USE_CUDA = torch.cuda.is_available()
from dqn import QLearner, compute_td_loss, ReplayBuffer
env_id = "PongNoFrameskip-v4" # established environment that will be played
env = make_atari(env_id)
env = wrap_deepmind(env)
env = wrap_pytorch(env)
num_frames = 1000000 # total frames that will be learning from
batch_size = 32 # the number of samples that are provided to the model for update services at a given time
gamma = 0.99 # the discount of future rewards
record_idx = 10000 #
replay_initial = 10000 # number frames that are held
replay_buffer = ReplayBuffer(100000)
model = QLearner(env, num_frames, batch_size, gamma, replay_buffer)
model.load_state_dict(
torch.load("model_pretrained.pth",
map_location='cpu')) #loading in the pretrained model
target_model = QLearner(env, num_frames, batch_size, gamma,
replay_buffer) #load in model
target_model.copy_from(model)
optimizer = optim.Adam(model.parameters(),
lr=0.0001) #learning rate set and optimizing the model
if USE_CUDA:
model = model.cuda() # sends model to gpu
target_model = target_model.cuda()
print("Using cuda")
import os.path as op
import matplotlib.pyplot as plt
env_id = "PongNoFrameskip-v4"
env = make_atari(env_id)
env = wrap_deepmind(env)
env = wrap_pytorch(env)
num_frames = 1000000
batch_size = 32
gamma = 0.99
replay_initial = 10000
replay_buffer = ReplayBuffer(100000)
model = QLearner(env, num_frames, batch_size, gamma, replay_buffer)
model.load_state_dict(torch.load('trained_model.pth'))
model.eval()
if USE_CUDA:
model = model.cuda()
epsilon_start = 1.0
epsilon_final = 0.01
epsilon_decay = 30000
losses = []
all_rewards = []
episode_reward = 0
loss_list = []
reward_list = []
USE_CUDA = torch.cuda.is_available()
from dqn import QLearner, compute_td_loss, ReplayBuffer
env_id = "PongNoFrameskip-v4"
env = make_atari(env_id)
env = wrap_deepmind(env)
env = wrap_pytorch(env)
num_frames = 1000000
batch_size = 32
gamma = 0.99
record_idx = 10000
replay_initial = 10000
replay_buffer = ReplayBuffer(100000)
model = QLearner(env, num_frames, batch_size, gamma, replay_buffer)
model.load_state_dict(torch.load("model_pretrained.pth", map_location='cpu'))
target_model = QLearner(env, num_frames, batch_size, gamma, replay_buffer)
target_model.copy_from(model)
optimizer = optim.Adam(model.parameters(), lr=0.00001)
if USE_CUDA:
model = model.cuda()
target_model = target_model.cuda()
print("Using cuda")
epsilon_start = 1.0
epsilon_final = 0.01
epsilon_decay = 30000
epsilon_by_frame = lambda frame_idx: epsilon_final + (
USE_CUDA = torch.cuda.is_available()
from dqn import QLearner, compute_td_loss, ReplayBuffer
env_id = "PongNoFrameskip-v4"
env = make_atari(env_id)
env = wrap_deepmind(env)
env = wrap_pytorch(env)
num_frames = 1000000
batch_size = 32
gamma = 0.99
replay_initial = 10000
replay_buffer = ReplayBuffer(100000)
model = QLearner(env, num_frames, batch_size, gamma, replay_buffer)
model.load_state_dict(torch.load(sys.argv[1], map_location='cpu'))
model.eval()
if USE_CUDA:
model = model.cuda()
print("Using cuda")
model.load_state_dict(torch.load(pthname, map_location='cpu'))
env.seed(1)
state = env.reset()
done = False
games_won = 0
while not done:
USE_CUDA = torch.cuda.is_available()
from dqn import QLearner, compute_td_loss, ReplayBuffer
import sys
env_id = "PongNoFrameskip-v4"
env = make_atari(env_id)
env = wrap_deepmind(env)
env = wrap_pytorch(env)
num_frames = 1500000
batch_size = 32
gamma = 0.9
replay_initial = 10000
replay_buffer = ReplayBuffer(1000000)
model = QLearner(env, num_frames, batch_size, gamma, replay_buffer)
optimizer = optim.Adam(model.parameters(), lr=0.00001)
if USE_CUDA:
model = model.cuda()
epsilon_start = 0.1
epsilon_final = 0.1
epsilon_decay = 30000
epsilon_by_frame = lambda frame_idx: epsilon_final + (
epsilon_start - epsilon_final) * math.exp(-1. * frame_idx / epsilon_decay)
losses = []
all_rewards = []
episode_reward = 0
state = env.reset()
env = wrap_pytorch(env)
train_num_frames = 5000000
