m, rb.sample(sample_size), tgt, env.action_space.n, device
)
wandb.log(
{
"loss": loss.detach().cpu().item(),
"eps": eps,
"avg_reward": np.mean(episode_rewards),
},
step=step_num,
)
episode_rewards = []
epochs_since_tgt += 1
epochs_since_test += 1
if epochs_since_test > epochs_before_test:
rew, frames = run_test_episode(m, test_env, device)
# T, H, W, C
wandb.log({'test_reward': rew, 'test_video': wandb.Video(frames.transpose(0, 3, 1, 2), str(rew), fps=25, format='mp4')})
epochs_since_test = 0
if epochs_since_tgt > tgt_model_update:
print("updating target model")
update_tgt_model(m, tgt)
epochs_since_tgt = 0
torch.save(tgt.state_dict(), f"models/{step_num}.pth")
steps_since_train = 0
except KeyboardInterrupt:
pass
env.close()
def main(name=None, chkpt=None, test_run=False, local_run=False):
"Sync to wandb cloud as standard, but sync locally if local_run and not at all if test_run"
if not test_run:
if local_run:
os.environ["WANDB_MODE"] = "dryrun"
if name == None:
name = input("Name the run: ")
wandb.init(project="atari-breakout", name=name, config={
'memory_size': memory_size,
'min_rb_size': min_rb_size,
'sample_size': sample_size,
'lr': lr,
'eps_min': eps_min,
'eps_decay': eps_decay,
'discount_factor': discount_factor,
'env_steps_before_train': env_steps_before_train,
'epochs_before_tgt_model_update': epochs_before_tgt_model_update,
'epochs_before_test': epochs_before_test,
'episode_max_steps': episode_max_steps,
'optimizer_function': optimizer_function.__name__,
'exploration_method': exploration_method.__name__,
'env_type': env_type.__name__
})
"Create enviroments and reset"
env = env_type(gym.make("BreakoutDeterministic-v4"), 84, 84, 4)
test_env = env_type(gym.make("BreakoutDeterministic-v4"), 84, 84, 4)
last_observation = env.reset()
"Set the model and targetmodel"
m = ConvModel(env.observation_space.shape, env.action_space.n, lr=lr)
if chkpt is not None:
m.load_state_dict(torch.load(os.path.join(os.path.dirname(__file__), f"Models/{chkpt}")))
target = ConvModel(env.observation_space.shape, env.action_space.n)
update_target_model(m, target)
"Create replaybuffer and other variables"
rb = ReplayBuffer(memory_size)
steps_since_train = 0
epochs_since_tgt = 0
epochs_since_test = 0
step_num = -1 * min_rb_size # Want to run the iteration for min_rb_size before starting to actually learn
episode_rewards = []
total_reward = 0
tq = tqdm()
try:
while True:
if test_run:
env.render()
time.sleep(0.05)
tq.update(1)
"Updating epsilon"
eps = eps_decay ** (step_num)
if test_run:
eps = 0
elif eps < eps_min:
eps = eps_min
"Exploration vs exploitation, Boltzmann with eps_decay vs. Epsilon Greedy (defined in constants.py)"
action = exploration_method(model=m, env=env, last_observation=last_observation, eps=eps)
"Perform step and insert observation to replaybuffer"
observation, reward, done, _ = env.step(action)
total_reward += reward
rb.insert(GameInformation(last_observation, action, reward, observation, done))
last_observation = observation
"Reset and append total_reward to episode_rewards if done"
if done:
episode_rewards.append(total_reward)
if test_run:
print(total_reward)
total_reward = 0
observation = env.reset()
"Train if ran enough steps since last training"
steps_since_train += 1
step_num += 1
if ((not test_run) and rb.i > min_rb_size and steps_since_train > env_steps_before_train):
loss = train_step(m, rb.sample(sample_size), target, env.action_space.n)
if not local_run:
wandb.log(
{
