Python distributions.Categorical方法代码示例

# 需要导入模块: from torch import distributions [as 别名]
# 或者: from torch.distributions import Categorical [as 别名]
def sample_rl(self, images, seq_len, *args):
        device = images.device
        b_s = images.size(0)
        state = self.init_state(b_s, device)
        out = None

        outputs = []
        log_probs = []
        for t in range(seq_len):
            out, state = self.step(t, state, out, images, None, *args, mode='feedback')
            distr = distributions.Categorical(logits=out)
            out = distr.sample()
            outputs.append(out)
            log_probs.append(distr.log_prob(out))

        return torch.cat([o.unsqueeze(1) for o in outputs], 1), torch.cat([o.unsqueeze(1) for o in log_probs], 1) 

# 需要导入模块: from torch import distributions [as 别名]
# 或者: from torch.distributions import Categorical [as 别名]
def forward(self, state, hidden=None):
    explore = random.random() < self.epsilon
    state = state.unsqueeze(0)
    hidden = hidden.unsqueeze(0) if hidden else None
    logits = self.policy(
      state, hidden=hidden
    )
    outputs = [None]
    if isinstance(logits, tuple):
      logits, outputs = logits

    action = logits.argmax(dim=1)

    if explore:
      logits = torch.ones_like(logits)
      logits = logits / logits.size(1)
      action = Categorical(logits=logits).sample()

    return action[0], logits[0], outputs[0] 

# 需要导入模块: from torch import distributions [as 别名]
# 或者: from torch.distributions import Categorical [as 别名]
def select_action(self, agent_inputs, avail_actions, t_env, test_mode=False):

        # Assuming agent_inputs is a batch of Q-Values for each agent bav
        self.epsilon = self.schedule.eval(t_env)

        if test_mode:
            # Greedy action selection only
            self.epsilon = 0.0

        # mask actions that are excluded from selection
        masked_q_values = agent_inputs.clone()
        masked_q_values[avail_actions == 0.0] = -float("inf")  # should never be selected!

        random_numbers = th.rand_like(agent_inputs[:, :, 0])
        pick_random = (random_numbers < self.epsilon).long()
        random_actions = Categorical(avail_actions.float()).sample().long()

        picked_actions = pick_random * random_actions + (1 - pick_random) * masked_q_values.max(dim=2)[1]
        return picked_actions 

# 需要导入模块: from torch import distributions [as 别名]
# 或者: from torch.distributions import Categorical [as 别名]
def select_action(self, ob_id, state):
        r"""
        This function is mostly borrowed from the Reinforcement Learning example.
        See https://github.com/pytorch/examples/tree/master/reinforcement_learning
        The main difference is that instead of keeping all probs in one list,
        the agent keeps probs in a dictionary, one key per observer.

        NB: no need to enforce thread-safety here as GIL will serialize
        executions.
        """
        state = torch.from_numpy(state).float().unsqueeze(0)
        probs = self.policy(state)
        m = Categorical(probs)
        action = m.sample()
        self.saved_log_probs[ob_id].append(m.log_prob(action))
        return action.item() 

# 需要导入模块: from torch import distributions [as 别名]
# 或者: from torch.distributions import Categorical [as 别名]
def select_action_batch(agent_rref, ob_id, state):
        r"""
        Batching select_action: In each step, the agent waits for states from
        all observers, and process them together. This helps to reduce the
        number of CUDA kernels launched and hence speed up amortized inference
        speed.
        """
        self = agent_rref.local_value()
        self.states[ob_id].copy_(state)
        future_action = self.future_actions.then(
            lambda future_actions: future_actions.wait()[ob_id].item()
        )

        with self.lock:
            self.pending_states -= 1
            if self.pending_states == 0:
                self.pending_states = len(self.ob_rrefs)
                probs = self.policy(self.states.cuda())
                m = Categorical(probs)
                actions = m.sample()
                self.saved_log_probs.append(m.log_prob(actions).t()[0])
                future_actions = self.future_actions
                self.future_actions = torch.futures.Future()
                future_actions.set_result(actions.cpu())
        return future_action 

