def __init__(self, env, batch_size):
self.batch_size = batch_size
self.tau = 1e-2
memory_size = 1000000
self.gamma = 0.99
actor_learning_rate = 1e-4
critic_learning_rate = 1e-3
self.critic_loss_fn = nn.MSELoss()
self.actor = DdpgActor(env.observation_space.shape[0],
env.action_space.shape[0],
env.action_space.high, env.action_space.low)
self.actor_target = DdpgActor(env.observation_space.shape[0],
env.action_space.shape[0],
env.action_space.high,
env.action_space.low)
self.copy_networks(self.actor, self.actor_target)
self.critic = Critic(env.observation_space.shape[0],
env.action_space.shape[0])
self.critic_target = Critic(env.observation_space.shape[0],
env.action_space.shape[0])
self.copy_networks(self.critic, self.critic_target)
self.memory = Memory(memory_size)
self.actor_optimizer = optim.Adam(self.actor.parameters(),
lr=actor_learning_rate)
self.critic_optimizer = optim.Adam(self.critic.parameters(),
lr=critic_learning_rate)
def __init__(self):
self.capture_view = CaptureView.instance()
self.memory = Memory()
self.cursor = Cursor()
self.setup_pages()
self.setup_name_table()
self.monitor = 'a'
self.channel = 1
self.set_page('input_monitor.' + self.monitor)
def __init__(self, filename, raw_type, raw_base, raw_big_endian, database):
import capstone as CAPSTONE
self.capstone_inst = {} # capstone instruction cache
if database.loaded:
self.mem = database.mem
else:
self.mem = Memory()
database.mem = self.mem
self.binary = Binary(self.mem, filename, raw_type, raw_base,
raw_big_endian)
self.binary.load_section_names()
arch, mode = self.binary.get_arch()
if arch is None or mode is None:
raise ExcArch(self.binary.get_arch_string())
self.jmptables = database.jmptables
self.user_inline_comments = database.user_inline_comments
self.internal_inline_comments = database.internal_inline_comments
self.user_previous_comments = database.user_previous_comments
self.internal_previous_comments = database.internal_previous_comments
self.functions = database.functions
self.func_id = database.func_id
self.end_functions = database.end_functions
self.xrefs = database.xrefs
# TODO: is it a global constant or $gp can change during the execution ?
self.mips_gp = database.mips_gp
if database.loaded:
self.binary.symbols = database.symbols
self.binary.reverse_symbols = database.reverse_symbols
self.binary.imports = database.imports
else:
self.binary.load_symbols()
database.symbols = self.binary.symbols
database.reverse_symbols = self.binary.reverse_symbols
database.imports = self.binary.imports
self.capstone = CAPSTONE
self.md = CAPSTONE.Cs(arch, mode)
self.md.detail = True
self.arch = arch
self.mode = mode
for s in self.binary.iter_sections():
s.big_endian = self.mode & self.capstone.CS_MODE_BIG_ENDIAN
# TODO: useful ?
if not database.loaded:
self.mem.add(s.start, s.end, MEM_UNK)
def __load_memory(self, data):
self.mem = Memory()
try:
self.mem.code = data["mem_code"]
except:
# Not available in previous versions, this try will be
# removed in the future
pass
def __init__(self, env, batch_size):
self.batch_size = batch_size
self.tau = 1e-2
memory_size = 1000000
self.gamma = 0.99
self.q_lr = 3e-4
self.actor_lr = 3e-4
self.alpha_lr = 3e-3
self.update_step = 0
self.delay_step = 2
self.action_range = [env.action_space.low, env.action_space.high]
self.memory = Memory(memory_size)
# entropy temperature
self.alpha = 0.2
self.target_entropy = -torch.prod(torch.Tensor(
env.action_space.shape)).item()
self.log_alpha = torch.zeros(1, requires_grad=True)
self.alpha_optim = optim.Adam([self.log_alpha], lr=self.alpha_lr)
self.actor = SacActor(env.observation_space.shape[0],
env.action_space.shape[0])
self.actor_optimizer = optim.Adam(self.actor.parameters(),
lr=self.actor_lr)
self.q_net_1 = Critic(env.observation_space.shape[0],
env.action_space.shape[0])
self.q_net_1_target = Critic(env.observation_space.shape[0],
env.action_space.shape[0])
self.copy_networks(self.q_net_1, self.q_net_1_target)
self.q_net_1_optimizer = optim.Adam(self.q_net_1.parameters(),
lr=self.q_lr)
self.q_net_2 = Critic(env.observation_space.shape[0],
env.action_space.shape[0])
self.q_net_2_target = Critic(env.observation_space.shape[0],
env.action_space.shape[0])
self.copy_networks(self.q_net_2, self.q_net_2_target)
self.q_net_2_optimizer = optim.Adam(self.q_net_2.parameters(),
lr=self.q_lr)
def __init__(self, filename, raw_type, raw_base, raw_big_endian, database):
import capstone as CAPSTONE
self.capstone_inst = {} # capstone instruction cache
if database.loaded:
self.mem = database.mem
else:
self.mem = Memory()
database.mem = self.mem
self.binary = Binary(self.mem, filename, raw_type, raw_base, raw_big_endian)
self.binary.load_section_names()
arch, mode = self.binary.get_arch()
if arch is None or mode is None:
raise ExcArch(self.binary.get_arch_string())
self.jmptables = database.jmptables
self.user_inline_comments = database.user_inline_comments
self.internal_inline_comments = database.internal_inline_comments
self.user_previous_comments = database.user_previous_comments
self.internal_previous_comments = database.internal_previous_comments
self.functions = database.functions
self.func_id = database.func_id
self.end_functions = database.end_functions
self.xrefs = database.xrefs
# TODO: is it a global constant or $gp can change during the execution ?
self.mips_gp = database.mips_gp
if database.loaded:
self.binary.symbols = database.symbols
self.binary.reverse_symbols = database.reverse_symbols
self.binary.imports = database.imports
else:
self.binary.load_symbols()
database.symbols = self.binary.symbols
database.reverse_symbols = self.binary.reverse_symbols
database.imports = self.binary.imports
self.capstone = CAPSTONE
self.md = CAPSTONE.Cs(arch, mode)
self.md.detail = True
self.arch = arch
self.mode = mode
for s in self.binary.iter_sections():
