def _add_reg(self, model):
offset = self._check_bn_drop(model)
reg_layers = []
for k in self.reg_params.keys():
if k in ["all_l2", "all_l1"]:
l2_reg = False
if k == "all_l2":
l2_reg = True
num_lin_layers = int(((len(self.model) - 2) / 2) + 1)
j = 0
for i in range(num_lin_layers):
space = self.reg_params[k]
hyperp = sample_from(space)
reg_layers.append((j, hyperp, l2_reg))
j += 2 + offset
elif k.split('_', 1)[1] in ["l2", "l1"]:
layer_num = int(k.split('_', 1)[0])
layer_num += (layer_num // 2) * (offset)
l2_reg = True
if k.split('_', 1)[1] == "l1":
l2_reg = False
space = self.reg_params[k]
hyperp = sample_from(space)
reg_layers.append((layer_num, hyperp, l2_reg))
else:
pass
model.new_params["reg_layers"] = reg_layers
return reg_layers
def get_sequence(self, length):
"""Given a model (a `Grid`), return a sequence of observations and the corresponding states."""
H, W = self.shape
states, observations = [], []
for t in range(length):
# choose a random init state
if t == 0:
state = np.random.randint(H), np.random.randint(W)
else:
state = sample_from(self.get_neighbours(state))
o = sample_from(self.get_colors(state))
states.append(state)
observations.append(o)
return np.array(observations), states
def _get_optimizer(self, model):
lr = self.def_lr
name = self.def_optim
if "optim" in self.optim_params:
space = self.optim_params['optim']
name = sample_from(space)
if "lr" in self.optim_params:
space = self.optim_params['lr']
lr = sample_from(space)
if name == "sgd":
opt = SGD
elif name == "adam":
opt = Adam
model.new_params["optim"] = name
model.new_params["lr"] = lr
optim = opt(model.parameters(), lr=lr)
return optim
def run_config(self, model, num_iters):
"""
Train a particular hyperparameter configuration for a
given number of iterations and evaluate the loss on the
validation set.
For hyperparameters that have previously been evaluated,
resume from a previous checkpoint.
Args
----
- model: the mutated model to train.
- num_iters: an int indicating the number of iterations
to train the model for.
Returns
-------
- val_loss: the lowest validaton loss achieved.
"""
try:
ckpt = self._load_checkpoint(model.ckpt_name)
model.load_state_dict(ckpt['state_dict'])
except FileNotFoundError:
pass
model = model.to(self.device)
# parse reg params
reg_layers = self._add_reg(model)
# setup train loader
if self.data_loader is None:
self.batch_hyper = True
space = self.optim_params['batch_size']
batch_size = sample_from(space)
tqdm.write("batch size: {}".format(batch_size))
self.data_loader = get_train_valid_loader(
self.data_dir, self.args.name, batch_size,
self.args.valid_size, self.args.shuffle, **self.kwargs)
# training logic
min_val_loss = 999999
counter = 0
num_epochs = int(num_iters) if self.epoch_scale else 1
num_passes = None if self.epoch_scale else num_iters
for epoch in range(num_epochs):
self._train_one_epoch(model, num_passes, reg_layers)
val_loss = self._validate_one_epoch(model)
if val_loss < min_val_loss:
min_val_loss = val_loss
counter = 0
else:
counter += 1
if counter > self.patience:
tqdm.write("[!] early stopped!!")
model.early_stopped = True
return min_val_loss
if self.batch_hyper:
self.data_loader = None
state = {
'state_dict': model.state_dict(),
'min_val_loss': min_val_loss,
}
self._save_checkpoint(state, model.ckpt_name)
return min_val_loss
def get_random_config(self):
"""
Build a mutated version of the user's model that
incorporates the new hyperparameters settings defined
by `hyperparams`.