"loss": loss.detach().item(),
"eps": eps,
"avg_reward": np.mean(episode_rewards),
},
step=step_num,
)
episode_rewards = []
epochs_since_tgt += 1
epochs_since_test += 1
"Run test_run episode"
if epochs_since_test > epochs_before_test:
rew, frames = run_test_episode(m, test_env)
if not local_run:
wandb.log({'test_reward': rew, 'test_video': wandb.Video(frames.transpose(0, 3, 1, 2), str(rew), fps=25, format='mp4')})
epochs_since_test = 0
"Update target model"
if epochs_since_tgt > epochs_before_tgt_model_update:
print("updating target model")
update_target_model(m, target)
epochs_since_tgt = 0
torch.save(target.state_dict(), os.path.join(os.path.dirname(__file__), f"Models/{step_num}.pth"))
steps_since_train = 0
except KeyboardInterrupt:
pass
env.close()
import sys
from torch.nn import CrossEntropyLoss
from util import Trainer
from datasets import KPIDataset
from models import ConvModel
dataset = KPIDataset(
'../data/train_preprocessed.csv',
seq_length=1001,
step_width=1
)
model = ConvModel(1001)
args = {
"lr": 0.5e-4,
"betas": (0.9, 0.999),
"eps": 1e-8,
"weight_decay": 0.0
}
trainer = Trainer(
model,
dataset,
batch_size=512,
epochs=100,
log_nth=800,
validation_size=0.2,
optim_args=args,
dataloader = torch.utils.data.DataLoader(dataset, batch_size=batch_size, shuffle=True, drop_last=True)
# Map for labels, so that we can visualize the results better
label_map = {
0: 'Apu', 1: 'Bart', 2: 'Mr. Burns', 3: 'Chief Wiggum', 4: 'Edna', 5: 'Grandpa',
6: 'Homer', 7: 'Krusty', 8: 'Lisa', 9: 'Marge', 10: 'Milhouse', 11: 'Moe',
12: 'Flanders', 13: 'Nelson', 14: 'Patty', 15: 'Skinner', 16: 'Selma', 17: 'Smithers'
}
# Load test dataset
test_dataset = torchvision.datasets.ImageFolder('./data/test', transform=transform)
validation_loader = torch.utils.data.DataLoader(test_dataset, batch_size=batch_size, shuffle=True, drop_last=True)
from models import ConvModel
model = ConvModel(channels=3, num_classes=18).to(device)
criterion = torch.nn.CrossEntropyLoss()
optimizer = torch.optim.Adam(model.parameters(), lr=learn_rate)
print(model)
trainer = create_supervised_trainer(model, optimizer, criterion, device=device)
metrics = {
"accuracy": Accuracy(),
"loss": Loss(criterion)
}
evaluator = create_supervised_evaluator(model,metrics, device=device)
@trainer.on(Events.ITERATION_COMPLETED(every=10))
def log_training_loss(engine):
def dqnmain(project_name,
do_boltzman_exploration=False,
test=False,
chkpt=None,
hypeparams=hyperparameter_defaults,
steps=1000,
device='cuda'):
image_arr = []
if (not test):
wdbrun = wandb.init(project=project_name,
config=hypeparams,
name=hypeparams['run_name'],
reinit=True,
monitor_gym=False)
# run.save("*.pth")
config = wdbrun.config
max_reward = config.max_reward
max_steps = config.max_steps
memory_size = config.memory_size
min_rb_size = config.min_rb_size
sample_size = config.sample_size
env_steps_before_train = config.env_steps_before_train
tgt_model_update = config.tgt_model_update
reward_scaler = config.reward_scaler
eps_min = config.eps_min
eps_decay = config.eps_decay
gamma = config.gamma
learning_rate = config.learning_rate
else:
max_reward = hypeparams['max_reward']
max_steps = steps
memory_size = hypeparams['memory_size']
min_rb_size = hypeparams['min_rb_size']
sample_size = hypeparams['sample_size']
env_steps_before_train = hypeparams['env_steps_before_train']
tgt_model_update = hypeparams['tgt_model_update']
reward_scaler = hypeparams['reward_scaler']
eps_min = hypeparams['eps_min']