s.big_endian = self.mode & self.capstone.CS_MODE_BIG_ENDIAN
# TODO: useful ?
if not database.loaded:
self.mem.add(s.start, s.end, MEM_UNK)
def __load_memory(self, data):
self.mem = Memory()
try:
if self.version == -1:
self.mem.mm = data["mem_code"]
for ad in self.mem.mm:
self.mem.mm[ad].append(-1)
return
self.mem.mm = data["mem"]
except:
# Not available in previous versions, this try will be
# removed in the future
pass
class MemoryTest(unittest.TestCase):
def setUp(self):
self.memory = Memory(1)
def test_add_memory(self):
self.memory.add_memory("string")
self.assertEqual(self.memory.buffer, ["string"])
def test_memory_fill(self):
self.memory.add_memory("string 1")
self.memory.add_memory("string 2")
self.assertEqual(self.memory.buffer, ["string 2"])
def test_sample_memory(self):
self.memory.add_memory("string")
taken_sample = self.memory.sample_memory(3)
self.assertEqual(taken_sample, ["string"])
class Disassembler():
def __init__(self, filename, raw_type, raw_base, raw_big_endian, database):
import capstone as CAPSTONE
self.capstone_inst = {} # capstone instruction cache
self.binary = Binary(filename, raw_type, raw_base, raw_big_endian)
arch, mode = self.binary.get_arch()
if arch is None or mode is None:
raise ExcArch(self.binary.get_arch_string())
if database.loaded:
self.binary.symbols = database.symbols
self.binary.reverse_symbols = database.reverse_symbols
self.mem = database.mem
else:
self.binary.load_symbols()
database.symbols = self.binary.symbols
database.reverse_symbols = self.binary.reverse_symbols
self.mem = Memory()
database.mem = self.mem
self.jmptables = database.jmptables
self.user_inline_comments = database.user_inline_comments
self.internal_inline_comments = database.internal_inline_comments
self.user_previous_comments = database.user_previous_comments
self.internal_previous_comments = database.internal_previous_comments
self.functions = database.functions
self.end_functions = database.end_functions
# TODO: is it a global constant or $gp can change during the execution ?
self.mips_gp = database.mips_gp
self.binary.load_section_names()
self.capstone = CAPSTONE
self.md = CAPSTONE.Cs(arch, mode)
self.md.detail = True
self.arch = arch
self.mode = mode
for s in self.binary.iter_sections():
s.big_endian = self.mode & self.capstone.CS_MODE_BIG_ENDIAN
if not database.loaded:
self.mem.add(s.start, s.end, MEM_UNK)
def get_unpack_str(self, size_word):
if self.mode & self.capstone.CS_MODE_BIG_ENDIAN:
endian = ">"
else:
endian = "<"
if size_word == 1:
unpack_str = endian + "B"
elif size_word == 2:
unpack_str = endian + "H"
elif size_word == 4:
unpack_str = endian + "L"
elif size_word == 8:
unpack_str = endian + "Q"
else:
return None
return unpack_str
def add_symbol(self, addr, name):
if name in self.binary.symbols:
last = self.binary.symbols[name]
del self.binary.reverse_symbols[last]
self.binary.symbols[name] = [addr, SYM_UNK]
self.binary.reverse_symbols[addr] = [name, SYM_UNK]
return name
# TODO: create a function in SectionAbs
def read_array(self, ad, array_max_size, size_word, s=None):
unpack_str = self.get_unpack_str(size_word)
N = size_word * array_max_size
if s is None:
s = self.binary.get_section(ad)
array = []
l = 0
while l < array_max_size:
buf = s.read(ad, N)
if not buf:
break
i = 0
while i < len(buf):
b = buf[i:i + size_word]
if ad > s.end or len(b) != size_word:
return array
w = struct.unpack(unpack_str, b)[0]
array.append(w)
ad += size_word
i += size_word
l += 1
if l >= array_max_size:
return array
return array
def load_arch_module(self):
if self.arch == self.capstone.CS_ARCH_X86:
import lib.arch.x86 as ARCH
elif self.arch == self.capstone.CS_ARCH_ARM:
import lib.arch.arm as ARCH
elif self.arch == self.capstone.CS_ARCH_MIPS:
import lib.arch.mips as ARCH
else:
raise NotImplementedError
return ARCH
def get_addr_from_string(self, opt_addr, raw=False):
if opt_addr is None:
if raw:
return 0
search = ["main", "_main"]
else:
search = [opt_addr]
for s in search:
if s.startswith("0x"):
try:
a = int(opt_addr, 16)
except:
raise ExcSymNotFound(search[0])
else:
a = self.binary.symbols.get(s, -1)
if a == -1:
a = self.binary.section_names.get(s, -1)
else:
a = a[0] # it contains [ad, type]
if a != -1:
return a
raise ExcSymNotFound(search[0])
def dump_asm(self, ctx, lines=NB_LINES_TO_DISASM, until=-1):
from capstone import CS_OP_IMM
ARCH = self.load_arch_module()
ARCH_UTILS = ARCH.utils
ARCH_OUTPUT = ARCH.output
ad = ctx.entry_addr
s = self.binary.get_section(ctx.entry_addr)
if s is None:
# until is != -1 only from the visual mode
# It allows to not go before the first section.
if until != -1:
return None
# Get the next section, it's not mandatory that sections
# are consecutives !
s = self.binary.get_next_section(ad)
if s is None:
return None
ad = s.start
o = ARCH_OUTPUT.Output(ctx)
o._new_line()
o.section_prefix = True
o.curr_section = s
l = 0
while 1:
if ad == s.start:
o._new_line()
o._dash()
o._section(s.name)
o._add(" 0x%x -> 0x%x" % (s.start, s.end))
o._new_line()
o._new_line()
while ((l < lines and until == -1) or (ad != until and until != -1)) \
and ad <= s.end:
if self.mem.is_code(ad): # TODO optimize
if ad in self.functions:
if not o.is_last_2_line_empty():
o._new_line()
o._dash()
o._user_comment("; SUBROUTINE")
o._new_line()
o._dash()
i = self.lazy_disasm(ad, s.start)
o._asm_inst(i)
if ad in self.end_functions:
for e in self.end_functions[ad]:
sy = self.binary.reverse_symbols[e][0]
o._user_comment("; end function %s" % sy)
o._new_line()
o._new_line()
ad += i.size
else:
if o.is_symbol(ad):
o._symbol(ad)
o._new_line()
o._address(ad)
o._db(s.read_byte(ad))
o._new_line()
ad += 1
l += 1
if (l >= lines and until == -1) or (ad == until and until != -1):
break
s = self.binary.get_section(ad)
if s is None:
# Get the next section, it's not mandatory that sections
# are consecutives !
s = self.binary.get_next_section(ad)
if s is None:
break
ad = s.start
if ad == until:
break
o.curr_section = s
if until in self.functions:
o._new_line()
# remove the last empty line
o.lines.pop(-1)
o.token_lines.pop(-1)
o.join_lines()
# TODO: move it in the analyzer
if self.binary.type == T_BIN_PE:
# TODO: if ret != 0 : database is modified
self.binary.pe_reverse_stripped_symbols(self, o.addr_line)
return o
def find_addr_before(self, ad):
l = 0
s = self.binary.get_section(ad)
while l < NB_LINES_TO_DISASM:
if self.mem.is_code(ad):
size = self.mem.code[ad][0]
l += 1
l -= size
else:
l += 1
if ad == s.start:
s = self.binary.get_prev_section(ad)
if s is None:
return ad
ad = s.end
ad -= 1
return ad
def dump_data_ascii(self, ctx, lines):
N = 128 # read by block of 128 bytes
addr = ctx.entry_addr
s = self.binary.get_section(ctx.entry_addr)
s.print_header()
l = 0
ascii_str = []
addr_str = -1
while l < lines:
buf = s.read(addr, N)
if not buf:
break
i = 0
while i < len(buf):
if addr > s.end:
return
j = i
while j < len(buf):
c = buf[j]
if c not in BYTES_PRINTABLE_SET:
break
if addr_str == -1:
addr_str = addr
ascii_str.append(c)
j += 1
if c != 0 and j == len(buf):
addr += j - i
break
if c == 0 and len(ascii_str) >= 2:
print_no_end(color_addr(addr_str))
print_no_end(color_string(
"\"" + "".join(map(get_char, ascii_str)) + "\""))
print(", 0")
addr += j - i
i = j
else:
print_no_end(color_addr(addr))
print("0x%.2x " % buf[i])
addr += 1
i += 1
addr_str = -1
ascii_str = []
l += 1
if l >= lines:
return
def dump_data(self, ctx, lines, size_word):
s = self.binary.get_section(ctx.entry_addr)
s.print_header()
ad = ctx.entry_addr
for w in self.read_array(ctx.entry_addr, lines, size_word, s):
if ad in self.binary.reverse_symbols:
print(color_symbol(self.binary.reverse_symbols[ad][0]))
print_no_end(color_addr(ad))
print_no_end("0x%.2x" % w)
section = self.binary.get_section(w)
if section is not None:
print_no_end(" (")
print_no_end(color_section(section.name))
print_no_end(")")
if size_word >= 4 and w in self.binary.reverse_symbols:
print_no_end(" ")
print_no_end(color_symbol(self.binary.reverse_symbols[w][0]))
ad += size_word
print()
def print_calls(self, ctx):
ARCH = self.load_arch_module()
ARCH_UTILS = ARCH.utils
ARCH_OUTPUT = ARCH.output
s = self.binary.get_section(ctx.entry_addr)
s.print_header()
o = ARCH_OUTPUT.Output(ctx)
o._new_line()
ad = s.start
while ad < s.end:
i = self.lazy_disasm(ad, s.start)
if i is None:
ad += 1
else:
ad += i.size
if ARCH_UTILS.is_call(i):
o._asm_inst(i)
o.print()
#
# sym_filter : search a symbol, non case-sensitive
# if it starts with '-', it prints non-matching symbols
#
def print_symbols(self, print_sections, sym_filter=None, only_func=False):
if sym_filter is not None:
sym_filter = sym_filter.lower()
if sym_filter[0] == "-":
invert_match = True
sym_filter = sym_filter[1:]
else:
invert_match = False
total = 0
# TODO: race condition with the analyzer
for sy in list(self.binary.symbols):
addr, ty = self.binary.symbols[sy]
if only_func and ty != SYM_FUNC:
continue
if sym_filter is None or \
(invert_match and sym_filter not in sy.lower()) or \
(not invert_match and sym_filter in sy.lower()):
if sy:
section = self.binary.get_section(addr)
print_no_end(color_addr(addr) + " " + sy)
if print_sections and section is not None:
print_no_end(" (" + color_section(section.name) + ")")
print()
total += 1
print("Total:", total)
def lazy_disasm(self, addr, stay_in_section=-1, s=None):
s = self.binary.get_section(addr)
if s is None:
return None
# if stay_in_section != -1 and s.start != stay_in_section:
# return None, s
if addr in self.capstone_inst:
return self.capstone_inst[addr]
# TODO: remove when it's too big ?
if len(self.capstone_inst) > CAPSTONE_CACHE_SIZE:
self.capstone_inst.clear()
# Disassemble by block of N bytes
N = 128
d = s.read(addr, N)
gen = self.md.disasm(d, addr)
try:
first = next(gen)
except StopIteration:
return None
self.capstone_inst[first.address] = first
for i in gen:
if i.address in self.capstone_inst:
break
self.capstone_inst[i.address] = i
return first
def __prefetch_inst(self, inst):
return self.lazy_disasm(inst.address + inst.size)
# Generate a flow graph of the given function (addr)
def get_graph(self, entry_addr):
from capstone import CS_OP_IMM, CS_ARCH_MIPS
ARCH_UTILS = self.load_arch_module().utils
gph = Graph(self, entry_addr)
stack = [entry_addr]
start = time()
prefetch = None
addresses = set()
# WARNING: this assume that on every architectures the jump
# address is the last operand (operands[-1])
# Here each instruction is a node. Blocks will be created in the
# function __simplify.
while stack:
ad = stack.pop()
inst = self.lazy_disasm(ad)
if inst is None:
# Remove all previous instructions which have a link
# to this instruction.
if ad in gph.link_in:
for i in gph.link_in[ad]:
gph.link_out[i].remove(ad)
for i in gph.link_in[ad]:
if not gph.link_out[i]:
del gph.link_out[i]
del gph.link_in[ad]
continue
if gph.exists(inst):
continue
addresses.add(ad)
if ARCH_UTILS.is_ret(inst):
if self.arch == CS_ARCH_MIPS:
prefetch = self.__prefetch_inst(inst)
addresses.add(prefetch.address)
gph.new_node(inst, prefetch, None)
elif ARCH_UTILS.is_uncond_jump(inst):
if self.arch == CS_ARCH_MIPS:
prefetch = self.__prefetch_inst(inst)
addresses.add(prefetch.address)
gph.uncond_jumps_set.add(ad)
op = inst.operands[-1]
if op.type == CS_OP_IMM:
nxt = op.value.imm
stack.append(nxt)
gph.new_node(inst, prefetch, [nxt])
else:
if inst.address in self.jmptables:
table = self.jmptables[inst.address].table
stack += table
gph.new_node(inst, prefetch, table)
else:
# Can't interpret jmp ADDR|reg
gph.new_node(inst, prefetch, None)
elif ARCH_UTILS.is_cond_jump(inst):
if self.arch == CS_ARCH_MIPS:
prefetch = self.__prefetch_inst(inst)
addresses.add(prefetch.address)
gph.cond_jumps_set.add(ad)
op = inst.operands[-1]
if op.type == CS_OP_IMM:
if self.arch == CS_ARCH_MIPS:
direct_nxt = prefetch.address + prefetch.size
else:
direct_nxt = inst.address + inst.size
nxt_jmp = op.value.imm
stack.append(direct_nxt)
stack.append(nxt_jmp)
gph.new_node(inst, prefetch, [direct_nxt, nxt_jmp])
else:
# Can't interpret jmp ADDR|reg
gph.new_node(inst, prefetch, None)
else:
nxt = inst.address + inst.size
stack.append(nxt)
gph.new_node(inst, None, [nxt])
if len(gph.nodes) == 0:
return None, 0
if self.binary.type == T_BIN_PE:
nb_new_syms = self.binary.pe_reverse_stripped_symbols(self, addresses)
else:
nb_new_syms = 0
elapsed = time()
elapsed = elapsed - start
debug__("Graph built in %fs (%d instructions)" % (elapsed, len(gph.nodes)))
return gph, nb_new_syms
def add_jmptable(self, inst_addr, table_addr, entry_size, nb_entries):
name = self.add_symbol(table_addr, "jmptable_0x%x" % table_addr)
table = self.read_array(table_addr, nb_entries, entry_size)
self.jmptables[inst_addr] = Jmptable(inst_addr, table_addr, table, name)
self.internal_inline_comments[inst_addr] = "switch statement %s" % name
all_cases = {}
for ad in table:
all_cases[ad] = []
case = 0
for ad in table:
all_cases[ad].append(case)
case += 1
for ad in all_cases:
self.internal_previous_comments[ad] = \
["case %s %s" % (
", ".join(map(str, all_cases[ad])),
name
)]
class Mixer(object):
def __init__(self):
self.capture_view = CaptureView.instance()
self.memory = Memory()
self.cursor = Cursor()
self.setup_pages()
self.setup_name_table()
self.monitor = 'a'
self.channel = 1
self.set_page('input_monitor.' + self.monitor)
def setup_pages(self):
self.pages = {
"input_monitor.a": InputPage(self, 'a'),
"input_monitor.b": InputPage(self, 'b'),
"input_monitor.c": InputPage(self, 'c'),
"input_monitor.d": InputPage(self, 'd'),
"daw_monitor.a": OutputPage(self, 'a'),
"daw_monitor.b": OutputPage(self, 'b'),
"daw_monitor.c": OutputPage(self, 'c'),
"daw_monitor.d": OutputPage(self, 'd'),
"preamp": PreampPage(self),
"compressor": CompressorPage(self),
"line": LinePage(self),
"reverb": ReverbPage(self),
"patchbay": Patchbay(self),
}
for ch in range(0, 16):
self.pages |= {"channel.%d" % (ch + 1): ChannelPage(self, ch + 1)}
def setup_controls(self):
self.page = self.pages[self.page_name]
controls = []
self.header = self.page.get_header()
self.controls = self.page.get_controls()
def setup_name_table(self):
for name, page in self.pages.items():
for row in page.get_controls():
for control in row:
if control is None:
continue
self.capture_view.add_name_to_table(control)
def height(self):
return len(self.controls)
def width(self):
return len(self.controls[self.cursor.y])
def set_page(self, page):
self.page_name = page
self.setup_controls()
self.cursor.y = min(self.cursor.y, self.height() - 1)
self.cursor.x = min(self.cursor.x, self.width() - 1)
def set_monitor(self, m):
self.monitor = m
if 'monitor.' in self.page_name:
self.set_page(self.page_name[:-1] + self.monitor)
else:
print(self.page_name)
def set_channel(self, ch):
self.channel = ch
if 'channel.' in self.page_name:
self.set_page("channel.%d" % ch)
else:
print(self.page_name)
def cursor_down(self):
w = self.width()
if self.cursor.y + 1 < self.height():
self.cursor.y += 1
if self.width() >= w * 2:
self.cursor.x *= 2
def cursor_up(self):
w = self.width()
if self.cursor.y > 0:
self.cursor.y -= 1
if self.width() <= w // 2:
self.cursor.x //= 2
def cursor_left(self):
if self.cursor.x > 0:
self.cursor.x -= 1
def cursor_right(self):
if self.cursor.x + 1 < self.width():
self.cursor.x += 1
def get_selected_control(self):
row = self.cursor.y
col = self.cursor.x
return self.controls[row][col]
def get_selected_addr(self):
control = self.get_selected_control()
if control is None:
return None
return Capture.get_addr(control)
def get_memory_value(self, control):
addr = Capture.get_addr(control)
return self.memory.get_value(addr)
def decrement_selected(self):
addr = self.get_selected_addr()
data = self.memory.decrement(addr) if addr else None
return addr, data
def increment_selected(self):
addr = self.get_selected_addr()
data = self.memory.increment(addr) if addr else None
return addr, data
def zero_selected(self):
addr = self.get_selected_addr()
data = self.memory.zero(addr) if addr else None
return addr, data
def set_memory_value(self, name, value):
addr = Capture.get_addr(name)
data = self.memory.set_value(addr, value)
return addr, data
import unittest
from math import inf
from lib.memory import Memory
from lib.types import Volume
mem = Memory()
class TestMemory(unittest.TestCase):
@classmethod
def setUpClass(cls):
cls.mem = mem
cls.mem.capture_view.add_name_to_table(
'input_monitor.a.channel.1.volume')
def test_memory_get_empty(self):
self.assertEqual(None, self.mem.get(0))
def test_memory_set_get_erase(self):
self.mem.set(0, 1)
self.assertEqual(mem.get(0), 1)
self.mem.set(1, 2)
self.assertEqual(mem.get(0), 1)
self.assertEqual(mem.get(1), 2)
self.mem.set(0, 3)
self.assertEqual(mem.get(0), 3)
self.assertEqual(mem.get(1), 2)
self.mem.erase(0)
self.mem.erase(1)
self.assertIsNone(self.mem.get(0))
class SacAgent(object):
actor_store_dir = 'actor'
q_net_1_store_dir = 'q_1'
q_net_2_store_dir = 'q_2'
def __init__(self, env, batch_size):
self.batch_size = batch_size
self.tau = 1e-2
