def _step(self, inputs, input_sizes, targets):
"""
Make a single gradient update. This is called by train() and should not
be called manually.
Parameters
----------
inputs:
inputs_sizes:
targets:
"""
output = self.model(inputs, input_sizes)
loss = self.criterion(output, targets.long())
loss = loss / inputs.size(0) # average the loss by minibatch
if self.distributed:
loss = loss.to(self.device)
loss_value = reduce_tensor(loss, self.world_size).item()
else:
loss_value = loss.item()
# Check to ensure valid loss was calculated
valid_loss, error = check_loss(loss, loss_value)
if valid_loss:
self.optimizer.zero_grad()
with amp.scale_loss(loss, self.optimizer) as scaled_loss:
scaled_loss.backward()
torch.nn.utils.clip_grad_norm_(amp.master_params(self.optimizer),
self.max_norm)
self.optimizer.step()
else:
print(error)
print('Skipping grad update')
loss_value = 0
return output, loss_value
def step(self, loss):
if self.distributed:
loss = loss.to(self.device)
loss_value = reduce_tensor(loss, self.world_size).item()
else:
loss_value = loss.item()
valid_loss, error = check_loss(loss, loss_value)
if valid_loss:
self.optimizer.zero_grad()
with amp.scale_loss(loss, self.optimizer) as scaled_loss:
scaled_loss.backward()
torch.nn.utils.clip_grad_norm_(amp.master_params(self.optimizer), self.max_norm)
self.optimizer.step()
else:
print(error)
print('Skipping grad update')
return False
self.avg_loss += loss_value
return True
def state():
board.populate()
utils.spawn_new()
while not utils.check_loss():
# print(utils.elapsed_time())
# thresh = round(utils.elapsed_time() % globs.delay, 2)
# print(thresh)
# if (utils.elapsed_time() % globs.delay == 0):
t = str(utils.elapsed_time())
# print(t)
drop_nums = globs.drop_lists[globs.current_level - 1]
if (any(x in t for x in drop_nums)):
Canvas.draw_board()
if (not globs.current_piece.collision()):
globs.current_piece.move_down()
utils.insert_piece(globs.current_piece)
else:
# utils.reset_piece(globs.preview_pc)
utils.de_select_piece(globs.current_piece)
utils.clear_board()
# print("DONE CLEARING")
# time.sleep(1)
utils.spawn_new()
# print("SPAWNED")
globs.num_placed += 1
utils.check_level()
# time.sleep(0.1)
# give player 0.2 seconds to make a last-second adjustment
if globs.current_piece.collision() and not globs.dropped:
Canvas.draw_board()
kb.actions(kb.get_dir(0.2))
# give player 0.5 seconds to slide piece once made contact
# TODO set timer, only stop accepting input once timer runs out
# reset timer when key is pressed
else:
kb.actions(kb.get_dir(0.05))
float_out = out.float() # ensure float32 for loss
#print(float_out.to('cpu'))
#break
loss = criterion(float_out.to('cpu'), targets, output_sizes,
target_sizes).to(device)
loss = loss / inputs.size(0) # average the loss by minibatchi
if args.distributed:
loss = loss.to(device)
loss_value = reduce_tensor(loss, args.world_size).item()
else:
loss_value = loss.item()
# Check to ensure valid loss was calculated
valid_loss, error = check_loss(loss, loss_value)
if valid_loss:
optimizer.zero_grad()
# compute gradient
with amp.scale_loss(loss, optimizer) as scaled_loss:
scaled_loss.backward()
torch.nn.utils.clip_grad_norm_(model.parameters(),
args.max_norm)
optimizer.step()
#if i%16 == 15:
# print('step')
# optimizer.step()
# optimizer.zero_grad()
else:
print(error)