# 需要导入模块: from torch import distributions [as 别名]
# 或者: from torch.distributions import Categorical [as 别名]
def forward(self, input, params=None):
        if params is None:
            params = OrderedDict(self.named_parameters())

        output = input
        for i in range(1, self.num_layers):
            output = F.linear(output,
                              weight=params['layer{0}.weight'.format(i)],
                              bias=params['layer{0}.bias'.format(i)])
            output = self.nonlinearity(output)

        logits = F.linear(output,
                          weight=params['layer{0}.weight'.format(self.num_layers)],
                          bias=params['layer{0}.bias'.format(self.num_layers)])

        return Categorical(logits=logits) 

# 需要导入模块: from torch import distributions [as 别名]
# 或者: from torch.distributions import Categorical [as 别名]
def sample_model_spec(self, num):
        """
        Sample model specs by number.
        :param num:
        :return: list, num x [architecture ]
        """
        alpha_topology = self.alpha_topology.detach().clone()
        alpha_ops = self.alpha_ops.detach().clone()
        sample_archs = []
        sample_ops = []
        with torch.no_grad():
            for i in range(self.num_intermediate_nodes):
                # align with topoligy weights
                probs = nn.functional.softmax(alpha_topology[: i+2, i], dim=0)
                sample_archs.append(Categorical(probs))
                probs_op = nn.functional.softmax(alpha_ops[:, i], dim=0)
                sample_ops.append(Categorical(probs_op))
            return self._sample_model_spec(num, sample_archs, sample_ops) 

# 需要导入模块: from torch import distributions [as 别名]
# 或者: from torch.distributions import Categorical [as 别名]
def sample_model_spec(self, num):
        """
        Override, sample the alpha via gumbel softmax instead of normal softmax.
        :param num:
        :return:
        """
        alpha_topology = self.alpha_topology.detach().clone()
        alpha_ops = self.alpha_ops.detach().clone()
        sample_archs = []
        sample_ops = []
        gumbel_dist = Gumbel(torch.tensor([.0]), torch.tensor([1.0]))
        with torch.no_grad():
            for i in range(self.num_intermediate_nodes):
                # align with topoligy weights
                probs = gumbel_softmax(alpha_topology[: i+2, i], self.temperature(), gumbel_dist)
                sample_archs.append(Categorical(probs))
                probs_op = gumbel_softmax(alpha_ops[:, i], self.temperature(), gumbel_dist)
                sample_ops.append(Categorical(probs_op))

            return self._sample_model_spec(num, sample_archs, sample_ops) 

# 需要导入模块: from torch import distributions [as 别名]
# 或者: from torch.distributions import Categorical [as 别名]
def sample_rl(self, statics, *args):
        device = statics[0].device
        b_s = statics[0].size(0)
        state = self.init_state(b_s, device)

        outputs = []
        log_probs = []
        for t in range(self.seq_len):
            prev_outputs = outputs[-1] if t > 0 else None
            outs, state = self.step(t, state, prev_outputs, statics, None, *args, mode='feedback')
            outputs.append([])
            log_probs.append([])
            for out in outs:
                distr = distributions.Categorical(logits=out)
                sample = distr.sample()
                outputs[-1].append(sample)
                log_probs[-1].append(distr.log_prob(sample))

        outputs = list(zip(*outputs))
        outputs = tuple(torch.cat([oo.unsqueeze(1) for oo in o], 1) for o in outputs)
        log_probs = list(zip(*log_probs))
        log_probs = tuple(torch.cat([oo.unsqueeze(1) for oo in o], 1) for o in log_probs)
        return outputs, log_probs 

# 需要导入模块: from torch import distributions [as 别名]
# 或者: from torch.distributions import Categorical [as 别名]
def gen_one_noTcond(self, eos_index, max_T):
        hidden = self.init_hidden(1)
        device = hidden[0].device
        
        last_rnn_outp = hidden[0][-1] # [1, C]
        generation = []

        for t in range(max_T):
            scores = self.output_embedding(last_rnn_outp) # [1, V]
            word_dist = Categorical(logits=scores)
            selected_index = word_dist.sample() # [1]

            if selected_index == eos_index:
                break

            generation.append(selected_index)
            inp_embeddings = self.input_embedding(selected_index) # [1, inp_E]
            last_rnn_outp, hidden = self.rnn(inp_embeddings[None, :, :], hidden)
            last_rnn_outp = last_rnn_outp[0]

        return torch.tensor(generation, dtype=torch.long, device=device) 