memory_size = 1000000
self.gamma = 0.99
self.q_lr = 3e-4
self.actor_lr = 3e-4
self.alpha_lr = 3e-3
self.update_step = 0
self.delay_step = 2
self.action_range = [env.action_space.low, env.action_space.high]
self.memory = Memory(memory_size)
# entropy temperature
self.alpha = 0.2
self.target_entropy = -torch.prod(torch.Tensor(
env.action_space.shape)).item()
self.log_alpha = torch.zeros(1, requires_grad=True)
self.alpha_optim = optim.Adam([self.log_alpha], lr=self.alpha_lr)
self.actor = SacActor(env.observation_space.shape[0],
env.action_space.shape[0])
self.actor_optimizer = optim.Adam(self.actor.parameters(),
lr=self.actor_lr)
self.q_net_1 = Critic(env.observation_space.shape[0],
env.action_space.shape[0])
self.q_net_1_target = Critic(env.observation_space.shape[0],
env.action_space.shape[0])
self.copy_networks(self.q_net_1, self.q_net_1_target)
self.q_net_1_optimizer = optim.Adam(self.q_net_1.parameters(),
lr=self.q_lr)
self.q_net_2 = Critic(env.observation_space.shape[0],
env.action_space.shape[0])
self.q_net_2_target = Critic(env.observation_space.shape[0],
env.action_space.shape[0])
self.copy_networks(self.q_net_2, self.q_net_2_target)
self.q_net_2_optimizer = optim.Adam(self.q_net_2.parameters(),
lr=self.q_lr)
def copy_networks(self, org_net, dest_net):
for dest_param, param in zip(dest_net.parameters(),
org_net.parameters()):
dest_param.data.copy_(param.data)
def get_test_action(self, state):
# 100% deterministic. It is not always the best option to do it this way
state = torch.FloatTensor(state).unsqueeze(0)
mean, log_std = self.actor.forward(state)
action = torch.tanh(mean)
action = action.detach().squeeze(0).numpy()
return self.rescale_action(action)
def get_action(self, state):
state = torch.FloatTensor(state).unsqueeze(0)
action, log_pi = self.actor.sample(state)
action = action.detach().squeeze(0).numpy()
return self.rescale_action(action)
def rescale_action(self, action):
return action * (self.action_range[1] - self.action_range[0]) / 2.0 +\
(self.action_range[1] + self.action_range[0]) / 2.0
def save(self, state, action, reward, new_state, cost, fail):
self.memory.push(state, action, reward, new_state, cost, fail)
def save_model(self, data_dir):
actor_dir = os.path.join(data_dir, self.actor_store_dir)
torch.save(self.actor, actor_dir)
q_net_1_dir = os.path.join(data_dir, self.q_net_1_store_dir)
torch.save(self.q_net_1, q_net_1_dir)
q_net_2_dir = os.path.join(data_dir, self.q_net_2_store_dir)
torch.save(self.q_net_2, q_net_2_dir)
def load_model(self, data_dir):
actor_dir = os.path.join(data_dir, self.actor_store_dir)
self.actor = torch.load(actor_dir)
q_net_1_dir = os.path.join(data_dir, self.q_net_1_store_dir)
self.q_net_1 = torch.load(q_net_1_dir)
self.copy_networks(self.q_net_1, self.q_net_1_target)
q_net_2_dir = os.path.join(data_dir, self.q_net_2_store_dir)
self.q_net_2 = torch.load(q_net_2_dir)
self.copy_networks(self.q_net_2, self.q_net_2_target)
def update(self, num=1):
for _ in range(num):
self.__one_update()
def __one_update(self):
if (len(self.memory) < self.batch_size):
return
states, actions, rewards, next_states, costs, fails = self.memory.get_batch(
self.batch_size)
not_fails = (fails == 0)
next_actions, next_log_pi = self.actor.sample(next_states)
next_q_1 = self.q_net_1_target(next_states, next_actions)
next_q_2 = self.q_net_2_target(next_states, next_actions)
next_q_target = torch.min(next_q_1,
next_q_2) - self.alpha * next_log_pi
expected_q = rewards - costs + not_fails * self.gamma * next_q_target
curr_q_1 = self.q_net_1.forward(states, actions)
curr_q_2 = self.q_net_2.forward(states, actions)
q1_loss = F.mse_loss(curr_q_1, expected_q.detach())
q2_loss = F.mse_loss(curr_q_2, expected_q.detach())
self.q_net_1_optimizer.zero_grad()
q1_loss.backward()
self.q_net_1_optimizer.step()
self.q_net_2_optimizer.zero_grad()
q2_loss.backward()
self.q_net_2_optimizer.step()
# delayed update for policy network and target q networks
new_actions, log_pi = self.actor.sample(states)
if self.update_step % self.delay_step == 0:
min_q = torch.min(self.q_net_1.forward(states, new_actions),
self.q_net_2.forward(states, new_actions))
actor_loss = (self.alpha * log_pi - min_q).mean()
self.actor_optimizer.zero_grad()
actor_loss.backward()
self.actor_optimizer.step()
# target networks
for target_param, param in zip(self.q_net_1_target.parameters(),
self.q_net_1.parameters()):
target_param.data.copy_(self.tau * param +
(1 - self.tau) * target_param)
for target_param, param in zip(self.q_net_2_target.parameters(),
self.q_net_2.parameters()):
target_param.data.copy_(self.tau * param +
(1 - self.tau) * target_param)
# update temperature
alpha_loss = (self.log_alpha *
(-log_pi - self.target_entropy).detach()).mean()
self.alpha_optim.zero_grad()
alpha_loss.backward()
self.alpha_optim.step()
self.alpha = self.log_alpha.exp()
self.update_step += 1
class Disassembler():
def __init__(self, filename, raw_type, raw_base, raw_big_endian, database):
import capstone as CAPSTONE
self.capstone_inst = {} # capstone instruction cache
if database.loaded:
self.mem = database.mem
else:
self.mem = Memory()
database.mem = self.mem
self.binary = Binary(self.mem, filename, raw_type, raw_base,
raw_big_endian)
self.binary.load_section_names()
arch, mode = self.binary.get_arch()
if arch is None or mode is None:
raise ExcArch(self.binary.get_arch_string())
self.jmptables = database.jmptables
self.user_inline_comments = database.user_inline_comments
self.internal_inline_comments = database.internal_inline_comments
self.user_previous_comments = database.user_previous_comments
self.internal_previous_comments = database.internal_previous_comments
self.functions = database.functions
self.func_id = database.func_id
self.end_functions = database.end_functions
self.xrefs = database.xrefs
# TODO: is it a global constant or $gp can change during the execution ?
self.mips_gp = database.mips_gp
if database.loaded:
self.binary.symbols = database.symbols
self.binary.reverse_symbols = database.reverse_symbols
self.binary.imports = database.imports
else:
self.binary.load_symbols()
database.symbols = self.binary.symbols
database.reverse_symbols = self.binary.reverse_symbols
database.imports = self.binary.imports
self.capstone = CAPSTONE
self.md = CAPSTONE.Cs(arch, mode)
self.md.detail = True
self.arch = arch
self.mode = mode
for s in self.binary.iter_sections():