eps_decay = hypeparams['eps_decay']
gamma = hypeparams['gamma']
learning_rate = hypeparams['learning_rate']
env = gym.make(hypeparams['env_name'])
if hypeparams['env_name'] == 'Breakout-v0':
#TODO
env = FrameStackingAndResizingEnv(env, 84, 84,
4) # change stack size here
env._max_episode_steps = 4000
test_env = gym.make(hypeparams['env_name'])
if hypeparams['env_name'] == 'Breakout-v0':
#TODO
test_env = FrameStackingAndResizingEnv(test_env, 84, 84,
4) # change stack size here
test_env._max_episode_steps = 4000
last_observation = env.reset()
if hypeparams['env_name'] == 'Breakout-v0':
m = ConvModel(env.observation_space.shape, env.action_space.n,
learning_rate).to(device)
else:
m = Model(env.observation_space.shape, env.action_space.n,
learning_rate).to(device)
if chkpt is not None:
m.load_state_dict(torch.load(chkpt))
if hypeparams['env_name'] == 'Breakout-v0':
tgt = ConvModel(env.observation_space.shape,
env.action_space.n).to(device)
else:
tgt = Model(env.observation_space.shape, env.action_space.n).to(
device) # target model, gets update fewer times
update_tgt_model(m, tgt)
rb = ReplayBuffer(memory_size)
steps_since_train = 0
epochs_since_tgt = 0
step_num = -1 * min_rb_size
i = 0
episode_rewards = []
rolling_reward = 0
solved = False
try:
while (not solved) and step_num < max_steps:
if test:
screen = env.render('rgb_array')
image_arr.append(screen)
eps = 0
else:
eps = eps_decay**(step_num)
if do_boltzman_exploration:
if hypeparams['env_name'] == 'Breakout-v0':
logits = m(
torch.Tensor(last_observation).unsqueeze(0).to(
device))[0]
action = torch.distributions.Categorical(
logits=logits).sample().item()
else:
logits = m(torch.Tensor(last_observation).to(device))[0]
action = torch.distributions.Categorical(
logits=logits).sample().item()
else:
if random.random() < eps:
action = env.action_space.sample()
else:
if hypeparams['env_name'] == 'Breakout-v0':
action = m(
torch.Tensor(last_observation).unsqueeze(0).to(
device)).max(-1)[-1].item()
else:
action = m(torch.Tensor(last_observation).to(
device)).max(-1)[-1].item()
observation, reward, done, info = env.step(action)
rolling_reward += reward
reward = reward / reward_scaler
rb.insert(SARS(last_observation, action, reward, done,
observation))
last_observation = observation
if done:
episode_rewards.append(rolling_reward)
if test:
print(rolling_reward)
rolling_reward = 0
observation = env.reset()
steps_since_train += 1
i += 1
step_num += 1
if (
not test
) and rb.idx > min_rb_size and steps_since_train > env_steps_before_train:
loss = train_step(m, rb.sample(sample_size), tgt,
env.action_space.n, gamma, device)
ave_reward = np.mean(episode_rewards)
wdbrun.log(
{
'loss': loss.detach().cpu().item(),
'epsilon': eps,
'avg_reward': ave_reward
},
step=step_num)
if ave_reward >= max_reward:
solved = True
episode_rewards = []
epochs_since_tgt += 1
# print(step_num, loss.detach().item())
if epochs_since_tgt > tgt_model_update:
# print('updating target model')
update_tgt_model(m, tgt)
rew, frames = run_test_episode(m, test_env, device)
# frames.shape == (T, H, W, C)
# wandb.log({'test_reward': rew, 'test_video': wandb.Video(frames.transpose(0, 3, 1, 2), str(rew), fps=25, format='mp4')})
wandb.log({'test_reward': rew})
epochs_since_tgt = 0
torch.save(
tgt.state_dict(),
f"{wandb.run.dir}/{hypeparams['run_name']}_{step_num}.pth"
)
steps_since_train = 0
if ave_reward >= max_reward:
solved = True
wandb.join()
env.close()
except KeyboardInterrupt:
sys.exit()