sample_num_frames = 50000
batch_size = 32
gamma = 0.99
target_update = 50000
epsilon_start = 1.0
epsilon_final = 0.01
epsilon_decay = 1000000
replay_initial = 10000
learning_rate = 1e-5
train_replay_buffer = ReplayBuffer(100000)
analysis_replay_buffer = ReplayBuffer(100000)
policy_model = QLearner(env, train_num_frames, batch_size, gamma,
train_replay_buffer)
target_model = QLearner(env, train_num_frames, batch_size, gamma,
train_replay_buffer)
target_model.load_state_dict(policy_model.state_dict())
target_model.eval()
optimizer = optim.Adam(policy_model.parameters(), lr=learning_rate)
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
if USE_CUDA:
policy_model = policy_model.to(device)
target_model = target_model.to(device)
epsilon_by_frame = lambda frame_idx: epsilon_final + (
epsilon_start - epsilon_final) * math.exp(-1. * frame_idx / epsilon_decay)
USE_CUDA = torch.cuda.is_available()
from dqn import QLearner, compute_td_loss, ReplayBuffer
env_id = "PongNoFrameskip-v4"
env = make_atari(env_id)
env = wrap_deepmind(env)
env = wrap_pytorch(env)
num_frames = 1000000
batch_size = 32
gamma = 0.99
record_idx = 10000
replay_initial = 10000
replay_buffer = ReplayBuffer(100000)
model = QLearner(env, num_frames, batch_size, gamma, replay_buffer)
model.load_state_dict(torch.load("model_pretrained.pth", map_location='cpu'))
target_model = QLearner(env, num_frames, batch_size, gamma, replay_buffer)
target_model.copy_from(model)
optimizer = optim.Adam(model.parameters(), lr=0.00001)
if USE_CUDA:
model = model.cuda()
target_model = target_model.cuda()
print("Using cuda")
epsilon_start = 1.0
epsilon_final = 0.01
epsilon_decay = 30000
epsilon_by_frame = lambda frame_idx: epsilon_final + (
plt.figure(2)
plt.plot([reward[0] for reward in all_rewards], [reward[1] for reward in all_rewards])
plt.xlabel('Frame #')
plt.ylabel('Episode Reward')
plt.savefig(f'rewards_lr={lr}.pdf')
USE_CUDA = torch.cuda.is_available()
# Set up game
env_id = "PongNoFrameskip-v4"
env = make_atari(env_id)
env = wrap_deepmind(env)
env = wrap_pytorch(env)
replay_buffer = ReplayBuffer(replay_buff_size) # Buffer size
model = QLearner(env, num_frames, batch_size, gamma, replay_buffer) # Create model
model.load_state_dict(torch.load("model_pretrained.pth", map_location='cpu'))
model.eval()
target_model = QLearner(env, num_frames, batch_size, gamma, replay_buffer) # Create target model
target_model.copy_from(model)
# Optimize model's parameters
optimizer = optim.Adam(model.parameters(), lr=lr)
if USE_CUDA:
model = model.cuda()
target_model = target_model.cuda()
print("Using cuda")
# Neg exp func. Start exploring then exploiting according to frame_indx
epsilon_by_frame = lambda frame_idx: epsilon_final + (epsilon_start - epsilon_final) * math.exp(-1. * frame_idx / epsilon_decay)
import torch.nn.functional as F
USE_CUDA = torch.cuda.is_available()
from dqn import QLearner, compute_td_loss, ReplayBuffer
env_id = "PongNoFrameskip-v4"
env = make_atari(env_id)
env = wrap_deepmind(env)
env = wrap_pytorch(env)
num_frames = 1400000
batch_size = 32
gamma = 0.99
replay_initial = 10000
replay_buffer = ReplayBuffer(100000)
model = QLearner(env, num_frames, batch_size, gamma, replay_buffer)
optimizer = optim.Adam(model.parameters(), lr=0.00001)
if USE_CUDA:
model = model.cuda()
epsilon_start = 1.0
epsilon_final = 0.01
epsilon_decay = 30000
epsilon_by_frame = lambda frame_idx: epsilon_final + (epsilon_start - epsilon_final) * math.exp(-1. * frame_idx / epsilon_decay)
losses = []
all_rewards = []
episode_reward = 0
state = env.reset()
USE_CUDA = torch.cuda.is_available()
from dqn import QLearner, compute_td_loss, ReplayBuffer
env_id = "PongNoFrameskip-v4"
env = make_atari(env_id)
env = wrap_deepmind(env)
env = wrap_pytorch(env)
num_frames = 1000000 # total frames will be lerning from
batch_size = 32 # num samples provided to for update ppurposes
gamma = 0.99
record_idx = 10000
replay_initial = 10000 # num frames it will hold??