s.big_endian = self.mode & self.capstone.CS_MODE_BIG_ENDIAN
# TODO: useful ?
if not database.loaded:
self.mem.add(s.start, s.end, MEM_UNK)
def get_unpack_str(self, size_word):
if self.mode & self.capstone.CS_MODE_BIG_ENDIAN:
endian = ">"
else:
endian = "<"
if size_word == 1:
unpack_str = endian + "B"
elif size_word == 2:
unpack_str = endian + "H"
elif size_word == 4:
unpack_str = endian + "L"
elif size_word == 8:
unpack_str = endian + "Q"
else:
return None
return unpack_str
def add_xref(self, from_ad, to_ad):
if isinstance(to_ad, list):
for x in to_ad:
if x in self.xrefs:
if from_ad not in self.xrefs[x]:
self.xrefs[x].append(from_ad)
else:
self.xrefs[x] = [from_ad]
else:
if to_ad in self.xrefs:
if from_ad not in self.xrefs[to_ad]:
self.xrefs[to_ad].append(from_ad)
else:
self.xrefs[to_ad] = [from_ad]
def add_symbol(self, ad, name):
if name in self.binary.symbols:
last = self.binary.symbols[name]
del self.binary.reverse_symbols[last]
if ad in self.binary.reverse_symbols:
last = self.binary.reverse_symbols[ad]
del self.binary.symbols[last]
self.binary.symbols[name] = ad
self.binary.reverse_symbols[ad] = name
if not self.mem.exists(ad):
self.mem.add(ad, 1, MEM_UNK)
return name
# TODO: create a function in SectionAbs
def read_array(self, ad, array_max_size, size_word, s=None):
unpack_str = self.get_unpack_str(size_word)
N = size_word * array_max_size
if s is None:
s = self.binary.get_section(ad)
array = []
l = 0
while l < array_max_size:
buf = s.read(ad, N)
if not buf:
break
i = 0
while i < len(buf):
b = buf[i:i + size_word]
if ad > s.end or len(b) != size_word:
return array
w = struct.unpack(unpack_str, b)[0]
array.append(w)
ad += size_word
i += size_word
l += 1
if l >= array_max_size:
return array
return array
def load_arch_module(self):
if self.arch == self.capstone.CS_ARCH_X86:
import lib.arch.x86 as ARCH
elif self.arch == self.capstone.CS_ARCH_ARM:
import lib.arch.arm as ARCH
elif self.arch == self.capstone.CS_ARCH_MIPS:
import lib.arch.mips as ARCH
else:
raise NotImplementedError
return ARCH
def dump_xrefs(self, ctx, ad):
ARCH = self.load_arch_module()
ARCH_OUTPUT = ARCH.output
o = ARCH_OUTPUT.Output(ctx)
o._new_line()
o.print_labels = False
for x in ctx.gctx.dis.xrefs[ad]:
s = self.binary.get_section(x)
if self.mem.is_code(x):
func_id = self.mem.get_func_id(x)
if func_id != -1:
fad = self.func_id[func_id]
o._label(fad)
diff = x - fad
if diff >= 0:
o._add(" + %d " % diff)
else:
o._add(" - %d " % (-diff))
o._pad_width(20)
i = self.lazy_disasm(x, s.start)
o._asm_inst(i)
else:
o._address(x)
o._new_line()
# remove the last empty line
o.lines.pop(-1)
o.token_lines.pop(-1)
o.join_lines()
return o
def is_label(self, ad):
return ad in self.binary.reverse_symbols or ad in self.xrefs
def get_symbol(self, ad):
s = self.binary.reverse_symbols.get(ad, None)
if s is None:
ty = self.mem.get_type(ad)
if ty == MEM_FUNC:
return "sub_%x" % ad
if ty == MEM_CODE:
return "loc_%x" % ad
if ty == MEM_UNK:
return "unk_%x" % ad
return s
def dump_asm(self, ctx, lines=NB_LINES_TO_DISASM, until=-1):
ARCH = self.load_arch_module()
ARCH_OUTPUT = ARCH.output
ad = ctx.entry
s = self.binary.get_section(ad)
if s is None:
# until is != -1 only from the visual mode
# It allows to not go before the first section.
if until != -1:
return None
# Get the next section, it's not mandatory that sections
# are consecutives !
s = self.binary.get_next_section(ad)
if s is None:
return None
ad = s.start
o = ARCH_OUTPUT.Output(ctx)
o._new_line()
o.section_prefix = True
o.curr_section = s
o.mode_dump = True
l = 0
while 1:
if ad == s.start:
o._new_line()
o._dash()
o._section(s.name)
o._add(" 0x%x -> 0x%x" % (s.start, s.end))
o._new_line()
o._new_line()
while ((l < lines and until == -1) or (ad < until and until != -1)) \
and ad <= s.end:
# A PE import should not be displayed as a subroutine
if not(self.binary.type == T_BIN_PE and ad in self.binary.imports) \
and self.mem.is_code(ad):
is_func = ad in self.functions and self.functions[ad][
0] != -1
if is_func:
if not o.is_last_2_line_empty():
o._new_line()
o._dash()
o._user_comment("; SUBROUTINE")
o._new_line()
o._dash()
i = self.lazy_disasm(ad, s.start)
if not is_func and ad in self.xrefs and \
not o.is_last_2_line_empty():
o._new_line()
o._asm_inst(i)
if ad in self.end_functions:
for fad in self.end_functions[ad]:
sy = self.get_symbol(fad)
o._user_comment("; end function %s" % sy)
o._new_line()
o._new_line()
ad += i.size
else:
o._label_and_address(ad)
o.set_line(ad)
o._db(s.read_byte(ad))
o._new_line()
ad += 1
l += 1
if (l >= lines and until == -1) or (ad >= until and until != -1):
break
s = self.binary.get_section(ad)
if s is None:
# Get the next section, it's not mandatory that sections
# are consecutives !
s = self.binary.get_next_section(ad)
if s is None:
break
ad = s.start
if until != -1 and ad >= until:
break
o.curr_section = s
if until in self.functions:
o._new_line()
# remove the last empty line
o.lines.pop(-1)
o.token_lines.pop(-1)
o.join_lines()
return o
def find_addr_before(self, ad):
l = 0
s = self.binary.get_section(ad)
while l < NB_LINES_TO_DISASM:
if self.mem.is_code(ad):
size = self.mem.mm[ad][0]
l += 1
l -= size
else:
l += 1
if ad == s.start:
s = self.binary.get_prev_section(ad)
if s is None:
return ad
ad = s.end
ad -= 1
return ad
def dump_data_ascii(self, ctx, lines):
N = 128 # read by block of 128 bytes
ad = ctx.entry
s = self.binary.get_section(ad)
print(hex(ad))
s.print_header()
l = 0
ascii_str = []
ad_str = -1
while l < lines:
buf = s.read(ad, N)
if not buf:
break
i = 0
while i < len(buf):
if ad > s.end:
return
j = i
while j < len(buf):
c = buf[j]
if c not in BYTES_PRINTABLE_SET:
break
if ad_str == -1:
ad_str = ad
ascii_str.append(c)
j += 1
if c != 0 and j == len(buf):
ad += j - i
break
if c == 0 and len(ascii_str) >= 2:
if self.is_label(ad_str):
print(color_symbol(self.get_symbol(ad_str)))
print_no_end(color_addr(ad_str))
print_no_end(
color_string("\"" + "".join(map(get_char, ascii_str)) +
"\""))
print(", 0")
ad += j - i
i = j
else:
if self.is_label(ad):
print(color_symbol(self.get_symbol(ad)))
print_no_end(color_addr(ad))
print("0x%.2x " % buf[i])
ad += 1
i += 1
ad_str = -1
ascii_str = []
l += 1
if l >= lines:
return
def dump_data(self, ctx, lines, size_word):
ad = ctx.entry
s = self.binary.get_section(ad)
s.print_header()
for w in self.read_array(ad, lines, size_word, s):
if self.is_label(ad):
print(color_symbol(self.get_symbol(ad)))
print_no_end(color_addr(ad))
print_no_end("0x%.2x" % w)
section = self.binary.get_section(w)
if section is not None:
print_no_end(" (")
print_no_end(color_section(section.name))
print_no_end(")")
if size_word >= 4 and self.is_label(w):
print_no_end(" ")
print_no_end(color_symbol(self.get_symbol(w)))
ad += size_word
print()
def print_functions(self):