# 需要导入模块: from torch import distributions [as 别名]
# 或者: from torch.distributions import Categorical [as 别名]
def sample(self, time: int, outputs: torch.Tensor) -> torch.LongTensor:
        del time  # unused by sample_fn
        # Outputs are logits, we sample from the top-k candidates
        if not torch.is_tensor(outputs):
            raise TypeError(
                f"Expected outputs to be a single Tensor, got: {type(outputs)}")
        if self._softmax_temperature is None:
            logits = outputs
        else:
            logits = outputs / self._softmax_temperature

        logits = _top_k_logits(logits, k=self._top_k)

        sample_id_sampler = Categorical(logits=logits)
        sample_ids = sample_id_sampler.sample()

        return sample_ids 

# 需要导入模块: from torch import distributions [as 别名]
# 或者: from torch.distributions import Categorical [as 别名]
def pi_beta_sample(self, state, beta, action, **kwargs):
        # 1. obtain probabilities
        # note: detach is to block gradient
        beta_probs = beta(state.detach(), action=action)
        pi_probs = self.forward(state)

        # 2. probabilities -> categorical distribution.
        beta_categorical = Categorical(beta_probs)
        pi_categorical = Categorical(pi_probs)

        # 3. sample the actions
        # See this issue: https://github.com/awarebayes/RecNN/issues/7
        # usually it works like:
        # pi_action = pi_categorical.sample(); beta_action = beta_categorical.sample();
        # but changing the action_source to {pi: beta, beta: beta} can be configured to be:
        # pi_action = beta_categorical.sample(); beta_action = beta_categorical.sample();
        available_actions = {'pi': pi_categorical.sample(), 'beta': beta_categorical.sample()}
        pi_action = available_actions[self.action_source['pi']]
        beta_action = available_actions[self.action_source['beta']]

        # 4. calculate stuff we need
        pi_log_prob = pi_categorical.log_prob(pi_action)
        beta_log_prob = beta_categorical.log_prob(beta_action)

        return pi_log_prob, beta_log_prob, pi_probs 

# 需要导入模块: from torch import distributions [as 别名]
# 或者: from torch.distributions import Categorical [as 别名]
def sampling_decode_single(self, pz_dec_outs, u_enc_out, m_tm1, u_input_np, last_hidden, degree_input, bspan_index):
        decoded = []
        reward_sum = 0
        log_probs = []
        rewards = []
        bspan_index_np = np.array(bspan_index).reshape(-1, 1)
        for t in range(self.max_ts):
            # reward
            reward, finished = self.reward(m_tm1.data.view(-1), decoded, bspan_index)
            reward_sum += reward
            rewards.append(reward)
            if t == self.max_ts - 1:
                finished = True
            if finished:
                loss = self.finish_episode(log_probs, rewards)
                return loss
            # action
            proba, last_hidden, _ = self.m_decoder(pz_dec_outs, u_enc_out, u_input_np, m_tm1,
                                                   degree_input, last_hidden, bspan_index_np)
            proba = proba.squeeze(0)  # [B,V]
            dis = Categorical(proba)
            action = dis.sample()
            log_probs.append(dis.log_prob(action))
            mt_index = action.data.view(-1)
            decoded.append(mt_index.clone())

            for i in range(mt_index.size(0)):
                if mt_index[i] >= cfg.vocab_size:
                    mt_index[i] = 2  # unk

            m_tm1 = cuda_(Variable(mt_index).view(1, -1)) 

# 需要导入模块: from torch import distributions [as 别名]
# 或者: from torch.distributions import Categorical [as 别名]
def __init__(self, probs=None, logits=None, validate_args=None):
        self.categoricals = []
        if probs is None:
            probs = [None] * len(logits)
        elif logits is None:
            logits = [None] * len(probs)
        else:
            raise ValueError('Either probs or logits must be None')

        for sub_probs, sub_logits in zip(probs, logits):
            categorical = distributions.Categorical(probs=sub_probs, logits=sub_logits, validate_args=validate_args)
            self.categoricals.append(categorical) 

# 需要导入模块: from torch import distributions [as 别名]
# 或者: from torch.distributions import Categorical [as 别名]
def _sample_event(self, output, greedy=True, temperature=1.0):
        if greedy:
            return output.argmax(-1)
        else:
            output = output / temperature
            probs = self.output_fc_activation(output)
            return Categorical(probs).sample()