"""
self.all_batchnorm = False
self.all_drop = False
new_params = {}
if not self.net_params:
mutated = self.model
else:
layers = []
used_acts = []
all_act = False
all_drop = False
all_batchnorm = False
num_layers = len(self.model)
i = 0
used_acts.append(self.model[1].__str__())
for layer_hp in self.net_params.keys():
layer, hp = layer_hp.split('_', 1)
if layer.isdigit():
layer_num = int(layer)
diff = layer_num - i
if diff > 0:
for j in range(diff + 1):
layers.append(self.model[i + j])
i += diff
if hp == 'act':
space = find_key(self.net_params,
'{}_act'.format(layer_num))
hyperp = sample_from(space)
new_params["act"] = hyperp
new_act = str2act(hyperp)
used_acts.append(new_act.__str__())
layers.append(new_act)
i += 1
elif hp == 'dropout':
layers.append(self.model[i])
space = find_key(self.net_params,
'{}_drop'.format(layer_num))
hyperp = sample_from(space)
new_params["drop"] = hyperp
layers.append(nn.Dropout(p=hyperp))
else:
pass
elif diff == 0:
layers.append(self.model[i])
if hp == 'act':
space = find_key(self.net_params,
'{}_act'.format(layer_num))
hyperp = sample_from(space)
new_params["act"] = hyperp
new_act = str2act(hyperp)
used_acts.append(new_act.__str__())
layers.append(new_act)
i += 1
elif hp == 'dropout':
i += 1
layers.append(self.model[i])
space = find_key(self.net_params,
'{}_drop'.format(layer_num))
hyperp = sample_from(space)
new_params["drop"] = hyperp
layers.append(nn.Dropout(p=hyperp))
else:
pass
else:
if hp == 'act':
space = find_key(self.net_params,
'{}_act'.format(layer_num))
hyperp = sample_from(space)
new_params["act"] = hyperp
new_act = str2act(hyperp)
used_acts.append(new_act.__str__())
layers[i] = new_act
elif hp == 'dropout':
space = find_key(self.net_params,
'{}_drop'.format(layer_num))
hyperp = sample_from(space)
new_params["drop"] = hyperp
layers.append(nn.Dropout(p=hyperp))
layers.append(self.model[i])
else:
pass
i += 1
else:
if (i < num_layers) and (len(layers) < num_layers):
for j in range(num_layers - i):
layers.append(self.model[i + j])
i += 1
if layer == "all":
if hp == "act":
space = self.net_params['all_act']
hyperp = sample_from(space)
all_act = False if hyperp == [0] else True
elif hp == "dropout":
space = self.net_params['all_dropout']
hyperp = sample_from(space)
all_drop = False if hyperp == [0] else True
elif hp == "batchnorm":
space = self.net_params['all_batchnorm']
hyperp = sample_from(space)
all_batchnorm = True if hyperp == 1 else False
else:
pass
used_acts = sorted(set(used_acts), key=used_acts.index)
if all_act:
old_act = used_acts[0]
space = self.net_params['all_act'][1][1]
hyperp = sample_from(space)
new_params["all_act"] = hyperp
new_act = str2act(hyperp)
used_acts.append(new_act.__str__())
for i, l in enumerate(layers):
if l.__str__() == old_act:
layers[i] = new_act
if all_batchnorm:
self.all_batchnorm = True
new_params["all_batch"] = True
target_acts = used_acts if not all_act else used_acts[1:]
for i, l in enumerate(layers):
if l.__str__() in target_acts:
if 'Linear' in layers[i - 1].__str__():
bn = nn.BatchNorm2d(layers[i - 1].out_features)
else:
bn = nn.BatchNorm2d(layers[i - 1].out_channels)
layers.insert(i + 1, bn)
if 'Linear' in layers[-2].__str__():
bn = nn.BatchNorm2d(layers[i - 1].out_features)
else:
bn = nn.BatchNorm2d(layers[i - 1].out_channels)
layers.insert(-1, bn)
if all_drop:
self.all_drop = True
new_params["all_drop"] = True
target_acts = used_acts if not all_act else used_acts[1:]
space = self.net_params['all_dropout'][1][1]
hyperp = sample_from(space)
for i, l in enumerate(layers):
if l.__str__() in target_acts:
layers.insert(i + 1 + all_batchnorm,
nn.Dropout(p=hyperp))
sizes = {}
for k, v in self.size_params.items():
layer_num = int(k.split("_", 1)[0])
layer_num += (layer_num // 2) * (self.all_batchnorm +
self.all_drop)
hyperp = sample_from(v)
new_params["{}_hidden_size".format(layer_num)] = hyperp
sizes[layer_num] = hyperp
for layer, size in sizes.items():
in_dim = layers[layer].in_features
layers[layer] = nn.Linear(in_dim, size)
if self.all_batchnorm:
layers[layer + 2] = nn.BatchNorm2d(size)
next_layer = layer + (2 + self.all_batchnorm + self.all_drop)
out_dim = layers[next_layer].out_features
layers[next_layer] = nn.Linear(size, out_dim)
mutated = nn.Sequential(*layers)
self._init_weights_biases(mutated)
mutated.ckpt_name = str(uuid.uuid4().hex)
mutated.new_params = new_params
mutated.early_stopped = False
return mutated
def sample_observation(self, state: int) -> int:
return sample_from([(o, self.B[state, o]) for o in range(self.M)])
def sample_transition(self, from_state: int) -> int:
return sample_from([(s, self.A[from_state, s]) for s in range(self.N)])
def sample_initial(self) -> float:
return sample_from([(s, self.pi[s]) for s in range(self.N)])