total = 0
# TODO: race condition with the analyzer ?
for ad in list(self.functions):
print(color_addr(ad) + " " + self.get_symbol(ad))
total += 1
print("Total:", total)
#
# sym_filter : search a symbol, non case-sensitive
# if it starts with '-', it prints non-matching symbols
#
def print_symbols(self, print_sections, sym_filter=None):
if sym_filter is not None:
sym_filter = sym_filter.lower()
if sym_filter[0] == "-":
invert_match = True
sym_filter = sym_filter[1:]
else:
invert_match = False
total = 0
# TODO: race condition with the analyzer ?
for sy in list(self.binary.symbols):
ad = self.binary.symbols[sy]
if sym_filter is None or \
(invert_match and sym_filter not in sy.lower()) or \
(not invert_match and sym_filter in sy.lower()):
if sy:
section = self.binary.get_section(ad)
print_no_end(color_addr(ad) + " " + sy)
if print_sections and section is not None:
print_no_end(" (" + color_section(section.name) + ")")
print()
total += 1
print("Total:", total)
def lazy_disasm(self, ad, stay_in_section=-1, s=None):
s = self.binary.get_section(ad)
if s is None:
return None
# if stay_in_section != -1 and s.start != stay_in_section:
# return None, s
if ad in self.capstone_inst:
return self.capstone_inst[ad]
# TODO: remove when it's too big ?
if len(self.capstone_inst) > CAPSTONE_CACHE_SIZE:
self.capstone_inst.clear()
# Disassemble by block of N bytes
N = 128
d = s.read(ad, N)
gen = self.md.disasm(d, ad)
try:
first = next(gen)
except StopIteration:
return None
self.capstone_inst[first.address] = first
for i in gen:
if i.address in self.capstone_inst:
break
self.capstone_inst[i.address] = i
return first
def __add_prefetch(self, addr_set, inst):
if self.arch == self.CS_ARCH_MIPS:
prefetch = self.lazy_disasm(inst.address + inst.size)
addr_set.add(prefetch.address)
return prefetch
return None
def is_noreturn(self, ad):
return self.functions[ad][1] & FUNC_FLAG_NORETURN
# Generate a flow graph of the given function (addr)
def get_graph(self, entry):
from capstone import CS_OP_IMM, CS_ARCH_MIPS
self.CS_ARCH_MIPS = CS_ARCH_MIPS
ARCH_UTILS = self.load_arch_module().utils
gph = Graph(self, entry)
stack = [entry]
start = time()
prefetch = None
addresses = set()
# WARNING: this assume that on every architectures the jump
# address is the last operand (operands[-1])
# Here each instruction is a node. Blocks will be created in the
# function __simplify.
while stack:
ad = stack.pop()
inst = self.lazy_disasm(ad)
if inst is None:
# Remove all previous instructions which have a link
# to this instruction.
if ad in gph.link_in:
for i in gph.link_in[ad]:
gph.link_out[i].remove(ad)
for i in gph.link_in[ad]:
if not gph.link_out[i]:
del gph.link_out[i]
del gph.link_in[ad]
continue
if gph.exists(inst):
continue
addresses.add(ad)
if ARCH_UTILS.is_ret(inst):
prefetch = self.__add_prefetch(addresses, inst)
gph.new_node(inst, prefetch, None)
elif ARCH_UTILS.is_uncond_jump(inst):
prefetch = self.__add_prefetch(addresses, inst)
gph.uncond_jumps_set.add(ad)
op = inst.operands[-1]
if op.type == CS_OP_IMM:
nxt = op.value.imm
if nxt in self.functions:
gph.new_node(inst, prefetch, None)
else:
stack.append(nxt)
gph.new_node(inst, prefetch, [nxt])
else:
if inst.address in self.jmptables:
table = self.jmptables[inst.address].table
stack += table
gph.new_node(inst, prefetch, table)
else:
# Can't interpret jmp ADDR|reg
gph.new_node(inst, prefetch, None)
elif ARCH_UTILS.is_cond_jump(inst):
prefetch = self.__add_prefetch(addresses, inst)
gph.cond_jumps_set.add(ad)
op = inst.operands[-1]
if op.type == CS_OP_IMM:
if prefetch is None:
direct_nxt = inst.address + inst.size
else:
direct_nxt = prefetch.address + prefetch.size
nxt_jmp = op.value.imm
stack.append(direct_nxt)
if nxt_jmp in self.functions:
gph.new_node(inst, prefetch, [direct_nxt])
else:
stack.append(nxt_jmp)
gph.new_node(inst, prefetch, [direct_nxt, nxt_jmp])
else:
# Can't interpret jmp ADDR|reg
gph.new_node(inst, prefetch, None)
else:
if ad != entry and ARCH_UTILS.is_call(inst):
op = inst.operands[0]
if op.type == CS_OP_IMM:
imm = op.value.imm
if imm in self.functions and self.is_noreturn(imm):
prefetch = self.__add_prefetch(addresses, inst)
gph.new_node(inst, prefetch, None)
continue
nxt = inst.address + inst.size
stack.append(nxt)
gph.new_node(inst, None, [nxt])
if len(gph.nodes) == 0:
return None, 0
if self.binary.type == T_BIN_PE:
nb_new_syms = self.binary.pe_reverse_stripped_list(self, addresses)
else:
nb_new_syms = 0
elapsed = time()
elapsed = elapsed - start
debug__("Graph built in %fs (%d instructions)" %
(elapsed, len(gph.nodes)))
return gph, nb_new_syms
def add_jmptable(self, inst_addr, table_addr, entry_size, nb_entries):
name = self.add_symbol(table_addr, "jmptable_%x" % table_addr)
table = self.read_array(table_addr, nb_entries, entry_size)
self.jmptables[inst_addr] = Jmptable(inst_addr, table_addr, table,
name)
self.internal_inline_comments[inst_addr] = "switch statement %s" % name
all_cases = {}
for ad in table:
all_cases[ad] = []
case = 0
for ad in table:
all_cases[ad].append(case)
case += 1
for ad in all_cases:
self.internal_previous_comments[ad] = \
["case %s %s" % (
", ".join(map(str, all_cases[ad])),
name
)]
class DdpgAgent(object):
actor_store_dir = 'actor'
critic_store_dir = 'critic'
def __init__(self, env, batch_size):
self.batch_size = batch_size
self.tau = 1e-2
memory_size = 1000000
self.gamma = 0.99
actor_learning_rate = 1e-4
critic_learning_rate = 1e-3
self.critic_loss_fn = nn.MSELoss()
self.actor = DdpgActor(env.observation_space.shape[0],
env.action_space.shape[0],