replay_buffer = ReplayBuffer(100000)
model = QLearner(env, num_frames, batch_size, gamma,
replay_buffer) #from dpn.py; pth file??v
model.load_state_dict(torch.load("model_pretrained.pth", map_location='cpu'))
target_model = QLearner(env, num_frames, batch_size, gamma, replay_buffer)
target_model.copy_from(model)
optimizer = optim.Adam(model.parameters(), lr=0.00001) #lr == learning rate
if USE_CUDA: # gpu to use
model = model.cuda()
target_model = target_model.cuda()
print("Using cuda")
epsilon_start = 1.0 # in textbook, figure out what it is used fro
epsilon_final = 0.01 # go towards this from above
epsilon_decay = 30000
epsilon_by_frame = lambda frame_idx: epsilon_final + (
import sys
Variable = lambda *args, **kwargs: autograd.Variable(*args, **kwargs).cuda(
) if USE_CUDA else autograd.Variable(*args, **kwargs)
import hashlib
# initialize the game environment
env_id = "PongNoFrameskip-v4"
env = make_atari(env_id)
env = wrap_deepmind(env)
env = wrap_pytorch(env)
# initialize learner with the same parameters that it was trained with
batch_size = 32
gamma = 0.99
replay_buffer = ReplayBuffer(100000)
model = QLearner(env, 5000000, batch_size, gamma, replay_buffer)
if USE_CUDA:
model = model.cuda()
losses = []
all_rewards = []
episode_reward = 0
# load the saved model
if len(sys.argv) > 1:
checkpoint = torch.load(sys.argv[1])
model.load_state_dict(checkpoint['state_dict'])
model.eval()
# no random actions during this eval stage
epsilon = -1
env = wrap_pytorch(env)
num_frames = 100000
#num_frames = 20000
batch_size = 32
gamma = 0.99
replay_initial = 10000
replay_buffer = ReplayBuffer(100000)
epsilon_start = 1.0
epsilon_final = 0.01
epsilon_decay = 30000
epsilon_by_frame = lambda frame_idx: epsilon_final + (epsilon_start - epsilon_final) * math.exp(-1. * frame_idx / epsilon_decay)
"""loading the saved model"""
device = torch.device("cuda")
model = QLearner(env, num_frames, batch_size, gamma, replay_buffer)
filename = 'newdqnModel.pt'
model.load_state_dict(torch.load(filename))
model.to(device)
model.eval()
"""choosing 1000 frames randomly!"""