env.action_space.high, env.action_space.low)
self.actor_target = DdpgActor(env.observation_space.shape[0],
env.action_space.shape[0],
env.action_space.high,
env.action_space.low)
self.copy_networks(self.actor, self.actor_target)
self.critic = Critic(env.observation_space.shape[0],
env.action_space.shape[0])
self.critic_target = Critic(env.observation_space.shape[0],
env.action_space.shape[0])
self.copy_networks(self.critic, self.critic_target)
self.memory = Memory(memory_size)
self.actor_optimizer = optim.Adam(self.actor.parameters(),
lr=actor_learning_rate)
self.critic_optimizer = optim.Adam(self.critic.parameters(),
lr=critic_learning_rate)
def copy_networks(self, org_net, dest_net):
for dest_param, param in zip(dest_net.parameters(),
org_net.parameters()):
dest_param.data.copy_(param.data)
def get_action(self, state):
tensor_state = Variable(torch.from_numpy(state).float().unsqueeze(0))
tensor_action = self.actor.noisy_forward(tensor_state)
#tensor_action = self.actor.forward(tensor_state)
return tensor_action.detach().numpy()[0]
def get_test_action(self, state):
tensor_state = Variable(torch.from_numpy(state).float().unsqueeze(0))
tensor_action = self.actor.forward(tensor_state)
return tensor_action.detach().numpy()[0]
def save(self, state, action, reward, new_state, cost, fail):
self.memory.push(state, action, reward, new_state, cost, fail)
def save_model(self, data_dir):
actor_dir = os.path.join(data_dir, self.actor_store_dir)
torch.save(self.actor, actor_dir)
critic_dir = os.path.join(data_dir, self.critic_store_dir)
torch.save(self.critic, critic_dir)
def load_model(self, data_dir):
actor_dir = os.path.join(data_dir, self.actor_store_dir)
self.actor = torch.load(actor_dir)
self.copy_networks(self.actor, self.actor_target)
critic_dir = os.path.join(data_dir, self.critic_store_dir)
self.critic = torch.load(critic_dir)
self.copy_networks(self.critic, self.critic_target)
def update(self, num=1):
for _ in range(num):
self.__one_update()
self.actor.reset_noise()
def __one_update(self):
if (len(self.memory) < self.batch_size):
return
states, actions, rewards, next_states, costs, fails = self.memory.get_batch(
self.batch_size)
states_q_values = self.critic.forward(states, actions)
next_actions = self.actor_target.forward(next_states)
next_states_q_value = self.critic_target.forward(
next_states, next_actions.detach())
not_fails = (fails == 0)
next_states_q_value = next_states_q_value * not_fails
new_q_value = rewards - costs + (self.gamma * next_states_q_value)
critic_loss = self.critic_loss_fn(states_q_values, new_q_value)
actor_loss = -self.critic.forward(states,
self.actor.forward(states)).mean()
self.actor_optimizer.zero_grad()
actor_loss.backward()
self.actor_optimizer.step()
self.critic_optimizer.zero_grad()
critic_loss.backward()
self.critic_optimizer.step()
for target_param, param in zip(self.actor_target.parameters(),
self.actor.parameters()):
target_param.data.copy_(param.data * self.tau + target_param.data *
(1.0 - self.tau))
for target_param, param in zip(self.critic_target.parameters(),
self.critic.parameters()):
target_param.data.copy_(param.data * self.tau + target_param.data *
(1.0 - self.tau))
def setUp(self):
self.memory = Memory(1)
def main():
memory_bank = Memory(MEMORY_SIZE)
pong_game = Game(GAME_LENGTH, GAME_STEP_TIME)
champion = Network(3,
7,
hidden_layer_size=HIDDEN_LAYER_SIZE,
no_hidden_layers=NO_HIDDEN_LAYERS,
learning_rate=LEARNING_RATE)
competitor = Network(3,
7,
hidden_layer_size=HIDDEN_LAYER_SIZE,
no_hidden_layers=NO_HIDDEN_LAYERS)
trainer = Trainer(pong_game,
memory_bank,
champion,
competitor,
MAX_EPSILON,
MIN_EPSILON,
EPSILON_DECAY,
GAMMA,
RETURNS_DECAY,
WINNERS_GROWTH,
batch_size=BATCH_SIZE)
champion.save_network(DIRECTORY + '/version_' + str(STARTING_VERSION))
for version in range(STARTING_VERSION,
STARTING_VERSION + NUMBER_OF_TRAINING_SESSIONS):
start_time = time.time()
for _ in range(GAMES_PER_TRAINING_SESSION):
print('New game')
trainer.run_game()
trainer.game = Game(GAME_LENGTH, GAME_STEP_TIME)
print("Time taken for training session: ", time.time() - start_time)
champion.save_network(DIRECTORY + '/version_' + str(version + 1))
current_epsilon = trainer.epsilon
current_returns_parameter = trainer.returns_parameter
current_winners_parameter = trainer.winners_parameter
trainer = Trainer(Game(GAME_LENGTH, GAME_STEP_TIME),
memory_bank,
champion,
competitor,
current_epsilon,
MIN_EPSILON,
EPSILON_DECAY,
GAMMA,
RETURNS_DECAY,
WINNERS_GROWTH,
returns_parameter=current_returns_parameter,
winners_parameter=current_winners_parameter,
batch_size=BATCH_SIZE)
test_score = trainer.test_game()
if test_score < 0:
print('Competitor wins, score was ' + str(test_score))
competitor.save_network(DIRECTORY + '/competitor_save')
champion.load_network(DIRECTORY + '/competitor_save')
else:
print('Champion continues, score was ' + str(test_score))
new_competitor_version = random.randint(max(0, version - 5), version)
print('New competitor version: ' + str(new_competitor_version))
competitor.load_network(DIRECTORY + '/version_' +
str(new_competitor_version))
current_epsilon = trainer.epsilon
print('epsilon is ' + str(current_epsilon))
current_returns_parameter = trainer.returns_parameter
current_winners_parameter = trainer.winners_parameter
trainer = Trainer(Game(GAME_LENGTH, GAME_STEP_TIME),
memory_bank,
champion,
competitor,
current_epsilon,
MIN_EPSILON,
EPSILON_DECAY,
GAMMA,
RETURNS_DECAY,
WINNERS_GROWTH,
returns_parameter=current_returns_parameter,
winners_parameter=current_winners_parameter,
batch_size=BATCH_SIZE)