frame_range = 50000
frame_list = set(random.sample(range(1, frame_range), 1000))
vis_feature_matrix = []
vis_rewards = []
vis_actions = []
state = env.reset()
indx = 0
episode_reward = 0
USE_CUDA = torch.cuda.is_available()
from dqn import QLearner, compute_td_loss, ReplayBuffer
import pickle as pkl
env_id = "PongNoFrameskip-v4"
env = make_atari(env_id)
env = wrap_deepmind(env)
env = wrap_pytorch(env)
num_frames = 2000000
batch_size = 32
gamma = 0.99
replay_initial = 20000
replay_buffer = ReplayBuffer(200000)
model = QLearner(env, num_frames, batch_size, gamma, replay_buffer)
optimizer = optim.Adam(model.parameters(), lr=0.00001)
if USE_CUDA:
model = model.cuda()
epsilon_start = 1.0
epsilon_final = 0.01
epsilon_decay = 30000
epsilon_by_frame = lambda frame_idx: epsilon_final + (
epsilon_start - epsilon_final) * math.exp(-1. * frame_idx / epsilon_decay)
losses = []
all_rewards = []
episode_reward = 0
state = env.reset()
USE_CUDA = torch.cuda.is_available()
from dqn import QLearner, compute_td_loss, ReplayBuffer
env_id = "PongNoFrameskip-v4"
env = make_atari(env_id)
env = wrap_deepmind(env)
env = wrap_pytorch(env)
num_frames = 1000000
batch_size = 32 #num samples provided to model at a time to update
gamma = 0.99
record_idx = 10000
replay_initial = 10000
replay_buffer = ReplayBuffer(100000)
model = QLearner(env, num_frames, batch_size, gamma, replay_buffer)
model.load_state_dict(torch.load("modelsave.pth", map_location='cpu'))
target_model = QLearner(env, num_frames, batch_size, gamma, replay_buffer)
target_model.copy_from(model) #prevents too many radical changes
optimizer = optim.Adam(model.parameters(), lr=0.0001)
if USE_CUDA: #GPU stuff
model = model.cuda()
target_model = target_model.cuda()
print("Using cuda")
epsilon_start = 0.5 # Q-learning "error"
epsilon_final = 0.01
epsilon_decay = 30000
epsilon_by_frame = lambda frame_idx: epsilon_final + (
USE_CUDA = torch.cuda.is_available()
from dqn import QLearner, compute_td_loss, ReplayBuffer
env_id = "PongNoFrameskip-v4"
env = make_atari(env_id)
env = wrap_deepmind(env)
env = wrap_pytorch(env)
num_frames = 1000000
batch_size = 32
gamma = 0.99
replay_initial = 10000
replay_buffer = ReplayBuffer(100000)
t_replay_buffer = ReplayBuffer(100000)
model = QLearner(env, num_frames, batch_size, gamma, replay_buffer)
target_model = QLearner(env, num_frames, batch_size, gamma, t_replay_buffer)
target_model.load_state_dict(model.state_dict())
optimizer = optim.Adam(model.parameters(), lr=0.00001)
if USE_CUDA:
model = model.cuda()
target_model = target_model.cuda()
epsilon_start = 1.0
epsilon_final = 0.01
epsilon_decay = 30000
epsilon_by_frame = lambda frame_idx: epsilon_final + (
epsilon_start - epsilon_final) * math.exp(-1. * frame_idx / epsilon_decay)
losses = []
# DISCOUNTED REWARD CONSTANT gamma = 0.99 # ... record_idx = 10000 # ... replay_initial = 10000 # STORES MOMENTS FROM TRAINING EXPERIENCES, RANDOMLY SAMPLED? # STORES INPUTS AND TRANSITIONS / REWARDS FOR ANY GIVEN FRAME replay_buffer = ReplayBuffer(100000) # CONTAINS THE NETWORK THAT APPROXIMATES THE Q-TABLE model = QLearner(env, num_frames, batch_size, gamma, replay_buffer) # LOADS THE SAVED NN model.load_state_dict(torch.load(file_name, map_location='cpu')) # CONTAINS THE TARGET NETWORK THAT APPROXIMATES THE Q-TABLE target_model = QLearner(env, num_frames, batch_size, gamma, replay_buffer) # DUPLICATE THE POLICY NETWORK INTO THE TARGET (APPROXIMATES Q*) target_model.copy_from(model) # INITIALIZE OPTIMIZER, USING ADAM METHOD TO UPDATE GRADIENTS optimizer = optim.Adam(model.parameters(), lr=0.0001) # CUDA (GFX CARD) SETUP if USE_CUDA:
