repo
stringlengths
1
99
file
stringlengths
13
215
code
stringlengths
12
59.2M
file_length
int64
12
59.2M
avg_line_length
float64
3.82
1.48M
max_line_length
int64
12
2.51M
extension_type
stringclasses
1 value
bonito
bonito-master/bonito/reader.py
""" Bonito Read Utils """ from glob import iglob from collections import OrderedDict from importlib import import_module import torch import numpy as np from scipy.signal import find_peaks __formats__ = ["fast5", "pod5"] # Normalisation parameters for kit 14 DNA # Different parameters can be specified in the 'normalisation' section # of a bonito config file __default_norm_params__ = {'quantile_a' : 0.2, 'quantile_b' : 0.9, 'shift_multiplier' : 0.51, 'scale_multiplier' : 0.53} class Reader: def __init__(self, directory, recursive=False): self.fmt = None for fmt in __formats__: pattern = f"**/*.{fmt}" if recursive else f"*.{fmt}" match = next(iglob(directory + "/" + pattern, recursive=True), None) if match is not None: self.fmt = fmt break else: raise FileNotFoundError() _reader = import_module(f"bonito.{self.fmt}") self._get_reads = getattr(_reader, "get_reads") self._get_read_groups = getattr(_reader, "get_read_groups") def get_reads(self, *args, **kwargs): return self._get_reads(*args, **kwargs) def get_read_groups(self, *args, **kwargs): return self._get_read_groups(*args, **kwargs) class Read: def __init__(self, read, filename, meta=False): ... def __repr__(self): return "Read('%s')" % self.read_id def readgroup(self, model): self._groupdict = OrderedDict([ ('ID', f"{self.run_id}_{model}"), ('PL', f"ONT"), ('DT', f"{self.exp_start_time}"), ('PU', f"{self.flow_cell_id}"), ('PM', f"{self.device_id}"), ('LB', f"{self.sample_id}"), ('SM', f"{self.sample_id}"), ('DS', f"%s" % ' '.join([ f"run_id={self.run_id}", f"basecall_model={model}", ])) ]) return '\t'.join(["@RG", *[f"{k}:{v}" for k, v in self._groupdict.items()]]) def tagdata(self): return [ f"mx:i:{self.mux}", f"ch:i:{self.channel}", f"st:Z:{self.start_time}", f"du:f:{self.duration}", f"rn:i:{self.read_number}", f"f5:Z:{self.filename}", f"sm:f:{self.shift}", f"sd:f:{self.scale}", f"sv:Z:quantile", ] class ReadChunk: def __init__(self, read, chunk, i, n): self.read_id = "%s:%i:%i" % (read.read_id, i, n) self.run_id = read.run_id self.filename = read.filename self.mux = read.mux self.channel = read.channel self.start = read.start self.duration = read.duration self.template_start = self.start self.template_duration = self.duration self.signal = chunk def __repr__(self): return "ReadChunk('%s')" % self.read_id def read_chunks(read, chunksize=4000, overlap=400): """ Split a Read in fixed sized ReadChunks """ if len(read.signal) < chunksize: return _, offset = divmod(len(read.signal) - chunksize, chunksize - overlap) signal = torch.from_numpy(read.signal[offset:]) blocks = signal.unfold(0, chunksize, chunksize - overlap) for i, block in enumerate(blocks): yield ReadChunk(read, block.numpy(), i+1, blocks.shape[0]) def trim(signal, window_size=40, threshold=2.4, min_trim=10, min_elements=3, max_samples=8000, max_trim=0.3): seen_peak = False num_windows = min(max_samples, len(signal)) // window_size for pos in range(num_windows): start = pos * window_size + min_trim end = start + window_size window = signal[start:end] if len(window[window > threshold]) > min_elements or seen_peak: seen_peak = True if window[-1] > threshold: continue if end >= min(max_samples, len(signal)) or end / len(signal) > max_trim: return min_trim return end return min_trim def normalisation(sig, norm_params=None): """ Calculate signal shift and scale factors for normalisation.. """ if norm_params is None: norm_params = __default_norm_params__ qa, qb = np.quantile(sig, [norm_params['quantile_a'], norm_params['quantile_b']]) shift = max(10, norm_params['shift_multiplier'] * (qa + qb)) scale = max(1.0, norm_params['scale_multiplier'] * (qb - qa)) return shift, scale
4,550
29.543624
109
py
bonito
bonito-master/bonito/training.py
""" Bonito train """ import math import os import re from glob import glob from functools import partial from time import perf_counter from collections import OrderedDict from datetime import datetime from bonito.schedule import linear_warmup_cosine_decay from bonito.util import accuracy, decode_ref, permute, concat, match_names import bonito import torch import numpy as np import torch.nn as nn from tqdm import tqdm import torch.cuda.amp as amp def load_state(dirname, device, model, optim=None): """ Load a model state dict from disk """ model.to(device) if hasattr(model, "module"): model = model.module weight_no = optim_no = None optim_files = glob(os.path.join(dirname, "optim_*.tar")) optim_nos = {int(re.sub(".*_([0-9]+).tar", "\\1", w)) for w in optim_files} weight_files = glob(os.path.join(dirname, "weights_*.tar")) weight_nos = {int(re.sub(".*_([0-9]+).tar", "\\1", w)) for w in weight_files} if optim is not None: weight_no = optim_no = max(optim_nos & weight_nos, default=None) else: weight_no = max(weight_nos, default=None) to_load = [] if weight_no: to_load.append(("weights", model)) if optim_no: to_load.append(("optim", optim)) if to_load: print("[picking up %s state from epoch %s]" % (', '.join([n for n, _ in to_load]), weight_no)) for name, obj in to_load: state_dict = torch.load( os.path.join(dirname, '%s_%s.tar' % (name, weight_no)), map_location=device ) if name == "weights": state_dict = {k2: state_dict[k1] for k1, k2 in match_names(state_dict, obj).items()} new_state_dict = OrderedDict() for k, v in state_dict.items(): name = k.replace('module.', '') new_state_dict[name] = v state_dict = new_state_dict obj.load_state_dict(state_dict) epoch = weight_no else: epoch = 0 return epoch class ClipGrad: def __init__(self, quantile=0.5, factor=2.0, buffer_size=100): self.buffer = np.full(buffer_size, fill_value=1e6) self.quantile = quantile self.factor = factor self.i = 0 def append(self, grad_norm): self.buffer[self.i] = grad_norm self.i = (self.i + 1) % len(self.buffer) def __call__(self, parameters): max_norm = self.factor * np.quantile(self.buffer, self.quantile) grad_norm = torch.nn.utils.clip_grad_norm_(parameters, max_norm=max_norm).item() if not math.isnan(grad_norm): self.append(grad_norm) return grad_norm class Trainer: def __init__( self, model, device, train_loader, valid_loader, criterion=None, use_amp=True, lr_scheduler_fn=None, restore_optim=False, save_optim_every=10, grad_accum_split=1, quantile_grad_clip=False ): self.model = model.to(device) self.device = device self.train_loader = train_loader self.valid_loader = valid_loader self.criterion = criterion or model.loss self.use_amp = use_amp self.lr_scheduler_fn = lr_scheduler_fn or linear_warmup_cosine_decay() self.restore_optim = restore_optim self.save_optim_every = save_optim_every self.grad_accum_split = grad_accum_split self.scaler = torch.cuda.amp.GradScaler(enabled=use_amp) self.optimizer = None if quantile_grad_clip: self.clip_grad = ClipGrad() else: self.clip_grad = lambda parameters: torch.nn.utils.clip_grad_norm_(parameters, max_norm=2.0).item() def train_one_step(self, batch): self.optimizer.zero_grad() losses = None with amp.autocast(enabled=self.use_amp): for batch_ in zip( *map(lambda t: t.chunk(self.grad_accum_split, dim=0), batch) ): data_, targets_, lengths_, *args = (x.to(self.device) for x in batch_) scores_ = self.model(data_, *args) losses_ = self.criterion(scores_, targets_, lengths_) if not isinstance(losses_, dict): losses_ = {'loss': losses_} total_loss = losses_.get('total_loss', losses_['loss']) / self.grad_accum_split self.scaler.scale(total_loss).backward() losses = { k: ((v.item() / self.grad_accum_split) if losses is None else (v.item() / self.grad_accum_split) + losses[k]) for k, v in losses_.items() } self.scaler.unscale_(self.optimizer) grad_norm = self.clip_grad(self.model.parameters()) self.scaler.step(self.optimizer) self.scaler.update() return losses, grad_norm def train_one_epoch(self, loss_log, lr_scheduler): t0 = perf_counter() chunks = 0 self.model.train() progress_bar = tqdm( total=len(self.train_loader), desc='[0/{}]'.format(len(self.train_loader.sampler)), ascii=True, leave=True, ncols=100, bar_format='{l_bar}{bar}| [{elapsed}{postfix}]' ) smoothed_loss = None with progress_bar: for batch in self.train_loader: chunks += batch[0].shape[0] losses, grad_norm = self.train_one_step(batch) smoothed_loss = losses['loss'] if smoothed_loss is None else (0.01 * losses['loss'] + 0.99 * smoothed_loss) progress_bar.set_postfix(loss='%.4f' % smoothed_loss) progress_bar.set_description("[{}/{}]".format(chunks, len(self.train_loader.sampler))) progress_bar.update() if loss_log is not None: lr = lr_scheduler.get_last_lr() if lr_scheduler is not None else [pg["lr"] for pg in optim.param_groups] if len(lr) == 1: lr = lr[0] loss_log.append({ 'chunks': chunks, 'time': perf_counter() - t0, 'grad_norm': grad_norm, 'lr': lr, **losses }) if lr_scheduler is not None: lr_scheduler.step() return smoothed_loss, perf_counter() - t0 def validate_one_step(self, batch): data, targets, lengths, *args = batch scores = self.model(data.to(self.device), *(x.to(self.device) for x in args)) losses = self.criterion(scores, targets.to(self.device), lengths.to(self.device)) losses = {k: v.item() for k, v in losses.items()} if isinstance(losses, dict) else losses.item() if hasattr(self.model, 'decode_batch'): seqs = self.model.decode_batch(scores) else: seqs = [self.model.decode(x) for x in permute(scores, 'TNC', 'NTC')] refs = [decode_ref(target, self.model.alphabet) for target in targets] n_pre = getattr(self.model, "n_pre_context_bases", 0) n_post = getattr(self.model, "n_post_context_bases", 0) if n_pre > 0 or n_post > 0: refs = [ref[n_pre:len(ref)-n_post] for ref in refs] accs = [ accuracy(ref, seq, min_coverage=0.5) if len(seq) else 0. for ref, seq in zip(refs, seqs) ] return seqs, refs, accs, losses def validate_one_epoch(self): self.model.eval() with torch.no_grad(): seqs, refs, accs, losses = zip(*(self.validate_one_step(batch) for batch in self.valid_loader)) seqs, refs, accs = (sum(x, []) for x in (seqs, refs, accs)) loss = np.mean([(x['loss'] if isinstance(x, dict) else x) for x in losses]) return loss, np.mean(accs), np.median(accs) def init_optimizer(self, lr, **kwargs): if isinstance(lr, (list, tuple)): if len(list(self.model.children())) != len(lr): raise ValueError('Number of lrs does not match number of model children') param_groups = [{'params': list(m.parameters()), 'lr': v} for (m, v) in zip(self.model.children(), lr)] self.optimizer = torch.optim.AdamW(param_groups, lr=lr[0], **kwargs) else: self.optimizer = torch.optim.AdamW(self.model.parameters(), lr=lr, **kwargs) def get_lr_scheduler(self, epochs, last_epoch=0): return self.lr_scheduler_fn(self.optimizer, self.train_loader, epochs, last_epoch) def fit(self, workdir, epochs=1, lr=2e-3, **optim_kwargs): if self.optimizer is None: self.init_optimizer(lr, **optim_kwargs) last_epoch = load_state(workdir, self.device, self.model, self.optimizer if self.restore_optim else None) if self.restore_optim: # override learning rate to new value for i, pg in enumerate(self.optimizer.param_groups): pg["initial_lr"] = pg["lr"] = lr[i] if isinstance(lr, (list, tuple)) else lr lr_scheduler = self.get_lr_scheduler(epochs, last_epoch=last_epoch) for epoch in range(1 + last_epoch, epochs + 1): try: with bonito.io.CSVLogger(os.path.join(workdir, 'losses_{}.csv'.format(epoch))) as loss_log: train_loss, duration = self.train_one_epoch(loss_log, lr_scheduler) model_state = self.model.module.state_dict() if hasattr(self.model, 'module') else self.model.state_dict() torch.save(model_state, os.path.join(workdir, "weights_%s.tar" % epoch)) if epoch % self.save_optim_every == 0: torch.save(self.optimizer.state_dict(), os.path.join(workdir, "optim_%s.tar" % epoch)) val_loss, val_mean, val_median = self.validate_one_epoch() except KeyboardInterrupt: break print("[epoch {}] directory={} loss={:.4f} mean_acc={:.3f}% median_acc={:.3f}%".format( epoch, workdir, val_loss, val_mean, val_median )) with bonito.io.CSVLogger(os.path.join(workdir, 'training.csv')) as training_log: training_log.append({ 'time': datetime.today(), 'duration': int(duration), 'epoch': epoch, 'train_loss': train_loss, 'validation_loss': val_loss, 'validation_mean': val_mean, 'validation_median': val_median })
10,481
38.258427
129
py
bonito
bonito-master/bonito/data.py
import importlib import os from pathlib import Path import numpy as np from torch.utils.data import DataLoader class ChunkDataSet: def __init__(self, chunks, targets, lengths): self.chunks = np.expand_dims(chunks, axis=1) self.targets = targets self.lengths = lengths def __getitem__(self, i): return ( self.chunks[i].astype(np.float32), self.targets[i].astype(np.int64), self.lengths[i].astype(np.int64), ) def __len__(self): return len(self.lengths) def load_script(directory, name="dataset", suffix=".py", **kwargs): directory = Path(directory) filepath = (directory / name).with_suffix(suffix) spec = importlib.util.spec_from_file_location(name, filepath) module = importlib.util.module_from_spec(spec) spec.loader.exec_module(module) loader = module.Loader(**kwargs) return loader.train_loader_kwargs(**kwargs), loader.valid_loader_kwargs(**kwargs) def load_numpy(limit, directory, valid_chunks=None): """ Returns training and validation DataLoaders for data in directory. """ train_data = load_numpy_datasets(limit=limit, directory=directory) if os.path.exists(os.path.join(directory, 'validation')): valid_data = load_numpy_datasets(limit=valid_chunks, directory=os.path.join(directory, 'validation') ) else: print("[validation set not found: splitting training set]") if valid_chunks is None: split = np.floor(len(train_data[0]) * 0.97).astype(np.int32) else: split = max(0, len(train_data[0]) - valid_chunks) valid_data = [x[split:] for x in train_data] train_data = [x[:split] for x in train_data] train_loader_kwargs = {"dataset": ChunkDataSet(*train_data), "shuffle": True} valid_loader_kwargs = {"dataset": ChunkDataSet(*valid_data), "shuffle": False} return train_loader_kwargs, valid_loader_kwargs def load_numpy_datasets(limit=None, directory=None): """ Returns numpy chunks, targets and lengths arrays. """ if directory is None: directory = default_data chunks = np.load(os.path.join(directory, "chunks.npy"), mmap_mode='r') targets = np.load(os.path.join(directory, "references.npy"), mmap_mode='r') lengths = np.load(os.path.join(directory, "reference_lengths.npy"), mmap_mode='r') indices = os.path.join(directory, "indices.npy") if os.path.exists(indices): idx = np.load(indices, mmap_mode='r') idx = idx[idx < lengths.shape[0]] if limit: idx = idx[:limit] return chunks[idx, :], targets[idx, :], lengths[idx] if limit: chunks = chunks[:limit] targets = targets[:limit] lengths = lengths[:limit] return np.array(chunks), np.array(targets), np.array(lengths)
2,868
32.752941
86
py
bonito
bonito-master/bonito/nn.py
""" Bonito nn modules. """ import torch from torch.nn import Module from torch.nn.init import orthogonal_ from torch.nn.utils.fusion import fuse_conv_bn_eval layers = {} def register(layer): layer.name = layer.__name__.lower() layers[layer.name] = layer return layer register(torch.nn.ReLU) register(torch.nn.Tanh) @register class Linear(Module): def __init__(self, in_features, out_features, bias=True): super().__init__() self.in_features = in_features self.out_features = out_features self.bias = bias self.linear = torch.nn.Linear( in_features=in_features, out_features=out_features, bias=bias ) def forward(self, x): return self.linear(x) def to_dict(self, include_weights=False): res = { "in_features": self.in_features, "out_features": self.out_features, "bias": self.bias, } if include_weights: res['params'] = { 'W': self.linear.weight, 'b': self.linear.bias if self.bias is not None else [] } return res @register class Swish(torch.nn.SiLU): pass @register class Clamp(Module): def __init__(self, min, max): super().__init__() self.min = min self.max = max def forward(self, x): return torch.clamp(x, min=self.min, max=self.max) def to_dict(self, include_weights=False): return { 'min': self.min, 'max': self.max } @register class Serial(torch.nn.Sequential): def __init__(self, sublayers): super().__init__(*sublayers) def forward(self, x, return_features=False): if return_features: fmaps = [] for layer in self: x = layer(x) fmaps.append(x) return x, fmaps return super().forward(x) def to_dict(self, include_weights=False): return { 'sublayers': [to_dict(layer, include_weights) for layer in self._modules.values()] } @register class Reverse(Module): def __init__(self, sublayers): super().__init__() self.layer = Serial(sublayers) if isinstance(sublayers, list) else sublayers def forward(self, x): return self.layer(x.flip(0)).flip(0) def to_dict(self, include_weights=False): if isinstance(self.layer, Serial): return self.layer.to_dict(include_weights) else: return {'sublayers': to_dict(self.layer, include_weights)} @register class BatchNorm(Module): def __init__(self, num_features, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True): super().__init__() self.bn = torch.nn.BatchNorm1d(num_features, eps, momentum, affine, track_running_stats) def forward(self, x): return self.bn(x) def to_dict(self, include_weights=False): res = { "num_features": self.bn.num_features, "eps": self.bn.eps, "momentum": self.bn.momentum, "affine": self.bn.affine, "track_running_stats": self.bn.track_running_stats } if include_weights: params = {} if res["affine"]: params["W"] = self.bn.weight params["b"] = self.bn.bias if res["track_running_stats"]: params["running_mean"] = self.bn.running_mean params["running_var"] = self.bn.running_var res["params"] = params return res @register class Convolution(Module): def __init__(self, insize, size, winlen, stride=1, padding=0, bias=True, activation=None, norm=None): super().__init__() self.conv = torch.nn.Conv1d(insize, size, winlen, stride=stride, padding=padding, bias=bias) self.activation = layers.get(activation, lambda: activation)() if isinstance(norm, dict): self.norm = from_dict(norm) elif isinstance(norm, str): self.norm = layers[norm](size) else: self.norm = norm def forward(self, x): h = self.conv(x) if self.norm is not None: h = self.norm(h) if self.activation is not None: h = self.activation(h) return h def to_dict(self, include_weights=False): res = { "insize": self.conv.in_channels, "size": self.conv.out_channels, "bias": self.conv.bias is not None, "winlen": self.conv.kernel_size[0], "stride": self.conv.stride[0], "padding": self.conv.padding[0], "activation": self.activation.name if self.activation else None, "norm": self.norm.to_dict(include_weights) if self.norm is not None else None } if include_weights: res['params'] = { 'W': self.conv.weight, 'b': self.conv.bias if self.conv.bias is not None else [] } return res @register class LinearCRFEncoder(Module): def __init__(self, insize, n_base, state_len, bias=True, scale=None, activation=None, blank_score=None, expand_blanks=True): super().__init__() self.scale = scale self.n_base = n_base self.state_len = state_len self.blank_score = blank_score self.expand_blanks = expand_blanks size = (n_base + 1) * n_base**state_len if blank_score is None else n_base**(state_len + 1) self.linear = torch.nn.Linear(insize, size, bias=bias) self.activation = layers.get(activation, lambda: activation)() def forward(self, x): scores = self.linear(x) if self.activation is not None: scores = self.activation(scores) if self.scale is not None: scores = scores * self.scale if self.blank_score is not None and self.expand_blanks: T, N, C = scores.shape scores = torch.nn.functional.pad( scores.view(T, N, C // self.n_base, self.n_base), (1, 0, 0, 0, 0, 0, 0, 0), value=self.blank_score ).view(T, N, -1) return scores def to_dict(self, include_weights=False): res = { 'insize': self.linear.in_features, 'n_base': self.n_base, 'state_len': self.state_len, 'bias': self.linear.bias is not None, 'scale': self.scale, 'activation': self.activation.name if self.activation else None, 'blank_score': self.blank_score, } if include_weights: res['params'] = { 'W': self.linear.weight, 'b': self.linear.bias if self.linear.bias is not None else [] } return res def extra_repr(self): return 'n_base={}, state_len={}, scale={}, blank_score={}, expand_blanks={}'.format( self.n_base, self.state_len, self.scale, self.blank_score, self.expand_blanks ) @register class Permute(Module): def __init__(self, dims): super().__init__() self.dims = dims def forward(self, x): return x.permute(*self.dims) def to_dict(self, include_weights=False): return {'dims': self.dims} def extra_repr(self): return 'dims={}'.format(self.dims) def truncated_normal(size, dtype=torch.float32, device=None, num_resample=5): x = torch.empty(size + (num_resample,), dtype=torch.float32, device=device).normal_() i = ((x < 2) & (x > -2)).max(-1, keepdim=True)[1] return torch.clamp_(x.gather(-1, i).squeeze(-1), -2, 2) class RNNWrapper(Module): def __init__( self, rnn_type, *args, reverse=False, orthogonal_weight_init=True, disable_state_bias=True, bidirectional=False, **kwargs ): super().__init__() if reverse and bidirectional: raise Exception("'reverse' and 'bidirectional' should not both be set to True") self.reverse = reverse self.rnn = rnn_type(*args, bidirectional=bidirectional, **kwargs) self.init_orthogonal(orthogonal_weight_init) self.init_biases() if disable_state_bias: self.disable_state_bias() def forward(self, x): if self.reverse: x = x.flip(0) y, h = self.rnn(x) if self.reverse: y = y.flip(0) return y def init_biases(self, types=('bias_ih',)): for name, param in self.rnn.named_parameters(): if any(k in name for k in types): with torch.no_grad(): param.set_(0.5*truncated_normal(param.shape, dtype=param.dtype, device=param.device)) def init_orthogonal(self, types=True): if not types: return if types == True: types = ('weight_ih', 'weight_hh') for name, x in self.rnn.named_parameters(): if any(k in name for k in types): for i in range(0, x.size(0), self.rnn.hidden_size): orthogonal_(x[i:i+self.rnn.hidden_size]) def disable_state_bias(self): for name, x in self.rnn.named_parameters(): if 'bias_hh' in name: x.requires_grad = False x.zero_() def extra_repr(self): return 'reverse={}'.format(bool(self.reverse)) @register class LSTM(RNNWrapper): def __init__(self, size, insize, bias=True, reverse=False): super().__init__(torch.nn.LSTM, insize, size, bias=bias, reverse=reverse) def to_dict(self, include_weights=False): res = { 'size': self.rnn.hidden_size, 'insize': self.rnn.input_size, 'bias': self.rnn.bias, 'reverse': self.reverse, } if include_weights: res['params'] = { 'iW': self.rnn.weight_ih_l0.reshape(4, self.rnn.hidden_size, self.rnn.input_size), 'sW': self.rnn.weight_hh_l0.reshape(4, self.rnn.hidden_size, self.rnn.hidden_size), 'b': self.rnn.bias_ih_l0.reshape(4, self.rnn.hidden_size) } return res def to_dict(layer, include_weights=False): if hasattr(layer, 'to_dict'): return {'type': layer.name, **layer.to_dict(include_weights)} return {'type': layer.name} def from_dict(model_dict, layer_types=None): model_dict = model_dict.copy() if layer_types is None: layer_types = layers type_name = model_dict.pop('type') typ = layer_types[type_name] if 'sublayers' in model_dict: sublayers = model_dict['sublayers'] model_dict['sublayers'] = [ from_dict(x, layer_types) for x in sublayers ] if isinstance(sublayers, list) else from_dict(sublayers, layer_types) try: layer = typ(**model_dict) except Exception as e: raise Exception(f'Failed to build layer of type {typ} with args {model_dict}') from e return layer def fuse_bn_(m): """ Sets the module m to eval mode and if a Convolution fuses any batchnorm layer. """ m.training = False if isinstance(m, Convolution) and isinstance(m.norm, BatchNorm): m.conv = fuse_conv_bn_eval(m.conv, m.norm.bn) m.norm = None
11,194
30.713881
133
py
bonito
bonito-master/bonito/util.py
""" Bonito utils """ import os import re import sys import random from glob import glob from itertools import groupby from operator import itemgetter from importlib import import_module from collections import deque, defaultdict, OrderedDict from torch.utils.data import DataLoader import toml import torch import koi.lstm import parasail import numpy as np from torch.cuda import get_device_capability try: from claragenomics.bindings import cuda from claragenomics.bindings.cudapoa import CudaPoaBatch except ImportError: pass __dir__ = os.path.dirname(os.path.realpath(__file__)) __data__ = os.path.join(__dir__, "data") __models__ = os.path.join(__dir__, "models") __configs__ = os.path.join(__dir__, "models/configs") split_cigar = re.compile(r"(?P<len>\d+)(?P<op>\D+)") default_data = os.path.join(__data__, "dna_r9.4.1") default_config = os.path.join(__configs__, "dna_r9.4.1@v3.1.toml") def init(seed, device, deterministic=True): """ Initialise random libs and setup cudnn https://pytorch.org/docs/stable/notes/randomness.html """ random.seed(seed) np.random.seed(seed) torch.manual_seed(seed) if device == "cpu": return torch.backends.cudnn.enabled = True torch.backends.cudnn.deterministic = deterministic torch.backends.cudnn.benchmark = (not deterministic) assert(torch.cuda.is_available()) def permute(x, input_layout, output_layout): """ Permute `x` from `input_layout` to `output_layout` >>> permute(x, 'TNC', 'NTC') """ if input_layout == output_layout: return x return x.permute(*[input_layout.index(x) for x in output_layout]) def concat(xs, dim=0): """ Type agnostic concat. """ if isinstance(xs[0], torch.Tensor): return torch.cat(xs, dim=dim) elif isinstance(xs[0], np.ndarray): return np.concatenate(xs, axis=dim) elif isinstance(xs[0], list): return [x for l in xs for x in l] elif isinstance(xs[0], str): return ''.join(xs) elif isinstance(xs[0], dict): return {k: concat([x[k] for x in xs], dim) for k in xs[0].keys()} else: raise TypeError def select_range(x, start, end, dim=0): """ Type agnostic range select. """ if isinstance(x, dict): return {k: select_range(v, start, end, dim) for (k, v) in x.items()} if dim == 0 or isinstance(x, list): return x[start:end] return x[(*(slice(None),)*dim, slice(start, end))] def size(x, dim=0): """ Type agnostic size. """ if hasattr(x, 'shape'): return x.shape[dim] elif dim == 0: return len(x) raise TypeError def half_supported(): """ Returns whether FP16 is support on the GPU """ try: return get_device_capability()[0] >= 7 except: return False def phred(prob, scale=1.0, bias=0.0): """ Converts `prob` into a ascii encoded phred quality score between 0 and 40. """ p = max(1 - prob, 1e-4) q = -10 * np.log10(p) * scale + bias return chr(int(np.round(q) + 33)) def mean_qscore_from_qstring(qstring): """ Convert qstring into a mean qscore """ if len(qstring) == 0: return 0.0 qs = (np.array(qstring, 'c').view(np.uint8) - 33) mean_err = np.exp(qs * (-np.log(10) / 10.)).mean() return -10 * np.log10(max(mean_err, 1e-4)) def decode_ref(encoded, labels): """ Convert a integer encoded reference into a string and remove blanks """ return ''.join(labels[e] for e in encoded.tolist() if e) def column_to_set(filename, idx=0, skip_header=False): """ Pull a column from a file and return a set of the values. """ if filename and os.path.isfile(filename): with open(filename, 'r') as tsv: if skip_header: next(tsv) return {line.strip().split()[idx] for line in tsv.readlines()} def chunk(signal, chunksize, overlap): """ Convert a read into overlapping chunks before calling """ if signal.ndim == 1: signal = signal.unsqueeze(0) T = signal.shape[-1] if chunksize == 0: chunks = signal[None, :] elif T < chunksize: n, overhang = divmod(chunksize, T) # np.tile operates only on dimension -1 by default, # whereas torch.repeat requires explicit listing of all input dimensions eg (1,n) or (1,1,n) chunks = torch.cat((torch.from_numpy(np.tile(signal,n)), signal[...,:overhang]), dim=-1)[None, :] else: stub = (T - overlap) % (chunksize - overlap) chunks = signal[...,stub:].unfold(-1, chunksize, chunksize - overlap).movedim(-2,0) if stub > 0: chunks = torch.cat([signal[None, ..., :chunksize], chunks], dim=0) return chunks def stitch(chunks, chunksize, overlap, length, stride, reverse=False): """ Stitch chunks together with a given overlap """ if chunks.shape[0] == 1: return chunks.squeeze(0) semi_overlap = overlap // 2 start, end = semi_overlap // stride, (chunksize - semi_overlap) // stride stub = (length - overlap) % (chunksize - overlap) first_chunk_end = (stub + semi_overlap) // stride if (stub > 0) else end if reverse: chunks = list(chunks) return concat([ chunks[-1][:-start], *(x[-end:-start] for x in reversed(chunks[1:-1])), chunks[0][-first_chunk_end:] ]) else: return concat([ chunks[0, :first_chunk_end], *chunks[1:-1, start:end], chunks[-1, start:] ]) def batchify(items, batchsize, dim=0): """ Batch up items up to `batch_size`. """ stack, pos = [], 0 for k, v in items: breaks = range(batchsize - pos, size(v, dim), batchsize) for start, end in zip([0, *breaks], [*breaks, size(v, dim)]): sub_batch = select_range(v, start, end, dim) stack.append(((k, (pos, pos + end - start)), sub_batch)) if pos + end - start == batchsize: ks, vs = zip(*stack) yield ks, concat(vs, dim) stack, pos = [], 0 else: pos += end - start if len(stack): ks, vs = zip(*stack) yield ks, concat(vs, dim) def unbatchify(batches, dim=0): """ Reconstruct batches. """ batches = ( (k, select_range(v, start, end, dim)) for sub_batches, v in batches for k, (start, end) in sub_batches ) return ( (k, concat([v for (k, v) in group], dim)) for k, group in groupby(batches, itemgetter(0)) ) def load_symbol(config, symbol): """ Dynamic load a symbol from module specified in model config. """ if not isinstance(config, dict): if not os.path.isdir(config) and os.path.isdir(os.path.join(__models__, config)): dirname = os.path.join(__models__, config) else: dirname = config config = toml.load(os.path.join(dirname, 'config.toml')) imported = import_module(config['model']['package']) return getattr(imported, symbol) def match_names(state_dict, model): keys_and_shapes = lambda state_dict: zip(*[ (k, s) for s, i, k in sorted([(v.shape, i, k) for i, (k, v) in enumerate(state_dict.items())]) ]) k1, s1 = keys_and_shapes(state_dict) k2, s2 = keys_and_shapes(model.state_dict()) assert s1 == s2 remap = dict(zip(k1, k2)) return OrderedDict([(k, remap[k]) for k in state_dict.keys()]) def get_last_checkpoint(dirname): weight_files = glob(os.path.join(dirname, "weights_*.tar")) if not weight_files: raise FileNotFoundError("no model weights found in '%s'" % dirname) weights = max([int(re.sub(".*_([0-9]+).tar", "\\1", w)) for w in weight_files]) return os.path.join(dirname, 'weights_%s.tar' % weights) def set_config_defaults(config, chunksize=None, batchsize=None, overlap=None, quantize=False): basecall_params = config.get("basecaller", {}) # use `value or dict.get(key)` rather than `dict.get(key, value)` to make # flags override values in config basecall_params["chunksize"] = chunksize or basecall_params.get("chunksize", 4000) basecall_params["overlap"] = overlap if overlap is not None else basecall_params.get("overlap", 500) basecall_params["batchsize"] = batchsize or basecall_params.get("batchsize", 64) basecall_params["quantize"] = basecall_params.get("quantize") if quantize is None else quantize config["basecaller"] = basecall_params return config def load_model(dirname, device, weights=None, half=None, chunksize=None, batchsize=None, overlap=None, quantize=False, use_koi=False): """ Load a model config and weights off disk from `dirname`. """ if not os.path.isdir(dirname) and os.path.isdir(os.path.join(__models__, dirname)): dirname = os.path.join(__models__, dirname) weights = get_last_checkpoint(dirname) if weights is None else os.path.join(dirname, 'weights_%s.tar' % weights) config = toml.load(os.path.join(dirname, 'config.toml')) config = set_config_defaults(config, chunksize, batchsize, overlap, quantize) return _load_model(weights, config, device, half, use_koi) def _load_model(model_file, config, device, half=None, use_koi=False): device = torch.device(device) Model = load_symbol(config, "Model") model = Model(config) config["basecaller"]["chunksize"] -= config["basecaller"]["chunksize"] % model.stride # overlap must be even multiple of stride for correct stitching config["basecaller"]["overlap"] -= config["basecaller"]["overlap"] % (model.stride * 2) if use_koi: model.use_koi( batchsize=config["basecaller"]["batchsize"], chunksize=config["basecaller"]["chunksize"], quantize=config["basecaller"]["quantize"], ) state_dict = torch.load(model_file, map_location=device) state_dict = {k2: state_dict[k1] for k1, k2 in match_names(state_dict, model).items()} new_state_dict = OrderedDict() for k, v in state_dict.items(): name = k.replace('module.', '') new_state_dict[name] = v model.load_state_dict(new_state_dict) if half is None: half = half_supported() if half: model = model.half() model.eval() model.to(device) return model def parasail_to_sam(result, seq): """ Extract reference start and sam compatible cigar string. :param result: parasail alignment result. :param seq: query sequence. :returns: reference start coordinate, cigar string. """ cigstr = result.cigar.decode.decode() first = re.search(split_cigar, cigstr) first_count, first_op = first.groups() prefix = first.group() rstart = result.cigar.beg_ref cliplen = result.cigar.beg_query clip = '' if cliplen == 0 else '{}S'.format(cliplen) if first_op == 'I': pre = '{}S'.format(int(first_count) + cliplen) elif first_op == 'D': pre = clip rstart = int(first_count) else: pre = '{}{}'.format(clip, prefix) mid = cigstr[len(prefix):] end_clip = len(seq) - result.end_query - 1 suf = '{}S'.format(end_clip) if end_clip > 0 else '' new_cigstr = ''.join((pre, mid, suf)) return rstart, new_cigstr def accuracy(ref, seq, balanced=False, min_coverage=0.0): """ Calculate the accuracy between `ref` and `seq` """ alignment = parasail.sw_trace_striped_32(seq, ref, 8, 4, parasail.dnafull) counts = defaultdict(int) q_coverage = len(alignment.traceback.query) / len(seq) r_coverage = len(alignment.traceback.ref) / len(ref) if r_coverage < min_coverage: return 0.0 _, cigar = parasail_to_sam(alignment, seq) for count, op in re.findall(split_cigar, cigar): counts[op] += int(count) if balanced: accuracy = (counts['='] - counts['I']) / (counts['='] + counts['X'] + counts['D']) else: accuracy = counts['='] / (counts['='] + counts['I'] + counts['X'] + counts['D']) return accuracy * 100 def print_alignment(ref, seq): """ Print the alignment between `ref` and `seq` """ alignment = parasail.sw_trace_striped_32(seq, ref, 8, 4, parasail.dnafull) print(alignment.traceback.ref) print(alignment.traceback.comp) print(alignment.traceback.query) print(" Score=%s" % alignment.score) return alignment.score def poa(groups, max_poa_sequences=100, gpu_mem_per_batch=0.9): """ Generate consensus for POA groups. Args: groups : A list of lists of sequences for which consensus is to be generated. """ free, total = cuda.cuda_get_mem_info(cuda.cuda_get_device()) gpu_mem_per_batch *= free batch = CudaPoaBatch(max_poa_sequences, gpu_mem_per_batch, stream=None, output_type="consensus") results = [] for i, group in enumerate(groups, start=1): group_status, seq_status = batch.add_poa_group(group) # Once batch is full, run POA processing if group_status == 1 or i == len(groups): batch.generate_poa() consensus, coverage, status = batch.get_consensus() results.extend(consensus) batch.reset() group_status, seq_status = batch.add_poa_group(group) return results
13,291
30.49763
134
py
bonito
bonito-master/bonito/schedule.py
import math import numpy as np from torch.optim.lr_scheduler import LambdaLR def linear_warmup_cosine_decay(end_ratio=0.01, warmup_steps=500, **kwargs): """ Linear warmup, cosine decay scheduler """ return lambda optimizer, train_loader, epochs, last_epoch: func_scheduler( optimizer=optimizer, func=cosine_decay_schedule(1.0, end_ratio), total_steps=epochs * len(train_loader), warmup_steps=warmup_steps, start_step=last_epoch * len(train_loader), ) def linear_warmup_const_inverse_sqrt_decay( warmup_steps=1000, decay_start_epoch=10, decay_scale=1.0, linear_cooldown_n_epochs=0, linear_cooldown_end_ratio=0.0, **kwargs ): """ Linear warmup, hold const, inverse sqrt decay, optional cooldown scheduler """ def gen_sched(optimizer, train_loader, epochs, last_epoch): steps_per_epoch = len(train_loader) start_step = steps_per_epoch*last_epoch total_steps = steps_per_epoch * epochs n_decay_epochs = epochs - decay_start_epoch - linear_cooldown_n_epochs decay_sched = inverse_sqrt_decay_schedule(decay_scale*n_decay_epochs) func = piecewise_schedule( [ warmup_steps / total_steps, decay_start_epoch / epochs, (epochs - linear_cooldown_n_epochs) / epochs ], [ linear_schedule(0.0, 1.0), const_schedule(1.0), decay_sched, linear_schedule( decay_sched(1.0), linear_cooldown_end_ratio ) ] ) return LambdaLR(optimizer, (lambda step: func((step + start_step) / total_steps))) return gen_sched def linear_cooldown(end_ratio=0.0, **kwargs): """ Linear Cooldown Scheduler """ return lambda optimizer, train_loader, epochs, last_epoch: func_scheduler( optimizer=optimizer, func=linear_schedule(1.0, end_ratio), total_steps=epochs * len(train_loader), start_step=0, ) #------------------------------------------------------------------------------- def const_schedule(y): """ Constant Scheduler """ return lambda t: y def linear_schedule(y0, y1): """ Linear Scheduler """ return lambda t: y0 + (y1 - y0) * t def cosine_decay_schedule(y0, y1): """ Cosine Decay Scheduler """ return lambda t: y1 + 0.5 * (y0 - y1) * (np.cos(t * np.pi) + 1.0) def piecewise_schedule(knots, funcs): """ Piecewise Scheduler """ def f(t): i = np.searchsorted(knots, t) t0 = 0.0 if i == 0 else knots[i - 1] t1 = 1.0 if i == len(knots) else knots[i] return funcs[i]((t - t0) / (t1 - t0)) return f def inverse_sqrt_decay_schedule(scale): return lambda t: 1.0 / math.sqrt(1 + scale*t) def func_scheduler(optimizer, func, total_steps, warmup_steps=None, warmup_ratio=0.1, start_step=0): """ Learning Rate Scheduler """ if warmup_steps: y0 = func(0.0) func = piecewise_schedule( [warmup_steps / total_steps], [linear_schedule(warmup_ratio * y0, y0), func] ) return LambdaLR(optimizer, (lambda step: func((step + start_step) / total_steps)))
3,330
26.528926
100
py
bonito
bonito-master/bonito/ctc/basecall.py
""" Bonito basecall """ import torch import numpy as np from functools import partial from bonito.multiprocessing import process_map from bonito.util import mean_qscore_from_qstring from bonito.util import chunk, stitch, batchify, unbatchify, permute def basecall(model, reads, beamsize=5, chunksize=0, overlap=0, batchsize=1, qscores=False, reverse=None): """ Basecalls a set of reads. """ chunks = ( (read, chunk(torch.tensor(read.signal), chunksize, overlap)) for read in reads ) scores = unbatchify( (k, compute_scores(model, v)) for k, v in batchify(chunks, batchsize) ) scores = ( (read, {'scores': stitch(v, chunksize, overlap, len(read.signal), model.stride)}) for read, v in scores ) decoder = partial(decode, decode=model.decode, beamsize=beamsize, qscores=qscores, stride=model.stride) basecalls = process_map(decoder, scores, n_proc=4) return basecalls def compute_scores(model, batch): """ Compute scores for model. """ with torch.no_grad(): device = next(model.parameters()).device chunks = batch.to(torch.half).to(device) probs = permute(model(chunks), 'TNC', 'NTC') return probs.cpu().to(torch.float32) def decode(scores, decode, beamsize=5, qscores=False, stride=1): """ Convert the network scores into a sequence. """ # do a greedy decode to get a sensible qstring to compute the mean qscore from seq, path = decode(scores['scores'], beamsize=1, qscores=True, return_path=True) seq, qstring = seq[:len(path)], seq[len(path):] mean_qscore = mean_qscore_from_qstring(qstring) # beam search will produce a better sequence but doesn't produce a sensible qstring/path if not (qscores or beamsize == 1): try: seq = decode(scores['scores'], beamsize=beamsize) path = None qstring = '*' except: pass return {'sequence': seq, 'qstring': qstring, 'stride': stride, 'moves': path}
2,018
31.564516
111
py
bonito
bonito-master/bonito/ctc/model.py
""" Bonito Model template """ import numpy as np from bonito.nn import Permute, layers import torch from torch.nn.functional import log_softmax, ctc_loss from torch.nn import Module, ModuleList, Sequential, Conv1d, BatchNorm1d, Dropout from fast_ctc_decode import beam_search, viterbi_search class Model(Module): """ Model template for QuartzNet style architectures https://arxiv.org/pdf/1910.10261.pdf """ def __init__(self, config): super(Model, self).__init__() if 'qscore' not in config: self.qbias = 0.0 self.qscale = 1.0 else: self.qbias = config['qscore']['bias'] self.qscale = config['qscore']['scale'] self.config = config self.stride = config['block'][0]['stride'][0] self.alphabet = config['labels']['labels'] self.features = config['block'][-1]['filters'] self.encoder = Encoder(config) self.decoder = Decoder(self.features, len(self.alphabet)) def forward(self, x): encoded = self.encoder(x) return self.decoder(encoded) def decode(self, x, beamsize=5, threshold=1e-3, qscores=False, return_path=False): x = x.exp().cpu().numpy().astype(np.float32) if beamsize == 1 or qscores: seq, path = viterbi_search(x, self.alphabet, qscores, self.qscale, self.qbias) else: seq, path = beam_search(x, self.alphabet, beamsize, threshold) if return_path: return seq, path return seq def ctc_label_smoothing_loss(self, log_probs, targets, lengths, weights=None): T, N, C = log_probs.shape weights = weights or torch.cat([torch.tensor([0.4]), (0.1 / (C - 1)) * torch.ones(C - 1)]) log_probs_lengths = torch.full(size=(N, ), fill_value=T, dtype=torch.int64) loss = ctc_loss(log_probs.to(torch.float32), targets, log_probs_lengths, lengths, reduction='mean') label_smoothing_loss = -((log_probs * weights.to(log_probs.device)).mean()) return {'total_loss': loss + label_smoothing_loss, 'loss': loss, 'label_smooth_loss': label_smoothing_loss} def loss(self, log_probs, targets, lengths): return self.ctc_label_smoothing_loss(log_probs, targets, lengths) class Encoder(Module): """ Builds the model encoder """ def __init__(self, config): super(Encoder, self).__init__() self.config = config features = self.config['input']['features'] activation = layers[self.config['encoder']['activation']]() encoder_layers = [] for layer in self.config['block']: encoder_layers.append( Block( features, layer['filters'], activation, repeat=layer['repeat'], kernel_size=layer['kernel'], stride=layer['stride'], dilation=layer['dilation'], dropout=layer['dropout'], residual=layer['residual'], separable=layer['separable'], ) ) features = layer['filters'] self.encoder = Sequential(*encoder_layers) def forward(self, x): return self.encoder(x) class TCSConv1d(Module): """ Time-Channel Separable 1D Convolution """ def __init__(self, in_channels, out_channels, kernel_size, stride=1, padding=0, dilation=1, groups=1, bias=False, separable=False): super(TCSConv1d, self).__init__() self.separable = separable if separable: self.depthwise = Conv1d( in_channels, in_channels, kernel_size=kernel_size, stride=stride, padding=padding, dilation=dilation, bias=bias, groups=in_channels ) self.pointwise = Conv1d( in_channels, out_channels, kernel_size=1, stride=1, dilation=dilation, bias=bias, padding=0 ) else: self.conv = Conv1d( in_channels, out_channels, kernel_size=kernel_size, stride=stride, padding=padding, dilation=dilation, bias=bias ) def forward(self, x): if self.separable: x = self.depthwise(x) x = self.pointwise(x) else: x = self.conv(x) return x class Block(Module): """ TCSConv, Batch Normalisation, Activation, Dropout """ def __init__(self, in_channels, out_channels, activation, repeat=5, kernel_size=1, stride=1, dilation=1, dropout=0.0, residual=False, separable=False): super(Block, self).__init__() self.use_res = residual self.conv = ModuleList() _in_channels = in_channels padding = self.get_padding(kernel_size[0], stride[0], dilation[0]) # add the first n - 1 convolutions + activation for _ in range(repeat - 1): self.conv.extend( self.get_tcs( _in_channels, out_channels, kernel_size=kernel_size, stride=stride, dilation=dilation, padding=padding, separable=separable ) ) self.conv.extend(self.get_activation(activation, dropout)) _in_channels = out_channels # add the last conv and batch norm self.conv.extend( self.get_tcs( _in_channels, out_channels, kernel_size=kernel_size, stride=stride, dilation=dilation, padding=padding, separable=separable ) ) # add the residual connection if self.use_res: self.residual = Sequential(*self.get_tcs(in_channels, out_channels)) # add the activation and dropout self.activation = Sequential(*self.get_activation(activation, dropout)) def get_activation(self, activation, dropout): return activation, Dropout(p=dropout) def get_padding(self, kernel_size, stride, dilation): if stride > 1 and dilation > 1: raise ValueError("Dilation and stride can not both be greater than 1") return (kernel_size // 2) * dilation def get_tcs(self, in_channels, out_channels, kernel_size=1, stride=1, dilation=1, padding=0, bias=False, separable=False): return [ TCSConv1d( in_channels, out_channels, kernel_size, stride=stride, dilation=dilation, padding=padding, bias=bias, separable=separable ), BatchNorm1d(out_channels, eps=1e-3, momentum=0.1) ] def forward(self, x): _x = x for layer in self.conv: _x = layer(_x) if self.use_res: _x = _x + self.residual(x) return self.activation(_x) class Decoder(Module): """ Decoder """ def __init__(self, features, classes): super(Decoder, self).__init__() self.layers = Sequential( Conv1d(features, classes, kernel_size=1, bias=True), Permute([2, 0, 1]) ) def forward(self, x): return log_softmax(self.layers(x), dim=-1)
7,110
33.1875
155
py
bonito
bonito-master/bonito/cli/export.py
""" Bonito Export """ import io import os import re import sys import json import toml import torch import bonito import hashlib import numpy as np from glob import glob import base64 from argparse import ArgumentParser, ArgumentDefaultsHelpFormatter from bonito.nn import fuse_bn_ from bonito.util import _load_model, get_last_checkpoint, set_config_defaults class JsonEncoder(json.JSONEncoder): def default(self, obj): if isinstance(obj, np.integer): return int(obj) elif isinstance(obj, np.floating): return float(obj) elif isinstance(obj, np.ndarray): return obj.tolist() elif isinstance(obj, torch.nn.Parameter): return obj.data elif isinstance(obj, torch.Tensor): return obj.detach().numpy() elif isinstance(obj, bytes): return obj.decode('ascii') else: return super(JsonEncoder, self).default(obj) def file_md5(filename, nblock=1024): """ Get md5 string from file. """ hasher = hashlib.md5() block_size = nblock * hasher.block_size with open(filename, "rb") as fh: for blk in iter((lambda: fh.read(block_size)), b""): hasher.update(blk) return hasher.hexdigest() def save_tensor(directory, name, tensor): """ Save a tensor `x` to `fn.tensor` for use with libtorch. """ module = torch.nn.Module() param = torch.nn.Parameter(tensor, requires_grad=False) module.register_parameter("0", param) tensors = torch.jit.script(module) tensors.save(f"{directory}/{name}.tensor") def reformat_output_layer(layer_dict, v4=True): n_base, state_len, blank_score = [layer_dict.pop(k) for k in ['n_base', 'state_len', 'blank_score']] layer_dict['size'] = (n_base + 1) * n_base**state_len if blank_score is not None: if v4: layer_dict['type'] = 'GlobalNormTransducer' params = layer_dict['params'] params['W'] = torch.nn.functional.pad( params['W'].reshape([n_base**state_len, n_base, -1]), (0, 0, 1, 0), value=0. ).reshape((n_base + 1) * n_base**state_len, -1) if layer_dict['bias'] is False: params['b'] = torch.zeros(n_base**state_len * (n_base + 1)) params['b'][0::5] = np.arctanh(blank_score / 5.0) else: params['b'] = torch.nn.functional.pad( params['b'].reshape(n_base**state_len, n_base), (1, 0), value=0. ).reshape(-1) layer_dict['activation'] = 'identity' layer_dict['scale'] = 1.0 layer_dict['stay_score'] = blank_score else: layer_dict['type'] = 'GlobalNormTransducer' assert layer_dict['activation'] == 'tanh' params = layer_dict['params'] params['W'] = torch.nn.functional.pad( params['W'].reshape([n_base**state_len, n_base, -1]), (0, 0, 1, 0), value=0. ).reshape((n_base + 1) * n_base**state_len, -1) params['b'] = torch.nn.functional.pad( params['b'].reshape(n_base**state_len, n_base), (1, 0), value=np.arctanh(blank_score / layer_dict['scale']) ).reshape(-1) return layer_dict def to_guppy_feed_forward(layer): layer['type'] = 'feed-forward' layer['insize'] = layer['in_features'] layer['size'] = layer['out_features'] layer['activation'] = 'identity' del layer['in_features'] del layer['out_features'] return layer def to_guppy_dict(model, include_weights=True, binary_weights=True, v4=True): guppy_dict = bonito.nn.to_dict(model.encoder, include_weights=include_weights) guppy_dict['sublayers'] = [x for x in guppy_dict['sublayers'] if x['type'] != 'permute'] guppy_dict['sublayers'] = [dict(x, type='LSTM', activation='tanh', gate='sigmoid') if x['type'] == 'lstm' else x for x in guppy_dict['sublayers']] guppy_dict['sublayers'] = [dict(x, padding=(x['padding'], x['padding'])) if x['type'] == 'convolution' else x for x in guppy_dict['sublayers']] guppy_dict['sublayers'] = [to_guppy_feed_forward(x) if x['type'] == 'linear' else x for x in guppy_dict['sublayers']] idx = -1 if guppy_dict['sublayers'][-1]['type'] == 'linearcrfencoder' else -2 guppy_dict['sublayers'][idx] = reformat_output_layer(guppy_dict['sublayers'][idx], v4=v4) if binary_weights: for layer_dict in guppy_dict['sublayers']: if 'params' in layer_dict: layer_dict['params'] = { f'{k}_binary': base64.b64encode(v.data.detach().numpy().astype(np.float32).tobytes()) for (k, v) in layer_dict['params'].items() } guppy_dict['sublayers'] = [{'type': 'reverse', 'sublayers': x} if x.pop('reverse', False) else x for x in guppy_dict['sublayers']] return guppy_dict def main(args): model_file = get_last_checkpoint(args.model) if os.path.isdir(args.model) else args.model if args.config is None: args.config = os.path.join(os.path.dirname(model_file), "config.toml") config = toml.load(args.config) config = set_config_defaults(config) model = _load_model(model_file, config, device='cpu') if args.fuse_bn: # model weights might be saved in half when training and PyTorch's bn fusion # code uses an op (rsqrt) that currently (1.11) only has a float implementation model = model.to(torch.float32).apply(fuse_bn_) if args.format == 'guppy': v4 = True if 'type' in config['encoder'] else False jsn = to_guppy_dict(model, v4=v4) jsn["md5sum"] = file_md5(model_file) json.dump(jsn, sys.stdout, cls=JsonEncoder) elif args.format == 'dorado': for name, tensor in model.encoder.state_dict().items(): save_tensor(args.model, name, tensor) elif args.format == 'torchscript': tmp_tensor = torch.rand(10, 1, 1000) model = model.float() traced_script_module = torch.jit.trace(model, tmp_tensor) buffer = io.BytesIO() torch.jit.save(traced_script_module, buffer) buffer.seek(0) sys.stdout.buffer.write(buffer.getvalue()) sys.stdout.flush() else: raise NotImplementedError("Export format not supported") def argparser(): parser = ArgumentParser( formatter_class=ArgumentDefaultsHelpFormatter, add_help=False ) parser.add_argument('model') parser.add_argument('--format', choices=['guppy', 'dorado', 'torchscript'], default='guppy') parser.add_argument('--config', default=None, help='config file to read settings from') parser.add_argument('--fuse-bn', default=True, help='fuse batchnorm layers') return parser
6,898
35.696809
150
py
bonito
bonito-master/bonito/cli/evaluate.py
""" Bonito model evaluator """ import os import time import torch import numpy as np from itertools import starmap from argparse import ArgumentParser, ArgumentDefaultsHelpFormatter from pathlib import Path from bonito.data import load_numpy, load_script from bonito.util import accuracy, poa, decode_ref, half_supported from bonito.util import init, load_model, concat, permute from torch.utils.data import DataLoader def main(args): poas = [] init(args.seed, args.device) print("* loading data") try: _, valid_loader_kwargs = load_numpy(args.chunks, args.directory) except FileNotFoundError: _, valid_loader_kwargs = load_script( args.directory, seed=args.seed, chunks=args.chunks, valid_chunks=args.chunks ) dataloader = DataLoader( batch_size=args.batchsize, num_workers=4, pin_memory=True, **valid_loader_kwargs ) accuracy_with_cov = lambda ref, seq: accuracy(ref, seq, min_coverage=args.min_coverage) for w in [int(i) for i in args.weights.split(',')]: seqs = [] print("* loading model", w) model = load_model(args.model_directory, args.device, weights=w) print("* calling") t0 = time.perf_counter() targets = [] with torch.no_grad(): for data, target, *_ in dataloader: targets.extend(torch.unbind(target, 0)) if half_supported(): data = data.type(torch.float16).to(args.device) else: data = data.to(args.device) log_probs = model(data) if hasattr(model, 'decode_batch'): seqs.extend(model.decode_batch(log_probs)) else: seqs.extend([model.decode(p) for p in permute(log_probs, 'TNC', 'NTC')]) duration = time.perf_counter() - t0 refs = [decode_ref(target, model.alphabet) for target in targets] accuracies = [accuracy_with_cov(ref, seq) if len(seq) else 0. for ref, seq in zip(refs, seqs)] if args.poa: poas.append(sequences) print("* mean %.2f%%" % np.mean(accuracies)) print("* median %.2f%%" % np.median(accuracies)) print("* time %.2f" % duration) print("* samples/s %.2E" % (args.chunks * data.shape[2] / duration)) if args.poa: print("* doing poa") t0 = time.perf_counter() # group each sequence prediction per model together poas = [list(seq) for seq in zip(*poas)] consensuses = poa(poas) duration = time.perf_counter() - t0 accuracies = list(starmap(accuracy_with_coverage_filter, zip(references, consensuses))) print("* mean %.2f%%" % np.mean(accuracies)) print("* median %.2f%%" % np.median(accuracies)) print("* time %.2f" % duration) def argparser(): parser = ArgumentParser( formatter_class=ArgumentDefaultsHelpFormatter, add_help=False ) parser.add_argument("model_directory") parser.add_argument("--directory", type=Path) parser.add_argument("--device", default="cuda") parser.add_argument("--seed", default=9, type=int) parser.add_argument("--weights", default="0", type=str) parser.add_argument("--chunks", default=1000, type=int) parser.add_argument("--batchsize", default=96, type=int) parser.add_argument("--beamsize", default=5, type=int) parser.add_argument("--poa", action="store_true", default=False) parser.add_argument("--min-coverage", default=0.5, type=float) return parser
3,647
31.283186
102
py
bonito
bonito-master/bonito/cli/train.py
#!/usr/bin/env python3 """ Bonito training. """ import os from argparse import ArgumentParser from argparse import ArgumentDefaultsHelpFormatter from pathlib import Path from importlib import import_module from bonito.data import load_numpy, load_script from bonito.util import __models__, default_config, default_data from bonito.util import load_model, load_symbol, init, half_supported from bonito.training import load_state, Trainer import toml import torch import numpy as np from torch.utils.data import DataLoader def main(args): workdir = os.path.expanduser(args.training_directory) if os.path.exists(workdir) and not args.force: print("[error] %s exists, use -f to force continue training." % workdir) exit(1) init(args.seed, args.device, (not args.nondeterministic)) device = torch.device(args.device) if not args.pretrained: config = toml.load(args.config) else: dirname = args.pretrained if not os.path.isdir(dirname) and os.path.isdir(os.path.join(__models__, dirname)): dirname = os.path.join(__models__, dirname) pretrain_file = os.path.join(dirname, 'config.toml') config = toml.load(pretrain_file) if 'lr_scheduler' in config: print(f"[ignoring 'lr_scheduler' in --pretrained config]") del config['lr_scheduler'] argsdict = dict(training=vars(args)) print("[loading model]") if args.pretrained: print("[using pretrained model {}]".format(args.pretrained)) model = load_model(args.pretrained, device, half=False) else: model = load_symbol(config, 'Model')(config) print("[loading data]") try: train_loader_kwargs, valid_loader_kwargs = load_numpy( args.chunks, args.directory, valid_chunks = args.valid_chunks ) except FileNotFoundError: train_loader_kwargs, valid_loader_kwargs = load_script( args.directory, seed=args.seed, chunks=args.chunks, valid_chunks=args.valid_chunks, n_pre_context_bases=getattr(model, "n_pre_context_bases", 0), n_post_context_bases=getattr(model, "n_post_context_bases", 0), ) loader_kwargs = { "batch_size": args.batch, "num_workers": 4, "pin_memory": True } train_loader = DataLoader(**loader_kwargs, **train_loader_kwargs) valid_loader = DataLoader(**loader_kwargs, **valid_loader_kwargs) os.makedirs(workdir, exist_ok=True) toml.dump({**config, **argsdict}, open(os.path.join(workdir, 'config.toml'), 'w')) if config.get("lr_scheduler"): sched_config = config["lr_scheduler"] lr_scheduler_fn = getattr( import_module(sched_config["package"]), sched_config["symbol"] )(**sched_config) else: lr_scheduler_fn = None trainer = Trainer( model, device, train_loader, valid_loader, use_amp=half_supported() and not args.no_amp, lr_scheduler_fn=lr_scheduler_fn, restore_optim=args.restore_optim, save_optim_every=args.save_optim_every, grad_accum_split=args.grad_accum_split, quantile_grad_clip=args.quantile_grad_clip ) if (',' in args.lr): lr = [float(x) for x in args.lr.split(',')] else: lr = float(args.lr) trainer.fit(workdir, args.epochs, lr) def argparser(): parser = ArgumentParser( formatter_class=ArgumentDefaultsHelpFormatter, add_help=False ) parser.add_argument("training_directory") group = parser.add_mutually_exclusive_group() group.add_argument('--config', default=default_config) group.add_argument('--pretrained', default="") parser.add_argument("--directory", type=Path) parser.add_argument("--device", default="cuda") parser.add_argument("--lr", default='2e-3') parser.add_argument("--seed", default=25, type=int) parser.add_argument("--epochs", default=5, type=int) parser.add_argument("--batch", default=64, type=int) parser.add_argument("--chunks", default=0, type=int) parser.add_argument("--valid-chunks", default=None, type=int) parser.add_argument("--no-amp", action="store_true", default=False) parser.add_argument("-f", "--force", action="store_true", default=False) parser.add_argument("--restore-optim", action="store_true", default=False) parser.add_argument("--nondeterministic", action="store_true", default=False) parser.add_argument("--save-optim-every", default=10, type=int) parser.add_argument("--grad-accum-split", default=1, type=int) parser.add_argument("--quantile-grad-clip", action="store_true", default=False) return parser
4,704
35.472868
91
py
bonito
bonito-master/bonito/crf/basecall.py
""" Bonito CRF basecalling """ import torch import numpy as np from koi.decode import beam_search, to_str from bonito.multiprocessing import thread_iter from bonito.util import chunk, stitch, batchify, unbatchify, half_supported def stitch_results(results, length, size, overlap, stride, reverse=False): """ Stitch results together with a given overlap. """ if isinstance(results, dict): return { k: stitch_results(v, length, size, overlap, stride, reverse=reverse) for k, v in results.items() } if length < size: return results[0, :int(np.floor(length / stride))] return stitch(results, size, overlap, length, stride, reverse=reverse) def compute_scores(model, batch, beam_width=32, beam_cut=100.0, scale=1.0, offset=0.0, blank_score=2.0, reverse=False): """ Compute scores for model. """ with torch.inference_mode(): device = next(model.parameters()).device dtype = torch.float16 if half_supported() else torch.float32 scores = model(batch.to(dtype).to(device)) if reverse: scores = model.seqdist.reverse_complement(scores) with torch.cuda.device(scores.device): sequence, qstring, moves = beam_search( scores, beam_width=beam_width, beam_cut=beam_cut, scale=scale, offset=offset, blank_score=blank_score ) return { 'moves': moves, 'qstring': qstring, 'sequence': sequence, } def fmt(stride, attrs, rna=False): fliprna = (lambda x:x[::-1]) if rna else (lambda x:x) return { 'stride': stride, 'moves': attrs['moves'].numpy(), 'qstring': fliprna(to_str(attrs['qstring'])), 'sequence': fliprna(to_str(attrs['sequence'])), } def basecall(model, reads, chunksize=4000, overlap=100, batchsize=32, reverse=False, rna=False): """ Basecalls a set of reads. """ chunks = thread_iter( ((read, 0, read.signal.shape[-1]), chunk(torch.from_numpy(read.signal), chunksize, overlap)) for read in reads ) batches = thread_iter(batchify(chunks, batchsize=batchsize)) scores = thread_iter( (read, compute_scores(model, batch, reverse=reverse)) for read, batch in batches ) results = thread_iter( (read, stitch_results(scores, end - start, chunksize, overlap, model.stride, reverse)) for ((read, start, end), scores) in unbatchify(scores) ) return thread_iter( (read, fmt(model.stride, attrs, rna)) for read, attrs in results )
2,632
30.345238
119
py
bonito
bonito-master/bonito/crf/model.py
""" Bonito CTC-CRF Model. """ import torch import numpy as np import koi from koi.ctc import SequenceDist, Max, Log, semiring from koi.ctc import logZ_cu, viterbi_alignments, logZ_cu_sparse, bwd_scores_cu_sparse, fwd_scores_cu_sparse from bonito.nn import Module, Convolution, LinearCRFEncoder, Serial, Permute, layers, from_dict def get_stride(m): children = list(m.children()) if len(children) == 0: if hasattr(m, "stride"): stride = m.stride if isinstance(stride, int): return stride return np.prod(stride) return 1 return np.prod([get_stride(c) for c in children]) class CTC_CRF(SequenceDist): def __init__(self, state_len, alphabet, n_pre_context_bases=0, n_post_context_bases=0): super().__init__() self.alphabet = alphabet self.state_len = state_len self.n_pre_context_bases = n_pre_context_bases self.n_post_context_bases = n_post_context_bases self.n_base = len(alphabet[1:]) self.idx = torch.cat([ torch.arange(self.n_base**(self.state_len))[:, None], torch.arange( self.n_base**(self.state_len) ).repeat_interleave(self.n_base).reshape(self.n_base, -1).T ], dim=1).to(torch.int32) def n_score(self): return len(self.alphabet) * self.n_base**(self.state_len) def logZ(self, scores, S:semiring=Log): T, N, _ = scores.shape Ms = scores.reshape(T, N, -1, len(self.alphabet)) alpha_0 = Ms.new_full((N, self.n_base**(self.state_len)), S.one) beta_T = Ms.new_full((N, self.n_base**(self.state_len)), S.one) return logZ_cu_sparse(Ms, self.idx, alpha_0, beta_T, S) def normalise(self, scores): return (scores - self.logZ(scores)[:, None] / len(scores)) def forward_scores(self, scores, S: semiring=Log): T, N, _ = scores.shape Ms = scores.reshape(T, N, -1, self.n_base + 1) alpha_0 = Ms.new_full((N, self.n_base**(self.state_len)), S.one) return fwd_scores_cu_sparse(Ms, self.idx, alpha_0, S, K=1) def backward_scores(self, scores, S: semiring=Log): T, N, _ = scores.shape Ms = scores.reshape(T, N, -1, self.n_base + 1) beta_T = Ms.new_full((N, self.n_base**(self.state_len)), S.one) return bwd_scores_cu_sparse(Ms, self.idx, beta_T, S, K=1) def compute_transition_probs(self, scores, betas): T, N, C = scores.shape # add bwd scores to edge scores log_trans_probs = (scores.reshape(T, N, -1, self.n_base + 1) + betas[1:, :, :, None]) # transpose from (new_state, dropped_base) to (old_state, emitted_base) layout log_trans_probs = torch.cat([ log_trans_probs[:, :, :, [0]], log_trans_probs[:, :, :, 1:].transpose(3, 2).reshape(T, N, -1, self.n_base) ], dim=-1) # convert from log probs to probs by exponentiating and normalising trans_probs = torch.softmax(log_trans_probs, dim=-1) #convert first bwd score to initial state probabilities init_state_probs = torch.softmax(betas[0], dim=-1) return trans_probs, init_state_probs def reverse_complement(self, scores): T, N, C = scores.shape expand_dims = T, N, *(self.n_base for _ in range(self.state_len)), self.n_base + 1 scores = scores.reshape(*expand_dims) blanks = torch.flip(scores[..., 0].permute( 0, 1, *range(self.state_len + 1, 1, -1)).reshape(T, N, -1, 1), [0, 2] ) emissions = torch.flip(scores[..., 1:].permute( 0, 1, *range(self.state_len, 1, -1), self.state_len +2, self.state_len + 1).reshape(T, N, -1, self.n_base), [0, 2, 3] ) return torch.cat([blanks, emissions], dim=-1).reshape(T, N, -1) def viterbi(self, scores): traceback = self.posteriors(scores, Max) a_traceback = traceback.argmax(2) moves = (a_traceback % len(self.alphabet)) != 0 paths = 1 + (torch.div(a_traceback, len(self.alphabet), rounding_mode="floor") % self.n_base) return torch.where(moves, paths, 0) def path_to_str(self, path): alphabet = np.frombuffer(''.join(self.alphabet).encode(), dtype='u1') seq = alphabet[path[path != 0]] return seq.tobytes().decode() def prepare_ctc_scores(self, scores, targets): # convert from CTC targets (with blank=0) to zero indexed targets = torch.clamp(targets - 1, 0) T, N, C = scores.shape scores = scores.to(torch.float32) n = targets.size(1) - (self.state_len - 1) stay_indices = sum( targets[:, i:n + i] * self.n_base ** (self.state_len - i - 1) for i in range(self.state_len) ) * len(self.alphabet) move_indices = stay_indices[:, 1:] + targets[:, :n - 1] + 1 stay_scores = scores.gather(2, stay_indices.expand(T, -1, -1)) move_scores = scores.gather(2, move_indices.expand(T, -1, -1)) return stay_scores, move_scores def ctc_loss(self, scores, targets, target_lengths, loss_clip=None, reduction='mean', normalise_scores=True): if normalise_scores: scores = self.normalise(scores) stay_scores, move_scores = self.prepare_ctc_scores(scores, targets) logz = logZ_cu(stay_scores, move_scores, target_lengths + 1 - self.state_len) loss = - (logz / target_lengths) if loss_clip: loss = torch.clamp(loss, 0.0, loss_clip) if reduction == 'mean': return loss.mean() elif reduction in ('none', None): return loss else: raise ValueError('Unknown reduction type {}'.format(reduction)) def ctc_viterbi_alignments(self, scores, targets, target_lengths): stay_scores, move_scores = self.prepare_ctc_scores(scores, targets) return viterbi_alignments(stay_scores, move_scores, target_lengths + 1 - self.state_len) def conv(c_in, c_out, ks, stride=1, bias=False, activation=None, norm=None): return Convolution(c_in, c_out, ks, stride=stride, padding=ks//2, bias=bias, activation=activation, norm=norm) def rnn_encoder(n_base, state_len, insize=1, stride=5, winlen=19, activation='swish', rnn_type='lstm', features=768, scale=5.0, blank_score=None, expand_blanks=True, num_layers=5, norm=None): rnn = layers[rnn_type] return Serial([ conv(insize, 4, ks=5, bias=True, activation=activation, norm=norm), conv(4, 16, ks=5, bias=True, activation=activation, norm=norm), conv(16, features, ks=winlen, stride=stride, bias=True, activation=activation, norm=norm), Permute([2, 0, 1]), *(rnn(features, features, reverse=(num_layers - i) % 2) for i in range(num_layers)), LinearCRFEncoder( features, n_base, state_len, activation='tanh', scale=scale, blank_score=blank_score, expand_blanks=expand_blanks ) ]) class SeqdistModel(Module): def __init__(self, encoder, seqdist, n_pre_post_context_bases=None): super().__init__() self.seqdist = seqdist self.encoder = encoder self.stride = get_stride(encoder) self.alphabet = seqdist.alphabet if n_pre_post_context_bases is None: self.n_pre_context_bases = self.seqdist.state_len - 1 self.n_post_context_bases = 1 else: self.n_pre_context_bases, self.n_post_context_bases = n_pre_post_context_bases def forward(self, x, *args): return self.encoder(x) def decode_batch(self, x): scores = self.seqdist.posteriors(x.to(torch.float32)) + 1e-8 tracebacks = self.seqdist.viterbi(scores.log()).to(torch.int16).T return [self.seqdist.path_to_str(x) for x in tracebacks.cpu().numpy()] def decode(self, x): return self.decode_batch(x.unsqueeze(1))[0] def loss(self, scores, targets, target_lengths, **kwargs): return self.seqdist.ctc_loss(scores.to(torch.float32), targets, target_lengths, **kwargs) def use_koi(self, **kwargs): self.encoder = koi.lstm.update_graph( self.encoder, batchsize=kwargs["batchsize"], chunksize=kwargs["chunksize"] // self.stride, quantize=kwargs["quantize"], ) class Model(SeqdistModel): def __init__(self, config): seqdist = CTC_CRF( state_len=config['global_norm']['state_len'], alphabet=config['labels']['labels'] ) if 'type' in config['encoder']: #new-style config encoder = from_dict(config['encoder']) else: #old-style encoder = rnn_encoder(seqdist.n_base, seqdist.state_len, insize=config['input']['features'], **config['encoder']) super().__init__(encoder, seqdist, n_pre_post_context_bases=config['input'].get('n_pre_post_context_bases')) self.config = config
8,951
40.831776
191
py
ReAgent
ReAgent-master/reagent/ope/utils.py
#!/usr/bin/env python3 import math from collections import OrderedDict from typing import Sequence, Union import numpy as np import torch DEFAULT_MIN = float("-inf") DEFAULT_MAX = float("inf") def convert_to_one_hots(a, num_classes: int, dtype=torch.int, device=None): """ Convert class index array (num_sample,) to an one hots array (num_sample, num_classes) Args: a: index array num_classes: number of classes dtype: data type Returns: one hots array in shape of (a.shape[0], num_classes) """ one_hots = torch.zeros((len(a), num_classes), dtype=dtype, device=device) one_hots[torch.arange(one_hots.shape[0]), a] = 1 return one_hots class LRUCache(OrderedDict): def __init__(self, maxsize=2 ** 10, *args, **kwds): self.maxsize = maxsize super().__init__(*args, **kwds) def __getitem__(self, key): value = super().__getitem__(key) self.move_to_end(key) return value def __setitem__(self, key, value): super().__setitem__(key, value) if len(self) > self.maxsize: del self[next(iter(self))] class RunningAverage: def __init__(self, init_val: float = float("nan")): self._average = init_val self._count = 0 if math.isnan(init_val) else 1 def add(self, value) -> "RunningAverage": if not math.isnan(value) and not math.isinf(value): if self._count == 0: self._average = 0.0 self._count += 1 self._average = self._average + (float(value) - self._average) / self._count return self @property def average(self): return self._average @property def count(self): return self._count @property def total(self): return self._average * self._count def __float__(self): return self._average class Clamper: def __init__(self, min_v: float = DEFAULT_MIN, max_v: float = DEFAULT_MAX): self._min = min_v self._max = max_v if self._min >= self._max: raise ValueError(f"min[{min}] greater than max[{max}]") def __call__( self, v: Union[float, Sequence[float], torch.Tensor, np.ndarray] ) -> Union[float, Sequence[float], torch.Tensor, np.ndarray]: if isinstance(v, torch.Tensor): return v.clamp(self._min, self._max) elif isinstance(v, np.ndarray): return v.clip(self._min, self._max) elif isinstance(v, Sequence): return [max(self._min, min(self._max, float(i))) for i in v] else: return max(self._min, min(self._max, float(v))) def __repr__(self): return f"Clamper({self._min},{self._max})"
2,743
26.717172
88
py
ReAgent
ReAgent-master/reagent/ope/estimators/slate_estimators.py
#!/usr/bin/env python3 import logging import math import random import time from abc import ABC, abstractmethod from dataclasses import dataclass from typing import ( Iterable, Mapping, MutableMapping, MutableSequence, Optional, Sequence, Set, Tuple, Union, ) import numpy as np import torch from reagent.ope.estimators.estimator import ( Estimator, EstimatorResult, EstimatorSampleResult, ) from reagent.ope.estimators.types import ( Action, Items, Objects, Probability, Reward, Trainer, TrainingData, TypeWrapper, Values, ValueType, ) from reagent.ope.utils import Clamper, RunningAverage from torch import Tensor # Main algorithms are from two papers: # 1. Offline Evaluation of Ranking Policies with Click Models # https://arxiv.org/abs/1804.10488 # 2. Off-policy evaluation for slate recommendation # https://arxiv.org/abs/1605.04812 # Types for slates SlateSlotType = Union[int, Tuple[int], float, Tuple[float], np.ndarray, Tensor] SlateSlot = TypeWrapper[SlateSlotType] logger = logging.getLogger(__name__) class SlateSlotValues(Values[SlateSlot]): """ Map from a slot to a value """ def _to_key(self, k: int) -> SlateSlot: return SlateSlot(k) class SlateSlots(Items[SlateSlot]): """ List of slot """ def _new_item(self, i: int) -> SlateSlot: return SlateSlot(i) # pyre-fixme[15]: `fill` overrides method defined in `Items` inconsistently. def fill( self, values: Union[Mapping[SlateSlot, float], Sequence[float], np.ndarray, Tensor], ) -> SlateSlotValues: """ Map slots to given values Args: values: given values Returns: Map from slots to given values """ return SlateSlotValues(super().fill(values)) class SlateSlotObjects(Objects[SlateSlot, ValueType]): def __init__( self, values: Union[MutableMapping[SlateSlot, ValueType], MutableSequence[ValueType]], ): assert (len(values)) > 0 super().__init__(values) def _to_key(self, k: int) -> SlateSlot: return SlateSlot(k) @property def slots(self) -> SlateSlots: if self.is_sequence: # pyre-fixme[16]: `SlateSlotObjects` has no attribute `_values`. return SlateSlots(len(self._values)) else: return SlateSlots(list(self._key_to_index.keys())) @property def objects(self) -> Sequence[ValueType]: return super().values def fill( self, values: Sequence[ValueType] ) -> Union[Mapping[SlateSlot, ValueType], Sequence[ValueType]]: # pyre-fixme[16]: `SlateSlotObjects` has no attribute `_values`. assert len(values) >= len(self._values) if self._key_to_index is None: return values[: len(self._values)] else: return {s: v for s, v in zip(self.slots, values[: len(self._values)])} # type of valid slate candidates, e.g., doc id SlateItem = Action class SlateItems(Items[SlateItem]): def _new_item(self, i: int) -> SlateItem: return SlateItem(i) class SlateItemValues(Values[SlateItem]): def _to_key(self, k: int) -> SlateItem: return SlateItem(k) @property def items(self) -> SlateItems: if self.is_sequence: return SlateItems(len(self)) else: return SlateItems(super().keys) class SlateItemFeatures(Objects[SlateItem, Tensor]): def __init__( self, values: Union[Mapping[SlateItem, Tensor], Sequence[Tensor], Tensor, np.ndarray], ): # pyre-fixme[6]: Expected # `Union[Mapping[Variable[reagent.ope.estimators.types.KeyType], # Variable[ValueType]], Sequence[Variable[ValueType]]]` for 1st param but got # `Union[Mapping[TypeWrapper[Union[Tuple[float], Tuple[int], Tensor, float, # int, np.ndarray]], Tensor], Sequence[Tensor], Tensor, np.ndarray]`. super().__init__(values) def _init_values( self, values: Union[Mapping[SlateItem, Tensor], Sequence[Tensor], Tensor, np.ndarray], ): if isinstance(values, Tensor): # pyre-fixme[16]: `SlateItemFeatures` has no attribute `_values`. self._values = values.to(dtype=torch.double) elif isinstance(values, np.ndarray): self._values = torch.as_tensor(values, dtype=torch.double) elif isinstance(values, Sequence): # pyre-fixme[6]: Expected `Union[typing.List[Tensor], # typing.Tuple[Tensor, ...]]` for 1st param but got `Sequence[Tensor]`. self._values = torch.stack(values).to(dtype=torch.double) elif isinstance(values, Mapping): self._key_to_index = dict(zip(values.keys(), range(len(values)))) self._index_to_key = list(values.keys()) self._values = torch.stack(list(values.values())).to(dtype=torch.double) else: raise TypeError(f"Unsupported values type {type(values)}") def _to_key(self, k: int) -> SlateItem: return SlateItem(k) @property def items(self) -> SlateItems: if self.is_sequence: return SlateItems(len(self)) else: return SlateItems(super().keys) # SlateSlotFeatures = SlateSlotObjects[Tensor] class SlateSlotFeatures(SlateSlotObjects[Tensor]): @property def features(self) -> Tensor: # pyre-fixme[16]: `SlateSlotFeatures` has no attribute `_values`. return torch.stack(self._values) class Slate(SlateSlotObjects[SlateItem]): """ Class represents a slate: map from slots to items/docs """ def one_hots(self, items: SlateItems, device=None) -> Tensor: oh = torch.zeros((len(self), len(items)), dtype=torch.double, device=device) # pyre-fixme[16]: `Slate` has no attribute `_values`. for t, i in zip(oh, self._values): t[items.index_of(i)] = 1.0 return oh @property def items(self) -> Sequence[SlateItem]: return super().values def slot_values(self, item_values: SlateItemValues) -> SlateSlotValues: """ Map items in the slate to given values Args: item_values: Map from all items to some values Returns: List of values in the slate """ if self._key_to_index is None: # pyre-fixme[16]: `Slate` has no attribute `_values`. return SlateSlotValues([item_values[i] for i in self._values]) else: return SlateSlotValues({k: item_values[i] for k, i in self._key_to_index}) def slot_features(self, item_features: SlateItemFeatures) -> SlateSlotFeatures: """ Map items in the slate to given values Args: item_values: Map from all items to some values Returns: List of values in the slate """ if self._key_to_index is None: return SlateSlotFeatures( # pyre-fixme[16]: `Slate` has no attribute `_values`. [item_features[i].detach().clone() for i in self._values] ) else: return SlateSlotFeatures( {k: item_features[i].detach().clone() for k, i in self._key_to_index} ) def __repr__(self): return f"{self.__class__.__name__}{{value[{self._values}]}}" def make_slate(slots: SlateSlots, items: Sequence[SlateItem]) -> Slate: """ Assign items to slots to make a slate """ assert len(items) >= len(slots) if slots.is_sequence: return Slate(list(items[: len(slots)])) else: return Slate(dict(zip(slots, items[: len(slots)]))) class SlateSlotItemValues(SlateSlotObjects[SlateItemValues]): def __init__( self, values: Union[ MutableMapping[SlateSlot, SlateItemValues], MutableSequence[SlateItemValues] ], ): super().__init__(values) # pyre-fixme[16]: `SlateSlotItemValues` has no attribute `_values`. self._item_size = len(self._values[0]) for v in self._values[1:]: assert self._item_size == len(v) def values_tensor(self, device=None) -> Tensor: # pyre-fixme[16]: `SlateSlotItemValues` has no attribute `_values`. dist = [v.values for v in self._values] return torch.stack(dist).to(device=device) class SlateSlotItemExpectations(SlateSlotItemValues): def expected_rewards( self, item_rewards: SlateItemValues, device=None ) -> SlateSlotValues: """ Calculate expected relevances of each slot, given each item's relevances, under this distribution Args: item_rewards: device: Returns: Map of slots to their expected relevance """ dist = self.values_tensor(device) rewards = item_rewards.values.to(device=device) rewards = torch.mm(dist, rewards.unsqueeze(0).t()).squeeze() if self.is_sequence: return SlateSlotValues(rewards) else: return SlateSlotValues(dict(zip(self.slots, rewards.tolist()))) @property def expectations(self) -> Sequence[SlateItemValues]: return super().values def make_slot_item_distributions( slots: SlateSlots, dists: Sequence[SlateItemValues] ) -> SlateSlotItemExpectations: assert len(dists) >= len(slots) if slots.is_sequence: return SlateSlotItemExpectations(list(dists[: len(slots)])) else: return SlateSlotItemExpectations(dict(zip(slots, dists[: len(slots)]))) def is_to_calculate_expectation(slate_size: int, item_size: int) -> bool: """ Switch between calculating and sampling expectations, balanced by execution time and accuracy Return: True to calculate False to sample """ return ( slate_size < 4 or (slate_size == 4 and item_size < 182) or (slate_size == 5 and item_size < 47) or (slate_size == 6 and item_size < 22) or (slate_size == 7 and item_size < 15) ) def _calculate_slot_expectation( d_out: Tensor, probs: Sequence[float], buffer: Iterable[Tuple[Set[int], float, float, float]], ) -> Iterable[Tuple[Set[int], float, float, float]]: """ A helper function to calculate items' expectations for a slot """ assert d_out.shape[0] == len(probs) next_buffer = [] for b0, b1, b2, _ in buffer: # memory buffer for all ordered combinations so far, list of tuples of # b0: all the items in this ordered combination # b1: cumulative probability of b0 # b2: sum of the probabilities of b0 # b3: = b1 / (1.0 - b2) cached value for faster computation for i, i_prob in enumerate(probs): # only add i if it's not already in if i in b0: continue # nb* are next buffer values nb2 = b2 + i_prob # due to precision errors, sometimes nb2 becomes 1, in this # case, discard the combination if nb2 < 1.0: nb1 = b1 * i_prob / (1.0 - b2) next_buffer.append(({*b0, i}, nb1, nb2, nb1 / (1.0 - nb2))) for i, i_prob in enumerate(probs): p = 0.0 for b0, _, _, b3 in next_buffer: if i in b0: continue p += b3 d_out[i] = p * i_prob return next_buffer class SlateItemProbabilities(SlateItemValues): """ Probabilities of each item being selected into the slate """ def __init__( self, values: Union[Mapping[SlateItem, float], Sequence[float], np.ndarray, Tensor], greedy: bool = False, ): super().__init__(values) self._greedy = greedy self._slot_item_expectations = None def _to_key(self, k: int) -> SlateItem: return SlateItem(k) def _reset(self): super()._reset() self._slot_item_expectations = None def slate_probability(self, slate: Slate) -> Probability: """ Calculate probability of a slate under this distribution Args: slate: Returns: probability """ if self._greedy: items = super().greedy(len(slate)) for i1, i2 in zip(items, slate.items): if i1 != i2: return 0.0 return 1.0 else: # pyre-fixme[16]: `SlateItemProbabilities` has no attribute `_values`. clamped = torch.clamp(self._values, 0.0) indices = [self.index_of(item) for _, item in slate] probs = clamped[indices] sums = clamped[indices] clamped[indices] = 0.0 sums = sums.flip(0).cumsum(0).flip(0) + clamped.sum() return Probability((probs / sums).prod().item()) def slot_item_expectations(self, slots: SlateSlots) -> SlateSlotItemExpectations: slate_size = len(slots) if ( self._slot_item_expectations is not None and len(self._slot_item_expectations) >= slate_size ): return self._slot_item_expectations item_size = len(self) assert item_size >= slate_size if self._greedy: self._slot_item_expectations = make_slot_item_distributions( slots, # pyre-fixme[6]: Expected `Sequence[SlateItemValues]` for 2nd param # but got `List[Values[typing.Any]]`. [ self.replace(torch.zeros(item_size, dtype=torch.double)) for _ in range(len(self)) ], ) sorted_items, _ = self.sort() for item, ds in zip( sorted_items, self._slot_item_expectations.expectations ): ds[item] = 1.0 else: self._normalize() if is_to_calculate_expectation(len(slots), len(self)): self._calculate_expectations(slots) else: self._sample_expectations(slots, 20000) return self._slot_item_expectations def _sample_expectations(self, slots: SlateSlots, num_samples: int): slate_size = len(slots) item_size = len(self) dm = torch.zeros((slate_size, item_size), dtype=torch.double) ri = torch.arange(slate_size) # pyre-fixme[16]: `SlateItemProbabilities` has no attribute `_probabilities`. ws = self._probabilities.repeat((num_samples, 1)) for _ in range(item_size): samples = torch.multinomial(ws, slate_size) for sample in samples: dm[ri, sample] += 1 dm /= num_samples * item_size self._slot_item_expectations = make_slot_item_distributions( slots, # pyre-fixme[6]: Expected `Sequence[SlateItemValues]` for 2nd param but # got `List[Values[typing.Any]]`. [self.replace(vs) for vs in dm], ) def _calculate_expectations(self, slots: SlateSlots): """ A brute-force way to calculate each item's expectations at each slot by going through all l-choose-m (l!/(l-m)!) possible slates. """ slate_size = len(slots) item_size = len(self) dm = torch.zeros((slate_size, item_size), dtype=torch.double) # pyre-fixme[16]: `SlateItemProbabilities` has no attribute `_probabilities`. dm[0] = self._probabilities buffer = [(set(), 1.0, 0.0, 1.0)] probs = self._probabilities.tolist() for d in dm[1:]: buffer = _calculate_slot_expectation(d, probs, buffer) self._slot_item_expectations = make_slot_item_distributions( slots, # pyre-fixme[6]: Expected `Sequence[SlateItemValues]` for 2nd param but # got `List[Values[typing.Any]]`. [self.replace(vs) for vs in dm], ) def sample_slate(self, slots: SlateSlots) -> Slate: slate_size = len(slots) if self._greedy: items = super().greedy(slate_size) else: items = super().sample(slate_size) return make_slate(slots, items) @property def is_deterministic(self) -> bool: return self._greedy def slate_space( self, slots: SlateSlots, max_size: int = -1 ) -> Iterable[Tuple[Sequence[SlateItem], float]]: """Return all possible slates and their probabilities The algorithm is similar to :func:`~_calculate_expectations`, but has less value to cache thus save both space and computation Args: slots: slots to be filled max_size: max number of samples to be returned <= 0 return all samples """ slate_size = len(slots) item_size = len(self) assert item_size >= slate_size if self._greedy: items = super().greedy(slate_size) return [(items, 1.0)] else: buffer = [([], 1.0, 0.0)] # pyre-fixme[16]: `SlateItemProbabilities` has no attribute # `_probabilities`. probs = self._probabilities.tolist() for _ in range(slate_size): next_buffer = [] for b0, b1, b2 in buffer: # memory buffer for all ordered combinations so far, list of tuples of # b0: all the items in this ordered combination # b1: cumulative probability of b0 # b2: sum of the probabilities of b0 for i, i_prob in enumerate(probs): if i in b0: continue nb2 = b2 + i_prob if nb2 < 1.0: nb1 = b1 * i_prob / (1.0 - b2) next_buffer.append(([*b0, i], nb1, nb2)) if max_size <= 0 or max_size > len(next_buffer): buffer = next_buffer else: buffer = random.sample(next_buffer, max_size) return [([SlateItem(i) for i in b[0]], b[1]) for b in buffer] class SlateSlotItemProbabilities(SlateSlotItemValues): def __init__( self, values: Union[ MutableMapping[SlateSlot, SlateItemValues], MutableSequence[SlateItemValues] ], greedy: bool = False, ): super().__init__(values) self._greedy = greedy self._slot_item_distributions = None self._slot_item_expectations = None def slate_probability(self, slate: Slate) -> Probability: """ Calculate probability of a slate under this distribution Args: slate: Returns: probability """ assert len(slate) <= len(self) if self._greedy: for slot, item in slate: probs = self[slot] its, _ = probs.sort() if its[0] != item: return 0.0 return 1.0 else: p = 1.0 last_items = [] for slot, item in slate: item_probs = self[slot] w = 1.0 for last_item in last_items: w -= item_probs.probability(last_item) if math.fabs(w - 0.0) < 1.0e-10: return 0.0 p *= item_probs.probability(item) / w last_items.append(item) return p def slot_item_expectations(self, samples: int = 20000) -> SlateSlotItemExpectations: slate_size = len(self.slots) if ( self._slot_item_expectations is not None and len(self._slot_item_expectations) >= slate_size ): return self._slot_item_expectations # pyre-fixme[16]: `SlateSlotItemProbabilities` has no attribute `_values`. item_size = len(self._values[0]) assert item_size >= slate_size ps = self.values_tensor() if self._greedy: dists = [] for i, value in zip(range(slate_size), self._values): item = ps[i].argmax().item() dist = torch.zeros(item_size, dtype=torch.double) dist[item] = 1.0 dists.append(value.replace(dist)) ps[torch.arange(i + 1, slate_size), item] = 0.0 self._slot_item_expectations = make_slot_item_distributions( self.slots, dists ) else: if is_to_calculate_expectation(slate_size, item_size): self._calculate_expectations() else: self._sample_expectations(samples * item_size) return self._slot_item_expectations def _sample_expectations(self, num_samples: int): slate_size = len(self.slots) # pyre-fixme[16]: `SlateSlotItemProbabilities` has no attribute `_values`. item_size = len(self._values[0]) dm = torch.zeros((slate_size, item_size), dtype=torch.double) ri = torch.arange(slate_size) for _ in range(num_samples): ps = self.values_tensor() sample = [] for i in range(slate_size): item = ps[i].multinomial(1) sample.append(item) ps[torch.arange(i + 1, slate_size), item] = 0.0 dm[ri, sample] += 1 dm /= num_samples self._slot_item_expectations = make_slot_item_distributions( self.slots, [ivs.replace(vs) for ivs, vs in zip(self._values, dm)] ) def _calculate_expectations(self): slate_size = len(self.slots) item_size = len(self._values[0]) dm = torch.zeros((slate_size, item_size), dtype=torch.double) prob_list = [] for v in self._values: v._normalize() prob_list.append(v._probabilities.detach().clone()) dm[0] = prob_list[0] buffer = [({}, 1.0, 0.0, 1.0)] for d, probs in zip(dm[1:], prob_list[1:]): buffer = _calculate_slot_expectation(d, probs.tolist(), buffer) self._slot_item_expectations = make_slot_item_distributions( self.slots, [its.replace(vs) for its, vs in zip(self._values, dm)] ) def sample_slate(self, slots: SlateSlots) -> Slate: slate_size = len(slots) ps = self.values_tensor() items = [] if self._greedy: # pyre-fixme[16]: `SlateSlotItemProbabilities` has no attribute `_values`. for i, value in zip(range(slate_size), self._values): item = ps[i].argmax().item() items.append(value.items[item]) ps[torch.arange(i + 1, slate_size), item] = 0.0 else: for i, value in zip(range(slate_size), self._values): item = ps[i].multinomial(1).item() items.append(value.items[item]) ps[torch.arange(i + 1, slate_size), item] = 0.0 return make_slate(slots, items) class RewardDistribution(ABC): """ Return customized probability distribution according to rewards """ def __init__(self, deterministic: bool = False): self._deterministic = deterministic @abstractmethod def distribution(self, rewards: Tensor) -> Tensor: pass def __call__(self, rewards: SlateItemValues) -> SlateItemProbabilities: dist = self.distribution(rewards.values) return SlateItemProbabilities(rewards.items.fill(dist), self._deterministic) @property @abstractmethod def name(self) -> str: pass class PassThruDistribution(RewardDistribution): """ No-op distribution, probability determined by reward """ def distribution(self, rewards: Tensor) -> Tensor: return rewards.detach().clone() @property def name(self) -> str: return f"{self._deterministic}" def __repr__(self): return f"PassThruDistribution[deterministic={self._deterministic}]" class RankingDistribution(RewardDistribution): """ Ranking distribution according to https://arxiv.org/abs/1605.04812 """ def __init__(self, alpha: float = -1.0, deterministic: bool = False): super().__init__(deterministic) self._alpha = alpha def distribution(self, rewards: Tensor) -> Tensor: dist = rewards.detach().clone() if self._alpha >= 0: _, ids = torch.sort(rewards, descending=True) rank = torch.arange(1, ids.shape[0] + 1, dtype=torch.double) dist[ids] = torch.pow( 2.0, # pyre-fixme[16]: `float` has no attribute `floor_`. (-1.0 * (self._alpha * torch.log2(rank)).floor_()), ) return dist @property def name(self) -> str: return f"ranking_{self._alpha}_{self._deterministic}" def __repr__(self): return ( f"RankingDistribution[alpha={self._alpha}" f",deterministic={self._deterministic}]" ) class FrechetDistribution(RewardDistribution): """ Frechet distribution """ def __init__(self, shape: float, deterministic: bool = False): super().__init__(deterministic) self._shape = shape def distribution(self, rewards: Tensor) -> Tensor: return torch.pow(rewards, self._shape) @property def name(self) -> str: return f"frechet_{self._shape}_{self._deterministic}" def __repr__(self): return ( f"FrechetDistribution[shape={self._shape}]" f",deterministic={self._deterministic}]" ) SlateQueryType = Union[Tuple[int], Tuple[float], np.ndarray, Tensor, Tuple[int, int]] SlateQuery = TypeWrapper[SlateQueryType] @dataclass(frozen=True) class SlateContext: query: SlateQuery slots: SlateSlots params: object = None class SlatePolicy(ABC): """ Policy interface """ def __init__(self, device=None): self.device = device @abstractmethod def _query(self, context: SlateContext) -> SlateItemProbabilities: pass def __call__(self, context: SlateContext) -> SlateItemProbabilities: return self._query(context) class SlateMetric: """ Metric calculator for a slate: weights (dot) rewards Base class is just sum of the all item rewards """ def __init__(self, device=None): self._device = device def calculate_reward( self, slots: SlateSlots, rewards: Optional[SlateSlotValues] = None, slot_values: Optional[SlateSlotValues] = None, slot_weights: Optional[SlateSlotValues] = None, ) -> float: if slot_values is None: assert rewards is not None slot_values = self.slot_values(rewards) values = slot_values.values.to(device=self._device) if slot_weights is None: slot_weights = self.slot_weights(slots) weights = slot_weights.values.to(device=self._device) return torch.tensordot(values, weights, dims=([0], [0])).item() def __call__(self, slots: SlateSlots, rewards: SlateSlotValues) -> float: return self.calculate_reward(slots, rewards) def slot_weights(self, slots: SlateSlots) -> SlateSlotValues: return slots.fill([1.0] * len(slots)) def slot_values(self, rewards: SlateSlotValues) -> SlateSlotValues: return rewards class DCGSlateMetric(SlateMetric): _weights: Optional[Tensor] = None def _get_discount(self, slate_size: int) -> Tensor: weights = DCGSlateMetric._weights if ( weights is None or weights.shape[0] < slate_size or weights.device != self._device ): DCGSlateMetric._weights = torch.reciprocal( torch.log2( torch.arange( 2, slate_size + 2, dtype=torch.double, device=self._device ) ) ) weights = DCGSlateMetric._weights assert weights is not None return weights[:slate_size] def slot_weights(self, slots: SlateSlots) -> SlateSlotValues: return slots.fill(self._get_discount(len(slots))) def slot_values(self, rewards: SlateSlotValues) -> SlateSlotValues: # pyre-fixme[7]: Expected `SlateSlotValues` but got `Values[typing.Any]`. return rewards.replace(torch.pow(2.0, rewards.values) - 1.0) class NDCGSlateMetric(DCGSlateMetric): def __init__(self, item_rewards: SlateItemValues, device=None): super().__init__(device) self._sorted_items, _ = item_rewards.sort() self._item_rewards = item_rewards self._idcg = {} def slot_weights(self, slots: SlateSlots) -> SlateSlotValues: slate_size = len(slots) assert len(self._sorted_items) >= slate_size if slate_size not in self._idcg: i_slate = make_slate(slots, self._sorted_items[:slate_size]) idcg = super().calculate_reward( slots, i_slate.slot_values(self._item_rewards), None, super().slot_weights(slots), ) self._idcg[slate_size] = idcg else: idcg = self._idcg[slate_size] return slots.fill( torch.zeros(slate_size, dtype=torch.double) if idcg == 0 else self._get_discount(slate_size) / idcg ) class ERRSlateMetric(SlateMetric): def __init__(self, max_reward: float, device=None): super().__init__(device) self._max_reward = max_reward def slot_weights(self, slots: SlateSlots) -> SlateSlotValues: return slots.fill([1.0 / (r + 1) for r in range(len(slots))]) def slot_values(self, rewards: SlateSlotValues) -> SlateSlotValues: d = torch.tensor(self._max_reward, device=self._device).pow(2.0) r = (torch.pow(2.0, rewards.values.clamp(0.0, self._max_reward)) - 1.0) / d p = 1.0 err = torch.zeros(len(rewards), dtype=torch.double, device=self._device) for i in range(len(rewards)): ri = r[i] err[i] = p * ri p = p * (1.0 - ri.item()) # pyre-fixme[7]: Expected `SlateSlotValues` but got `Values[typing.Any]`. return rewards.replace(err) class SlateModel(ABC): """ Model providing item relevance/reward, slot examination (click) distribution """ @abstractmethod def item_rewards(self, context: SlateContext) -> SlateItemValues: """ Returns each item's relevance under the context Args: context: Returns: Item relevances """ pass def slot_probabilities(self, context: SlateContext) -> SlateSlotValues: """ Returns each slot/positions's probability independent of showing item, used in PBM estimator Args: context: Returns: """ return context.slots.fill(torch.ones(len(context.slots), dtype=torch.double)) @dataclass(frozen=True) class LogSample: context: SlateContext metric: SlateMetric log_slate: Slate log_reward: Reward _log_slate_probability: Probability = float("nan") # probability for each item being places at each slot _log_slot_item_probabilities: Optional[SlateSlotItemProbabilities] = None # item probability distribution from behavior policy _log_item_probabilities: Optional[SlateItemProbabilities] = None _tgt_slate_probability: Probability = float("nan") _tgt_slot_item_probabilities: Optional[SlateSlotItemProbabilities] = None # item probability distribution from target policy _tgt_item_probabilities: Optional[SlateItemProbabilities] = None # gt_item_rewards: Optional[SlateItemValues] = None # pre-calculated ground truth for target policy ground_truth_reward: Reward = float("nan") # context dependent slot weights (e.g. DCG or ERR weights), used by PBM slot_weights: Optional[SlateSlotValues] = None # item/action independent examination probabilities of each slot, used by PBM slot_probabilities: Optional[SlateSlotValues] = None # features associated with the slate, to train direct model item_features: Optional[SlateItemFeatures] = None def validate(self): slate_size = len(self.context.slots) item_size = len(self.items) assert len(self.log_slate) == slate_size assert ( math.isnan(self._log_slate_probability) or self._log_slate_probability <= 1.0 ) assert ( math.isnan(self._tgt_slate_probability) or self._tgt_slate_probability <= 1.0 ) assert ( self._log_slot_item_probabilities is None or len(self._log_slot_item_probabilities) == slate_size ) assert ( self._log_item_probabilities is None or len(self._log_item_probabilities) == item_size ) assert ( self._tgt_slot_item_probabilities is None or len(self._tgt_slot_item_probabilities) == slate_size ) assert ( self._tgt_item_probabilities is None or len(self._tgt_item_probabilities) == item_size ) assert self.slot_weights is None or len(self.slot_weights) == slate_size assert ( self.slot_probabilities is None or len(self.slot_probabilities) == slate_size ) def log_slot_item_expectations( self, slots: SlateSlots ) -> Optional[SlateSlotItemExpectations]: if self._log_slot_item_probabilities is not None: return self._log_slot_item_probabilities.slot_item_expectations() if self._log_item_probabilities is not None: return self._log_item_probabilities.slot_item_expectations(slots) return None def log_slate_probability(self, slate: Optional[Slate] = None) -> float: if not math.isnan(self._log_slate_probability): return self._log_slate_probability if slate is None: slate = self.log_slate if self._log_slot_item_probabilities is not None: return self._log_slot_item_probabilities.slate_probability(slate) if self._log_item_probabilities is not None: return self._log_item_probabilities.slate_probability(slate) return 0.0 def tgt_slot_expectations( self, slots: SlateSlots ) -> Optional[SlateSlotItemExpectations]: if self._tgt_slot_item_probabilities is not None: return self._tgt_slot_item_probabilities.slot_item_expectations() if self._tgt_item_probabilities is not None: return self._tgt_item_probabilities.slot_item_expectations(slots) return None def tgt_slate_probability(self) -> float: if not math.isnan(self._tgt_slate_probability): return self._tgt_slate_probability if self._tgt_slot_item_probabilities is not None: return self._tgt_slot_item_probabilities.slate_probability(self.log_slate) if self._tgt_item_probabilities is not None: return self._tgt_item_probabilities.slate_probability(self.log_slate) return 0.0 def tgt_slate_space( self, slots: SlateSlots ) -> Iterable[Tuple[Sequence[SlateItem], float]]: if self._tgt_item_probabilities is not None: return self._tgt_item_probabilities.slate_space(slots) return [] @property def items(self) -> SlateItems: if self._log_slot_item_probabilities is not None: # pyre-fixme[16]: `SlateSlotItemProbabilities` has no attribute `_values`. return self._log_slot_item_probabilities._values[0].items if self._log_item_probabilities is not None: return self._log_item_probabilities.items return SlateItems(0) @dataclass(frozen=True) class SlateEstimatorInput: samples: Sequence[LogSample] def validate(self): for s in self.samples: s.validate() class SlateEstimator(Estimator): @abstractmethod def _evaluate_sample(self, sample: LogSample) -> Optional[EstimatorSampleResult]: pass class DMEstimator(SlateEstimator): """ Direct Method estimator """ def __init__(self, trainer: Trainer, training_sample_ratio: float, device=None): super().__init__(device) self._trainer = trainer self._training_sample_ratio = training_sample_ratio def _train_model( self, samples: Sequence[LogSample] ) -> Optional[Iterable[LogSample]]: if self._trainer is None: logger.error("Target model trainer is none, DM is not available") return None self._trainer.reset() logger.info(" training direct model...") st = time.perf_counter() sample_size = len(samples) if self._training_sample_ratio > 0.0 and self._training_sample_ratio < 1.0: training_samples = range(int(sample_size * self._training_sample_ratio)) else: training_samples = range(sample_size) train_x = [] train_y = [] vali_mask = [True] * len(samples) for i in training_samples: sample = samples[i] if sample.item_features is None: continue slate_features = sample.log_slate.slot_features(sample.item_features) train_x.append(slate_features.features.flatten()) train_y.append(sample.log_reward) vali_mask[i] = False if len(train_x) == 0: logger.error("Slate features not provided, DM is not available") return None train_x = torch.stack(train_x) train_y = torch.tensor(train_y, dtype=torch.double, device=train_x.device) vali_x = [] vali_y = [] evaluate_samples = [] for mask, sample in zip(vali_mask, samples): if not mask or sample.item_features is None: continue slate_features = sample.log_slate.slot_features(sample.item_features) vali_x.append(slate_features.features.flatten()) vali_y.append(sample.log_reward) evaluate_samples.append(sample) if len(vali_x) == 0: vali_x = train_x.detach().clone() vali_y = train_y.detach().clone() evaluate_samples = samples else: vali_x = torch.stack(vali_x) vali_y = torch.tensor(vali_y, dtype=torch.double, device=vali_x.device) training_data = TrainingData(train_x, train_y, None, vali_x, vali_y, None) self._trainer.train(training_data) logger.info(f" training direct model done: {time.perf_counter() - st}s") return evaluate_samples def _evaluate_sample(self, sample: LogSample) -> Optional[EstimatorSampleResult]: slots = sample.context.slots tgt_slate_space = sample.tgt_slate_space(slots) features = [] probs = [] for items, prob in tgt_slate_space: slate = make_slate(slots, items) assert sample.item_features is not None slate_features = slate.slot_features(sample.item_features) features.append(slate_features.features.flatten()) probs.append(prob) preds = self._trainer.predict(torch.stack(features), device=self._device) tgt_reward = torch.dot( preds.scores, torch.tensor(probs, dtype=torch.double, device=self._device) ) return EstimatorSampleResult( sample.log_reward, tgt_reward.item(), sample.ground_truth_reward, float("nan"), ) # pyre-fixme[14]: `evaluate` overrides method defined in `Estimator` inconsistently. def evaluate( self, input: SlateEstimatorInput, *kwargs ) -> Optional[EstimatorResult]: input.validate() samples = self._train_model(input.samples) if samples is None: return None log_avg = RunningAverage() tgt_avg = RunningAverage() gt_avg = RunningAverage() for sample in samples: result = self._evaluate_sample(sample) if result is None: continue log_avg.add(result.log_reward) tgt_avg.add(result.target_reward) gt_avg.add(result.ground_truth_reward) return EstimatorResult( log_avg.average, tgt_avg.average, gt_avg.average, tgt_avg.count ) def __repr__(self): return ( f"DMEstimator(trainer({self._trainer.name})" f",ratio({self._training_sample_ratio}),device({self._device}))" ) class IPSEstimator(SlateEstimator): def __init__( self, weight_clamper: Optional[Clamper] = None, weighted: bool = True, device=None, ): super().__init__(device) self._weight_clamper = ( weight_clamper if weight_clamper is not None else Clamper() ) self._weighted = weighted def _evaluate_sample(self, sample: LogSample) -> Optional[EstimatorSampleResult]: tgt_prob = sample.tgt_slate_probability() log_prob = sample.log_slate_probability(sample.log_slate) if tgt_prob == log_prob: weight = 1.0 elif tgt_prob <= 0.0: weight = 0.0 elif log_prob <= 0.0: return None else: weight = self._weight_clamper(tgt_prob / log_prob) return EstimatorSampleResult( sample.log_reward, sample.log_reward * weight, sample.ground_truth_reward, weight, ) # pyre-fixme[14]: `evaluate` overrides method defined in `Estimator` inconsistently. def evaluate( self, input: SlateEstimatorInput, *kwargs ) -> Optional[EstimatorResult]: input.validate() log_avg = RunningAverage() tgt_avg = RunningAverage() acc_weight = RunningAverage() gt_avg = RunningAverage() zw = 0 for sample in input.samples: result = self._evaluate_sample(sample) if result is None: zw += 1 continue log_avg.add(result.log_reward) tgt_avg.add(result.target_reward) gt_avg.add(result.ground_truth_reward) acc_weight.add(result.weight) if result.weight == 0.0: zw += 1 logging.info( f"IPSEstimator invalid sample pct: {zw * 100 / len(input.samples)}%" ) if tgt_avg.count == 0: return None if self._weighted: estimated = tgt_avg.total / acc_weight.total return EstimatorResult( log_avg.average, estimated, gt_avg.average, acc_weight.average ) else: return EstimatorResult( log_avg.average, tgt_avg.average, gt_avg.average, tgt_avg.count ) def __repr__(self): return ( f"IPSEstimator(weight_clamper({self._weight_clamper})" f",weighted({self._weighted}),device({self._device}))" ) class DoublyRobustEstimator(DMEstimator): def __init__( self, trainer: Trainer, training_sample_ratio: float, weight_clamper: Optional[Clamper] = None, weighted: bool = False, device=None, ): super().__init__(trainer, training_sample_ratio, device) self._weight_clamper = ( weight_clamper if weight_clamper is not None else Clamper() ) self._weighted = weighted def _evaluate_sample(self, sample: LogSample) -> Optional[EstimatorSampleResult]: slots = sample.context.slots if self._trainer.is_trained: tgt_slate_space = sample.tgt_slate_space(slots) features = [] probs = [] for items, prob in tgt_slate_space: slate = make_slate(slots, items) assert sample.item_features is not None slate_features = slate.slot_features(sample.item_features) features.append(slate_features.features.flatten()) probs.append(prob) preds = self._trainer.predict(torch.stack(features), device=self._device) dm_reward = torch.dot( preds.scores, torch.tensor(probs, dtype=torch.double, device=self._device), ).item() assert sample.item_features is not None log_slate_feature = sample.log_slate.slot_features(sample.item_features) pred = self._trainer.predict( torch.unsqueeze(log_slate_feature.features.flatten(), dim=0), device=self._device, ) log_dm_reward = pred.scores[0].item() else: dm_reward = 0.0 log_dm_reward = 0.0 tgt_prob = sample.tgt_slate_probability() log_prob = sample.log_slate_probability(sample.log_slate) if tgt_prob == log_prob: weight = 1.0 elif tgt_prob <= 0.0: weight = 0.0 elif log_prob <= 0.0: return None else: weight = self._weight_clamper(tgt_prob / log_prob) target_reward = (sample.log_reward - log_dm_reward) * weight + dm_reward return EstimatorSampleResult( sample.log_reward, target_reward, sample.ground_truth_reward, weight ) def evaluate( self, input: SlateEstimatorInput, *kwargs ) -> Optional[EstimatorResult]: input.validate() samples = self._train_model(input.samples) if samples is None: samples = input.samples log_avg = RunningAverage() tgt_avg = RunningAverage() acc_weight = RunningAverage() gt_avg = RunningAverage() for sample in samples: result = self._evaluate_sample(sample) if result is None: continue log_avg.add(result.log_reward) tgt_avg.add(result.target_reward) acc_weight.add(result.weight) gt_avg.add(result.ground_truth_reward) if self._weighted: estimated = tgt_avg.total / acc_weight.total return EstimatorResult( log_avg.average, estimated, gt_avg.average, acc_weight.average ) else: return EstimatorResult( log_avg.average, tgt_avg.average, gt_avg.average, tgt_avg.count ) def __repr__(self): return ( f"DoublyRobustEstimator(trainer({self._trainer.name})" f",ratio({self._training_sample_ratio})" f",weight_clamper({self._weight_clamper})" f",weighted({self._weighted}),device({self._device}))" ) class PseudoInverseEstimator(SlateEstimator): """ Estimator from reference 2 """ def __init__( self, weight_clamper: Optional[Clamper] = None, weighted: bool = True, device=None, ): super().__init__(device) self._weight_clamper = ( weight_clamper if weight_clamper is not None else Clamper() ) self._weighted = weighted def _evaluate_sample(self, sample: LogSample) -> Optional[EstimatorSampleResult]: log_slot_expects = sample.log_slot_item_expectations(sample.context.slots) if log_slot_expects is None: logger.warning("Log slot distribution not available") return None tgt_slot_expects = sample.tgt_slot_expectations(sample.context.slots) if tgt_slot_expects is None: logger.warning("Target slot distribution not available") return None log_indicator = log_slot_expects.values_tensor(self._device) tgt_indicator = tgt_slot_expects.values_tensor(self._device) lm = len(sample.context.slots) * len(sample.items) gamma = torch.as_tensor( np.linalg.pinv( torch.mm( log_indicator.view((lm, 1)), log_indicator.view((1, lm)) ).numpy() ) ) # torch.pinverse is not very stable # gamma = torch.pinverse( # torch.mm(log_indicator.view((lm, 1)), log_indicator.view((1, lm))) # ) ones = sample.log_slate.one_hots(sample.items, self._device) weight = self._weight_clamper( torch.mm(tgt_indicator.view((1, lm)), torch.mm(gamma, ones.view((lm, 1)))) ).item() return EstimatorSampleResult( sample.log_reward, sample.log_reward * weight, sample.ground_truth_reward, weight, ) # pyre-fixme[14]: `evaluate` overrides method defined in `Estimator` inconsistently. def evaluate( self, input: SlateEstimatorInput, *kwargs ) -> Optional[EstimatorResult]: input.validate() log_avg = RunningAverage() tgt_avg = RunningAverage() acc_weight = RunningAverage() gt_avg = RunningAverage() zw = 0 for sample in input.samples: result = self._evaluate_sample(sample) if result is None: zw += 1 continue log_avg.add(result.log_reward) tgt_avg.add(result.target_reward) gt_avg.add(result.ground_truth_reward) acc_weight.add(result.weight) if result.weight == 0.0: zw += 1 if tgt_avg.count % 1000 == 0: logger.info(f" PseudoInverseEstimator: processed {tgt_avg.count}") logging.info( f"PseudoInverseEstimator invalid sample pct: {zw * 100 / len(input.samples)}%" ) if tgt_avg.count == 0: return None if self._weighted: estimated = tgt_avg.total / acc_weight.total return EstimatorResult( log_avg.average, estimated, gt_avg.average, acc_weight.average ) else: return EstimatorResult( log_avg.average, tgt_avg.average, gt_avg.average, tgt_avg.count ) def __repr__(self): return ( f"PseudoInverseEstimator(weight_clamper({self._weight_clamper})" f",weighted({self._weighted}),device({self._device}))" ) class PBMEstimator(SlateEstimator): """ Estimator from reference 1: Position-Based Click Model """ def __init__( self, weight_clamper: Optional[Clamper] = None, weighted: bool = True, device=None, ): super().__init__(device) self._weight_clamper = ( weight_clamper if weight_clamper is not None else Clamper() ) self._weighted = weighted def _evaluate_sample(self, sample: LogSample) -> Optional[EstimatorSampleResult]: log_slot_expects = sample.log_slot_item_expectations(sample.context.slots) if log_slot_expects is None: logger.warning(" Log slot distribution not available") return None tgt_slot_expects = sample.tgt_slot_expectations(sample.context.slots) if tgt_slot_expects is None: logger.warning(" Target slot distribution not available") return None slate_size = len(sample.context.slots) slot_weights = sample.slot_weights if slot_weights is None: slot_weights = SlateSlotValues(torch.ones(slate_size, dtype=torch.double)) weights = slot_weights.values.to(device=self._device) if sample.slot_probabilities is not None: weights *= sample.slot_probabilities.values h = torch.zeros(slate_size, dtype=torch.double, device=self._device) p = torch.zeros(slate_size, dtype=torch.double, device=self._device) i = 0 for slot, item in sample.log_slate: h[i] = tgt_slot_expects[slot][item] p[i] = log_slot_expects[slot][item] i += 1 nu = torch.tensordot(h, weights, dims=([0], [0])) de = torch.tensordot(p, weights, dims=([0], [0])) if nu == de: weight = 1.0 elif nu == 0: weight = 0.0 elif de == 0: return None else: weight = self._weight_clamper(nu / de) return EstimatorSampleResult( sample.log_reward, sample.log_reward * weight, sample.ground_truth_reward, weight, ) # pyre-fixme[14]: `evaluate` overrides method defined in `Estimator` inconsistently. def evaluate( self, input: SlateEstimatorInput, *kwargs ) -> Optional[EstimatorResult]: input.validate() log_avg = RunningAverage() tgt_avg = RunningAverage() acc_weight = RunningAverage() gt_avg = RunningAverage() zw = 0 for sample in input.samples: result = self._evaluate_sample(sample) if result is None: zw += 1 continue log_avg.add(result.log_reward) tgt_avg.add(result.target_reward) gt_avg.add(result.ground_truth_reward) acc_weight.add(result.weight) if result.weight == 0.0: zw += 1 if tgt_avg.count % 1000 == 0: logger.info(f" PBMEstimator: processed {tgt_avg.count}") logging.info( f"PBMEstimator invalid sample pct: {zw * 100 / len(input.samples)}%" ) if tgt_avg.count == 0: return None if self._weighted: estimated = tgt_avg.total / acc_weight.total return EstimatorResult( log_avg.average, estimated, gt_avg.average, acc_weight.average ) else: return EstimatorResult( log_avg.average, tgt_avg.average, gt_avg.average, tgt_avg.count ) def __repr__(self): return ( f"PBMEstimator(weight_clamper({self._weight_clamper})" f",weighted({self._weighted}),device({self._device}))" )
54,387
34.225389
90
py
ReAgent
ReAgent-master/reagent/ope/estimators/sequential_estimators.py
#!/usr/bin/env python3 import copy import logging import random import time import typing from abc import ABC, abstractmethod from copy import deepcopy from dataclasses import dataclass from enum import Enum from functools import reduce from itertools import count, zip_longest from typing import Callable, Dict, Iterable, Mapping, Optional, Sequence, Tuple, Union import numpy as np import torch from reagent.ope.estimators.estimator import ( Estimator, EstimatorResult, EstimatorResults, ) from reagent.ope.estimators.types import ( Action, ActionDistribution, ActionSpace, Probability, Reward, TypeWrapper, ) from reagent.ope.trainers.linear_trainers import LinearNet from reagent.ope.utils import Clamper, RunningAverage from torch import Tensor StateType = Union[float, Tuple[float], Tuple[int], np.ndarray, Tensor] @dataclass(frozen=True) class State(TypeWrapper[StateType]): is_terminal: bool = False def __repr__(self): return super().__repr__()[:-1] + f",is_terminal[{self.is_terminal}]}}" @dataclass(frozen=True) class StateReward: state: Optional[State] = None reward: Reward = 0.0 @dataclass(frozen=True) class RewardProbability: reward: Reward = 0.0 prob: Probability = 0.0 # State distribution: State -> (reward, probability) StateDistribution = Mapping[State, RewardProbability] @dataclass(frozen=True) class Transition: class Status(Enum): NOOP = 0 NORMAL = 1 TERMINATED = 2 last_state: Optional[State] = None # from state action: Optional[Action] = None # action action_prob: float = 0.0 # action probability state: Optional[State] = None # to state reward: float = 0.0 status: Status = Status.NORMAL # MDP sequence Mdp = Sequence[Transition] class RLPolicy(ABC): """ Policy interface """ def __init__(self, action_space: ActionSpace, device=None): self._action_space = action_space self._device = device @abstractmethod def action_dist(self, state: State) -> ActionDistribution: pass def __call__(self, state: State) -> ActionDistribution: return self.action_dist(state) @property def action_space(self): return self._action_space class RandomRLPolicy(RLPolicy): """ A random policy which return an action according to uniform distribution """ def __init__(self, action_space: ActionSpace, device=None): super().__init__(action_space, device) self._prob = 1.0 / len(action_space) def action_dist(self, state: State) -> ActionDistribution: # pyre-fixme[6]: Expected `Union[Mapping[TypeWrapper[Union[Tuple[float], # Tuple[int], Tensor, float, int, np.ndarray]], float], Sequence[float], # Tensor, np.ndarray]` for 1st param but got `int`. return self._action_space.distribution([self._prob] * len(self._action_space)) class EpsilonGreedyRLPolicy(RLPolicy): """ A wrapper policy: Skewing the wrapped policy action distribution by epsilon Number of total actions must be given, and wrapped policy should calculate probabilities for all actions """ def __init__(self, policy: RLPolicy, epsilon: float = 0.0): assert policy is not None and 0.0 <= epsilon < 1.0 super().__init__(policy._device) self._policy = policy self._exploitation_prob = 1.0 - epsilon self._exploration_prob = epsilon / len(policy.action_space) def action_dist(self, state) -> ActionDistribution: new_dist = deepcopy(self._policy(state)) for a, p in new_dist: new_dist[a] = p * self._exploitation_prob + self._exploration_prob return new_dist class Model(ABC): """ Model interface """ @abstractmethod def next_state_reward_dist(self, state: State, action: Action) -> StateDistribution: pass def __call__(self, state: State, action: Action) -> StateDistribution: return self.next_state_reward_dist(state, action) class ValueFunction(ABC): """ Value function to calculate state and state-action values """ @abstractmethod def state_action_value(self, state: State, action: Action) -> float: pass @abstractmethod def state_value(self, state: State) -> float: pass @abstractmethod def reset(self): pass def __call__(self, state: State, action: Optional[Action] = None) -> float: return ( self.state_action_value(state, action) if action is not None else self.state_value(state) ) @dataclass(frozen=True) class RLEstimatorInput: gamma: float log: Sequence[Mdp] target_policy: RLPolicy value_function: Optional[ValueFunction] = None ground_truth: Optional[ValueFunction] = None horizon: int = -1 discrete_states: bool = True class RLEstimator(Estimator): def _log_reward(self, gamma: float, mdps: Sequence[Mdp]) -> float: avg = RunningAverage() for mdp in mdps: discount = 1.0 r = 0.0 for t in mdp: r += discount * t.reward discount *= gamma avg.add(r) return avg.average def _estimate_value( self, gamma: float, mdps: Sequence[Mdp], value_function: ValueFunction ) -> float: avg = RunningAverage() for mdp in mdps: discount = 1.0 r = 0.0 for t in mdp: if t.last_state is None: break r += discount * value_function(t.last_state) discount *= gamma avg.add(r) return avg.average class DMEstimator(RLEstimator): """ Direct Method estimator """ def evaluate(self, input: RLEstimatorInput, **kwargs) -> EstimatorResults: # kwargs is part of the function signature, so to satisfy pyre it must be included assert input.value_function is not None logging.info(f"{self}: start evaluating") stime = time.process_time() results = EstimatorResults() estimate = self._estimate_value(input.gamma, input.log, input.value_function) if input.ground_truth is not None: gt = self._estimate_value(input.gamma, input.log, input.ground_truth) results.append( EstimatorResult( self._log_reward(input.gamma, input.log), estimate, # pyre-fixme[61]: `gt` may not be initialized here. None if input.ground_truth is None else gt, ) ) logging.info( f"{self}: finishing evaluating[" f"process_time={time.process_time() - stime}]" ) return results class IPSEstimator(RLEstimator): """ IPS estimator """ def __init__( self, weight_clamper: Optional[Clamper] = None, weighted: bool = True, device=None, ): super().__init__(device) self._weight_clamper = ( weight_clamper if weight_clamper is not None else Clamper() ) self._weighted = weighted def _calc_weights( self, episodes: int, horizon: int, mdp_transitions: Iterable[Iterable[Transition]], policy: RLPolicy, ) -> torch.Tensor: pi_e = torch.ones((episodes, horizon)) pi_b = torch.ones((episodes, horizon)) mask = torch.ones((episodes, horizon)) j = 0 for ts in mdp_transitions: i = 0 for t in ts: if t is not None and t.action is not None and t.action_prob > 0.0: assert t.last_state is not None pi_e[i, j] = policy(t.last_state)[t.action] pi_b[i, j] = t.action_prob else: mask[i, j] = 0.0 i += 1 j += 1 pi_e = pi_e.to(device=self._device) pi_b = pi_b.to(device=self._device) mask = mask.to(device=self._device) rho = pi_e.div_(pi_b).cumprod(1).mul_(mask) if self._weighted: weight = rho.sum(0) else: weight = mask.sum(0) weight.add_(weight.lt(1.0e-15) * episodes) ws = rho / weight return self._weight_clamper(ws) def evaluate(self, input: RLEstimatorInput, **kwargs) -> EstimatorResults: # kwargs is part of the function signature, so to satisfy pyre it must be included logging.info(f"{self}: start evaluating") stime = time.process_time() results = EstimatorResults() n = len(input.log) horizon = len(reduce(lambda a, b: a if len(a) > len(b) else b, input.log)) weights = self._calc_weights( n, horizon, zip_longest(*input.log), input.target_policy ) discount = torch.full((horizon,), input.gamma, device=self._device) discount[0] = 1.0 discount = discount.cumprod(0) rewards = torch.zeros((n, horizon)) j = 0 for ts in zip_longest(*input.log): i = 0 for t in ts: if t is not None: rewards[i, j] = t.reward i += 1 j += 1 rewards = rewards.to(device=self._device) estimate = weights.mul(rewards).sum(0).mul(discount).sum().item() results.append( EstimatorResult( self._log_reward(input.gamma, input.log), estimate, None if input.ground_truth is None else self._estimate_value(input.gamma, input.log, input.ground_truth), ) ) logging.info( f"{self}: finishing evaluating[" f"process_time={time.process_time() - stime}]" ) return results def __repr__(self): return super().__repr__()[0:-1] + f",weighted[{self._weighted}]}}" class DoublyRobustEstimator(IPSEstimator): """ Doubly Robust estimator """ def evaluate(self, input: RLEstimatorInput, **kwargs) -> EstimatorResults: # kwargs is part of the function signature, so to satisfy pyre it must be included logging.info(f"{self}: start evaluating") stime = time.process_time() results = EstimatorResults() n = len(input.log) horizon = len(reduce(lambda a, b: a if len(a) > len(b) else b, input.log)) ws = self._calc_weights( n, horizon, zip_longest(*input.log), input.target_policy ) last_ws = torch.zeros((n, horizon), device=self._device) last_ws[:, 0] = 1.0 / n last_ws[:, 1:] = ws[:, :-1] discount = torch.full((horizon,), input.gamma, device=self._device) discount[0] = 1.0 discount = discount.cumprod(0) rs = torch.zeros((n, horizon)) vs = torch.zeros((n, horizon)) qs = torch.zeros((n, horizon)) for ts, j in zip(zip_longest(*input.log), count()): for t, i in zip(ts, count()): if t is not None and t.action is not None: assert input.value_function is not None qs[i, j] = input.value_function(t.last_state, t.action) vs[i, j] = input.value_function(t.last_state) rs[i, j] = t.reward vs = vs.to(device=self._device) qs = qs.to(device=self._device) rs = rs.to(device=self._device) estimate = ((ws * (rs - qs) + last_ws * vs).sum(0) * discount).sum().item() results.append( EstimatorResult( self._log_reward(input.gamma, input.log), estimate, None if input.ground_truth is None else self._estimate_value(input.gamma, input.log, input.ground_truth), ) ) logging.info( f"{self}: finishing evaluating[" f"process_time={time.process_time() - stime}]" ) return results class MAGICEstimator(IPSEstimator): """ Algorithm from https://arxiv.org/abs/1604.00923, appendix G.3 """ def __init__(self, weight_clamper: Optional[Clamper] = None, device=None): super().__init__(weight_clamper, True, device) def evaluate(self, input: RLEstimatorInput, **kwargs) -> EstimatorResults: assert input.value_function is not None logging.info(f"{self}: start evaluating") stime = time.process_time() results = EstimatorResults() num_resamples = kwargs["num_resamples"] if "num_resamples" in kwargs else 200 loss_threshold = ( kwargs["loss_threshold"] if "loss_threshold" in kwargs else 0.00001 ) lr = kwargs["lr"] if "lr" in kwargs else 0.0001 logging.info( f" params: num_resamples[{num_resamples}], " f"loss_threshold[{loss_threshold}], " f"lr[{lr}]" ) # Compute MAGIC estimate n = len(input.log) horizon = len(reduce(lambda a, b: a if len(a) > len(b) else b, input.log)) ws = self._calc_weights( n, horizon, zip_longest(*input.log), input.target_policy ) last_ws = torch.zeros((n, horizon), device=self._device) last_ws[:, 0] = 1.0 / n last_ws[:, 1:] = ws[:, :-1] discount = torch.full((horizon,), input.gamma, device=self._device) discount[0] = 1.0 discount = discount.cumprod(0) rs = torch.zeros((n, horizon)) vs = torch.zeros((n, horizon)) qs = torch.zeros((n, horizon)) for ts, j in zip(zip_longest(*input.log), count()): for t, i in zip(ts, count()): if t is not None and t.action is not None: qs[i, j] = input.value_function(t.last_state, t.action) vs[i, j] = input.value_function(t.last_state) rs[i, j] = t.reward vs = vs.to(device=self._device) qs = qs.to(device=self._device) rs = rs.to(device=self._device) wdrs = ((ws * (rs - qs) + last_ws * vs) * discount).cumsum(1) wdr = wdrs[:, -1].sum(0) next_vs = torch.zeros((n, horizon), device=self._device) next_vs[:, :-1] = vs[:, 1:] gs = wdrs + ws * next_vs * discount gs_normal = gs.sub(torch.mean(gs, 0)) assert n > 1 omiga = (n / (n - 1.0)) * torch.einsum("ij,ik->jk", gs_normal, gs_normal) resample_wdrs = torch.zeros((num_resamples,)) for i in range(num_resamples): samples = random.choices(range(n), k=n) sws = ws[samples, :] last_sws = last_ws[samples, :] srs = rs[samples, :] svs = vs[samples, :] sqs = qs[samples, :] resample_wdrs[i] = ( ((sws * (srs - sqs) + last_sws * svs).sum(0) * discount).sum().item() ) resample_wdrs, _ = resample_wdrs.to(device=self._device).sort(0) lb = torch.min(wdr, resample_wdrs[int(round(0.05 * num_resamples))]) ub = torch.max(wdr, resample_wdrs[int(round(0.95 * num_resamples)) - 1]) b = torch.tensor( list( map( lambda a: a - ub if a > ub else (a - lb if a < lb else 0.0), gs.sum(0), ) ), device=self._device, ) b.unsqueeze_(0) bb = b * b.t() cov = omiga + bb # x = torch.rand((1, horizon), device=self.device, requires_grad=True) x = torch.zeros((1, horizon), device=self._device, requires_grad=True) # using SGD to find min x optimizer = torch.optim.SGD([x], lr=lr) last_y = 0.0 for i in range(100): x = torch.nn.functional.softmax(x, dim=1) y = torch.mm(torch.mm(x, cov), x.t()) if abs(y.item() - last_y) < loss_threshold: print(f"{i}: {last_y} -> {y.item()}") break last_y = y.item() optimizer.zero_grad() y.backward(retain_graph=True) optimizer.step() x = torch.nn.functional.softmax(x, dim=1) estimate = torch.mm(x, gs.sum(0, keepdim=True).t()).cpu().item() results.append( EstimatorResult( self._log_reward(input.gamma, input.log), estimate, None if input.ground_truth is None else self._estimate_value(input.gamma, input.log, input.ground_truth), ) ) logging.info( f"{self}: finishing evaluating[" f"process_time={time.process_time() - stime}]" ) return results @dataclass class NeuralDualDICE(RLEstimator): # See https://arxiv.org/pdf/1906.04733.pdf sections 4, 5, A # Google's implementation: https://github.com/google-research/google-research/tree/master/dual_dice """ Args: state_dim: The dimensionality of the state vectors action_dim: The number of discrete actions deterministic_env: Whether or not the environment is determinstic. Can help with stability of training. average_next_v: Whether or not to average the next nu value over all possible actions. Can help with stability of training. polynomial_degree: The degree of the convex function f(x) = 1/p * |x|^p value_lr: The learning rate for nu zeta_lr: The learning rate for zeta hidden_dim: The dimensionality of the hidden layers for zeta and v hidden_layers: The number of hidden layers for zeta and v activation: The activation function for zeta and v training_samples: The number of batches to train zeta and v for batch_size: The number of samples in each batch loss_callback_fn: A function that will be called every reporting_frequency batches, giving the average zeta loss, average nu loss, and self reporting_frequency: The number of batches between outputting the state of the training """ state_dim: int action_dim: int deterministic_env: bool average_next_v: bool = False polynomial_degree: float = 1.5 value_lr: float = 0.01 zeta_lr: float = 0.01 hidden_dim: int = 64 hidden_layers: int = 2 activation = torch.nn.Tanh training_samples: int = 100000 batch_size: int = 2048 device: typing.Any = None loss_callback_fn: Optional[Callable[[float, float, RLEstimator], None]] = None reporting_frequency: int = 1000 # These are initialized in __post_init__() and calms Pyre v: typing.Any = None zeta: typing.Any = None f: typing.Any = None fconjugate: typing.Any = None zeta_net: typing.Any = None v_net: typing.Any = None def __post_init__(self): conjugate_exponent = self.polynomial_degree / (self.polynomial_degree - 1) self.f = self._get_convex_f(self.polynomial_degree) self.fconjugate = self._get_convex_f(conjugate_exponent) self.reset() def _get_convex_f(self, degree): return lambda x: (torch.abs(x) ** degree) / degree # pyre-fixme[56]: Decorator `torch.no_grad(...)` could not be called, because # its type `no_grad` is not callable. @torch.no_grad() def _mdps_value(self, mdps: Sequence[Mdp], gamma: float) -> float: self.zeta_net.eval() avg = RunningAverage() for mdp in mdps: discount = 1.0 r = 0.0 for t in mdp: assert t.last_state is not None, "Expected last_state, got None" assert t.action is not None, "Expected action, got None" zeta = self.zeta( torch.tensor(t.last_state.value, dtype=torch.float) .reshape(-1, self.state_dim) .to(self.device), torch.nn.functional.one_hot( torch.tensor(t.action.value, dtype=torch.long), self.action_dim ) .reshape(-1, self.action_dim) .float() .to(self.device), ) r += discount * t.reward * zeta.cpu().item() discount *= gamma avg.add(r) self.zeta_net.train() return avg.average # pyre-fixme[56]: Decorator `torch.no_grad(...)` could not be called, because # its type `no_grad` is not callable. @torch.no_grad() def _compute_estimates(self, input: RLEstimatorInput) -> EstimatorResults: results = EstimatorResults() estimate = self._mdps_value(input.log, input.gamma) results.append( EstimatorResult( self._log_reward(input.gamma, input.log), estimate, None if input.ground_truth is None else self._estimate_value(input.gamma, input.log, input.ground_truth), ) ) return results def _compute_average_v(self, transition): next_vs = [ transition["tgt_action_props"][:, a].reshape(-1, 1) * self.v( transition["state"], torch.nn.functional.one_hot( torch.tensor(a, dtype=torch.long), self.action_dim ) .reshape(1, -1) .float() .to(self.device) .repeat(transition["state"].shape[0], 1), ) for a in range(self.action_dim) ] return sum(next_vs) def _compute_loss( self, gamma: float, transition: Dict, compute_determ_v_loss: bool ): if self.average_next_v: next_v = self._compute_average_v(transition) else: next_v = self.v(transition["state"], transition["next_action"]) delta_v = ( self.v(transition["last_state"], transition["log_action"]) - gamma * next_v ) init_v = self.v(transition["init_state"], transition["init_action"]) if compute_determ_v_loss: unweighted_loss = self.f(delta_v) - (1 - gamma) * init_v else: zeta = self.zeta(transition["last_state"], transition["log_action"]) unweighted_loss = ( delta_v * zeta - self.fconjugate(zeta) - (1 - gamma) * init_v ) weights = torch.full( (unweighted_loss.shape[0], 1), gamma, dtype=torch.float ).to(device=self.device) ** transition["timestep"].reshape((-1, 1)) return torch.sum(weights * unweighted_loss) / torch.sum(weights) def reset(self): self.v_net = LinearNet( self.state_dim + self.action_dim, self.hidden_dim, 1, self.hidden_layers, self.activation, ) self.zeta_net = copy.deepcopy(self.v_net) self.v_net.to(self.device) self.zeta_net.to(self.device) self.v = self._build_function(self.v_net) self.zeta = self._build_function(self.zeta_net) def _build_function(self, net: torch.nn.Module): return lambda s, a: net(torch.cat((s, a), dim=1)) def _collect_data(self, input: RLEstimatorInput): samples = { "init_state": [], "init_action": [], "last_state": [], "state": [], "log_action": [], "next_action": [], "tgt_action_props": [], "timestep": [], "reward": [], } for mdp in input.log: state = mdp[0].last_state assert state is not None, "Expected initial state, got None" tgt_init_action = input.target_policy.action_dist(state).sample()[0] for i, t in enumerate(mdp): assert ( t.state is not None and t.last_state is not None and t.action is not None ), "Expected all fields to be present" tgt_dist = input.target_policy.action_dist(t.state) tgt_action = tgt_dist.sample()[0] samples["init_state"].append(state.value) samples["init_action"].append( torch.nn.functional.one_hot( torch.tensor(tgt_init_action.value, dtype=torch.long), self.action_dim, ).float() ) samples["last_state"].append(t.last_state.value) samples["state"].append(t.state.value) samples["log_action"].append( torch.nn.functional.one_hot( torch.tensor(t.action.value, dtype=torch.long), self.action_dim ).float() ) samples["next_action"].append( torch.nn.functional.one_hot( torch.tensor(tgt_action.value, dtype=torch.long), self.action_dim, ).float() ) samples["tgt_action_props"].append(tgt_dist.values) samples["timestep"].append(i) samples["reward"].append(t.reward) return { k: torch.stack(v).to(self.device) if "action" in k else torch.tensor(v, dtype=torch.float).to(self.device) for k, v in samples.items() } def _sample_batch(self, dataset): idxs = np.random.choice(dataset["init_state"].shape[0], self.batch_size) return {k: v[idxs] for k, v in dataset.items()} def evaluate(self, input: RLEstimatorInput, **kwargs) -> EstimatorResults: stime = time.process_time() dataset = self._collect_data(input) logging.info(f"Data loading time: {time.process_time() - stime}") zeta_optim = torch.optim.Adam(self.zeta_net.parameters(), lr=self.zeta_lr) v_optim = torch.optim.Adam(self.v_net.parameters(), lr=self.value_lr) avg_zeta_loss = RunningAverage() avg_v_loss = RunningAverage() sample_time = time.process_time() for sampled in range(self.training_samples): sample = self._sample_batch(dataset) zeta_loss = -(self._compute_loss(input.gamma, sample, False)) # Populate zeta gradients and optimize zeta_optim.zero_grad() zeta_loss.backward() zeta_optim.step() if self.deterministic_env: v_loss = self._compute_loss(input.gamma, sample, True) else: v_loss = self._compute_loss(*sample) # Populate value gradients and optimize v_optim.zero_grad() v_loss.backward() v_optim.step() avg_zeta_loss.add(zeta_loss.cpu().item()) avg_v_loss.add(v_loss.cpu().item()) if sampled % self.reporting_frequency == 0: report_time = time.process_time() - sample_time callback_time = None if self.loss_callback_fn is not None: # Pyre gets angry if we don't make callback local callback = self.loss_callback_fn assert callback is not None stime = time.process_time() callback(avg_zeta_loss.average, avg_v_loss.average, self) callback_time = abs(time.process_time() - stime) logging.info( f"Samples {sampled}, " f"Avg Zeta Loss {avg_zeta_loss.average}, " f"Avg Value Loss {avg_v_loss.average},\n" f"Time per {self.reporting_frequency} samples: {report_time}" + ( "" if callback_time is None else f", Time for callback: {callback_time}" ) ) avg_zeta_loss = RunningAverage() avg_v_loss = RunningAverage() sample_time = time.process_time() return self._compute_estimates(input)
28,267
34.964377
103
py
ReAgent
ReAgent-master/reagent/ope/estimators/types.py
#!/usr/bin/env python3 import logging import pickle from abc import ABC, abstractmethod from copy import deepcopy from dataclasses import dataclass from typing import Generic, Mapping, Optional, Sequence, Tuple, TypeVar, Union import numpy as np import torch from torch import Tensor def is_array(obj): return isinstance(obj, Tensor) or isinstance(obj, np.ndarray) Type = TypeVar("Type") KeyType = TypeVar("KeyType") ValueType = TypeVar("ValueType") @dataclass(frozen=True) class TypeWrapper(Generic[ValueType]): value: ValueType def __index__(self): try: return int(self.value) except Exception: raise ValueError(f"{self} cannot be used as index") def __int__(self): try: return int(self.value) except Exception: raise ValueError(f"{self} cannot be converted to int") def __hash__(self): if ( isinstance(self.value, int) or isinstance(self.value, float) or isinstance(self.value, tuple) ): return hash(self.value) elif isinstance(self.value, Tensor): return hash(tuple(self.value.numpy().flatten())) elif isinstance(self.value, np.ndarray): return hash(tuple(self.value.flatten())) elif isinstance(self.value, list): return hash(tuple(self.value)) else: raise TypeError def __eq__(self, other): if not isinstance(other, TypeWrapper): return False if isinstance(self.value, Tensor): if isinstance(other.value, Tensor): return torch.equal(self.value, other.value) else: raise TypeError(f"{self} cannot be compared with non-tensor") elif isinstance(self.value, np.ndarray): return np.array_equal(self.value, other.value) else: return self.value == other.value def __ne__(self, other): return not self.__eq__(other) def __lt__(self, other): if not isinstance(other, TypeWrapper): return False if isinstance(self.value, Tensor) and isinstance(other.value, Tensor): return torch.lt(self.value, other.value).prod().item() elif isinstance(self.value, np.ndarray) and isinstance(other.value, np.ndarray): return np.less(self.value, other.value).prod() else: return self.value < other.value def __repr__(self): return f"{self.__class__.__name__}{{value[{self.value}]}}" class Objects(Generic[KeyType, ValueType], ABC): """ Generic class for a map from item to its value. It supports [] indexing, and iterator protocol Attributes: items: list of items values: list of their values """ def __init__(self, values: Union[Mapping[KeyType, ValueType], Sequence[ValueType]]): self._key_to_index = None self._index_to_key = None self._init_values(values) self._reset() def _init_values( self, values: Union[Mapping[KeyType, ValueType], Sequence[ValueType]] ): if isinstance(values, Sequence): # pyre-fixme[16]: `Objects` has no attribute `_values`. self._values = list(values) elif isinstance(values, Mapping): self._key_to_index = dict(zip(values.keys(), range(len(values)))) self._index_to_key = list(values.keys()) self._values = list(values.values()) else: raise TypeError(f"Unsupported values type {type(values)}") def _reset(self): self._unzipped = None self._keys = None def __getitem__(self, key: KeyType) -> ValueType: if self._key_to_index is not None: # pyre-fixme[16]: `Objects` has no attribute `_values`. return self._values[self._key_to_index[key]] else: return self._values[key] def __setitem__(self, key: KeyType, value: ValueType): if self._key_to_index is not None: # pyre-fixme[16]: `Objects` has no attribute `_values`. self._values[self._key_to_index[key]] = value else: self._values[key] = value self._reset() @abstractmethod def _to_key(self, k: int) -> KeyType: pass def _to_value(self, v) -> ValueType: return v def __iter__(self): if self._key_to_index is not None: return ( (k, self._to_value(self._values[i])) for k, i in self._key_to_index.items() ) else: return ( (self._to_key(i), self._to_value(v)) for i, v in enumerate(self._values) ) def __len__(self) -> int: # pyre-fixme[16]: `Objects` has no attribute `_values`. return len(self._values) @property def is_sequence(self): return self._key_to_index is None @property def _values_copy(self) -> Sequence[ValueType]: # pyre-fixme[16]: `Objects` has no attribute `_values`. return list(self._values) def index_of(self, key: KeyType) -> int: if self._key_to_index is None: try: # pyre-fixme[6]: Expected `Union[_SupportsIndex, bytes, str, # typing.SupportsInt]` for 1st param but got `KeyType`. index = int(key) if 0 <= index < len(self): return index else: raise ValueError(f"{key} is not valid") except Exception: raise ValueError(f"{key} is not valid") elif self._key_to_index is not None: try: return self._key_to_index[key] except Exception: raise ValueError(f"{key} is not valid") else: raise ValueError(f"{key} is not valid") @property def keys(self) -> Sequence[KeyType]: if self._keys is None: if self._key_to_index is not None: self._keys = list(self._key_to_index.keys()) else: self._keys = [self._to_key(i) for i in range(len(self))] return self._keys @property def values(self): return self._values_copy def __repr__(self): return f"{self.__class__.__name__}{{values[{self._values}]}}" class Values(Objects[KeyType, float]): """ Generic class for a map from item to its value. It supports [] indexing, and iterator protocol Attributes: items: list of items values: list of their values """ def __init__( self, values: Union[Mapping[KeyType, float], Sequence[float], np.ndarray, Tensor], ): # pyre-fixme[6]: Expected `Union[Mapping[Variable[KeyType], # Variable[ValueType]], Sequence[Variable[ValueType]]]` for 1st param but got # `Union[Mapping[Variable[KeyType], float], Sequence[float], Tensor, # np.ndarray]`. super().__init__(values) def _init_values( self, values: Union[Mapping[KeyType, float], Sequence[float], np.ndarray, Tensor], ): if isinstance(values, Tensor): # pyre-fixme[16]: `Values` has no attribute `_values`. self._values = values.to(dtype=torch.double) elif isinstance(values, np.ndarray): self._values = torch.as_tensor(values, dtype=torch.double) elif isinstance(values, Sequence): self._values = torch.tensor(values, dtype=torch.double) elif isinstance(values, Mapping): self._key_to_index = dict(zip(values.keys(), range(len(values)))) self._index_to_key = list(values.keys()) self._values = torch.tensor(list(values.values()), dtype=torch.double) else: raise TypeError(f"Unsupported values type {type(values)}") def _reset(self): super()._reset() self._probabilities = None self._is_normalized = False self._sorted = None def __getitem__(self, key: KeyType) -> float: return super().__getitem__(key).item() def _to_value(self, v: Tensor) -> float: return v.item() def __len__(self) -> int: # pyre-fixme[16]: `Values` has no attribute `_values`. return self._values.shape[0] def sort(self, descending: bool = True) -> Tuple[Sequence[KeyType], Tensor]: """ Sort based on values Args: descending: sorting order Returns: Tuple of sorted indices and values """ # pyre-fixme[16]: `Values` has no attribute `_sorted`. if self._sorted is None: # pyre-fixme[16]: `Values` has no attribute `_values`. rs, ids = torch.sort(self._values, descending=descending) if self._index_to_key is not None: self._sorted = ( [self._index_to_key[i.item()] for i in ids], rs.detach(), ) else: self._sorted = ([self._to_key(i.item()) for i in ids], rs.detach()) return self._sorted @property def _values_copy(self) -> Tensor: # pyre-fixme[16]: `Values` has no attribute `_values`. return self._values.clone().detach() def replace( self, values: Union[Mapping[ValueType, float], Sequence[float], Tensor, np.ndarray], ) -> "Values": """ Replace current values with new values, and returns the new copy. Current Values object is not changed Args: values: new value Returns: Values object with new values """ copy = deepcopy(self) if isinstance(values, Tensor): # pyre-fixme[16]: `Values` has no attribute `_values`. assert values.shape[0] == copy._values.shape[0] copy._values = values.to(dtype=torch.double) elif isinstance(values, np.ndarray): assert values.shape[0] == copy._values.shape[0] copy._values = torch.as_tensor(values, dtype=torch.double) elif isinstance(values, Sequence): assert len(values) == copy._values.shape[0] copy._values = torch.tensor(values, dtype=torch.double) elif isinstance(values, Mapping): if copy._key_to_index is None: for k, v in values.items(): copy._values[k] = v else: for k, v in values.items(): copy._values[copy._key_to_index[k]] = v else: raise TypeError(f"Unsupported values type {type(values)}") return copy def _normalize(self): if self._is_normalized: if self._probabilities is None: raise ValueError(f"Invalid distribution {type(self._values)}") return self._is_normalized = True self._probabilities = None try: dist = self._values.detach().clamp(min=0.0) dist /= dist.sum() self._probabilities = dist except ZeroDivisionError: pass def probability(self, key: ValueType) -> float: self._normalize() # pyre-fixme[16]: `Values` has no attribute `_probabilities`. if self._probabilities is not None: if self._key_to_index is not None: return self._probabilities[self._key_to_index[key]].item() else: return self._probabilities[key].item() else: return 0.0 def sample(self, size=1) -> Sequence[KeyType]: self._normalize() if self._index_to_key is not None: l = [ self._index_to_key[k.item()] # pyre-fixme[16]: `Values` has no attribute `_probabilities`. for k in torch.multinomial(self._probabilities, size) ] else: l = [ self._to_key(k.item()) for k in torch.multinomial(self._probabilities, size) ] return l def greedy(self, size=1) -> Sequence[KeyType]: sorted_keys, _ = self.sort() return sorted_keys[:size] class Items(Generic[ValueType], ABC): """ List of items """ def __init__(self, items: Union[Sequence[ValueType], int]): if isinstance(items, int): assert items > 0 self._items = [self._new_item(i) for i in range(items)] self._reverse_lookup = None else: self._items = items self._reverse_lookup = {v: i for i, v in enumerate(items)} def __getitem__(self, i) -> ValueType: return self._items[i] def __len__(self): return len(self._items) def __iter__(self): return iter(self._items) def __int__(self): if self._reverse_lookup is None: return len(self._items) else: return 0 @abstractmethod def _new_item(self, i: int) -> ValueType: pass @property def is_sequence(self): return self._reverse_lookup is None def index_of(self, item: ValueType) -> int: if self._reverse_lookup is None: # pyre-fixme[16]: `ValueType` has no attribute `value`. int_val = int(item.value) if 0 <= int_val < len(self._items): return int_val else: raise ValueError(f"{item} is not valid") elif self._reverse_lookup is not None: try: return self._reverse_lookup[item] except Exception: raise ValueError(f"{item} is not valid") else: raise ValueError(f"{item} is not valid") def fill( self, values: Union[Mapping[ValueType, float], Sequence[float], np.ndarray, Tensor], ) -> Union[Sequence[float], Mapping[ValueType, float]]: if self._reverse_lookup is None: if isinstance(values, Mapping): ds = [] for a in self._items: if a in values: ds.append(values[a]) else: ds.append(0.0) return ds else: ds = [0.0] * len(self._items) l = min(len(self._items), len(values)) ds[:l] = values[:l] return ds elif isinstance(values, Mapping): ds = {} for a in self._items: if a in values: ds[a] = values[a] else: ds[a] = 0.0 return ds else: ds = {} for a in self._items: try: ds[a] = values[self._reverse_lookup[a]] except Exception: ds[a] = 0.0 return ds # action type ActionType = Union[int, Tuple[int], float, Tuple[float], np.ndarray, Tensor] Action = TypeWrapper[ActionType] Reward = float Probability = float # Action distribution: Action -> probability # if action can be indexed, the type is either sequence of float or 1-D tensor, # with the indices being the action class ActionDistribution(Values[Action]): def _to_key(self, k: int) -> Action: return Action(k) class ActionSpace(Items[Action]): def _new_item(self, i: int) -> Action: return Action(i) @property def space(self) -> Sequence[Action]: return self._items def distribution( self, dist: Union[Mapping[Action, float], Sequence[float], np.ndarray, Tensor] ) -> ActionDistribution: return ActionDistribution(super().fill(dist)) class Policy(ABC): """ Policy interface """ def __init__(self, action_space: ActionSpace, device=None): self._action_space = action_space self._device = device @abstractmethod def _query(self, context) -> Tuple[Action, ActionDistribution]: pass def __call__(self, context) -> Tuple[Action, ActionDistribution]: return self._query(context) @property def action_space(self): return self._action_space @dataclass(frozen=True) class TrainingData: train_x: Tensor train_y: Tensor train_weight: Optional[Tensor] validation_x: Tensor validation_y: Tensor validation_weight: Optional[Tensor] @dataclass(frozen=True) class PredictResults: predictions: Optional[Tensor] # shape = [num_samples] scores: Tensor # shape = [num_samples] probabilities: Optional[Tensor] = None class Trainer(ABC): def __init__(self): self._model = None @staticmethod def _sample( x: Tensor, y: Tensor, weight: Optional[Tensor] = None, num_samples: int = 0, fortran_order: bool = False, ) -> Tuple[np.ndarray, np.ndarray, np.ndarray]: assert x.shape[0] == y.shape[0] x_na = x.numpy() if fortran_order: x_na = x_na.reshape(x.shape, order="F") y_na = y.numpy() w_na = weight.numpy() if weight is not None else None if num_samples > 0 and num_samples < x.shape[0]: cs = np.random.choice(x.shape[0], num_samples, replace=False) x_na = x_na[cs, :] y_na = y_na[cs] w_na = w_na[cs] if w_na is not None else None return x_na, y_na, w_na def reset(self): self._model = None @property @abstractmethod def name(self) -> str: pass @property def is_trained(self) -> bool: return self._model is not None @abstractmethod def train(self, data: TrainingData, iterations: int = 1, num_samples: int = 0): pass @abstractmethod def predict(self, x: Tensor, device=None) -> PredictResults: pass @abstractmethod def score(self, x: Tensor, y: Tensor, weight: Optional[Tensor] = None) -> float: pass def save_model(self, file: str): if self._model is None: logging.error(f"{self.__class__.__name__}.save_model: _model is None ") return try: with open(file, "wb") as f: pickle.dump(self._model, f, protocol=pickle.HIGHEST_PROTOCOL) except Exception: logging.error(f"{file} cannot be accessed.") def load_model(self, file: str): try: logging.info(f"{self.__class__.__name__}.load_model: {file}") with open(file, "rb") as f: self._model = pickle.load(f) except Exception: logging.error(f"{file} cannot be read.")
18,646
30.712585
88
py
ReAgent
ReAgent-master/reagent/ope/estimators/contextual_bandits_estimators.py
#!/usr/bin/env python3 import logging import time from abc import ABC, abstractmethod from dataclasses import dataclass from typing import Optional, Sequence, Tuple, Union import numpy as np import torch from reagent.ope.estimators.estimator import Estimator, EstimatorResult from reagent.ope.estimators.types import ( Action, ActionDistribution, ActionSpace, Reward, Trainer, TrainingData, Values, ) from reagent.ope.utils import Clamper, RunningAverage from torch import Tensor logger = logging.getLogger(__name__) Actions = Union[Sequence[Action], Tensor, np.ndarray] PROPENSITY_THRESHOLD = 1e-6 class ActionRewards(Values[Action]): def _to_key(self, k: int) -> Action: return Action(k) class BanditsModel(ABC): @abstractmethod def _action_rewards(self, context) -> ActionRewards: """ Calculate reward for each action based on context Args: context: task specific context Returns: action -> reward map """ pass def __call__(self, context) -> ActionRewards: return self._action_rewards(context) @dataclass(frozen=True) class ModelOutputs: tgt_reward_from_log_action: Reward tgt_rewards: Sequence[Reward] @dataclass(frozen=True) class LogSample: # task specific context context: object # log log_action: Action log_reward: Reward log_action_probabilities: ActionDistribution # result from target policy tgt_action_probabilities: ActionDistribution tgt_action: Action model_outputs: Optional[ModelOutputs] = None ground_truth_reward: Reward = float("nan") item_feature: Optional[Tensor] = None @dataclass(frozen=True) class BanditsEstimatorInput: action_space: ActionSpace samples: Sequence[LogSample] has_model_outputs: bool class DMEstimator(Estimator): TRAINING_VALIDATION_SPLIT = 0.8 """ Estimating using Direct Method (DM), assuming a reward model is trained """ def __init__(self, trainer: Optional[Trainer] = None, device=None): super().__init__(device) self._trainer = trainer def _train_model( self, samples: Sequence[LogSample], force_train: bool = False ) -> bool: if self._trainer is None: logger.error("Target model trainer not set") return False trainer = self._trainer assert trainer is not None if trainer.is_trained and not force_train: return True logger.info(" training direct model...") st = time.perf_counter() sample_size = len(samples) training_size = int(sample_size * DMEstimator.TRAINING_VALIDATION_SPLIT) train_x = [] train_y = [] for i in range(training_size): sample = samples[i] if sample.item_feature is None: continue train_x.append( torch.cat( ( torch.tensor( sample.log_action.value, dtype=torch.float ).flatten(), sample.item_feature.flatten(), ) ) ) train_y.append(sample.log_reward) if len(train_x) == 0: logger.error("Item features not provided, DM is not available") return False train_x = torch.stack(train_x) train_y = torch.tensor(train_y, dtype=torch.float, device=train_x.device) vali_x = [] vali_y = [] for i in range(training_size, sample_size): sample = samples[i] if sample.item_feature is None: continue vali_x.append( torch.cat( ( torch.tensor( sample.log_action.value, dtype=torch.float ).flatten(), sample.item_feature.flatten(), ) ) ) vali_y.append(sample.log_reward) if len(vali_x) == 0: vali_x = train_x.detach().clone() vali_y = train_y.detach().clone() else: vali_x = torch.stack(vali_x) vali_y = torch.tensor(vali_y, dtype=torch.float, device=vali_x.device) training_data = TrainingData(train_x, train_y, None, vali_x, vali_y, None) trainer.train(training_data) logger.info(f" training direct model done: {time.perf_counter() - st}s") return True def _calc_dm_reward( self, action_space: ActionSpace, sample: LogSample ) -> Tuple[Optional[Reward], torch.Tensor, torch.Tensor]: if sample.model_outputs is not None: return ( sample.model_outputs.tgt_reward_from_log_action, torch.tensor( sample.model_outputs.tgt_rewards, dtype=torch.float, device=self._device, ), torch.tensor( # pyre-fixme[16]: `ActionDistribution` has no attribute `_values`. sample.tgt_action_probabilities._values, dtype=torch.float, device=self._device, ), ) trainer = self._trainer if trainer is None or not trainer.is_trained: return 0.0, torch.zeros(), torch.zeros() assert sample.item_feature is not None item_feature = sample.item_feature.flatten() features = [] probs = [] idx = -1 for action in action_space: if idx < 0 and action == sample.log_action: idx = len(features) features.append( torch.cat( ( torch.tensor(action.value, dtype=torch.float).flatten(), item_feature, ) ) ) probs.append(sample.tgt_action_probabilities[action]) preds = trainer.predict(torch.stack(features).float(), device=self._device) return ( preds.scores[idx].item(), preds.scores, torch.tensor(probs, dtype=torch.float, device=self._device), ) def _evaluate( self, input: BanditsEstimatorInput, train_samples: Sequence[LogSample], eval_samples: Sequence[LogSample], force_train: bool = False, **kwargs, ) -> Optional[EstimatorResult]: logger.info("OPE DM Evaluating") if ( not self._train_model(train_samples, force_train) and not input.has_model_outputs ): return None log_avg = RunningAverage() tgt_avg = RunningAverage() tgt_vals = [] gt_avg = RunningAverage() for sample in eval_samples: log_avg.add(sample.log_reward) _, tgt_scores, tgt_probs = self._calc_dm_reward(input.action_space, sample) tgt_reward = torch.dot(tgt_scores.reshape(-1), tgt_probs.reshape(-1)).item() tgt_avg.add(tgt_reward) tgt_vals.append(tgt_reward) gt_avg.add(sample.ground_truth_reward) ( tgt_score_normalized, tgt_std_err, tgt_std_err_normalized, ) = self._compute_metric_data(torch.tensor(tgt_vals), log_avg.average) return EstimatorResult( log_reward=log_avg.average, estimated_reward=tgt_avg.average, ground_truth_reward=gt_avg.average, estimated_weight=tgt_avg.count, estimated_reward_normalized=tgt_score_normalized, estimated_reward_std_error=tgt_std_err, estimated_reward_normalized_std_error=tgt_std_err_normalized, ) @staticmethod def _calc_optional_avg(a: Optional[float], b: Optional[float]) -> Optional[float]: # Annoying but Pyre would only take it like this return None if a is None else (None if b is None else (a + b) / 2) def evaluate( self, input: BanditsEstimatorInput, **kwargs ) -> Optional[EstimatorResult]: if input.has_model_outputs: return self._evaluate( input, input.samples, input.samples, force_train=True, **kwargs ) log_avg = RunningAverage() gt_avg = RunningAverage() for sample in input.samples: log_avg.add(sample.log_reward) gt_avg.add(sample.ground_truth_reward) # 2-fold cross "validation" as used by https://arxiv.org/pdf/1612.01205.pdf shuffled = list(input.samples) np.random.shuffle(shuffled) lower_half = shuffled[: len(shuffled) // 2] upper_half = shuffled[len(shuffled) // 2 :] er_lower = self._evaluate( input, lower_half, upper_half, force_train=True, **kwargs ) er_upper = self._evaluate( input, upper_half, lower_half, force_train=True, **kwargs ) if er_lower is None or er_upper is None: return None return EstimatorResult( log_reward=log_avg.average, estimated_reward=( (er_lower.estimated_reward + er_upper.estimated_reward) / 2 ), estimated_reward_normalized=( DMEstimator._calc_optional_avg( er_lower.estimated_reward_normalized, er_upper.estimated_reward_normalized, ) ), estimated_reward_normalized_std_error=( DMEstimator._calc_optional_avg( er_lower.estimated_reward_normalized_std_error, er_upper.estimated_reward_normalized_std_error, ) ), estimated_reward_std_error=( DMEstimator._calc_optional_avg( er_lower.estimated_reward_std_error, er_upper.estimated_reward_std_error, ) ), ground_truth_reward=gt_avg.average, ) def __repr__(self): return f"DMEstimator(trainer({None if self._trainer is None else self._trainer.name},device({self._device}))" class IPSEstimator(Estimator): """ Inverse Propensity Scoring (IPS) estimator """ def __init__( self, weight_clamper: Optional[Clamper] = None, weighted: bool = False, device=None, ): super().__init__(device) self._weight_clamper = Clamper() if weight_clamper is None else weight_clamper self._weighted = weighted def evaluate( self, input: BanditsEstimatorInput, **kwargs ) -> Optional[EstimatorResult]: logger.info("OPE IPS Evaluating") log_avg = RunningAverage() logged_vals = [] tgt_avg = RunningAverage() tgt_vals = [] acc_weight = RunningAverage() gt_avg = RunningAverage() for sample in input.samples: log_avg.add(sample.log_reward) logged_vals.append(sample.log_reward) weight = 0.0 tgt_result = 0.0 if sample.log_action.value is not None: weight = ( 0.0 if sample.log_action_probabilities[sample.log_action] < PROPENSITY_THRESHOLD else sample.tgt_action_probabilities[sample.log_action] / sample.log_action_probabilities[sample.log_action] ) weight = self._weight_clamper(weight) tgt_result = sample.log_reward * weight tgt_avg.add(tgt_result) tgt_vals.append(tgt_result) acc_weight.add(weight) gt_avg.add(sample.ground_truth_reward) ( tgt_score_normalized, tgt_std_err, tgt_std_err_normalized, ) = self._compute_metric_data(torch.tensor(tgt_vals), log_avg.average) return EstimatorResult( log_reward=log_avg.average, estimated_reward=tgt_avg.average if not self._weighted else tgt_avg.average / acc_weight.total, ground_truth_reward=gt_avg.average, estimated_weight=tgt_avg.count, estimated_reward_normalized=tgt_score_normalized, estimated_reward_std_error=tgt_std_err, estimated_reward_normalized_std_error=tgt_std_err_normalized, ) def __repr__(self): return ( f"IPSEstimator(weight_clamper({self._weight_clamper})" f",weighted({self._weighted}),device({self._device}))" ) class DoublyRobustEstimator(DMEstimator): """ Doubly Robust (DR) estimator: reference: https://arxiv.org/abs/1103.4601 (deterministic reward model) https://arxiv.org/abs/1612.01205 (distributed reward model) """ def __init__( self, trainer: Optional[Trainer] = None, weight_clamper: Optional[Clamper] = None, device=None, ): super().__init__(trainer, device) self._weight_clamper = Clamper() if weight_clamper is None else weight_clamper def _evaluate( self, input: BanditsEstimatorInput, train_samples: Sequence[LogSample], eval_samples: Sequence[LogSample], force_train: bool = False, **kwargs, ) -> Optional[EstimatorResult]: logger.info("OPE DR Evaluating") self._train_model(train_samples, force_train) log_avg = RunningAverage() tgt_avg = RunningAverage() tgt_vals = [] gt_avg = RunningAverage() for sample in eval_samples: log_avg.add(sample.log_reward) dm_action_reward, dm_scores, dm_probs = self._calc_dm_reward( input.action_space, sample ) dm_reward = torch.dot(dm_scores.reshape(-1), dm_probs.reshape(-1)).item() tgt_result = 0.0 weight = 0.0 if sample.log_action.value is not None: weight = ( 0.0 if sample.log_action_probabilities[sample.log_action] < PROPENSITY_THRESHOLD else sample.tgt_action_probabilities[sample.log_action] / sample.log_action_probabilities[sample.log_action] ) weight = self._weight_clamper(weight) assert dm_action_reward is not None assert dm_reward is not None tgt_result += ( sample.log_reward - dm_action_reward ) * weight + dm_reward else: tgt_result = dm_reward tgt_avg.add(tgt_result) tgt_vals.append(tgt_result) gt_avg.add(sample.ground_truth_reward) ( tgt_score_normalized, tgt_std_err, tgt_std_err_normalized, ) = self._compute_metric_data(torch.tensor(tgt_vals), log_avg.average) return EstimatorResult( log_reward=log_avg.average, estimated_reward=tgt_avg.average, ground_truth_reward=gt_avg.average, estimated_weight=tgt_avg.count, estimated_reward_normalized=tgt_score_normalized, estimated_reward_std_error=tgt_std_err, estimated_reward_normalized_std_error=tgt_std_err_normalized, ) def __repr__(self): return ( f"DoublyRobustEstimator(trainer({None if self._trainer is None else self._trainer.name})" f",weight_clamper({self._weight_clamper}),device({self._device}))" ) class SwitchEstimator(DMEstimator): # For details, visit https://arxiv.org/abs/1612.01205 sections 4, 5 CANDIDATES = 21 EXP_BASE = 1.5 EPSILON = 1e-6 def __init__( self, trainer: Optional[Trainer] = None, weight_clamper: Optional[Clamper] = None, rmax: Optional[Reward] = None, device=None, ): """ rmax is an a priori upper bound on any possible reward. The tighter the bound, the better the estimator can estimate its bias. If not provided, the estimator will use the max reward seen in the sample data. """ super().__init__(trainer, device) self._rmax = rmax self._weight_clamper = Clamper() if weight_clamper is None else weight_clamper def _estimate_rmax(self, input: BanditsEstimatorInput) -> Reward: rmax = float("-inf") for sample in input.samples: _, dm_scores, dm_probs = self._calc_dm_reward(input.action_space, sample) max_sample_r = max(sample.log_reward, torch.max(dm_scores).item()) rmax = max(rmax, max_sample_r) return rmax def _calc_weight_reward_tensors( self, input: BanditsEstimatorInput, eval_samples: Sequence[LogSample] ) -> Tuple[ torch.Tensor, torch.Tensor, torch.Tensor, torch.Tensor, torch.Tensor, torch.Tensor, torch.Tensor, RunningAverage, RunningAverage, ]: n = len(eval_samples) ws = torch.ones((n, len(input.action_space))) rs = torch.zeros((n, 1)) r_est = torch.zeros((n, len(input.action_space))) r_est_for_logged_action = torch.zeros((n, 1)) actions = torch.zeros((n, len(input.action_space))) expected_rmax = torch.zeros((n, len(input.action_space))) propensities = torch.zeros((n, len(input.action_space))) log_avg = RunningAverage() gt_avg = RunningAverage() priori_rmax = self._estimate_rmax(input) if self._rmax is None else self._rmax assert priori_rmax is not None for i, sample in enumerate(eval_samples): dm_score_for_logged_action, dm_scores, dm_probs = self._calc_dm_reward( input.action_space, sample ) for a in input.action_space: weight = ( 0.0 if sample.log_action_probabilities[a] < PROPENSITY_THRESHOLD else sample.tgt_action_probabilities[a] / sample.log_action_probabilities[a] ) ws[i, a] = self._weight_clamper(weight) propensities[i, a] = sample.tgt_action_probabilities[a] expected_rmax[i, a] = sample.tgt_action_probabilities[a] * priori_rmax actions[i, a] = float(a == sample.log_action) rs[i, 0] = sample.log_reward r_est[i] = dm_scores.reshape(-1) r_est_for_logged_action[i] = dm_score_for_logged_action log_avg.add(sample.log_reward) gt_avg.add(sample.ground_truth_reward) return ( actions, ws, rs, r_est, r_est_for_logged_action, propensities, expected_rmax, log_avg, gt_avg, ) def _calc_estimated_values( self, logged_rewards: torch.Tensor, weights: torch.Tensor, actions: torch.Tensor, threshold: float, est_rewards: torch.Tensor, est_rewards_for_logged_action: torch.Tensor, tgt_props: torch.Tensor, ) -> torch.Tensor: ips_scores = (weights * actions).sum(dim=1, keepdim=True) return logged_rewards * ips_scores * (ips_scores <= threshold).float() + ( est_rewards * tgt_props * (weights > threshold).float() ).sum(dim=1, keepdim=True) def _evaluate( self, input: BanditsEstimatorInput, train_samples: Sequence[LogSample], eval_samples: Sequence[LogSample], force_train: bool = False, **kwargs, ) -> Optional[EstimatorResult]: logger.info("OPE Switch Evaluating") self._train_model(train_samples, force_train) if "exp_base" in kwargs: exp_base = kwargs["exp_base"] else: exp_base = SwitchEstimator.EXP_BASE if "candidates" in kwargs: num_candidates = kwargs["candidates"] else: num_candidates = SwitchEstimator.CANDIDATES ( actions, ws, rs, r_est, r_est_for_logged_action, propensities, expected_rmax, log_avg, gt_avg, ) = self._calc_weight_reward_tensors(input, eval_samples) min_w, max_w = float(torch.min(ws).item()), float(torch.max(ws).item()) diff = max_w - min_w # The threshold lies in the range [min ips, max ips] # Picking a small threshold -> using mainly the model-based estimator # Picking a large threshold -> using mainly the ips-based estimator candidates = [ min_w + ((exp_base ** x) / (exp_base ** (num_candidates - 1))) * diff for x in range(num_candidates) ] # This prevents the edge case where nearly all scores being min_w prevents # switch from trying a purely DM estimate tau = min_w - SwitchEstimator.EPSILON loss = float("inf") for candidate in candidates: estimated_values = self._calc_estimated_values( rs, ws, actions, candidate, r_est, r_est_for_logged_action, propensities ) var = (1.0 / (estimated_values.shape[0] ** 2)) * torch.sum( (estimated_values - torch.mean(estimated_values)) ** 2 ).item() bias = torch.mean( torch.sum(expected_rmax * (ws > candidate).float(), dim=1, keepdim=True) ).item() cand_loss = var + bias * bias if cand_loss < loss: tau = candidate loss = cand_loss estimated_values = self._calc_estimated_values( rs, ws, actions, tau, r_est, r_est_for_logged_action, propensities ) ( tgt_score_normalized, tgt_std_err, tgt_std_err_normalized, ) = self._compute_metric_data(estimated_values.detach(), log_avg.average) return EstimatorResult( log_reward=log_avg.average, estimated_reward=torch.mean(estimated_values).item(), ground_truth_reward=gt_avg.average, estimated_weight=float(estimated_values.shape[0]), estimated_reward_normalized=tgt_score_normalized, estimated_reward_std_error=tgt_std_err, estimated_reward_normalized_std_error=tgt_std_err_normalized, ) def __repr__(self): return ( f"SwitchEstimator(trainer({None if self._trainer is None else self._trainer.name})" f",weight_clamper({self._weight_clamper}),device({self._device}))" ) class SwitchDREstimator(SwitchEstimator): # For details, visit https://arxiv.org/abs/1612.01205 sections 4, 5 def _calc_estimated_values( self, logged_rewards: torch.Tensor, weights: torch.Tensor, actions: torch.Tensor, threshold: float, est_rewards: torch.Tensor, est_rewards_for_logged_action: torch.Tensor, tgt_props: torch.Tensor, ) -> torch.Tensor: ips_scores = (weights * actions).sum(dim=1, keepdim=True) dr = ips_scores * (logged_rewards - est_rewards_for_logged_action) + ( tgt_props * est_rewards ).sum(dim=1, keepdim=True) return dr * (ips_scores <= threshold).float() + ( est_rewards * tgt_props * (weights > threshold).float() ).sum(dim=1, keepdim=True) def __repr__(self): return ( f"SwitchDREstimator(trainer({None if self._trainer is None else self._trainer.name})" f",weight_clamper({self._weight_clamper}),device({self._device}))" )
23,956
34.971471
117
py
ReAgent
ReAgent-master/reagent/ope/estimators/estimator.py
#!/usr/bin/env python3 import logging import math import pickle import tempfile from abc import ABC, abstractmethod from dataclasses import dataclass, field from multiprocessing import Pool from typing import Iterable, List, Mapping, Optional, Tuple, Union import torch from reagent.evaluation.cpe import bootstrapped_std_error_of_mean from torch import Tensor logger = logging.getLogger(__name__) SCORE_THRESHOLD = 1e-6 class ResultDiffs: """ Statistics for differences, e.g., estimates vs ground truth """ def __init__(self, diffs: Tensor): self._diffs = diffs self._rmse = None self._bias = None self._variance = None @property def rmse(self) -> Tensor: if self._rmse is None: self._rmse = (self._diffs ** 2.0).mean().sqrt() return self._rmse @property def bias(self) -> Tensor: if self._bias is None: self._bias = self._diffs.mean() return self._bias @property def variance(self) -> Tensor: if self._variance is None: # pyre-fixme[16]: `Tensor` has no attribute `var`. self._variance = self._diffs.var() return self._variance def __repr__(self): return ( f"samples={self._diffs.shape[0]}, rmse={self.rmse.item()}" f", bias={self.bias}, variance={self.variance}" ) @dataclass(frozen=True) class EstimatorResult: log_reward: float estimated_reward: float ground_truth_reward: Optional[float] = 0.0 estimated_weight: float = 1.0 estimated_reward_normalized: Optional[float] = None estimated_reward_std_error: Optional[float] = None estimated_reward_normalized_std_error: Optional[float] = None @dataclass class EstimatorResults: """ Estimator results """ results: List[EstimatorResult] = field(default_factory=list) device = None def append(self, result: EstimatorResult): """Append a data point Args: result: result from an experimental run """ er = result.estimated_reward if math.isnan(er) or math.isinf(er): logging.warning(f" Invalid estimate: {er}") return lr = result.log_reward gr = ( result.ground_truth_reward if result.ground_truth_reward is not None else 0.0 ) logging.info( f" Append estimate [{len(self.results) + 1}]: " f"log={lr}, estimated={er}, ground_truth={gr}" ) self.results.append( EstimatorResult( log_reward=result.log_reward, estimated_reward=result.estimated_reward, ground_truth_reward=gr, estimated_weight=result.estimated_weight, ) ) def report(self): ert = torch.tensor( [res.estimated_reward for res in self.results], dtype=torch.double, device=self.device, ) lrt = torch.tensor( [res.log_reward for res in self.results], dtype=torch.double, device=self.device, ) grt = torch.tensor( [ res.ground_truth_reward if res.ground_truth_reward is not None else 0.0 for res in self.results ], dtype=torch.double, device=self.device, ) self._estimated_log_diff = ResultDiffs(ert - lrt) self._estimated_ground_truth_diff = ResultDiffs(ert - grt) return ( lrt.mean().item(), ert.mean().item(), grt.mean().item(), ResultDiffs(ert - grt), ResultDiffs(ert - lrt), torch.tensor([float(res.estimated_weight) for res in self.results]) .mean() .item(), ) @dataclass(frozen=True) class EstimatorSampleResult: log_reward: float target_reward: float ground_truth_reward: float weight: float def __repr__(self): return ( f"EstimatorSampleResult(log={self.log_reward}" f",tgt={self.target_reward},gt={self.ground_truth_reward}" f",wgt={self.weight}" ) class Estimator(ABC): """ Estimator interface """ def __init__(self, device=None): self._device = device def _compute_metric_data( self, tgt_rewards: Tensor, logged_score: float ) -> Tuple[float, float, float]: """ Given a sequence of scores, normalizes the target score by the average logged score and computes the standard error of the target score. Normalizing by the logged score can provide a better metric to compare models against. """ if len(tgt_rewards.shape) > 1: assert tgt_rewards.shape[1] == 1 tgt_rewards = tgt_rewards.reshape((tgt_rewards.shape[0],)) if logged_score < SCORE_THRESHOLD: normalizer = 0.0 else: normalizer = 1.0 / logged_score std_err = bootstrapped_std_error_of_mean(tgt_rewards) return ( torch.mean(tgt_rewards).item() * normalizer, std_err, std_err * normalizer, ) @abstractmethod def evaluate( self, input, **kwargs ) -> Optional[Union[EstimatorResult, EstimatorResults]]: pass def __repr__(self): return f"{self.__class__.__name__}(device({self._device}))" def run_evaluation( file_name: str, ) -> Optional[Mapping[str, Iterable[EstimatorResults]]]: logger.info(f"received filename {file_name}") try: with open(file_name, "rb") as fp: estimators, inputs = pickle.load(fp) except Exception as err: return None results = {} for estimator in estimators: estimator_name = repr(estimator) estimator_results = [] for input in inputs: try: estimator_results.append(estimator.evaluate(input)) except Exception as err: logger.error(f"{estimator_name} error {err}") results[repr(estimator)] = estimator_results return results class Evaluator: """ Multiprocessing evaluator """ def __init__( self, experiments: Iterable[Tuple[Iterable[Estimator], object]], max_num_workers: int, ): """ Args: estimators: estimators to be evaluated experiments: max_num_workers: <= 0 no multiprocessing otherwise create max_num_workers processes """ self._experiments = experiments self._tasks = None if max_num_workers > 0: self._tasks = [[] for _ in range(max_num_workers)] for i, experiment in enumerate(experiments): self._tasks[i % max_num_workers].append(experiment) def evaluate(self) -> Mapping[str, EstimatorResults]: results = {} if self._tasks is None: for estimators, input in self._experiments: for estimator in estimators: estimator_name = repr(estimator) if estimator_name in results: result = results[estimator_name] else: result = EstimatorResults() results[estimator_name] = result result.append(estimator.evaluate(input)) else: tmp_files = [] tmp_file_names = [] for task in self._tasks: fp = tempfile.NamedTemporaryFile() pickle.dump(task, fp, protocol=pickle.HIGHEST_PROTOCOL) fp.flush() tmp_files.append(fp) tmp_file_names.append(fp.name) with Pool(len(tmp_file_names)) as pool: evaluation_results = pool.map(run_evaluation, tmp_file_names) for tmp_file in tmp_files: tmp_file.close() for evaluation_result in evaluation_results: if evaluation_result is None: continue for estimator_name, estimator_results in evaluation_result.items(): if estimator_name in results: result = results[estimator_name] else: result = EstimatorResults() results[estimator_name] = result for estimator_result in estimator_results: result.append(estimator_result) return results @staticmethod def report_results(results: Mapping[str, EstimatorResults]): for name, result in results.items(): log_r, tgt_r, gt_r, tgt_gt, tgt_log, weight = result.report() print( f"{name} rewards: log_reward{log_r} tgt_reward[{tgt_r}] gt_reward[{gt_r}]" f", diffs: tgt-gt[{tgt_gt}] tgt-log[{tgt_log}]", flush=True, )
9,051
30.430556
92
py
ReAgent
ReAgent-master/reagent/ope/test/mslr_slate.py
#!/usr/bin/env python3 import argparse import itertools import json import logging import os import pickle import random import sys import time from collections import OrderedDict from typing import Iterable, List, Optional, Tuple import numpy as np import torch import torch.multiprocessing as mp from reagent.ope.estimators.estimator import Evaluator from reagent.ope.estimators.slate_estimators import ( DCGSlateMetric, DMEstimator, DoublyRobustEstimator, ERRSlateMetric, IPSEstimator, LogSample, NDCGSlateMetric, PassThruDistribution, PBMEstimator, PseudoInverseEstimator, RankingDistribution, RewardDistribution, SlateContext, SlateEstimator, SlateEstimatorInput, SlateItemFeatures, SlateItemValues, SlateModel, SlateQuery, SlateSlots, ) from reagent.ope.estimators.types import Trainer, TrainingData from reagent.ope.trainers.linear_trainers import DecisionTreeTrainer, LassoTrainer from reagent.ope.utils import Clamper from torch import Tensor # Slate test using Microsoft Learning to Rank Datasets (MSLR-WEB30K): # https://www.microsoft.com/en-us/research/project/mslr/ class MSLRDatasets: def __init__( self, params, num_columns: int, anchor_url_features: List[int], body_features: List[int], dataset_name: str = "", device=None, ): if "folder" not in params: raise Exception('Please define "folder" in "dataset"') self._folder = params["folder"] self._source_file = ( params["source_file"] if "source_file" in params else ["train.txt"] ) self._cache_file = params["cache_file"] if "cache_file" in params else "" self._num_columns = num_columns self._anchor_url_features = anchor_url_features self._body_features = body_features self._device = device self._dict = None self._queries = None self._features = None self._relevances = None self._sample_weights = None self._train_data = None self._validation_data = None self._test_data = None self._name = dataset_name @property def name(self) -> str: return self._name def _add(self, qid: Optional[int], feature_list: List[Tuple[float, Tensor]]): if qid is None or len(feature_list) == 0: return if qid in self._dict: self._dict[qid].extend(feature_list) else: self._dict[qid] = feature_list def load(self): pickle_file = os.path.join(self._folder, self._cache_file) if len(self._cache_file) > 0 and os.access(pickle_file, os.R_OK): logging.info(f"loading {pickle_file}") with open(pickle_file, "rb") as f: self._queries, self._features, self._relevances = pickle.load(f) self._cache_file = "" del f else: self._dict = OrderedDict() text_file = os.path.join(self._folder, self._source_file) logging.info(f"loading {text_file}") if not os.access(text_file, os.R_OK): logging.warning(f"{text_file} cannot be accessed.") return with open(text_file, "r") as f: c = 0 st = time.process_time() # feature index starts with 1, so leave features[0] as padding features = list(range(self._num_columns - 1)) features_list = [] prev_qid = None for line in f.readlines(): tokens = line.strip().split() if len(tokens) != self._num_columns: continue rel = int(tokens[0]) qid = int(tokens[1].split(":")[1]) for i in range(2, self._num_columns): feature = tokens[i].split(":") features[i - 1] = float(feature[1]) f_tensor = torch.tensor(features, device=self._device) if prev_qid is None: prev_qid = qid features_list.append((rel, f_tensor)) elif prev_qid != qid: self._add(prev_qid, features_list) prev_qid = qid features_list = [] features_list.append((rel, f_tensor)) else: features_list.append((rel, f_tensor)) c += 1 if c % 100000 == 0: print(f"{c} - {(time.process_time() - st) / c}") self._add(prev_qid, features_list) def save(self): if len(self._cache_file) == 0 or self._dict is None: return pickle_file = os.path.join(self._folder, self._cache_file) try: with open(pickle_file, "wb") as f: self._load_features() pickle.dump( (self.queries, self._features, self.relevances), f, protocol=pickle.HIGHEST_PROTOCOL, ) except Exception: logging.error(f"{pickle_file} cannot be accessed.") @property def queries(self) -> Tensor: if self._queries is None: rows = [] c = 0 for i in self._dict.items(): s = len(i[1]) rows.append([i[0], c, s]) c += s self._queries = torch.tensor(rows, dtype=torch.int, device=self._device) return self._queries def _load_features(self): if self._features is None: self._features = torch.stack([r[1] for v in self._dict.values() for r in v]) @property def features(self) -> Tensor: self._load_features() return self._features[:, 1:] @property def all_features(self) -> Tensor: return self.features @property def anchor_url_features(self) -> Tensor: self._load_features() return ( self._features[:, self._anchor_url_features] if self._anchor_url_features is not None else None ) @property def body_features(self) -> Tensor: self._load_features() return ( self._features[:, self._body_features] if self._body_features is not None else None ) @property def relevances(self) -> Tensor: if self._relevances is None: self._relevances = torch.tensor( [r[0] for r in itertools.chain(self._dict.values())], device=self._device, ) return self._relevances @property def sample_weights(self) -> Tensor: if self._sample_weights is None: samples = self.queries[:, 2] self._sample_weights = torch.repeat_interleave( samples.to(dtype=torch.float).reciprocal(), samples.to(dtype=torch.long) ) return self._sample_weights @property def folder(self) -> str: return self._folder @property def source_file(self) -> List[str]: return self._source_file @property def cache_file(self) -> str: return self._cache_file def train( trainer: Trainer, train_dataset: MSLRDatasets, vali_dataset: MSLRDatasets, prefix: str = "", ): logging.info("training all features...") st = time.process_time() training_data = TrainingData( train_dataset.all_features, train_dataset.relevances, train_dataset.sample_weights, vali_dataset.all_features, vali_dataset.relevances, vali_dataset.sample_weights, ) trainer.train(training_data) logging.info(f" training time: {time.process_time() - st}") trainer.save_model( os.path.join( train_dataset.folder, trainer.name + "_" + prefix + "_all_features.pickle" ) ) logging.info("scoring...") score = trainer.score( vali_dataset.all_features, vali_dataset.relevances, vali_dataset.sample_weights ) logging.info(f" score: {score}") logging.info("training anchor_url features...") st = time.process_time() trainer.train( TrainingData( train_dataset.anchor_url_features, train_dataset.relevances, train_dataset.sample_weights, vali_dataset.anchor_url_features, vali_dataset.relevances, vali_dataset.sample_weights, ) ) logging.info(f" training time: {time.process_time() - st}") trainer.save_model( os.path.join( train_dataset.folder, trainer.name + "_" + prefix + "_anchor_url_features.pickle", ) ) logging.info("scoring...") score = trainer.score( vali_dataset.anchor_url_features, vali_dataset.relevances, vali_dataset.sample_weights, ) logging.info(f" score: {score}") logging.info("training body features...") st = time.process_time() trainer.train( TrainingData( train_dataset.body_features, train_dataset.relevances, train_dataset.sample_weights, vali_dataset.body_features, vali_dataset.relevances, vali_dataset.sample_weights, ) ) logging.info(f" training time: {time.process_time() - st}") trainer.save_model( os.path.join( train_dataset.folder, trainer.name + "_" + prefix + "_body_features.pickle" ) ) logging.info("scoring...") score = trainer.score( vali_dataset.body_features, vali_dataset.relevances, vali_dataset.sample_weights ) logging.info(f" score: {score}") def load_dataset( params, num_columns, anchor_url_features, body_features, dataset_name="" ) -> MSLRDatasets: logging.info(f"loading {params['source_file']}") dataset = MSLRDatasets( params, num_columns, anchor_url_features, body_features, dataset_name ) st = time.process_time() dataset.load() logging.info(f" load time: {time.process_time() - st}") st = time.process_time() dataset.save() logging.info(f" save time: {time.process_time() - st}") logging.info( f" queries: {dataset.queries.shape}" f", features: {dataset.features.shape}" f", sample_weights: {dataset.sample_weights.shape}" f", relevance: {dataset.relevances.shape}" f", anchor_url: {dataset.anchor_url_features.shape}" f", body: {dataset.body_features.shape}" ) return dataset def train_all(train_dataset, vali_dataset, prefix: str = ""): # train(DecisionTreeClassifierTrainer(), train_dataset, vali_dataset) train(DecisionTreeTrainer(), train_dataset, vali_dataset, prefix) train(LassoTrainer(), train_dataset, vali_dataset, prefix) # train(LogisticRegressionTrainer(), train_dataset, vali_dataset) # train(SGDClassifierTrainer(), train_dataset, vali_dataset) def train_models(params): all_dataset = load_dataset( params["all_set"], num_columns, anchor_url_features, body_features ) half_dataset = load_dataset( params["first_set"], num_columns, anchor_url_features, body_features ) vali_dataset = load_dataset( params["vali_set"], num_columns, anchor_url_features, body_features ) train_all(all_dataset, vali_dataset, "all") train_all(half_dataset, vali_dataset, "half") class MSLRModel(SlateModel): def __init__(self, relevances: Tensor, device=None): self._relevances = relevances self._device = device def item_relevances(self, context: SlateContext) -> Tensor: qv = context.query.value if context.params is None: relevances = self._relevances[qv[1] : (qv[1] + qv[2])].detach().clone() else: relevances = ( self._relevances[qv[1] : (qv[1] + qv[2])][context.params] .detach() .clone() ) return relevances def item_rewards(self, context: SlateContext) -> SlateItemValues: return SlateItemValues(self.item_relevances(context)) def evaluate( experiments: Iterable[Tuple[Iterable[SlateEstimator], int]], dataset: MSLRDatasets, slate_size: int, item_size: int, metric_func: str, log_trainer: Trainer, log_distribution: RewardDistribution, log_features: str, tgt_trainer: Trainer, tgt_distribution: RewardDistribution, tgt_features: str, dm_features: str, max_num_workers: int, device=None, ): assert slate_size < item_size print( f"Evaluate All:" f" slate_size={slate_size}, item_size={item_size}, metric={metric_func}" f", Log=[{log_trainer.name}, {log_distribution}, {log_features}]" f", Target=[{tgt_trainer.name}, {tgt_distribution}, {tgt_features}]" f", DM=[{dm_features}]" f", Workers={max_num_workers}, device={device}", flush=True, ) logging.info("Preparing models and policies...") st = time.perf_counter() log_trainer.load_model( os.path.join( dataset.folder, log_trainer.name + "_all_" + log_features + ".pickle" ) ) # calculate behavior model scores log_pred = log_trainer.predict(getattr(dataset, log_features)) tgt_trainer.load_model( os.path.join( dataset.folder, tgt_trainer.name + "_all_" + tgt_features + ".pickle" ) ) # calculate target model scores tgt_pred = tgt_trainer.predict(getattr(dataset, tgt_features)) dm_train_features = getattr(dataset, dm_features) slots = SlateSlots(slate_size) dt = time.perf_counter() - st logging.info(f"Preparing models and policies done: {dt}s") total_samples = 0 for _, num_samples in experiments: total_samples += num_samples logging.info(f"Generating log: total_samples={total_samples}") st = time.perf_counter() tasks = [] samples_generated = 0 total_queries = dataset.queries.shape[0] for estimators, num_samples in experiments: samples = [] for _ in range(num_samples): # randomly sample a query q = dataset.queries[random.randrange(total_queries)] doc_size = int(q[2]) if doc_size < item_size: # skip if number of docs is less than item_size continue si = int(q[1]) ei = si + doc_size # using top item_size docs for logging log_scores, item_choices = log_pred.scores[si:ei].sort( dim=0, descending=True ) log_scores = log_scores[:item_size] item_choices = item_choices[:item_size] log_item_probs = log_distribution(SlateItemValues(log_scores)) tgt_scores = tgt_pred.scores[si:ei][item_choices].detach().clone() tgt_item_probs = tgt_distribution(SlateItemValues(tgt_scores)) tgt_slot_expectation = tgt_item_probs.slot_item_expectations(slots) gt_item_rewards = SlateItemValues(dataset.relevances[si:ei][item_choices]) gt_rewards = tgt_slot_expectation.expected_rewards(gt_item_rewards) if metric_func == "dcg": metric = DCGSlateMetric(device=device) elif metric_func == "err": metric = ERRSlateMetric(4.0, device=device) else: metric = NDCGSlateMetric(gt_item_rewards, device=device) query = SlateQuery((si, ei)) context = SlateContext(query, slots, item_choices) slot_weights = metric.slot_weights(slots) gt_reward = metric.calculate_reward(slots, gt_rewards, None, slot_weights) if tgt_item_probs.is_deterministic: tgt_slate_prob = 1.0 log_slate = tgt_item_probs.sample_slate(slots) log_reward = gt_reward else: tgt_slate_prob = float("nan") log_slate = log_item_probs.sample_slate(slots) log_rewards = log_slate.slot_values(gt_item_rewards) log_reward = metric.calculate_reward( slots, log_rewards, None, slot_weights ) log_slate_prob = log_item_probs.slate_probability(log_slate) item_features = SlateItemFeatures(dm_train_features[si:ei][item_choices]) sample = LogSample( context, metric, log_slate, log_reward, log_slate_prob, None, log_item_probs, tgt_slate_prob, None, tgt_item_probs, gt_reward, slot_weights, None, item_features, ) samples.append(sample) samples_generated += 1 if samples_generated % 1000 == 0: logging.info( f" samples generated: {samples_generated}, {100 * samples_generated / total_samples:.1f}%" ) tasks.append((estimators, SlateEstimatorInput(samples))) dt = time.perf_counter() - st logging.info(f"Generating log done: {total_samples} samples in {dt}s") logging.info("start evaluating...") st = time.perf_counter() evaluator = Evaluator(tasks, max_num_workers) Evaluator.report_results(evaluator.evaluate()) logging.info(f"evaluating done in {time.perf_counter() - st}s") if __name__ == "__main__": mp.set_start_method("spawn") logging.basicConfig( format="%(asctime)-15s_%(levelname)s: %(message)s", level=logging.INFO ) logging.info(f"working dir - {os.getcwd()}") random.seed(1234) np.random.seed(1234) torch.random.manual_seed(1234) parser = argparse.ArgumentParser(description="Read command line parameters.") parser.add_argument("-p", "--parameters", help="Path to config file.") args = parser.parse_args(sys.argv[1:]) with open(args.parameters, "r") as f: params = json.load(f) if "train_set" not in params: logging.error('"train_set" not defined') exit(1) if "vali_set" not in params: logging.error('"vali_set" not defined') exit(1) if "test_set" not in params: logging.error('"test_set" not defined') exit(1) # device = torch.device("cuda") if torch.cuda.is_available() else None device = None num_columns = params["num_columns"] if "num_columns" in params else 138 anchor_url_features = ( params["anchor_url_features"] if "anchor_url_features" in params else None ) body_features = params["body_features"] if "body_features" in params else None # uncomment to train behavior and target models # train_models(params) test_dataset = load_dataset( params["second_set"], num_columns, anchor_url_features, body_features, "second_set", ) weight_clamper = Clamper(min_v=0.0) estimators = [ DMEstimator(DecisionTreeTrainer(), 0.5, device=device), IPSEstimator(weight_clamper=weight_clamper, device=device), DoublyRobustEstimator( DecisionTreeTrainer(), 0.5, weight_clamper, False, device ), DoublyRobustEstimator(DecisionTreeTrainer(), 0.5, weight_clamper, True, device), PseudoInverseEstimator(weight_clamper=weight_clamper, device=device), PBMEstimator(weight_clamper=weight_clamper, device=device), ] metrics = ["ndcg", "err"] alphas = [0.0, 1.0, 2.0] trainers = [ (DecisionTreeTrainer(), LassoTrainer()), (LassoTrainer(), DecisionTreeTrainer()), ] for log_trainer, tgt_trainers in trainers: for metric in metrics: for alpha in alphas: evaluate( [(estimators, 200)] * 4, test_dataset, 5, 20, metric, log_trainer, RankingDistribution(alpha), "anchor_url_features", tgt_trainers, PassThruDistribution(), "body_features", "all_features", 4, )
20,558
32.648118
111
py
ReAgent
ReAgent-master/reagent/ope/test/cartpole.py
#!/usr/bin/env python3 import logging import gym import torch from reagent.ope.estimators.sequential_estimators import ( Action, ActionDistribution, ActionSpace, IPSEstimator, Model, NeuralDualDICE, RandomRLPolicy, RewardProbability, RLEstimatorInput, RLPolicy, State, StateDistribution, Transition, ) from reagent.ope.utils import RunningAverage logging.basicConfig(level=logging.INFO) logger = logging.getLogger(__name__) NUM_EPISODES = 200 MAX_HORIZON = 250 GAMMA = 0.99 class ComboPolicy(RLPolicy): # Weighted combination between two given policies def __init__(self, action_space: ActionSpace, weights, policies): assert len(weights) == len(policies) self._weights = weights self._policies = policies self._action_space = action_space def action_dist(self, state: State) -> ActionDistribution: weighted_policies = [ w * p(state).values for w, p in zip(self._weights, self._policies) ] weighted = torch.stack(weighted_policies).sum(0) return self._action_space.distribution(weighted) class PyTorchPolicy(RLPolicy): def __init__(self, action_space: ActionSpace, model): self._action_space = action_space self._model = model self._softmax = torch.nn.Softmax(dim=0) def action_dist(self, state: State) -> ActionDistribution: self._model.eval() dist = self._model(torch.tensor(state.value, dtype=torch.float).reshape(1, -1))[ 0 ] return self._action_space.distribution(self._softmax(dist)) class EnvironmentModel(torch.nn.Module): def __init__(self, state_dim, action_dim, hidden_dim, hidden_layers, activation): super(EnvironmentModel.Network, self).__init__() self._state_dim = state_dim self._action_dim = action_dim self._hidden_dim = hidden_dim self._hidden_layers = hidden_layers self._activation = activation self.layers = [] dim = self._state_dim + self._action_dim for _ in range(self._hidden_layers): self.layers.append(torch.nn.Linear(dim, self._hidden_dim)) self.layers.append(self._activation()) dim = self._hidden_dim # Output is the next state and its reward self.layers.append(torch.nn.Linear(dim, self._state_dim + 1)) self.model = torch.nn.Sequential(*self.layers) def forward(self, state: torch.Tensor, action: torch.Tensor): x = torch.cat((state, action), dim=1) return self.model(x) class ModelWrapper(Model): def __init__(self, model: EnvironmentModel, device=None): self._model = model self._device = device self._model.to(self._device) def next_state_reward_dist(self, state: State, action: Action) -> StateDistribution: self._model.eval() state_reward_tensor = ( self._model( torch.tensor(state.value, dtype=torch.float) .reshape(-1, self._model._state_dim) .to(self._device), torch.nn.functional.one_hot( torch.tensor(action.value, dtype=torch.long), self._model._action_dim, ) .reshape(-1, self._model._action_dim) .float() .to(self._device), ) .reshape(-1) .cpu() ) return { State(state_reward_tensor[: self._model._state_dim]): RewardProbability( state_reward_tensor[-1].item() ) } def to(self, device): self._model.to(device) def generate_logs(episodes: int, max_horizon: int, policy: RLPolicy): """ Args: episodes: number of episodes to generate max_horizon: max horizon of each episode policy: RLPolicy which uses real-valued states """ log = [] env = gym.make("CartPole-v0") for _ in range(episodes): init_state = env.reset() cur_state = init_state mdp = [] for _ in range(max_horizon): action_dist = policy(State(cur_state)) action = action_dist.sample()[0].value action_prob = action_dist.probability(Action(action)) next_state, _, done, _ = env.step(action) mdp.append( Transition( last_state=State(cur_state), action=Action(action), action_prob=action_prob, state=State(next_state), reward=1.0, status=Transition.Status.NORMAL, ) ) cur_state = next_state if done: log.append(mdp) break log.append(mdp) return log def zeta_nu_loss_callback(losses, estimated_values, input: RLEstimatorInput): def callback_fn(zeta_loss, nu_loss, estimator): losses.append((zeta_loss, nu_loss)) estimated_values.append(estimator._compute_estimates(input)) return callback_fn def estimate_value(episodes: int, max_horizon: int, policy: RLPolicy, gamma: float): avg = RunningAverage() env = gym.make("CartPole-v0") for _ in range(episodes): init_state = env.reset() cur_state = init_state r = 0.0 discount = 1.0 for _ in range(max_horizon): action_dist = policy(State(cur_state)) action = action_dist.sample()[0].value next_state, _, done, _ = env.step(action) reward = 1.0 r += reward * discount discount *= gamma if done: break cur_state = next_state avg.add(r) return avg.average def run_dualdice_test(model_path: str, alpha: float): device = torch.device("cuda") if torch.cuda.is_available() else None logger.info(f"Device - {device}") model = torch.jit.load(model_path) model = model.dqn_with_preprocessor.model random_policy = RandomRLPolicy(ActionSpace(2)) model_policy = PyTorchPolicy(ActionSpace(2), model) target_policy = ComboPolicy( ActionSpace(2), [0.7, 0.3], [model_policy, random_policy] ) behavior_policy = ComboPolicy( ActionSpace(2), [0.55 + 0.15 * alpha, 0.45 - 0.15 * alpha], [model_policy, random_policy], ) ground_truth = estimate_value(NUM_EPISODES, MAX_HORIZON, target_policy, GAMMA) log_policy_value = estimate_value(NUM_EPISODES, MAX_HORIZON, behavior_policy, GAMMA) trained_policy_value = estimate_value( NUM_EPISODES, MAX_HORIZON, model_policy, GAMMA ) logger.info(f"Target Policy Ground Truth value: {ground_truth}") logger.info(f"Behavior Policy Ground Truth value: {log_policy_value}") logger.info(f"Model Policy Ground Truth value: {trained_policy_value}") log = generate_logs(NUM_EPISODES, MAX_HORIZON, behavior_policy) inp = RLEstimatorInput( gamma=GAMMA, log=log, target_policy=target_policy, discrete_states=False ) ips = IPSEstimator() dualdice_losses = [] dualdice_values = [] dualdice = NeuralDualDICE( state_dim=4, action_dim=2, deterministic_env=True, average_next_v=False, value_lr=0.003, zeta_lr=0.003, batch_size=2048, reporting_frequency=1000, training_samples=100000, loss_callback_fn=zeta_nu_loss_callback(dualdice_losses, dualdice_values, inp), device=device, ) ips_result = ips.evaluate(inp) dd_result = dualdice.evaluate(inp) return { "ips_estimate": ips_result, "dualdice_estimate": dd_result, "ground_truth": ground_truth, "dualdice_losses": dualdice_losses, "dualdice_estimates_per_epoch": dualdice_values, } if __name__ == "__main__": run_dualdice_test( "/mnt/vol/gfsfblearner-nebraska/flow/data/2020-07-27/a56cd422-794b-4866-9b73-5de95fb65700/207851498_207851498_0.pt", 0.0, )
8,083
31.079365
124
py
ReAgent
ReAgent-master/reagent/ope/test/gridworld.py
#!/usr/bin/env python3 import logging import random from typing import Iterable, Optional, Sequence, Tuple import numpy as np import torch from reagent.ope.estimators.sequential_estimators import ( DMEstimator, DoublyRobustEstimator, EpsilonGreedyRLPolicy, IPSEstimator, MAGICEstimator, NeuralDualDICE, RandomRLPolicy, RewardProbability, RLEstimatorInput, State, StateDistribution, StateReward, ValueFunction, ) from reagent.ope.estimators.types import Action, ActionSpace from reagent.ope.test.envs import Environment, PolicyLogGenerator from reagent.ope.trainers.rl_tabular_trainers import ( DPTrainer, DPValueFunction, EstimatedStateValueFunction, TabularPolicy, ) class GridWorld(Environment): def __init__( self, size: Tuple[int, int], start: Tuple[int, int], goal: Tuple[int, int], max_horizon: int = -1, walls: Iterable[Tuple[int, int]] = (), use_taxicab_reward: bool = False, ): super().__init__(max_horizon) self.size = size self.start = start self.goal = goal self.walls = set(walls) self.use_taxicab_reward = use_taxicab_reward self.reset() @classmethod def from_grid( cls, grid: Sequence[Sequence[str]], max_horizon: int = -1, use_taxicab_reward: bool = False, ): size = (len(grid), len(grid[0])) start = (0, 0) goal = (0, 0) walls = [] for x, r in enumerate(grid): for y, c in enumerate(r): g = c.lower() if g == "s": start = (x, y) elif g == "g": goal = (x, y) elif g == "w": walls += ((x, y),) return cls(size, start, goal, max_horizon, walls, use_taxicab_reward) @classmethod def random_grid( cls, length: int, max_horizon: int = -1, wall_prob: float = 0.1, use_taxicab_reward: bool = False, ): """ Generates a random grid of size length x length with start = (0, 0) and goal = (length-1, length-1) """ size = (length, length) start = (0, 0) goal = (length - 1, length - 1) walls = [] for r in range(length): for c in range(length): if (r, c) == start or (r, c) == goal: continue else: if random.uniform(0, 1) < wall_prob: walls.append((r, c)) return cls(size, start, goal, max_horizon, walls, use_taxicab_reward) def reset(self, state: Optional[State] = None): super().reset(state) if self._current_state is None: self._current_state = State(self.start) return self._current_state def close(self): pass def _validate(self, pos: Tuple[int, int]) -> bool: return ( 0 <= pos[0] < self.size[0] and 0 <= pos[1] < self.size[1] and pos not in self.walls ) def _transit( self, from_pos: Tuple[int, int], to_pos: Tuple[int, int] ) -> Tuple[Tuple[int, int], float, bool]: if not self._validate(to_pos): return from_pos, 0.0, False elif to_pos == self.goal: return to_pos, 1.0, True else: return ( to_pos, 0.0 if not self.use_taxicab_reward else np.exp(-2 * self._taxi_distance(to_pos, self.goal) / self.size[0]), False, ) def _taxi_distance( self, from_pos: Tuple[int, int], to_pos: Tuple[int, int] ) -> float: return abs(from_pos[0] - to_pos[0]) + abs(from_pos[1] - to_pos[1]) def _next_state_reward(self, state: State, action: Action) -> StateReward: value = state.value assert isinstance(value, tuple), f"got type {type(value)} instead of tuple" # pyre-fixme[23]: Unable to unpack single value, 2 were expected. (x, y) = value assert isinstance(x, int) and isinstance( y, int ), "Gridworld expects states to be Tuple[int, int]" if state.value in self.walls or state.value == self.goal: return StateReward(State((x, y), state.is_terminal), 0.0) if action.value == 0: to_pos, reward, is_end = self._transit((x, y), (x + 1, y)) elif action.value == 1: to_pos, reward, is_end = self._transit((x, y), (x, y + 1)) elif action.value == 2: to_pos, reward, is_end = self._transit((x, y), (x - 1, y)) else: to_pos, reward, is_end = self._transit((x, y), (x, y - 1)) return StateReward(State(to_pos, is_end), reward) def next_state_reward_dist(self, state: State, action: Action) -> StateDistribution: sr = self._next_state_reward(state, action) assert sr.state is not None return {sr.state: RewardProbability(sr.reward, 1.0)} @property def observation_space(self): return (2,), ((0, self.size[0]), (0, self.size[1])) @property def states(self): for x in range(self.size[0]): for y in range(self.size[1]): state = (x, y) if state != self.goal and state not in self.walls: yield State((x, y)) def __repr__(self): dump = "" for x in range(self.size[0]): for y in range(self.size[1]): pos = (x, y) if pos == self.start: dump += "\u2b55" elif pos == self.goal: dump += "\u2b50" elif pos in self.walls: dump += "\u2b1b" else: dump += "\u2b1c" dump += "\n" return dump def dump_state_values(self, state_values) -> str: dump = "" for x in range(self.size[0]): for y in range(self.size[1]): pos = State((x, y)) value = 0.0 if pos in state_values: value = state_values[pos] dump += "{:6.3}".format(value) dump += "\n" return dump def dump_value_func(self, valfunc: ValueFunction) -> str: dump = "" for x in range(self.size[0]): for y in range(self.size[1]): dump += "{:6.3}".format(valfunc(State((x, y)))) dump += "\n" return dump def dump_policy(self, policy) -> str: dump = "" for x in range(self.size[0]): for y in range(self.size[1]): pos = (x, y) if pos == self.start: dump += "\u2b28" elif pos == self.goal: dump += "\u2b27" elif pos in self.walls: dump += "\u2588" else: action = policy(State(pos)).greedy()[0] if action.value == 0: dump += "\u21e9" elif action.value == 1: dump += "\u21e8" elif action.value == 2: dump += "\u21e7" else: dump += "\u21e6" dump += "\n" return dump class ThomasGridWorld(GridWorld): """ GridWorld set up in https://people.cs.umass.edu/~pthomas/papers/Thomas2015c.pdf """ def __init__(self): super().__init__((4, 4), (0, 0), (3, 3), 100) def _transit( self, from_pos: Tuple[int, int], to_pos: Tuple[int, int] ) -> Tuple[Tuple[int, int], float, bool]: if not self._validate(to_pos): return from_pos, 0.0, False elif to_pos == (1, 2): return to_pos, -10.0, False elif to_pos == (1, 3): return to_pos, 1.0, False elif to_pos == self.goal: return to_pos, 10.0, True else: return to_pos, -1.0, False class NoiseGridWorldModel(Environment): def __init__( self, gridworld: GridWorld, action_space: ActionSpace, epsilon: float = 0.1, max_horizon: int = -1, ): super().__init__(max_horizon) self._gridworld = gridworld self.action_space = action_space self.epsilon = epsilon self.noise_prob = epsilon / (len(action_space) - 1) def reset(self, state: Optional[State] = None): self._gridworld.reset(state) self._steps_taken = 0 return self._gridworld.current_state def close(self): pass def next_state_reward_dist(self, state: State, action: Action) -> StateDistribution: probs = [self.noise_prob] * len(self.action_space) assert isinstance( action.value, int ), f"got type {type(action.value)} instead of int" # pyre-fixme[16]: `int` has no attribute `__setitem__`. probs[action.value] = 1 - self.epsilon states = {} for a in self.action_space: sr = self._gridworld._next_state_reward(state, a) if sr.state in states: rp = states[sr.state] states[sr.state] = RewardProbability( rp.reward + sr.reward, # pyre-fixme[16]: `int` has no attribute `__getitem__`. rp.prob + probs[a.value], ) else: states[sr.state] = RewardProbability(sr.reward, probs[a.value]) return states @property def observation_space(self): return self._gridworld.observation_space @property def states(self): return self._gridworld.states @property def current_state(self): return self._gridworld.current_state @current_state.setter def current_state(self, state: Optional[None]): self._gridworld._current_state = state GAMMA = 0.9 USE_DP_VALUE_FUNC = True if __name__ == "__main__": logging.basicConfig(level=logging.INFO) random.seed(1234) np.random.seed(1234) torch.random.manual_seed(1234) device = torch.device("cuda") if torch.cuda.is_available() else None print(f"device - {device}") gridworld = GridWorld.random_grid(10, max_horizon=250, use_taxicab_reward=True) logging.info(f"GridWorld:\n{gridworld}") action_space = ActionSpace(4) opt_policy = TabularPolicy(action_space) trainer = DPTrainer(gridworld, opt_policy) value_func = trainer.train(gamma=GAMMA) logging.info(f"Opt Policy:\n{gridworld.dump_policy(opt_policy)}") logging.info(f"Opt state values:\n{gridworld.dump_value_func(value_func)}") # behavivor_policy = RandomRLPolicy(action_space) behavivor_policy = EpsilonGreedyRLPolicy(opt_policy, 0.7) target_policy = EpsilonGreedyRLPolicy(opt_policy, 0.3) model = NoiseGridWorldModel(gridworld, action_space, epsilon=0.1, max_horizon=1000) value_func = DPValueFunction(target_policy, model, GAMMA) ground_truth: Optional[ValueFunction] = None if USE_DP_VALUE_FUNC: ground_truth = DPValueFunction(target_policy, gridworld, GAMMA) else: ground_truth = EstimatedStateValueFunction(target_policy, gridworld, GAMMA) logging.info( f"Target Policy ground truth values:\n" f"{gridworld.dump_value_func(ground_truth)}" ) logging.info( f"Logging Policy values:\n" f"{gridworld.dump_value_func(DPValueFunction(behavivor_policy, model, GAMMA))}" ) log = [] log_generator = PolicyLogGenerator(gridworld, behavivor_policy) num_episodes = 50 for state in gridworld.states: for _ in range(num_episodes): log.append(log_generator.generate_log(state)) logging.info(f"Generated {num_episodes} logs for {state}") estimator_input = RLEstimatorInput( gamma=GAMMA, log=log, target_policy=target_policy, value_function=value_func, ground_truth=ground_truth, ) NeuralDualDICE( device=device, state_dim=2, action_dim=4, deterministic_env=True, average_next_v=False, value_lr=0.001, zeta_lr=0.0001, batch_size=512, ).evaluate(estimator_input) DMEstimator(device=device).evaluate(estimator_input) IPSEstimator(weight_clamper=None, weighted=False, device=device).evaluate( estimator_input ) IPSEstimator(weight_clamper=None, weighted=True, device=device).evaluate( estimator_input ) DoublyRobustEstimator(weight_clamper=None, weighted=False, device=device).evaluate( estimator_input ) DoublyRobustEstimator(weight_clamper=None, weighted=True, device=device).evaluate( estimator_input ) MAGICEstimator(device=device).evaluate( estimator_input, num_resamples=10, loss_threhold=0.0000001, lr=0.00001 )
13,049
30.90709
88
py
ReAgent
ReAgent-master/reagent/ope/test/multiclass_bandits.py
#!/usr/bin/env python3 import argparse import json import logging import os import random import sys from dataclasses import dataclass from pathlib import PurePath from typing import Iterable, Tuple import numpy as np import pandas as pd import torch from reagent.ope.estimators.contextual_bandits_estimators import ( Action, ActionDistribution, ActionRewards, BanditsEstimatorInput, BanditsModel, DMEstimator, DoublyRobustEstimator, IPSEstimator, LogSample, ) from reagent.ope.estimators.estimator import Estimator, Evaluator from reagent.ope.estimators.types import ActionSpace, Policy, Trainer, TrainingData from reagent.ope.trainers.linear_trainers import ( DecisionTreeTrainer, LogisticRegressionTrainer, SGDClassifierTrainer, ) from torch import Tensor @dataclass(frozen=True) class MultiClassDataRow(object): feature: torch.Tensor label: torch.Tensor one_hot: torch.Tensor class UCIMultiClassDataset: """ To load and hold UCI classification datasets: https://archive.ics.uci.edu/ml/datasets.php?task=cla&sort=nameUp&view=table Also to convert it to contextual bandits problems References: https://arxiv.org/abs/1103.4601 """ def __init__(self, params, device=None): if "file" not in params: raise Exception('Please define "file" in "dataset"') if "label_col" not in params: raise Exception('Please define "label_col" in "dataset"') index_col = params["index_col"] if "index_col" in params else None label_col = params["label_col"] sep = params["sep"] if "sep" in params else "," self._config_file = params["file"] self._data_frame = pd.read_csv( self._config_file, sep=sep, header=None, index_col=index_col if index_col is not None else False, ) file_col_len = self._data_frame.shape[1] + (1 if index_col is not None else 0) if label_col < 0: label_col = file_col_len + label_col frame_label_col = label_col if index_col is not None and label_col > index_col: label_col = label_col - 1 self._features = torch.as_tensor( self._data_frame.iloc[ :, [i for i in range(self._data_frame.shape[1]) if i != label_col] ].values, dtype=torch.float32, device=device, ) self._classes = self._data_frame[frame_label_col].unique() self._classes.sort() self._labels = self._data_frame[frame_label_col].values self._class_indices = torch.tensor( [np.where(self._classes == i)[0][0] for i in self._labels], dtype=torch.long, device=device, ) self._one_hots = torch.zeros( (self._class_indices.shape[0], len(self._classes)), dtype=torch.int, device=device, ) self._one_hots[ torch.arange(self._one_hots.shape[0], dtype=torch.long), self._class_indices ] = 1 self.device = device def __len__(self): return len(self._data_frame) def __getitem__(self, idx) -> MultiClassDataRow: return MultiClassDataRow( self._features[idx], self._class_indices[idx], self._one_hots[idx] ) @property def config_file(self) -> str: return self._config_file @property def num_features(self) -> int: return self._features.shape[1] @property def num_actions(self) -> int: return len(self._classes) @property def features(self) -> torch.Tensor: return self._features @property def labels(self) -> torch.Tensor: return self._class_indices @property def one_hots(self) -> torch.Tensor: return self._one_hots def train_val_test_split( self, ratios: Tuple[float, float] = (0.8, 0.8), device=None ): total_len = len(self._data_frame) train_len = int(total_len * ratios[0]) train_choices = random.sample(range(total_len), train_len) train_x = np.take(self._features, train_choices, axis=0) train_y = np.take(self._class_indices, train_choices) train_r = np.take(self._one_hots, train_choices, axis=0) fit_len = int(train_len * ratios[1]) fit_choices = random.sample(range(train_len), fit_len) fit_x = np.take(train_x, fit_choices, axis=0) fit_y = np.take(train_y, fit_choices) fit_r = np.take(train_r, fit_choices, axis=0) val_x = np.delete(train_x, fit_choices, axis=0) val_y = np.delete(train_y, fit_choices) val_r = np.delete(train_r, fit_choices, axis=0) test_x = np.delete(self._features, train_choices, axis=0) test_y = np.delete(self._class_indices, train_choices) test_r = np.delete(self._one_hots, train_choices, axis=0) return ( torch.as_tensor(fit_x, dtype=torch.float, device=device), torch.as_tensor(fit_y, dtype=torch.float, device=device), torch.as_tensor(fit_r, dtype=torch.float, device=device), torch.as_tensor(val_x, dtype=torch.float, device=device), torch.as_tensor(val_y, dtype=torch.float, device=device), torch.as_tensor(val_r, dtype=torch.float, device=device), torch.as_tensor(test_x, dtype=torch.float, device=device), torch.as_tensor(test_y, dtype=torch.float, device=device), torch.as_tensor(test_r, dtype=torch.float, device=device), train_choices, ) @dataclass(frozen=True) class MultiClassContext: query_id: int class MultiClassModel(BanditsModel): def __init__(self, features: Tensor, rewards: Tensor): self._features = features self._rewards = rewards def _action_rewards(self, context: MultiClassContext) -> ActionRewards: return ActionRewards(self._rewards[context.query_id]) class MultiClassPolicy(Policy): def __init__( self, action_space: ActionSpace, action_distributions: Tensor, epsilon: float, device=None, ): super().__init__(action_space, device) self._action_distributions = action_distributions self._exploitation_prob = 1.0 - epsilon self._exploration_prob = epsilon / len(self.action_space) def _query(self, context: int) -> Tuple[Action, ActionDistribution]: dist = self._action_distributions[context] if len(dist.shape) > 1 and dist.shape[0] == 1: dist = dist[0] if dist.shape[0] < len(self.action_space): dist = torch.cat( (dist, torch.zeros([len(self.action_space) - dist.shape[0]])) ) dist = dist * self._exploitation_prob + self._exploration_prob action = torch.multinomial(dist, 1).item() return Action(action), ActionDistribution(dist) def evaluate_all( experiments: Iterable[Tuple[Iterable[Estimator], int]], dataset: UCIMultiClassDataset, log_trainer: Trainer, log_epsilon: float, tgt_trainer: Trainer, tgt_epsilon: float, max_num_workers: int, random_reward_prob: float = 0.0, device=None, ): action_space = ActionSpace(dataset.num_actions) config_path = PurePath(dataset.config_file) data_name = config_path.stem log_model_name = data_name + "_" + log_trainer.__class__.__name__ + ".pickle" log_model_file = str(config_path.with_name(log_model_name)) tgt_model_name = data_name + "_" + tgt_trainer.__class__.__name__ + ".pickle" tgt_model_file = str(config_path.with_name(tgt_model_name)) log_trainer.load_model(log_model_file) tgt_trainer.load_model(tgt_model_file) if not log_trainer.is_trained or not tgt_trainer.is_trained: ( train_x, train_y, train_r, val_x, val_y, val_r, test_x, test_y, test_r, train_choices, ) = dataset.train_val_test_split((0.2, 0.8)) trainer_data = TrainingData(train_x, train_y, None, val_x, val_y, None) if not log_trainer.is_trained: log_trainer.train(trainer_data) log_trainer.save_model(log_model_file) if not tgt_trainer.is_trained: tgt_trainer.train(trainer_data) tgt_trainer.save_model(tgt_model_file) log_results = log_trainer.predict(dataset.features) assert log_results.probabilities is not None log_policy = MultiClassPolicy(action_space, log_results.probabilities, log_epsilon) tgt_results = tgt_trainer.predict(dataset.features) assert tgt_results.probabilities is not None tgt_policy = MultiClassPolicy(action_space, tgt_results.probabilities, tgt_epsilon) tasks = [] # pyre-fixme[61]: `train_choices` may not be initialized here. test_queries = list(set(range(len(dataset))) - set(train_choices)) for estimators, num_samples in experiments: samples = [] for _ in range(num_samples): qid = random.sample(test_queries, 1) label = int(dataset.labels[qid].item()) log_action, log_action_probabilities = log_policy(qid) log_reward = 1.0 if log_action.value == label else 0.0 tgt_action, tgt_action_probabilities = tgt_policy(qid) ground_truth_reward = 1.0 if tgt_action.value == label else 0.0 item_feature = dataset.features[qid] random_reward = random.random() < random_reward_prob samples.append( LogSample( context=qid, log_action=log_action, log_reward=random.randint(0, 1) if random_reward else log_reward, log_action_probabilities=log_action_probabilities, tgt_action_probabilities=tgt_action_probabilities, tgt_action=tgt_action, ground_truth_reward=ground_truth_reward, item_feature=item_feature, ) ) tasks.append((estimators, BanditsEstimatorInput(action_space, samples, False))) evaluator = Evaluator(tasks, max_num_workers) results = evaluator.evaluate() Evaluator.report_results(results) return results DEFAULT_ITERATIONS = 500 if __name__ == "__main__": logging.basicConfig(level=logging.INFO) logging.info(f"working dir - {os.getcwd()}") parser = argparse.ArgumentParser(description="Read command line parameters.") parser.add_argument("-p", "--parameters", help="Path to config file.") args = parser.parse_args(sys.argv[1:]) with open(args.parameters, "r") as f: params = json.load(f) if "dataset" not in params: raise Exception('Please define "dataset" in config file') random.seed(1234) np.random.seed(1234) torch.random.manual_seed(1234) dataset = UCIMultiClassDataset(params["dataset"]) log_trainer = LogisticRegressionTrainer() log_epsilon = 0.1 tgt_trainer = SGDClassifierTrainer() tgt_epsilon = 0.1 dm_trainer = DecisionTreeTrainer() experiments = [ ( ( DMEstimator(DecisionTreeTrainer()), IPSEstimator(), DoublyRobustEstimator(DecisionTreeTrainer()), ), 1000, ) for _ in range(100) ] evaluate_all( experiments, dataset, log_trainer, log_epsilon, tgt_trainer, tgt_epsilon, 0 )
11,560
33.927492
88
py
ReAgent
ReAgent-master/reagent/ope/test/yandex_web_search.py
#!/usr/bin/env python3 import argparse import json import logging import os import pickle import random import sys import time from typing import ( Dict, Iterable, List, Mapping, MutableMapping, Optional, Sequence, Tuple, Union, ) import numpy as np import torch import torch.multiprocessing as mp from reagent.ope.estimators.estimator import Evaluator from reagent.ope.estimators.slate_estimators import ( DCGSlateMetric, ERRSlateMetric, FrechetDistribution, IPSEstimator, LogSample, NDCGSlateMetric, PBMEstimator, PseudoInverseEstimator, RankingDistribution, RewardDistribution, SlateContext, SlateEstimator, SlateEstimatorInput, SlateItemValues, SlateModel, SlateQuery, SlateSlots, SlateSlotValues, ) from reagent.ope.utils import RunningAverage # Slate test using Yandex Personalized Web Search Dataset: # https://www.kaggle.com/c/yandex-personalized-web-search-challenge/ RELEVANT_THRESHOLD = 49 HIGHLY_RELEVANT_THRESHOLD = 399 MAX_SLATE_SIZE = 10 MIN_QUERY_COUNT = 10 def click_to_relevances( clicks: Iterable[Tuple[int, int]], urls: Sequence[Tuple[int, int]] ) -> Tuple[List[float], Mapping[Tuple[int, int], float]]: position_relevances = [0.0] * max(len(urls), MAX_SLATE_SIZE) url_relevances = {url: 0.0 for url in urls} for i, dt in clicks: r = 0.0 if dt > HIGHLY_RELEVANT_THRESHOLD: r = 2.0 elif dt > RELEVANT_THRESHOLD: r = 1.0 position_relevances[i] = r url_relevances[urls[i]] = r return position_relevances, url_relevances class LoggedQuery: def __init__( self, user_id: int, query_id: int, query_terms: Tuple[int], list: Sequence[Tuple[int, int]], ): self._user_id = user_id self._query_id = query_id self._query_terms = query_terms self._list = list self._clicks: List[Tuple[int, int]] = [] self._position_relevances: Optional[List[float]] = None self._url_relevances: Optional[MutableMapping[Tuple[int, int], float]] = None def click(self, url_id: int, dwell_time: int): self._position_relevances = None self._url_relevances = None i = 0 for r in self.list: if url_id == r[0]: self.clicks.append((i, dwell_time)) break i += 1 @property def user_id(self): return self._user_id @property def query_id(self): return self._query_id @property def query_terms(self): return self._query_terms @property def list(self): return self._list @property def clicks(self): return self._clicks def _click_to_relevances(self): self._position_relevances = [0.0] * max(len(self._list), MAX_SLATE_SIZE) self._url_relevances = {url: 0.0 for url in self._list} for i, dt in self.clicks: r = 0.0 if dt > HIGHLY_RELEVANT_THRESHOLD: r = 2.0 elif dt > RELEVANT_THRESHOLD: r = 1.0 self._position_relevances[i] = r self._url_relevances[self._list[i]] = r @property def position_relevances(self): if self._position_relevances is None: self._click_to_relevances() return self._position_relevances @property def url_relevances(self): if self._url_relevances is None: self._click_to_relevances() return self._url_relevances class TrainingQuery: def __init__(self, query_id: int, query_terms: Tuple[int]): self._query_id = query_id self._query_terms = query_terms self._count = 0 self._url_relevances: MutableMapping[Tuple[int, int], RunningAverage] = {} self._position_relevances = [RunningAverage() for _ in range(MAX_SLATE_SIZE)] def add(self, query: LoggedQuery): self._count += 1 urs = query.url_relevances for item_id, r in urs.items(): if item_id not in self._url_relevances: self._url_relevances[item_id] = RunningAverage(r) else: self._url_relevances[item_id].add(r) prs = query.position_relevances for i in range(MAX_SLATE_SIZE): self._position_relevances[i].add(prs[i]) def merge(self, other: "TrainingQuery"): for i, r in other.url_relevances.items(): if i not in self._url_relevances: self._url_relevances[i] = RunningAverage(r) else: self._url_relevances[i].add(r) for i in range(MAX_SLATE_SIZE): self._position_relevances[i].add(other.position_relevances[i]) def finalize(self): self._url_relevances = {k: v.average for k, v in self._url_relevances.items()} self._position_relevances = [v.average for v in self._position_relevances] def pack(self): if isinstance(self._url_relevances, Mapping): self._url_relevances = list(self._url_relevances.items()) def _unpack(self): if isinstance(self._url_relevances, Sequence): self._url_relevances = {v[0]: v[1] for v in self._url_relevances} @property def count(self): return self._count @property def query_id(self): return self._query_id @property def query_terms(self): return self._query_terms @property def url_relevances(self): self._unpack() return self._url_relevances @property def position_relevances(self): return self._position_relevances def create_cache(params): if "folder" not in params: raise Exception('Please define "folder" in "raw_data"') folder = params["folder"] if "folder" in params else "" if len(folder) == 0: folder = os.getcwd() cache_folder = params["cache_folder"] if "cache_folder" in params else folder if len(cache_folder) == 0: cache_folder = folder source_file = params["source_file"] if "source_file" in params else "" if len(source_file) == 0: raise Exception('"source_file" not defined!') total_days = params["total_days"] if "total_days" in params else 27 text_file = os.path.join(folder, source_file) logging.info(f"loading {text_file}") if not os.access(text_file, os.R_OK): logging.warning(f"{text_file} cannot be accessed.") return for d in range(1, total_days + 1): pickle_file = os.path.join(cache_folder, f"{source_file}_{d:02}.pickle") logging.info(f"creating cache for day {d:02}: {pickle_file}") queries = [] st = time.process_time() with open(text_file, "r") as f: curr_sess = None curr_user = -1 num_sess = 0 last_click = None for line in f: tokens = line.strip().split() tlen = len(tokens) if tlen == 4 and tokens[1] == "M": if last_click is not None: query = curr_sess[2][last_click[0]] query.click(last_click[1], 10000) last_click = None day = int(tokens[2]) if day != d: continue num_sess += 1 if num_sess % 100000 == 0: logging.info(f" {num_sess} session processed...") if curr_sess is not None: qids = set() for q in curr_sess[2].values(): if len(q.clicks) > 0: queries.append(q) elif q.query_id not in qids: queries.append(q) qids.add(q.query_id) del qids curr_sess = (int(tokens[0]), int(tokens[2]), {}) curr_user = int(tokens[3]) elif ( curr_sess is not None and tlen > 4 and int(tokens[0]) == curr_sess[0] ): t = int(tokens[1]) if last_click is not None: query = curr_sess[2][last_click[0]] query.click(last_click[1], t - last_click[2]) last_click = None if tokens[2] == "Q": serp_id = int(tokens[3]) query_id = int(tokens[4]) query_terms = tuple([int(s) for s in tokens[5].split(",")]) results = [] for r in tokens[6:]: rs = r.split(",") results.append((int(rs[0]), int(rs[1]))) query = LoggedQuery(curr_user, query_id, query_terms, results) curr_sess[2][serp_id] = query elif tokens[2] == "C": last_click = (int(tokens[3]), int(tokens[4]), t) else: logging.warning(f"unknown record type: {tokens[2]}") logging.info(f" loading time: {time.process_time() - st}") st = time.process_time() try: with open(pickle_file, "wb") as f: pickle.dump(queries, f, protocol=pickle.HIGHEST_PROTOCOL) except Exception: logging.error(f"{pickle_file} cannot be accessed.") logging.info(f" saving time: {time.process_time() - st}") def load_logged_queries(params) -> Sequence[TrainingQuery]: logging.info("loading logged queries...") if "folder" not in params: raise Exception('Please define "folder" in "raw_data"') folder = params["folder"] if "folder" in params else "" if len(folder) == 0: folder = os.getcwd() cache_file_name = params["cache_file_name"] if "cache_file_name" in params else "" cache_file = os.path.join(folder, f"{cache_file_name}.pickle") if len(cache_file_name) > 0 and os.access(cache_file, os.R_OK): logging.info(f" loading {cache_file}") try: st = time.perf_counter() with open(cache_file, "rb") as f: logged_queries = pickle.load(f) logging.info(f" loading time {time.perf_counter() - st}") return logged_queries except Exception as err: logging.warning(f" loading error {err}") base_file_name = params["base_file_name"] if "base_file_name" in params else "" if len(base_file_name) == 0: raise Exception('"base_file_name" not defined!') days = params["days"] if "days" in params else [] all_queries = {} st = time.perf_counter() for day in days: pickle_file = os.path.join(folder, f"{base_file_name}_{day:02}.pickle") if os.access(pickle_file, os.R_OK): logging.info(f" loading {pickle_file}") with open(pickle_file, "rb") as f: queries = pickle.load(f) if queries is None: logging.warning(f" loading {pickle_file} failed!") else: logging.info(f" loaded queries: {len(queries)}") for q in queries: if q.query_id in all_queries: tq = all_queries[q.query_id] else: tq = TrainingQuery(q.query_id, q.query_terms) all_queries[q.query_id] = tq tq.add(q) else: logging.warning(f" {pickle_file} not accessible!") logging.info(f" loading time {time.perf_counter() - st}") logged_queries = tuple(all_queries.values()) for v in logged_queries: v.finalize() if len(cache_file_name) > 0: logging.info(f" saving logged queries to {cache_file}") try: st = time.perf_counter() with open(cache_file, "wb") as f: pickle.dump(logged_queries, f, protocol=pickle.HIGHEST_PROTOCOL) logging.info(f" saving time {time.perf_counter() - st}") except Exception: logging.warning(f" {cache_file} not accessible!") return logged_queries class TrainingDataset: def __init__(self, params, device=None): if "folder" not in params: raise Exception('Please define "folder" in "dataset"') self._folder = params["folder"] self._days = params["days"] if "days" in params else [] self._base_file_name = ( params["base_file_name"] if "base_file_name" in params else "train" ) self._min_query_count = ( params["min_query_count"] if "min_query_count" in params else MIN_QUERY_COUNT ) self._cache_file = params["cache_file"] if "cache_file" in params else "" self._device = device self._queries = None self._query_ids = None self._query_terms = None self._position_relevances = None def load_queries(self, reload=False): logging.info("loading training queries...") pickle_file = os.path.join(self._folder, self._cache_file) if not reload and len(self._cache_file) > 0 and os.access(pickle_file, os.R_OK): logging.info(f" loading {pickle_file}") st = time.process_time() with open(pickle_file, "rb") as f: ( min_query_count, days, queries, query_ids, query_terms, position_relevances, ) = pickle.load(f) if min_query_count != self._min_query_count or days != self._days: logging.info(" updated config from last cache, reload") self.load_queries(True) else: self._queries = queries self._query_ids = query_ids self._query_terms = query_terms self._position_relevances = position_relevances logging.info( f" loaded {len(self._queries)}, " f" time {time.process_time() - st}" ) else: all_queries = {} for d in self._days: cache_file = os.path.join( self._folder, f"{self._base_file_name}_{d:02}.pickle" ) if os.access(cache_file, os.R_OK): logging.info(f" loading {cache_file}") st = time.process_time() with open(cache_file, "rb") as f: queries = pickle.load(f) if queries is None: logging.warning(f" loading {cache_file} failed!") continue logging.info(f" loaded queries: {len(queries)}") logging.info(f" loading time {time.process_time() - st}") st = time.process_time() for q in queries: if q.query_id not in all_queries: qr = TrainingQuery(q.query_id, q.query_terms) all_queries[q.query_id] = qr else: qr = all_queries[q.query_id] qr.add(q) logging.info(f" process time {time.process_time() - st}") else: logging.warning(f" {cache_file} not accessible!") self._queries = [] for v in all_queries.values(): if v.count >= self._min_query_count: v.finalize() v.pack() self._queries.append(v) self._query_ids = None self._query_terms = None self._position_relevances = None if len(self._cache_file) > 0: logging.info(f"saving training queries to {pickle_file}") try: st = time.process_time() with open(pickle_file, "wb") as f: self._process_training_queries() pickle.dump( ( self._min_query_count, self._days, self._queries, self._query_ids, self._query_terms, self._position_relevances, ), f, protocol=pickle.HIGHEST_PROTOCOL, ) logging.info(f" saving time {time.process_time() - st}") except Exception: logging.warning(f" {pickle_file} not accessible!") # self._query_ids = None # self._query_terms = None # self._position_relevances = None logging.info(f"loaded training queries: {len(self._queries)}") def _process_training_queries(self): if ( self._query_ids is not None and self._query_terms is not None and self._position_relevances is not None ): return logging.info("processing training queries...") st = time.process_time() self._query_ids = {} self._query_terms = {} self._position_relevances = [RunningAverage() for _ in range(MAX_SLATE_SIZE)] for q in self._queries: self._query_ids[q.query_id] = q for t in q.query_terms: if t in self._query_terms: self._query_terms[t].merge(q) else: mq = TrainingQuery(0, (t,)) mq.merge(q) self._query_terms[t] = mq for ra, r in zip(self._position_relevances, q.position_relevances): ra.add(r) for q in self._query_terms.values(): q.finalize() self._position_relevances = [v.average for v in self._position_relevances] logging.info(f"processing time {time.process_time() - st}") @property def training_queries(self): return self._queries def item_relevances( self, query_id: int, query_terms: Tuple[int], items: Iterable[Tuple[int, int]] ) -> SlateItemValues: self._process_training_queries() if query_id in self._query_ids: q = self._query_ids[query_id] rels = q.url_relevances else: ras = {} for t in query_terms: if t in self._query_terms: q = self._query_terms[t] for i, r in q.url_relevances: if i in ras: ra = ras[i] else: ra = RunningAverage() ras[i] = ra ra.add(r) rels = {i: r.average for i, r in ras.items()} item_rels = {} for i in items: if i in rels: item_rels[i] = rels[i] else: item_rels[i] = 0.0 return SlateItemValues(item_rels) def slot_relevances(self, slots: SlateSlots) -> SlateSlotValues: return SlateSlotValues(self._position_relevances[: len(slots)]) class YandexSlateModel(SlateModel): def __init__(self, dataset: TrainingDataset): self._dataset = dataset def item_rewards(self, context: SlateContext) -> SlateItemValues: query = context.query.value # pyre-fixme[20]: Call `TrainingDataset.item_relevances` expects argument `items`. return self._dataset.item_relevances(query[0], query[1:]) def slot_probabilities(self, context: SlateContext) -> SlateSlotValues: return self._dataset.slot_relevances(context.slots) def evaluate( experiments: Iterable[Tuple[Iterable[SlateEstimator], int]], log_dataset: TrainingDataset, log_distribution: RewardDistribution, tgt_dataset: TrainingDataset, tgt_distribution: RewardDistribution, log_queries: Sequence[TrainingQuery], slate_size: int, item_size: int, metric_func: str, max_num_workers: int, device=None, ): log_length = len(log_queries) slots = SlateSlots(slate_size) logging.info("Generating log...") st = time.perf_counter() tasks = [] total_samples = 0 for estimators, num_samples in experiments: samples = [] if num_samples * 10 > log_length: logging.warning(f"not enough log data, needs {num_samples * 10}") continue query_choices = np.random.choice(log_length, num_samples, replace=False) for i in query_choices: q = log_queries[i] # pyre-fixme[60]: Expected to unpack an iterable, but got `unknown`. context = SlateContext(SlateQuery((q.query_id, *(q.query_terms))), slots) url_relevances = q.url_relevances if len(url_relevances) > item_size: url_relevances = { k: v for k, v in sorted( url_relevances.items(), key=lambda item: item[1] )[:item_size] } items = url_relevances.keys() log_item_rewards = log_dataset.item_relevances( q.query_id, q.query_terms, items ) log_item_probs = log_distribution(log_item_rewards) tgt_item_rewards = tgt_dataset.item_relevances( q.query_id, q.query_terms, items ) tgt_item_probs = tgt_distribution(tgt_item_rewards) tgt_slot_expectation = tgt_item_probs.slot_item_expectations(slots) gt_item_rewards = SlateItemValues(url_relevances) if metric_func == "dcg": metric = DCGSlateMetric(device=device) elif metric_func == "err": metric = ERRSlateMetric(4.0, device=device) else: metric = NDCGSlateMetric(gt_item_rewards, device=device) slot_weights = metric.slot_weights(slots) if tgt_item_probs.is_deterministic: tgt_slate_prob = 1.0 log_slate = tgt_item_probs.sample_slate(slots) else: tgt_slate_prob = float("nan") log_slate = log_item_probs.sample_slate(slots) log_slate_prob = log_item_probs.slate_probability(log_slate) log_rewards = log_slate.slot_values(gt_item_rewards) log_reward = metric.calculate_reward(slots, log_rewards, None, slot_weights) gt_slot_rewards = tgt_slot_expectation.expected_rewards(gt_item_rewards) gt_reward = metric.calculate_reward( slots, gt_slot_rewards, None, slot_weights ) samples.append( LogSample( context, metric, log_slate, log_reward, log_slate_prob, None, log_item_probs, tgt_slate_prob, None, tgt_item_probs, gt_reward, slot_weights, ) ) total_samples += 1 tasks.append((estimators, SlateEstimatorInput(samples))) dt = time.perf_counter() - st logging.info(f"Generating log done: {total_samples} samples in {dt}s") logging.info("start evaluating...") st = time.perf_counter() evaluator = Evaluator(tasks, max_num_workers) Evaluator.report_results(evaluator.evaluate()) logging.info(f"evaluating done in {time.perf_counter() - st}s") if __name__ == "__main__": mp.set_start_method("spawn") logging.basicConfig( format="%(asctime)-15s_%(levelname)s: %(message)s", level=logging.INFO ) logging.info(f"working dir - {os.getcwd()}") random.seed(1234) np.random.seed(1234) torch.random.manual_seed(1234) parser = argparse.ArgumentParser(description="Read command line parameters.") parser.add_argument("-p", "--parameters", help="Path to config file.") args = parser.parse_args(sys.argv[1:]) with open(args.parameters, "r") as f: params = json.load(f) # uncomment to create cache for faster data loading # create_cache(params["raw_data"]) # device = torch.device("cuda") if torch.cuda.is_available() else torch.device("cpu") device = None logging.info('loading "log_data"') log_dataset = TrainingDataset(params["log_data"]) st = time.perf_counter() log_dataset.load_queries() logging.info(f"load time: {time.perf_counter() - st}") logging.info('loading "target_data"') tgt_dataset = TrainingDataset(params["target_data"]) st = time.perf_counter() tgt_dataset.load_queries() logging.info(f"load time: {time.perf_counter() - st}") logging.info('loading "test_data"') st = time.perf_counter() log_queries = load_logged_queries(params["test_data"]) logging.info(f"load time: {time.perf_counter() - st}") estimators = [IPSEstimator(), PseudoInverseEstimator(), PBMEstimator()] evaluate( [(estimators, 200)] * 4, log_dataset, RankingDistribution(1.0), tgt_dataset, FrechetDistribution(2.0, True), log_queries, 5, 10, "ndcg", 2, )
25,761
35.855508
90
py
ReAgent
ReAgent-master/reagent/ope/test/unit_tests/test_slate_estimators.py
#!/usr/bin/env python3 import random import unittest import torch from reagent.ope.estimators.slate_estimators import ( DCGSlateMetric, NDCGSlateMetric, Slate, SlateItem, SlateItemProbabilities, SlateItemValues, SlateSlotItemProbabilities, SlateSlots, ) class TestEstimator(unittest.TestCase): def setUp(self) -> None: random.seed(1234) torch.random.manual_seed(1234) self._item_relevances = [1.0, 0.5, 2.5, 2.0, 5.0] self._slot_item_relevances = [ [1.0, 0.5, 2.5, 2.0, 5.0], [1.5, 1.0, 2.0, 1.0, 4.0], [3.0, 2.5, 0.5, 4.0, 2.0], ] self._item_rewards = [3.0, 8.0, 0.0, 4.0, 1.0] self._item_size = len(self._item_relevances) self._slate_size = 3 self._slots = SlateSlots(self._slate_size) def test_slate_item_probabilities(self): probs = SlateItemProbabilities(self._item_relevances) slate = probs.sample_slate(self._slots) slate_prob = probs.slate_probability(slate) self.assertAlmostEqual(slate_prob, 0.017825312, places=2) slot_item_expectations = probs.slot_item_expectations(self._slots) slot_rewards = slot_item_expectations.expected_rewards( SlateItemValues(self._item_rewards) ) diff = slot_rewards.values - torch.tensor([1.81818, 2.13736, 2.66197]) self.assertAlmostEqual(diff.sum().item(), 0, places=5) def test_slate_slot_item_probabilities(self): probs = SlateSlotItemProbabilities( [SlateItemValues(vs) for vs in self._slot_item_relevances] ) slate = probs.sample_slate(self._slots) slate_prob = probs.slate_probability(slate) self.assertAlmostEqual(slate_prob, 0.02139037, places=2) slot_item_expectations = probs.slot_item_expectations() slot_rewards = slot_item_expectations.expected_rewards( SlateItemValues(self._item_rewards) ) diff = slot_rewards.values - torch.tensor([1.81818, 2.51352, 7.36929]) self.assertAlmostEqual(diff.sum().item(), 0, places=5) def test_metrics(self): dcg = DCGSlateMetric() ndcg = NDCGSlateMetric(SlateItemValues([1.0, 2.5, 2.0, 3.0, 1.5, 0.0])) item_rewards = SlateItemValues([2.0, 1.0, 0.0, 3.0, 1.5, 2.5]) slate = Slate([SlateItem(1), SlateItem(3), SlateItem(2)]) reward = dcg(slate.slots, slate.slot_values(item_rewards)) self.assertAlmostEqual(reward, 5.416508275) reward = ndcg(slate.slots, slate.slot_values(item_rewards)) self.assertAlmostEqual(reward, 0.473547669) slate = Slate([SlateItem(5), SlateItem(0), SlateItem(4)]) reward = dcg(slate.slots, slate.slot_values(item_rewards)) self.assertAlmostEqual(reward, 7.463857073) reward = ndcg(slate.slots, slate.slot_values(item_rewards)) self.assertAlmostEqual(reward, 0.652540703)
2,939
38.72973
79
py
ReAgent
ReAgent-master/reagent/ope/test/unit_tests/test_types.py
#!/usr/bin/env python3 import unittest from typing import Tuple, Union import numpy as np import torch from reagent.ope.estimators.types import ( ActionDistribution as Distribution, TypeWrapper, Values, ) class TestTypes(unittest.TestCase): TestType = Union[int, Tuple[int], float, Tuple[float], np.ndarray, torch.Tensor] TestClass = TypeWrapper[TestType] def setUp(self) -> None: self._test_list = [0, 1, 2, 3, 5] def test_int_type(self): int_val = TestTypes.TestClass(3) self.assertEqual(self._test_list[int_val], 3) self.assertEqual(hash(int_val), hash(3)) int_val_other = TestTypes.TestClass(3) self.assertEqual(int_val, int_val_other) int_val_other = TestTypes.TestClass(4) self.assertNotEqual(int_val, int_val_other) def test_float_type(self): float_val = TestTypes.TestClass(3.2) self.assertEqual(self._test_list[float_val], 3) self.assertEqual(hash(float_val), hash(3.2)) float_val_other = TestTypes.TestClass(3.2) self.assertEqual(float_val, float_val_other) float_val_other = TestTypes.TestClass(4.3) self.assertNotEqual(float_val, float_val_other) def test_tuple_int_type(self): tuple_int_val = TestTypes.TestClass((1, 2, 3)) with self.assertRaises(ValueError): self._test_list[tuple_int_val] = 1 self.assertEqual(hash(tuple_int_val), hash((1, 2, 3))) tuple_int_val_other = TestTypes.TestClass((1, 2, 3)) self.assertEqual(tuple_int_val, tuple_int_val_other) tuple_int_val_other = TestTypes.TestClass((2, 3, 1)) self.assertNotEqual(tuple_int_val, tuple_int_val_other) def test_tuple_float_type(self): tuple_float_val = TestTypes.TestClass((1.1, 2.2, 3.3)) with self.assertRaises(ValueError): self._test_list[tuple_float_val] = 1 self.assertEqual(hash(tuple_float_val), hash((1.1, 2.2, 3.3))) tuple_float_val_other = TestTypes.TestClass((1.1, 2.2, 3.3)) self.assertEqual(tuple_float_val, tuple_float_val_other) tuple_float_val_other = TestTypes.TestClass((2.2, 3.3, 1.1)) self.assertNotEqual(tuple_float_val, tuple_float_val_other) def test_ndarray_type(self): ndarray_val = TestTypes.TestClass(np.array(3)) self.assertEqual(self._test_list[ndarray_val], 3) self.assertEqual(hash(ndarray_val), hash((3,))) ndarray_val_other = TestTypes.TestClass(np.array(3)) self.assertEqual(ndarray_val, ndarray_val_other) int_val_other = TestTypes.TestClass(3) self.assertEqual(ndarray_val, int_val_other) ndarray_val_other = TestTypes.TestClass(np.array(4)) self.assertNotEqual(ndarray_val, ndarray_val_other) ndarray_val = TestTypes.TestClass(np.array(((1, 2), (3, 4)))) with self.assertRaises(ValueError): self._test_list[ndarray_val] = 1 self.assertEqual(hash(ndarray_val), hash((1, 2, 3, 4))) ndarray_val_other = TestTypes.TestClass(((1, 2), (3, 4))) self.assertEqual(ndarray_val, ndarray_val_other) ndarray_val_other = TestTypes.TestClass(np.ndarray((1, 2, 3, 4))) self.assertNotEqual(ndarray_val, ndarray_val_other) def test_tensor_type(self): tensor_val = TestTypes.TestClass(torch.tensor(3)) self.assertEqual(self._test_list[tensor_val], 3) self.assertEqual(hash(tensor_val), hash((3,))) tensor_val_other = TestTypes.TestClass(torch.tensor(3)) self.assertEqual(tensor_val, tensor_val_other) int_val_other = TestTypes.TestClass(3) with self.assertRaises(TypeError): _ = tensor_val == int_val_other tensor_val_other = TestTypes.TestClass(torch.tensor(4)) self.assertNotEqual(tensor_val, tensor_val_other) tensor_val = TestTypes.TestClass(torch.tensor(((1, 2), (3, 4)))) with self.assertRaises(ValueError): self._test_list[tensor_val] = 1 self.assertEqual(hash(tensor_val), hash((1, 2, 3, 4))) tensor_val_other = TestTypes.TestClass(torch.tensor((1, 2, 3, 4))) self.assertNotEqual(tensor_val, tensor_val_other) class TestValues(unittest.TestCase): TestIntType = TypeWrapper[int] TestTupleFloatType = TypeWrapper[Tuple[float]] class TestIntKeyValues(Values[TestIntType]): def _new_key(self, k: int): return TestValues.TestIntType(k) class TestTupleFloatKeyValues(Values[TestTupleFloatType]): def _new_key(self, k: int): raise TypeError( f"value {k} invalid for " f"{TestValues.TestTupleFloatType.__name__}" ) def setUp(self) -> None: self._int_float_values = TestValues.TestIntKeyValues([2.2, 4.4, 1.1, 3.3]) self._tuple_float_float_values = TestValues.TestTupleFloatKeyValues( { TestValues.TestTupleFloatType((1.0, 2.0)): 2.2, TestValues.TestTupleFloatType((3.0, 4.0)): 4.4, TestValues.TestTupleFloatType((5.0, 6.0)): 1.1, TestValues.TestTupleFloatType((7.0, 8.0)): 3.3, } ) self._int_array_values = TestValues.TestIntKeyValues( np.array((2.2, 4.4, 1.1, 3.3)) ) self._int_tensor_values = TestValues.TestIntKeyValues( torch.tensor((2.2, 4.4, 1.1, 3.3)) ) def test_indexing(self): self.assertEqual(self._int_float_values[2], 1.1) self.assertEqual(self._int_float_values[TestValues.TestIntType(2)], 1.1) self.assertEqual( self._tuple_float_float_values[TestValues.TestTupleFloatType((3.0, 4.0))], 4.4, ) def test_sort(self): keys, values = self._int_float_values.sort() self.assertEqual( keys, [ TestValues.TestIntType(1), TestValues.TestIntType(3), TestValues.TestIntType(0), TestValues.TestIntType(2), ], ) self.assertEqual(values, [4.4, 3.3, 2.2, 1.1]) keys, values = self._tuple_float_float_values.sort() self.assertEqual( keys, [ TestValues.TestTupleFloatType((3.0, 4.0)), TestValues.TestTupleFloatType((7.0, 8.0)), TestValues.TestTupleFloatType((1.0, 2.0)), TestValues.TestTupleFloatType((5.0, 6.0)), ], ) self.assertEqual(values, [4.4, 3.3, 2.2, 1.1]) keys, values = self._int_array_values.sort() self.assertEqual( keys, [ TestValues.TestIntType(1), TestValues.TestIntType(3), TestValues.TestIntType(0), TestValues.TestIntType(2), ], ) self.assertTrue(np.array_equal(values, np.array([4.4, 3.3, 2.2, 1.1]))) keys, values = self._int_tensor_values.sort() self.assertEqual( keys, [ TestValues.TestIntType(1), TestValues.TestIntType(3), TestValues.TestIntType(0), TestValues.TestIntType(2), ], ) self.assertTrue(torch.equal(values, torch.tensor([4.4, 3.3, 2.2, 1.1]))) def test_unzip(self): items = self._int_float_values.items values = self._int_float_values.values self.assertEqual( items, [ TestValues.TestIntType(0), TestValues.TestIntType(1), TestValues.TestIntType(2), TestValues.TestIntType(3), ], ) self.assertEqual(values, [2.2, 4.4, 1.1, 3.3]) items = self._tuple_float_float_values.items values = self._tuple_float_float_values.values self.assertEqual( items, [ TestValues.TestTupleFloatType((1.0, 2.0)), TestValues.TestTupleFloatType((3.0, 4.0)), TestValues.TestTupleFloatType((5.0, 6.0)), TestValues.TestTupleFloatType((7.0, 8.0)), ], ) self.assertEqual(values, [2.2, 4.4, 1.1, 3.3]) items = self._int_array_values.items values = self._int_array_values.values self.assertEqual( items, [ TestValues.TestIntType(0), TestValues.TestIntType(1), TestValues.TestIntType(2), TestValues.TestIntType(3), ], ) self.assertTrue(np.array_equal(values, np.array([2.2, 4.4, 1.1, 3.3]))) items = self._int_tensor_values.items values = self._int_tensor_values.values self.assertEqual( items, [ TestValues.TestIntType(0), TestValues.TestIntType(1), TestValues.TestIntType(2), TestValues.TestIntType(3), ], ) self.assertTrue(torch.equal(values, torch.tensor([2.2, 4.4, 1.1, 3.3]))) def test_copy(self): copy = self._int_float_values.copy() for i, c in zip(self._int_float_values, copy): self.assertEqual(i, c) copy[1] = 2.1 self.assertNotEqual(copy[1], self._int_float_values[1]) copy = self._tuple_float_float_values.copy() for i, c in zip(self._tuple_float_float_values, copy): self.assertEqual(i, c) key = TestValues.TestTupleFloatType((3.0, 4.0)) copy[key] = 2.1 self.assertNotEqual(copy[key], self._tuple_float_float_values[key]) copy = self._int_array_values.copy() for i, c in zip(self._int_array_values, copy): self.assertEqual(i, c) copy[1] = 2.1 self.assertNotEqual(copy[1], self._int_array_values[1]) copy = self._int_tensor_values.copy() for i, c in zip(self._int_tensor_values, copy): self.assertEqual(i, c) copy[1] = 2.1 self.assertNotEqual(copy[1], self._int_tensor_values[1]) def test_conversion(self): float_list_val = [1.1, 2.2, 3.3] tensor_val = torch.tensor([1.1, 2.2, 3.3], dtype=torch.double) array_val = np.array([1.1, 2.2, 3.3], dtype=np.float64) self.assertTrue( torch.equal( Values.to_tensor(float_list_val, dtype=torch.double), tensor_val ) ) self.assertTrue( torch.equal(Values.to_tensor(tensor_val, dtype=torch.double), tensor_val) ) self.assertTrue( torch.equal(Values.to_tensor(array_val, dtype=torch.double), tensor_val) ) self.assertTrue(np.array_equal(Values.to_ndarray(float_list_val), array_val)) self.assertTrue( np.array_equal(Values.to_ndarray(tensor_val, dtype=np.float64), array_val) ) self.assertTrue(np.array_equal(Values.to_ndarray(array_val), array_val)) self.assertEqual(Values.to_sequence(float_list_val), float_list_val) self.assertEqual(Values.to_sequence(tensor_val), float_list_val) self.assertEqual(Values.to_sequence(array_val), float_list_val) class TestDistribution(unittest.TestCase): class TestIntKeyDistribution(Distribution[int]): def _new_key(self, k: int): return k def setUp(self) -> None: self._tensor_distribution = TestDistribution.TestIntKeyDistribution( torch.tensor([1.0, 2.0, 3.0, 4.0]) ) self._array_distribution = TestDistribution.TestIntKeyDistribution( np.array([1.0, 2.0, 3.0, 4.0]) ) self._list_distribution = TestDistribution.TestIntKeyDistribution( [1.0, 2.0, 3.0, 4.0] ) self._map_distribution = TestDistribution.TestIntKeyDistribution( {0: 1.0, 1: 2.0, 2: 3.0, 3: 4.0} ) def test_values(self): self.assertTrue( torch.equal( self._tensor_distribution.values, torch.tensor([0.1, 0.2, 0.3, 0.4]) ) ) self.assertTrue( np.array_equal( self._array_distribution.values, np.array([0.1, 0.2, 0.3, 0.4]) ) ) self.assertEqual(self._list_distribution.values, [0.1, 0.2, 0.3, 0.4]) self.assertTrue(self._map_distribution.values, [0.1, 0.2, 0.3, 0.4]) def _test_sample(self, distribution: Distribution): counts = [0] * 4 total = 100000 for _ in range(total): counts[distribution.sample()[0]] += 1 self.assertAlmostEqual(counts[0] / total, 0.1, places=2) self.assertAlmostEqual(counts[1] / total, 0.2, places=2) self.assertAlmostEqual(counts[2] / total, 0.3, places=2) self.assertAlmostEqual(counts[3] / total, 0.4, places=2) def test_sample(self): self._test_sample(self._tensor_distribution) self.assertEqual(self._tensor_distribution.greedy(4), [3, 2, 1, 0]) self._test_sample(self._array_distribution) self.assertEqual(self._array_distribution.greedy(4), [3, 2, 1, 0]) self._test_sample(self._list_distribution) self.assertEqual(self._list_distribution.greedy(4), [3, 2, 1, 0]) self._test_sample(self._map_distribution) self.assertEqual(self._map_distribution.greedy(4), [3, 2, 1, 0]) if __name__ == "__main__": np.random.seed(1234) torch.random.manual_seed(1234) unittest.main()
13,486
38.667647
86
py
ReAgent
ReAgent-master/reagent/ope/test/unit_tests/test_contextual_bandit_estimators.py
#!/usr/bin/env python3 import random import unittest import numpy as np import torch from reagent.ope.estimators.contextual_bandits_estimators import ( Action, ActionDistribution, ActionSpace, BanditsEstimatorInput, DMEstimator, DoublyRobustEstimator, IPSEstimator, LogSample, ModelOutputs, SwitchDREstimator, SwitchEstimator, ) class TestSwitchEstimators(unittest.TestCase): """ These unit tests verify basic properties of the Switch estimators, in that when the threshold is low, the model-based DM estimator is used and when the threshold is high, the propensity score estimator is used. """ NUM_ACTIONS = 2 DR_EPSILON = 0.05 def setUp(self) -> None: random.seed(0) torch.random.manual_seed(0) np.random.seed(0) self.action_space = ActionSpace(TestSwitchEstimators.NUM_ACTIONS) self.sample1 = LogSample( context=0, log_action=Action(0), log_reward=1.0, log_action_probabilities=ActionDistribution(torch.tensor([0.7, 0.3])), tgt_action_probabilities=ActionDistribution([0.6, 0.4]), tgt_action=Action(1), model_outputs=ModelOutputs(0.5, [0.4, 0.5]), ) self.sample2 = LogSample( context=0, log_action=Action(1), log_reward=0.0, log_action_probabilities=ActionDistribution([0.5, 0.5]), tgt_action_probabilities=ActionDistribution([0.7, 0.3]), tgt_action=Action(0), model_outputs=ModelOutputs(0.0, [0.0, 0.0]), ) self.bandit_input = BanditsEstimatorInput( self.action_space, [self.sample1, self.sample2], True ) SwitchEstimator.EXP_BASE = 1.5 SwitchEstimator.CANDIDATES = 21 def test_switch_equal_to_ips(self): """ Switch with tau set at the max value should be equal to IPS """ # Setting the base to 1 will cause all candidates to be the maximum threshold SwitchEstimator.EXP_BASE = 1 switch = SwitchEstimator(rmax=1.0).evaluate(self.bandit_input) ips = IPSEstimator().evaluate(self.bandit_input) self.assertAlmostEqual(ips.estimated_reward, switch.estimated_reward) def test_switch_dr_equal_to_dr(self): """ Switch-DR with tau set at the max value should be equal to DR """ # Setting the base to 1 will cause all candidates to be the maximum threshold SwitchEstimator.EXP_BASE = 1 switch = SwitchDREstimator(rmax=1.0).evaluate(self.bandit_input) dr = DoublyRobustEstimator().evaluate(self.bandit_input) self.assertAlmostEqual( dr.estimated_reward, switch.estimated_reward, delta=TestSwitchEstimators.DR_EPSILON, ) def test_switch_equal_to_dm(self): """ Switch with tau set at the min value should be equal to DM """ # Setting candidates to 0 will default to tau being the minimum threshold SwitchEstimator.CANDIDATES = 0 switch = SwitchEstimator(rmax=1.0).evaluate(self.bandit_input) dm = DMEstimator().evaluate(self.bandit_input) self.assertAlmostEqual(dm.estimated_reward, switch.estimated_reward) def test_switch_dr_equal_to_dm(self): """ Switch-DR with tau set at the min value should be equal to DM """ # Setting candidates to 0 will default to tau being the minimum threshold SwitchEstimator.CANDIDATES = 0 switch = SwitchDREstimator(rmax=1.0).evaluate(self.bandit_input) dm = DMEstimator().evaluate(self.bandit_input) self.assertAlmostEqual(dm.estimated_reward, switch.estimated_reward)
3,772
34.933333
85
py
ReAgent
ReAgent-master/reagent/ope/test/unit_tests/test_utils.py
#!/usr/bin/env python3 import unittest import numpy as np import torch from reagent.ope.utils import Clamper, RunningAverage class TestUtils(unittest.TestCase): def test_running_average(self): ra = RunningAverage() ra.add(1.0).add(2.0).add(3.0).add(4.0) self.assertEqual(ra.count, 4) self.assertEqual(ra.average, 2.5) self.assertEqual(ra.total, 10.0) def test_clamper(self): with self.assertRaises(ValueError): clamper = Clamper(1.0, 0.0) list_value = [-1.1, 0.9, 0.0, 1.1, -0.9] tensor_value = torch.tensor(list_value) array_value = np.array(list_value) clamper = Clamper() self.assertEqual(clamper(list_value), list_value) self.assertTrue(torch.equal(clamper(tensor_value), tensor_value)) self.assertTrue(np.array_equal(clamper(array_value), array_value)) clamper = Clamper(-1.0, 1.0) self.assertEqual(clamper(list_value), [-1.0, 0.9, 0.0, 1.0, -0.9]) self.assertTrue( torch.equal( clamper(tensor_value), torch.tensor([-1.0, 0.9, 0.0, 1.0, -0.9]) ) ) self.assertTrue( np.array_equal(clamper(array_value), np.array([-1.0, 0.9, 0.0, 1.0, -0.9])) ) if __name__ == "__main__": unittest.main()
1,326
30.595238
87
py
ReAgent
ReAgent-master/reagent/ope/datasets/logged_dataset.py
#!/usr/bin/env python3 from abc import ABC, abstractmethod from dataclasses import dataclass import torch class BanditsDataset(ABC): """ Base class for logged, aka behavior, dataset """ @abstractmethod def __len__(self) -> int: """ Returns: length of the dataset """ pass @abstractmethod def __getitem__(self, idx) -> dataclass: """ Args: idx: index of the sample Returns: tuple of features, action, and reward at idx """ pass @property @abstractmethod def num_features(self) -> int: """ Returns: number of features """ pass @property @abstractmethod def num_actions(self) -> int: """ Returns: number of total possible actions """ pass @property @abstractmethod def features(self) -> torch.Tensor: """ Returns: all features in the dataset as numpy array """ pass @property @abstractmethod def actions(self) -> torch.Tensor: """ Returns: all actions in the dataset as numpy array """ pass @property @abstractmethod def rewards(self) -> torch.Tensor: """ Returns: all rewards in the dataset as numpy array """ pass
1,441
17.727273
56
py
ReAgent
ReAgent-master/reagent/ope/trainers/linear_trainers.py
#!/usr/bin/env python3 import logging import math import time from typing import Optional import numpy as np import torch from reagent.ope.estimators.types import PredictResults, Trainer, TrainingData from sklearn.linear_model import Lasso, LogisticRegression, SGDClassifier from sklearn.metrics import accuracy_score, mean_squared_error from sklearn.tree import DecisionTreeClassifier, DecisionTreeRegressor from torch import Tensor class LinearTrainer(Trainer): def __init__(self, is_classifier: bool = False): super().__init__() self._is_classifier = is_classifier def predict(self, x: Tensor, device=None) -> PredictResults: if self._model is not None: if hasattr(self._model, "predict_proba"): proba = torch.as_tensor( self._model.predict_proba(x), dtype=torch.float, device=device ) score = (proba * torch.arange(proba.shape[1])).sum(dim=1) return PredictResults(torch.argmax(proba, 1), score, proba) elif hasattr(self._model, "predict"): return PredictResults( None, torch.as_tensor( self._model.predict(x), dtype=torch.float, device=device ), None, ) else: raise AttributeError("model doesn't have predict_proba or predict") else: raise Exception("model not trained") def _score(self, y_true: np.ndarray, y_pred: np.ndarray, weight=None) -> float: if self._is_classifier: return accuracy_score(y_true, y_pred, sample_weight=weight) else: return 1.0 / math.pow( 2, mean_squared_error(y_true, y_pred, sample_weight=weight), ) def score(self, x: Tensor, y: Tensor, weight: Optional[Tensor] = None) -> float: y_pred = self._model.predict(x) w = weight.numpy() if weight is not None else None return self._score(y.numpy(), y_pred, weight=w) class LassoTrainer(LinearTrainer): @property def name(self) -> str: return "lasso" def train(self, data: TrainingData, iterations: int = 1, num_samples: int = 0): logging.info("LassoTrainer.train...") self._model = None best_score = float("-inf") for _ in range(iterations): x, y, _ = super()._sample( data.train_x, data.train_y, data.train_weight, num_samples, True ) sx, sy, ssw = super()._sample( data.validation_x, data.validation_y, data.validation_weight ) for alpha in np.logspace(-4, 2, num=7, base=10): model = Lasso( alpha=alpha, fit_intercept=False, copy_X=True, max_iter=10000, warm_start=False, selection="random", ) model.fit(x, y) y_pred = model.predict(sx) score = self._score(sy, y_pred, weight=ssw) # score = model.score(sx, sy, ssw) logging.info(f" alpha: {alpha}, score: {score}") if score > best_score: best_score = score self._model = model class DecisionTreeTrainer(LinearTrainer): @property def name(self) -> str: return "decision_tree" def train(self, data: TrainingData, iterations: int = 1, num_samples: int = 0): logging.info("DecisionTreeTrainer.train...") self._model = None best_score = float("-inf") for _ in range(iterations): x, y, sw = super()._sample( data.train_x, data.train_y, data.train_weight, num_samples, True ) sx, sy, ssw = super()._sample( data.validation_x, data.validation_y, data.validation_weight ) if self._model is None: self._model = DecisionTreeRegressor( criterion="mse", splitter="random", max_depth=None, min_samples_split=4, min_samples_leaf=4, ) self._model.fit(x, y, sw) y_pred = self._model.predict(sx) best_score = self._score(sy, y_pred, weight=ssw) logging.info(f" max_depth: None, score: {best_score}") for depth in range(3, 21, 3): model = DecisionTreeRegressor( criterion="mse", splitter="random", max_depth=depth, min_samples_split=4, min_samples_leaf=4, ) model.fit(x, y, sw) y_pred = model.predict(sx) score = self._score(sy, y_pred, weight=ssw) # score = model.score(sx, sy, ssw) logging.info(f" max_depth: {depth}, score: {score}") if score > best_score: best_score = score self._model = model class DecisionTreeClassifierTrainer(LinearTrainer): def __init__(self): super().__init__(True) @property def name(self) -> str: return "decision_tree_classifier" def train(self, data: TrainingData, iterations: int = 1, num_samples: int = 0): logging.info("DecisionTreeClassifierTrainer.train...") self._model = None best_score = float("-inf") for _ in range(iterations): x, y, sw = super()._sample( data.train_x, data.train_y, data.train_weight, num_samples, True ) sx, sy, ssw = super()._sample( data.validation_x, data.validation_y, data.validation_weight ) for depth in range(3, 21, 3): model = DecisionTreeClassifier( criterion="entropy", splitter="random", max_depth=depth, min_samples_split=4, min_samples_leaf=4, ) model.fit(x, y, sw) score = model.score(sx, sy, ssw) logging.info(f" max_depth: {depth}, score: {score}") if score > best_score: best_score = score self._model = model class LogisticRegressionTrainer(LinearTrainer): def __init__(self, solver: str = "lbfgs"): super().__init__(True) self._solver = solver @property def name(self) -> str: return "logistic_regression" def train(self, data: TrainingData, iterations: int = 1, num_samples: int = 0): logging.info("LogisticRegressionTrainer.train...") self._model = None best_score = float("-inf") for _ in range(iterations): x, y, sw = super()._sample( data.train_x, data.train_y, data.train_weight, num_samples, True ) sx, sy, ssw = super()._sample( data.validation_x, data.validation_y, data.validation_weight ) for c in np.logspace(-5, 4, num=10, base=10): model = LogisticRegression( C=c, fit_intercept=False, solver=self._solver, max_iter=1000, multi_class="auto", n_jobs=-1, ) model.fit(x, y, sw) score = model.score(sx, sy, ssw) logging.info(f" C: {c}, score: {score}") if score > best_score: best_score = score self._model = model class SGDClassifierTrainer(LinearTrainer): def __init__(self, loss: str = "log", max_iter: int = 1000): super().__init__(True) self._loss = loss self._max_iter = max_iter @property def name(self) -> str: return "sgd_classifier" def train(self, data: TrainingData, iterations: int = 1, num_samples: int = 0): logging.info("SGDClassifierTrainer.train...") self._model = None best_score = float("-inf") for _ in range(iterations): x, y, _ = super()._sample( data.train_x, data.train_y, data.train_weight, num_samples, True ) sx, sy, ssw = super()._sample( data.validation_x, data.validation_y, data.validation_weight ) for alpha in np.logspace(-8, -1, num=8, base=10): model = SGDClassifier( loss=self._loss, alpha=alpha, random_state=0, max_iter=self._max_iter, ) model.fit(x, y) score = model.score(sx, sy, ssw) logging.info(f" alpha: {alpha}, score: {score}") if score > best_score: best_score = score self._model = model class LinearNet(torch.nn.Module): def __init__( self, D_in: int, H: int, D_out: int, hidden_layers: int = 2, activation=torch.nn.ReLU, ): super(LinearNet, self).__init__() self._hidden_dim = H self._hidden_layers = hidden_layers self._activation = activation self._out_dim = D_out self.layers = [] dim = D_in for _ in range(self._hidden_layers): self.layers.append(torch.nn.Linear(dim, self._hidden_dim)) self.layers.append(self._activation()) dim = self._hidden_dim self.layers.append(torch.nn.Linear(dim, self._out_dim)) self.model = torch.nn.Sequential(*self.layers) def forward(self, x: torch.Tensor): x = x.requires_grad_(True) return self.model(x) class NNTrainer(Trainer): def __init__(self, device=None): super().__init__() self._device = device self._loss_fn: Optional[torch.nn.MSELoss] = None @property def name(self) -> str: return "linear_net" def train( self, data: TrainingData, iterations: int = 100, epochs: int = 1, num_samples: int = 0, ): d_in, d_out = ( data.train_x.shape[1], data.train_y.shape[1] if len(data.train_y.shape) > 1 else 1, ) if d_in == 0 or d_out == 0: return None h = 500 n = data.train_x.shape[0] // 200 logging.info(f"start training...") logging.info(f" d_in = {d_in}, h = {h}, d_out = {d_out}, n = {n}") st = time.process_time() self._model = LinearNet(d_in, h, d_out) if self._device is not None and self._device.type == "cuda": self._model = self._model.cuda() self._loss_fn = torch.nn.MSELoss(reduction="mean") learning_rate = 1e-3 optimizer = torch.optim.Adam(self._model.parameters(), lr=learning_rate) scheduler = torch.optim.lr_scheduler.ReduceLROnPlateau( optimizer, "min", patience=5, verbose=True, threshold=1e-5 ) for _ in range(epochs): for t in range(iterations): x, y, _ = super()._sample( data.train_x, data.train_y, data.train_weight, num_samples, True ) x = torch.as_tensor(x, device=self._device) y = torch.as_tensor(y, device=self._device) if len(y.shape) == 1: y = y.reshape(-1, 1) y_pred = self._model(x) # pyre-fixme[29]: `Optional[torch.nn.MSELoss]` is not a function. loss = self._loss_fn(y_pred, y) if (t + 1) % 10 == 0: scheduler.step(loss.item()) logging.info(f" step [{t + 1}]: loss={loss.item()}") optimizer.zero_grad() loss.backward() optimizer.step() logging.info(f" training time {time.process_time() - st}") def predict(self, x: Tensor, device=None) -> PredictResults: if self._model is not None: self._model.eval() proba = torch.as_tensor(self._model(x), dtype=torch.float, device=device) return PredictResults(torch.argmax(proba, 1), proba) else: raise Exception("mode not trained") def score(self, x: Tensor, y: Tensor, weight: Optional[Tensor] = None) -> float: if self._loss_fn is not None: return self._loss_fn(y, x).item() else: raise Exception("mode not trained")
12,799
35.571429
85
py
ReAgent
ReAgent-master/reagent/ope/trainers/rl_tabular_trainers.py
#!/usr/bin/env python3 import pickle from functools import reduce from typing import List, Mapping, Sequence import torch from reagent.ope.estimators.sequential_estimators import ( Model, RLPolicy, State, ValueFunction, ) from reagent.ope.estimators.types import Action, ActionDistribution, ActionSpace from reagent.ope.test.envs import Environment, PolicyLogGenerator from reagent.ope.utils import RunningAverage class TabularPolicy(RLPolicy): def __init__(self, action_space: ActionSpace, epsilon: float = 0.0, device=None): super().__init__(action_space, device) self._epsilon = epsilon as_size = len(action_space) self._exploitation_prob = 1.0 - epsilon self._exploration_prob = epsilon / len(action_space) self._uniform_probs: List[float] = as_size * [1.0 / as_size] self._state_space = {} def update(self, state: State, actions: Sequence[float]) -> float: old_dist = self._uniform_probs if state in self._state_space: old_dist = self._state_space[state] self._state_space[state] = actions return ( reduce( lambda a, b: a + b, map(lambda a: (a[0] - a[1]) ** 2, zip(old_dist, actions)), ) ** 0.5 ) def action_dist(self, state: State) -> ActionDistribution: if state in self._state_space: actions = self._state_space[state] probs = list( map( lambda p: p * self._exploitation_prob + self._exploration_prob, actions, ) ) else: probs = self._uniform_probs return self._action_space.distribution(probs) def save(self, path) -> bool: try: with open(path, "wb") as f: pickle.dump((self._action_space, self._epsilon, self._state_space), f) except Exception: return False else: return True def load(self, path) -> bool: try: with open(path, "rb") as f: self._action_space, self._epsilon, self._state_space = pickle.load(f) except Exception: return False else: return True class TabularValueFunction(ValueFunction): def __init__(self, policy: RLPolicy, model: Model, gamma=0.99): self._policy = policy self._model = model self._gamma = gamma self._state_values = {} def _state_value(self, state: State) -> float: return ( 0.0 if (state is None or state.is_terminal) else (0.0 if state not in self._state_values else self._state_values[state]) ) def state_action_value(self, state: State, action: Action) -> float: value = 0.0 sr_dist = self._model.next_state_reward_dist(state, action) for _, rp in sr_dist.items(): value += rp.prob * (rp.reward + self._gamma * self.state_value(state)) return value def state_value(self, state: State) -> float: pass def reset(self, clear_state_values: bool = False): pass class EstimatedStateValueFunction(ValueFunction): def __init__( self, policy: RLPolicy, env: Environment, gamma: float, num_episodes: int = 100 ): self._policy = policy self._env = env self._gamma = gamma self._num_episodes = num_episodes self._state_values = {} self._estimate_value() def _estimate_value(self): tgt_generator = PolicyLogGenerator(self._env, self._policy) log = {} for state in self._env.states: mdps = [] for _ in range(self._num_episodes): mdps.append(tgt_generator.generate_log(state)) log[state] = mdps for state, mdps in log.items(): avg = RunningAverage() for mdp in mdps: discount = 1.0 r = 0.0 for t in mdp: r += discount * t.reward discount *= self._gamma avg.add(r) self._state_values[state] = avg.average def state_action_value(self, state: State, action: Action) -> float: return 0.0 def state_value(self, state: State) -> float: return self._state_values[state] def reset(self): self._state_values = {} class DPValueFunction(TabularValueFunction): def __init__( self, policy: RLPolicy, env: Environment, gamma: float = 0.99, threshold: float = 0.0001, ): super().__init__(policy, env, gamma) self._env = env self._threshold = threshold self._evaluated = False def state_value(self, state: State, horizon: int = -1) -> float: if not self._evaluated: self._evaluate() return self._state_value(state) def reset(self, clear_state_values: bool = False): self._evaluated = False if clear_state_values: self._state_values.clear() def _evaluate(self): delta = float("inf") while delta >= self._threshold: delta = 0.0 for state in self._env.states: old_value = self._state_value(state) new_value = 0.0 a_dist = self._policy(state) for a, ap in a_dist: s_dist = self._model(state, a) a_value = 0.0 for s, rp in s_dist.items(): a_value += rp.prob * ( rp.reward + self._gamma * self._state_value(s) ) new_value += ap * a_value delta = max(delta, abs(old_value - new_value)) self._state_values[state] = new_value self._evaluated = True class DPTrainer(object): def __init__(self, env: Environment, policy: TabularPolicy): self._env = env self._policy = policy @staticmethod def _state_value(state: State, state_values: Mapping[State, float]) -> float: return 0.0 if state not in state_values else state_values[state] def train(self, gamma: float = 0.9, threshold: float = 0.0001): stable = False valfunc = DPValueFunction(self._policy, self._env, gamma, threshold) while not stable: stable = True for state in self._env.states: new_actions = [] max_value = float("-inf") for action in self._policy.action_space: s_dist = self._env(state, action) value = 0.0 for s, rp in s_dist.items(): value += rp.prob * rp.reward if s is not None: value += rp.prob * gamma * valfunc(s) if value > max_value: max_value = value new_actions = [action] elif value == max_value: new_actions.append(action) prob = 1.0 / len(new_actions) actions = [0.0] * len(self._policy.action_space) for a in new_actions: actions[a.value] = prob if self._policy.update(state, actions) >= 1.0e-6: stable = False valfunc.reset() return valfunc class MonteCarloValueFunction(TabularValueFunction): def __init__( self, policy: RLPolicy, env: Environment, gamma: float = 0.99, first_visit: bool = True, count_threshold: int = 100, max_iteration: int = 200, ): super().__init__(policy, env, gamma) self._env = env self._first_visit = first_visit self._count_threshold = count_threshold self._max_iteration = max_iteration self._log_generator = PolicyLogGenerator(env, policy) self._state_counts = {} def _state_value(self, state: State): i = 0 state_count = self._state_counts[state] if state in self._state_counts else 0 while state_count < self._count_threshold and i < self._max_iteration: i += 1 mdp = self._log_generator.generate_log(state) if self._first_visit: state_counts = {} for t in mdp: if t.last_state is None: continue if t.last_state in state_counts: state_counts[t.last_state] += 1 else: state_counts[t.last_state] = 1 g = 0 for t in reversed(mdp): if t.last_state is None: continue g = self._gamma * g + t.reward counts = state_counts[t.last_state] if counts > 1: self._update_state_value(t.last_state, g) counts -= 1 if counts == 0: del state_counts[t.last_state] else: state_counts[t.last_state] = counts else: g = 0 for t in reversed(mdp): if t.last_state is None: continue g = self._gamma * g + t.reward self._update_state_value(t.last_state, g) state_count = ( self._state_counts[state] if state in self._state_counts else 0 ) return super()._state_value(state) def _update_state_value(self, state: State, g: float): sv = super()._state_value(state) sc = self._state_counts[state] if state in self._state_counts else 0 sc += 1 sv = sv + (g - sv) / sc self._state_values[state] = sv self._state_counts[state] = sc def state_value(self, state: State) -> float: return self._state_value(state) def reset(self, clear_state_values: bool = False): if clear_state_values: self._state_values.clear() self._state_counts.clear() class MonteCarloTrainer(object): def __init__(self, env: Environment, policy: TabularPolicy): self._env = env self._policy = policy self._log_generator = PolicyLogGenerator(env, policy) def train( self, iterations: int, gamma: float = 0.9, first_visit: bool = True, update_interval: int = 20, ): i = 0 value_counts = {} while i < iterations: i += 1 for state in self._env.states: mdp = self._log_generator.generate_log(state) if first_visit: vcounts = {} for t in mdp: if t.last_state is None or t.action is None: continue key = (t.last_state, t.action) if key in vcounts: vcounts[key] += 1 else: vcounts[key] = 1 g = 0 for t in reversed(mdp): if t.last_state is None or t.action is None: continue g = gamma * g + t.reward key = (t.last_state, t.action) vc = vcounts[key] if vc > 1: self._update_state_value( value_counts, t.last_state, t.action, g ) vc -= 1 if vc == 0: del vcounts[key] else: vcounts[key] = vc else: g = 0 for t in reversed(mdp): if t.last_state is None or t.action is None: continue g = gamma * g + t.reward self._update_state_value( value_counts, t.last_state, t.action, g ) if i % update_interval == 0 and self._update_policy(value_counts): break def _update_state_value(self, value_counts, state, action, g: float): key = (state, action) sv, sc = value_counts[key] if key in value_counts else (0.0, 0) sc += 1 sv = sv + (g - sv) / sc value_counts[key] = (sv, sc) def _update_policy(self, value_counts) -> bool: stable = True for state in self._env.states: probs = [] for a in self._policy.action_space: key = (state, a) if key not in value_counts: probs.append(0.0) else: v, c = value_counts[key] probs.append(v * c) probs = torch.nn.functional.softmax(torch.tensor(probs), dim=0).tolist() if self._policy.update(state, probs) >= 1.0e-6: stable = False return stable
13,359
34.157895
88
py
ReAgent
ReAgent-master/reagent/mab/ucb.py
import math from abc import ABC, abstractmethod from typing import Union, Optional, List import torch from torch import Tensor def _get_arm_indices( ids_of_all_arms: List[Union[str, int]], ids_of_arms_in_batch: List[Union[str, int]] ) -> List[int]: arm_idxs = [] for i in ids_of_arms_in_batch: try: arm_idxs.append(ids_of_all_arms.index(i)) except ValueError: raise ValueError(f"Unknown arm_id {i}. Known arm ids: {ids_of_all_arms}") return arm_idxs def _place_values_at_indices(values: Tensor, idxs: List[int], total_len: int) -> Tensor: """ TODO: maybe replace with sparse vector function? Args: values (Tensor): The values idxs (List[int]): The indices at which the values have to be placed total_len (int): Length of the array """ assert len(values) == len(idxs) ret = torch.zeros(total_len) ret[idxs] = values return ret class BaseUCB(torch.nn.Module, ABC): """ Base class for UCB-like Multi-Armed Bandits (MAB) """ def __init__( self, *, n_arms: Optional[int] = None, arm_ids: Optional[List[Union[str, int]]] = None, ): super().__init__() if n_arms is not None: self.arm_ids = list(range(n_arms)) self.n_arms = n_arms if arm_ids is not None: self.arm_ids = arm_ids self.n_arms = len(arm_ids) self.total_n_obs_all_arms = 0 self.total_n_obs_per_arm = torch.zeros(self.n_arms) self.total_sum_reward_per_arm = torch.zeros(self.n_arms) def add_batch_observations( self, n_obs_per_arm: Tensor, sum_reward_per_arm: Tensor, arm_ids: Optional[List[Union[str, int]]] = None, ): if not isinstance(n_obs_per_arm, Tensor): n_obs_per_arm = torch.tensor(n_obs_per_arm, dtype=torch.float) if not isinstance(sum_reward_per_arm, Tensor): sum_reward_per_arm = torch.tensor(sum_reward_per_arm, dtype=torch.float) if arm_ids is None or arm_ids == self.arm_ids: # assume that the observations are for all arms in the default order arm_ids = self.arm_ids arm_idxs = list(range(self.n_arms)) else: assert len(arm_ids) == len( set(arm_ids) ) # make sure no duplicates in arm IDs # get the indices of the arms arm_idxs = _get_arm_indices(self.arm_ids, arm_ids) # put elements from the batch in the positions specified by `arm_ids` (missing arms will be zero) n_obs_per_arm = _place_values_at_indices( n_obs_per_arm, arm_idxs, self.n_arms ) sum_reward_per_arm = _place_values_at_indices( sum_reward_per_arm, arm_idxs, self.n_arms ) self.total_n_obs_per_arm += n_obs_per_arm self.total_sum_reward_per_arm += sum_reward_per_arm self.total_n_obs_all_arms += int(n_obs_per_arm.sum()) def add_single_observation(self, arm_id: int, reward: float): assert arm_id in self.arm_ids arm_idx = self.arm_ids.index(arm_id) self.total_n_obs_per_arm[arm_idx] += 1 self.total_sum_reward_per_arm[arm_idx] += reward self.total_n_obs_all_arms += 1 def get_avg_reward_values(self) -> Tensor: return self.total_sum_reward_per_arm / self.total_n_obs_per_arm def get_action(self) -> Union[str, int]: """ Get the id of the action chosen by the UCB algorithm Returns: int: The integer ID of the chosen action """ ucb_scores = self.get_ucb_scores() return self.arm_ids[torch.argmax(ucb_scores)] @classmethod def get_ucb_scores_from_batch( cls, n_obs_per_arm: Tensor, sum_reward_per_arm: Tensor, *args, **kwargs, ) -> Tensor: """ A utility method used to create the bandit, feed in a batch of observations and get the UCB scores in one function call Args: n_obs_per_arm (Tensor): An array of counts of per-arm numbers of observations sum_reward_per_arm (Tensor): An array of sums of rewards for each arm (additional arguments can be provided for specific concrete class implementations) Returns: Tensor: Array of per-arm UCB scores """ n_arms = len(n_obs_per_arm) b = cls(n_arms=n_arms) b.add_batch_observations(n_obs_per_arm, sum_reward_per_arm, *args, **kwargs) return b.get_ucb_scores() @abstractmethod def get_ucb_scores(self): pass def __repr__(self): t = ", ".join( f"{v:.3f} ({int(n)})" for v, n in zip(self.get_avg_reward_values(), self.total_n_obs_per_arm) ) return f"UCB({self.n_arms} arms; {t}" def forward(self): return self.get_ucb_scores() class UCB1(BaseUCB): """ Canonical implementation of UCB1 Reference: https://www.cs.bham.ac.uk/internal/courses/robotics/lectures/ucb1.pdf """ def get_ucb_scores(self): """ Get per-arm UCB scores. The formula is UCB_i = AVG([rewards_i]) + SQRT(2*LN(T)/N_i) Returns: Tensor: An array of UCB scores (one per arm) """ avg_rewards = self.get_avg_reward_values() log_t_over_ni = ( math.log(self.total_n_obs_all_arms + 1) / self.total_n_obs_per_arm ) ucb = avg_rewards + torch.sqrt(2 * log_t_over_ni) return torch.where( self.total_n_obs_per_arm > 0, ucb, torch.tensor(torch.inf, dtype=torch.float), ) class UCBTuned(BaseUCB): """ Implementation of the UCB-Tuned algorithm from Section 4 of https://link.springer.com/content/pdf/10.1023/A:1013689704352.pdf Biggest difference from basic UCB is that per-arm reward variance is estimated. """ # _fields_for_saving = BaseUCB._fields_for_saving + [ # "total_sum_reward_squared_per_arm" # ] def __init__( self, n_arms: Optional[int] = None, arm_ids: Optional[List[Union[str, int]]] = None, ): super(UCBTuned, self).__init__(n_arms=n_arms, arm_ids=arm_ids) self.total_sum_reward_squared_per_arm = torch.zeros(self.n_arms) def add_batch_observations( self, n_obs_per_arm: Tensor, sum_reward_per_arm: Tensor, sum_reward_squared_per_arm: Tensor, arm_ids: Optional[List[Union[str, int]]] = None, ): """ Add information about arm rewards in a batched form. Args: n_obs_per_arm (Tensor): An array of counts of per-arm numbers of observations sum_reward_per_arm (Tensor): An array of sums of rewards for each arm sum_reward_squared_per_arm (Tensor): An array of sums of squares of rewards for each arm arm_ids (Optional[List[Union[str, int]]]): A list of ids of arms in the same order as the elements of previous arrays """ assert len(sum_reward_per_arm) == len(sum_reward_squared_per_arm) super().add_batch_observations( n_obs_per_arm, sum_reward_per_arm, arm_ids=arm_ids ) if not isinstance(sum_reward_per_arm, Tensor): sum_reward_squared_per_arm = torch.tensor( sum_reward_squared_per_arm, dtype=torch.float ) if arm_ids is None or arm_ids == self.arm_ids: # assume that the observations are for all arms in the default order arm_ids = self.arm_ids arm_idxs = list(range(self.n_arms)) else: assert len(arm_ids) == len( set(arm_ids) ) # make sure no duplicates in arm IDs # get the indices of the arms arm_idxs = _get_arm_indices(self.arm_ids, arm_ids) # put elements from the batch in the positions specified by `arm_ids` (missing arms will be zero) sum_reward_squared_per_arm = _place_values_at_indices( sum_reward_squared_per_arm, arm_idxs, self.n_arms ) self.total_sum_reward_squared_per_arm += sum_reward_squared_per_arm def add_single_observation(self, arm_id: int, reward: float): """ Add a single observation (arm played, reward) to the bandit Args: arm_id (int): Which arm was played reward (float): Reward renerated by the arm """ super().add_single_observation(arm_id, reward) arm_idx = self.arm_ids.index(arm_id) self.total_sum_reward_squared_per_arm[arm_idx] += reward ** 2 def get_ucb_scores(self) -> Tensor: """ Get per-arm UCB scores. The formula is UCB_i = AVG([rewards_i]) + SQRT(LN(T)/N_i * V_i) where V_i is a conservative variance estimate of arm i: V_i = AVG([rewards_i**2]) - AVG([rewards_i])**2 + sqrt(2ln(t) / n_i) Nore that we don't apply the min(1/4, ...) operator to the variance because this bandit is meant for non-Bernoulli applications as well Returns: Tensor: An array of UCB scores (one per arm) """ avg_rewards = self.get_avg_reward_values() log_t_over_ni = ( math.log(self.total_n_obs_all_arms + 1) / self.total_n_obs_per_arm ) per_arm_var_est = ( self.total_sum_reward_squared_per_arm / self.total_n_obs_per_arm - avg_rewards ** 2 + torch.sqrt( 2 * log_t_over_ni ) # additional term to make the estimate conservative (unlikely to underestimate) ) ucb = avg_rewards + torch.sqrt(log_t_over_ni * per_arm_var_est) return torch.where( self.total_n_obs_per_arm > 0, ucb, torch.tensor(torch.inf, dtype=torch.float), ) class UCBTunedBernoulli(UCBTuned): def add_batch_observations( self, n_obs_per_arm: Tensor, num_success_per_arm: Tensor, arm_ids: Optional[List[Union[str, int]]] = None, ): """ Add a batch of observations to the UCBTuned bandit, assuming Bernoulli distribution of rewards. Because of the Bernoulli assumption, we don't need to provide the squared rewards separately Args: n_obs_per_arm (Tensor): An array of counts of per-arm numbers of observations num_success_per_arm (Tensor): An array of counts of per-arm numbers of successes """ super().add_batch_observations( n_obs_per_arm, num_success_per_arm, num_success_per_arm, arm_ids=arm_ids ) class MetricUCB(BaseUCB): """ This is an improvement over UCB1 which uses a more precise confidence radius, especially for small expected rewards. Reference: https://arxiv.org/pdf/0809.4882.pdf """ def get_ucb_scores(self): """ Get per-arm UCB scores. The formula is UCB_i = AVG([rewards_i]) + SQRT(AVG([rewards_i]) * LN(T+1)/N_i) + LN(T+1)/N_i Returns: Tensor: An array of UCB scores (one per arm) """ avg_rewards = self.get_avg_reward_values() log_t_over_ni = ( math.log(self.total_n_obs_all_arms + 1) / self.total_n_obs_per_arm ) ucb = avg_rewards + torch.sqrt(avg_rewards * log_t_over_ni) + log_t_over_ni return torch.where( self.total_n_obs_per_arm > 0, ucb, torch.tensor(torch.inf, dtype=torch.float), ) def get_bernoulli_tuned_ucb_scores(n_obs_per_arm, num_success_per_arm): # a minimalistic function that implements Tuned UCB for Bernoulli bandit avg_rewards = n_obs_per_arm / num_success_per_arm log_t_over_ni = torch.log(torch.sum(n_obs_per_arm)) / num_success_per_arm per_arm_var_est = ( avg_rewards - avg_rewards ** 2 + torch.sqrt( 2 * log_t_over_ni ) # additional term to make the estimate conservative (unlikely to underestimate) ) return avg_rewards + torch.sqrt(log_t_over_ni * per_arm_var_est)
12,188
34.85
143
py
ReAgent
ReAgent-master/reagent/gym/utils.py
#!/usr/bin/env python3 # Copyright (c) Facebook, Inc. and its affiliates. All rights reserved. import logging import random from typing import Dict, List, Optional import gym import numpy as np import pandas as pd import torch # @manual import torch.nn.functional as F from gym import spaces from reagent.core.parameters import NormalizationData, NormalizationKey, ProblemDomain from reagent.gym.agents.agent import Agent from reagent.gym.agents.post_step import add_replay_buffer_post_step from reagent.gym.envs import EnvWrapper from reagent.gym.normalizers import ( only_continuous_normalizer, discrete_action_normalizer, only_continuous_action_normalizer, ) from reagent.gym.policies.random_policies import make_random_policy_for_env from reagent.gym.runners.gymrunner import run_episode from reagent.replay_memory import ReplayBuffer from tqdm import tqdm logger = logging.getLogger(__name__) SEED = 0 try: from reagent.gym.envs import RecSim # noqa HAS_RECSIM = True except ImportError: HAS_RECSIM = False def fill_replay_buffer( env, replay_buffer: ReplayBuffer, desired_size: int, agent: Agent ): """Fill replay buffer with transitions until size reaches desired_size.""" assert ( 0 < desired_size and desired_size <= replay_buffer._replay_capacity ), f"It's not true that 0 < {desired_size} <= {replay_buffer._replay_capacity}." assert replay_buffer.size < desired_size, ( f"Replay buffer already has {replay_buffer.size} elements. " f"(more than desired_size = {desired_size})" ) logger.info( f" Starting to fill replay buffer using policy to size: {desired_size}." ) post_step = add_replay_buffer_post_step(replay_buffer, env=env) agent.post_transition_callback = post_step max_episode_steps = env.max_steps with tqdm( total=desired_size - replay_buffer.size, desc=f"Filling replay buffer from {replay_buffer.size} to size {desired_size}", ) as pbar: mdp_id = 0 while replay_buffer.size < desired_size: last_size = replay_buffer.size max_steps = desired_size - replay_buffer.size if max_episode_steps is not None: max_steps = min(max_episode_steps, max_steps) run_episode(env=env, agent=agent, mdp_id=mdp_id, max_steps=max_steps) size_delta = replay_buffer.size - last_size # The assertion below is commented out because it can't # support input samples which has seq_len>1. This should be # treated as a bug, and need to be fixed in the future. # assert ( # size_delta >= 0 # ), f"size delta is {size_delta} which should be non-negative." pbar.update(n=size_delta) mdp_id += 1 if size_delta <= 0: # replay buffer size isn't increasing... so stop early break if replay_buffer.size >= desired_size: logger.info(f"Successfully filled replay buffer to size: {replay_buffer.size}!") else: logger.info( f"Stopped early and filled replay buffer to size: {replay_buffer.size}." ) def build_state_normalizer(env: EnvWrapper): if isinstance(env.observation_space, spaces.Box): assert ( len(env.observation_space.shape) == 1 ), f"{env.observation_space.shape} has dim > 1, and is not supported." return only_continuous_normalizer( list(range(env.observation_space.shape[0])), env.observation_space.low, env.observation_space.high, ) elif isinstance(env.observation_space, spaces.Dict): # assuming env.observation_space is image return None else: raise NotImplementedError(f"{env.observation_space} not supported") def build_action_normalizer(env: EnvWrapper): action_space = env.action_space if isinstance(action_space, spaces.Discrete): return discrete_action_normalizer(list(range(action_space.n))) elif isinstance(action_space, spaces.Box): assert ( len(action_space.shape) == 1 ), f"Box action shape {action_space.shape} not supported." action_dim = action_space.shape[0] return only_continuous_action_normalizer( list(range(action_dim)), min_value=action_space.low, max_value=action_space.high, ) else: raise NotImplementedError(f"{action_space} not supported.") def build_normalizer(env: EnvWrapper) -> Dict[str, NormalizationData]: try: return env.normalization_data except AttributeError: # TODO: make this a property of EnvWrapper? # pyre-fixme[16]: Module `envs` has no attribute `RecSim`. if HAS_RECSIM and isinstance(env, RecSim): return { NormalizationKey.STATE: NormalizationData( dense_normalization_parameters=only_continuous_normalizer( list(range(env.observation_space["user"].shape[0])) ) ), NormalizationKey.ITEM: NormalizationData( dense_normalization_parameters=only_continuous_normalizer( list(range(env.observation_space["doc"]["0"].shape[0])) ) ), } return { NormalizationKey.STATE: NormalizationData( dense_normalization_parameters=build_state_normalizer(env) ), NormalizationKey.ACTION: NormalizationData( dense_normalization_parameters=build_action_normalizer(env) ), } def create_df_from_replay_buffer( env, problem_domain: ProblemDomain, desired_size: int, multi_steps: Optional[int], ds: str, shuffle_df: bool = True, ) -> pd.DataFrame: # fill the replay buffer set_seed(env, SEED) if multi_steps is None: update_horizon = 1 return_as_timeline_format = False else: update_horizon = multi_steps return_as_timeline_format = True is_multi_steps = multi_steps is not None # The last element of replay buffer always lacks # next_action and next_possible_actions. # To get full data for every returned sample, we create # replay buffer of desired_size + 1 and discard the last element. replay_buffer = ReplayBuffer( replay_capacity=desired_size + 1, batch_size=1, update_horizon=update_horizon, return_as_timeline_format=return_as_timeline_format, ) random_policy = make_random_policy_for_env(env) agent = Agent.create_for_env(env, policy=random_policy) fill_replay_buffer(env, replay_buffer, desired_size + 1, agent) batch = replay_buffer.sample_transition_batch( batch_size=desired_size, indices=torch.arange(desired_size) ) n = batch.state.shape[0] logger.info(f"Creating df of size {n}.") def discrete_feat_transform(elem) -> str: """query data expects str format""" return str(elem.item()) def continuous_feat_transform(elem: List[float]) -> Dict[int, float]: """query data expects sparse format""" assert isinstance(elem, torch.Tensor), f"{type(elem)} isn't tensor" assert len(elem.shape) == 1, f"{elem.shape} isn't 1-dimensional" return {i: s.item() for i, s in enumerate(elem)} def make_parametric_feat_transform(one_hot_dim: int): """one-hot and then continuous_feat_transform""" def transform(elem) -> Dict[int, float]: elem_tensor = torch.tensor(elem.item()) one_hot_feat = F.one_hot(elem_tensor, one_hot_dim).float() return continuous_feat_transform(one_hot_feat) return transform state_features = feature_transform(batch.state, continuous_feat_transform) next_state_features = feature_transform( batch.next_state, continuous_feat_transform, is_next_with_multi_steps=is_multi_steps, ) if problem_domain == ProblemDomain.DISCRETE_ACTION: # discrete action is str action = feature_transform(batch.action, discrete_feat_transform) next_action = feature_transform( batch.next_action, discrete_feat_transform, is_next_with_multi_steps=is_multi_steps, replace_when_terminal="", terminal=batch.terminal, ) elif problem_domain == ProblemDomain.PARAMETRIC_ACTION: # continuous action is Dict[int, double] assert isinstance(env.action_space, gym.spaces.Discrete) parametric_feat_transform = make_parametric_feat_transform(env.action_space.n) action = feature_transform(batch.action, parametric_feat_transform) next_action = feature_transform( batch.next_action, parametric_feat_transform, is_next_with_multi_steps=is_multi_steps, replace_when_terminal={}, terminal=batch.terminal, ) elif problem_domain == ProblemDomain.CONTINUOUS_ACTION: action = feature_transform(batch.action, continuous_feat_transform) next_action = feature_transform( batch.next_action, continuous_feat_transform, is_next_with_multi_steps=is_multi_steps, replace_when_terminal={}, terminal=batch.terminal, ) elif problem_domain == ProblemDomain.MDN_RNN: action = feature_transform(batch.action, discrete_feat_transform) assert multi_steps is not None next_action = feature_transform( batch.next_action, discrete_feat_transform, is_next_with_multi_steps=True, replace_when_terminal="", terminal=batch.terminal, ) else: raise NotImplementedError(f"model type: {problem_domain}.") if multi_steps is None: time_diff = [1] * n reward = batch.reward.squeeze(1).tolist() metrics = [{"reward": r} for r in reward] else: time_diff = [[1] * len(ns) for ns in next_state_features] reward = [reward_list.tolist() for reward_list in batch.reward] metrics = [ [{"reward": r.item()} for r in reward_list] for reward_list in batch.reward ] # TODO(T67265031): change this to int mdp_id = [str(i.item()) for i in batch.mdp_id] sequence_number = batch.sequence_number.squeeze(1).tolist() # in the product data, all sequence_number_ordinal start from 1. # So to be consistent with the product data. sequence_number_ordinal = (batch.sequence_number.squeeze(1) + 1).tolist() action_probability = batch.log_prob.exp().squeeze(1).tolist() df_dict = { "state_features": state_features, "next_state_features": next_state_features, "action": action, "next_action": next_action, "reward": reward, "action_probability": action_probability, "metrics": metrics, "time_diff": time_diff, "mdp_id": mdp_id, "sequence_number": sequence_number, "sequence_number_ordinal": sequence_number_ordinal, "ds": [ds] * n, } if problem_domain == ProblemDomain.PARAMETRIC_ACTION: # Possible actions are List[Dict[int, float]] assert isinstance(env.action_space, gym.spaces.Discrete) possible_actions = [{i: 1.0} for i in range(env.action_space.n)] elif problem_domain == ProblemDomain.DISCRETE_ACTION: # Possible actions are List[str] assert isinstance(env.action_space, gym.spaces.Discrete) possible_actions = [str(i) for i in range(env.action_space.n)] elif problem_domain == ProblemDomain.MDN_RNN: # Possible actions are List[str] assert isinstance(env.action_space, gym.spaces.Discrete) possible_actions = [str(i) for i in range(env.action_space.n)] # these are fillers, which should have correct shape pa_features = range(n) pna_features = time_diff if problem_domain in ( ProblemDomain.DISCRETE_ACTION, ProblemDomain.PARAMETRIC_ACTION, ProblemDomain.MDN_RNN, ): def pa_transform(x): return possible_actions df_dict["possible_actions"] = feature_transform(pa_features, pa_transform) df_dict["possible_next_actions"] = feature_transform( pna_features, pa_transform, is_next_with_multi_steps=is_multi_steps, replace_when_terminal=[], terminal=batch.terminal, ) df = pd.DataFrame(df_dict) # validate df validate_mdp_ids_seq_nums(df) if shuffle_df: # shuffling (sample the whole batch) df = df.reindex(np.random.permutation(df.index)) return df def set_seed(env: gym.Env, seed: int): np.random.seed(seed) random.seed(seed) torch.manual_seed(seed) env.seed(seed) env.action_space.seed(seed) def feature_transform( features, single_elem_transform, is_next_with_multi_steps=False, replace_when_terminal=None, terminal=None, ): """feature_transform is a method on a single row. We assume features is List[features] (batch of features). This can also be called for next_features with multi_steps which we assume to be List[List[features]]. First List is denoting that it's a batch, second List is denoting that a single row consists of a list of features. """ if is_next_with_multi_steps: if terminal is None: return [ [single_elem_transform(feat) for feat in multi_steps_features] for multi_steps_features in features ] else: # for next features where we replace them when terminal assert replace_when_terminal is not None return [ [single_elem_transform(feat) for feat in multi_steps_features] if not terminal[idx] else [single_elem_transform(feat) for feat in multi_steps_features[:-1]] + [replace_when_terminal] for idx, multi_steps_features in enumerate(features) ] else: if terminal is None: return [single_elem_transform(feat) for feat in features] else: assert replace_when_terminal is not None return [ single_elem_transform(feat) if not terminal[idx] else replace_when_terminal for idx, feat in enumerate(features) ] def validate_mdp_ids_seq_nums(df): mdp_ids = list(df["mdp_id"]) sequence_numbers = list(df["sequence_number"]) unique_mdp_ids = set(mdp_ids) prev_mdp_id, prev_seq_num = None, None mdp_count = 0 for mdp_id, seq_num in zip(mdp_ids, sequence_numbers): if prev_mdp_id is None or mdp_id != prev_mdp_id: mdp_count += 1 prev_mdp_id = mdp_id else: assert seq_num == prev_seq_num + 1, ( f"For mdp_id {mdp_id}, got {seq_num} <= {prev_seq_num}." f"Sequence number must be in increasing order.\n" f"Zip(mdp_id, seq_num): " f"{list(zip(mdp_ids, sequence_numbers))}" ) prev_seq_num = seq_num assert len(unique_mdp_ids) == mdp_count, "MDPs are broken up. {} vs {}".format( len(unique_mdp_ids), mdp_count ) return
15,498
36.167866
88
py
ReAgent
ReAgent-master/reagent/gym/types.py
#!/usr/bin/env python3 # Copyright (c) Facebook, Inc. and its affiliates. All rights reserved. # Please DO NOT import gym in here. We might have installation without gym depending on # this module for typing from abc import ABC, abstractmethod from dataclasses import asdict, dataclass, field, fields from typing import Any, Callable, Dict, List, Optional, Union import numpy as np import reagent.core.types as rlt import torch import torch.nn.functional as F @dataclass class Transition(rlt.BaseDataClass): mdp_id: int sequence_number: int observation: Any action: Any reward: float terminal: bool log_prob: Optional[float] = None possible_actions_mask: Optional[np.ndarray] = None info: Optional[Dict] = None # Same as asdict but filters out none values. def asdict(self): return {k: v for k, v in asdict(self).items() if v is not None} def get_optional_fields(cls) -> List[str]: """return list of optional annotated fields""" ret: List[str] = [] for f in fields(cls): # Check if exactly two arguments exists and one of them are None type if hasattr(f.type, "__args__") and type(None) in f.type.__args__: ret.append(f.name) return ret @dataclass class Trajectory(rlt.BaseDataClass): transitions: List[Transition] = field(default_factory=list) def __post_init__(self): self.optional_field_exist: Dict[str, bool] = { f: False for f in get_optional_fields(Transition) } def __len__(self): return len(self.transitions) def add_transition(self, transition: Transition): if len(self) == 0: # remember which optional fields should be filled for f in self.optional_field_exist: val = getattr(transition, f, None) if val is not None: self.optional_field_exist[f] = True # check that later additions also fill the same optional fields for f, should_exist in self.optional_field_exist.items(): val = getattr(transition, f, None) if (val is not None) != should_exist: raise ValueError( f"Field {f} given val {val} whereas should_exist is {should_exist}." ) self.transitions.append(transition) def __getattr__(self, attr: str): ret = [] for transition in self.transitions: ret.append(getattr(transition, attr)) return ret def calculate_cumulative_reward(self, gamma: float = 1.0): """Return (discounted) sum of rewards.""" num_transitions = len(self) assert num_transitions > 0, "called on empty trajectory" rewards = self.reward discounts = [gamma ** i for i in range(num_transitions)] return sum(reward * discount for reward, discount in zip(rewards, discounts)) def to_dict(self): d = {"action": F.one_hot(torch.from_numpy(np.stack(self.action)), 2)} for f in [ "observation", "reward", "terminal", "log_prob", "possible_actions_mask", ]: if self.optional_field_exist.get(f, True): f_value = getattr(self, f) if np.isscalar(f_value[0]): # scalar values d[f] = torch.tensor(f_value) else: # vector values, need to stack d[f] = torch.from_numpy(np.stack(f_value)).float() return d class Sampler(ABC): """Given scores, select the action.""" @abstractmethod def sample_action(self, scores: Any) -> rlt.ActorOutput: raise NotImplementedError() @abstractmethod def log_prob(self, scores: Any, action: torch.Tensor) -> torch.Tensor: raise NotImplementedError() def update(self) -> None: """Call to update internal parameters (e.g. decay epsilon)""" pass # From preprocessed observation, produce scores for sampler to select action DiscreteScorer = Callable[[Any, Optional[torch.Tensor]], Any] ContinuousScorer = Callable[[Any], Any] Scorer = Union[DiscreteScorer, ContinuousScorer] # Transform ReplayBuffer's transition batch to trainer.train TrainerPreprocessor = Callable[[Any], Any] """ Called after env.step(action) Args: (state, action, reward, terminal, log_prob) """ PostStep = Callable[[Transition], None] """ Called after end of episode """ PostEpisode = Callable[[Trajectory], None] @dataclass class GaussianSamplerScore(rlt.BaseDataClass): loc: torch.Tensor scale_log: torch.Tensor
4,633
30.52381
88
py
ReAgent
ReAgent-master/reagent/gym/preprocessors/default_preprocessors.py
#!/usr/bin/env python3 # Copyright (c) Facebook, Inc. and its affiliates. All rights reserved. """ Get default preprocessors for training time. """ import logging from typing import List, Optional, Tuple import numpy as np import reagent.core.types as rlt import torch import torch.nn.functional as F from gym import Env, spaces logger = logging.getLogger(__name__) ####################################### ### Default obs preprocessors. ### These should operate on single obs. ####################################### class RecsimObsPreprocessor: def __init__( self, *, num_docs: int, discrete_keys: List[Tuple[str, int]], box_keys: List[Tuple[str, int]], ): self.num_docs = num_docs self.discrete_keys = discrete_keys self.box_keys = box_keys @classmethod def create_from_env(cls, env: Env, **kwargs): obs_space = env.observation_space assert isinstance(obs_space, spaces.Dict) user_obs_space = obs_space["user"] if not isinstance(user_obs_space, spaces.Box): raise NotImplementedError( f"User observation space {type(user_obs_space)} is not supported" ) doc_obs_space = obs_space["doc"] if not isinstance(doc_obs_space, spaces.Dict): raise NotImplementedError( f"Doc space {type(doc_obs_space)} is not supported" ) # Assume that all docs are in the same space discrete_keys: List[Tuple[str, int]] = [] box_keys: List[Tuple[str, int]] = [] key_0 = next(iter(doc_obs_space.spaces)) doc_0_space = doc_obs_space[key_0] if isinstance(doc_0_space, spaces.Dict): for k, v in doc_obs_space[key_0].spaces.items(): if isinstance(v, spaces.Discrete): if v.n > 0: discrete_keys.append((k, v.n)) elif isinstance(v, spaces.Box): shape_dim = len(v.shape) if shape_dim == 0: box_keys.append((k, 1)) elif shape_dim == 1: box_keys.append((k, v.shape[0])) else: raise NotImplementedError else: raise NotImplementedError( f"Doc feature {k} with the observation space of {type(v)}" " is not supported" ) elif isinstance(doc_0_space, spaces.Box): pass else: raise NotImplementedError(f"Unknown space {doc_0_space}") return cls( num_docs=len(doc_obs_space.spaces), discrete_keys=sorted(discrete_keys), box_keys=sorted(box_keys), **kwargs, ) def __call__(self, obs): user = torch.tensor(obs["user"]).float().unsqueeze(0) doc_obs = obs["doc"] if self.discrete_keys or self.box_keys: # Dict space discrete_features: List[torch.Tensor] = [] for k, n in self.discrete_keys: vals = torch.tensor([v[k] for v in doc_obs.values()]) assert vals.shape == (self.num_docs,) discrete_features.append(F.one_hot(vals, n).float()) box_features: List[torch.Tensor] = [] for k, d in self.box_keys: vals = np.vstack([v[k] for v in doc_obs.values()]) assert vals.shape == (self.num_docs, d) box_features.append(torch.tensor(vals).float()) doc_features = torch.cat(discrete_features + box_features, dim=1).unsqueeze( 0 ) else: # Simply a Box space vals = np.vstack(list(doc_obs.values())) doc_features = torch.tensor(vals).float().unsqueeze(0) # This comes from ValueWrapper value = ( torch.tensor([v["value"] for v in obs["augmentation"].values()]) .float() .unsqueeze(0) ) candidate_docs = rlt.DocList(float_features=doc_features, value=value) return rlt.FeatureData(float_features=user, candidate_docs=candidate_docs)
4,247
32.714286
88
py
ReAgent
ReAgent-master/reagent/gym/preprocessors/trainer_preprocessor.py
#!/usr/bin/env python3 # Copyright (c) Facebook, Inc. and its affiliates. All rights reserved. """ Get default preprocessors for training time. """ import inspect import logging from typing import Dict, Optional import gym import numpy as np import reagent.core.types as rlt import torch import torch.nn.functional as F from reagent.core.parameters import CONTINUOUS_TRAINING_ACTION_RANGE from reagent.gym.types import Trajectory from reagent.preprocessing.types import InputColumn from reagent.training.reagent_lightning_module import ReAgentLightningModule from reagent.training.utils import rescale_actions logger = logging.getLogger(__name__) # This is here to make typechecker happpy, sigh ONLINE_MAKER_MAP = {} REPLAY_BUFFER_MAKER_MAP = {} def make_trainer_preprocessor( trainer: ReAgentLightningModule, device: torch.device, env: gym.Env, maker_map: Dict, ): assert isinstance(trainer, ReAgentLightningModule), f"{type(trainer)}" sig = inspect.signature(trainer.train_step_gen) logger.info(f"Deriving trainer_preprocessor from {sig.parameters}") # Assuming training_batch is in the first position (excluding self) assert ( list(sig.parameters.keys())[0] == "training_batch" ), f"{sig.parameters} doesn't have training batch in first position." training_batch_type = sig.parameters["training_batch"].annotation assert training_batch_type != inspect.Parameter.empty try: maker = maker_map[training_batch_type].create_for_env(env) except KeyError: logger.error(f"Unknown type: {training_batch_type}") raise def trainer_preprocessor(batch): retval = maker(batch) return retval.to(device) return trainer_preprocessor def make_trainer_preprocessor_online( trainer: ReAgentLightningModule, device: torch.device, env: gym.Env ): return make_trainer_preprocessor(trainer, device, env, ONLINE_MAKER_MAP) def make_replay_buffer_trainer_preprocessor( trainer: ReAgentLightningModule, device: torch.device, env: gym.Env ): return make_trainer_preprocessor(trainer, device, env, REPLAY_BUFFER_MAKER_MAP) def one_hot_actions( num_actions: int, action: torch.Tensor, next_action: torch.Tensor, terminal: torch.Tensor, ): """ One-hot encode actions and non-terminal next actions. Input shape is (batch_size, 1). Output shape is (batch_size, num_actions) """ assert ( len(action.shape) == 2 and action.shape[1] == 1 and next_action.shape == action.shape ), ( f"Must be action with stack_size = 1, but " f"got shapes {action.shape}, {next_action.shape}" ) action = F.one_hot(action, num_actions).squeeze(1).float() # next action is garbage for terminal transitions (so just zero them) next_action_res = torch.zeros_like(action) non_terminal_indices = (terminal == 0).squeeze(1) next_action_res[non_terminal_indices] = ( F.one_hot(next_action[non_terminal_indices], num_actions).squeeze(1).float() ) return action, next_action_res class DiscreteDqnInputMaker: def __init__(self, num_actions: int, trainer_preprocessor=None): self.num_actions = num_actions self.trainer_preprocessor = trainer_preprocessor @classmethod def create_for_env(cls, env: gym.Env): action_space = env.action_space assert isinstance(action_space, gym.spaces.Discrete) try: return cls( num_actions=action_space.n, # pyre-fixme[16]: `Env` has no attribute `trainer_preprocessor`. trainer_preprocessor=env.trainer_preprocessor, ) except AttributeError: return cls(num_actions=action_space.n) def __call__(self, batch): not_terminal = 1.0 - batch.terminal.float() action, next_action = one_hot_actions( self.num_actions, batch.action, batch.next_action, batch.terminal ) if self.trainer_preprocessor is not None: state = self.trainer_preprocessor(batch.state) next_state = self.trainer_preprocessor(batch.next_state) else: state = rlt.FeatureData(float_features=batch.state) next_state = rlt.FeatureData(float_features=batch.next_state) try: possible_actions_mask = batch.possible_actions_mask.float() except AttributeError: possible_actions_mask = torch.ones_like(action).float() try: possible_next_actions_mask = batch.next_possible_actions_mask.float() except AttributeError: possible_next_actions_mask = torch.ones_like(next_action).float() return rlt.DiscreteDqnInput( state=state, action=action, next_state=next_state, next_action=next_action, possible_actions_mask=possible_actions_mask, possible_next_actions_mask=possible_next_actions_mask, reward=batch.reward, not_terminal=not_terminal, step=None, time_diff=None, extras=rlt.ExtraData( mdp_id=None, sequence_number=None, action_probability=batch.log_prob.exp(), max_num_actions=None, metrics=None, ), ) class PolicyNetworkInputMaker: def __init__(self, action_low: np.ndarray, action_high: np.ndarray): self.action_low = torch.tensor(action_low) self.action_high = torch.tensor(action_high) (train_low, train_high) = CONTINUOUS_TRAINING_ACTION_RANGE self.train_low = torch.tensor(train_low) self.train_high = torch.tensor(train_high) @classmethod def create_for_env(cls, env: gym.Env): action_space = env.action_space assert isinstance(action_space, gym.spaces.Box) return cls(action_space.low, action_space.high) def __call__(self, batch): not_terminal = 1.0 - batch.terminal.float() # normalize actions action = rescale_actions( batch.action, new_min=self.train_low, new_max=self.train_high, prev_min=self.action_low, prev_max=self.action_high, ) # only normalize non-terminal non_terminal_indices = (batch.terminal == 0).squeeze(1) next_action = torch.zeros_like(action) next_action[non_terminal_indices] = rescale_actions( batch.next_action[non_terminal_indices], new_min=self.train_low, new_max=self.train_high, prev_min=self.action_low, prev_max=self.action_high, ) dict_batch = { InputColumn.STATE_FEATURES: batch.state, InputColumn.NEXT_STATE_FEATURES: batch.next_state, InputColumn.ACTION: action, InputColumn.NEXT_ACTION: next_action, InputColumn.REWARD: batch.reward, InputColumn.NOT_TERMINAL: not_terminal, InputColumn.STEP: None, InputColumn.TIME_DIFF: None, InputColumn.EXTRAS: rlt.ExtraData( mdp_id=None, sequence_number=None, action_probability=batch.log_prob.exp(), max_num_actions=None, metrics=None, ), } has_candidate_features = False try: dict_batch.update( { InputColumn.CANDIDATE_FEATURES: batch.doc, InputColumn.NEXT_CANDIDATE_FEATURES: batch.next_doc, } ) has_candidate_features = True except AttributeError: pass output = rlt.PolicyNetworkInput.from_dict(dict_batch) if has_candidate_features: output.state = rlt._embed_states(output.state) output.next_state = rlt._embed_states(output.next_state) return output class SlateQInputMaker: def __init__(self): self.metric = "watch_time" @classmethod def create_for_env(cls, env: gym.Env): return cls() def __call__(self, batch): n = batch.state.shape[0] item_mask = torch.ones(batch.doc.shape[:2]) next_item_mask = torch.ones(batch.doc.shape[:2]) # TODO: abs value to make probability? item_probability = batch.augmentation_value # .unsqueeze(2) next_item_probability = batch.next_augmentation_value # .unsqueeze(2) # concat null action null_action = torch.tensor([batch.action.shape[1]] * n, dtype=torch.int64).view( n, 1 ) action = torch.cat([batch.action, null_action], dim=1) next_action = torch.cat([batch.next_action, null_action], dim=1) # concat null reward to position wise reward position_reward = getattr(batch, f"response_{self.metric}") null_reward = torch.zeros((n, 1)) position_reward = torch.cat([position_reward, null_reward], dim=1) # concat null mask when nothing clicked reward_mask = batch.response_click null_mask = (reward_mask.sum(dim=1) == 0).view(n, 1) reward_mask = torch.cat([reward_mask.to(torch.bool), null_mask], dim=1) dict_batch = { "state_features": batch.state, "next_state_features": batch.next_state, "candidate_features": batch.doc, "next_candidate_features": batch.next_doc, "item_mask": item_mask, "next_item_mask": next_item_mask, "item_probability": item_probability, "next_item_probability": next_item_probability, "action": action, "next_action": next_action, "position_reward": position_reward, "reward_mask": reward_mask, "time_diff": None, "not_terminal": ~batch.terminal, } return rlt.SlateQInput.from_dict(dict_batch) class MemoryNetworkInputMaker: def __init__(self, num_actions: Optional[int] = None): self.num_actions = num_actions @classmethod def create_for_env(cls, env: gym.Env): action_space = env.action_space if isinstance(action_space, gym.spaces.Discrete): return cls(action_space.n) elif isinstance(action_space, gym.spaces.Box): return cls() else: raise NotImplementedError() def __call__(self, batch): action = batch.action if self.num_actions is not None: assert len(action.shape) == 2, f"{action.shape}" # one hot makes shape (batch_size, stack_size, feature_dim) action = F.one_hot(batch.action, self.num_actions).float() # make shape to (batch_size, feature_dim, stack_size) action = action.transpose(1, 2) # For (1-dimensional) vector fields, RB returns (batch_size, state_dim) # or (batch_size, state_dim, stack_size). # We want these to all be (stack_size, batch_size, state_dim), so # unsqueeze the former case; Note this only happens for stack_size = 1. # Then, permute. permutation = [2, 0, 1] vector_fields = { "state": batch.state, "action": action, "next_state": batch.next_state, } for name, tensor in vector_fields.items(): if len(tensor.shape) == 2: tensor = tensor.unsqueeze(2) assert len(tensor.shape) == 3, f"{name} has shape {tensor.shape}" vector_fields[name] = tensor.permute(permutation) # For scalar fields, RB returns (batch_size), or (batch_size, stack_size) # Do same as above, except transpose instead. scalar_fields = { "reward": batch.reward, "not_terminal": 1.0 - batch.terminal.float(), } for name, tensor in scalar_fields.items(): if len(tensor.shape) == 1: tensor = tensor.unsqueeze(1) assert len(tensor.shape) == 2, f"{name} has shape {tensor.shape}" scalar_fields[name] = tensor.transpose(0, 1) # stack_size > 1, so let's pad not_terminal with 1's, since # previous states couldn't have been terminal.. if scalar_fields["reward"].shape[0] > 1: batch_size = scalar_fields["reward"].shape[1] assert scalar_fields["not_terminal"].shape == ( 1, batch_size, ), f"{scalar_fields['not_terminal'].shape}" stacked_not_terminal = torch.ones_like(scalar_fields["reward"]) stacked_not_terminal[-1] = scalar_fields["not_terminal"] scalar_fields["not_terminal"] = stacked_not_terminal dict_batch = { "state": vector_fields["state"], "next_state": vector_fields["next_state"], "action": vector_fields["action"], "reward": scalar_fields["reward"], "not_terminal": scalar_fields["not_terminal"], "step": None, "time_diff": None, } return rlt.MemoryNetworkInput.from_dict(dict_batch) def get_possible_actions_for_gym(batch_size: int, num_actions: int) -> rlt.FeatureData: """ tiled_actions should be (batch_size * num_actions, num_actions) forall i in [batch_size], tiled_actions[i*num_actions:(i+1)*num_actions] should be I[num_actions] where I[n] is the n-dimensional identity matrix. NOTE: this is only the case for when we convert discrete action to parametric action via one-hot encoding. """ possible_actions = torch.eye(num_actions).repeat(repeats=(batch_size, 1)) return rlt.FeatureData(float_features=possible_actions) class ParametricDqnInputMaker: def __init__(self, num_actions: int): self.num_actions = num_actions @classmethod def create_for_env(cls, env: gym.Env): action_space = env.action_space assert isinstance(action_space, gym.spaces.Discrete) return cls(action_space.n) def __call__(self, batch): not_terminal = 1.0 - batch.terminal.float() assert ( len(batch.state.shape) == 2 ), f"{batch.state.shape} is not (batch_size, state_dim)." batch_size, _ = batch.state.shape action, next_action = one_hot_actions( self.num_actions, batch.action, batch.next_action, batch.terminal ) possible_actions = get_possible_actions_for_gym(batch_size, self.num_actions) possible_next_actions = possible_actions.clone() possible_actions_mask = torch.ones((batch_size, self.num_actions)) possible_next_actions_mask = possible_actions_mask.clone() return rlt.ParametricDqnInput( state=rlt.FeatureData(float_features=batch.state), action=rlt.FeatureData(float_features=action), next_state=rlt.FeatureData(float_features=batch.next_state), next_action=rlt.FeatureData(float_features=next_action), possible_actions=possible_actions, possible_actions_mask=possible_actions_mask, possible_next_actions=possible_next_actions, possible_next_actions_mask=possible_next_actions_mask, reward=batch.reward, not_terminal=not_terminal, step=None, time_diff=None, extras=rlt.ExtraData( mdp_id=None, sequence_number=None, action_probability=batch.log_prob.exp(), max_num_actions=None, metrics=None, ), ) class PolicyGradientInputMaker: def __init__(self, num_actions: Optional[int] = None, recsim_obs: bool = False): self.num_actions = num_actions self.recsim_obs = recsim_obs @classmethod def create_for_env(cls, env: gym.Env): action_space = env.action_space if isinstance(action_space, gym.spaces.Discrete): return cls(action_space.n) elif isinstance(action_space, gym.spaces.Box): return cls() elif isinstance(action_space, gym.spaces.MultiDiscrete): return cls(recsim_obs=True) else: raise NotImplementedError() def _get_recsim_state(self, observation): def _stack(slates): obs = rlt.FeatureData( float_features=torch.from_numpy( np.stack(np.array([slate["user"] for slate in slates])) ), candidate_docs=rlt.DocList( float_features=torch.from_numpy( np.stack(np.array([slate["doc"] for slate in slates])) ) ), ) return obs def _stack_slate(slate): return { "user": slate["user"], "doc": np.stack(np.array(list(slate["doc"].values()))), } return _stack([_stack_slate(slate) for slate in observation]) def __call__(self, trajectory: Trajectory): action = torch.from_numpy(np.stack(trajectory.action).squeeze()) if self.num_actions is not None: action = F.one_hot(action, self.num_actions).float() assert len(action.shape) == 2, f"{action.shape}" # one hot makes shape (batch_size, num_actions) state = ( self._get_recsim_state(trajectory.observation) if self.recsim_obs else rlt.FeatureData( torch.from_numpy(np.stack(trajectory.observation)).float() ) ) return rlt.PolicyGradientInput( state=state, action=action, reward=torch.tensor(trajectory.reward), log_prob=torch.tensor(trajectory.log_prob), ) ONLINE_MAKER_MAP = {rlt.PolicyGradientInput: PolicyGradientInputMaker} REPLAY_BUFFER_MAKER_MAP = { rlt.DiscreteDqnInput: DiscreteDqnInputMaker, rlt.PolicyNetworkInput: PolicyNetworkInputMaker, rlt.MemoryNetworkInput: MemoryNetworkInputMaker, rlt.ParametricDqnInput: ParametricDqnInputMaker, rlt.SlateQInput: SlateQInputMaker, }
18,178
36.637681
88
py
ReAgent
ReAgent-master/reagent/gym/envs/oracle_pvm.py
#!/usr/bin/env python3 # Copyright (c) Facebook, Inc. and its affiliates. All rights reserved. import logging from collections import OrderedDict from typing import Callable, Dict, List import gym import numpy as np import reagent.core.types as rlt import torch from reagent.core.dataclasses import dataclass from reagent.gym.envs import RecSim from reagent.gym.preprocessors.default_preprocessors import RecsimObsPreprocessor from scipy import stats logger = logging.getLogger(__name__) # score function takes user and doc features, and outputs a score SCORE_FUNCTION_T = Callable[[np.ndarray, np.ndarray], float] def make_default_score_fn(fn_i: int) -> SCORE_FUNCTION_T: """ Make ith score_fn (constructor of ith score) """ def fn(user: np.ndarray, doc: np.ndarray) -> float: return doc[fn_i] # user = user ** (fn_i + 1) # doc = doc ** (fn_i + 1) # return np.inner(user, doc) # return user[fn_i] * doc[fn_i] return fn VM_WEIGHT_LOW = -1.0 VM_WEIGHT_HIGH = 1.0 MATCH_REWARD_BOOST = 3.0 def get_default_score_fns(num_weights): return [make_default_score_fn(i) for i in range(num_weights)] def get_ground_truth_weights(num_weights): return np.array([1] * num_weights) @dataclass class OraclePVM(RecSim): """ Wrapper over RecSim for simulating (Personalized) VM Tuning. The state is the same as for RecSim (user feature + candidate features). There are num_weights VM weights to tune, and so action space is a vector of length num_weights. OraclePVM hides num_weights number of (1) score_fns (akin to VM models), that take in user + candidate_i feature and produces a score for candidate_i. (2) ground_truth_weights, that are used to produce "ground truth", a.k.a. "Oracle", rankings. Reward is the Kendall-Tau between ground truth and the ranking created from the weights given by action. If the rankings match exactly, the reward is boosted to 3. NOTE: This environment only tests if the Agent can learn the hidden ground truth weights, which may be far from optimal (in terms of RecSim's rewards, which we're ignoring). This is easier for unit tests, but in the real world we will be trying to learn the optimal weights, and the reward signal would reflect that. TODO: made environment easier to learn from by not using RecSim. """ user_feat_dim: int = 1 candidate_feat_dim: int = 3 num_weights: int = 3 def __post_init_post_parse__(self): assert ( self.slate_size == self.num_candidates ), f"Must be equal (slate_size) {self.slate_size} != (num_candidates) {self.num_candidates}" super().__post_init_post_parse__() self.score_fns: List[SCORE_FUNCTION_T] = get_default_score_fns(self.num_weights) self.ground_truth_weights: List[float] = get_ground_truth_weights( self.num_weights ) assert len(self.score_fns) == len( self.ground_truth_weights ), f"{len(self.score_fns)} != {len(self.ground_truth_weights)}" assert ( len(self.ground_truth_weights) == self.num_weights ), f"{self.ground_truth_weights.shape} != {self.num_weights}" def reset(self): self.prev_obs = super().reset() self.prev_obs.update( { "user": np.random.rand(self.user_feat_dim), "doc": OrderedDict( [ (str(i), np.random.rand(self.candidate_feat_dim)) for i in range(self.num_candidates) ] ), } ) return self.prev_obs def step(self, action): user_feat = self.prev_obs["user"] doc_feats = self.prev_obs["doc"] scores = self._get_scores(user_feat, doc_feats) ground_truth_ranking = self._get_ranking(scores, self.ground_truth_weights) policy_ranking = self._get_ranking(scores, action) t = True # comment out to avoid non-stationary # self.prev_obs, _, t, i = super().step(policy_ranking) num_matches = (ground_truth_ranking == policy_ranking).sum() if num_matches == self.slate_size: reward = MATCH_REWARD_BOOST else: reward, _p_value = stats.kendalltau(ground_truth_ranking, policy_ranking) return self.prev_obs, reward, t, None def is_match(self, reward): # for evaluation, return true iff the reward represents a match return reward > (MATCH_REWARD_BOOST - 1e-6) @property def action_space(self): return gym.spaces.Box( low=VM_WEIGHT_LOW, high=VM_WEIGHT_HIGH, shape=(self.num_weights,) ) @action_space.setter def action_space(self, val): pass def _get_scores( self, user_feat: np.ndarray, doc_feats: Dict[str, np.ndarray] ) -> np.ndarray: # num_docs x num_scores where i,j coordinate is jth score for ith doc scores = np.array( [ [score_fn(user_feat, doc_feat) for score_fn in self.score_fns] for _k, doc_feat in doc_feats.items() ] ) return scores def _get_ranking(self, scores: np.ndarray, weights: np.ndarray): assert weights.shape == (scores.shape[1],), f"{weights.shape}, {scores.shape}" weighted_scores = scores * weights values = weighted_scores.sum(axis=1) indices = np.argsort(-values) return indices[: self.slate_size] def obs_preprocessor(self, obs: np.ndarray) -> rlt.FeatureData: preprocessor = RecsimObsPreprocessor.create_from_env(self) preprocessed_obs = preprocessor(obs) return rlt._embed_states(preprocessed_obs) def serving_obs_preprocessor(self, obs: np.ndarray): preprocessor = RecsimObsPreprocessor.create_from_env(self) x = preprocessor(obs) # user was batch_size x state_size, stack user = x.float_features.unsqueeze(1).repeat_interleave( self.num_candidates, dim=1 ) candidates = x.candidate_docs.float_features combined = torch.cat([user, candidates], dim=2).squeeze(0) return (combined, torch.ones_like(combined, dtype=torch.uint8))
6,305
35.034286
100
py
ReAgent
ReAgent-master/reagent/gym/envs/changing_arms.py
#!/usr/bin/env python3 # Copyright (c) Facebook, Inc. and its affiliates. All rights reserved. """ Traditional MAB setup has sequence length = 1 always. In this setup, the distributions of the arms rewards changes every round, and the agent is presented with some information and control about how the arms will change. In particular, the observation includes "mu_changes", which is the possible changes to mu; only the arm picked by agent will have it's mu_changes reflected. This way, the next state depend on (only) the previous state and action; hence this a MDP. The reward for picking an action is the change in mu corresponding to that arm. With following set-up (where ARM_INIT_VALUE = 100 and NUM_ARMS = 5), the optimal policy can accumulate a reward of 500 per run. Note that if the policy picks an illegal action at any time, the game ends. """ import random import gym import numpy as np import reagent.core.types as rlt import torch from reagent.core.dataclasses import dataclass from reagent.core.parameters import NormalizationData, NormalizationKey from reagent.gym.envs.env_wrapper import EnvWrapper from reagent.gym.normalizers import only_continuous_normalizer ABS_LOW = -1000.0 ABS_HIGH = 1000.0 MU_LOW = 0.0 MU_HIGH = 1000.0 # illegal move causes game to end with a big BOOM!!! INVALID_MOVE_PENALTY = -1000.0 IDLE_PENALTY = -500.0 NUM_ARMS = 5 # keep these constant for now ARM_INIT_VALUE = 100.0 ARM_MU_DECREASE = 10.0 MAX_STEPS = 49 # in the real world, IDs are not indices into embedding table # thus, we offset vals to test hashing mechanism ID_LIST_OFFSET = 1000000 ID_SCORE_LIST_OFFSET = 1500000 ID_LIST_FEATURE_ID = 100 ID_SCORE_LIST_FEATURE_ID = 1000 def get_initial_mus(num_arms): return torch.tensor([ARM_INIT_VALUE] * num_arms) def get_mu_changes(num_arms): return torch.tensor([-ARM_MU_DECREASE] * num_arms) def get_legal_indices_mask(num_arms): # FIXME: hardcoded for now assert num_arms == 5, f"unsupported num_arms = {num_arms}, should be 5" LEGAL_PROBS = torch.tensor([0.95, 1.0, 0.95, 0.8, 0.8]) return torch.bernoulli(LEGAL_PROBS).to(torch.uint8) @dataclass class ChangingArms(EnvWrapper): num_arms: int = NUM_ARMS def make(self) -> gym.Env: return ChangingArmsEnv(self.num_arms) def _split_state(self, obs: np.ndarray): assert obs.shape == (3, self.num_arms), f"{obs.shape}." dense_val = torch.tensor(obs[0, :]).view(1, self.num_arms) id_list_val = torch.tensor(obs[1, :]).nonzero(as_tuple=True)[0].to(torch.long) id_score_list_val = torch.tensor(obs[2, :]) return dense_val, id_list_val, id_score_list_val def obs_preprocessor(self, obs: np.ndarray) -> rlt.FeatureData: dense_val, id_list_val, id_score_list_val = self._split_state(obs) return rlt.FeatureData( # dense value float_features=dense_val, # (offset, value) id_list_features={ "legal": (torch.tensor([0], dtype=torch.long), id_list_val) }, # (offset, key, value) id_score_list_features={ "mu_changes": ( torch.tensor([0], dtype=torch.long), torch.arange(self.num_arms, dtype=torch.long), id_score_list_val, ) }, ) def serving_obs_preprocessor(self, obs: np.ndarray) -> rlt.ServingFeatureData: dense_val, id_list_val, id_score_list_val = self._split_state(obs) return rlt.ServingFeatureData( float_features_with_presence=( dense_val, torch.ones_like(dense_val, dtype=torch.uint8), ), id_list_features={ ID_LIST_FEATURE_ID: ( torch.tensor([0], dtype=torch.long), id_list_val + ID_LIST_OFFSET, ) }, id_score_list_features={ ID_SCORE_LIST_FEATURE_ID: ( torch.tensor([0], dtype=torch.long), torch.arange(self.num_arms, dtype=torch.long) + ID_SCORE_LIST_OFFSET, id_score_list_val, ) }, ) def split_state_transform(self, elem: torch.Tensor): """For generate data""" dense_val, id_list_val, id_score_list_val = self._split_state(elem.numpy()) return ( {i: s.item() for i, s in enumerate(dense_val.view(-1))}, {ID_LIST_FEATURE_ID: (id_list_val + ID_LIST_OFFSET).tolist()}, { ID_SCORE_LIST_FEATURE_ID: { i + ID_SCORE_LIST_OFFSET: s.item() for i, s in enumerate(id_score_list_val) } }, ) @property def normalization_data(self): return { NormalizationKey.STATE: NormalizationData( dense_normalization_parameters=only_continuous_normalizer( list(range(self.num_arms)), MU_LOW, MU_HIGH ) ) } def trainer_preprocessor(self, obs: torch.Tensor): batch_size = obs.shape[0] assert obs.shape == (batch_size, 3, self.num_arms), f"{obs.shape}" dense_val = obs[:, 0, :].view(batch_size, self.num_arms) # extract one-hot encoded values from id_list batch_indices, id_list_val = obs[:, 1, :].nonzero(as_tuple=True) offsets = [] prev_batch_idx = -1 for i, batch_idx in enumerate(batch_indices.tolist()): if batch_idx > prev_batch_idx: offsets.extend([i] * (batch_idx - prev_batch_idx)) prev_batch_idx = batch_idx else: assert batch_idx == prev_batch_idx # handle the case of trailing empty batches # pyre-fixme[61]: `batch_idx` may not be initialized here. if batch_idx < batch_size - 1: # pyre-fixme[61]: `batch_idx` may not be initialized here. offsets.extend([i] * (batch_size - 1 - batch_idx)) assert len(offsets) == batch_size, f"{len(offsets)} != {batch_size}." id_list_offsets = torch.tensor(offsets) # id_score_list is easier because not one-hot encoded id_score_list_offsets = torch.tensor( list(range(0, batch_size * self.num_arms, self.num_arms)) ) id_score_list_keys = torch.arange(self.num_arms).repeat(batch_size) id_score_list_vals = obs[:, 2, :].reshape(-1) return rlt.FeatureData( # dense value float_features=dense_val, # (offset, value) id_list_features={"legal": (id_list_offsets, id_list_val)}, # (offset, key, value) id_score_list_features={ "mu_changes": ( id_score_list_offsets, id_score_list_keys, id_score_list_vals, ) }, ) class ChangingArmsEnv(gym.Env): """This is just the gym environment, without extra functionality""" def __init__(self, num_arms): self.seed(0) self.num_arms = num_arms self.max_steps = MAX_STEPS def step(self, action): if isinstance(action, np.ndarray): action = action.item() assert ( 0 <= action and action <= self.num_arms ), f"out-of-bounds action {action}." reached_max_steps = self.num_steps >= self.max_steps self.num_steps += 1 # idle action if action == self.num_arms: # simply return new state, without updating distributions # this is ideal when there aren't any legal actions, this # would generate a new batch of legal actions return self.state, IDLE_PENALTY, reached_max_steps, None # illegal action if action not in self.legal_indices: return self.state, INVALID_MOVE_PENALTY, True, None # update states for only the action selected prev = self.mus[action].item() self.mus[action] = prev + self.mu_changes[action] if self.mus[action] <= MU_LOW: self.legal_indices_mask[action] = 0 reward = prev - self.mus[action].item() return self.state, reward, reached_max_steps, None def seed(self, seed: int): random.seed(seed) torch.manual_seed(seed) def reset(self): # initialize the distributions self.num_steps = 0 self.mus = get_initial_mus(self.num_arms) # these are turned off when an arm has been "exhausted" self.legal_indices_mask = torch.tensor([1] * self.num_arms).to(torch.uint8) return self.state @property def state(self): """ State comprises of: - initial mus - legal_indices mask - randomly-generated mu changes """ self.mu_changes = get_mu_changes(self.num_arms) legal_indices_mask = ( get_legal_indices_mask(self.num_arms) & self.legal_indices_mask ) self.legal_indices = legal_indices_mask.nonzero(as_tuple=True)[0] result = torch.stack([self.mus, legal_indices_mask, self.mu_changes]) return result.numpy() @property def observation_space(self): """ It should really be a Dict, but we return them all stacked since it's more convenient for RB. """ return gym.spaces.Box(ABS_LOW, ABS_HIGH, shape=(3, self.num_arms)) @property def action_space(self): # Selecting 0,1,2...,num_arms-1 is selecting an arm. # If action is invalid, agent incurs a penalty. # If action is valid, action is an idx i, and reward # is a sample from ith distribution. At the same time # the ith distribution is updated with the changes. # Alternatively, can choose NULL (i.e. do-nothing) action # if action = num_arms return gym.spaces.Discrete(self.num_arms + 1)
10,024
34.9319
86
py
ReAgent
ReAgent-master/reagent/gym/envs/gym.py
#!/usr/bin/env python3 import logging from typing import Optional, Tuple import gym import numpy as np import reagent.core.types as rlt import torch from gym import spaces from gym_minigrid.wrappers import ReseedWrapper from reagent.core.dataclasses import dataclass from reagent.gym.envs.env_wrapper import EnvWrapper from reagent.gym.envs.wrappers.simple_minigrid import SimpleObsWrapper logger = logging.getLogger(__name__) @dataclass class Gym(EnvWrapper): env_name: str set_max_steps: Optional[int] = None def make(self) -> gym.Env: kwargs = {} if self.set_max_steps is not None: kwargs["max_steps"] = self.set_max_steps env: gym.Env = gym.make(self.env_name, **kwargs) if self.env_name.startswith("MiniGrid-"): # Wrap in minigrid simplifier env = SimpleObsWrapper(ReseedWrapper(env)) return env def obs_preprocessor(self, obs: np.ndarray) -> rlt.FeatureData: obs_space = self.observation_space if isinstance(obs_space, spaces.Box): return rlt.FeatureData(torch.tensor(obs).float().unsqueeze(0)) else: raise NotImplementedError(f"{obs_space} obs space not supported for Gym.") # TODO: make return serving feature data # pyre-fixme[15]: `serving_obs_preprocessor` overrides method defined in # `EnvWrapper` inconsistently. def serving_obs_preprocessor( self, obs: np.ndarray ) -> Tuple[torch.Tensor, torch.Tensor]: obs_space = self.observation_space if not isinstance(obs_space, spaces.Box): raise NotImplementedError(f"{obs_space} not supported!") if len(obs_space.shape) != 1: raise NotImplementedError(f"Box shape {obs_space.shape} not supported!") state_dim = obs_space.shape[0] obs_tensor = torch.tensor(obs).float().view(1, state_dim) presence_tensor = torch.ones_like(obs_tensor) return (obs_tensor, presence_tensor)
1,987
33.275862
86
py
ReAgent
ReAgent-master/reagent/gym/envs/env_wrapper.py
#!/usr/bin/env python3 # Copyright (c) Facebook, Inc. and its affiliates. All rights reserved. import abc import logging from typing import Callable, Optional import gym import numpy as np import reagent.core.types as rlt import torch from gym import spaces from reagent.core.dataclasses import dataclass from reagent.core.parameters import CONTINUOUS_TRAINING_ACTION_RANGE from reagent.core.registry_meta import RegistryMeta from reagent.training.utils import rescale_actions # types for reference ObsPreprocessor = Callable[[np.ndarray], rlt.FeatureData] ServingObsPreprocessor = Callable[[np.ndarray], rlt.ServingFeatureData] ActionExtractor = Callable[[rlt.ActorOutput], np.ndarray] ServingActionExtractor = ActionExtractor CONTINUOUS_MODEL_LOW = torch.tensor(CONTINUOUS_TRAINING_ACTION_RANGE[0]) CONTINUOUS_MODEL_HIGH = torch.tensor(CONTINUOUS_TRAINING_ACTION_RANGE[1]) logger = logging.getLogger(__name__) @dataclass class EnvWrapper(gym.core.Wrapper, metaclass=RegistryMeta): """Wrapper around it's environment, to simplify configuration.""" def __post_init_post_parse__(self): super().__init__(self.make()) logger.info( f"Env: {self.env};\n" f"observation_space: {self.env.observation_space};\n" f"action_space: {self.env.action_space};" ) @abc.abstractmethod def make(self) -> gym.Env: pass @abc.abstractmethod def obs_preprocessor(self, obs: np.ndarray) -> rlt.FeatureData: pass @abc.abstractmethod def serving_obs_preprocessor(self, obs: np.ndarray) -> rlt.ServingFeatureData: pass def get_obs_preprocessor(self, *ctor_args, **ctor_kwargs): # ctor_args go to .to call ctor_kwargs["non_blocking"] = True return lambda *args, **kwargs: self.obs_preprocessor(*args, **kwargs).to( *ctor_args, **ctor_kwargs ) def get_serving_obs_preprocessor(self): return lambda *args, **kwargs: self.serving_obs_preprocessor(*args, **kwargs) def action_extractor(self, actor_output: rlt.ActorOutput) -> torch.Tensor: action = actor_output.action action_space = self.action_space # Canonical rule to return one-hot encoded actions for discrete assert ( len(action.shape) == 2 and action.shape[0] == 1 ), f"{action} (shape: {action.shape}) is not a single action!" if isinstance(action_space, spaces.Discrete): # pyre-fixme[16]: `Tensor` has no attribute `argmax`. return action.squeeze(0).argmax() elif isinstance(action_space, spaces.MultiDiscrete): return action.squeeze(0) # Canonical rule to scale actions to CONTINUOUS_TRAINING_ACTION_RANGE elif isinstance(action_space, spaces.Box): assert len(action_space.shape) == 1, f"{action_space} not supported." return rescale_actions( action.squeeze(0), new_min=torch.tensor(action_space.low), new_max=torch.tensor(action_space.high), prev_min=CONTINUOUS_MODEL_LOW, prev_max=CONTINUOUS_MODEL_HIGH, ) else: raise NotImplementedError(f"Unsupported action space: {action_space}") def serving_action_extractor(self, actor_output: rlt.ActorOutput) -> torch.Tensor: action = actor_output.action action_space = self.action_space assert ( len(action.shape) == 2 and action.shape[0] == 1 ), f"{action.shape} isn't (1, action_dim)" if isinstance(action_space, spaces.Discrete): # pyre-fixme[16]: `Tensor` has no attribute `argmax`. return action.squeeze(0).argmax().view([]) elif isinstance(action_space, spaces.MultiDiscrete): return action.squeeze(0) elif isinstance(action_space, spaces.Box): assert ( len(action_space.shape) == 1 ), f"Unsupported Box with shape {action_space.shape}" return action.squeeze(0) else: raise NotImplementedError(f"Unsupported action space: {action_space}") def get_action_extractor(self): return ( lambda *args, **kwargs: self.action_extractor(*args, **kwargs).cpu().numpy() ) def get_serving_action_extractor(self): return ( lambda *args, **kwargs: self.serving_action_extractor(*args, **kwargs) .cpu() .numpy() ) # TODO: add more methods to simplify gym code # e.g. normalization, specific preprocessor, etc. # This can move a lot of the if statements from create_from_env methods. @property def max_steps(self) -> Optional[int]: possible_keys = [ # gym should have _max_episode_steps "_max_episode_steps", # Minigrid should have max_steps "max_steps", ] for key in possible_keys: res = getattr(self.env, key, None) if res is not None: return res return None @property def possible_actions_mask(self) -> Optional[np.ndarray]: ret = getattr(self.env, "possible_actions_mask", None) if ret is not None: ret = ret.copy() return ret
5,302
35.572414
88
py
ReAgent
ReAgent-master/reagent/gym/envs/pomdp/string_game_v1.py
#!/usr/bin/env python3 # Copyright (c) Facebook, Inc. and its affiliates. All rights reserved. """ A game with a stochastic length of the MDP but no longer than 3. An agent can choose one character to reveal (either "A" or "B") as the action, and the next state is exactly the action just taken (i.e., the transition function only depends on the action). Each episode is limited to 3 steps. There is some probability to terminate at any step (but the agent must terminate if making 3 steps) If the current state is "A", the agent has 0.5 probability to make to the next step. If the current state is "B", the agent has 0.9 probability to make to the next step. The reward is given at the terminal state, based on the accumulated observation (a string). If the agent observes "AAA" (survive the first 2 steps and terminate at the last step no matter what action taken), it receives +5 reward. If the agent observes "BA" (survive the first step and terminate at the second step), it receives +4 reward. For all other scenarios, the agent receives 0 reward. If we plan for 3 steps ahead from the beginning, "A" is the better action to take first. If we plan with consideration of termination probabilities, "B" is better. Because: The expected Q-value of "A" = 0.5 * 0 + 0.5 * max(0.5 * 0 + 0.5 * max(5, 0), 0) = 1.25 The expected Q-value of "B" = 0.1 * 0 + 0.9 * max(0.5 * 4 + 0.5 * max(0, 0), 0) = 1.8 """ import logging from collections import deque, defaultdict import numpy as np import torch from gym import Env from gym.spaces import Box, Discrete logger = logging.getLogger(__name__) MAX_STEP = 3 CHARACTERS = ["A", "B"] STATE_DIM = ACTION_DIM = len(CHARACTERS) class StringGameEnvV1(Env): def __init__(self, max_steps=MAX_STEP): np.random.seed(123) torch.manual_seed(123) self.max_steps = max_steps self.reward_map = defaultdict(float) self.terminal_probs = defaultdict(float) self._init_reward_and_terminal_probs() self.recent_actions = deque([], maxlen=MAX_STEP) self.action_space = Discrete(ACTION_DIM) self.observation_space = Box(low=0, high=1, shape=(STATE_DIM,)) self.step_cnt = 0 self.reset() def _init_reward_and_terminal_probs(self): self.reward_map["AAA"] = 5.0 self.reward_map["BA"] = 4.0 self.terminal_probs["A"] = 0.5 self.terminal_probs["B"] = 0.1 def seed(self, seed=None): np.random.seed(seed) torch.manual_seed(seed) @staticmethod def random_action(): return np.random.randint(0, ACTION_DIM) def get_reward(self): """ The function you can write to customize rewards. In this specific environment, the reward only depends on action history """ recent_characters = [CHARACTERS[c] for c in list(self.recent_actions)] string = "".join(recent_characters) if not self.done: reward = 0 else: reward = self.reward_map[string] return reward, string def step(self, action): assert self.action_space.contains(action) assert self.done is False self.step_cnt += 1 self.recent_actions.append(action) if self.step_cnt >= self.max_steps: self.done = True else: self.done = self.sample_terminal(action) reward, info = self.get_reward() ob = self.get_observation() return ob, reward, self.done, {"reward_str": info} def sample_terminal(self, action): terminal_probability = self.terminal_probs[CHARACTERS[action]] if np.random.rand() < terminal_probability: return True return False def get_observation(self): """ The function you can write to customize transitions. In this specific environment, the next state is exactly the latest action taken. The initial observation is all zeros. """ ob = np.zeros(STATE_DIM) if len(self.recent_actions) > 0: ob[self.recent_actions[-1]] = 1 return ob def reset(self): self.done = False self.recent_actions = deque([], maxlen=MAX_STEP) self.step_cnt = 0 ob = self.get_observation() return ob def print_internal_state(self): action_str = "".join([CHARACTERS[c] for c in self.recent_actions]) logger.debug( f"Step {self.step_cnt}, recent actions {action_str}, terminal={self.done}" ) @staticmethod def print_ob(ob): return str(ob) @staticmethod def print_action(action): return CHARACTERS[action]
4,664
33.051095
91
py
ReAgent
ReAgent-master/reagent/gym/envs/pomdp/state_embed_env.py
#!/usr/bin/env python3 # Copyright (c) Facebook, Inc. and its affiliates. All rights reserved. """ This file shows an example of using embedded states to feed to RL models in partially observable environments (POMDPs). Embedded states are generated by a world model which learns how to encode past n observations into a low-dimension vector.Embedded states improve performance in POMDPs compared to just using one-step observations as states because they encode more historical information than one-step observations. """ import logging from collections import deque from typing import Optional import gym import numpy as np import reagent.core.types as rlt import torch from gym.spaces import Box from reagent.gym.envs import EnvWrapper from reagent.models.world_model import MemoryNetwork logger = logging.getLogger(__name__) class StateEmbedEnvironment(gym.Env): def __init__( self, gym_env: EnvWrapper, mdnrnn: MemoryNetwork, max_embed_seq_len: int, state_min_value: Optional[float] = None, state_max_value: Optional[float] = None, ): self.env = gym_env self.unwrapped.spec = self.env.unwrapped.spec self.max_embed_seq_len = max_embed_seq_len self.mdnrnn = mdnrnn self.embed_size = self.mdnrnn.num_hiddens self.raw_state_dim = self.env.observation_space.shape[0] # type: ignore self.state_dim = self.embed_size + self.raw_state_dim if isinstance(self.env.action_space, gym.spaces.Discrete): self.is_discrete_action = True self.action_dim = self.env.action_space.n elif isinstance(self.env.action_space, gym.spaces.Box): self.is_discrete_action = False self.action_dim = self.env.action_space.shape[0] self.action_space = self.env.action_space # only need to set up if needed if state_min_value is None or state_max_value is None: state_min_value = np.min(gym_env.observation_space.low) state_max_value = np.max(gym_env.observation_space.high) self.observation_space = Box( # type: ignore low=state_min_value, high=state_max_value, shape=(self.state_dim,) ) self.cur_raw_state = None self.recent_states = deque([], maxlen=self.max_embed_seq_len) # type: ignore self.recent_actions = deque([], maxlen=self.max_embed_seq_len) # type: ignore def seed(self, seed): self.env.seed(seed) def __getattr__(self, name): return getattr(self.env, name) @torch.no_grad() def embed_state(self, state): """Embed state after either reset() or step()""" assert len(self.recent_states) == len(self.recent_actions) old_mdnrnn_mode = self.mdnrnn.mdnrnn.training self.mdnrnn.mdnrnn.eval() # Embed the state as the hidden layer's output # until the previous step + current state if len(self.recent_states) == 0: mdnrnn_state = np.zeros((1, self.raw_state_dim)) mdnrnn_action = np.zeros((1, self.action_dim)) else: mdnrnn_state = np.array(list(self.recent_states)) mdnrnn_action = np.array(list(self.recent_actions)) mdnrnn_state = torch.tensor(mdnrnn_state, dtype=torch.float).unsqueeze(1) mdnrnn_action = torch.tensor(mdnrnn_action, dtype=torch.float).unsqueeze(1) mdnrnn_output = self.mdnrnn( rlt.FeatureData(mdnrnn_state), rlt.FeatureData(mdnrnn_action) ) hidden_embed = ( mdnrnn_output.all_steps_lstm_hidden[-1].squeeze().detach().cpu().numpy() ) state_embed = np.hstack((hidden_embed, state)) self.mdnrnn.mdnrnn.train(old_mdnrnn_mode) logger.debug( f"Embed_state\nrecent states: {np.array(self.recent_states)}\n" f"recent actions: {np.array(self.recent_actions)}\n" f"state_embed{state_embed}\n" ) return state_embed def reset(self): next_raw_state = self.env.reset() self.recent_states = deque([], maxlen=self.max_embed_seq_len) self.recent_actions = deque([], maxlen=self.max_embed_seq_len) self.cur_raw_state = next_raw_state next_embed_state = self.embed_state(next_raw_state) return next_embed_state def step(self, action): if self.is_discrete_action: action_np = np.zeros(self.action_dim) action_np[action] = 1.0 else: action_np = action self.recent_states.append(self.cur_raw_state) self.recent_actions.append(action_np) next_raw_state, reward, terminal, info = self.env.step(action) logger.debug("action {}, reward {}\n".format(action, reward)) self.cur_raw_state = next_raw_state next_embed_state = self.embed_state(next_raw_state) return next_embed_state, reward, terminal, info
4,927
38.424
86
py
ReAgent
ReAgent-master/reagent/gym/envs/pomdp/string_game.py
#!/usr/bin/env python3 # Copyright (c) Facebook, Inc. and its affiliates. All rights reserved. """ The agent can observe a character at one time. But the reward is given based on last n (n>1) steps' observation (a string). In this environment, the agent can observe a character ("A", "B") at each time step, but the reward it receives actually depends on past 3 steps: if the agent observes "ABB" in the past 3 steps, it receives +5 reward; if the agent observes "BBB", it receives -5 reward; otherwise, the agent receives 0. The action is the next character the agent wants to reveal, and the next state is exactly the action just taken (i.e., the transition function only depends on the action). Each episode is limited to 6 steps. Therefore, the optimal policy is to choose actions "ABBABB" in sequence which results to +10 reward. """ import itertools import logging from collections import deque import numpy as np import torch from gym import Env from gym.spaces import Box, Discrete logger = logging.getLogger(__name__) MAX_STEP = 6 CHARACTERS = ["A", "B"] STATE_DIM = ACTION_DIM = len(CHARACTERS) SEQ_LEN = 3 class StringGameEnv(Env): def __init__(self, max_steps=MAX_STEP): np.random.seed(123) torch.manual_seed(123) self.max_steps = max_steps self.reward_map = {} self._init_reward() logger.debug(self.reward_map) self.recent_actions = deque([], maxlen=SEQ_LEN) self.recent_states = deque([], maxlen=SEQ_LEN) self.cur_state = None self.action_space = Discrete(ACTION_DIM) self.observation_space = Box(low=0, high=1, shape=(STATE_DIM,)) self.step_cnt = 0 self.reset() def _init_reward(self): for seq_len in range(1, SEQ_LEN + 1): for k in itertools.product(CHARACTERS, repeat=seq_len): self.reward_map["".join(k)] = 0 self.reward_map["ABB"] = 5 self.reward_map["BBB"] = -5 def seed(self, seed=None): np.random.seed(seed) torch.manual_seed(seed) @staticmethod def random_action(): return np.random.randint(0, ACTION_DIM) def get_reward(self): """ The function you can write to customize rewards. In this specific environment, the reward only depends on action history """ recent_characters = [CHARACTERS[c] for c in list(self.recent_actions)] string = "".join(recent_characters) reward = self.reward_map[string] return reward, string def step(self, action): assert self.action_space.contains(action) assert self.done is False self.step_cnt += 1 self.recent_states.append(self.cur_state) self.recent_actions.append(action) reward, info = self.get_reward() if self.step_cnt >= self.max_steps: self.done = True ob = self.get_observation() self.cur_state = ob return ob, reward, self.done, {"reward_str": info} def get_observation(self): """ The function you can write to customize transitions. In this specific environment, the next state is exactly the latest action taken. The initial observation is all zeros. """ ob = np.zeros(STATE_DIM) if len(self.recent_actions) > 0: ob[self.recent_actions[-1]] = 1 return ob def reset(self): self.done = False self.recent_states = deque([], maxlen=SEQ_LEN) self.recent_actions = deque([], maxlen=SEQ_LEN) self.step_cnt = 0 ob = self.get_observation() self.cur_state = ob return ob def print_internal_state(self): print("Step", self.step_cnt) def state_to_chr(s): state_index = np.nonzero(s)[0] if len(state_index) != 1: # initial state return "I" return CHARACTERS[state_index.item()] state_str = "".join([state_to_chr(s) for s in self.recent_states]) action_str = "".join([CHARACTERS[c] for c in self.recent_actions]) print( "Internal state: recent states {}, recent actions {}".format( state_str, action_str ) ) @staticmethod def print_ob(ob): return str(ob) @staticmethod def print_action(action): return CHARACTERS[action]
4,391
31.533333
80
py
ReAgent
ReAgent-master/reagent/gym/policies/predictor_policies.py
#!/usr/bin/env python3 # Copyright (c) Facebook, Inc. and its affiliates. All rights reserved. from typing import Optional, Tuple, Union import numpy as np import reagent.core.types as rlt import torch from reagent.core.fb_checker import IS_FB_ENVIRONMENT from reagent.core.parameters import RLParameters from reagent.gym.policies import Policy from reagent.gym.policies.samplers.discrete_sampler import ( GreedyActionSampler, SoftmaxActionSampler, ) from reagent.gym.policies.samplers.top_k_sampler import TopKSampler from reagent.gym.policies.scorers.discrete_scorer import ( discrete_dqn_serving_scorer, parametric_dqn_serving_scorer, ) from reagent.gym.policies.scorers.slate_q_scorer import slate_q_serving_scorer from reagent.models.actor import LOG_PROB_MIN, LOG_PROB_MAX if IS_FB_ENVIRONMENT: from reagent.fb.prediction.fb_predictor_wrapper import ( FbActorPredictorUnwrapper as ActorPredictorUnwrapper, FbDiscreteDqnPredictorUnwrapper as DiscreteDqnPredictorUnwrapper, FbParametricPredictorUnwrapper as ParametricDqnPredictorUnwrapper, ) else: from reagent.prediction.predictor_wrapper import ( ActorPredictorUnwrapper, DiscreteDqnPredictorUnwrapper, ParametricDqnPredictorUnwrapper, ) def create_predictor_policy_from_model(serving_module, **kwargs) -> Policy: """ serving_module is the result of ModelManager.build_serving_module(). This function creates a Policy for gym environments. """ module_name = serving_module.original_name if module_name.endswith("DiscreteDqnPredictorWrapper"): rl_parameters = kwargs.get("rl_parameters", None) return DiscreteDQNPredictorPolicy(serving_module, rl_parameters) elif module_name.endswith("ActorPredictorWrapper"): return ActorPredictorPolicy(predictor=ActorPredictorUnwrapper(serving_module)) elif module_name.endswith("ParametricDqnPredictorWrapper"): # TODO: remove this dependency max_num_actions = kwargs.get("max_num_actions", None) assert ( max_num_actions is not None ), f"max_num_actions not given for Parametric DQN." q_network = ParametricDqnPredictorUnwrapper(serving_module) # TODO: write SlateQ Wrapper slate_size = kwargs.get("slate_size", None) if slate_size is not None: scorer = slate_q_serving_scorer( num_candidates=max_num_actions, q_network=q_network ) sampler = TopKSampler(k=slate_size) else: sampler = GreedyActionSampler() scorer = parametric_dqn_serving_scorer( max_num_actions=max_num_actions, q_network=q_network ) return Policy(scorer=scorer, sampler=sampler) else: raise NotImplementedError( f"Predictor policy for serving module {serving_module} not available." ) class DiscreteDQNPredictorPolicy(Policy): def __init__(self, wrapped_dqn_predictor, rl_parameters: Optional[RLParameters]): if rl_parameters and rl_parameters.softmax_policy: self.sampler = SoftmaxActionSampler(temperature=rl_parameters.temperature) else: self.sampler = GreedyActionSampler() self.scorer = discrete_dqn_serving_scorer( q_network=DiscreteDqnPredictorUnwrapper(wrapped_dqn_predictor) ) # pyre-fixme[56]: Decorator `torch.no_grad(...)` could not be called, because # its type `no_grad` is not callable. @torch.no_grad() def act( self, obs: Union[rlt.ServingFeatureData, Tuple[torch.Tensor, torch.Tensor]], possible_actions_mask: Optional[torch.Tensor], ) -> rlt.ActorOutput: """Input is either state_with_presence, or ServingFeatureData (in the case of sparse features)""" assert isinstance(obs, tuple) if isinstance(obs, rlt.ServingFeatureData): state: rlt.ServingFeatureData = obs else: state = rlt.ServingFeatureData( float_features_with_presence=obs, id_list_features={}, id_score_list_features={}, ) scores = self.scorer(state, possible_actions_mask) return self.sampler.sample_action(scores).cpu().detach() class ActorPredictorPolicy(Policy): def __init__(self, predictor): self.predictor = predictor # pyre-fixme[56]: Decorator `torch.no_grad(...)` could not be called, because # its type `no_grad` is not callable. @torch.no_grad() def act( self, obs: Union[rlt.ServingFeatureData, Tuple[torch.Tensor, torch.Tensor]], possible_actions_mask: Optional[torch.Tensor] = None, ) -> rlt.ActorOutput: """Input is either state_with_presence, or ServingFeatureData (in the case of sparse features)""" assert isinstance(obs, tuple) if isinstance(obs, rlt.ServingFeatureData): state: rlt.ServingFeatureData = obs else: state = rlt.ServingFeatureData( float_features_with_presence=obs, id_list_features={}, id_score_list_features={}, ) output = self.predictor(*state) if isinstance(output, tuple): action, log_prob = output log_prob = log_prob.clamp(LOG_PROB_MIN, LOG_PROB_MAX) return rlt.ActorOutput(action=action.cpu(), log_prob=log_prob.cpu()) else: return rlt.ActorOutput(action=output.cpu())
5,542
38.592857
86
py
ReAgent
ReAgent-master/reagent/gym/policies/policy.py
#!/usr/bin/env python3 # Copyright (c) Facebook, Inc. and its affiliates. All rights reserved. from typing import Any, Optional import reagent.core.types as rlt import torch from reagent.gym.types import Sampler, Scorer class Policy: def __init__(self, scorer: Scorer, sampler: Sampler): """ The Policy composes the scorer and sampler to create actions. Args: scorer: given preprocessed input, outputs intermediate scores used for sampling actions sampler: given scores (from the scorer), samples an action. """ self.scorer = scorer self.sampler = sampler def act( self, obs: Any, possible_actions_mask: Optional[torch.Tensor] = None ) -> rlt.ActorOutput: """ Performs the composition described above. These are the actions being put into the replay buffer, not necessary the actions taken by the environment! """ scorer_inputs = (obs,) if possible_actions_mask is not None: scorer_inputs += (possible_actions_mask,) scores = self.scorer(*scorer_inputs) actor_output = self.sampler.sample_action(scores) return actor_output.cpu().detach()
1,245
31.789474
77
py
ReAgent
ReAgent-master/reagent/gym/policies/random_policies.py
#!/usr/bin/env python3 # Copyright (c) Facebook, Inc. and its affiliates. All rights reserved. from typing import List, Optional import gym import numpy as np import reagent.core.types as rlt import torch import torch.nn.functional as F from reagent.core.parameters import CONTINUOUS_TRAINING_ACTION_RANGE from reagent.gym.policies.policy import Policy from reagent.gym.policies.scorers.discrete_scorer import apply_possible_actions_mask def make_random_policy_for_env(env: gym.Env) -> Policy: if isinstance(env.action_space, gym.spaces.Discrete): # discrete action space return DiscreteRandomPolicy.create_for_env(env) elif isinstance(env.action_space, gym.spaces.Box): # continuous action space return ContinuousRandomPolicy.create_for_env(env) elif isinstance(env.action_space, gym.spaces.MultiDiscrete): return MultiDiscreteRandomPolicy.create_for_env(env) else: raise NotImplementedError(f"{env.action_space} not supported") class DiscreteRandomPolicy(Policy): def __init__(self, num_actions: int): """Random actor for accumulating random offline data.""" self.num_actions = num_actions @classmethod def create_for_env(cls, env: gym.Env): action_space = env.action_space if isinstance(action_space, gym.spaces.Discrete): return cls(num_actions=action_space.n) elif isinstance(action_space, gym.spaces.Box): raise NotImplementedError(f"Try continuous random policy instead") else: raise NotImplementedError(f"action_space is {type(action_space)}") def act( self, obs: rlt.FeatureData, possible_actions_mask: Optional[torch.Tensor] = None ) -> rlt.ActorOutput: """Act randomly regardless of the observation.""" # pyre-fixme[35]: Target cannot be annotated. obs: torch.Tensor = obs.float_features assert obs.dim() >= 2, f"obs has shape {obs.shape} (dim < 2)" assert obs.shape[0] == 1, f"obs has shape {obs.shape} (0th dim != 1)" batch_size = obs.shape[0] scores = torch.ones((batch_size, self.num_actions)) scores = apply_possible_actions_mask( scores, possible_actions_mask, invalid_score=0.0 ) # sample a random action m = torch.distributions.Categorical(scores) raw_action = m.sample() action = F.one_hot(raw_action, self.num_actions) log_prob = m.log_prob(raw_action).float() return rlt.ActorOutput(action=action, log_prob=log_prob) class MultiDiscreteRandomPolicy(Policy): def __init__(self, num_action_vec: List[int]): self.num_action_vec = num_action_vec self.dists = [ torch.distributions.Categorical(torch.ones(n) / n) for n in self.num_action_vec ] @classmethod def create_for_env(cls, env: gym.Env): action_space = env.action_space if not isinstance(action_space, gym.spaces.MultiDiscrete): raise ValueError(f"Invalid action space: {action_space}") return cls(action_space.nvec.tolist()) # TODO: consider possible_actions_mask def act( self, obs: rlt.FeatureData, possible_actions_mask: Optional[torch.Tensor] = None ) -> rlt.ActorOutput: # pyre-fixme[35]: Target cannot be annotated. obs: torch.Tensor = obs.float_features batch_size, _ = obs.shape actions = [] log_probs = [] for m in self.dists: actions.append(m.sample((batch_size, 1))) log_probs.append(m.log_prob(actions[-1]).float()) return rlt.ActorOutput( action=torch.cat(actions, dim=1), log_prob=torch.cat(log_probs, dim=1).sum(1, keepdim=True), ) class ContinuousRandomPolicy(Policy): def __init__(self, low: torch.Tensor, high: torch.Tensor): self.low = low self.high = high assert ( low.shape == high.shape ), f"low.shape = {low.shape}, high.shape = {high.shape}" self.dist = torch.distributions.uniform.Uniform(self.low, self.high) @classmethod def create_for_env(cls, env: gym.Env): action_space = env.action_space if isinstance(action_space, gym.spaces.Discrete): raise NotImplementedError( f"Action space is discrete. Try using DiscreteRandomPolicy instead." ) elif isinstance(action_space, gym.spaces.Box): assert ( len(action_space.shape) == 1 ), f"Box with shape {action_space.shape} not supported." low, high = CONTINUOUS_TRAINING_ACTION_RANGE # broadcast low and high to shape np_ones = np.ones(action_space.shape) return cls( low=torch.tensor(low * np_ones), high=torch.tensor(high * np_ones) ) else: raise NotImplementedError(f"action_space is {type(action_space)}") def act( self, obs: rlt.FeatureData, possible_actions_mask: Optional[torch.Tensor] = None ) -> rlt.ActorOutput: """Act randomly regardless of the observation.""" # pyre-fixme[35]: Target cannot be annotated. obs: torch.Tensor = obs.float_features assert obs.dim() >= 2, f"obs has shape {obs.shape} (dim < 2)" batch_size = obs.size(0) # pyre-fixme[6]: Expected `Union[torch.Size, torch.Tensor]` for 1st param # but got `Tuple[int]`. action = self.dist.sample((batch_size,)) # sum over action_dim (since assuming i.i.d. per coordinate) log_prob = self.dist.log_prob(action).sum(1) return rlt.ActorOutput(action=action, log_prob=log_prob)
5,717
38.434483
88
py
ReAgent
ReAgent-master/reagent/gym/policies/scorers/slate_q_scorer.py
#!/usr/bin/env python3 # Copyright (c) Facebook, Inc. and its affiliates. All rights reserved. import reagent.core.types as rlt import torch import torch.nn.functional as F from reagent.gym.types import Scorer from reagent.models.base import ModelBase def slate_q_scorer(num_candidates: int, q_network: ModelBase) -> Scorer: @torch.no_grad() def score(state: rlt.FeatureData) -> torch.Tensor: tiled_state = state.repeat_interleave(repeats=num_candidates, axis=0) candidate_docs = state.candidate_docs assert candidate_docs is not None actions = candidate_docs.as_feature_data() q_network.eval() scores = q_network(tiled_state, actions).view(-1, num_candidates) q_network.train() select_prob = F.softmax(candidate_docs.value, dim=1) assert select_prob.shape == scores.shape return select_prob * scores return score def slate_q_serving_scorer(num_candidates: int, q_network: torch.nn.Module) -> Scorer: @torch.no_grad() def score(state: rlt.FeatureData) -> torch.Tensor: # pyre-fixme[28]: Unexpected keyword argument `axis`. tiled_state = state.float_features.repeat_interleave( repeats=num_candidates, axis=0 ) candidate_docs = state.candidate_docs assert candidate_docs is not None actions = candidate_docs.as_feature_data().float_features q_network.eval() action_names, q_values = q_network( (tiled_state, torch.ones_like(tiled_state)), (actions, torch.ones_like(actions)), ) scores = q_values.view(-1, num_candidates) q_network.train() select_prob = F.softmax(candidate_docs.value, dim=1) assert select_prob.shape == scores.shape return select_prob * scores return score
1,838
31.839286
86
py
ReAgent
ReAgent-master/reagent/gym/policies/scorers/continuous_scorer.py
#!/usr/bin/env python3 # Copyright (c) Facebook, Inc. and its affiliates. All rights reserved. import reagent.core.types as rlt import torch from reagent.gym.types import GaussianSamplerScore, Scorer from reagent.models.base import ModelBase def sac_scorer(actor_network: ModelBase) -> Scorer: @torch.no_grad() def score(preprocessed_obs: rlt.FeatureData) -> GaussianSamplerScore: actor_network.eval() # pyre-fixme[29]: `Union[torch.Tensor, torch.nn.Module]` is not a function. loc, scale_log = actor_network._get_loc_and_scale_log(preprocessed_obs) actor_network.train() return GaussianSamplerScore(loc=loc, scale_log=scale_log) return score
699
34
83
py
ReAgent
ReAgent-master/reagent/gym/policies/scorers/discrete_scorer.py
#!/usr/bin/env python3 # Copyright (c) Facebook, Inc. and its affiliates. All rights reserved. from typing import Optional, Tuple import numpy as np import reagent.core.types as rlt import torch from reagent.gym.preprocessors.trainer_preprocessor import get_possible_actions_for_gym from reagent.gym.types import Scorer from reagent.models.base import ModelBase NEG_INF = float("-inf") def apply_possible_actions_mask( scores: torch.Tensor, possible_actions_mask: Optional[torch.Tensor] = None, invalid_score: float = NEG_INF, ) -> torch.Tensor: if possible_actions_mask is None: return scores possible_actions_mask = possible_actions_mask.unsqueeze(0) assert ( scores.shape == possible_actions_mask.shape ), f"{scores.shape} != {possible_actions_mask.shape}" scores[~possible_actions_mask] = invalid_score # pyre-ignore[16] return scores def discrete_dqn_scorer(q_network: ModelBase) -> Scorer: @torch.no_grad() def score( preprocessed_obs: rlt.FeatureData, possible_actions_mask: Optional[torch.Tensor] = None, ) -> torch.Tensor: q_network.eval() scores = q_network(preprocessed_obs) # qrdqn returns (batchsize, num_actions, num_atoms) if scores.dim() == 3: scores = scores.mean(dim=2) assert scores.dim() == 2, f"{scores.shape} isn't (batchsize, num_actions)." q_network.train() scores = apply_possible_actions_mask(scores, possible_actions_mask) return scores return score def discrete_dqn_serving_scorer(q_network: torch.nn.Module) -> Scorer: @torch.no_grad() def score( state: rlt.ServingFeatureData, possible_actions_mask: Optional[torch.Tensor] = None, ) -> torch.Tensor: action_names, q_values = q_network(*state) q_values = apply_possible_actions_mask(q_values, possible_actions_mask) return q_values return score def get_parametric_input(max_num_actions: int, obs: rlt.FeatureData): assert ( len(obs.float_features.shape) == 2 ), f"{obs.float_features.shape} is not (batch_size, state_dim)." batch_size, _ = obs.float_features.shape possible_actions = get_possible_actions_for_gym(batch_size, max_num_actions).to( obs.float_features.device ) return obs.get_tiled_batch(max_num_actions), possible_actions def parametric_dqn_scorer(max_num_actions: int, q_network: ModelBase) -> Scorer: @torch.no_grad() def score(preprocessed_obs: rlt.FeatureData) -> torch.Tensor: tiled_state, possible_actions = get_parametric_input( max_num_actions, preprocessed_obs ) q_network.eval() scores = q_network(tiled_state, possible_actions) q_network.train() return scores.view(-1, max_num_actions) return score def parametric_dqn_serving_scorer( max_num_actions: int, q_network: torch.nn.Module ) -> Scorer: @torch.no_grad() def score(value_presence: Tuple[torch.Tensor, torch.Tensor]) -> torch.Tensor: obs = value_presence[0] tiled_state, possible_actions = get_parametric_input( max_num_actions, rlt.FeatureData(obs) ) tiled_state_tensor = tiled_state.float_features possible_actions_tensor = possible_actions.float_features action_names, q_values = q_network( state_with_presence=( tiled_state_tensor, torch.ones_like(tiled_state_tensor), ), action_with_presence=( possible_actions_tensor, torch.ones_like(possible_actions_tensor), ), ) return q_values.view(-1, max_num_actions) return score
3,740
32.106195
87
py
ReAgent
ReAgent-master/reagent/gym/policies/samplers/continuous_sampler.py
#!/usr/bin/env python3 # Copyright (c) Facebook, Inc. and its affiliates. All rights reserved. import reagent.core.types as rlt import torch from reagent.gym.types import GaussianSamplerScore, Sampler class GaussianSampler(Sampler): def __init__(self, actor_network): self.actor_network = actor_network def _sample_action(self, loc: torch.Tensor, scale_log: torch.Tensor): r = torch.randn_like(scale_log, device=scale_log.device) # pyre-fixme[16]: `Tensor` has no attribute `exp`. action = torch.tanh(loc + r * scale_log.exp()) # Since each dim are independent, log-prob is simply sum log_prob = self.actor_network._log_prob(r, scale_log) squash_correction = self.actor_network._squash_correction(action) log_prob = torch.sum(log_prob - squash_correction, dim=1) return action, log_prob # pyre-fixme[56]: Decorator `torch.no_grad(...)` could not be called, because # its type `no_grad` is not callable. @torch.no_grad() def sample_action(self, scores: GaussianSamplerScore) -> rlt.ActorOutput: self.actor_network.eval() unscaled_actions, log_prob = self._sample_action(scores.loc, scores.scale_log) self.actor_network.train() return rlt.ActorOutput(action=unscaled_actions, log_prob=log_prob) def _log_prob( self, loc: torch.Tensor, scale_log: torch.Tensor, squashed_action: torch.Tensor ): # This is not getting exported; we can use it # pyre-fixme[16]: `Tensor` has no attribute `exp`. n = torch.distributions.Normal(loc, scale_log.exp()) raw_action = self.actor_network._atanh(squashed_action) log_prob = n.log_prob(raw_action) squash_correction = self.actor_network._squash_correction(squashed_action) log_prob = torch.sum(log_prob - squash_correction, dim=1) return log_prob # pyre-fixme[56]: Decorator `torch.no_grad(...)` could not be called, because # its type `no_grad` is not callable. @torch.no_grad() def log_prob( self, scores: GaussianSamplerScore, squashed_action: torch.Tensor ) -> torch.Tensor: self.actor_network.eval() # pyre-fixme[20]: Argument `squashed_action` expected. log_prob = self._log_prob(scores.loc, scores.scale_log) self.actor_network.train() return log_prob
2,378
41.482143
87
py
ReAgent
ReAgent-master/reagent/gym/policies/samplers/top_k_sampler.py
#!/usr/bin/env python3 # Copyright (c) Facebook, Inc. and its affiliates. All rights reserved. import reagent.core.types as rlt import torch from reagent.gym.types import Sampler class TopKSampler(Sampler): def __init__(self, k: int): self.k = k def sample_action(self, scores: torch.Tensor) -> rlt.ActorOutput: top_values, item_idxs = torch.topk(scores, self.k, dim=1) return rlt.ActorOutput( action=item_idxs, log_prob=torch.zeros(item_idxs.shape[0], 1) ) def log_prob(self, scores: torch.Tensor, action: torch.Tensor) -> torch.Tensor: raise NotImplementedError
634
27.863636
83
py
ReAgent
ReAgent-master/reagent/gym/policies/samplers/discrete_sampler.py
#!/usr/bin/env python3 # Copyright (c) Facebook, Inc. and its affiliates. All rights reserved. import reagent.core.types as rlt import torch import torch.nn.functional as F from reagent.gym.types import Sampler from reagent.models.dqn import INVALID_ACTION_CONSTANT class SoftmaxActionSampler(Sampler): """ Softmax sampler. Equation: http://incompleteideas.net/book/first/ebook/node17.html The action scores are logits. Supports decaying the temperature over time. Args: temperature: A measure of how uniformly random the distribution looks. The higher the temperature, the more uniform the sampling. temperature_decay: A multiplier by which temperature is reduced at each .update() call minimum_temperature: Minimum temperature, below which the temperature is not decayed further """ def __init__( self, temperature: float = 1.0, temperature_decay: float = 1.0, minimum_temperature: float = 0.1, ): assert temperature > 0, f"Invalid non-positive temperature {temperature}." self.temperature = temperature self.temperature_decay = temperature_decay self.minimum_temperature = minimum_temperature assert ( temperature_decay <= 1.0 ), f"Invalid temperature_decay>1: {temperature_decay}." assert ( minimum_temperature <= temperature ), f"minimum_temperature ({minimum_temperature}) exceeds initial temperature ({temperature})" def _get_distribution( self, scores: torch.Tensor ) -> torch.distributions.Categorical: return torch.distributions.Categorical(logits=scores / self.temperature) # pyre-fixme[56]: Decorator `torch.no_grad(...)` could not be called, because # its type `no_grad` is not callable. @torch.no_grad() def sample_action(self, scores: torch.Tensor) -> rlt.ActorOutput: assert ( scores.dim() == 2 ), f"scores shape is {scores.shape}, not (batch_size, num_actions)" batch_size, num_actions = scores.shape m = self._get_distribution(scores) raw_action = m.sample() assert raw_action.shape == ( batch_size, ), f"{raw_action.shape} != ({batch_size}, )" action = F.one_hot(raw_action, num_actions) assert action.ndim == 2 log_prob = m.log_prob(raw_action) assert log_prob.ndim == 1 return rlt.ActorOutput(action=action, log_prob=log_prob) def log_prob(self, scores: torch.Tensor, action: torch.Tensor) -> torch.Tensor: assert len(scores.shape) == 2, f"{scores.shape}" assert scores.shape == action.shape, f"{scores.shape} != {action.shape}" m = self._get_distribution(scores) # pyre-fixme[16]: `Tensor` has no attribute `argmax`. return m.log_prob(action.argmax(dim=1)) def entropy(self, scores: torch.Tensor) -> torch.Tensor: """ Returns average policy entropy. Simple unweighted average across the batch. """ assert len(scores.shape) == 2, f"{scores.shape}" m = self._get_distribution(scores) return m.entropy().mean() def update(self) -> None: self.temperature *= self.temperature_decay self.temperature = max(self.temperature, self.minimum_temperature) class GreedyActionSampler(Sampler): """ Return the highest scoring action. """ def _get_greedy_indices(self, scores: torch.Tensor) -> torch.Tensor: assert ( len(scores.shape) == 2 ), f"scores shape is {scores.shape}, not (batchsize, num_actions)" # pyre-fixme[16]: `Tensor` has no attribute `argmax`. return scores.argmax(dim=1) # pyre-fixme[56]: Decorator `torch.no_grad(...)` could not be called, because # its type `no_grad` is not callable. @torch.no_grad() def sample_action(self, scores: torch.Tensor) -> rlt.ActorOutput: batch_size, num_actions = scores.shape raw_action = self._get_greedy_indices(scores) action = F.one_hot(raw_action, num_actions) assert action.shape == (batch_size, num_actions) return rlt.ActorOutput( action=action, log_prob=torch.zeros_like(raw_action, dtype=torch.float) ) # pyre-fixme[56]: Decorator `torch.no_grad(...)` could not be called, because # its type `no_grad` is not callable. @torch.no_grad() def log_prob(self, scores: torch.Tensor, action: torch.Tensor) -> torch.Tensor: greedy_indices = self._get_greedy_indices(scores) # pyre-fixme[16]: `Tensor` has no attribute `argmax`. match = greedy_indices == action.argmax(-1) lp = torch.zeros(scores.shape[0]).float() lp[match] = -float("inf") return lp class EpsilonGreedyActionSampler(Sampler): """ Epsilon-Greedy Policy With probability epsilon, a random action is sampled. Otherwise, the highest scoring (greedy) action is chosen. Call update() to decay the amount of exploration by lowering `epsilon` by a factor of `epsilon_decay` (<=1) until we reach `minimum_epsilon` """ def __init__( self, epsilon: float, epsilon_decay: float = 1.0, minimum_epsilon: float = 0.0 ): self.epsilon = float(epsilon) assert epsilon_decay <= 1 self.epsilon_decay = epsilon_decay assert minimum_epsilon <= epsilon_decay self.minimum_epsilon = minimum_epsilon def sample_action(self, scores: torch.Tensor) -> rlt.ActorOutput: assert scores.dim() == 2, ( "scores dim is %d" % scores.dim() ) # batch_size x num_actions batch_size, num_actions = scores.shape # pyre-fixme[16]: `Tensor` has no attribute `argmax`. argmax = F.one_hot(scores.argmax(dim=1), num_actions).bool() valid_actions_ind = (scores > INVALID_ACTION_CONSTANT).bool() num_valid_actions = valid_actions_ind.float().sum(1, keepdim=True) rand_prob = self.epsilon / num_valid_actions p = torch.full_like(scores, rand_prob) greedy_prob = 1 - self.epsilon + rand_prob p[argmax] = greedy_prob.squeeze() p[~valid_actions_ind] = 0.0 assert torch.isclose(p.sum(1) == torch.ones(p.shape[0])) m = torch.distributions.Categorical(probs=p) raw_action = m.sample() action = F.one_hot(raw_action, num_actions) assert action.shape == (batch_size, num_actions) log_prob = m.log_prob(raw_action) assert log_prob.shape == (batch_size,) return rlt.ActorOutput(action=action, log_prob=log_prob) def log_prob(self, scores: torch.Tensor, action: torch.Tensor) -> torch.Tensor: max_index = self.sample_action(scores).argmax(-1) # pyre-fixme[16]: `Tensor` has no attribute `argmax`. opt = max_index == action.argmax(-1) n = len(scores) lp = torch.ones(n) * self.epsilon / n lp[opt] = 1 - self.epsilon + self.epsilon / n return lp def update(self) -> None: self.epsilon *= self.epsilon_decay if self.minimum_epsilon is not None: self.epsilon = max(self.epsilon, self.minimum_epsilon)
7,223
37.425532
101
py
ReAgent
ReAgent-master/reagent/gym/datasets/replay_buffer_dataset.py
#!/usr/bin/env python3 import logging from typing import Optional, Callable import torch from reagent.gym.agents.agent import Agent from reagent.gym.envs import EnvWrapper from reagent.gym.preprocessors import ( make_replay_buffer_inserter, make_replay_buffer_trainer_preprocessor, ) from reagent.gym.types import Transition, Trajectory from reagent.replay_memory.circular_replay_buffer import ReplayBuffer logger = logging.getLogger(__name__) class ReplayBufferDataset(torch.utils.data.IterableDataset): def __init__( self, env: EnvWrapper, agent: Agent, replay_buffer: ReplayBuffer, batch_size: int, training_frequency: int = 1, num_episodes: Optional[int] = None, max_steps: Optional[int] = None, post_episode_callback: Optional[Callable] = None, trainer_preprocessor=None, replay_buffer_inserter=None, ): super().__init__() self._env = env self._agent = agent self._replay_buffer = replay_buffer self._batch_size = batch_size self._training_frequency = training_frequency self._num_episodes = num_episodes self._max_steps = max_steps self._post_episode_callback = post_episode_callback self._trainer_preprocessor = trainer_preprocessor assert replay_buffer_inserter is not None self._replay_buffer_inserter = replay_buffer_inserter # TODO: Just use kwargs here? @classmethod def create_for_trainer( cls, trainer, env: EnvWrapper, agent: Agent, replay_buffer: ReplayBuffer, batch_size: int, training_frequency: int = 1, num_episodes: Optional[int] = None, max_steps: Optional[int] = None, post_episode_callback: Optional[Callable] = None, trainer_preprocessor=None, replay_buffer_inserter=None, device=None, ): device = device or torch.device("cpu") if trainer_preprocessor is None: trainer_preprocessor = make_replay_buffer_trainer_preprocessor( trainer, device, env ) if replay_buffer_inserter is None: replay_buffer_inserter = make_replay_buffer_inserter(env) return cls( env=env, agent=agent, replay_buffer=replay_buffer, batch_size=batch_size, training_frequency=training_frequency, num_episodes=num_episodes, max_steps=max_steps, post_episode_callback=post_episode_callback, trainer_preprocessor=trainer_preprocessor, replay_buffer_inserter=replay_buffer_inserter, ) def __iter__(self): mdp_id = 0 global_num_steps = 0 rewards = [] # TODO: We probably should put member vars into local vars to # reduce indirection, improving perf while self._num_episodes is None or mdp_id < self._num_episodes: obs = self._env.reset() possible_actions_mask = self._env.possible_actions_mask terminal = False num_steps = 0 episode_reward_sum = 0 trajectory = Trajectory() while not terminal: action, log_prob = self._agent.act(obs, possible_actions_mask) next_obs, reward, terminal, info = self._env.step(action) next_possible_actions_mask = self._env.possible_actions_mask if self._max_steps is not None and num_steps >= self._max_steps: terminal = True # Only partially filled. Agent can fill in more fields. transition = Transition( mdp_id=mdp_id, sequence_number=num_steps, observation=obs, action=action, reward=float(reward), terminal=bool(terminal), log_prob=log_prob, possible_actions_mask=possible_actions_mask, ) trajectory.add_transition(transition) self._replay_buffer_inserter(self._replay_buffer, transition) episode_reward_sum += reward if ( global_num_steps % self._training_frequency == 0 and self._replay_buffer.size >= self._batch_size ): train_batch = self._replay_buffer.sample_transition_batch( batch_size=self._batch_size ) if self._trainer_preprocessor: train_batch = self._trainer_preprocessor(train_batch) yield train_batch obs = next_obs possible_actions_mask = next_possible_actions_mask num_steps += 1 global_num_steps += 1 if self._post_episode_callback: self._post_episode_callback(trajectory, info) rewards.append(episode_reward_sum) mdp_id += 1 logger.info( f"Training episode: {mdp_id}, total episode reward = {episode_reward_sum}" ) logger.info(f"Episode rewards during training: {rewards}") class OfflineReplayBufferDataset(torch.utils.data.IterableDataset): """ Simply sampling from the replay buffer """ def __init__( self, env: EnvWrapper, replay_buffer: ReplayBuffer, batch_size: int, num_batches: int, trainer_preprocessor=None, ): super().__init__() self._env = env self._replay_buffer = replay_buffer self._batch_size = batch_size self._num_batches = num_batches self._trainer_preprocessor = trainer_preprocessor # TODO: Just use kwargs here? @classmethod def create_for_trainer( cls, trainer, env: EnvWrapper, replay_buffer: ReplayBuffer, batch_size: int, num_batches: int, trainer_preprocessor=None, device=None, ): device = device or torch.device("cpu") if trainer_preprocessor is None: trainer_preprocessor = make_replay_buffer_trainer_preprocessor( trainer, device, env ) return cls( env=env, replay_buffer=replay_buffer, batch_size=batch_size, num_batches=num_batches, trainer_preprocessor=trainer_preprocessor, ) def __iter__(self): for _ in range(self._num_batches): train_batch = self._replay_buffer.sample_transition_batch( batch_size=self._batch_size ) if self._trainer_preprocessor: train_batch = self._trainer_preprocessor(train_batch) yield train_batch
6,916
33.242574
90
py
ReAgent
ReAgent-master/reagent/gym/datasets/episodic_dataset.py
#!/usr/bin/env python3 # Copyright (c) Facebook, Inc. and its affiliates. All rights reserved. import logging from typing import Optional import torch from reagent.gym.agents.agent import Agent from reagent.gym.envs.gym import Gym from reagent.gym.runners.gymrunner import run_episode logger = logging.getLogger(__name__) class EpisodicDataset(torch.utils.data.IterableDataset): def __init__( self, env: Gym, agent: Agent, num_episodes: int, seed: int = 0, max_steps: Optional[int] = None, ): self.env = env self.agent = agent self.num_episodes = num_episodes self.seed = seed self.max_steps = max_steps def __iter__(self): self.env.reset() for i in range(self.num_episodes): trajectory = run_episode( self.env, self.agent, max_steps=self.max_steps, mdp_id=i ) yield trajectory.to_dict() def __len__(self): return self.num_episodes
1,020
23.902439
72
py
ReAgent
ReAgent-master/reagent/gym/agents/agent.py
#!/usr/bin/env python3 # Copyright (c) Facebook, Inc. and its affiliates. All rights reserved. from typing import Any, Optional, Tuple, Union import numpy as np import torch from reagent.gym.envs.env_wrapper import EnvWrapper from reagent.gym.policies.policy import Policy from reagent.gym.policies.random_policies import make_random_policy_for_env from reagent.gym.types import PostEpisode, PostStep, Trajectory, Transition def _id(x): return x class Agent: def __init__( self, policy: Policy, post_transition_callback: Optional[PostStep] = None, post_episode_callback: Optional[PostEpisode] = None, obs_preprocessor=_id, action_extractor=_id, device: Optional[torch.device] = None, ): """ The Agent orchestrates the interactions on our RL components, given the interactions with the environment. Args: policy: Policy that acts given preprocessed input action_preprocessor: preprocesses action for environment post_step: called after env.step(action). Default post_step is to do nothing. """ device = device or torch.device("cpu") self.policy = policy self.obs_preprocessor = obs_preprocessor self.action_extractor = action_extractor self.post_transition_callback = post_transition_callback self.post_episode_callback = post_episode_callback self.device = device @classmethod def create_for_env( cls, env: EnvWrapper, policy: Optional[Policy], *, device: Union[str, torch.device] = "cpu", obs_preprocessor=None, action_extractor=None, **kwargs, ): """ If `policy` is not given, we will try to create a random policy """ if isinstance(device, str): device = torch.device(device) if obs_preprocessor is None: obs_preprocessor = env.get_obs_preprocessor(device=device) if action_extractor is None: action_extractor = env.get_action_extractor() if policy is None: policy = make_random_policy_for_env(env) return cls( policy, obs_preprocessor=obs_preprocessor, action_extractor=action_extractor, device=device, **kwargs, ) @classmethod def create_for_env_with_serving_policy( cls, env: EnvWrapper, serving_policy: Policy, *, obs_preprocessor=None, action_extractor=None, **kwargs, ): # device shouldn't be provided as serving is CPU only if obs_preprocessor is None: obs_preprocessor = env.get_serving_obs_preprocessor() if action_extractor is None: action_extractor = env.get_serving_action_extractor() return cls( serving_policy, obs_preprocessor=obs_preprocessor, action_extractor=action_extractor, **kwargs, ) def act( self, obs: Any, possible_actions_mask: Optional[np.ndarray] = None ) -> Tuple[Any, Optional[float]]: """Act on a single observation""" # preprocess and convert to batch data preprocessed_obs = self.obs_preprocessor(obs) if possible_actions_mask is not None: possible_actions_mask = torch.tensor( possible_actions_mask, device=self.device ) # store intermediate actor output for post_step actor_output = self.policy.act(preprocessed_obs, possible_actions_mask) log_prob = actor_output.log_prob if log_prob is not None: log_prob = log_prob.cpu().squeeze(0).item() return self.action_extractor(actor_output), log_prob def post_step(self, transition: Transition): """to be called after step(action)""" if self.post_transition_callback is not None: self.post_transition_callback(transition) def post_episode(self, trajectory: Trajectory): """to be called after step(action)""" if self.post_episode_callback is not None: self.post_episode_callback(trajectory)
4,248
31.435115
79
py
ReAgent
ReAgent-master/reagent/gym/tests/test_gym_offline.py
#!/usr/bin/env python3 # Copyright (c) Facebook, Inc. and its affiliates. All rights reserved. import logging import os import pprint import unittest import uuid import numpy as np import pytest import pytorch_lightning as pl import torch from parameterized import parameterized from reagent.gym.agents.agent import Agent from reagent.gym.datasets.replay_buffer_dataset import OfflineReplayBufferDataset from reagent.gym.envs import Gym from reagent.gym.policies.random_policies import make_random_policy_for_env from reagent.gym.runners.gymrunner import evaluate_for_n_episodes from reagent.gym.utils import build_normalizer, fill_replay_buffer from reagent.model_managers.union import ModelManager__Union from reagent.replay_memory.circular_replay_buffer import ReplayBuffer from reagent.test.base.horizon_test_base import HorizonTestBase from reagent.workflow.types import RewardOptions # for seeding the environment SEED = 0 logger = logging.getLogger(__name__) logger.setLevel(logging.INFO) """ These are trained offline. """ GYM_TESTS = [ ("CEM Cartpole", "configs/world_model/cem_cartpole_offline.yaml"), ( "CEM Single World Model Linear Dynamics", "configs/world_model/cem_single_world_model_linear_dynamics_offline.yaml", ), ( "CEM Many World Models Linear Dynamics", "configs/world_model/cem_many_world_models_linear_dynamics_offline.yaml", ), ] curr_dir = os.path.dirname(__file__) class TestGymOffline(HorizonTestBase): # pyre-fixme[16]: Module `parameterized` has no attribute `expand`. @parameterized.expand(GYM_TESTS) def test_gym_offline_cpu(self, name: str, config_path: str): self.run_from_config( run_test=run_test_offline, config_path=os.path.join(curr_dir, config_path), use_gpu=False, ) logger.info(f"{name} passes!") # pyre-fixme[16]: Module `parameterized` has no attribute `expand`. @parameterized.expand(GYM_TESTS) @pytest.mark.serial @unittest.skipIf(not torch.cuda.is_available(), "CUDA not available") def test_gym_offline_gpu(self, name: str, config_path: str): self.run_from_config( run_test=run_test_offline, config_path=os.path.join(curr_dir, config_path), use_gpu=True, ) logger.info(f"{name} passes!") def evaluate_cem(env, manager, trainer_module, num_eval_episodes: int): # NOTE: for CEM, serving isn't implemented policy = manager.create_policy(trainer_module, serving=False) agent = Agent.create_for_env(env, policy) return evaluate_for_n_episodes( n=num_eval_episodes, env=env, agent=agent, max_steps=env.max_steps ) def identity_collate(batch): assert isinstance(batch, list) and len(batch) == 1, f"Got {batch}" return batch[0] def run_test_offline( env_name: str, model: ModelManager__Union, replay_memory_size: int, num_batches_per_epoch: int, num_train_epochs: int, passing_score_bar: float, num_eval_episodes: int, minibatch_size: int, use_gpu: bool, ): env = Gym(env_name=env_name) env.seed(SEED) env.action_space.seed(SEED) normalization = build_normalizer(env) logger.info(f"Normalization is: \n{pprint.pformat(normalization)}") manager = model.value trainer = manager.build_trainer( use_gpu=use_gpu, reward_options=RewardOptions(), normalization_data_map=normalization, ) # first fill the replay buffer to burn_in replay_buffer = ReplayBuffer( replay_capacity=replay_memory_size, batch_size=minibatch_size ) # always fill full RB random_policy = make_random_policy_for_env(env) agent = Agent.create_for_env(env, policy=random_policy) fill_replay_buffer( env=env, replay_buffer=replay_buffer, desired_size=replay_memory_size, agent=agent, ) device = torch.device("cuda") if use_gpu else None dataset = OfflineReplayBufferDataset.create_for_trainer( trainer, env, replay_buffer, batch_size=minibatch_size, num_batches=num_batches_per_epoch, device=device, ) data_loader = torch.utils.data.DataLoader(dataset, collate_fn=identity_collate) pl_trainer = pl.Trainer( max_epochs=num_train_epochs, gpus=int(use_gpu), deterministic=True, default_root_dir=f"lightning_log_{str(uuid.uuid4())}", ) pl_trainer.fit(trainer, data_loader) logger.info(f"Evaluating after training for {num_train_epochs} epochs: ") eval_rewards = evaluate_cem(env, manager, trainer, num_eval_episodes) mean_rewards = np.mean(eval_rewards) assert ( mean_rewards >= passing_score_bar ), f"{mean_rewards} doesn't pass the bar {passing_score_bar}." if __name__ == "__main__": unittest.main()
4,868
30.412903
83
py
ReAgent
ReAgent-master/reagent/gym/tests/test_gym.py
#!/usr/bin/env python3 # Copyright (c) Facebook, Inc. and its affiliates. All rights reserved. import logging import os import pprint import unittest import uuid from typing import Optional import numpy as np import pytest import pytorch_lightning as pl import torch from parameterized import parameterized from reagent.gym.agents.agent import Agent from reagent.gym.datasets.episodic_dataset import ( EpisodicDataset, ) from reagent.gym.datasets.replay_buffer_dataset import ReplayBufferDataset from reagent.gym.envs import Env__Union from reagent.gym.envs.env_wrapper import EnvWrapper from reagent.gym.policies.policy import Policy from reagent.gym.policies.random_policies import make_random_policy_for_env from reagent.gym.runners.gymrunner import evaluate_for_n_episodes from reagent.gym.utils import build_normalizer, fill_replay_buffer from reagent.model_managers.union import ModelManager__Union from reagent.replay_memory.circular_replay_buffer import ReplayBuffer from reagent.test.base.horizon_test_base import HorizonTestBase # for seeding the environment SEED = 0 # exponential moving average parameter for tracking reward progress REWARD_DECAY = 0.8 logger = logging.getLogger(__name__) logger.setLevel(logging.INFO) """ Put on-policy gym tests here in the format (test name, path to yaml config). Format path to be: "configs/<env_name>/<model_name>_<env_name>_online.yaml." NOTE: These tests should ideally finish quickly (within 10 minutes) since they are unit tests which are run many times. """ REPLAY_BUFFER_GYM_TESTS_1 = [ ("Discrete CRR Cartpole", "configs/cartpole/discrete_crr_cartpole_online.yaml"), ("Discrete DQN Cartpole", "configs/cartpole/discrete_dqn_cartpole_online.yaml"), ("Discrete C51 Cartpole", "configs/cartpole/discrete_c51_cartpole_online.yaml"), ("Discrete QR Cartpole", "configs/cartpole/discrete_qr_cartpole_online.yaml"), ( "Discrete DQN Open Gridworld", "configs/open_gridworld/discrete_dqn_open_gridworld.yaml", ), ("SAC Pendulum", "configs/pendulum/sac_pendulum_online.yaml"), ("Continuous CRR Pendulum", "configs/pendulum/continuous_crr_pendulum_online.yaml"), ("TD3 Pendulum", "configs/pendulum/td3_pendulum_online.yaml"), ] REPLAY_BUFFER_GYM_TESTS_2 = [ ("Parametric DQN Cartpole", "configs/cartpole/parametric_dqn_cartpole_online.yaml"), ( "Parametric SARSA Cartpole", "configs/cartpole/parametric_sarsa_cartpole_online.yaml", ), # Disabled for now because flaky. # ( # "Sparse DQN Changing Arms", # "configs/sparse/discrete_dqn_changing_arms_online.yaml", # ), ("SlateQ RecSim", "configs/recsim/slate_q_recsim_online.yaml"), ( "SlateQ RecSim with Discount Scaled by Time Diff", "configs/recsim/slate_q_recsim_online_with_time_scale.yaml", ), ( "SlateQ RecSim multi selection", "configs/recsim/slate_q_recsim_online_multi_selection.yaml", ), ( "SlateQ RecSim multi selection average by current slate size", "configs/recsim/slate_q_recsim_online_multi_selection_avg_curr.yaml", ), ("PossibleActionsMask DQN", "configs/functionality/dqn_possible_actions_mask.yaml"), ] ONLINE_EPISODE_GYM_TESTS = [ ( "REINFORCE Cartpole online", "configs/cartpole/discrete_reinforce_cartpole_online.yaml", ), ( "PPO Cartpole online", "configs/cartpole/discrete_ppo_cartpole_online.yaml", ), ] curr_dir = os.path.dirname(__file__) class TestGym(HorizonTestBase): # pyre-fixme[16]: Module `parameterized` has no attribute `expand`. @parameterized.expand(REPLAY_BUFFER_GYM_TESTS_1) def test_replay_buffer_gym_cpu_1(self, name: str, config_path: str): self._test_replay_buffer_gym_cpu(name, config_path) # pyre-fixme[16]: Module `parameterized` has no attribute `expand`. @parameterized.expand(REPLAY_BUFFER_GYM_TESTS_2) def test_replay_buffer_gym_cpu_2(self, name: str, config_path: str): self._test_replay_buffer_gym_cpu(name, config_path) def _test_replay_buffer_gym_cpu(self, name: str, config_path: str): logger.info(f"Starting {name} on CPU") self.run_from_config( run_test=run_test_replay_buffer, config_path=os.path.join(curr_dir, config_path), use_gpu=False, ) logger.info(f"{name} passes!") # pyre-fixme[16]: Module `parameterized` has no attribute `expand`. @parameterized.expand(REPLAY_BUFFER_GYM_TESTS_1) @pytest.mark.serial @unittest.skipIf(not torch.cuda.is_available(), "CUDA not available") def test_replay_buffer_gym_gpu_1(self, name: str, config_path: str): self._test_replay_buffer_gym_gpu(name, config_path) # pyre-fixme[16]: Module `parameterized` has no attribute `expand`. @parameterized.expand(REPLAY_BUFFER_GYM_TESTS_2) @pytest.mark.serial @unittest.skipIf(not torch.cuda.is_available(), "CUDA not available") def test_replay_buffer_gym_gpu_2(self, name: str, config_path: str): self._test_replay_buffer_gym_gpu(name, config_path) def _test_replay_buffer_gym_gpu(self, name: str, config_path: str): logger.info(f"Starting {name} on GPU") self.run_from_config( run_test=run_test_replay_buffer, config_path=os.path.join(curr_dir, config_path), use_gpu=True, ) logger.info(f"{name} passes!") # pyre-fixme[16]: Module `parameterized` has no attribute `expand`. @parameterized.expand(ONLINE_EPISODE_GYM_TESTS) def test_online_episode_gym_cpu(self, name: str, config_path: str): logger.info(f"Starting {name} on CPU") self.run_from_config( run_test=run_test_online_episode, config_path=os.path.join(curr_dir, config_path), use_gpu=False, ) logger.info(f"{name} passes!") def eval_policy( env: EnvWrapper, serving_policy: Policy, num_eval_episodes: int, serving: bool = True, ) -> np.ndarray: agent = ( Agent.create_for_env_with_serving_policy(env, serving_policy) if serving else Agent.create_for_env(env, serving_policy) ) eval_rewards = evaluate_for_n_episodes( n=num_eval_episodes, env=env, agent=agent, max_steps=env.max_steps, num_processes=1, ).squeeze(1) logger.info("============Eval rewards==============") logger.info(eval_rewards) mean_eval = np.mean(eval_rewards) logger.info(f"average: {mean_eval};\tmax: {np.max(eval_rewards)}") return np.array(eval_rewards) def identity_collate(batch): assert isinstance(batch, list) and len(batch) == 1, f"Got {batch}" return batch[0] def run_test_replay_buffer( env: Env__Union, model: ModelManager__Union, replay_memory_size: int, train_every_ts: int, train_after_ts: int, num_train_episodes: int, passing_score_bar: float, num_eval_episodes: int, use_gpu: bool, minibatch_size: Optional[int] = None, ): """ Run an online learning test with a replay buffer. The replay buffer is pre-filled, then the training starts. Each transition is added to the replay buffer immediately after it takes place. """ env = env.value pl.seed_everything(SEED) env.seed(SEED) env.action_space.seed(SEED) normalization = build_normalizer(env) logger.info(f"Normalization is: \n{pprint.pformat(normalization)}") manager = model.value trainer = manager.build_trainer( use_gpu=use_gpu, normalization_data_map=normalization, ) training_policy = manager.create_policy(trainer, serving=False) if not isinstance(trainer, pl.LightningModule): if minibatch_size is None: minibatch_size = trainer.minibatch_size assert minibatch_size == trainer.minibatch_size assert minibatch_size is not None replay_buffer = ReplayBuffer( replay_capacity=replay_memory_size, batch_size=minibatch_size ) device = torch.device("cuda") if use_gpu else torch.device("cpu") # first fill the replay buffer using random policy train_after_ts = max(train_after_ts, minibatch_size) random_policy = make_random_policy_for_env(env) agent = Agent.create_for_env(env, policy=random_policy) fill_replay_buffer( env=env, replay_buffer=replay_buffer, desired_size=train_after_ts, agent=agent, ) agent = Agent.create_for_env(env, policy=training_policy, device=device) # TODO: Simplify this setup by creating LightningDataModule dataset = ReplayBufferDataset.create_for_trainer( trainer, env, agent, replay_buffer, batch_size=minibatch_size, training_frequency=train_every_ts, num_episodes=num_train_episodes, max_steps=200, device=device, ) data_loader = torch.utils.data.DataLoader(dataset, collate_fn=identity_collate) pl_trainer = pl.Trainer( max_epochs=1, gpus=int(use_gpu), deterministic=True, default_root_dir=f"lightning_log_{str(uuid.uuid4())}", ) # Note: the fit() function below also evaluates the agent along the way # and adds the new transitions to the replay buffer, so it is training # on incrementally larger and larger buffers. pl_trainer.fit(trainer, data_loader) # TODO: Also check train_reward serving_policy = manager.create_policy( trainer, serving=True, normalization_data_map=normalization ) eval_rewards = eval_policy(env, serving_policy, num_eval_episodes, serving=True) assert ( eval_rewards.mean() >= passing_score_bar ), f"Eval reward is {eval_rewards.mean()}, less than < {passing_score_bar}.\n" def run_test_online_episode( env: Env__Union, model: ModelManager__Union, num_train_episodes: int, passing_score_bar: float, num_eval_episodes: int, use_gpu: bool, ): """ Run an online learning test. At the end of each episode training is run on the trajectory. """ env = env.value pl.seed_everything(SEED) env.seed(SEED) env.action_space.seed(SEED) normalization = build_normalizer(env) logger.info(f"Normalization is: \n{pprint.pformat(normalization)}") manager = model.value trainer = manager.build_trainer( use_gpu=use_gpu, normalization_data_map=normalization, ) policy = manager.create_policy(trainer, serving=False) device = torch.device("cuda") if use_gpu else torch.device("cpu") agent = Agent.create_for_env(env, policy, device=device) pl_trainer = pl.Trainer( max_epochs=1, gpus=int(use_gpu), deterministic=True, default_root_dir=f"lightning_log_{str(uuid.uuid4())}", ) dataset = EpisodicDataset( env=env, agent=agent, num_episodes=num_train_episodes, seed=SEED ) data_loader = torch.utils.data.DataLoader(dataset, collate_fn=identity_collate) pl_trainer.fit(trainer, data_loader) eval_rewards = evaluate_for_n_episodes( n=num_eval_episodes, env=env, agent=agent, max_steps=env.max_steps, num_processes=1, ).squeeze(1) assert ( eval_rewards.mean() >= passing_score_bar ), f"Eval reward is {eval_rewards.mean()}, less than < {passing_score_bar}.\n" if __name__ == "__main__": unittest.main()
11,439
33.251497
112
py
ReAgent
ReAgent-master/reagent/gym/tests/test_world_model.py
#!/usr/bin/env python3 # Copyright (c) Facebook, Inc. and its affiliates. All rights reserved. import logging import os import unittest from typing import Dict, List, Optional import gym import numpy as np import reagent.core.types as rlt import torch from reagent.evaluation.world_model_evaluator import ( FeatureImportanceEvaluator, FeatureSensitivityEvaluator, ) from reagent.gym.agents.agent import Agent from reagent.gym.envs import EnvWrapper, Gym from reagent.gym.envs.pomdp.state_embed_env import StateEmbedEnvironment from reagent.gym.policies.random_policies import make_random_policy_for_env from reagent.gym.preprocessors import make_replay_buffer_trainer_preprocessor from reagent.gym.runners.gymrunner import evaluate_for_n_episodes from reagent.gym.utils import build_normalizer, fill_replay_buffer from reagent.model_managers.union import ModelManager__Union from reagent.models.world_model import MemoryNetwork from reagent.replay_memory.circular_replay_buffer import ReplayBuffer from reagent.test.base.horizon_test_base import HorizonTestBase from reagent.training.world_model.mdnrnn_trainer import MDNRNNTrainer from reagent.workflow.types import RewardOptions from tqdm import tqdm logging.basicConfig(level=logging.INFO) logger = logging.getLogger(__name__) logger.setLevel(logging.INFO) curr_dir = os.path.dirname(__file__) SEED = 0 def calculate_feature_importance( env: gym.Env, trainer: MDNRNNTrainer, use_gpu: bool, test_batch: rlt.MemoryNetworkInput, ): assert isinstance(env.action_space, gym.spaces.Discrete) assert isinstance(env.observation_space, gym.spaces.Box) assert len(env.observation_space.shape) == 1 state_dim = env.observation_space.shape[0] action_dim = env.action_space.n feature_importance_evaluator = FeatureImportanceEvaluator( trainer, discrete_action=True, state_feature_num=state_dim, action_feature_num=action_dim, sorted_state_feature_start_indices=list(range(state_dim)), sorted_action_feature_start_indices=list(range(action_dim)), ) feature_loss_vector = feature_importance_evaluator.evaluate(test_batch)[ "feature_loss_increase" ] feature_importance_map = {} for i in range(action_dim): print( "action {}, feature importance: {}".format(i, feature_loss_vector[i].item()) ) feature_importance_map[f"action{i}"] = feature_loss_vector[i].item() for i in range(state_dim): print( "state {}, feature importance: {}".format( i, feature_loss_vector[i + action_dim].item() ) ) feature_importance_map[f"state{i}"] = feature_loss_vector[i + action_dim].item() return feature_importance_map def calculate_feature_sensitivity( env: EnvWrapper, trainer: MDNRNNTrainer, use_gpu: bool, test_batch: rlt.MemoryNetworkInput, ): assert isinstance(env.action_space, gym.spaces.Discrete) assert isinstance(env.observation_space, gym.spaces.Box) assert len(env.observation_space.shape) == 1 state_dim = env.observation_space.shape[0] feature_sensitivity_evaluator = FeatureSensitivityEvaluator( trainer, state_feature_num=state_dim, sorted_state_feature_start_indices=list(range(state_dim)), ) feature_sensitivity_vector = feature_sensitivity_evaluator.evaluate(test_batch)[ "feature_sensitivity" ] feature_sensitivity_map = {} for i in range(state_dim): feature_sensitivity_map["state" + str(i)] = feature_sensitivity_vector[i].item() print( "state {}, feature sensitivity: {}".format( i, feature_sensitivity_vector[i].item() ) ) return feature_sensitivity_map def train_mdnrnn( env: EnvWrapper, trainer: MDNRNNTrainer, trainer_preprocessor, num_train_transitions: int, seq_len: int, batch_size: int, num_train_epochs: int, # for optional validation test_replay_buffer=None, ): train_replay_buffer = ReplayBuffer( replay_capacity=num_train_transitions, batch_size=batch_size, stack_size=seq_len, return_everything_as_stack=True, ) random_policy = make_random_policy_for_env(env) agent = Agent.create_for_env(env, policy=random_policy) fill_replay_buffer(env, train_replay_buffer, num_train_transitions, agent) num_batch_per_epoch = train_replay_buffer.size // batch_size logger.info("Made RBs, starting to train now!") optimizer = trainer.configure_optimizers()[0] for _ in range(num_train_epochs): for i in range(num_batch_per_epoch): batch = train_replay_buffer.sample_transition_batch(batch_size=batch_size) preprocessed_batch = trainer_preprocessor(batch) loss = next(trainer.train_step_gen(preprocessed_batch, i)) optimizer.zero_grad() loss.backward() optimizer.step() # validation if test_replay_buffer is not None: with torch.no_grad(): trainer.memory_network.mdnrnn.eval() test_batch = test_replay_buffer.sample_transition_batch( batch_size=batch_size ) preprocessed_test_batch = trainer_preprocessor(test_batch) valid_losses = trainer.get_loss(preprocessed_test_batch) trainer.memory_network.mdnrnn.train() return trainer def train_mdnrnn_and_compute_feature_stats( env_name: str, model: ModelManager__Union, num_train_transitions: int, num_test_transitions: int, seq_len: int, batch_size: int, num_train_epochs: int, use_gpu: bool, saved_mdnrnn_path: Optional[str] = None, ): """Train MDNRNN Memory Network and compute feature importance/sensitivity.""" env: gym.Env = Gym(env_name=env_name) env.seed(SEED) manager = model.value trainer = manager.build_trainer( use_gpu=use_gpu, reward_options=RewardOptions(), normalization_data_map=build_normalizer(env), ) device = "cuda" if use_gpu else "cpu" # pyre-fixme[6]: Expected `device` for 2nd param but got `str`. trainer_preprocessor = make_replay_buffer_trainer_preprocessor(trainer, device, env) test_replay_buffer = ReplayBuffer( replay_capacity=num_test_transitions, batch_size=batch_size, stack_size=seq_len, return_everything_as_stack=True, ) random_policy = make_random_policy_for_env(env) agent = Agent.create_for_env(env, policy=random_policy) fill_replay_buffer(env, test_replay_buffer, num_test_transitions, agent) if saved_mdnrnn_path is None: # train from scratch trainer = train_mdnrnn( env=env, trainer=trainer, trainer_preprocessor=trainer_preprocessor, num_train_transitions=num_train_transitions, seq_len=seq_len, batch_size=batch_size, num_train_epochs=num_train_epochs, test_replay_buffer=test_replay_buffer, ) else: # load a pretrained model, and just evaluate it trainer.memory_network.mdnrnn.load_state_dict(torch.load(saved_mdnrnn_path)) with torch.no_grad(): trainer.memory_network.mdnrnn.eval() test_batch = test_replay_buffer.sample_transition_batch( batch_size=test_replay_buffer.size ) preprocessed_test_batch = trainer_preprocessor(test_batch) feature_importance = calculate_feature_importance( env=env, trainer=trainer, use_gpu=use_gpu, test_batch=preprocessed_test_batch, ) feature_sensitivity = calculate_feature_sensitivity( env=env, trainer=trainer, use_gpu=use_gpu, test_batch=preprocessed_test_batch, ) trainer.memory_network.mdnrnn.train() return feature_importance, feature_sensitivity def create_embed_rl_dataset( env: EnvWrapper, memory_network: MemoryNetwork, num_state_embed_transitions: int, batch_size: int, seq_len: int, hidden_dim: int, use_gpu: bool, ): assert isinstance(env.action_space, gym.spaces.Discrete) assert isinstance(env.observation_space, gym.spaces.Box) assert len(env.observation_space.shape) == 1 logger.info("Starting to create embedded RL Dataset!") # seqlen+1 because MDNRNN embeds the first seq_len steps and then # the embedded state will be concatenated with the last step # Ie.. (o1,o2,...,on) -> RNN -> h1,h2,...,hn # and we set s_{n+1} = [o_{n+1}, h_n] embed_env = StateEmbedEnvironment( gym_env=env, mdnrnn=memory_network, max_embed_seq_len=seq_len + 1 ) # now create a filled replay buffer of embeddings # new obs shape dim = state_dim + hidden_dim embed_rb = ReplayBuffer( replay_capacity=num_state_embed_transitions, batch_size=batch_size, stack_size=1 ) random_policy = make_random_policy_for_env(env) agent = Agent.create_for_env(env, policy=random_policy) fill_replay_buffer( env=embed_env, replay_buffer=embed_rb, desired_size=num_state_embed_transitions, agent=agent, ) batch = embed_rb.sample_transition_batch(batch_size=num_state_embed_transitions) state_min = min(batch.state.min(), batch.next_state.min()).item() state_max = max(batch.state.max(), batch.next_state.max()).item() logger.info( f"Finished making embed dataset with size {embed_rb.size}, " f"min {state_min}, max {state_max}" ) return embed_rb, state_min, state_max def train_mdnrnn_and_train_on_embedded_env( env_name: str, embedding_model: ModelManager__Union, num_embedding_train_transitions: int, seq_len: int, batch_size: int, num_embedding_train_epochs: int, train_model: ModelManager__Union, num_state_embed_transitions: int, num_agent_train_epochs: int, num_agent_eval_epochs: int, use_gpu: bool, passing_score_bar: float, # pyre-fixme[9]: saved_mdnrnn_path has type `str`; used as `None`. saved_mdnrnn_path: str = None, ): """Train an agent on embedded states by the MDNRNN.""" env = Gym(env_name=env_name) env.seed(SEED) embedding_manager = embedding_model.value embedding_trainer = embedding_manager.build_trainer( use_gpu=use_gpu, reward_options=RewardOptions(), normalization_data_map=build_normalizer(env), ) device = "cuda" if use_gpu else "cpu" embedding_trainer_preprocessor = make_replay_buffer_trainer_preprocessor( embedding_trainer, # pyre-fixme[6]: Expected `device` for 2nd param but got `str`. device, env, ) if saved_mdnrnn_path is None: # train from scratch embedding_trainer = train_mdnrnn( env=env, trainer=embedding_trainer, trainer_preprocessor=embedding_trainer_preprocessor, num_train_transitions=num_embedding_train_transitions, seq_len=seq_len, batch_size=batch_size, num_train_epochs=num_embedding_train_epochs, ) else: # load a pretrained model, and just evaluate it embedding_trainer.memory_network.mdnrnn.load_state_dict( torch.load(saved_mdnrnn_path) ) # create embedding dataset embed_rb, state_min, state_max = create_embed_rl_dataset( env=env, memory_network=embedding_trainer.memory_network, num_state_embed_transitions=num_state_embed_transitions, batch_size=batch_size, seq_len=seq_len, hidden_dim=embedding_trainer.params.hidden_size, use_gpu=use_gpu, ) embed_env = StateEmbedEnvironment( gym_env=env, mdnrnn=embedding_trainer.memory_network, max_embed_seq_len=seq_len, state_min_value=state_min, state_max_value=state_max, ) agent_manager = train_model.value agent_trainer = agent_manager.build_trainer( use_gpu=use_gpu, reward_options=RewardOptions(), # pyre-fixme[6]: Expected `EnvWrapper` for 1st param but got # `StateEmbedEnvironment`. normalization_data_map=build_normalizer(embed_env), ) device = "cuda" if use_gpu else "cpu" agent_trainer_preprocessor = make_replay_buffer_trainer_preprocessor( agent_trainer, # pyre-fixme[6]: Expected `device` for 2nd param but got `str`. device, env, ) num_batch_per_epoch = embed_rb.size // batch_size # FIXME: This has to be wrapped in dataloader for epoch in range(num_agent_train_epochs): for _ in tqdm(range(num_batch_per_epoch), desc=f"epoch {epoch}"): batch = embed_rb.sample_transition_batch(batch_size=batch_size) preprocessed_batch = agent_trainer_preprocessor(batch) # FIXME: This should be fitted with Lightning's trainer agent_trainer.train(preprocessed_batch) # evaluate model rewards = [] policy = agent_manager.create_policy(agent_trainer, serving=False) # pyre-fixme[6]: Expected `EnvWrapper` for 1st param but got # `StateEmbedEnvironment`. agent = Agent.create_for_env(embed_env, policy=policy, device=device) # num_processes=1 needed to avoid workers from dying on CircleCI tests rewards = evaluate_for_n_episodes( n=num_agent_eval_epochs, # pyre-fixme[6]: Expected `EnvWrapper` for 2nd param but got # `StateEmbedEnvironment`. env=embed_env, agent=agent, num_processes=1, ) assert ( np.mean(rewards) >= passing_score_bar ), f"average reward doesn't pass our bar {passing_score_bar}" return rewards class TestWorldModel(HorizonTestBase): @staticmethod def verify_result(result_dict: Dict[str, float], expected_top_features: List[str]): top_feature = max(result_dict, key=result_dict.get) assert ( top_feature in expected_top_features ), f"top_feature: {top_feature}, expected_top_features: {expected_top_features}" def test_mdnrnn(self): """Test MDNRNN feature importance and feature sensitivity.""" config_path = "configs/world_model/cartpole_features.yaml" feature_importance, feature_sensitivity = self.run_from_config( run_test=train_mdnrnn_and_compute_feature_stats, config_path=os.path.join(curr_dir, config_path), use_gpu=False, ) TestWorldModel.verify_result(feature_importance, ["state1", "state3"]) TestWorldModel.verify_result(feature_sensitivity, ["state3"]) logger.info("MDNRNN feature test passes!") @unittest.skip("This test has to be migrated to Lightning") def test_world_model(self): """Train DQN on POMDP given features from world model.""" config_path = "configs/world_model/discrete_dqn_string.yaml" HorizonTestBase.run_from_config( run_test=train_mdnrnn_and_train_on_embedded_env, config_path=os.path.join(curr_dir, config_path), use_gpu=False, ) logger.info("World model test passes!") if __name__ == "__main__": unittest.main()
15,345
35.451306
88
py
ReAgent
ReAgent-master/reagent/gym/tests/preprocessors/test_default_preprocessors.py
#!/usr/bin/env python3 # Copyright (c) Facebook, Inc. and its affiliates. All rights reserved. import unittest import gym import numpy.testing as npt import torch import torch.nn.functional as F from reagent.gym.envs import Gym try: from reagent.gym.envs import RecSim HAS_RECSIM = True except ModuleNotFoundError: HAS_RECSIM = False class TestMakeDefaultObsPreprocessor(unittest.TestCase): def test_box(self): env = Gym(env_name="CartPole-v0") obs_preprocessor = env.get_obs_preprocessor() obs = env.reset() state = obs_preprocessor(obs) self.assertTrue(state.has_float_features_only) self.assertEqual(state.float_features.shape, (1, obs.shape[0])) self.assertEqual(state.float_features.dtype, torch.float32) self.assertEqual(state.float_features.device, torch.device("cpu")) npt.assert_array_almost_equal(obs, state.float_features.squeeze(0)) @unittest.skipIf(not torch.cuda.is_available(), "CUDA not available") def test_box_cuda(self): env = Gym(env_name="CartPole-v0") device = torch.device("cuda") obs_preprocessor = env.get_obs_preprocessor(device=device) obs = env.reset() state = obs_preprocessor(obs) self.assertTrue(state.has_float_features_only) self.assertEqual(state.float_features.shape, (1, obs.shape[0])) self.assertEqual(state.float_features.dtype, torch.float32) # `device` doesn't have index. So we need this. x = torch.zeros(1, device=device) self.assertEqual(state.float_features.device, x.device) npt.assert_array_almost_equal(obs, state.float_features.cpu().squeeze(0)) @unittest.skipIf(not HAS_RECSIM, "Recsim is not installed") def test_recsim_interest_evolution(self): num_candidate = 10 env = RecSim( num_candidates=num_candidate, slate_size=3, resample_documents=False ) obs_preprocessor = env.get_obs_preprocessor() obs = env.reset() state = obs_preprocessor(obs) self.assertFalse(state.has_float_features_only) self.assertEqual(state.float_features.shape, (1, obs["user"].shape[0])) self.assertEqual(state.float_features.dtype, torch.float32) self.assertEqual(state.float_features.device, torch.device("cpu")) npt.assert_array_almost_equal(obs["user"], state.float_features.squeeze(0)) doc_float_features = state.candidate_docs.float_features self.assertIsNotNone(doc_float_features) self.assertEqual( doc_float_features.shape, (1, num_candidate, obs["doc"]["0"].shape[0]) ) self.assertEqual(doc_float_features.dtype, torch.float32) self.assertEqual(doc_float_features.device, torch.device("cpu")) for i, v in enumerate(obs["doc"].values()): npt.assert_array_almost_equal(v, doc_float_features[0, i]) @unittest.skipIf(not HAS_RECSIM, "Recsim is not installed") def test_recsim_interest_exploration(self): num_candidate = 10 env = RecSim( num_candidates=num_candidate, slate_size=3, resample_documents=False, is_interest_exploration=True, ) obs_preprocessor = env.get_obs_preprocessor() obs = env.reset() state = obs_preprocessor(obs) self.assertFalse(state.has_float_features_only) self.assertEqual(state.float_features.shape, (1, obs["user"].shape[0])) self.assertEqual(state.float_features.dtype, torch.float32) self.assertEqual(state.float_features.device, torch.device("cpu")) npt.assert_array_almost_equal(obs["user"], state.float_features.squeeze(0)) doc_float_features = state.candidate_docs.float_features self.assertIsNotNone(doc_float_features) quality_len = 1 expected_doc_feature_length = ( env.observation_space["doc"]["0"]["cluster_id"].n + quality_len ) self.assertEqual( doc_float_features.shape, (1, num_candidate, expected_doc_feature_length) ) self.assertEqual(doc_float_features.dtype, torch.float32) self.assertEqual(doc_float_features.device, torch.device("cpu")) for i, v in enumerate(obs["doc"].values()): expected_doc_feature = torch.cat( [ F.one_hot(torch.tensor(v["cluster_id"]), 2).float(), # This needs unsqueeze because it's a scalar torch.tensor(v["quality"]).unsqueeze(0).float(), ], dim=0, ) npt.assert_array_almost_equal( expected_doc_feature, doc_float_features[0, i] )
4,748
40.657895
85
py
ReAgent
ReAgent-master/reagent/gym/tests/preprocessors/test_replay_buffer_inserters.py
#!/usr/bin/env python3 # Copyright (c) Facebook, Inc. and its affiliates. All rights reserved. import logging import unittest import gym import numpy as np import numpy.testing as npt import torch from reagent.gym.envs import EnvWrapper from reagent.gym.preprocessors import make_replay_buffer_inserter from reagent.gym.types import Transition from reagent.replay_memory import ReplayBuffer from reagent.test.base.horizon_test_base import HorizonTestBase logger = logging.getLogger(__name__) try: from reagent.gym.envs import RecSim HAS_RECSIM = True except ModuleNotFoundError: HAS_RECSIM = False def _create_replay_buffer_and_insert(env: EnvWrapper): env.seed(1) replay_buffer = ReplayBuffer(replay_capacity=6, batch_size=1) replay_buffer_inserter = make_replay_buffer_inserter(env) obs = env.reset() inserted = [] terminal = False i = 0 while not terminal and i < 5: logger.info(f"Iteration: {i}") action = env.action_space.sample() next_obs, reward, terminal, _ = env.step(action) inserted.append( { "observation": obs, "action": action, "reward": reward, "terminal": terminal, } ) transition = Transition( mdp_id=0, sequence_number=i, observation=obs, action=action, reward=reward, terminal=terminal, log_prob=0.0, ) replay_buffer_inserter(replay_buffer, transition) obs = next_obs i += 1 return replay_buffer, inserted class TestBasicReplayBufferInserter(HorizonTestBase): def test_cartpole(self): env = gym.make("CartPole-v0") replay_buffer, inserted = _create_replay_buffer_and_insert(env) batch = replay_buffer.sample_transition_batch(indices=torch.tensor([0])) npt.assert_array_almost_equal( inserted[0]["observation"], batch.state.squeeze(0) ) npt.assert_array_almost_equal( inserted[1]["observation"], batch.next_state.squeeze(0) ) npt.assert_array_equal(inserted[0]["action"], batch.action.squeeze(0)) npt.assert_array_equal(inserted[1]["action"], batch.next_action.squeeze(0)) class TestRecSimReplayBufferInserter(HorizonTestBase): @unittest.skipIf(not HAS_RECSIM, "RecSim not installed") def test_recsim_interest_evolution(self): num_candidate = 10 slate_size = 3 env = RecSim( num_candidates=num_candidate, slate_size=slate_size, resample_documents=False, ) replay_buffer, inserted = _create_replay_buffer_and_insert(env) batch = replay_buffer.sample_transition_batch(indices=torch.tensor([0])) npt.assert_array_almost_equal( inserted[0]["observation"]["user"], batch.state.squeeze(0) ) npt.assert_array_almost_equal( inserted[1]["observation"]["user"], batch.next_state.squeeze(0) ) docs = list(inserted[0]["observation"]["doc"].values()) next_docs = list(inserted[1]["observation"]["doc"].values()) for i in range(num_candidate): npt.assert_array_equal(docs[i], batch.doc.squeeze(0)[i]) npt.assert_array_equal(next_docs[i], batch.next_doc.squeeze(0)[i]) npt.assert_array_equal(inserted[0]["action"], batch.action.squeeze(0)) npt.assert_array_equal(inserted[1]["action"], batch.next_action.squeeze(0)) npt.assert_array_equal([0, 0, 0], batch.response_click.squeeze(0)) npt.assert_array_equal([0, 0, 0], batch.response_cluster_id.squeeze(0)) npt.assert_array_equal([0, 0, 0], batch.response_liked.squeeze(0)) npt.assert_array_equal([0.0, 0.0, 0.0], batch.response_quality.squeeze(0)) npt.assert_array_equal([0.0, 0.0, 0.0], batch.response_watch_time.squeeze(0)) resp = inserted[1]["observation"]["response"] for i in range(slate_size): npt.assert_array_equal( resp[i]["click"], batch.next_response_click.squeeze(0)[i] ) npt.assert_array_equal( resp[i]["cluster_id"], batch.next_response_cluster_id.squeeze(0)[i] ) npt.assert_array_equal( resp[i]["liked"], batch.next_response_liked.squeeze(0)[i] ) npt.assert_array_almost_equal( resp[i]["quality"], batch.next_response_quality.squeeze(0)[i] ) npt.assert_array_almost_equal( resp[i]["watch_time"], batch.next_response_watch_time.squeeze(0)[i] ) @unittest.skipIf(not HAS_RECSIM, "RecSim not installed") def test_recsim_interest_exploration(self): num_candidate = 10 slate_size = 3 env = RecSim( num_candidates=num_candidate, slate_size=slate_size, resample_documents=False, is_interest_exploration=True, ) replay_buffer, inserted = _create_replay_buffer_and_insert(env) batch = replay_buffer.sample_transition_batch(indices=torch.tensor([0])) npt.assert_array_almost_equal( inserted[0]["observation"]["user"].astype(np.float32), batch.state.squeeze(0), ) npt.assert_array_almost_equal( inserted[1]["observation"]["user"], batch.next_state.squeeze(0) ) docs = list(inserted[0]["observation"]["doc"].values()) next_docs = list(inserted[1]["observation"]["doc"].values()) for i in range(num_candidate): npt.assert_array_almost_equal( docs[i]["quality"], batch.doc_quality.squeeze(0)[i] ) npt.assert_array_almost_equal( next_docs[i]["quality"], batch.next_doc_quality.squeeze(0)[i] ) npt.assert_array_equal(inserted[0]["action"], batch.action.squeeze(0)) npt.assert_array_equal(inserted[1]["action"], batch.next_action.squeeze(0)) npt.assert_array_equal([0, 0, 0], batch.response_click.squeeze(0)) npt.assert_array_equal([0, 0, 0], batch.response_cluster_id.squeeze(0)) npt.assert_array_equal([0.0, 0.0, 0.0], batch.response_quality.squeeze(0)) resp = inserted[1]["observation"]["response"] for i in range(slate_size): npt.assert_array_equal( resp[i]["click"], batch.next_response_click.squeeze(0)[i] ) npt.assert_array_equal( resp[i]["cluster_id"], batch.next_response_cluster_id.squeeze(0)[i] ) npt.assert_array_almost_equal( resp[i]["quality"].astype(np.float32), batch.next_response_quality.squeeze(0)[i], )
6,822
38.668605
85
py
ReAgent
ReAgent-master/reagent/gym/runners/gymrunner.py
#!/usr/bin/env python3 # Copyright (c) Facebook, Inc. and its affiliates. All rights reserved. import logging import pickle from typing import Optional, Sequence import numpy as np import torch.multiprocessing as mp from reagent.core.multiprocess_utils import ( unwrap_function_outputs, wrap_function_arguments, ) from reagent.core.tensorboardX import SummaryWriterContext from reagent.gym.agents.agent import Agent from reagent.gym.envs import EnvWrapper from reagent.gym.types import Trajectory, Transition logger = logging.getLogger(__name__) def run_episode( env: EnvWrapper, agent: Agent, mdp_id: int = 0, max_steps: Optional[int] = None ) -> Trajectory: """ Return sum of rewards from episode. After max_steps (if specified), the environment is assumed to be terminal. Can also specify the mdp_id and gamma of episode. """ trajectory = Trajectory() obs = env.reset() possible_actions_mask = env.possible_actions_mask terminal = False num_steps = 0 while not terminal: action, log_prob = agent.act(obs, possible_actions_mask) next_obs, reward, terminal, _ = env.step(action) next_possible_actions_mask = env.possible_actions_mask if max_steps is not None and num_steps >= (max_steps - 1): terminal = True # Only partially filled. Agent can fill in more fields. transition = Transition( mdp_id=mdp_id, sequence_number=num_steps, observation=obs, action=action, reward=float(reward), terminal=bool(terminal), log_prob=log_prob, possible_actions_mask=possible_actions_mask, ) agent.post_step(transition) trajectory.add_transition(transition) SummaryWriterContext.increase_global_step() obs = next_obs possible_actions_mask = next_possible_actions_mask num_steps += 1 agent.post_episode(trajectory) return trajectory def evaluate_for_n_episodes( n: int, env: EnvWrapper, agent: Agent, max_steps: Optional[int] = None, gammas: Sequence[float] = (1.0,), num_processes: int = 0, ) -> np.ndarray: """Return an np array A of shape n x len(gammas) where A[i, j] = ith episode evaluated with gamma=gammas[j]. Runs environments on num_processes, via multiprocessing.Pool. """ num_processes = min(num_processes, n) def evaluate_one_episode( mdp_id: int, env: EnvWrapper, agent: Agent, max_steps: Optional[int], gammas: Sequence[float], ) -> np.ndarray: rewards = np.empty((len(gammas),)) trajectory = run_episode( env=env, agent=agent, mdp_id=mdp_id, max_steps=max_steps ) for i_gamma, gamma in enumerate(gammas): rewards[i_gamma] = trajectory.calculate_cumulative_reward(gamma) return rewards rewards = None if num_processes > 1: try: with mp.Pool(num_processes) as pool: rewards = unwrap_function_outputs( pool.map( wrap_function_arguments( evaluate_one_episode, env=env, agent=agent, max_steps=max_steps, gammas=gammas, ), range(n), ) ) except pickle.PickleError as e: # NOTE: Probably tried to perform mixed serialization of ScriptModule # and non-script modules. This isn't supported right now. logger.info(e) logger.info( "This is probably from trying to serialize a TorchScript module, " "wrapped in a non-script module. Mixed serialization is not supported." ) # if we didn't run multiprocessing, or it failed, try single-processing instead. if rewards is None: rewards = [] for i in range(n): rewards.append( evaluate_one_episode( mdp_id=i, env=env, agent=agent, max_steps=max_steps, gammas=gammas ) ) rewards = np.array(rewards) for i, gamma in enumerate(gammas): gamma_rewards = rewards[:, i] logger.info( f"For gamma={gamma}, average reward is {gamma_rewards.mean()}\n" f"Rewards list: {gamma_rewards}" ) return rewards
4,541
32.397059
87
py
ReAgent
ReAgent-master/reagent/evaluation/evaluation_data_page.py
#!/usr/bin/env python3 # Copyright (c) Facebook, Inc. and its affiliates. All rights reserved. from __future__ import annotations import logging import math from dataclasses import dataclass, fields from typing import TYPE_CHECKING, Optional, cast import numpy as np import torch import torch.nn as nn from reagent.core import types as rlt from reagent.core.torch_utils import masked_softmax from reagent.model_utils.seq2slate_utils import Seq2SlateMode from reagent.models.seq2slate import Seq2SlateTransformerNet if TYPE_CHECKING: from reagent.training import ParametricDQNTrainer, ReAgentLightningModule from reagent.training.dqn_trainer import DQNTrainer logger = logging.getLogger(__name__) @dataclass class EvaluationDataPage(rlt.TensorDataClass): mdp_id: Optional[torch.Tensor] sequence_number: Optional[torch.Tensor] logged_propensities: torch.Tensor logged_rewards: torch.Tensor action_mask: torch.Tensor model_propensities: torch.Tensor model_rewards: torch.Tensor model_rewards_for_logged_action: torch.Tensor model_values: Optional[torch.Tensor] = None possible_actions_mask: Optional[torch.Tensor] = None optimal_q_values: Optional[torch.Tensor] = None eval_action_idxs: Optional[torch.Tensor] = None logged_values: Optional[torch.Tensor] = None logged_metrics: Optional[torch.Tensor] = None logged_metrics_values: Optional[torch.Tensor] = None model_metrics: Optional[torch.Tensor] = None model_metrics_for_logged_action: Optional[torch.Tensor] = None model_metrics_values: Optional[torch.Tensor] = None model_metrics_values_for_logged_action: Optional[torch.Tensor] = None possible_actions_state_concat: Optional[torch.Tensor] = None contexts: Optional[torch.Tensor] = None @classmethod def create_from_training_batch( cls, tdb: rlt.PreprocessedRankingInput, trainer: ReAgentLightningModule, reward_network: Optional[nn.Module] = None, ): if isinstance(tdb, rlt.DiscreteDqnInput): discrete_training_input = cast(rlt.DiscreteDqnInput, tdb) return EvaluationDataPage.create_from_tensors_dqn( trainer, tdb.extras.mdp_id, tdb.extras.sequence_number, discrete_training_input.state, discrete_training_input.action, tdb.extras.action_probability, discrete_training_input.reward, discrete_training_input.possible_actions_mask, metrics=tdb.extras.metrics, ) elif isinstance(tdb, rlt.ParametricDqnInput): return EvaluationDataPage.create_from_tensors_parametric_dqn( trainer, tdb.extras.mdp_id, tdb.extras.sequence_number, tdb.state, tdb.action, tdb.extras.action_probability, tdb.reward, tdb.possible_actions_mask, tdb.possible_actions, tdb.extras.max_num_actions, metrics=tdb.extras.metrics, ) else: raise NotImplementedError(f"training_input type: {type(tdb)}") @classmethod # pyre-fixme[56]: Decorator `torch.no_grad(...)` could not be called, because # its type `no_grad` is not callable. @torch.no_grad() def create_from_tensors_seq2slate( cls, seq2slate_net: Seq2SlateTransformerNet, reward_network: nn.Module, training_input: rlt.PreprocessedRankingInput, eval_greedy: bool, mdp_ids: Optional[torch.Tensor] = None, sequence_numbers: Optional[torch.Tensor] = None, ): """ :param eval_greedy: If True, evaluate the greedy policy which always picks the most probable output sequence. If False, evaluate the stochastic ranking policy. """ assert ( training_input.slate_reward is not None and training_input.tgt_out_probs is not None and training_input.tgt_out_idx is not None and training_input.tgt_out_seq is not None ) ( batch_size, tgt_seq_len, candidate_dim, ) = training_input.tgt_out_seq.float_features.shape device = training_input.state.float_features.device rank_output = seq2slate_net( training_input, Seq2SlateMode.RANK_MODE, greedy=True ) assert rank_output.ranked_tgt_out_idx is not None if eval_greedy: model_propensities = torch.ones(batch_size, 1, device=device) action_mask = torch.all( # pyre-fixme[58]: `-` is not supported for operand types # `Optional[torch.Tensor]` and `int`. (training_input.tgt_out_idx - 2) == (rank_output.ranked_tgt_out_idx - 2), dim=1, keepdim=True, ).float() else: # Fully evaluating a non-greedy ranking model is too expensive because # we would need to compute propensities of all possible output sequences. # Here we only compute the propensity of the output sequences in the data. # As a result, we can still get a true IPS estimation but not correct # direct method / doubly-robust. model_propensities = torch.exp( seq2slate_net( training_input, Seq2SlateMode.PER_SEQ_LOG_PROB_MODE ).log_probs ) action_mask = torch.ones(batch_size, 1, device=device).float() model_rewards_for_logged_action = reward_network( training_input.state.float_features, training_input.src_seq.float_features, # pyre-fixme[16]: `Optional` has no attribute `float_features`. training_input.tgt_out_seq.float_features, training_input.src_src_mask, training_input.tgt_out_idx, ).reshape(-1, 1) ranked_tgt_out_seq = training_input.src_seq.float_features[ torch.arange(batch_size, device=device).repeat_interleave(tgt_seq_len), rank_output.ranked_tgt_out_idx.flatten() - 2, ].reshape(batch_size, tgt_seq_len, candidate_dim) # model_rewards refers to predicted rewards for the slate generated # greedily by the ranking model. It would be too expensive to # compute model_rewards for all possible slates model_rewards = reward_network( training_input.state.float_features, training_input.src_seq.float_features, ranked_tgt_out_seq, training_input.src_src_mask, rank_output.ranked_tgt_out_idx, ).reshape(-1, 1) # pyre-fixme[16]: `Optional` has no attribute `reshape`. logged_rewards = training_input.slate_reward.reshape(-1, 1) logged_propensities = training_input.tgt_out_probs.reshape(-1, 1) return cls( mdp_id=mdp_ids, sequence_number=sequence_numbers, model_propensities=model_propensities, model_rewards=model_rewards, action_mask=action_mask, logged_rewards=logged_rewards, model_rewards_for_logged_action=model_rewards_for_logged_action, logged_propensities=logged_propensities, ) @classmethod # pyre-fixme[56]: Decorator `torch.no_grad(...)` could not be called, because # its type `no_grad` is not callable. @torch.no_grad() def create_from_tensors_parametric_dqn( cls, trainer: ParametricDQNTrainer, mdp_ids: torch.Tensor, sequence_numbers: torch.Tensor, states: rlt.FeatureData, actions: rlt.FeatureData, propensities: torch.Tensor, rewards: torch.Tensor, possible_actions_mask: torch.Tensor, possible_actions: rlt.FeatureData, max_num_actions: int, metrics: Optional[torch.Tensor] = None, ): old_q_train_state = trainer.q_network.training old_reward_train_state = trainer.reward_network.training trainer.q_network.train(False) trainer.reward_network.train(False) tiled_state = states.float_features.repeat(1, max_num_actions).reshape( -1, states.float_features.shape[1] ) assert possible_actions is not None # Get Q-value of action taken possible_actions_state_concat = (rlt.FeatureData(tiled_state), possible_actions) # FIXME: model_values and model_metrics_values should be # calculated using q_network_cpe (as in discrete dqn). # q_network_cpe has not been added in parametric dqn yet. model_values = trainer.q_network(*possible_actions_state_concat) optimal_q_values, _ = trainer.get_detached_model_outputs( *possible_actions_state_concat ) eval_action_idxs = None assert ( model_values.shape[1] == 1 and model_values.shape[0] == possible_actions_mask.shape[0] * possible_actions_mask.shape[1] ), ( "Invalid shapes: " + str(model_values.shape) + " != " + str(possible_actions_mask.shape) ) model_values = model_values.reshape(possible_actions_mask.shape) optimal_q_values = optimal_q_values.reshape(possible_actions_mask.shape) model_propensities = masked_softmax( optimal_q_values, possible_actions_mask, trainer.rl_temperature ) rewards_and_metric_rewards = trainer.reward_network( *possible_actions_state_concat ) model_rewards = rewards_and_metric_rewards[:, :1] assert ( model_rewards.shape[0] * model_rewards.shape[1] == possible_actions_mask.shape[0] * possible_actions_mask.shape[1] ), ( "Invalid shapes: " + str(model_rewards.shape) + " != " + str(possible_actions_mask.shape) ) model_rewards = model_rewards.reshape(possible_actions_mask.shape) model_metrics = rewards_and_metric_rewards[:, 1:] model_metrics = model_metrics.reshape(possible_actions_mask.shape[0], -1) model_rewards_and_metrics_for_logged_action = trainer.reward_network( states, actions ) model_rewards_for_logged_action = model_rewards_and_metrics_for_logged_action[ :, :1 ] action_dim = possible_actions.float_features.shape[1] action_mask = torch.all( possible_actions.float_features.view(-1, max_num_actions, action_dim) == actions.float_features.unsqueeze(dim=1), dim=2, ).float() assert torch.all(action_mask.sum(dim=1) == 1) num_metrics = model_metrics.shape[1] // max_num_actions model_metrics_values = None model_metrics_for_logged_action = None model_metrics_values_for_logged_action = None if num_metrics > 0: # FIXME: calculate model_metrics_values when q_network_cpe is added # to parametric dqn model_metrics_values = model_values.repeat(1, num_metrics) trainer.q_network.train(old_q_train_state) trainer.reward_network.train(old_reward_train_state) return cls( mdp_id=mdp_ids, sequence_number=sequence_numbers, logged_propensities=propensities, logged_rewards=rewards, action_mask=action_mask, model_rewards=model_rewards, model_rewards_for_logged_action=model_rewards_for_logged_action, model_values=model_values, model_metrics_values=model_metrics_values, model_metrics_values_for_logged_action=model_metrics_values_for_logged_action, model_propensities=model_propensities, logged_metrics=metrics, model_metrics=model_metrics, model_metrics_for_logged_action=model_metrics_for_logged_action, # Will compute later logged_values=None, logged_metrics_values=None, possible_actions_mask=possible_actions_mask, optimal_q_values=optimal_q_values, eval_action_idxs=eval_action_idxs, ) @classmethod # pyre-fixme[56]: Decorator `torch.no_grad(...)` could not be called, because # its type `no_grad` is not callable. @torch.no_grad() def create_from_tensors_dqn( cls, trainer: DQNTrainer, mdp_ids: torch.Tensor, sequence_numbers: torch.Tensor, states: rlt.FeatureData, actions: rlt.FeatureData, propensities: torch.Tensor, rewards: torch.Tensor, possible_actions_mask: torch.Tensor, metrics: Optional[torch.Tensor] = None, ): old_q_train_state = trainer.q_network.training # pyre-fixme[16]: `DQNTrainer` has no attribute `reward_network`. old_reward_train_state = trainer.reward_network.training old_q_cpe_train_state = trainer.q_network_cpe.training trainer.q_network.train(False) # pyre-fixme[16]: `Tensor` has no attribute `train`. trainer.reward_network.train(False) trainer.q_network_cpe.train(False) num_actions = trainer.num_actions action_mask = actions.float() rewards = trainer.boost_rewards(rewards, actions) # pyre-fixme[29]: `Union[nn.Module, torch.Tensor]` is not a function. model_values = trainer.q_network_cpe(states)[:, 0:num_actions] # TODO: make generic get_action_idxs for each trainer class # Note: model_outputs are obtained from the q_network for DQN algorithms # and from the actor_network for CRR. model_outputs, _ = trainer.get_detached_model_outputs(states) # Note: eval_action_idxs is used in evaluate_post_training() function in evaluator.py eval_action_idxs = trainer.get_max_q_values( model_outputs, possible_actions_mask )[1] model_propensities = masked_softmax( model_outputs, possible_actions_mask, trainer.rl_temperature ) assert model_values.shape == actions.shape, ( "Invalid shape: " + str(model_values.shape) + " != " + str(actions.shape) ) assert model_values.shape == possible_actions_mask.shape, ( "Invalid shape: " + str(model_values.shape) + " != " + str(possible_actions_mask.shape) ) # pyre-fixme[29]: `Union[nn.Module, torch.Tensor]` is not a function. rewards_and_metric_rewards = trainer.reward_network(states) # In case we reuse the modular for Q-network if hasattr(rewards_and_metric_rewards, "q_values"): rewards_and_metric_rewards = rewards_and_metric_rewards model_rewards = rewards_and_metric_rewards[:, 0:num_actions] assert model_rewards.shape == actions.shape, ( "Invalid shape: " + str(model_rewards.shape) + " != " + str(actions.shape) ) model_rewards_for_logged_action = torch.sum( model_rewards * action_mask, dim=1, keepdim=True ) model_metrics = rewards_and_metric_rewards[:, num_actions:] assert model_metrics.shape[1] % num_actions == 0, ( "Invalid metrics shape: " + str(model_metrics.shape) + " " + str(num_actions) ) num_metrics = model_metrics.shape[1] // num_actions if num_metrics == 0: model_metrics_values = None model_metrics_for_logged_action = None model_metrics_values_for_logged_action = None else: # pyre-fixme[29]: `Union[nn.Module, torch.Tensor]` is not a function. model_metrics_values = trainer.q_network_cpe(states) # Backward compatility if hasattr(model_metrics_values, "q_values"): model_metrics_values = model_metrics_values model_metrics_values = model_metrics_values[:, num_actions:] assert model_metrics_values.shape[1] == num_actions * num_metrics, ( "Invalid shape: " + str(model_metrics_values.shape[1]) + " != " + str(actions.shape[1] * num_metrics) ) model_metrics_for_logged_action_list = [] model_metrics_values_for_logged_action_list = [] for metric_index in range(num_metrics): metric_start = metric_index * num_actions metric_end = (metric_index + 1) * num_actions model_metrics_for_logged_action_list.append( torch.sum( model_metrics[:, metric_start:metric_end] * action_mask, dim=1, keepdim=True, ) ) model_metrics_values_for_logged_action_list.append( torch.sum( model_metrics_values[:, metric_start:metric_end] * action_mask, dim=1, keepdim=True, ) ) model_metrics_for_logged_action = torch.cat( model_metrics_for_logged_action_list, dim=1 ) model_metrics_values_for_logged_action = torch.cat( model_metrics_values_for_logged_action_list, dim=1 ) trainer.q_network_cpe.train(old_q_cpe_train_state) trainer.q_network.train(old_q_train_state) trainer.reward_network.train(old_reward_train_state) return cls( mdp_id=mdp_ids, sequence_number=sequence_numbers, logged_propensities=propensities, logged_rewards=rewards, action_mask=action_mask, model_rewards=model_rewards, model_rewards_for_logged_action=model_rewards_for_logged_action, model_values=model_values, model_metrics_values=model_metrics_values, model_metrics_values_for_logged_action=model_metrics_values_for_logged_action, model_propensities=model_propensities, logged_metrics=metrics, model_metrics=model_metrics, model_metrics_for_logged_action=model_metrics_for_logged_action, # Will compute later logged_values=None, logged_metrics_values=None, possible_actions_mask=possible_actions_mask, optimal_q_values=model_outputs, eval_action_idxs=eval_action_idxs, ) def append(self, edp): new_edp = {} for x in fields(EvaluationDataPage): t = getattr(self, x.name) other_t = getattr(edp, x.name) assert int(t is not None) + int(other_t is not None) != 1, ( "Tried to append when a tensor existed in one training page but not the other: " + x.name ) if other_t is not None: if isinstance(t, torch.Tensor): new_edp[x.name] = torch.cat((t, other_t), dim=0) elif isinstance(t, np.ndarray): new_edp[x.name] = np.concatenate((t, other_t), axis=0) else: raise Exception("Invalid type in training data page") else: new_edp[x.name] = None return EvaluationDataPage(**new_edp) def sort(self): idxs = [] for i, (mdp_id, seq_num) in enumerate(zip(self.mdp_id, self.sequence_number)): idxs.append((mdp_id, int(seq_num), i)) sorted_idxs = [i for _mdp_id, _seq_num, i in sorted(idxs)] new_edp = {} for x in fields(EvaluationDataPage): t = getattr(self, x.name) new_edp[x.name] = t[sorted_idxs] if t is not None else None return EvaluationDataPage(**new_edp) def compute_values(self, gamma: float): assert self.mdp_id is not None and self.sequence_number is not None logged_values = EvaluationDataPage.compute_values_for_mdps( self.logged_rewards, self.mdp_id, self.sequence_number, gamma, ) if self.logged_metrics is not None: logged_metrics_values: Optional[ torch.Tensor ] = EvaluationDataPage.compute_values_for_mdps( self.logged_metrics, # pyre-fixme[6]: Expected `Tensor` for 2nd param but got # `Optional[torch.Tensor]`. self.mdp_id, # pyre-fixme[6]: Expected `Tensor` for 3rd param but got # `Optional[torch.Tensor]`. self.sequence_number, gamma, ) else: logged_metrics_values = None return self._replace( logged_values=logged_values, logged_metrics_values=logged_metrics_values ) @staticmethod def compute_values_for_mdps( rewards: torch.Tensor, mdp_ids: torch.Tensor, sequence_numbers: torch.Tensor, gamma: float, ) -> torch.Tensor: values = rewards.clone() for x in range(len(values) - 2, -1, -1): if mdp_ids[x] != mdp_ids[x + 1]: # Value = reward for new mdp_id continue values[x, 0] += values[x + 1, 0] * math.pow( gamma, float(sequence_numbers[x + 1, 0] - sequence_numbers[x, 0]) ) return values def validate(self): assert len(self.logged_propensities.shape) == 2 assert len(self.logged_rewards.shape) == 2 assert len(self.logged_values.shape) == 2 assert len(self.model_propensities.shape) == 2 assert len(self.model_rewards.shape) == 2 assert len(self.model_values.shape) == 2 assert self.logged_propensities.shape[1] == 1 assert self.logged_rewards.shape[1] == 1 assert self.logged_values.shape[1] == 1 num_actions = self.model_propensities.shape[1] assert self.model_rewards.shape[1] == num_actions assert self.model_values.shape[1] == num_actions assert self.action_mask.shape == self.model_propensities.shape if self.logged_metrics is not None: assert len(self.logged_metrics.shape) == 2 assert len(self.logged_metrics_values.shape) == 2 assert len(self.model_metrics.shape) == 2 assert len(self.model_metrics_values.shape) == 2 num_metrics = self.logged_metrics.shape[1] assert self.logged_metrics_values.shape[1] == num_metrics, ( "Invalid shape: " + str(self.logged_metrics_values.shape) + " != " + str(num_metrics) ) assert self.model_metrics.shape[1] == num_metrics * num_actions, ( "Invalid shape: " + str(self.model_metrics.shape) + " != " + str(num_metrics * num_actions) ) assert self.model_metrics_values.shape[1] == num_metrics * num_actions minibatch_size = self.logged_propensities.shape[0] logger.info("EvaluationDataPage data size: {}".format(minibatch_size)) assert minibatch_size == self.logged_rewards.shape[0] assert minibatch_size == self.logged_values.shape[0] assert minibatch_size == self.model_propensities.shape[0] assert minibatch_size == self.model_rewards.shape[0] assert minibatch_size == self.model_values.shape[0] if self.logged_metrics is not None: assert minibatch_size == self.logged_metrics.shape[0] assert minibatch_size == self.logged_metrics_values.shape[0] assert minibatch_size == self.model_metrics.shape[0] assert minibatch_size == self.model_metrics_values.shape[0] logger.info("Average logged reward = %s", self.logged_rewards.mean()) logger.info( "Average model propensity for action 0 = %s", self.model_propensities[:, 0].mean(), ) logger.info( "Average model propensity for action 1 = %s", self.model_propensities[:, 1].mean(), ) logger.info( "Average logged propensity = %s", self.logged_propensities.mean(), ) flatten_mdp_id = self.mdp_id.reshape(-1) unique_mdp_ids = set(flatten_mdp_id.tolist()) prev_mdp_id, prev_seq_num = None, None mdp_count = 0 for mdp_id, seq_num in zip(flatten_mdp_id, self.sequence_number): if prev_mdp_id is None or mdp_id != prev_mdp_id: mdp_count += 1 prev_mdp_id = mdp_id else: assert seq_num > prev_seq_num, ( f"For mdp_id {mdp_id}, got {seq_num} <= {prev_seq_num}." f"Sequence number must be in increasing order.\n" f"Zip(mdp_id, seq_num): " f"{list(zip(flatten_mdp_id, self.sequence_number))}" ) prev_seq_num = seq_num assert len(unique_mdp_ids) == mdp_count, "MDPs are broken up. {} vs {}".format( len(unique_mdp_ids), mdp_count ) def set_metric_as_reward(self, i: int, num_actions: int): assert self.logged_metrics is not None, "metrics must not be none" assert self.logged_metrics_values is not None, "metrics must not be none" assert self.model_metrics is not None, "metrics must not be none" assert self.model_metrics_values is not None, "metrics must not be none" return self._replace( logged_rewards=self.logged_metrics[:, i : i + 1], # pyre-fixme[16]: `Optional` has no attribute `__getitem__`. logged_values=self.logged_metrics_values[:, i : i + 1], model_rewards=self.model_metrics[ :, i * num_actions : (i + 1) * num_actions ], model_values=self.model_metrics_values[ :, i * num_actions : (i + 1) * num_actions ], logged_metrics=None, logged_metrics_values=None, model_metrics=None, model_metrics_values=None, )
26,496
40.401563
96
py
ReAgent
ReAgent-master/reagent/evaluation/world_model_evaluator.py
#!/usr/bin/env python3 # Copyright (c) Facebook, Inc. and its affiliates. All rights reserved. import logging from typing import Dict, List import torch from reagent.core.types import FeatureData, MemoryNetworkInput from reagent.training.world_model.mdnrnn_trainer import MDNRNNTrainer logger = logging.getLogger(__name__) class LossEvaluator(object): """Evaluate losses on data pages""" def __init__(self, trainer: MDNRNNTrainer, state_dim: int) -> None: self.trainer = trainer self.state_dim = state_dim def evaluate(self, tdp: MemoryNetworkInput) -> Dict[str, float]: self.trainer.memory_network.mdnrnn.eval() losses = self.trainer.get_loss(tdp, state_dim=self.state_dim) detached_losses = { "loss": losses["loss"].cpu().detach().item(), "gmm": losses["gmm"].cpu().detach().item(), "bce": losses["bce"].cpu().detach().item(), "mse": losses["mse"].cpu().detach().item(), } del losses self.trainer.memory_network.mdnrnn.train() return detached_losses class FeatureImportanceEvaluator(object): """Evaluate feature importance weights on data pages""" def __init__( self, trainer: MDNRNNTrainer, discrete_action: bool, state_feature_num: int, action_feature_num: int, sorted_action_feature_start_indices: List[int], sorted_state_feature_start_indices: List[int], ) -> None: """ :param sorted_action_feature_start_indices: the starting index of each action feature in the action vector (need this because some features (e.g., one-hot encoding enum) may take multiple components) :param sorted_state_feature_start_indices: the starting index of each state feature in the state vector """ self.trainer = trainer self.discrete_action = discrete_action self.state_feature_num = state_feature_num self.action_feature_num = action_feature_num self.sorted_action_feature_start_indices = sorted_action_feature_start_indices self.sorted_state_feature_start_indices = sorted_state_feature_start_indices def evaluate(self, batch: MemoryNetworkInput): """Calculate feature importance: setting each state/action feature to the mean value and observe loss increase.""" self.trainer.memory_network.mdnrnn.eval() state_features = batch.state.float_features action_features = batch.action seq_len, batch_size, state_dim = state_features.size() action_dim = action_features.size()[2] action_feature_num = self.action_feature_num state_feature_num = self.state_feature_num feature_importance = torch.zeros(action_feature_num + state_feature_num) orig_losses = self.trainer.get_loss(batch, state_dim=state_dim) orig_loss = orig_losses["loss"].cpu().detach().item() del orig_losses action_feature_boundaries = self.sorted_action_feature_start_indices + [ action_dim ] state_feature_boundaries = self.sorted_state_feature_start_indices + [state_dim] for i in range(action_feature_num): action_features = batch.action.reshape( (batch_size * seq_len, action_dim) ).data.clone() # if actions are discrete, an action's feature importance is the loss # increase due to setting all actions to this action if self.discrete_action: assert action_dim == action_feature_num action_vec = torch.zeros(action_dim) action_vec[i] = 1 action_features[:] = action_vec # if actions are continuous, an action's feature importance is the loss # increase due to masking this action feature to its mean value else: boundary_start, boundary_end = ( action_feature_boundaries[i], action_feature_boundaries[i + 1], ) action_features[ :, boundary_start:boundary_end ] = self.compute_median_feature_value( action_features[:, boundary_start:boundary_end] ) action_features = action_features.reshape((seq_len, batch_size, action_dim)) new_batch = MemoryNetworkInput( state=batch.state, action=action_features, next_state=batch.next_state, reward=batch.reward, time_diff=torch.ones_like(batch.reward).float(), not_terminal=batch.not_terminal, step=None, ) losses = self.trainer.get_loss(new_batch, state_dim=state_dim) feature_importance[i] = losses["loss"].cpu().detach().item() - orig_loss del losses for i in range(state_feature_num): state_features = batch.state.float_features.reshape( (batch_size * seq_len, state_dim) ).data.clone() boundary_start, boundary_end = ( state_feature_boundaries[i], state_feature_boundaries[i + 1], ) state_features[ :, boundary_start:boundary_end ] = self.compute_median_feature_value( state_features[:, boundary_start:boundary_end] ) state_features = state_features.reshape((seq_len, batch_size, state_dim)) new_batch = MemoryNetworkInput( state=FeatureData(float_features=state_features), action=batch.action, next_state=batch.next_state, reward=batch.reward, time_diff=torch.ones_like(batch.reward).float(), not_terminal=batch.not_terminal, step=None, ) losses = self.trainer.get_loss(new_batch, state_dim=state_dim) feature_importance[i + action_feature_num] = ( losses["loss"].cpu().detach().item() - orig_loss ) del losses self.trainer.memory_network.mdnrnn.train() logger.info( "**** Debug tool feature importance ****: {}".format(feature_importance) ) return {"feature_loss_increase": feature_importance.numpy()} def compute_median_feature_value(self, features): # enum type if features.shape[1] > 1: feature_counts = torch.sum(features, dim=0) median_feature_counts = torch.median(feature_counts) # no similar method as numpy.where in torch for i in range(features.shape[1]): if feature_counts[i] == median_feature_counts: break median_feature = torch.zeros(features.shape[1]) median_feature[i] = 1 # other types else: median_feature = features.mean(dim=0) return median_feature class FeatureSensitivityEvaluator(object): """Evaluate state feature sensitivity caused by varying actions""" def __init__( self, trainer: MDNRNNTrainer, state_feature_num: int, sorted_state_feature_start_indices: List[int], ) -> None: self.trainer = trainer self.state_feature_num = state_feature_num self.sorted_state_feature_start_indices = sorted_state_feature_start_indices def evaluate(self, batch: MemoryNetworkInput): """Calculate state feature sensitivity due to actions: randomly permutating actions and see how much the prediction of next state feature deviates.""" assert isinstance(batch, MemoryNetworkInput) self.trainer.memory_network.mdnrnn.eval() seq_len, batch_size, state_dim = batch.next_state.float_features.size() state_feature_num = self.state_feature_num feature_sensitivity = torch.zeros(state_feature_num) # the input of world_model has seq-len as the first dimension mdnrnn_output = self.trainer.memory_network( batch.state, FeatureData(batch.action) ) predicted_next_state_means = mdnrnn_output.mus shuffled_mdnrnn_output = self.trainer.memory_network( batch.state, # shuffle the actions FeatureData(batch.action[:, torch.randperm(batch_size), :]), ) shuffled_predicted_next_state_means = shuffled_mdnrnn_output.mus assert ( predicted_next_state_means.size() == shuffled_predicted_next_state_means.size() == (seq_len, batch_size, self.trainer.params.num_gaussians, state_dim) ) state_feature_boundaries = self.sorted_state_feature_start_indices + [state_dim] for i in range(state_feature_num): boundary_start, boundary_end = ( state_feature_boundaries[i], state_feature_boundaries[i + 1], ) abs_diff = torch.mean( torch.sum( torch.abs( shuffled_predicted_next_state_means[ :, :, :, boundary_start:boundary_end ] - predicted_next_state_means[ :, :, :, boundary_start:boundary_end ] ), dim=3, ) ) feature_sensitivity[i] = abs_diff.cpu().detach().item() self.trainer.memory_network.mdnrnn.train() logger.info( "**** Debug tool feature sensitivity ****: {}".format(feature_sensitivity) ) return {"feature_sensitivity": feature_sensitivity.numpy()}
9,841
39.336066
88
py
ReAgent
ReAgent-master/reagent/evaluation/sequential_doubly_robust_estimator.py
#!/usr/bin/env python3 # Copyright (c) Facebook, Inc. and its affiliates. All rights reserved. import logging from typing import List import numpy as np import torch from reagent.evaluation.cpe import CpeEstimate, bootstrapped_std_error_of_mean from reagent.evaluation.evaluation_data_page import EvaluationDataPage logger = logging.getLogger(__name__) class SequentialDoublyRobustEstimator: def __init__(self, gamma): self.gamma = gamma def estimate(self, edp: EvaluationDataPage) -> CpeEstimate: # For details, visit https://arxiv.org/pdf/1511.03722.pdf logged_rewards = edp.logged_rewards.squeeze() logged_propensities = edp.logged_propensities.squeeze() num_examples = edp.logged_rewards.shape[0] estimated_state_values = torch.sum( edp.model_propensities * edp.model_values, dim=1 ) estimated_q_values_for_logged_action = torch.sum( edp.model_values * edp.action_mask, dim=1 ) target_propensity_for_action = torch.sum( edp.model_propensities * edp.action_mask, dim=1 ) assert target_propensity_for_action.shape == logged_propensities.shape, ( "Invalid shape: " + str(target_propensity_for_action.shape) + " != " + str(logged_propensities.shape) ) assert ( target_propensity_for_action.shape == estimated_q_values_for_logged_action.shape ), ( "Invalid shape: " + str(target_propensity_for_action.shape) + " != " + str(estimated_q_values_for_logged_action.shape) ) assert target_propensity_for_action.shape == logged_rewards.shape, ( "Invalid shape: " + str(target_propensity_for_action.shape) + " != " + str(logged_rewards.shape) ) importance_weight = target_propensity_for_action / logged_propensities doubly_robusts: List[float] = [] episode_values: List[float] = [] assert edp.mdp_id is not None i = 0 last_episode_end = -1 while i < num_examples: # calculate the doubly-robust Q-value for one episode # pyre-ignore [16]: Optional type has no attribute `__getitem__` if i == num_examples - 1 or edp.mdp_id[i] != edp.mdp_id[i + 1]: episode_end = i episode_value = 0.0 doubly_robust = 0.0 for j in range(episode_end, last_episode_end, -1): doubly_robust = estimated_state_values[j] + importance_weight[j] * ( logged_rewards[j] + self.gamma * doubly_robust - estimated_q_values_for_logged_action[j] ) episode_value *= self.gamma episode_value += logged_rewards[j] doubly_robusts.append(float(doubly_robust)) episode_values.append(float(episode_value)) last_episode_end = episode_end i += 1 if len(doubly_robusts) == 0: torch.set_printoptions(profile="full") zipped_data = list( zip( *map( lambda x: x.tolist(), [ edp.mdp_id, logged_rewards, estimated_state_values, estimated_q_values_for_logged_action, importance_weight, ], ) ) ) raise RuntimeError( f"No valid doubly robusts data is generated.\n" f"mdp_ids x logged_rewards x estimated_state_values x " f"estimated_q_values_for_logged_action x importance_weight:\n" f"{zipped_data};\n" f"gamma={self.gamma};\n" f"Did you specify wrong metric names?" ) # pyre-fixme[9]: doubly_robusts has type `List[float]`; used as `ndarray`. doubly_robusts = np.array(doubly_robusts) dr_score = float(np.mean(doubly_robusts)) dr_score_std_error = bootstrapped_std_error_of_mean(doubly_robusts) # pyre-fixme[9]: episode_values has type `List[float]`; used as `ndarray`. episode_values = np.array(episode_values) logged_policy_score = np.mean(episode_values) if logged_policy_score < 1e-6: logger.warning( "Can't normalize SDR-CPE because of small" f" or negative logged_policy_score ({logged_policy_score})." f"Episode values: {episode_values}." ) return CpeEstimate( raw=dr_score, normalized=0.0, raw_std_error=dr_score_std_error, normalized_std_error=0.0, ) return CpeEstimate( raw=dr_score, normalized=dr_score / logged_policy_score, raw_std_error=dr_score_std_error, normalized_std_error=dr_score_std_error / logged_policy_score, )
5,254
36.535714
88
py
ReAgent
ReAgent-master/reagent/evaluation/doubly_robust_estimator.py
#!/usr/bin/env python3 # Copyright (c) Facebook, Inc. and its affiliates. All rights reserved. import itertools import logging from dataclasses import dataclass from typing import Dict, NamedTuple, Optional, Tuple, Union import numpy as np import torch from reagent.evaluation.cpe import CpeEstimate, bootstrapped_std_error_of_mean from reagent.evaluation.evaluation_data_page import EvaluationDataPage from torch import Tensor logger = logging.getLogger(__name__) DEFAULT_FRAC_TRAIN = 0.4 DEFAULT_FRAC_VALID = 0.1 class DoublyRobustHP(NamedTuple): frac_train: float = DEFAULT_FRAC_TRAIN frac_valid: float = DEFAULT_FRAC_VALID bootstrap_num_samples: int = 1000 bootstrap_sample_percent: float = 0.25 xgb_params: Optional[Dict[str, Union[float, int, str]]] = None bope_mode: Optional[str] = None bope_num_samples: Optional[int] = None class TrainValidEvalData(NamedTuple): contexts_dict: Dict[str, Tensor] model_propensities_dict: Dict[str, Tensor] actions_logged_dict: Dict[str, Tensor] action_mask_dict: Dict[str, Tensor] logged_rewards_dict: Dict[str, Tensor] model_rewards_dict: Dict[str, Tensor] model_rewards_for_logged_action_dict: Dict[str, Tensor] logged_propensities_dict: Dict[str, Tensor] num_examples_dict: Dict[str, int] @dataclass class EstimationData: contexts_actions_train: Optional[Tensor] policy_indicators_train: Optional[Tensor] weights_train: Optional[Tensor] contexts_actions_valid: Optional[Tensor] policy_indicators_valid: Optional[Tensor] weights_valid: Optional[Tensor] contexts_actions_eval: Optional[Tensor] contexts_train: Optional[Tensor] actions_logged_train: Optional[Tensor] contexts_valid: Optional[Tensor] actions_logged_valid: Optional[Tensor] contexts_eval: Optional[Tensor] actions_logged_eval: Optional[Tensor] model_propensities_eval: Tensor model_rewards_eval: Tensor action_mask_eval: Tensor logged_rewards_eval: Tensor model_rewards_for_logged_action_eval: Tensor logged_propensities_eval: Tensor def __post_init__(self): assert ( self.model_propensities_eval.shape == self.model_rewards_eval.shape == self.action_mask_eval.shape ) and len(self.model_propensities_eval.shape) == 2, ( f"{self.model_propensities_eval.shape} " f"{self.model_rewards_eval.shape} " f"{self.action_mask_eval.shape}" ) assert ( ( self.logged_rewards_eval.shape == self.model_rewards_for_logged_action_eval.shape == self.logged_propensities_eval.shape ) and len(self.logged_rewards_eval.shape) == 2 and self.logged_rewards_eval.shape[1] == 1 ), ( f"{self.logged_rewards_eval.shape} " f"{self.model_rewards_for_logged_action_eval.shape} " f"{self.logged_propensities_eval.shape}" ) class ImportanceSamplingData(NamedTuple): importance_weight: Tensor logged_rewards: Tensor model_rewards: Tensor model_rewards_for_logged_action: Tensor model_propensities: Tensor class DoublyRobustEstimator: """ For details, visit https://arxiv.org/pdf/1612.01205.pdf """ def _split_data( self, edp: EvaluationDataPage, frac_train: float = DEFAULT_FRAC_TRAIN, frac_valid: float = DEFAULT_FRAC_VALID, ) -> TrainValidEvalData: """ Split the data into training, validation and evaluation parts. Training and validation and used for model training to estimate the importance weights. Only evaluation data is used for policy estimation. This function is used for BOP-E and Estimated Propensity Score methods, but not for the standard Doubly Robust estimator. """ num_examples = edp.model_propensities.shape[0] # split data into training, validation and eval indices = np.random.permutation(num_examples) idx_train = indices[0 : int(frac_train * num_examples)] idx_valid = indices[ int(frac_train * num_examples) : int( (frac_train + frac_valid) * num_examples ) ] idx_eval = indices[int((frac_train + frac_valid) * num_examples) :] if edp.contexts is None: raise ValueError("contexts not provided in input") contexts_dict = { "train": edp.contexts[idx_train], # pyre-fixme[16]: `Optional` has no attribute `__getitem__`. "valid": edp.contexts[idx_valid], "eval": edp.contexts[idx_eval], } model_propensities_dict = { "train": edp.model_propensities[idx_train], "valid": edp.model_propensities[idx_valid], "eval": edp.model_propensities[idx_eval], } # edp.action_mask is N*N_actions tensor of indicators of which actions # were actually taken by the logged algo actions_logged = torch.max(edp.action_mask, dim=1, keepdim=True)[1].float() actions_logged_dict = { "train": actions_logged[idx_train], "valid": actions_logged[idx_valid], "eval": actions_logged[idx_eval], } action_mask_dict = { "train": edp.action_mask[idx_train], "valid": edp.action_mask[idx_valid], "eval": edp.action_mask[idx_eval], } logged_rewards_dict = { "train": edp.logged_rewards[idx_train], "valid": edp.logged_rewards[idx_valid], "eval": edp.logged_rewards[idx_eval], } model_rewards_dict = { "train": edp.model_rewards[idx_train], "valid": edp.model_rewards[idx_valid], "eval": edp.model_rewards[idx_eval], } model_rewards_for_logged_action_dict = { "train": edp.model_rewards_for_logged_action[idx_train], "valid": edp.model_rewards_for_logged_action[idx_valid], "eval": edp.model_rewards_for_logged_action[idx_eval], } logged_propensities_dict = { "train": edp.logged_propensities[idx_train], "valid": edp.logged_propensities[idx_valid], "eval": edp.logged_propensities[idx_eval], } num_examples_dict = { "train": int(frac_train * num_examples), "valid": int((frac_train + frac_valid) * num_examples) - int(frac_train * num_examples), "eval": num_examples - int((frac_train + frac_valid) * num_examples), } return TrainValidEvalData( contexts_dict=contexts_dict, model_propensities_dict=model_propensities_dict, actions_logged_dict=actions_logged_dict, action_mask_dict=action_mask_dict, logged_rewards_dict=logged_rewards_dict, model_rewards_dict=model_rewards_dict, model_rewards_for_logged_action_dict=model_rewards_for_logged_action_dict, logged_propensities_dict=logged_propensities_dict, num_examples_dict=num_examples_dict, ) def _prepare_data(self, edp: EvaluationDataPage) -> EstimationData: ed = EstimationData( contexts_actions_train=None, policy_indicators_train=None, weights_train=None, contexts_actions_valid=None, policy_indicators_valid=None, weights_valid=None, contexts_actions_eval=None, contexts_train=None, actions_logged_train=None, contexts_valid=None, actions_logged_valid=None, contexts_eval=edp.contexts, actions_logged_eval=None, model_propensities_eval=edp.model_propensities, model_rewards_eval=edp.model_rewards, action_mask_eval=edp.action_mask, logged_rewards_eval=edp.logged_rewards, model_rewards_for_logged_action_eval=edp.model_rewards_for_logged_action, logged_propensities_eval=edp.logged_propensities, ) return ed def _get_importance_sampling_inputs( self, ed: EstimationData ) -> ImportanceSamplingData: target_propensity_for_action = torch.sum( ed.model_propensities_eval * ed.action_mask_eval, dim=1, keepdim=True ) # target_propensity_for_action is N*1 tensor of target algo propensities # for historical actions # logged_propensities_eval is N*1 tensor of propensity scores for historical # actions by the prod algorithm at each context importance_weights = ( target_propensity_for_action / ed.logged_propensities_eval ).float() logger.info(f"Mean IPS weight on the eval dataset: {importance_weights.mean()}") return ImportanceSamplingData( importance_weight=importance_weights, logged_rewards=ed.logged_rewards_eval, model_rewards=ed.model_rewards_eval, model_rewards_for_logged_action=ed.model_rewards_for_logged_action_eval, model_propensities=ed.model_propensities_eval, ) def _get_importance_sampling_estimates( self, isd: ImportanceSamplingData, hp: DoublyRobustHP ) -> Tuple[CpeEstimate, CpeEstimate, CpeEstimate]: # The score we would get if we evaluate the logged policy against itself logged_policy_score = float( torch.mean(isd.logged_rewards) ) # logged_rewards is N*1 tensor of historical rewards if logged_policy_score < 1e-6: logger.warning( "Can't normalize DR-CPE because of small or negative " + "logged_policy_score" ) normalizer = 0.0 else: normalizer = 1.0 / logged_policy_score if isd.model_rewards is None: # Fill with zero, equivalent to just doing IPS direct_method_values = torch.zeros( [isd.model_propensities.shape[0], 1], dtype=torch.float32 ) else: # model rewards is (N_samples)*N_actions tensor of predicted # counterfactual rewards for each possible action at each # historical context direct_method_values = torch.sum( isd.model_propensities * isd.model_rewards, dim=1, keepdim=True ) direct_method_score = float(torch.mean(direct_method_values)) logger.info( f"Normalized Direct method score = {direct_method_score * normalizer}" ) direct_method_std_error = bootstrapped_std_error_of_mean( direct_method_values.squeeze(), sample_percent=hp.bootstrap_sample_percent, num_samples=hp.bootstrap_num_samples, ) direct_method_estimate = CpeEstimate( raw=direct_method_score, normalized=direct_method_score * normalizer, raw_std_error=direct_method_std_error, normalized_std_error=direct_method_std_error * normalizer, ) ips = isd.importance_weight * isd.logged_rewards # N*1 doubly_robust = ( isd.importance_weight * (isd.logged_rewards - isd.model_rewards_for_logged_action) ) + direct_method_values # model_rewards_for_logged_action is N*1 of estimated rewards for target # policy ips_score = float(torch.mean(ips)) logger.info(f"Normalized IPS score = {ips_score * normalizer}") ips_score_std_error = bootstrapped_std_error_of_mean( ips.squeeze(), sample_percent=hp.bootstrap_sample_percent, num_samples=hp.bootstrap_num_samples, ) inverse_propensity_estimate = CpeEstimate( raw=ips_score, normalized=ips_score * normalizer, raw_std_error=ips_score_std_error, normalized_std_error=ips_score_std_error * normalizer, ) dr_score = float(torch.mean(doubly_robust)) dr_score_std_error = bootstrapped_std_error_of_mean( doubly_robust.squeeze(), sample_percent=hp.bootstrap_sample_percent, num_samples=hp.bootstrap_num_samples, ) doubly_robust_estimate = CpeEstimate( raw=dr_score, normalized=dr_score * normalizer, raw_std_error=dr_score_std_error, normalized_std_error=dr_score_std_error * normalizer, ) return ( direct_method_estimate, inverse_propensity_estimate, doubly_robust_estimate, ) def estimate( self, edp: EvaluationDataPage, hp: Optional[DoublyRobustHP] = None ) -> Tuple[CpeEstimate, CpeEstimate, CpeEstimate]: hp = hp or DoublyRobustHP() ed = self._prepare_data(edp) isd = self._get_importance_sampling_inputs(ed) return self._get_importance_sampling_estimates(isd, hp=hp)
13,063
37.997015
88
py
ReAgent
ReAgent-master/reagent/evaluation/evaluator.py
#!/usr/bin/env python3 # Copyright (c) Facebook, Inc. and its affiliates. All rights reserved. import logging from collections import Counter from typing import Dict, List, Optional import torch import torch.nn.functional as F from reagent.core.tracker import observable from reagent.evaluation.cpe import CpeDetails, CpeEstimateSet from reagent.evaluation.doubly_robust_estimator import DoublyRobustEstimator from reagent.evaluation.evaluation_data_page import EvaluationDataPage from reagent.evaluation.sequential_doubly_robust_estimator import ( SequentialDoublyRobustEstimator, ) from reagent.evaluation.weighted_sequential_doubly_robust_estimator import ( WeightedSequentialDoublyRobustEstimator, ) logger = logging.getLogger(__name__) def get_tensor(x, dtype=None): """ Input: - x: list or a sequence - dtype: target data type of the elements in tensor [optional] It will be inferred automatically if not provided. Output: Tensor given a list or a sequence. If the input is None, it returns None If the input is a tensor it returns the tensor. If type is provides the output Tensor will have that type """ if x is None: return None if not isinstance(x, torch.Tensor): x = torch.tensor(x) if dtype is not None: x = x.type(dtype) return x def get_metrics_to_score(metric_reward_values: Optional[Dict[str, float]]) -> List[str]: if metric_reward_values is None: return [] return sorted([*metric_reward_values.keys()]) @observable(cpe_details=CpeDetails) class Evaluator: NUM_J_STEPS_FOR_MAGIC_ESTIMATOR = 25 def __init__(self, action_names, gamma, model, metrics_to_score=None) -> None: self.action_names = action_names self.metrics_to_score = metrics_to_score self.gamma = gamma self.model = model self.doubly_robust_estimator = DoublyRobustEstimator() self.sequential_doubly_robust_estimator = SequentialDoublyRobustEstimator(gamma) self.weighted_sequential_doubly_robust_estimator = ( WeightedSequentialDoublyRobustEstimator(gamma) ) def evaluate_post_training(self, edp: EvaluationDataPage) -> CpeDetails: cpe_details = CpeDetails() cpe_details.reward_estimates = self.score_cpe("Reward", edp) if ( self.metrics_to_score is not None and edp.logged_metrics is not None and self.action_names is not None ): for i, metric in enumerate(self.metrics_to_score): logger.info( "--------- Running CPE on metric: {} ---------".format(metric) ) metric_reward_edp = edp.set_metric_as_reward(i, len(self.action_names)) cpe_details.metric_estimates[metric] = self.score_cpe( metric, metric_reward_edp ) if self.action_names is not None: if edp.optimal_q_values is not None: value_means = edp.optimal_q_values.mean(dim=0) cpe_details.q_value_means = { action: float(value_means[i]) for i, action in enumerate(self.action_names) } # pyre-ignore [16]: `Optional` has no attribute `std` value_stds = edp.optimal_q_values.std(dim=0) cpe_details.q_value_stds = { action: float(value_stds[i]) for i, action in enumerate(self.action_names) } if edp.eval_action_idxs is not None: cpe_details.action_distribution = { # pyre-ignore [16]: `bool` has no attribute `sum` action: float((edp.eval_action_idxs == i).sum()) # pyre-ignore [16]: `Optional` has no attribute `shape` / edp.eval_action_idxs.shape[0] for i, action in enumerate(self.action_names) } # pyre-fixme[16]: `Evaluator` has no attribute `notify_observers`. self.notify_observers(cpe_details=cpe_details) return cpe_details def score_cpe(self, metric_name, edp: EvaluationDataPage): ( direct_method, inverse_propensity, doubly_robust, ) = self.doubly_robust_estimator.estimate(edp) sequential_doubly_robust = self.sequential_doubly_robust_estimator.estimate(edp) weighted_doubly_robust = ( self.weighted_sequential_doubly_robust_estimator.estimate( edp, num_j_steps=1, whether_self_normalize_importance_weights=True ) ) magic = self.weighted_sequential_doubly_robust_estimator.estimate( edp, num_j_steps=Evaluator.NUM_J_STEPS_FOR_MAGIC_ESTIMATOR, whether_self_normalize_importance_weights=True, ) return CpeEstimateSet( direct_method=direct_method, inverse_propensity=inverse_propensity, doubly_robust=doubly_robust, sequential_doubly_robust=sequential_doubly_robust, weighted_doubly_robust=weighted_doubly_robust, magic=magic, ) def _get_batch_logged_actions(self, arr): action_counter = Counter() for actions in arr: # torch.max() returns the element and the index. # The latter is the argmax equivalent _, argmax = torch.max(actions, dim=1) # Counter object does not work well with Tensors, hence casting back to numpy action_counter.update(Counter(argmax.numpy())) total_actions = 1.0 * sum(action_counter.values()) return ( { action_name: (action_counter[i] / total_actions) for i, action_name in enumerate(self.action_names) }, { action_name: action_counter[i] for i, action_name in enumerate(self.action_names) }, ) def get_target_distribution_error( self, actions, target_distribution, actual_distribution ): """Calculate MSE between actual and target action distribution.""" if not target_distribution: return None error = 0 for i, action in enumerate(actions): error += (target_distribution[i] - actual_distribution[action]) ** 2 return error / len(actions) @staticmethod def huberLoss(label, output): if abs(label - output) > 1: return abs(label - output) - 0.5 else: return 0.5 * (label - output) * (label - output)
6,718
35.516304
89
py
ReAgent
ReAgent-master/reagent/evaluation/cpe.py
#!/usr/bin/env python3 # Copyright (c) Facebook, Inc. and its affiliates. All rights reserved. import logging import math from typing import Dict, NamedTuple, Optional import numpy as np import torch from reagent.core.tensorboardX import SummaryWriterContext logger = logging.getLogger(__name__) class CpeEstimate(NamedTuple): raw: float normalized: float raw_std_error: float normalized_std_error: float class CpeEstimateSet(NamedTuple): direct_method: Optional[CpeEstimate] = None inverse_propensity: Optional[CpeEstimate] = None doubly_robust: Optional[CpeEstimate] = None sequential_doubly_robust: Optional[CpeEstimate] = None weighted_doubly_robust: Optional[CpeEstimate] = None magic: Optional[CpeEstimate] = None switch: Optional[CpeEstimate] = None switch_dr: Optional[CpeEstimate] = None def check_estimates_exist(self): assert self.direct_method is not None assert self.inverse_propensity is not None assert self.doubly_robust is not None assert self.sequential_doubly_robust is not None assert self.weighted_doubly_robust is not None assert self.magic is not None def log(self): self.check_estimates_exist() logger.info( "Reward Inverse Propensity Score : normalized {0:.3f} +/- {0:.3f} raw {1:.3f} +/- {1:.3f}".format( self.inverse_propensity.normalized, self.inverse_propensity.normalized_std_error, self.inverse_propensity.raw, self.inverse_propensity.raw_std_error, ) ) logger.info( "Reward Direct Method : normalized {0:.3f} +/- {0:.3f} raw {1:.3f} +/- {1:.3f}".format( self.direct_method.normalized, self.direct_method.normalized_std_error, self.direct_method.raw, self.direct_method.raw_std_error, ) ) logger.info( "Reward Doubly Robust P.E. : normalized {0:.3f} +/- {0:.3f} raw {1:.3f} +/- {1:.3f}".format( self.doubly_robust.normalized, self.doubly_robust.normalized_std_error, self.doubly_robust.raw, self.doubly_robust.raw_std_error, ) ) logger.info( "Value Weighted Doubly Robust P.E. : normalized {0:.3f} +/- {0:.3f} raw {1:.3f} +/- {1:.3f}".format( self.weighted_doubly_robust.normalized, self.weighted_doubly_robust.normalized_std_error, self.weighted_doubly_robust.raw, self.weighted_doubly_robust.raw_std_error, ) ) logger.info( "Value Sequential Doubly Robust P.E. : normalized {0:.3f} +/- {0:.3f} raw {1:.3f} +/- {1:.3f}".format( self.sequential_doubly_robust.normalized, self.sequential_doubly_robust.normalized_std_error, self.sequential_doubly_robust.raw, self.sequential_doubly_robust.raw_std_error, ) ) logger.info( "Value Magic Doubly Robust P.E. : normalized {0:.3f} +/- {0:.3f} raw {1:.3f} +/- {1:.3f}".format( self.magic.normalized, self.magic.normalized_std_error, self.magic.raw, self.magic.raw_std_error, ) ) def log_to_tensorboard(self, metric_name: str) -> None: self.check_estimates_exist() def none_to_zero(x: Optional[float]) -> float: if x is None or math.isnan(x): return 0.0 return x for name, value in [ ( "CPE/{}/Direct_Method_Reward".format(metric_name), # pyre-fixme[16]: `Optional` has no attribute `normalized`. self.direct_method.normalized, ), ( "CPE/{}/IPS_Reward".format(metric_name), self.inverse_propensity.normalized, ), ( "CPE/{}/Doubly_Robust_Reward".format(metric_name), self.doubly_robust.normalized, ), ( "CPE/{}/Sequential_Doubly_Robust".format(metric_name), self.sequential_doubly_robust.normalized, ), ( "CPE/{}/Weighted_Sequential_Doubly_Robust".format(metric_name), self.weighted_doubly_robust.normalized, ), ("CPE/{}/MAGIC".format(metric_name), self.magic.normalized), ]: SummaryWriterContext.add_scalar(name, none_to_zero(value)) def fill_empty_with_zero(self): retval = self for name, value in self._asdict().items(): if value is None: retval = retval._replace( **{ name: CpeEstimate( raw=0.0, normalized=0.0, raw_std_error=0.0, normalized_std_error=0.0, ) } ) return retval class CpeDetails: def __init__(self): self.reward_estimates: CpeEstimateSet = CpeEstimateSet() self.metric_estimates: Dict[str, CpeEstimateSet] = {} self.q_value_means: Optional[Dict[str, float]] = None self.q_value_stds: Optional[Dict[str, float]] = None self.action_distribution: Optional[Dict[str, float]] = None def log(self): logger.info("Reward Estimates:") logger.info("-----------------") self.reward_estimates.log() logger.info("-----------------") for metric in self.metric_estimates.keys(): logger.info(metric + " Estimates:") logger.info("-----------------") self.metric_estimates[metric].log() logger.info("-----------------") def log_to_tensorboard(self) -> None: self.reward_estimates.log_to_tensorboard("Reward") for metric_name, estimate_set in self.metric_estimates.items(): estimate_set.log_to_tensorboard(metric_name) def bootstrapped_std_error_of_mean(data, sample_percent=0.25, num_samples=1000): """ Compute bootstrapped standard error of mean of input data. :param data: Input data (1D torch tensor or numpy array). :param sample_percent: Size of sample to use to calculate bootstrap statistic. :param num_samples: Number of times to sample. """ if isinstance(data, torch.Tensor): data = data.cpu().numpy() sample_size = int(sample_percent * len(data)) means = [ np.mean(np.random.choice(data, size=sample_size, replace=True)) for i in range(num_samples) ] return np.std(means)
6,820
34.9
114
py
ReAgent
ReAgent-master/reagent/evaluation/weighted_sequential_doubly_robust_estimator.py
#!/usr/bin/env python3 # Copyright (c) Facebook, Inc. and its affiliates. All rights reserved. import itertools import logging import numpy as np import scipy as sp import torch from reagent.evaluation.cpe import CpeEstimate from reagent.evaluation.evaluation_data_page import EvaluationDataPage logger = logging.getLogger(__name__) class WeightedSequentialDoublyRobustEstimator: NUM_SUBSETS_FOR_CB_ESTIMATES = 25 CONFIDENCE_INTERVAL = 0.9 NUM_BOOTSTRAP_SAMPLES = 50 BOOTSTRAP_SAMPLE_PCT = 0.5 def __init__(self, gamma): self.gamma = gamma def estimate( self, edp: EvaluationDataPage, num_j_steps, whether_self_normalize_importance_weights, ) -> CpeEstimate: # For details, visit https://arxiv.org/pdf/1604.00923.pdf Section 5, 7, 8 assert edp.model_values is not None ( actions, rewards, logged_propensities, target_propensities, estimated_q_values, ) = WeightedSequentialDoublyRobustEstimator.transform_to_equal_length_trajectories( edp.mdp_id, edp.action_mask.cpu().numpy(), edp.logged_rewards.cpu().numpy().flatten(), edp.logged_propensities.cpu().numpy().flatten(), edp.model_propensities.cpu().numpy(), # pyre-ignore [16]: Optional type has no attribute `cpu` edp.model_values.cpu().numpy(), ) num_trajectories = actions.shape[0] trajectory_length = actions.shape[1] j_steps = [float("inf")] if num_j_steps > 1: j_steps.append(-1) if num_j_steps > 2: interval = trajectory_length // (num_j_steps - 1) j_steps.extend([i * interval for i in range(1, num_j_steps - 1)]) target_propensity_for_logged_action = np.sum( np.multiply(target_propensities, actions), axis=2 ) estimated_q_values_for_logged_action = np.sum( np.multiply(estimated_q_values, actions), axis=2 ) estimated_state_values = np.sum( np.multiply(target_propensities, estimated_q_values), axis=2 ) importance_weights = target_propensity_for_logged_action / logged_propensities importance_weights = np.cumprod(importance_weights, axis=1) importance_weights = ( WeightedSequentialDoublyRobustEstimator.normalize_importance_weights( importance_weights, whether_self_normalize_importance_weights ) ) importance_weights_one_earlier = ( np.ones([num_trajectories, 1]) * 1.0 / num_trajectories ) importance_weights_one_earlier = np.hstack( [importance_weights_one_earlier, importance_weights[:, :-1]] ) discounts = np.logspace( start=0, stop=trajectory_length - 1, num=trajectory_length, base=self.gamma ) j_step_return_trajectories = [] for j_step in j_steps: j_step_return_trajectories.append( WeightedSequentialDoublyRobustEstimator.calculate_step_return( rewards, discounts, importance_weights, importance_weights_one_earlier, estimated_state_values, estimated_q_values_for_logged_action, j_step, ) ) j_step_return_trajectories = np.array(j_step_return_trajectories) j_step_returns = np.sum(j_step_return_trajectories, axis=1) if len(j_step_returns) == 1: weighted_doubly_robust = j_step_returns[0] weighted_doubly_robust_std_error = 0.0 else: # break trajectories into several subsets to estimate confidence bounds infinite_step_returns = [] num_subsets = int( min( num_trajectories / 2, WeightedSequentialDoublyRobustEstimator.NUM_SUBSETS_FOR_CB_ESTIMATES, ) ) interval = num_trajectories / num_subsets for i in range(num_subsets): trajectory_subset = np.arange( int(i * interval), int((i + 1) * interval) ) importance_weights = ( target_propensity_for_logged_action[trajectory_subset] / logged_propensities[trajectory_subset] ) importance_weights = np.cumprod(importance_weights, axis=1) importance_weights = WeightedSequentialDoublyRobustEstimator.normalize_importance_weights( importance_weights, whether_self_normalize_importance_weights ) importance_weights_one_earlier = ( np.ones([len(trajectory_subset), 1]) * 1.0 / len(trajectory_subset) ) importance_weights_one_earlier = np.hstack( [importance_weights_one_earlier, importance_weights[:, :-1]] ) infinite_step_return = np.sum( WeightedSequentialDoublyRobustEstimator.calculate_step_return( rewards[trajectory_subset], discounts, importance_weights, importance_weights_one_earlier, estimated_state_values[trajectory_subset], estimated_q_values_for_logged_action[trajectory_subset], float("inf"), ) ) infinite_step_returns.append(infinite_step_return) # Compute weighted_doubly_robust mean point estimate using all data weighted_doubly_robust = self.compute_weighted_doubly_robust_point_estimate( j_steps, num_j_steps, j_step_returns, infinite_step_returns, j_step_return_trajectories, ) # Use bootstrapping to compute weighted_doubly_robust standard error bootstrapped_means = [] sample_size = int( WeightedSequentialDoublyRobustEstimator.BOOTSTRAP_SAMPLE_PCT * num_subsets ) for _ in range( WeightedSequentialDoublyRobustEstimator.NUM_BOOTSTRAP_SAMPLES ): random_idxs = np.random.choice(num_j_steps, sample_size, replace=False) random_idxs.sort() wdr_estimate = self.compute_weighted_doubly_robust_point_estimate( j_steps=[j_steps[i] for i in random_idxs], num_j_steps=sample_size, j_step_returns=j_step_returns[random_idxs], infinite_step_returns=infinite_step_returns, j_step_return_trajectories=j_step_return_trajectories[random_idxs], ) bootstrapped_means.append(wdr_estimate) weighted_doubly_robust_std_error = np.std(bootstrapped_means) episode_values = np.sum(np.multiply(rewards, discounts), axis=1) logged_policy_score = np.nanmean(episode_values) if logged_policy_score < 1e-6: logger.warning( "Can't normalize WSDR-CPE because of small or negative logged_policy_score" ) return CpeEstimate( raw=weighted_doubly_robust, normalized=0.0, raw_std_error=weighted_doubly_robust_std_error, normalized_std_error=0.0, ) return CpeEstimate( raw=weighted_doubly_robust, normalized=weighted_doubly_robust / logged_policy_score, raw_std_error=weighted_doubly_robust_std_error, normalized_std_error=weighted_doubly_robust_std_error / logged_policy_score, ) def compute_weighted_doubly_robust_point_estimate( self, j_steps, num_j_steps, j_step_returns, infinite_step_returns, j_step_return_trajectories, ): ( low_bound, high_bound, ) = WeightedSequentialDoublyRobustEstimator.confidence_bounds( infinite_step_returns, WeightedSequentialDoublyRobustEstimator.CONFIDENCE_INTERVAL, ) # decompose error into bias + variance j_step_bias = np.zeros([num_j_steps]) where_lower = np.where(j_step_returns < low_bound)[0] j_step_bias[where_lower] = low_bound - j_step_returns[where_lower] where_higher = np.where(j_step_returns > high_bound)[0] j_step_bias[where_higher] = j_step_returns[where_higher] - high_bound covariance = np.cov(j_step_return_trajectories) error = covariance + j_step_bias.T * j_step_bias # minimize mse error constraint = {"type": "eq", "fun": lambda x: np.sum(x) - 1.0} x = np.zeros([len(j_steps)]) res = sp.optimize.minimize( mse_loss, x, args=error, constraints=constraint, bounds=[(0, 1) for _ in range(x.shape[0])], ) x = np.array(res.x) return float(np.dot(x, j_step_returns)) @staticmethod def transform_to_equal_length_trajectories( mdp_ids, actions, rewards, logged_propensities, target_propensities, estimated_q_values, ): """ Take in samples (action, rewards, propensities, etc.) and output lists of equal-length trajectories (episodes) according to terminals. As the raw trajectories are of various lengths, the shorter ones are filled with zeros(ones) at the end. """ num_actions = len(target_propensities[0]) terminals = np.zeros(mdp_ids.shape[0]) for x in range(0, mdp_ids.shape[0]): if x + 1 == mdp_ids.shape[0] or mdp_ids[x, 0] != mdp_ids[x + 1, 0]: terminals[x] = 1 trajectories = [] episode_start = 0 episode_ends = np.nonzero(terminals)[0] if len(terminals) - 1 not in episode_ends: episode_ends = np.append(episode_ends, len(terminals) - 1) for episode_end in episode_ends: trajectories.append(np.arange(episode_start, episode_end + 1)) episode_start = episode_end + 1 action_trajectories = [] reward_trajectories = [] logged_propensity_trajectories = [] target_propensity_trajectories = [] Q_value_trajectories = [] for trajectory in trajectories: action_trajectories.append(actions[trajectory]) reward_trajectories.append(rewards[trajectory]) logged_propensity_trajectories.append(logged_propensities[trajectory]) target_propensity_trajectories.append(target_propensities[trajectory]) Q_value_trajectories.append(estimated_q_values[trajectory]) def to_equal_length(x, fill_value): x_equal_length = np.array( list(itertools.zip_longest(*x, fillvalue=fill_value)) ).swapaxes(0, 1) return x_equal_length action_trajectories = to_equal_length( action_trajectories, np.zeros([num_actions]) ) reward_trajectories = to_equal_length(reward_trajectories, 0) logged_propensity_trajectories = to_equal_length( logged_propensity_trajectories, 1 ) target_propensity_trajectories = to_equal_length( target_propensity_trajectories, np.zeros([num_actions]) ) Q_value_trajectories = to_equal_length( Q_value_trajectories, np.zeros([num_actions]) ) return ( action_trajectories, reward_trajectories, logged_propensity_trajectories, target_propensity_trajectories, Q_value_trajectories, ) @staticmethod def normalize_importance_weights( importance_weights, whether_self_normalize_importance_weights ): if whether_self_normalize_importance_weights: sum_importance_weights = np.sum(importance_weights, axis=0) where_zeros = np.where(sum_importance_weights == 0.0)[0] sum_importance_weights[where_zeros] = len(importance_weights) importance_weights[:, where_zeros] = 1.0 importance_weights /= sum_importance_weights return importance_weights else: importance_weights /= importance_weights.shape[0] return importance_weights @staticmethod def calculate_step_return( rewards, discounts, importance_weights, importance_weights_one_earlier, estimated_state_values, estimated_q_values, j_step, ): trajectory_length = len(rewards[0]) num_trajectories = len(rewards) j_step = int(min(j_step, trajectory_length - 1)) weighted_discounts = np.multiply(discounts, importance_weights) weighted_discounts_one_earlier = np.multiply( discounts, importance_weights_one_earlier ) importance_sampled_cumulative_reward = np.sum( np.multiply(weighted_discounts[:, : j_step + 1], rewards[:, : j_step + 1]), axis=1, ) if j_step < trajectory_length - 1: direct_method_value = ( weighted_discounts_one_earlier[:, j_step + 1] * estimated_state_values[:, j_step + 1] ) else: direct_method_value = np.zeros([num_trajectories]) control_variate = np.sum( np.multiply( weighted_discounts[:, : j_step + 1], estimated_q_values[:, : j_step + 1] ) - np.multiply( weighted_discounts_one_earlier[:, : j_step + 1], estimated_state_values[:, : j_step + 1], ), axis=1, ) j_step_return = ( importance_sampled_cumulative_reward + direct_method_value - control_variate ) return j_step_return @staticmethod def confidence_bounds(x, confidence): n = len(x) m, se = np.mean(x), sp.stats.sem(x) h = se * sp.stats.t._ppf((1 + confidence) / 2.0, n - 1) return m - h, m + h def mse_loss(x, error): return np.dot(np.dot(x, error), x.T)
14,541
36.576227
106
py
ReAgent
ReAgent-master/reagent/evaluation/ope_adapter.py
#!/usr/bin/env python3 # Copyright (c) Facebook, Inc. and its affiliates. All rights reserved. import logging import torch from reagent.evaluation.cpe import ( CpeEstimate, CpeEstimateSet, bootstrapped_std_error_of_mean, ) from reagent.evaluation.evaluation_data_page import EvaluationDataPage from reagent.evaluation.evaluator import Evaluator from reagent.evaluation.weighted_sequential_doubly_robust_estimator import ( WeightedSequentialDoublyRobustEstimator, ) from reagent.ope.estimators.contextual_bandits_estimators import ( BanditsEstimatorInput, DMEstimator, DoublyRobustEstimator, IPSEstimator, LogSample, ModelOutputs, ) from reagent.ope.estimators.estimator import ( Estimator, EstimatorResult, EstimatorResults, ) from reagent.ope.estimators.sequential_estimators import ( Action, ActionDistribution, DoublyRobustEstimator as SeqDREstimator, MAGICEstimator, RLEstimator, RLEstimatorInput, RLPolicy, State, Transition, ValueFunction, ) from reagent.ope.estimators.types import ActionSpace logger = logging.getLogger(__name__) class OPEstimatorAdapter: def __init__(self, ope_estimator: Estimator, device=None): self._ope_estimator = ope_estimator self._device = device @staticmethod def edp_to_contextual_bandit_log( edp: EvaluationDataPage, device=None ) -> BanditsEstimatorInput: log = [] n = edp.model_rewards.shape[0] for idx in range(n): # Action is only 1 if tgt policy and log policy took same action? action = torch.argmax(edp.action_mask[idx]).item() if edp.action_mask[idx][action] == 0.0: action = None logged_propensities = torch.zeros( edp.model_propensities[idx].shape, device=device ) if action is not None: logged_propensities[action] = edp.logged_propensities[idx] log.append( LogSample( context=None if edp.contexts is None else edp.contexts[idx], log_action=Action(action), log_reward=edp.logged_rewards[idx], log_action_probabilities=ActionDistribution(logged_propensities), tgt_action_probabilities=ActionDistribution( edp.model_propensities[idx] ), tgt_action=Action(action), model_outputs=ModelOutputs( tgt_reward_from_log_action=edp.model_rewards_for_logged_action[ idx ], tgt_rewards=edp.model_rewards[idx], ) # item features not specified as edp came from trained reward model ) ) return BanditsEstimatorInput(ActionSpace(edp.action_mask.shape[1]), log, True) @staticmethod def estimator_result_to_cpe_estimate(result: EstimatorResult) -> CpeEstimate: assert result.estimated_reward_normalized is not None assert result.estimated_reward_normalized is not None assert result.estimated_reward_std_error is not None assert result.estimated_reward_normalized_std_error is not None return CpeEstimate( raw=result.estimated_reward, normalized=result.estimated_reward_normalized, raw_std_error=result.estimated_reward_std_error, normalized_std_error=result.estimated_reward_normalized_std_error, ) def estimate(self, edp: EvaluationDataPage, **kwargs) -> CpeEstimate: result = self._ope_estimator.evaluate( OPEstimatorAdapter.edp_to_contextual_bandit_log(edp), **kwargs ) assert isinstance(result, EstimatorResult) logging.info(f"Got estimator result {result}, turning into cpe estimate") return OPEstimatorAdapter.estimator_result_to_cpe_estimate(result) class SequentialOPEstimatorAdapter: def __init__(self, seq_ope_estimator: RLEstimator, gamma: float, device=None): self.seq_ope_estimator = seq_ope_estimator self.gamma = gamma self._device = device class EDPSeqPolicy(RLPolicy): def __init__( self, num_actions: int, model_propensities: torch.Tensor, device=None ): super().__init__(ActionSpace(num_actions), device) self.model_propensities = model_propensities def action_dist(self, state: State) -> ActionDistribution: # "state" is (trajectory, step) return self.model_propensities[state.value] class EDPValueFunc(ValueFunction): def __init__( self, model_values: torch.Tensor, target_propensities: torch.Tensor ): self.model_values = model_values self.target_propensities = target_propensities def state_action_value(self, state: State, action: Action) -> float: return self.model_values[state.value][action].item() def state_value(self, state: State) -> float: return torch.dot( self.model_values[state.value], self.target_propensities[state.value] ).item() def reset(self): pass @staticmethod def edp_to_rl_input( edp: EvaluationDataPage, gamma, device=None ) -> RLEstimatorInput: assert edp.model_values is not None eq_len = WeightedSequentialDoublyRobustEstimator.transform_to_equal_length_trajectories( edp.mdp_id, edp.action_mask.cpu().numpy(), edp.logged_rewards.cpu().numpy().flatten(), edp.logged_propensities.cpu().numpy().flatten(), edp.model_propensities.cpu().numpy(), # pyre-ignore [16]: Optional type has no attribute `cpu` edp.model_values.cpu().numpy(), ) ( actions, rewards, logged_propensities, target_propensities, estimated_q_values, ) = ( torch.tensor(x, dtype=torch.double, device=device, requires_grad=True) for x in eq_len ) num_examples = logged_propensities.shape[0] horizon = logged_propensities.shape[1] log = [] for traj in range(num_examples): log.append( [ Transition( last_state=State((traj, i)), action=torch.argmax(actions[traj, i]).item(), action_prob=logged_propensities[traj, i].item(), state=State((traj, i + 1)), reward=rewards[traj, i].item(), ) for i in range(horizon - 1) if actions[traj, i][torch.argmax(actions[traj, i]).item()] != 0.0 ] ) return RLEstimatorInput( gamma=gamma, log=log, target_policy=SequentialOPEstimatorAdapter.EDPSeqPolicy( actions.shape[2], target_propensities ), value_function=SequentialOPEstimatorAdapter.EDPValueFunc( estimated_q_values, target_propensities ), ground_truth=None, horizon=horizon, ) @staticmethod def estimator_results_to_cpe_estimate( estimator_results: EstimatorResults, ) -> CpeEstimate: scores = torch.tensor( [r.estimated_reward for r in estimator_results.results], dtype=torch.double ) log_scores = torch.tensor( [r.log_reward for r in estimator_results.results], dtype=torch.double ) dr_score = float(torch.mean(scores).item()) dr_score_std_error = bootstrapped_std_error_of_mean(scores) log_score = float(torch.mean(log_scores).item()) if log_score < 1e-6: logger.warning( "Can't normalize SDR-CPE because of small" f" or negative logged_policy_score ({log_score})." f"Episode values: {log_scores}." ) return CpeEstimate( raw=dr_score, normalized=0.0, raw_std_error=dr_score_std_error, normalized_std_error=0.0, ) return CpeEstimate( raw=dr_score, normalized=dr_score / log_score, raw_std_error=dr_score_std_error, normalized_std_error=dr_score_std_error / log_score, ) def estimate(self, edp: EvaluationDataPage) -> CpeEstimate: estimator_results = self.seq_ope_estimator.evaluate( SequentialOPEstimatorAdapter.edp_to_rl_input(edp, self.gamma, self._device) ) assert isinstance(estimator_results, EstimatorResults) return SequentialOPEstimatorAdapter.estimator_results_to_cpe_estimate( estimator_results ) class OPEvaluator(Evaluator): def __init__( self, action_names, gamma, model, metrics_to_score=None, device=None ) -> None: super().__init__(action_names, gamma, model, metrics_to_score) self._device = device self.ope_dm_estimator = OPEstimatorAdapter(DMEstimator(device=self._device)) self.ope_ips_estimator = OPEstimatorAdapter(IPSEstimator(device=self._device)) self.ope_dr_estimator = OPEstimatorAdapter( DoublyRobustEstimator(device=self._device) ) self.ope_seq_dr_estimator = SequentialOPEstimatorAdapter( SeqDREstimator(device=self._device), gamma, device=self._device ) self.ope_seq_weighted_dr_estimator = SequentialOPEstimatorAdapter( SeqDREstimator(weighted=True, device=self._device), gamma, device=self._device, ) self.ope_seq_magic_estimator = SequentialOPEstimatorAdapter( MAGICEstimator(device=self._device), gamma ) def score_cpe(self, metric_name, edp: EvaluationDataPage): logger.info("Using OPE adapter") direct_method = self.ope_dm_estimator.estimate(edp) inverse_propensity = self.ope_ips_estimator.estimate(edp) doubly_robust = self.ope_dr_estimator.estimate(edp) sequential_doubly_robust = self.ope_seq_dr_estimator.estimate(edp) weighted_doubly_robust = self.ope_seq_weighted_dr_estimator.estimate(edp) magic = self.ope_seq_magic_estimator.estimate(edp) return CpeEstimateSet( direct_method=direct_method, inverse_propensity=inverse_propensity, doubly_robust=doubly_robust, sequential_doubly_robust=sequential_doubly_robust, weighted_doubly_robust=weighted_doubly_robust, magic=magic, )
10,869
36.353952
96
py
ReAgent
ReAgent-master/reagent/evaluation/feature_importance/feature_importance_perturbation.py
#!/usr/bin/env python3 # Copyright (c) Facebook, Inc. and its affiliates. All rights reserved. import copy import logging from collections import defaultdict from typing import Callable, Any, Optional import pandas as pd import torch import torch.nn as nn from reagent.core.dataclasses import dataclass from reagent.evaluation.feature_importance.feature_importance_base import ( FeatureImportanceBase, ) logger = logging.getLogger(__name__) @dataclass class FeatureImportancePerturbation(FeatureImportanceBase): data_loader: Any # Consume model (first arg) and data (second arg) to make model predictions # Expected to return a tensor of shape (batch_size, 1) pred_fn: Callable[[nn.Module, Any], torch.Tensor] # Perturb data (first arg) on a specific feature id (second arg) perturb_fn: Callable[[Any, int], Any] # How many rounds of perturbations for collecting feature importance for each batch # The higher it is, the less variance the result will have repeat: int = 1 def compute_feature_importance(self) -> pd.DataFrame: feature_importance_vals = defaultdict(list) for batch_idx, data in enumerate(self.data_loader): for r in range(self.repeat): pred_value = self.pred_fn(self.model, data) for feature_idx, feature_id in enumerate(self.sorted_feature_ids): copy_data = copy.deepcopy(data) perturbed_data = self.perturb_fn(copy_data, feature_idx) perturbed_pred_value = self.pred_fn(self.model, perturbed_data) feature_importance_vals[feature_id].append( torch.mean( torch.abs(perturbed_pred_value - pred_value) ).detach() ) logger.info(f"Processed {batch_idx} batches {r}-th time") feature_importance_mean = { k: torch.mean(torch.stack(v)).item() for k, v in feature_importance_vals.items() } result_df = pd.DataFrame.from_dict( feature_importance_mean, orient="index", columns=["feature_importance"] ).sort_values(by=["feature_importance"], ascending=False) # Fblearner UI can't show row names (index). So manually add names as a column result_df.insert(0, "feature_id", result_df.index) return result_df def create_default_perturb_fn(key: str): def default_perturb_fn( data, feature_idx, ): val_data, presence_data = data[key] batch_size = val_data.shape[0] random_idx = torch.randperm(batch_size) val_data[:, feature_idx] = val_data[:, feature_idx][random_idx] presence_data[:, feature_idx] = presence_data[:, feature_idx][random_idx] return data return default_perturb_fn
2,857
37.106667
87
py
ReAgent
ReAgent-master/reagent/evaluation/feature_importance/feature_importance_base.py
#!/usr/bin/env python3 # Copyright (c) Facebook, Inc. and its affiliates. All rights reserved. from typing import List import pandas as pd import torch.nn as nn from reagent.core.dataclasses import dataclass @dataclass class FeatureImportanceBase: model: nn.Module sorted_feature_ids: List[int] def compute_feature_importance(self) -> pd.DataFrame: raise NotImplementedError()
401
22.647059
71
py
ReAgent
ReAgent-master/reagent/core/aggregators.py
#!/usr/bin/env python3 # Copyright (c) Facebook, Inc. and its affiliates. All rights reserved. import logging from collections import deque from typing import Callable, Deque, Dict, List, Optional, Any import numpy as np import torch from reagent.core.tensorboardX import SummaryWriterContext from reagent.core.tracker import Aggregator logger = logging.getLogger(__name__) class TensorAggregator(Aggregator): def __call__(self, key: str, values): # Ensure that tensor is on cpu before aggregation. values = torch.cat(values, dim=0).cpu() return super().__call__(key, values) def _log_histogram_and_mean(log_key, val): try: SummaryWriterContext.add_histogram(log_key, val) SummaryWriterContext.add_scalar(f"{log_key}/mean", val.mean()) except ValueError: logger.warning( f"Cannot create histogram for key: {log_key}; " "this is likely because you have NULL value in your input; " f"value: {val}" ) raise class TensorBoardHistogramAndMeanAggregator(TensorAggregator): def __init__(self, key: str, log_key: str): super().__init__(key) self.log_key = log_key def aggregate(self, values): assert len(values.shape) == 1 or ( len(values.shape) == 2 and values.shape[1] == 1 ), f"Unexpected shape for {self.key}: {values.shape}" _log_histogram_and_mean(self.log_key, values) class TensorBoardActionHistogramAndMeanAggregator(TensorAggregator): def __init__( self, key: str, category: str, title: str, actions: List[str], log_key_prefix: Optional[str] = None, ): super().__init__(key) self.log_key_prefix = log_key_prefix or f"{category}/{title}" self.actions = actions SummaryWriterContext.add_custom_scalars_multilinechart( [f"{self.log_key_prefix}/{action_name}/mean" for action_name in actions], category=category, title=title, ) def aggregate(self, values): if not (len(values.shape) == 2 and values.shape[1] == len(self.actions)): raise ValueError( "Unexpected shape for {}: {}; actions: {}".format( self.key, values.shape, self.actions ) ) for i, action in enumerate(self.actions): _log_histogram_and_mean(f"{self.log_key_prefix}/{action}", values[:, i]) class TensorBoardActionCountAggregator(TensorAggregator): def __init__(self, key: str, title: str, actions: List[str]): super().__init__(key) self.log_key = f"actions/{title}" self.actions = actions SummaryWriterContext.add_custom_scalars_multilinechart( [f"{self.log_key}/{action_name}" for action_name in actions], category="actions", title=title, ) def aggregate(self, values): for i, action in enumerate(self.actions): SummaryWriterContext.add_scalar( f"{self.log_key}/{action}", (values == i).sum().item() ) class MeanAggregator(TensorAggregator): def __init__(self, key: str): super().__init__(key) self.values: List[float] = [] def aggregate(self, values): mean = values.mean().item() logger.info(f"{self.key}: {mean}") self.values.append(mean) class ListAggregator(Aggregator): def __init__(self, key: str): super().__init__(key) self.values: Optional[Any] = [] def aggregate(self, values): self.values.extend(values) class EpochListAggregator(TensorAggregator): def __init__(self, key: str): super().__init__(key) self.values: List = [] self.epoch_values: List = [] def aggregate(self, values): flattened = torch.flatten(values).tolist() self.values.extend(flattened) def flush(self): if self.values: self.epoch_values = self.values self.values = [] class FunctionsByActionAggregator(TensorAggregator): """ Aggregating the input by action, using the given functions. The input is assumed to be an `N x D` tensor, where each column is an action, and each row is an example. This takes a dictionary of functions so that the values only need to be concatenated once. Example: agg = FunctionByActionAggregator( "model_values", ["A", "B], {"mean": torch.mean, "std": torch.std} ) input = torch.tensor([ [0.9626, 0.7142], [0.7216, 0.5426], [0.4225, 0.9485], ]) agg(input) input2 = torch.tensor([ [0.0103, 0.0306], [0.9846, 0.8373], [0.4614, 0.0174], ]) agg(input2) print(agg.values) { "mean": { "A": [0.7022, 0.4854], "B": [0.7351, 0.2951], }, "std": { "A": [0.2706, 0.4876], "B": [0.2038, 0.4696], } } """ def __init__(self, key: str, actions: List[str], fns: Dict[str, Callable]): super().__init__(key) self.actions = actions self.values: Dict[str, Dict[str, List[float]]] = { fn: {action: [] for action in self.actions} for fn in fns } self.fns = fns def aggregate(self, values): for name, func in self.fns.items(): aggregated_values = func(values, dim=0) for action, value in zip(self.actions, aggregated_values): value = value.item() self.values[name][action].append(value) latest_values = { action: values[-1] for action, values in self.values[name].items() } logger.info(f"{name} {self.key} {latest_values}") class ActionCountAggregator(TensorAggregator): """ Counting the frequency of each action. Actions are indexed from `0` to `len(actions) - 1`. The input is assumed to contain action index. """ def __init__(self, key: str, actions: List[str]): super().__init__(key) self.actions = actions self.values: Dict[str, List[int]] = {action: [] for action in actions} def aggregate(self, values): for i, action in enumerate(self.actions): self.values[action].append((values == i).sum().item()) latest_count = {action: counts[-1] for action, counts in self.values.items()} logger.info(f"{self.key} {latest_count}") def get_distributions(self) -> Dict[str, List[float]]: """ Returns the action disributions in each aggregating step """ totals = np.array([sum(counts) for counts in zip(*self.values.values())]) return { action: (np.array(counts) / totals).tolist() for action, counts in self.values.items() } def get_cumulative_distributions(self) -> Dict[str, float]: """ Returns the cumulative distributions in each aggregating step """ totals = sum(sum(counts) for counts in zip(*self.values.values())) return {action: sum(counts) / totals for action, counts in self.values.items()} _RECENT_DEFAULT_SIZE = int(1e6) class RecentValuesAggregator(TensorAggregator): def __init__(self, key: str, size: int = _RECENT_DEFAULT_SIZE): super().__init__(key) self.values: Deque[float] = deque(maxlen=size) def aggregate(self, values): flattened = torch.flatten(values).tolist() self.values.extend(flattened)
7,651
30.751037
87
py
ReAgent
ReAgent-master/reagent/core/tracker.py
#!/usr/bin/env python3 # Copyright (c) Facebook, Inc. and its affiliates. All rights reserved. import functools import logging from typing import Dict, List, Type import torch logger = logging.getLogger(__name__) class Observer: """ Base class for observers """ def __init__(self, observing_keys: List[str]): super().__init__() assert isinstance(observing_keys, list) self.observing_keys = observing_keys def get_observing_keys(self) -> List[str]: return self.observing_keys def update(self, key: str, value): pass class Aggregator: def __init__(self, key: str): super().__init__() self.key = key def __call__(self, key: str, values): assert key == self.key, f"Got {key}; expected {self.key}" self.aggregate(values) def aggregate(self, values): pass def flush(self): pass class ObservableMixin: def __init__(self): super().__init__() self._observers = {v: [] for v in self._observable_value_types} @property def _observable_value_types(self) -> Dict[str, Type]: raise NotImplementedError def add_observer(self, observer: Observer): observing_keys = observer.get_observing_keys() unknown_keys = [ k for k in observing_keys if k not in self._observable_value_types ] if unknown_keys: logger.warning(f"{unknown_keys} cannot be observed in {type(self)}") for k in observing_keys: if k in self._observers and observer not in self._observers[k]: self._observers[k].append(observer) return self def add_observers(self, observers: List[Observer]): for observer in observers: self.add_observer(observer) return self def notify_observers(self, **kwargs): for key, value in kwargs.items(): if value is None: # Allow optional reporting continue assert key in self._observers, f"Unknown key: {key}" # TODO: Create a generic framework for type conversion if self._observable_value_types[key] == torch.Tensor: try: if not isinstance(value, torch.Tensor): value = torch.tensor(value) if len(value.shape) == 0: value = value.reshape(1) value = value.detach() except Exception: # Be lenient about conversion since ReporterBase # has inaccurate type pass for observer in self._observers[key]: observer.update(key, value) def observable(cls=None, **kwargs): # noqa: C901 """ Decorator to mark a class as producing observable values. The names of the observable values are the names of keyword arguments. The values of keyword arguments are the types of the value. The type is currently not used for anything. """ assert kwargs observable_value_types = kwargs def wrap(cls): assert not hasattr(cls, "add_observer") assert not hasattr(cls, "notify_observers") original_init = cls.__init__ @functools.wraps(original_init) def new_init(self, *args, **kwargs): original_init(self, *args, **kwargs) assert not hasattr(self, "_observable_value_types") assert not hasattr(self, "_observers") self._observable_value_types = observable_value_types self._observers = {v: [] for v in observable_value_types} cls.__init__ = new_init cls.add_observer = ObservableMixin.add_observer cls.add_observers = ObservableMixin.add_observers cls.notify_observers = ObservableMixin.notify_observers return cls if cls is None: return wrap return wrap(cls)
3,958
28.110294
80
py
ReAgent
ReAgent-master/reagent/core/utils.py
#!/usr/bin/env python3 from typing import Tuple, Optional import torch def get_rank() -> int: """ Returns the torch.distributed rank of the process. 0 represents the main process and is the default if torch.distributed isn't set up """ return ( torch.distributed.get_rank() if torch.distributed.is_available() and torch.distributed.is_initialized() else 0 ) class lazy_property(object): """ More or less copy-pasta: http://stackoverflow.com/a/6849299 Meant to be used for lazy evaluation of an object attribute. property should represent non-mutable data, as it replaces itself. """ def __init__(self, fget): self._fget = fget self.__doc__ = fget.__doc__ self.__name__ = fget.__name__ def __get__(self, obj, obj_cls_type): if obj is None: return None value = self._fget(obj) setattr(obj, self.__name__, value) return value
975
24.684211
82
py
ReAgent
ReAgent-master/reagent/core/configuration.py
#!/usr/bin/python3 import functools from dataclasses import MISSING, Field, fields from inspect import Parameter, isclass, signature from typing import List, Optional, Type, Union from reagent.core.dataclasses import dataclass from torch import nn BLOCKLIST_TYPES = [nn.Module] def _get_param_annotation(p): # if not annotated, infer type from default if p.annotation == Parameter.empty and p.default == Parameter.empty: raise ValueError( f"Param {p}: both annotation and default are empty, " "so cannot infer any useful annotation." ) if p.annotation != Parameter.empty: return p.annotation # case on default types if p.default is None: raise ValueError( f"Param {p}: default is None and annotation is empty, " "cannot infer useful annotation" ) if isinstance(p.default, tuple): raise ValueError(f"Param {p}: default is tuple, cannot infer type") if isinstance(p.default, dict): raise ValueError(f"Param{p}: default is tuple, cannot infer type") return type(p.default) def make_config_class( func, allowlist: Optional[List[str]] = None, blocklist: Optional[List[str]] = None, blocklist_types: List[Type] = BLOCKLIST_TYPES, ): """ Create a decorator to create dataclass with the arguments of `func` as fields. Only annotated arguments are converted to fields. If the default value is mutable, you must use `dataclass.field(default_factory=default_factory)` as default. In that case, the func has to be wrapped with @resolve_defaults below. `allowlist` & `blocklist` are mutually exclusive. """ parameters = signature(func).parameters assert ( allowlist is None or blocklist is None ), "allowlist & blocklist are mutually exclusive" blocklist_set = set(blocklist or []) def _is_type_blocklisted(t): if getattr(t, "__origin__", None) is Union: assert len(t.__args__) == 2 and t.__args__[1] == type( None ), "Only Unions of [X, None] (a.k.a. Optional[X]) are supported" t = t.__args__[0] if hasattr(t, "__origin__"): t = t.__origin__ assert isclass(t), f"{t} is not a class." return any(issubclass(t, blocklist_type) for blocklist_type in blocklist_types) def _is_valid_param(p): if p.name in blocklist_set: return False if p.annotation == Parameter.empty and p.default == Parameter.empty: return False ptype = _get_param_annotation(p) if _is_type_blocklisted(ptype): return False return True allowlist = allowlist or [p.name for p in parameters.values() if _is_valid_param(p)] def wrapper(config_cls): # Add __annotations__ for dataclass config_cls.__annotations__ = { field_name: _get_param_annotation(parameters[field_name]) for field_name in allowlist } # Set default values for field_name in allowlist: default = parameters[field_name].default if default != Parameter.empty: setattr(config_cls, field_name, default) # Add hashing to support hashing list and dict config_cls.__hash__ = param_hash # Add non-recursive asdict(). dataclasses.asdict() is recursive def asdict(self): return {field.name: getattr(self, field.name) for field in fields(self)} config_cls.asdict = asdict return dataclass(frozen=True)(config_cls) return wrapper def _resolve_default(val): if not isinstance(val, Field): return val if val.default != MISSING: return val.default if val.default_factory != MISSING: return val.default_factory() raise ValueError("No default value") def resolve_defaults(func): """ Use this decorator to resolve default field values in the constructor. """ func_params = list(signature(func).parameters.values()) @functools.wraps(func) def wrapper(*args, **kwargs): if len(args) > len(func_params): raise ValueError( f"There are {len(func_params)} parameters in total, " f"but args is {len(args)} long. \n" f"{args}" ) # go through unprovided default kwargs for p in func_params[len(args) :]: # only resolve defaults for Fields if isinstance(p.default, Field): if p.name not in kwargs: kwargs[p.name] = _resolve_default(p.default) return func(*args, **kwargs) return wrapper def param_hash(p): """ Use this to make parameters hashable. This is required because __hash__() is not inherited when subclass redefines __eq__(). We only need this when the parameter dataclass has a list or dict field. """ return hash(tuple(_hash_field(getattr(p, f.name)) for f in fields(p))) def _hash_field(val): """ Returns hashable value of the argument. A list is converted to a tuple. A dict is converted to a tuple of sorted pairs of key and value. """ if isinstance(val, list): return tuple(val) elif isinstance(val, dict): return tuple(sorted(val.items())) else: return val
5,364
31.515152
88
py
ReAgent
ReAgent-master/reagent/core/types.py
#!/usr/bin/env python3 # Copyright (c) Facebook, Inc. and its affiliates. All rights reserved. import dataclasses import logging # The dataclasses in this file should be vanilla dataclass to have minimal overhead from dataclasses import dataclass, field from typing import Dict, List, NamedTuple, Optional, Tuple, Union # Triggering registration to registries import reagent.core.result_types # noqa import torch import torch.nn.functional as F from reagent.core.base_dataclass import BaseDataClass from reagent.core.configuration import param_hash from reagent.core.dataclasses import dataclass as pydantic_dataclass from reagent.core.fb_checker import IS_FB_ENVIRONMENT from reagent.core.registry_meta import wrap_oss_with_dataclass from reagent.core.tagged_union import TaggedUnion from reagent.core.torch_utils import gather from reagent.model_utils.seq2slate_utils import DECODER_START_SYMBOL, subsequent_mask from reagent.preprocessing.types import InputColumn if IS_FB_ENVIRONMENT: import reagent.core.fb.fb_result_types # noqa class NoDuplicatedWarningLogger: def __init__(self, logger): self.logger = logger self.msg = set() def warning(self, msg): if msg not in self.msg: self.logger.warning(msg) self.msg.add(msg) logger = logging.getLogger(__name__) no_dup_logger = NoDuplicatedWarningLogger(logger) def isinstance_namedtuple(x): return isinstance(x, tuple) and hasattr(x, "_fields") @dataclass class TensorDataClass(BaseDataClass): def __getattr__(self, attr): if attr.startswith("__") and attr.endswith("__"): raise AttributeError tensor_attr = getattr(torch.Tensor, attr, None) if tensor_attr is None or not callable(tensor_attr): # TODO: can we get this working well with jupyter? logger.error( f"Attempting to call {self.__class__.__name__}.{attr} on " f"{type(self)} (instance of TensorDataClass)." ) if tensor_attr is None: raise AttributeError( f"{self.__class__.__name__}doesn't have {attr} attribute." ) else: raise RuntimeError(f"{self.__class__.__name__}.{attr} is not callable.") def continuation(*args, **kwargs): def f(v): # if possible, returns v.attr(*args, **kwargs). # otws, return v if isinstance(v, (torch.Tensor, TensorDataClass)): return getattr(v, attr)(*args, **kwargs) elif isinstance(v, dict): return {kk: f(vv) for kk, vv in v.items()} elif isinstance(v, tuple): return tuple(f(vv) for vv in v) return v return type(self)(**f(self.__dict__)) return continuation def cuda(self, *args, **kwargs): cuda_tensor = {} for k, v in self.__dict__.items(): # noqa F402 if isinstance(v, torch.Tensor): kwargs["non_blocking"] = kwargs.get("non_blocking", True) cuda_tensor[k] = v.cuda(*args, **kwargs) elif isinstance(v, TensorDataClass): cuda_tensor[k] = v.cuda(*args, **kwargs) else: cuda_tensor[k] = v return type(self)(**cuda_tensor) def cpu(self): cpu_tensor = {} for k, v in self.__dict__.items(): # noqa F402 if isinstance(v, (torch.Tensor, TensorDataClass)): cpu_tensor[k] = v.cpu() else: cpu_tensor[k] = v return type(self)(**cpu_tensor) # (offset, value) IdListFeatureValue = Tuple[torch.Tensor, torch.Tensor] # (offset, key, value) IdScoreListFeatureValue = Tuple[torch.Tensor, torch.Tensor, torch.Tensor] # name -> value IdListFeature = Dict[str, IdListFeatureValue] IdScoreListFeature = Dict[str, IdScoreListFeatureValue] # id -> value ServingIdListFeature = Dict[int, IdListFeatureValue] ServingIdScoreListFeature = Dict[int, IdScoreListFeatureValue] ##### # FIXME: These config types are misplaced but we need to write FBL config adapter # if we moved them. ###### @pydantic_dataclass class IdListFeatureConfig(BaseDataClass): name: str # integer feature ID feature_id: int # name of the embedding table to use id_mapping_name: str @pydantic_dataclass class IdScoreListFeatureConfig(BaseDataClass): name: str # integer feature ID feature_id: int # name of the embedding table to use id_mapping_name: str @pydantic_dataclass class FloatFeatureInfo(BaseDataClass): name: str feature_id: int @pydantic_dataclass class ExplicitMapping(object): __hash__ = param_hash ids: List[int] = field(default_factory=list) def __post_init_post_parse__(self): """ used in preprocessing ids list represents mapping from idx -> value we want the reverse: from feature to embedding table indices """ self._id2index: Dict[int, int] = {} @property def id2index(self) -> Dict[int, int]: # pyre-fixme[16]: `IdMapping` has no attribute `_id2index`. if not self._id2index: self._id2index = {id: i for i, id in enumerate(self.ids)} return self._id2index @property def table_size(self): return len(self.ids) @pydantic_dataclass class ModuloMapping: """ Map IDs to [0, table_size) via modulo `table_size` """ table_size: int @wrap_oss_with_dataclass class IdMappingUnion(TaggedUnion): explicit_mapping: Optional[ExplicitMapping] = None modulo: Optional[ModuloMapping] = None @pydantic_dataclass class ModelFeatureConfig(BaseDataClass): float_feature_infos: List[FloatFeatureInfo] = field(default_factory=list) # table name -> id mapping id_mapping_config: Dict[str, IdMappingUnion] = field(default_factory=dict) # id_list_feature_configs is feature_id -> list of values id_list_feature_configs: List[IdListFeatureConfig] = field(default_factory=list) # id_score_list_feature_configs is feature_id -> (keys -> values) id_score_list_feature_configs: List[IdScoreListFeatureConfig] = field( default_factory=list ) def __post_init_post_parse__(self): both_lists = self.id_list_feature_configs + self.id_score_list_feature_configs if not self.only_dense: # sanity check for keys in mapping config ids = [config.feature_id for config in both_lists] names = [config.name for config in both_lists] assert len(ids) == len(set(ids)), f"duplicates in ids: {ids}" assert len(names) == len(set(names)), f"duplicates in names: {names}" assert len(ids) == len(names), f"{len(ids)} != {len(names)}" self._id2name = {config.feature_id: config.name for config in both_lists} self._name2id = {config.name: config.feature_id for config in both_lists} self._id2config = {config.feature_id: config for config in both_lists} self._name2config = {config.name: config for config in both_lists} @property def only_dense(self): return not (self.id_list_feature_configs or self.id_score_list_feature_configs) @property def id2name(self): return self._id2name @property def name2id(self): return self._name2id @property def id2config(self): return self._id2config @property def name2config(self): return self._name2config ###### # dataclasses for internal API ###### @dataclass class ValuePresence(TensorDataClass): value: torch.Tensor presence: Optional[torch.Tensor] @dataclass class ActorOutput(TensorDataClass): action: torch.Tensor log_prob: Optional[torch.Tensor] = None squashed_mean: Optional[torch.Tensor] = None @dataclass class DocList(TensorDataClass): # the shape is (batch_size, num_candidates, num_document_features) float_features: torch.Tensor # the shapes below are (batch_size, num_candidates) # mask indicates whether the candidate is present or not; its dtype is torch.bool # pyre-fixme[8]: Attribute has type `Tensor`; used as `None`. mask: torch.Tensor = None # value is context dependent; it could be action probability or the score # of the document from another model # pyre-fixme[8]: Attribute has type `Tensor`; used as `None`. value: torch.Tensor = None def __post_init__(self): assert ( len(self.float_features.shape) == 3 ), f"Unexpected shape: {self.float_features.shape}" if self.mask is None: self.mask = self.float_features.new_ones( self.float_features.shape[:2], dtype=torch.bool ) if self.value is None: self.value = self.float_features.new_ones(self.float_features.shape[:2]) # pyre-fixme[56]: Decorator `torch.no_grad(...)` could not be called, because # its type `no_grad` is not callable. @torch.no_grad() def select_slate(self, action: torch.Tensor): row_idx = torch.repeat_interleave( torch.arange(action.shape[0]).unsqueeze(1), action.shape[1], dim=1 ) mask = self.mask[row_idx, action] float_features = self.float_features[row_idx, action] value = self.value[row_idx, action] return DocList(float_features, mask, value) def as_feature_data(self): _batch_size, _slate_size, feature_dim = self.float_features.shape return FeatureData(self.float_features.view(-1, feature_dim)) @dataclass class FeatureData(TensorDataClass): # For dense features, shape is (batch_size, feature_dim) float_features: torch.Tensor id_list_features: IdListFeature = dataclasses.field(default_factory=dict) id_score_list_features: IdScoreListFeature = dataclasses.field(default_factory=dict) # For sequence, shape is (stack_size, batch_size, feature_dim) stacked_float_features: Optional[torch.Tensor] = None # For ranking algos, candidate_docs: Optional[DocList] = None # Experimental: sticking this here instead of putting it in float_features # because a lot of places derive the shape of float_features from # normalization parameters. time_since_first: Optional[torch.Tensor] = None def __post_init__(self): def usage(): return ( "For sequence features, use `stacked_float_features`." "For document features, use `candidate_doc_float_features`." ) if self.float_features.ndim == 3: no_dup_logger.warning(f"`float_features` should be 2D.\n{usage()}") elif self.float_features.ndim != 2: raise ValueError( f"float_features should be 2D; got {self.float_features.shape}.\n{usage()}" ) @property def has_float_features_only(self) -> bool: return ( not self.id_list_features and self.time_since_first is None and self.candidate_docs is None ) def get_tiled_batch(self, num_tiles: int): assert ( self.has_float_features_only ), f"only works for float features now: {self}" """ tiled_feature should be (batch_size * num_tiles, feature_dim) forall i in [batch_size], tiled_feature[i*num_tiles:(i+1)*num_tiles] should be feat[i] """ feat = self.float_features assert ( len(feat.shape) == 2 ), f"Need feat shape to be (batch_size, feature_dim), got {feat.shape}." batch_size, _ = feat.shape tiled_feat = feat.repeat_interleave(repeats=num_tiles, dim=0) return FeatureData(float_features=tiled_feat) def concat_user_doc(self): assert not self.has_float_features_only, "only works when DocList present" assert self.float_features.dim() == 2 # batch_size x state_dim batch_size, state_dim = self.float_features.shape # batch_size x num_docs x candidate_dim assert self.candidate_docs.float_features.dim() == 3 assert len(self.candidate_docs.float_features) == batch_size _, num_docs, candidate_dim = self.candidate_docs.float_features.shape state_tiled = ( torch.repeat_interleave(self.float_features, num_docs, dim=0) .reshape(batch_size, num_docs, state_dim) .float() ) return torch.cat((state_tiled, self.candidate_docs.float_features), dim=2) def get_ranking_state(self, has_user_feat: bool): if has_user_feat: return self.concat_user_doc() else: # pyre-fixme[16]: `Optional` has no attribute `float_features`. return self.candidate_docs.float_features.float() def _embed_states(x: FeatureData) -> FeatureData: """ Get dense feature from float and doc features. TODO: make this an embedder. """ assert x.candidate_docs is not None def _concat_state_candidates(state: torch.Tensor, candidates: torch.Tensor): """ Expect state.shape = (n, state_dim), candidate.shape = (n, num_candidates, candidate_dim), Result has shape (n, state_dim + candidate_dim) [state, mean of candidates] """ n = state.shape[0] assert len(state.shape) == 2, f"{state.shape} != (batch_size, user_dim)" assert ( len(candidates.shape) == 3 ), f"{candidates.shape} != (batch_size, num_candidates, candidate_dim)" assert candidates.shape[0] == n, f"{candidates.shape} 0th dim != {n}" # TODO: have an embedder here # NOTE: mean aggregation is not very effective here candidates_embedding = candidates.view(n, -1) return torch.cat([state, candidates_embedding], dim=1) return FeatureData( float_features=_concat_state_candidates( x.float_features, x.candidate_docs.float_features, ) ) class TensorFeatureData(torch.nn.Module): """ Primarily for using in nn.Sequential """ def forward(self, input: torch.Tensor) -> FeatureData: assert isinstance(input, torch.Tensor) return FeatureData(input) class ServingFeatureData(NamedTuple): float_features_with_presence: Tuple[torch.Tensor, torch.Tensor] id_list_features: ServingIdListFeature id_score_list_features: ServingIdScoreListFeature @dataclass class ExtraData(TensorDataClass): mdp_id: Optional[torch.Tensor] = None sequence_number: Optional[torch.Tensor] = None action_probability: Optional[torch.Tensor] = None max_num_actions: Optional[int] = None metrics: Optional[torch.Tensor] = None @classmethod def from_dict(cls, d): return cls(**{f.name: d.get(f.name, None) for f in dataclasses.fields(cls)}) @dataclass class PreprocessedRankingInput(TensorDataClass): state: FeatureData src_seq: FeatureData src_src_mask: Optional[torch.Tensor] = None tgt_in_seq: Optional[FeatureData] = None tgt_out_seq: Optional[FeatureData] = None tgt_tgt_mask: Optional[torch.Tensor] = None slate_reward: Optional[torch.Tensor] = None position_reward: Optional[torch.Tensor] = None # all indices will be +2 to account for padding # symbol (0) and decoder_start_symbol (1) src_in_idx: Optional[torch.Tensor] = None tgt_in_idx: Optional[torch.Tensor] = None tgt_out_idx: Optional[torch.Tensor] = None tgt_out_probs: Optional[torch.Tensor] = None # store ground-truth target sequences optim_tgt_in_idx: Optional[torch.Tensor] = None optim_tgt_out_idx: Optional[torch.Tensor] = None optim_tgt_in_seq: Optional[FeatureData] = None optim_tgt_out_seq: Optional[FeatureData] = None extras: Optional[ExtraData] = field(default_factory=ExtraData) def batch_size(self) -> int: return self.state.float_features.size()[0] def __len__(self) -> int: return self.batch_size() @classmethod def from_input( cls, state: torch.Tensor, candidates: torch.Tensor, device: torch.device, action: Optional[torch.Tensor] = None, optimal_action: Optional[torch.Tensor] = None, logged_propensities: Optional[torch.Tensor] = None, slate_reward: Optional[torch.Tensor] = None, position_reward: Optional[torch.Tensor] = None, extras: Optional[ExtraData] = None, ): """ Build derived fields (indices & masks) from raw input """ # Shape checking assert len(state.shape) == 2 assert len(candidates.shape) == 3 state = state.to(device) candidates = candidates.to(device) if action is not None: assert len(action.shape) == 2 action = action.to(device) if logged_propensities is not None: assert ( len(logged_propensities.shape) == 2 and logged_propensities.shape[1] == 1 ) logged_propensities = logged_propensities.to(device) batch_size, candidate_num, candidate_dim = candidates.shape if slate_reward is not None: assert len(slate_reward.shape) == 2 and slate_reward.shape[1] == 1 slate_reward = slate_reward.to(device) if position_reward is not None: # pyre-fixme[16]: `Optional` has no attribute `shape`. assert position_reward.shape == action.shape position_reward = position_reward.to(device) src_in_idx = ( torch.arange(candidate_num, device=device).repeat(batch_size, 1) + 2 ) src_src_mask = ( (torch.ones(batch_size, candidate_num, candidate_num)) .type(torch.int8) .to(device) ) def process_tgt_seq(action): if action is not None: _, output_size = action.shape # Account for decoder starting symbol and padding symbol candidates_augment = torch.cat( ( torch.zeros(batch_size, 2, candidate_dim, device=device), candidates, ), dim=1, ) tgt_out_idx = action + 2 tgt_in_idx = torch.full( (batch_size, output_size), DECODER_START_SYMBOL, device=device ) tgt_in_idx[:, 1:] = tgt_out_idx[:, :-1] tgt_out_seq = gather(candidates_augment, tgt_out_idx) tgt_in_seq = torch.zeros( batch_size, output_size, candidate_dim, device=device ) tgt_in_seq[:, 1:] = tgt_out_seq[:, :-1] tgt_tgt_mask = subsequent_mask(output_size, device) else: tgt_in_idx = None tgt_out_idx = None tgt_in_seq = None tgt_out_seq = None tgt_tgt_mask = None return tgt_in_idx, tgt_out_idx, tgt_in_seq, tgt_out_seq, tgt_tgt_mask ( tgt_in_idx, tgt_out_idx, tgt_in_seq, tgt_out_seq, tgt_tgt_mask, ) = process_tgt_seq(action) ( optim_tgt_in_idx, optim_tgt_out_idx, optim_tgt_in_seq, optim_tgt_out_seq, _, ) = process_tgt_seq(optimal_action) return cls.from_tensors( state=state, src_seq=candidates, src_src_mask=src_src_mask, tgt_in_seq=tgt_in_seq, tgt_out_seq=tgt_out_seq, tgt_tgt_mask=tgt_tgt_mask, slate_reward=slate_reward, position_reward=position_reward, src_in_idx=src_in_idx, tgt_in_idx=tgt_in_idx, tgt_out_idx=tgt_out_idx, tgt_out_probs=logged_propensities, optim_tgt_in_idx=optim_tgt_in_idx, optim_tgt_out_idx=optim_tgt_out_idx, optim_tgt_in_seq=optim_tgt_in_seq, optim_tgt_out_seq=optim_tgt_out_seq, extras=extras, ) @classmethod def from_tensors( cls, state: torch.Tensor, src_seq: torch.Tensor, src_src_mask: Optional[torch.Tensor] = None, tgt_in_seq: Optional[torch.Tensor] = None, tgt_out_seq: Optional[torch.Tensor] = None, tgt_tgt_mask: Optional[torch.Tensor] = None, slate_reward: Optional[torch.Tensor] = None, position_reward: Optional[torch.Tensor] = None, src_in_idx: Optional[torch.Tensor] = None, tgt_in_idx: Optional[torch.Tensor] = None, tgt_out_idx: Optional[torch.Tensor] = None, tgt_out_probs: Optional[torch.Tensor] = None, optim_tgt_in_idx: Optional[torch.Tensor] = None, optim_tgt_out_idx: Optional[torch.Tensor] = None, optim_tgt_in_seq: Optional[torch.Tensor] = None, optim_tgt_out_seq: Optional[torch.Tensor] = None, extras: Optional[ExtraData] = None, **kwargs, ): assert isinstance(state, torch.Tensor) assert isinstance(src_seq, torch.Tensor) assert src_src_mask is None or isinstance(src_src_mask, torch.Tensor) assert tgt_in_seq is None or isinstance(tgt_in_seq, torch.Tensor) assert tgt_out_seq is None or isinstance(tgt_out_seq, torch.Tensor) assert tgt_tgt_mask is None or isinstance(tgt_tgt_mask, torch.Tensor) assert slate_reward is None or isinstance(slate_reward, torch.Tensor) assert position_reward is None or isinstance(position_reward, torch.Tensor) assert src_in_idx is None or isinstance(src_in_idx, torch.Tensor) assert tgt_in_idx is None or isinstance(tgt_in_idx, torch.Tensor) assert tgt_out_idx is None or isinstance(tgt_out_idx, torch.Tensor) assert tgt_out_probs is None or isinstance(tgt_out_probs, torch.Tensor) assert optim_tgt_out_idx is None or isinstance(optim_tgt_out_idx, torch.Tensor) assert optim_tgt_out_idx is None or isinstance(optim_tgt_out_idx, torch.Tensor) assert optim_tgt_in_seq is None or isinstance(optim_tgt_in_seq, torch.Tensor) assert optim_tgt_out_seq is None or isinstance(optim_tgt_out_seq, torch.Tensor) assert extras is None or isinstance(extras, ExtraData) return cls( state=FeatureData(float_features=state), src_seq=FeatureData(float_features=src_seq), src_src_mask=src_src_mask, tgt_in_seq=FeatureData(float_features=tgt_in_seq) if tgt_in_seq is not None else None, tgt_out_seq=FeatureData(float_features=tgt_out_seq) if tgt_out_seq is not None else None, tgt_tgt_mask=tgt_tgt_mask, slate_reward=slate_reward, position_reward=position_reward, src_in_idx=src_in_idx, tgt_in_idx=tgt_in_idx, tgt_out_idx=tgt_out_idx, tgt_out_probs=tgt_out_probs, optim_tgt_in_idx=optim_tgt_in_idx, optim_tgt_out_idx=optim_tgt_out_idx, optim_tgt_in_seq=FeatureData(float_features=optim_tgt_in_seq) if optim_tgt_in_seq is not None else None, optim_tgt_out_seq=FeatureData(float_features=optim_tgt_out_seq) if optim_tgt_out_seq is not None else None, extras=extras if extras is not None else None, ) def __post_init__(self): if ( isinstance(self.state, torch.Tensor) or isinstance(self.src_seq, torch.Tensor) or isinstance(self.tgt_in_seq, torch.Tensor) or isinstance(self.tgt_out_seq, torch.Tensor) or isinstance(self.optim_tgt_in_seq, torch.Tensor) or isinstance(self.optim_tgt_out_seq, torch.Tensor) ): raise ValueError( f"Use from_tensors() {type(self.state)} {type(self.src_seq)} " f"{type(self.tgt_in_seq)} {type(self.tgt_out_seq)} " f"{type(self.optim_tgt_in_seq)} {type(self.optim_tgt_out_seq)} " ) @dataclass class BaseInput(TensorDataClass): """ Base class for all inputs, both raw and preprocessed """ state: FeatureData next_state: FeatureData reward: torch.Tensor time_diff: torch.Tensor step: Optional[torch.Tensor] not_terminal: torch.Tensor def __len__(self): return self.state.float_features.size()[0] def batch_size(self): return len(self) def as_dict_shallow(self): return { "state": self.state, "next_state": self.next_state, "reward": self.reward, "time_diff": self.time_diff, "step": self.step, "not_terminal": self.not_terminal, } @staticmethod def from_dict(batch): id_list_features = batch.get(InputColumn.STATE_ID_LIST_FEATURES, None) or {} id_score_list_features = ( batch.get(InputColumn.STATE_ID_SCORE_LIST_FEATURES, None) or {} ) next_id_list_features = ( batch.get(InputColumn.NEXT_STATE_ID_LIST_FEATURES, None) or {} ) next_id_score_list_features = ( batch.get(InputColumn.NEXT_STATE_ID_SCORE_LIST_FEATURES, None) or {} ) # TODO: handle value/mask of DocList filler_mask_val = None doc_list = None candidate_features = batch.get(InputColumn.CANDIDATE_FEATURES, None) if candidate_features is not None: filler_mask_val = torch.zeros( (candidate_features.shape[0], candidate_features.shape[1]) ) doc_list = DocList( float_features=candidate_features, mask=filler_mask_val.clone().bool(), value=filler_mask_val.clone().float(), ) next_doc_list = None next_candidate_features = batch.get(InputColumn.NEXT_CANDIDATE_FEATURES, None) if next_candidate_features is not None: assert filler_mask_val is not None next_doc_list = DocList( float_features=next_candidate_features, mask=filler_mask_val.clone().bool(), value=filler_mask_val.clone().float(), ) return BaseInput( state=FeatureData( float_features=batch[InputColumn.STATE_FEATURES], id_list_features=id_list_features, id_score_list_features=id_score_list_features, candidate_docs=doc_list, ), next_state=FeatureData( float_features=batch[InputColumn.NEXT_STATE_FEATURES], id_list_features=next_id_list_features, id_score_list_features=next_id_score_list_features, candidate_docs=next_doc_list, ), reward=batch[InputColumn.REWARD], time_diff=batch[InputColumn.TIME_DIFF], step=batch.get(InputColumn.STEP, None), not_terminal=batch[InputColumn.NOT_TERMINAL], ) @dataclass class DiscreteDqnInput(BaseInput): """ See input_prototype for DQN expected input shapes """ action: torch.Tensor next_action: torch.Tensor possible_actions_mask: torch.Tensor possible_next_actions_mask: torch.Tensor extras: ExtraData @classmethod def input_prototype(cls, action_dim=2, batch_size=10, state_dim=3): return cls( state=FeatureData(float_features=torch.randn(batch_size, state_dim)), next_state=FeatureData(float_features=torch.randn(batch_size, state_dim)), reward=torch.rand(batch_size, 1), time_diff=torch.ones(batch_size, 1), step=torch.ones(batch_size, 1), not_terminal=torch.ones(batch_size, 1), action=F.one_hot( torch.randint(high=action_dim, size=(batch_size,)), num_classes=action_dim, ), next_action=F.one_hot( torch.randint(high=action_dim, size=(batch_size,)), num_classes=action_dim, ), possible_actions_mask=torch.ones(batch_size, action_dim), possible_next_actions_mask=torch.ones(batch_size, action_dim), extras=ExtraData(action_probability=torch.ones(batch_size, 1)), ) @classmethod def from_dict(cls, batch): base = super().from_dict(batch) return cls( action=batch[InputColumn.ACTION], next_action=batch[InputColumn.NEXT_ACTION], possible_actions_mask=batch[InputColumn.POSSIBLE_ACTIONS_MASK], possible_next_actions_mask=batch[InputColumn.POSSIBLE_NEXT_ACTIONS_MASK], extras=ExtraData.from_dict(batch), **base.as_dict_shallow(), ) @dataclass class SlateQInput(BaseInput): """ The shapes of `reward`, `reward_mask`, & `next_item_mask` are `(batch_size, slate_size)`. `reward_mask` indicated whether the reward could be observed, e.g., the item got into viewport or not. """ action: torch.Tensor next_action: torch.Tensor reward_mask: torch.Tensor extras: Optional[ExtraData] = None @classmethod def from_dict(cls, d): action = d["action"] next_action = d["next_action"] return cls( state=FeatureData( float_features=d["state_features"], candidate_docs=DocList( float_features=d["candidate_features"], mask=d["item_mask"], value=d["item_probability"], ), ), next_state=FeatureData( float_features=d["next_state_features"], candidate_docs=DocList( float_features=d["next_candidate_features"], mask=d["next_item_mask"], value=d["next_item_probability"], ), ), action=action, next_action=next_action, reward=d["position_reward"], reward_mask=d["reward_mask"], time_diff=d["time_diff"], not_terminal=d["not_terminal"], step=None, extras=ExtraData.from_dict(d), ) @dataclass class ParametricDqnInput(BaseInput): action: FeatureData next_action: FeatureData possible_actions: FeatureData possible_actions_mask: torch.Tensor possible_next_actions: FeatureData possible_next_actions_mask: torch.Tensor extras: Optional[ExtraData] = None @classmethod def from_dict(cls, batch): return cls( state=FeatureData(float_features=batch["state_features"]), action=FeatureData(float_features=batch["action"]), next_state=FeatureData(float_features=batch["next_state_features"]), next_action=FeatureData(float_features=batch["next_action"]), possible_actions=FeatureData(float_features=batch["possible_actions"]), possible_actions_mask=batch["possible_actions_mask"], possible_next_actions=FeatureData( float_features=batch["possible_next_actions"] ), possible_next_actions_mask=batch["possible_next_actions_mask"], reward=batch["reward"], not_terminal=batch["not_terminal"], time_diff=batch["time_diff"], step=batch["step"], extras=batch["extras"], ) @dataclass class PolicyNetworkInput(BaseInput): action: FeatureData next_action: FeatureData extras: Optional[ExtraData] = None @classmethod def from_dict(cls, batch): base = super().from_dict(batch) # TODO: Implement ExtraData.from_dict extras = batch.get("extras", None) return cls( action=FeatureData(float_features=batch["action"]), next_action=FeatureData(float_features=batch["next_action"]), extras=extras, **base.as_dict_shallow(), ) @dataclass class PolicyGradientInput(TensorDataClass): """ See input_prototype for expected input dimensions """ state: FeatureData action: torch.Tensor reward: torch.Tensor log_prob: torch.Tensor possible_actions_mask: Optional[torch.Tensor] = None @classmethod def input_prototype(cls, action_dim=2, batch_size=10, state_dim=3): return cls( state=FeatureData(float_features=torch.randn(batch_size, state_dim)), action=F.one_hot( torch.randint(high=action_dim, size=(batch_size,)), num_classes=action_dim, ), reward=torch.rand(batch_size), log_prob=torch.log(torch.rand(batch_size)), possible_actions_mask=torch.ones(batch_size, action_dim), ) @classmethod def from_dict(cls, d: Dict[str, torch.Tensor]): # TODO: rename "observation" to "state" in Transition and return cls(**d) return cls( state=FeatureData(float_features=d["observation"]), action=d["action"], reward=d["reward"], log_prob=d["log_prob"], possible_actions_mask=d.get("possible_actions_mask", None), ) def __len__(self): return len(self.action) @dataclass class BanditRewardModelInput(TensorDataClass): state: FeatureData action: torch.Tensor reward: torch.Tensor action_prob: Optional[torch.Tensor] = None @classmethod def from_dict(cls, batch: Dict[str, torch.Tensor]): return cls( state=FeatureData(float_features=batch["state_features"]), action=batch["action"], reward=batch["reward"], action_prob=batch.get("action_probability", None), ) @dataclass class MemoryNetworkInput(BaseInput): action: torch.Tensor valid_step: Optional[torch.Tensor] = None extras: ExtraData = field(default_factory=ExtraData) @classmethod def from_dict(cls, d): return cls( state=FeatureData( float_features=d["state"], ), next_state=FeatureData( float_features=d["next_state"], ), action=d["action"], reward=d["reward"], time_diff=d["time_diff"], not_terminal=d["not_terminal"], step=d["step"], extras=ExtraData.from_dict(d), ) def __len__(self): if len(self.state.float_features.size()) == 2: return self.state.float_features.size()[0] elif len(self.state.float_features.size()) == 3: return self.state.float_features.size()[1] else: raise NotImplementedError() @dataclass class PreprocessedTrainingBatch(TensorDataClass): training_input: Union[PreprocessedRankingInput] # TODO: deplicate this and move into individual ones. extras: ExtraData = field(default_factory=ExtraData) def batch_size(self): return self.training_input.state.float_features.size()[0] @dataclass class SlateScoreBatch: mdp_id: torch.Tensor sequence_number: torch.Tensor scores: torch.Tensor training_input: PolicyGradientInput @dataclass class MemoryNetworkOutput(TensorDataClass): mus: torch.Tensor sigmas: torch.Tensor logpi: torch.Tensor reward: torch.Tensor not_terminal: torch.Tensor last_step_lstm_hidden: torch.Tensor last_step_lstm_cell: torch.Tensor all_steps_lstm_hidden: torch.Tensor @dataclass class Seq2RewardOutput(TensorDataClass): acc_reward: torch.Tensor @dataclass class DqnPolicyActionSet(TensorDataClass): greedy: int softmax: Optional[int] = None greedy_act_name: Optional[str] = None softmax_act_name: Optional[str] = None softmax_act_prob: Optional[float] = None @dataclass class PlanningPolicyOutput(TensorDataClass): # best action to take next next_best_continuous_action: Optional[torch.Tensor] = None next_best_discrete_action_one_hot: Optional[torch.Tensor] = None next_best_discrete_action_idx: Optional[int] = None @dataclass class RankingOutput(TensorDataClass): # a tensor of integer indices w.r.t. to possible candidates # the values are offset by 2 to account for padding and decoder-starter symbol # shape: batch_size, tgt_seq_len # e.g., there are candidates C0, C1, C2, C3, C4, and the ranked order is # C4, C1, C2, C3, C0. Then the ranked_tgt_out_idx = [6, 3, 4, 5, 2] ranked_tgt_out_idx: Optional[torch.Tensor] = None # generative probability of ranked tgt sequences at each decoding step # shape: batch_size, tgt_seq_len, candidate_size ranked_per_symbol_probs: Optional[torch.Tensor] = None # generative probability of ranked tgt sequences # shape: batch_size, 1 ranked_per_seq_probs: Optional[torch.Tensor] = None # log probabilities of given tgt sequences are used in REINFORCE # shape: batch_size, 1 if Seq2SlateMode == PER_SEQ_LOG_PROB_MODE # shape: batch_size, tgt_seq_len if Seq2SlateMode == PER_SYMBOL_LOG_PROB_DIST_MODE log_probs: Optional[torch.Tensor] = None # encoder scores in tgt_out_idx order encoder_scores: Optional[torch.Tensor] = None @dataclass class RewardNetworkOutput(TensorDataClass): predicted_reward: torch.Tensor @dataclass class FrechetSortConfig: shape: float equiv_len: int topk: Optional[int] = None log_scores: bool = True
37,736
34.136872
91
py
ReAgent
ReAgent-master/reagent/core/torch_utils.py
#!/usr/bin/env python3 # Copyright (c) Facebook, Inc. and its affiliates. All rights reserved. from io import BytesIO from typing import Dict import numpy as np import torch def dict_to_tensor(batch: Dict[str, np.ndarray], device: str = "cpu"): return {k: torch.tensor(v).to(device) for k, v in batch.items()} def rescale_torch_tensor( tensor: torch.Tensor, new_min: torch.Tensor, new_max: torch.Tensor, prev_min: torch.Tensor, prev_max: torch.Tensor, ): """ Rescale column values in N X M torch tensor to be in new range. Each column m in input tensor will be rescaled from range [prev_min[m], prev_max[m]] to [new_min[m], new_max[m]] """ assert tensor.shape[1] == new_min.shape[1] == new_max.shape[1] assert tensor.shape[1] == prev_min.shape[1] == prev_max.shape[1] prev_range = prev_max - prev_min new_range = new_max - new_min return ((tensor - prev_min) / prev_range) * new_range + new_min def stack(mems): """ Stack a list of tensors Could use torch.stack here but torch.stack is much slower than torch.cat + view Submitted an issue for investigation: https://github.com/pytorch/pytorch/issues/22462 FIXME: Remove this function after the issue above is resolved """ shape = (-1, *mems[0].shape) return torch.cat(mems).view(*shape) def export_module_to_buffer(module) -> BytesIO: # traced_script_module = torch.jit.trace(module, module.input_prototype()) write_buffer = BytesIO() torch.jit.save(module, write_buffer) return write_buffer def softmax(x, temperature): """Compute softmax values for each sets of scores in x.""" x = x / temperature return torch.nn.functional.softmax(x, dim=1) def masked_softmax(x, mask, temperature): """Compute softmax values for each sets of scores in x.""" x = x / temperature mask_min_x = x - ((1.0 - mask) * 1e20) mask_min_x -= torch.max(mask_min_x, dim=1, keepdim=True)[0] e_x = torch.exp(mask_min_x) e_x *= mask out = e_x / e_x.sum(dim=1, keepdim=True) # Set NaN values to 0 (NaN happens when a full mask row is passed in) out[out != out] = 0 return out def gather(data, index_2d): """ Gather data alongs the second dim. Assume data is 3d with shape (batch_size, dim1, dim2), and index_2d's shape is (batch_size, dim1). output[i][j] = data[i][index_2d[i][j]] This function does not require data, output, or index_2d having the same shape, which is mandated by torch.gather. """ batch_size = data.shape[0] data_dim = data.shape[2] index_len = index_2d.shape[1] device = data.device res = data[ torch.arange(batch_size, device=device).repeat_interleave( # index_len has to be moved to the device explicitly, otherwise # error will throw during jit.trace torch.tensor([index_len], device=device) ), index_2d.flatten(), ].view(batch_size, index_len, data_dim) return res def get_device(model): return next(model.parameters()).device
3,085
29.86
93
py
ReAgent
ReAgent-master/reagent/core/parameters.py
#!/usr/bin/env python3 # Copyright (c) Facebook, Inc. and its affiliates. All rights reserved. import enum from typing import Dict, List, Optional from reagent.core.base_dataclass import BaseDataClass from reagent.core.configuration import param_hash from reagent.core.dataclasses import dataclass, field from reagent.core.parameters_seq2slate import ( IPSClamp, LearningMethod, SimulationParameters, ) # For TD3 and SAC: actions are normalized in this range for training and # rescaled back to action_space.low/high at serving time. CONTINUOUS_TRAINING_ACTION_RANGE = (-1.0, 1.0) class ProblemDomain(enum.Enum): CONTINUOUS_ACTION = "continuous_action" DISCRETE_ACTION = "discrete_action" PARAMETRIC_ACTION = "parametric_action" # I don't think the data generated for these 2 types are generic SEQ_TO_REWARD = "seq2reward" MDN_RNN = "mdn_rnn" class SlateOptMethod(enum.Enum): GREEDY = "greedy" TOP_K = "top_k" EXACT = "exact" @dataclass(frozen=True) class SlateOptParameters(BaseDataClass): __hash__ = param_hash method: SlateOptMethod = SlateOptMethod.TOP_K @dataclass(frozen=True) class RLParameters(BaseDataClass): __hash__ = param_hash gamma: float = 0.9 epsilon: float = 0.1 target_update_rate: float = 0.001 maxq_learning: bool = True reward_boost: Optional[Dict[str, float]] = None temperature: float = 0.01 softmax_policy: bool = False use_seq_num_diff_as_time_diff: bool = False q_network_loss: str = "mse" set_missing_value_to_zero: bool = False tensorboard_logging_freq: int = 0 predictor_atol_check: float = 0.0 predictor_rtol_check: float = 5e-5 time_diff_unit_length: float = 1.0 multi_steps: Optional[int] = None # for pytorch discrete model, specify the max number of prediction change # allowed during conversions between model frameworks in ratio ratio_different_predictions_tolerance: float = 0 @dataclass(frozen=True) class MDNRNNTrainerParameters(BaseDataClass): __hash__ = param_hash hidden_size: int = 64 num_hidden_layers: int = 2 learning_rate: float = 0.001 num_gaussians: int = 5 # weight in calculating world-model loss reward_loss_weight: float = 1.0 next_state_loss_weight: float = 1.0 not_terminal_loss_weight: float = 1.0 fit_only_one_next_step: bool = False action_dim: int = 2 action_names: Optional[List[str]] = None multi_steps: int = 1 @dataclass(frozen=True) class Seq2RewardTrainerParameters(BaseDataClass): __hash__ = param_hash learning_rate: float = 0.001 multi_steps: int = 1 action_names: List[str] = field(default_factory=lambda: []) compress_model_learning_rate: float = 0.001 gamma: float = 1.0 view_q_value: bool = False step_predict_net_size: int = 64 reward_boost: Optional[Dict[str, float]] = None @dataclass(frozen=True) class CEMTrainerParameters(BaseDataClass): __hash__ = param_hash plan_horizon_length: int = 0 num_world_models: int = 0 cem_population_size: int = 0 cem_num_iterations: int = 0 ensemble_population_size: int = 0 num_elites: int = 0 mdnrnn: MDNRNNTrainerParameters = MDNRNNTrainerParameters() rl: RLParameters = RLParameters() alpha: float = 0.25 epsilon: float = 0.001 @dataclass(frozen=True) class EvaluationParameters(BaseDataClass): calc_cpe_in_training: bool = True @dataclass(frozen=True) class EvolutionParameters(BaseDataClass): population_size: int = 1000 mutation_power: float = 0.1 learning_rate: float = 0.01 @dataclass(frozen=True) class StateFeatureParameters(BaseDataClass): __hash__ = param_hash state_feature_names_override: List[str] = field(default_factory=lambda: []) state_feature_hashes_override: List[int] = field(default_factory=lambda: []) @dataclass(frozen=True) class NormalizationParameters(BaseDataClass): __hash__ = param_hash feature_type: str boxcox_lambda: Optional[float] = None boxcox_shift: Optional[float] = None mean: Optional[float] = None stddev: Optional[float] = None possible_values: Optional[List[int]] = None # Assume present for ENUM type quantiles: Optional[ List[float] ] = None # Assume present for QUANTILE type and sorted min_value: Optional[float] = None max_value: Optional[float] = None class NormalizationKey(object): """Keys for dictionaries of NormalizationData""" STATE = "state" ACTION = "action" ITEM = "item" CANDIDATE = "candidate" @dataclass(frozen=True) class NormalizationData(BaseDataClass): __hash__ = param_hash dense_normalization_parameters: Dict[int, NormalizationParameters] @dataclass(frozen=True) class ConvNetParameters(BaseDataClass): conv_dims: List[int] conv_height_kernels: List[int] pool_types: List[str] pool_kernel_sizes: List[int] conv_width_kernels: Optional[List[int]] = None ################################################# # RL Ranking parameters # ################################################# @dataclass(frozen=True) class TransformerParameters(BaseDataClass): num_heads: int = 1 dim_model: int = 64 dim_feedforward: int = 32 num_stacked_layers: int = 2 state_embed_dim: Optional[int] = None @dataclass(frozen=True) class GRUParameters(BaseDataClass): dim_model: int num_stacked_layers: int @dataclass(frozen=True) class BaselineParameters(BaseDataClass): dim_feedforward: int num_stacked_layers: int warmup_num_batches: int = 0 @dataclass(frozen=True) class Seq2SlateParameters(BaseDataClass): on_policy: bool = True learning_method: LearningMethod = LearningMethod.REINFORCEMENT_LEARNING ips_clamp: Optional[IPSClamp] = None simulation: Optional[SimulationParameters] = None @dataclass(frozen=True) class RankingParameters(BaseDataClass): max_src_seq_len: int = 0 max_tgt_seq_len: int = 0 greedy_serving: bool = False
6,022
26.884259
80
py
ReAgent
ReAgent-master/reagent/core/oss_tensorboard_logger.py
from typing import Optional, Union, Dict, List, Tuple import torch from pytorch_lightning.loggers import TensorBoardLogger from pytorch_lightning.utilities import rank_zero_only class LocalCacheLogger: @staticmethod def store_metrics( tb_logger, metrics: Dict[ str, Union[float, torch.Tensor, Dict[str, Union[float, torch.Tensor]]] ], step: Optional[int] = None, ): for plot_name, plot_value_or_dict in metrics.items(): if isinstance(plot_value_or_dict, dict): if plot_name not in tb_logger.line_plot_buffer: tb_logger.line_plot_buffer[plot_name] = {} for line_name, plot_value in plot_value_or_dict.items(): LocalCacheLogger._add_point( tb_logger, plot_name, line_name, plot_value, step ) else: LocalCacheLogger._add_point( tb_logger, plot_name, "", plot_value_or_dict, step ) @staticmethod def _add_point( tb_logger, plot_name: str, line_name: str, plot_value: Union[float, torch.Tensor], step: Optional[int], ): """Adds a point to a multi-line plot given the plot name, the line name, and optionally the step (x coordinate).""" if isinstance(plot_value, torch.Tensor): plot_value = plot_value.item() if step is None: if ( plot_name in tb_logger.line_plot_buffer and line_name in tb_logger.line_plot_buffer[plot_name] ): x = tb_logger.line_plot_buffer[plot_name][line_name][-1][0] + 1.0 else: x = 0.0 else: x = float(step) LocalCacheLogger._create_plots_and_append( tb_logger.line_plot_buffer, plot_name, line_name, x, plot_value ) if len(tb_logger.line_plot_buffer[plot_name][line_name]) >= 50: mean = float( torch.mean( torch.FloatTensor( [ float(p[1]) for p in tb_logger.line_plot_buffer[plot_name][line_name] ] ) ).item() ) LocalCacheLogger._create_plots_and_append( tb_logger.line_plot_aggregated, plot_name, line_name, x, mean ) tb_logger.line_plot_buffer[plot_name][line_name].clear() @staticmethod def _create_plots_and_append( plot_store: Dict[str, Dict[str, List[Tuple[float, float]]]], plot_name: str, line_name: str, x: int, y: float, ): if plot_name in plot_store and line_name in plot_store[plot_name]: plot_store[plot_name][line_name].append((x, y)) elif plot_name in plot_store: plot_store[plot_name][line_name] = [(x, y)] else: plot_store[plot_name] = {line_name: [(x, y)]} class OssTensorboardLogger(TensorBoardLogger): """Wrapper around ManifoldTensorBoardLogger that collects the plot data in memory and can flush to create fblearner plot objects.""" def __init__( self, save_dir: str, name: Optional[str] = "default", version: Optional[Union[int, str]] = None, log_graph: bool = False, default_hp_metric: bool = True, prefix: str = "", **kwargs ): super().__init__( save_dir, name, version, log_graph, default_hp_metric, prefix, **kwargs, ) self.line_plot_aggregated: Dict[str, Dict[str, List[Tuple[float, float]]]] = {} self.line_plot_buffer: Dict[str, Dict[str, List[Tuple[float, float]]]] = {} @rank_zero_only def log_metrics( self, metrics: Dict[ str, Union[float, torch.Tensor, Dict[str, Union[float, torch.Tensor]]] ], step: Optional[int] = None, ) -> None: """Log a set of metrics. A metric is either a scalar or a set of scalars that will be plotted together""" super().log_metrics(metrics, step) LocalCacheLogger.store_metrics(self, metrics, step) def clear_local_data(self): # We don't call clear here because it's a lot of data and someone else probably owns it self.line_plot_aggregated = {} self.line_plot_buffer = {}
4,531
33.861538
136
py
ReAgent
ReAgent-master/reagent/core/tensorboardX.py
#!/usr/bin/env python3 # Copyright (c) Facebook, Inc. and its affiliates. All rights reserved. """ Context library to allow dropping tensorboardX anywhere in the codebase. If there is no SummaryWriter in the context, function calls will be no-op. Usage: writer = SummaryWriter() with summary_writer_context(writer): some_func() def some_func(): SummaryWriterContext.add_scalar("foo", tensor) """ import contextlib import logging from typing import Any, Dict, List from torch.utils.tensorboard import SummaryWriter logger = logging.getLogger(__name__) class SummaryWriterContextMeta(type): def __getattr__(cls, func): if func.startswith("__"): return super().__getattr__(func) if not cls._writer_stacks: def noop(*args, **kwargs): return return noop writer = cls._writer_stacks[-1] def call(*args, **kwargs): if "global_step" not in kwargs: kwargs["global_step"] = cls._global_step try: return getattr(writer, func)(*args, **kwargs) except Exception as e: if hasattr(writer, "exceptions_to_ignore") and isinstance( e, writer.exceptions_to_ignore ): logger.warning("Ignoring exception: {}".format(e)) if hasattr(writer, "exception_logging_func"): writer.exception_logging_func(e) return raise return call class SummaryWriterContext(metaclass=SummaryWriterContextMeta): _writer_stacks: List[SummaryWriter] = [] _global_step = 0 _custom_scalars: Dict[str, Any] = {} @classmethod def _reset_globals(cls): cls._global_step = 0 cls._custom_scalars = {} @classmethod def increase_global_step(cls): cls._global_step += 1 @classmethod def add_histogram(cls, key, val, *args, **kwargs): try: return cls.__getattr__("add_histogram")(key, val, *args, **kwargs) except ValueError: logger.warning(f"Cannot create histogram for {key}, got values: {val}") @classmethod def add_custom_scalars(cls, writer): """ Call this once you are satisfied setting up custom scalar """ writer.add_custom_scalars(cls._custom_scalars) @classmethod def add_custom_scalars_multilinechart(cls, tags, category=None, title=None): assert category and title, "category & title must be set" if category not in cls._custom_scalars: cls._custom_scalars[category] = {} assert ( title not in cls._custom_scalars[category] ), "Title ({}) is already in category ({})".format(title, category) cls._custom_scalars[category][title] = ["Multiline", tags] @classmethod def push(cls, writer): assert isinstance( writer, SummaryWriter ), "writer is not a SummaryWriter: {}".format(writer) cls._writer_stacks.append(writer) @classmethod def pop(cls): return cls._writer_stacks.pop() @contextlib.contextmanager def summary_writer_context(writer): if writer is not None: SummaryWriterContext.push(writer) try: yield finally: if writer is not None: SummaryWriterContext.pop()
3,412
26.97541
83
py
ReAgent
ReAgent-master/reagent/training/cem_trainer.py
#!/usr/bin/env python3 # Copyright (c) Facebook, Inc. and its affiliates. All rights reserved. """ The Trainer for Cross-Entropy Method. The idea is that an ensemble of world models are fitted to predict transitions and reward functions. A cross entropy method-based planner will then plan the best next action based on simulation data generated by the fitted world models. The idea is inspired by: https://arxiv.org/abs/1805.12114 """ import logging from typing import List import reagent.core.types as rlt import torch.nn as nn from reagent.core.parameters import CEMTrainerParameters from reagent.models.cem_planner import CEMPlannerNetwork from reagent.training.reagent_lightning_module import ReAgentLightningModule from reagent.training.world_model.mdnrnn_trainer import MDNRNNTrainer logger = logging.getLogger(__name__) def print_mdnrnn_losses(minibatch, model_index, losses) -> None: logger.info( f"{minibatch}-th minibatch {model_index}-th model: \n" f'loss={losses["loss"]}, bce={losses["bce"]}, ' f'gmm={losses["gmm"]}, mse={losses["mse"]}\n' ) class CEMTrainer(ReAgentLightningModule): def __init__( self, cem_planner_network: CEMPlannerNetwork, world_model_trainers: List[MDNRNNTrainer], parameters: CEMTrainerParameters, ) -> None: super().__init__() self.cem_planner_network = cem_planner_network self.world_model_trainers = nn.ModuleList(world_model_trainers) def configure_optimizers(self): return [o for t in self.world_model_trainers for o in t.configure_optimizers()] def train_step_gen(self, training_batch: rlt.MemoryNetworkInput, batch_idx: int): for t in self.world_model_trainers: yield from t.train_step_gen(training_batch, batch_idx)
1,805
35.12
87
py
ReAgent
ReAgent-master/reagent/training/qrdqn_trainer.py
#!/usr/bin/env python3 # Copyright (c) Facebook, Inc. and its affiliates. All rights reserved. import logging from typing import List, Tuple import reagent.core.types as rlt import torch from reagent.core.configuration import resolve_defaults from reagent.core.dataclasses import field from reagent.core.parameters import EvaluationParameters, RLParameters from reagent.optimizer import SoftUpdate from reagent.optimizer.union import Optimizer__Union from reagent.training.dqn_trainer_base import DQNTrainerBaseLightning logger = logging.getLogger(__name__) class QRDQNTrainer(DQNTrainerBaseLightning): """ Implementation of QR-DQN (Quantile Regression Deep Q-Network) See https://arxiv.org/abs/1710.10044 for details """ @resolve_defaults def __init__( self, q_network, q_network_target, metrics_to_score=None, reward_network=None, q_network_cpe=None, q_network_cpe_target=None, actions: List[str] = field(default_factory=list), # noqa: B008 rl: RLParameters = field(default_factory=RLParameters), # noqa: B008 double_q_learning: bool = True, num_atoms: int = 51, minibatch_size: int = 1024, minibatches_per_step: int = 1, optimizer: Optimizer__Union = field( # noqa: B008 default_factory=Optimizer__Union.default ), cpe_optimizer: Optimizer__Union = field( # noqa: B008 default_factory=Optimizer__Union.default ), evaluation: EvaluationParameters = field( # noqa: B008 default_factory=EvaluationParameters ), ) -> None: super().__init__( rl_parameters=rl, metrics_to_score=metrics_to_score, actions=actions, evaluation_parameters=evaluation, ) # TODO: check to ensure no rl parameter value is set that isn't actively used by class self.double_q_learning = double_q_learning self.minibatch_size = minibatch_size self.minibatches_per_step = minibatches_per_step self._actions = actions self.q_network = q_network self.q_network_target = q_network_target self.q_network_optimizer = optimizer self.num_atoms = num_atoms self.register_buffer("quantiles", None) self.quantiles = ( (0.5 + torch.arange(self.num_atoms).float()) / float(self.num_atoms) ).view(1, -1) self._initialize_cpe( reward_network, q_network_cpe, q_network_cpe_target, optimizer=cpe_optimizer ) self.register_buffer("reward_boosts", None) self.reward_boosts = torch.zeros([1, len(self._actions)]) if rl.reward_boost is not None: # pyre-fixme[16]: Optional type has no attribute `keys`. for k in rl.reward_boost.keys(): i = self._actions.index(k) # pyre-fixme[16]: Optional type has no attribute `__getitem__`. self.reward_boosts[0, i] = rl.reward_boost[k] def configure_optimizers(self): optimizers = [] target_params = list(self.q_network_target.parameters()) source_params = list(self.q_network.parameters()) optimizers.append( self.q_network_optimizer.make_optimizer_scheduler( self.q_network.parameters() ) ) if self.calc_cpe_in_training: ( cpe_target_params, cpe_source_params, cpe_optimizers, ) = self._configure_cpe_optimizers() target_params += cpe_target_params source_params += cpe_source_params optimizers += cpe_optimizers optimizers.append( SoftUpdate.make_optimizer_scheduler( target_params, source_params, tau=self.tau ) ) return optimizers def train_step_gen(self, training_batch: rlt.DiscreteDqnInput, batch_idx: int): self._check_input(training_batch) rewards = self.boost_rewards(training_batch.reward, training_batch.action) discount_tensor = torch.full_like(rewards, self.gamma) possible_next_actions_mask = training_batch.possible_next_actions_mask.float() possible_actions_mask = training_batch.possible_actions_mask.float() not_done_mask = training_batch.not_terminal.float() if self.use_seq_num_diff_as_time_diff: assert self.multi_steps is None discount_tensor = torch.pow(self.gamma, training_batch.time_diff.float()) if self.multi_steps is not None: assert training_batch.step is not None discount_tensor = torch.pow(self.gamma, training_batch.step.float()) next_qf = self.q_network_target(training_batch.next_state) if self.maxq_learning: # Select distribution corresponding to max valued action next_q_values = ( self.q_network(training_batch.next_state) if self.double_q_learning else next_qf ).mean(dim=2) next_action = self.argmax_with_mask( next_q_values, possible_next_actions_mask ) next_qf = next_qf[range(rewards.shape[0]), next_action.reshape(-1)] else: next_qf = (next_qf * training_batch.next_action.unsqueeze(-1)).sum(1) # Build target distribution target_Q = rewards + discount_tensor * not_done_mask * next_qf current_qf = self.q_network(training_batch.state) # for reporting only all_q_values = current_qf.mean(2).detach() current_qf = (current_qf * training_batch.action.unsqueeze(-1)).sum(1) # (batch, atoms) -> (atoms, batch, 1) -> (atoms, batch, atoms) td = target_Q.t().unsqueeze(-1) - current_qf loss = ( self.huber(td) * (self.quantiles - (td.detach() < 0).float()).abs() ).mean() yield loss # pyre-fixme[16]: `DQNTrainer` has no attribute `loss`. self.loss = loss.detach() # Get Q-values of next states, used in computing cpe all_next_action_scores = ( self.q_network(training_batch.next_state).detach().mean(dim=2) ) logged_action_idxs = torch.argmax(training_batch.action, dim=1, keepdim=True) yield from self._calculate_cpes( training_batch, training_batch.state, training_batch.next_state, all_q_values, all_next_action_scores, logged_action_idxs, discount_tensor, not_done_mask, ) model_action_idxs = self.argmax_with_mask( all_q_values, possible_actions_mask if self.maxq_learning else training_batch.action, ) self.reporter.log( td_loss=loss, logged_actions=logged_action_idxs, logged_propensities=training_batch.extras.action_probability, logged_rewards=rewards, logged_values=None, # Compute at end of each epoch for CPE model_values=all_q_values, model_values_on_logged_actions=None, # Compute at end of each epoch for CPE model_action_idxs=model_action_idxs, ) yield self.soft_update_result() # pyre-fixme[56]: Decorator `torch.no_grad(...)` could not be called, because # its type `no_grad` is not callable. @torch.no_grad() def boost_rewards( self, rewards: torch.Tensor, actions: torch.Tensor ) -> torch.Tensor: # Apply reward boost if specified reward_boosts = torch.sum( actions.float() * self.reward_boosts, dim=1, keepdim=True ) return rewards + reward_boosts def argmax_with_mask(self, q_values, possible_actions_mask): # Set q-values of impossible actions to a very large negative number. q_values = q_values.reshape(possible_actions_mask.shape) q_values = q_values + self.ACTION_NOT_POSSIBLE_VAL * (1 - possible_actions_mask) return q_values.argmax(1) # Used to prevent warning when a.shape != b.shape def huber(self, x): return torch.where(x.abs() < 1, 0.5 * x.pow(2), x.abs() - 0.5) # pyre-fixme[56]: Decorator `torch.no_grad(...)` could not be called, because # its type `no_grad` is not callable. @torch.no_grad() def get_detached_model_outputs( self, state: rlt.FeatureData ) -> Tuple[torch.Tensor, torch.Tensor]: """Gets the q values from the model and target networks""" q_values = self.q_network(state).mean(dim=2) q_values_target = self.q_network_target(state).mean(dim=2) return q_values, q_values_target
8,782
36.216102
94
py
ReAgent
ReAgent-master/reagent/training/reinforce_trainer.py
#!/usr/bin/env python3 # Copyright (c) Facebook, Inc. and its affiliates. All rights reserved. import inspect import logging import math from dataclasses import field from typing import List, Optional import reagent.core.types as rlt import torch import torch.optim from reagent.gym.policies.policy import Policy from reagent.models.base import ModelBase from reagent.optimizer.union import Optimizer__Union from reagent.training.reagent_lightning_module import ReAgentLightningModule from reagent.training.utils import discounted_returns, whiten logger = logging.getLogger(__name__) class ReinforceTrainer(ReAgentLightningModule): def __init__( self, policy: Policy, gamma: float = 0.0, optimizer: Optimizer__Union = field( # noqa: B008 default_factory=Optimizer__Union.default ), optimizer_value_net: Optimizer__Union = field( # noqa: B008 default_factory=Optimizer__Union.default ), actions: List[str] = field(default_factory=list), # noqa: B008 off_policy: bool = False, reward_clip: float = 1e6, clip_param: float = 1e6, normalize: bool = True, subtract_mean: bool = True, offset_clamp_min: bool = False, value_net: Optional[ModelBase] = None, ): super().__init__() self._actions = actions self.scorer = policy.scorer self.sampler = policy.sampler self.gamma = gamma self.off_policy = off_policy self.reward_clip = reward_clip self.clip_param = clip_param self.normalize = normalize self.subtract_mean = subtract_mean self.offset_clamp_min = offset_clamp_min self.optimizer = optimizer self.optimizer_value_net = optimizer_value_net if value_net is not None: self.value_net = value_net self.value_loss_fn = torch.nn.MSELoss(reduction="mean") else: self.value_net = None def _check_input(self, training_batch: rlt.PolicyGradientInput): assert training_batch.reward.ndim == 1 if self.off_policy: assert training_batch.log_prob.ndim == 1 def configure_optimizers(self): optimizers = [] # value net optimizer if self.value_net is not None: optimizers.append( self.optimizer_value_net.make_optimizer_scheduler( self.value_net.parameters() ) ) # policy optimizer optimizers.append( self.optimizer.make_optimizer_scheduler(self.scorer.parameters()) ) return optimizers def train_step_gen(self, training_batch: rlt.PolicyGradientInput, batch_idx: int): self._check_input(training_batch) actions = training_batch.action rewards = training_batch.reward.detach() scorer_inputs = [] if inspect.getattr_static(training_batch, "graph", None) is not None: # GNN scorer_inputs.append(training_batch.graph) else: scorer_inputs.append(training_batch.state) if training_batch.possible_actions_mask is not None: scorer_inputs.append(training_batch.possible_actions_mask) scores = self.scorer(*scorer_inputs) characteristic_eligibility = self.sampler.log_prob(scores, actions).float() offset_reinforcement = discounted_returns( torch.clamp(rewards, max=self.reward_clip).clone(), self.gamma ) if self.normalize: offset_reinforcement = whiten( offset_reinforcement, subtract_mean=self.subtract_mean ) if self.offset_clamp_min: offset_reinforcement = offset_reinforcement.clamp(min=0) if self.value_net is not None: if self.normalize: raise RuntimeError( "Can't apply a baseline and normalize rewards simultaneously" ) baselines = self.value_net(training_batch.state).squeeze() yield self.value_loss_fn(baselines, offset_reinforcement) # subtract learned value function baselines from rewards offset_reinforcement = offset_reinforcement - baselines if self.off_policy: characteristic_eligibility = torch.exp( torch.clamp( characteristic_eligibility - training_batch.log_prob, max=math.log(float(self.clip_param)), ) ).float() yield -(offset_reinforcement.float()) @ characteristic_eligibility # PG "loss"
4,640
37.040984
87
py
ReAgent
ReAgent-master/reagent/training/parametric_dqn_trainer.py
#!/usr/bin/env python3 # Copyright (c) Facebook, Inc. and its affiliates. All rights reserved. import logging from typing import Tuple import reagent.core.parameters as rlp import reagent.core.types as rlt import torch import torch.nn.functional as F from reagent.core.configuration import resolve_defaults from reagent.core.dataclasses import field from reagent.optimizer import Optimizer__Union, SoftUpdate from reagent.training.dqn_trainer_base import DQNTrainerMixin from reagent.training.reagent_lightning_module import ReAgentLightningModule from reagent.training.rl_trainer_pytorch import RLTrainerMixin logger = logging.getLogger(__name__) class ParametricDQNTrainer(DQNTrainerMixin, RLTrainerMixin, ReAgentLightningModule): @resolve_defaults def __init__( self, q_network, q_network_target, reward_network, # Start ParametricDQNTrainerParameters rl: rlp.RLParameters = field(default_factory=rlp.RLParameters), # noqa: B008 double_q_learning: bool = True, minibatches_per_step: int = 1, optimizer: Optimizer__Union = field( # noqa: B008 default_factory=Optimizer__Union.default ), ) -> None: super().__init__() self.rl_parameters = rl self.double_q_learning = double_q_learning self.minibatches_per_step = minibatches_per_step or 1 self.q_network = q_network self.q_network_target = q_network_target self.reward_network = reward_network self.optimizer = optimizer if rl.q_network_loss == "mse": self.q_network_loss = F.mse_loss elif rl.q_network_loss == "huber": self.q_network_loss = F.smooth_l1_loss elif rl.q_network_loss == "bce_with_logits": # The loss is only used when gamma = 0, reward is between 0 and 1 # and we need to calculate NE as metrics. assert ( rl.gamma == 0 ), "bce_with_logits loss is only supported when gamma is 0." self.q_network_loss = F.binary_cross_entropy_with_logits else: raise Exception( "Q-Network loss type {} not valid loss.".format(rl.q_network_loss) ) def configure_optimizers(self): optimizers = [] optimizers.append( self.optimizer.make_optimizer_scheduler(self.q_network.parameters()) ) optimizers.append( self.optimizer.make_optimizer_scheduler(self.reward_network.parameters()) ) # soft-update target_params = list(self.q_network_target.parameters()) source_params = list(self.q_network.parameters()) optimizers.append( SoftUpdate.make_optimizer_scheduler( target_params, source_params, tau=self.tau ) ) return optimizers # pyre-fixme[56]: Decorator `torch.no_grad(...)` could not be called, because # its type `no_grad` is not callable. @torch.no_grad() def get_detached_model_outputs( self, state, action ) -> Tuple[torch.Tensor, torch.Tensor]: """Gets the q values from the model and target networks""" q_values = self.q_network(state, action) q_values_target = self.q_network_target(state, action) return q_values, q_values_target def train_step_gen(self, training_batch: rlt.ParametricDqnInput, batch_idx: int): reward = training_batch.reward not_terminal = training_batch.not_terminal.float() discount_tensor = torch.full_like(reward, self.gamma) if self.use_seq_num_diff_as_time_diff: assert self.multi_steps is None discount_tensor = torch.pow(self.gamma, training_batch.time_diff.float()) if self.multi_steps is not None: # pyre-fixme[16]: Optional type has no attribute `float`. discount_tensor = torch.pow(self.gamma, training_batch.step.float()) if self.maxq_learning: # Assuming actions are parametrized in a k-dimensional space # tiled_state = (batch_size * max_num_action, state_dim) # possible_actions = (batch_size* max_num_action, k) # possible_actions_mask = (batch_size, max_num_action) product = training_batch.possible_next_actions.float_features.shape[0] batch_size = training_batch.possible_actions_mask.shape[0] assert product % batch_size == 0, ( f"batch_size * max_num_action {product} is " f"not divisible by batch_size {batch_size}" ) max_num_action = product // batch_size tiled_next_state = training_batch.next_state.get_tiled_batch(max_num_action) ( all_next_q_values, all_next_q_values_target, ) = self.get_detached_model_outputs( tiled_next_state, training_batch.possible_next_actions ) # Compute max a' Q(s', a') over all possible actions using target network next_q_values, _ = self.get_max_q_values_with_target( all_next_q_values, all_next_q_values_target, training_batch.possible_next_actions_mask.float(), ) assert ( len(next_q_values.shape) == 2 and next_q_values.shape[1] == 1 ), f"{next_q_values.shape}" else: # SARSA (Use the target network) _, next_q_values = self.get_detached_model_outputs( training_batch.next_state, training_batch.next_action ) assert ( len(next_q_values.shape) == 2 and next_q_values.shape[1] == 1 ), f"{next_q_values.shape}" target_q_values = reward + not_terminal * discount_tensor * next_q_values assert ( target_q_values.shape[-1] == 1 ), f"{target_q_values.shape} doesn't end with 1" # Get Q-value of action taken q_values = self.q_network(training_batch.state, training_batch.action) assert ( target_q_values.shape == q_values.shape ), f"{target_q_values.shape} != {q_values.shape}." td_loss = self.q_network_loss(q_values, target_q_values) yield td_loss # pyre-fixme[16]: Optional type has no attribute `metrics`. if training_batch.extras.metrics is not None: metrics_reward_concat_real_vals = torch.cat( (reward, training_batch.extras.metrics), dim=1 ) else: metrics_reward_concat_real_vals = reward # get reward estimates reward_estimates = self.reward_network( training_batch.state, training_batch.action ) reward_loss = F.mse_loss( reward_estimates.squeeze(-1), metrics_reward_concat_real_vals.squeeze(-1), ) yield reward_loss self.reporter.log( td_loss=td_loss.detach().cpu(), reward_loss=reward_loss.detach().cpu(), logged_rewards=reward, model_values_on_logged_actions=q_values.detach().cpu(), ) # Use the soft update rule to update target network yield self.soft_update_result()
7,284
39.027473
88
py
ReAgent
ReAgent-master/reagent/training/sac_trainer.py
#!/usr/bin/env python3 # Copyright (c) Facebook, Inc. and its affiliates. All rights reserved. import copy import logging from typing import List, Optional import numpy as np import reagent.core.types as rlt import torch import torch.nn.functional as F from reagent.core.configuration import resolve_defaults from reagent.core.dataclasses import dataclass from reagent.core.dataclasses import field from reagent.core.parameters import RLParameters from reagent.models.actor import LOG_PROB_MIN, LOG_PROB_MAX from reagent.optimizer import Optimizer__Union, SoftUpdate from reagent.training.reagent_lightning_module import ReAgentLightningModule from reagent.training.rl_trainer_pytorch import RLTrainerMixin logger = logging.getLogger(__name__) @dataclass class CRRWeightFn: # pick indicator or exponent indicator_fn_threshold: Optional[float] = None exponent_beta: Optional[float] = None exponent_clamp: Optional[float] = None def __post_init_post_parse__(self): assert self.exponent_beta or self.indicator_fn_threshold assert not (self.exponent_beta and self.indicator_fn_threshold) if self.exponent_beta: assert self.exponent_beta > 1e-6 if self.exponent_clamp: assert self.exponent_clamp > 1e-6 def get_weight_from_advantage(self, advantage): if self.indicator_fn_threshold: return (advantage >= self.indicator_fn_threshold).float() if self.exponent_beta: exp = torch.exp(advantage / self.exponent_beta) if self.exponent_clamp: exp = torch.clamp(exp, 0.0, self.exponent_clamp) return exp class SACTrainer(RLTrainerMixin, ReAgentLightningModule): """ Soft Actor-Critic trainer as described in https://arxiv.org/pdf/1801.01290 The actor is assumed to implement reparameterization trick. """ @resolve_defaults def __init__( self, actor_network, q1_network, q2_network=None, value_network=None, # Start SACTrainerParameters rl: RLParameters = field(default_factory=RLParameters), # noqa: B008 q_network_optimizer: Optimizer__Union = field( # noqa: B008 default_factory=Optimizer__Union.default ), value_network_optimizer: Optimizer__Union = field( # noqa: B008 default_factory=Optimizer__Union.default ), actor_network_optimizer: Optimizer__Union = field( # noqa: B008 default_factory=Optimizer__Union.default ), alpha_optimizer: Optional[Optimizer__Union] = field( # noqa: B008 default_factory=Optimizer__Union.default ), minibatch_size: int = 1024, entropy_temperature: float = 0.01, logged_action_uniform_prior: bool = True, target_entropy: float = -1.0, action_embedding_kld_weight: Optional[float] = None, apply_kld_on_mean: bool = False, action_embedding_mean: Optional[List[float]] = None, action_embedding_variance: Optional[List[float]] = None, crr_config: Optional[CRRWeightFn] = None, backprop_through_log_prob: bool = True, ) -> None: """ Args: actor_network: states -> actions, trained to maximize soft value, which is value + policy entropy. q1_network: states, action -> q-value q2_network (optional): double q-learning to stabilize training from overestimation bias value_network (optional): states -> value of state under actor # alpha in the paper; controlling explore & exploit backprop_through_log_prob: This is mostly for backward compatibility issue; we used to have a bug that does this and it yields a better result in some cases # TODO: finish """ super().__init__() self.rl_parameters = rl self.q1_network = q1_network self.q2_network = q2_network self.q_network_optimizer = q_network_optimizer self.value_network = value_network self.value_network_optimizer = value_network_optimizer if self.value_network is not None: self.value_network_target = copy.deepcopy(self.value_network) else: self.q1_network_target = copy.deepcopy(self.q1_network) self.q2_network_target = copy.deepcopy(self.q2_network) self.actor_network = actor_network self.actor_network_optimizer = actor_network_optimizer self.entropy_temperature = entropy_temperature self.alpha_optimizer = alpha_optimizer if alpha_optimizer is not None: self.target_entropy = target_entropy self.log_alpha = torch.nn.Parameter( torch.tensor([np.log(self.entropy_temperature)]) ) self.logged_action_uniform_prior = logged_action_uniform_prior self.add_kld_to_loss = bool(action_embedding_kld_weight) self.apply_kld_on_mean = apply_kld_on_mean if self.add_kld_to_loss: self.kld_weight = action_embedding_kld_weight # Calling register_buffer so that the tensors got moved to the right device self.register_buffer("action_emb_mean", None) self.register_buffer("action_emb_variance", None) # Assigning the values here instead of above so that typechecker wouldn't complain self.action_emb_mean = torch.tensor(action_embedding_mean) self.action_emb_variance = torch.tensor(action_embedding_variance) self.crr_config = crr_config if crr_config: assert self.value_network is not None self.backprop_through_log_prob = backprop_through_log_prob def configure_optimizers(self): optimizers = [] optimizers.append( self.q_network_optimizer.make_optimizer_scheduler( self.q1_network.parameters() ) ) if self.q2_network: optimizers.append( self.q_network_optimizer.make_optimizer_scheduler( self.q2_network.parameters() ) ) optimizers.append( self.actor_network_optimizer.make_optimizer_scheduler( self.actor_network.parameters() ) ) if self.alpha_optimizer is not None: optimizers.append( self.alpha_optimizer.make_optimizer_scheduler([self.log_alpha]) ) if self.value_network: optimizers.append( self.value_network_optimizer.make_optimizer_scheduler( self.value_network.parameters() ) ) # soft-update if self.value_network: target_params = self.value_network_target.parameters() source_params = self.value_network.parameters() else: target_params = list(self.q1_network_target.parameters()) source_params = list(self.q1_network.parameters()) if self.q2_network: target_params += list(self.q2_network_target.parameters()) source_params += list(self.q2_network.parameters()) optimizers.append( SoftUpdate.make_optimizer_scheduler( target_params, source_params, tau=self.tau ) ) return optimizers def train_step_gen(self, training_batch: rlt.PolicyNetworkInput, batch_idx: int): """ IMPORTANT: the input action here is assumed to match the range of the output of the actor. """ assert isinstance(training_batch, rlt.PolicyNetworkInput) state = training_batch.state action = training_batch.action reward = training_batch.reward discount = torch.full_like(reward, self.gamma) not_done_mask = training_batch.not_terminal # # First, optimize Q networks; minimizing MSE between # Q(s, a) & r + discount * V'(next_s) # if self.value_network is not None: next_state_value = self.value_network_target(training_batch.next_state) else: next_state_actor_output = self.actor_network(training_batch.next_state) next_state_actor_action = ( training_batch.next_state, rlt.FeatureData(next_state_actor_output.action), ) next_state_value = self.q1_network_target(*next_state_actor_action) if self.q2_network is not None: target_q2_value = self.q2_network_target(*next_state_actor_action) next_state_value = torch.min(next_state_value, target_q2_value) log_prob_a = self.actor_network.get_log_prob( training_batch.next_state, next_state_actor_output.action ).clamp(LOG_PROB_MIN, LOG_PROB_MAX) next_state_value -= self.entropy_temperature * log_prob_a if self.gamma > 0.0: target_q_value = ( reward + discount * next_state_value * not_done_mask.float() ) else: # This is useful in debugging instability issues target_q_value = reward q1_value = self.q1_network(state, action) q1_loss = F.mse_loss(q1_value, target_q_value) yield q1_loss if self.q2_network: q2_value = self.q2_network(state, action) q2_loss = F.mse_loss(q2_value, target_q_value) yield q2_loss # Second, optimize the actor; minimizing KL-divergence between # propensity & softmax of value. Due to reparameterization trick, # it ends up being log_prob(actor_action) - Q(s, actor_action) actor_output = self.actor_network(state) state_actor_action = (state, rlt.FeatureData(actor_output.action)) q1_actor_value = self.q1_network(*state_actor_action) min_q_actor_value = q1_actor_value if self.q2_network: q2_actor_value = self.q2_network(*state_actor_action) min_q_actor_value = torch.min(q1_actor_value, q2_actor_value) actor_log_prob = actor_output.log_prob.clamp(LOG_PROB_MIN, LOG_PROB_MAX) if not self.backprop_through_log_prob: actor_log_prob = actor_log_prob.detach() if self.crr_config is not None: cur_value = self.value_network(training_batch.state) advantage = (min_q_actor_value - cur_value).detach() # pyre-fixme[16]: `Optional` has no attribute `get_weight_from_advantage`. crr_weight = self.crr_config.get_weight_from_advantage(advantage) assert ( actor_log_prob.shape == crr_weight.shape ), f"{actor_log_prob.shape} != {crr_weight.shape}" actor_loss = -(actor_log_prob * crr_weight.detach()) else: actor_loss = self.entropy_temperature * actor_log_prob - min_q_actor_value # Do this in 2 steps so we can log histogram of actor loss actor_loss_mean = actor_loss.mean() if self.add_kld_to_loss: if self.apply_kld_on_mean: action_batch_m = torch.mean(actor_output.squashed_mean, axis=0) action_batch_v = torch.var(actor_output.squashed_mean, axis=0) else: action_batch_m = torch.mean(actor_output.action, axis=0) action_batch_v = torch.var(actor_output.action, axis=0) kld = ( 0.5 * ( (action_batch_v + (action_batch_m - self.action_emb_mean) ** 2) / self.action_emb_variance - 1 + self.action_emb_variance.log() - action_batch_v.log() ).sum() ) actor_loss_mean += self.kld_weight * kld yield actor_loss_mean # Optimize Alpha if self.alpha_optimizer is not None: alpha_loss = -( ( self.log_alpha * ( actor_output.log_prob.clamp(LOG_PROB_MIN, LOG_PROB_MAX) + self.target_entropy ).detach() ).mean() ) yield alpha_loss self.entropy_temperature = self.log_alpha.exp() # # Lastly, if applicable, optimize value network; minimizing MSE between # V(s) & E_a~pi(s) [ Q(s,a) - log(pi(a|s)) ] # if self.value_network is not None: state_value = self.value_network(state) if self.logged_action_uniform_prior: log_prob_a = torch.zeros_like(min_q_actor_value) target_value = min_q_actor_value else: log_prob_a = actor_output.log_prob.clamp(LOG_PROB_MIN, LOG_PROB_MAX) target_value = min_q_actor_value - self.entropy_temperature * log_prob_a value_loss = F.mse_loss(state_value, target_value.detach()) yield value_loss self.logger.log_metrics( { "td_loss": q1_loss, "logged_rewards": reward.mean(), "model_values_on_logged_actions": q1_value.mean(), "q1_value": q1_value.mean(), "entropy_temperature": self.entropy_temperature, "log_prob_a": log_prob_a.mean(), "next_state_value": next_state_value.mean(), "target_q_value": target_q_value.mean(), "min_q_actor_value": min_q_actor_value.mean(), "actor_output_log_prob": actor_output.log_prob.mean(), "actor_loss": actor_loss.mean(), }, step=self.all_batches_processed, ) if self.q2_network: self.logger.log_metrics( {"q2_value": q2_value.mean()}, step=self.all_batches_processed, ) if self.value_network: self.logger.log_metrics( {"target_state_value": target_value.mean()}, step=self.all_batches_processed, ) if self.add_kld_to_loss: self.logger.log_metrics( { "action_batch_mean": action_batch_m.mean(), "action_batch_var": action_batch_v.mean(), # pyre-fixme[61]: `kld` may not be initialized here. "kld": kld, }, step=self.all_batches_processed, ) # Use the soft update rule to update the target networks result = self.soft_update_result() self.log("td_loss", q1_loss, prog_bar=True) yield result
14,825
38.118734
94
py
ReAgent
ReAgent-master/reagent/training/utils.py
#!/usr/bin/env python3 # Copyright (c) Facebook, Inc. and its affiliates. All rights reserved. import numpy as np import torch import torch.nn.functional as F EPS = np.finfo(float).eps.item() def rescale_actions( actions: torch.Tensor, new_min: torch.Tensor, new_max: torch.Tensor, prev_min: torch.Tensor, prev_max: torch.Tensor, ) -> torch.Tensor: """Scale from [prev_min, prev_max] to [new_min, new_max]""" assert torch.all(prev_min <= actions) and torch.all( actions <= prev_max ), f"{actions} has values outside of [{prev_min}, {prev_max}]." assert torch.all( new_min <= new_max ), f"{new_min} is (has coordinate) greater than {new_max}." prev_range = prev_max - prev_min new_range = new_max - new_min return ((actions - prev_min) / prev_range) * new_range + new_min def whiten(x: torch.Tensor, subtract_mean: bool) -> torch.Tensor: numer = x if subtract_mean: numer -= x.mean() return numer / (x.std() + EPS) def discounted_returns(rewards: torch.Tensor, gamma: float = 0) -> torch.Tensor: """Perform rollout to compute reward to go and do a baseline subtraction.""" if gamma == 0: return rewards.float() else: R = 0 returns = [] for r in rewards.numpy()[::-1]: R = r + gamma * R returns.insert(0, R) return torch.tensor(returns).float() def gen_permutations(seq_len: int, num_action: int) -> torch.Tensor: """ generate all seq_len permutations for a given action set the return shape is (SEQ_LEN, PERM_NUM, ACTION_DIM) """ all_permut = torch.cartesian_prod(*[torch.arange(num_action)] * seq_len) if seq_len == 1: all_permut = all_permut.unsqueeze(1) all_permut = F.one_hot(all_permut, num_action).transpose(0, 1) return all_permut.float()
1,864
29.080645
80
py
ReAgent
ReAgent-master/reagent/training/reagent_lightning_module.py
#!/usr/bin/env python3 import inspect import logging import pytorch_lightning as pl import torch from reagent.core.tensorboardX import SummaryWriterContext from reagent.core.utils import lazy_property from typing_extensions import final logger = logging.getLogger(__name__) class ReAgentLightningModule(pl.LightningModule): def __init__(self, automatic_optimization=True): super().__init__() self._automatic_optimization = automatic_optimization self._training_step_generator = None self._reporter = pl.loggers.base.DummyExperiment() # For the generator API self._verified_steps = False # For summary_writer property self._summary_writer_logger = None self._summary_writer = None # To enable incremental training self.register_buffer("_next_stopping_epoch", None) self.register_buffer("_cleanly_stopped", None) self._next_stopping_epoch = torch.tensor([-1]).int() self._cleanly_stopped = torch.ones(1) self._setup_input_type() self.batches_processed_this_epoch = 0 self.all_batches_processed = 0 def _setup_input_type(self): self._training_batch_type = None sig = inspect.signature(self.train_step_gen) assert "training_batch" in sig.parameters param = sig.parameters["training_batch"] annotation = param.annotation if annotation == inspect.Parameter.empty: return if hasattr(annotation, "from_dict"): self._training_batch_type = annotation def set_reporter(self, reporter): if reporter is None: reporter = pl.loggers.base.DummyExperiment() self._reporter = reporter return self @property def reporter(self): return self._reporter def set_clean_stop(self, clean_stop: bool): self._cleanly_stopped[0] = int(clean_stop) def increase_next_stopping_epochs(self, num_epochs: int): self._next_stopping_epoch += num_epochs self.set_clean_stop(False) return self def train_step_gen(self, training_batch, batch_idx: int): """ Implement training step as generator here """ raise NotImplementedError def soft_update_result(self) -> torch.Tensor: """ A dummy loss to trigger soft-update """ one = torch.ones(1, requires_grad=True) return one + one @property def summary_writer(self): """ Accessor to TensorBoard's SummaryWriter """ if self._summary_writer_logger is self.logger: # If self.logger doesn't change between call, then return cached result return self._summary_writer # Invalidate self._summary_writer = None self._summary_writer_logger = self.logger if isinstance(self.logger, pl.loggers.base.LoggerCollection): for logger in self.logger._logger_iterable: if isinstance(logger, pl.loggers.tensorboard.TensorBoardLogger): self._summary_writer = logger.experiment break elif isinstance(logger, pl.loggers.tensorboard.TensorBoardLogger): self._summary_writer = logger.experiment return self._summary_writer # pyre-fixme[14]: `training_step` overrides method defined in `LightningModule` # inconsistently. def training_step(self, batch, batch_idx: int, optimizer_idx: int = 0): assert (optimizer_idx == 0) or (self._num_optimizing_steps > 1) if self._training_step_generator is None: if self._training_batch_type and isinstance(batch, dict): batch = self._training_batch_type.from_dict(batch) self._training_step_generator = self.train_step_gen(batch, batch_idx) ret = next(self._training_step_generator) if optimizer_idx == self._num_optimizing_steps - 1: if not self._verified_steps: try: next(self._training_step_generator) except StopIteration: self._verified_steps = True if not self._verified_steps: raise RuntimeError( "training_step_gen() yields too many times." "The number of yields should match the number of optimizers," f" in this case {self._num_optimizing_steps}" ) self._training_step_generator = None SummaryWriterContext.increase_global_step() return ret def optimizers(self, use_pl_optimizer: bool = True): o = super().optimizers(use_pl_optimizer) if isinstance(o, list): return o return [o] @lazy_property def _num_optimizing_steps(self) -> int: return len(self.configure_optimizers()) @final def on_epoch_end(self): logger.info( f"Finished epoch with {self.batches_processed_this_epoch} batches processed" ) self.batches_processed_this_epoch = 0 # Flush the reporter which has accumulated data in # training/validation/test self.reporter.flush(self.current_epoch) # Tell the trainer to stop. if self.current_epoch == self._next_stopping_epoch.item(): self.trainer.should_stop = True @final def on_train_batch_end(self, *args, **kwargs): self.batches_processed_this_epoch += 1 self.all_batches_processed += 1 @final def on_validation_batch_end(self, *args, **kwargs): self.batches_processed_this_epoch += 1 @final def on_test_batch_end(self, *args, **kwargs): self.batches_processed_this_epoch += 1 def train(self, *args): # trainer.train(batch) was the old, pre-Lightning ReAgent trainer API. # make sure that nobody is trying to call trainer.train() this way. # trainer.train() or trainer.train(True/False) is allowed - this puts the network into training/eval mode. if (len(args) == 0) or ((len(args) == 1) and (isinstance(args[0], bool))): super().train(*args) else: raise NotImplementedError( "Method .train() is not used for ReAgent Lightning trainers. Please use .fit() method of the pl.Trainer instead" ) class StoppingEpochCallback(pl.Callback): """ We use this callback to control the number of training epochs in incremental training. Epoch & step counts are not reset in the checkpoint. If we were to set `max_epochs` on the trainer, we would have to keep track of the previous `max_epochs` and add to it manually. This keeps the infomation in one place. Note that we need to set `_cleanly_stopped` back to True before saving the checkpoint. This is done in `ModelManager.save_trainer()`. """ def __init__(self, num_epochs): super().__init__() self.num_epochs = num_epochs def on_pretrain_routine_end(self, trainer, pl_module): assert isinstance(pl_module, ReAgentLightningModule) cleanly_stopped = pl_module._cleanly_stopped.item() logger.info(f"cleanly stopped: {cleanly_stopped}") if cleanly_stopped: pl_module.increase_next_stopping_epochs(self.num_epochs) def has_test_step_override(trainer_module: ReAgentLightningModule): """Detect if a subclass of LightningModule has test_step overridden""" return type(trainer_module).test_step != pl.LightningModule.test_step
7,568
35.921951
128
py
ReAgent
ReAgent-master/reagent/training/c51_trainer.py
#!/usr/bin/env python3 # Copyright (c) Facebook, Inc. and its affiliates. All rights reserved. from typing import List import reagent.core.types as rlt import torch from reagent.core.configuration import resolve_defaults from reagent.core.dataclasses import field from reagent.core.parameters import RLParameters from reagent.optimizer import Optimizer__Union, SoftUpdate from reagent.training.reagent_lightning_module import ReAgentLightningModule from reagent.training.rl_trainer_pytorch import RLTrainerMixin class C51Trainer(RLTrainerMixin, ReAgentLightningModule): """ Implementation of 51 Categorical DQN (C51) See https://arxiv.org/abs/1707.06887 for details """ @resolve_defaults def __init__( self, q_network, q_network_target, actions: List[str] = field(default_factory=list), # noqa: B008 rl: RLParameters = field(default_factory=RLParameters), # noqa: B008 double_q_learning: bool = True, minibatch_size: int = 1024, minibatches_per_step: int = 1, num_atoms: int = 51, qmin: float = -100, qmax: float = 200, optimizer: Optimizer__Union = field( # noqa: B008 default_factory=Optimizer__Union.default ), ) -> None: """ Args: q_network: states, action -> q-value q_network_target: model that provides targets actions(optional): list of agent's actions rl (optional): an instance of the RLParameter class, which defines relevant hyperparameters double_q_learning (optional): whether or not double Q learning, enabled by default, minibatch_size (optional): the size of the minibatch minibatches_per_step (optional): the number of minibatch updates per training step num_atoms (optional): number of "canonical returns"in the discretized value distributions qmin (optional): minimum q-value qmax (optional): maximum q-value optimizer (optional): the optimizer class and optimizer hyperparameters for the q network(s) optimizer """ super().__init__() self.double_q_learning = double_q_learning self.minibatch_size = minibatch_size self.minibatches_per_step = minibatches_per_step self._actions = actions self.q_network = q_network self.q_network_target = q_network_target self.q_network_optimizer = optimizer self.qmin = qmin self.qmax = qmax self.num_atoms = num_atoms self.rl_parameters = rl self.register_buffer("support", None) self.support = torch.linspace(self.qmin, self.qmax, self.num_atoms) self.scale_support = (self.qmax - self.qmin) / (self.num_atoms - 1.0) self.register_buffer("reward_boosts", None) self.reward_boosts = torch.zeros([1, len(self._actions)]) if self.rl_parameters.reward_boost is not None: # pyre-fixme[16]: Optional type has no attribute `keys`. for k in self.rl_parameters.reward_boost.keys(): i = self._actions.index(k) # pyre-fixme[16]: Optional type has no attribute `__getitem__`. self.reward_boosts[0, i] = self.rl_parameters.reward_boost[k] def configure_optimizers(self): optimizers = [ self.q_network_optimizer.make_optimizer_scheduler( self.q_network.parameters() ) ] # soft-update target_params = list(self.q_network_target.parameters()) source_params = list(self.q_network.parameters()) optimizers.append( SoftUpdate.make_optimizer_scheduler( target_params, source_params, tau=self.tau ) ) return optimizers def train_step_gen(self, training_batch: rlt.DiscreteDqnInput, batch_idx: int): rewards = self.boost_rewards(training_batch.reward, training_batch.action) discount_tensor = torch.full_like(rewards, self.gamma) possible_next_actions_mask = training_batch.possible_next_actions_mask.float() possible_actions_mask = training_batch.possible_actions_mask.float() not_terminal = training_batch.not_terminal.float() if self.use_seq_num_diff_as_time_diff: assert self.multi_steps is None discount_tensor = torch.pow(self.gamma, training_batch.time_diff.float()) if self.multi_steps is not None: assert training_batch.step is not None discount_tensor = torch.pow(self.gamma, training_batch.step.float()) next_dist = self.q_network_target.log_dist(training_batch.next_state).exp() if self.maxq_learning: # Select distribution corresponding to max valued action if self.double_q_learning: next_q_values = ( self.q_network.log_dist(training_batch.next_state).exp() * self.support ).sum(2) else: next_q_values = (next_dist * self.support).sum(2) next_action = self.argmax_with_mask( next_q_values, possible_next_actions_mask ) next_dist = next_dist[range(rewards.shape[0]), next_action.reshape(-1)] else: next_dist = (next_dist * training_batch.next_action.unsqueeze(-1)).sum(1) # Build target distribution target_Q = rewards + discount_tensor * not_terminal * self.support target_Q = target_Q.clamp(self.qmin, self.qmax) # rescale to indicies [0, 1, ..., N-1] b = (target_Q - self.qmin) / self.scale_support lo = b.floor().to(torch.int64) up = b.ceil().to(torch.int64) # handle corner cases of l == b == u # without the following, it would give 0 signal, whereas we want # m to add p(s_t+n, a*) to index l == b == u. # So we artificially adjust l and u. # (1) If 0 < l == u < N-1, we make l = l-1, so b-l = 1 # (2) If 0 == l == u, we make u = 1, so u-b=1 # (3) If l == u == N-1, we make l = N-2, so b-1 = 1 # This first line handles (1) and (3). lo[(up > 0) * (lo == up)] -= 1 # Note: l has already changed, so the only way l == u is possible is # if u == 0, in which case we let u = 1 # I don't even think we need the first condition in the next line up[(lo < (self.num_atoms - 1)) * (lo == up)] += 1 # distribute the probabilities # m_l = m_l + p(s_t+n, a*)(u - b) # m_u = m_u + p(s_t+n, a*)(b - l) m = torch.zeros_like(next_dist) # pyre-fixme[16]: `Tensor` has no attribute `scatter_add_`. m.scatter_add_(dim=1, index=lo, src=next_dist * (up.float() - b)) m.scatter_add_(dim=1, index=up, src=next_dist * (b - lo.float())) log_dist = self.q_network.log_dist(training_batch.state) # for reporting only all_q_values = (log_dist.exp() * self.support).sum(2).detach() model_action_idxs = self.argmax_with_mask( all_q_values, possible_actions_mask if self.maxq_learning else training_batch.action, ) log_dist = (log_dist * training_batch.action.unsqueeze(-1)).sum(1) loss = -(m * log_dist).sum(1).mean() if batch_idx % self.trainer.log_every_n_steps == 0: self.reporter.log( td_loss=loss, logged_actions=torch.argmax(training_batch.action, dim=1, keepdim=True), logged_propensities=training_batch.extras.action_probability, logged_rewards=rewards, model_values=all_q_values, model_action_idxs=model_action_idxs, ) self.log("td_loss", loss, prog_bar=True) yield loss result = self.soft_update_result() yield result # pyre-fixme[56]: Decorator `torch.no_grad(...)` could not be called, because # its type `no_grad` is not callable. @torch.no_grad() def boost_rewards( self, rewards: torch.Tensor, actions: torch.Tensor ) -> torch.Tensor: # Apply reward boost if specified reward_boosts = torch.sum( actions.float() * self.reward_boosts, dim=1, keepdim=True ) return rewards + reward_boosts def argmax_with_mask(self, q_values, possible_actions_mask): # Set q-values of impossible actions to a very large negative number. q_values = q_values.reshape(possible_actions_mask.shape) q_values = q_values + RLTrainerMixin.ACTION_NOT_POSSIBLE_VAL * ( 1 - possible_actions_mask ) return q_values.argmax(1)
8,784
41.033493
101
py
ReAgent
ReAgent-master/reagent/training/multi_stage_trainer.py
#!/usr/bin/env python3 import bisect import functools import itertools from collections import OrderedDict from typing import List, Dict, Tuple import torch.nn as nn from pytorch_lightning.loops.optimization.optimizer_loop import ClosureResult from reagent.core.utils import lazy_property from .reagent_lightning_module import ReAgentLightningModule class MultiStageTrainer(ReAgentLightningModule): def __init__( self, trainers: List[ReAgentLightningModule], epochs: List[int], assign_reporter_function=None, flush_reporter_function=None, automatic_optimization=True, ): super().__init__(automatic_optimization=automatic_optimization) # NB: wrapping in a ModuleList so the state can be saved self._trainers = nn.ModuleList(trainers) self._assign_reporter_function = assign_reporter_function self._flush_reporter_function = ( functools.partial(flush_reporter_function, self) if flush_reporter_function else self._flush_reporter ) self._in_testing_loop = False # Cumulative sum of number of epochs up to the index (of trainers) self._trainer_acc_epochs = [0] + epochs for i in range(1, len(epochs) + 1): self._trainer_acc_epochs[i] += self._trainer_acc_epochs[i - 1] # Num of epochs for each trainer. Used to check if the sum of them # equals to num_epochs used in pytorch-lightning trainer self.trainer_epoch_mapping = OrderedDict() for t, e in zip(trainers, epochs): trainer_name = type(t).__name__ self.trainer_epoch_mapping[trainer_name] = e @property def multi_stage_total_epochs(self): return self._trainer_acc_epochs[-1] def set_reporter(self, reporter): super().set_reporter(reporter) if self._assign_reporter_function: self._assign_reporter_function(self._trainers, reporter) else: # By default, assume CompoundReporter with the same # number of reporters as trainers assert len(self._trainers) == len( reporter._reporters ), f"{len(self._trainers)} != {len(reporter._reporters)}" for t, r in zip(self._trainers, reporter._reporters): t.set_reporter(r) @lazy_property def _optimizer_step_to_trainer_idx(self) -> Dict[int, Tuple[int, int]]: mapping = {} offset = 0 for i, t in enumerate(self._trainers): num_optimizing_steps = t._num_optimizing_steps for j in range(num_optimizing_steps): mapping[offset + j] = (i, offset) offset += num_optimizing_steps return mapping def _flush_reporter(self, reporter, epoch): """ By default, assume CompoundReporter with the same number of reporters as trainers """ if not self._in_testing_loop: epoch_trainer_idx = self._get_trainer_idx_from_epoch() reporter._reporters[epoch_trainer_idx].flush(epoch) else: for r in reporter._reporters: r.flush(epoch) def on_fit_start(self): self._starting_epoch = self.trainer.current_epoch # Connecting pl.Trainer to stage trainers for t in self._trainers: t.trainer = self.trainer t.on_fit_start() self.reporter.set_flush_function(self._flush_reporter_function) def on_fit_end(self): del self._starting_epoch # Disconnecting for t in self._trainers: t.on_fit_end() del t.trainer self.reporter.set_flush_function(None) def on_test_start(self): self._starting_epoch = self.trainer.current_epoch self._in_testing_loop = True for t in self._trainers: t.on_test_start() def on_test_end(self): del self._starting_epoch self._in_testing_loop = False for t in self._trainers: t.on_test_end() def _get_trainer_idx_from_epoch(self): # Cycling through the trainers epoch = (self.trainer.current_epoch - self._starting_epoch) % ( self._trainer_acc_epochs[-1] ) trainer_idx = bisect.bisect_right(self._trainer_acc_epochs, epoch) - 1 return trainer_idx def configure_optimizers(self): # FIXME: Doesn't support LRScheduler yet return list( itertools.chain(*[t.configure_optimizers() for t in self._trainers]) ) def training_step(self, batch, batch_idx: int, optimizer_idx: int = 0): trainer_idx, offset = self._optimizer_step_to_trainer_idx[optimizer_idx] epoch_trainer_idx = self._get_trainer_idx_from_epoch() assert ( trainer_idx == epoch_trainer_idx ), f"Got {trainer_idx}; expected {epoch_trainer_idx}" return self._trainers[trainer_idx].training_step( batch, batch_idx, optimizer_idx - offset ) def training_epoch_end(self, outputs): epoch_trainer_idx = self._get_trainer_idx_from_epoch() self._trainers[epoch_trainer_idx].training_epoch_end(outputs) def validation_step(self, *args, **kwargs): epoch_trainer_idx = self._get_trainer_idx_from_epoch() return self._trainers[epoch_trainer_idx].validation_step(*args, **kwargs) def validation_epoch_end(self, outputs): epoch_trainer_idx = self._get_trainer_idx_from_epoch() self._trainers[epoch_trainer_idx].validation_epoch_end(outputs) def test_step(self, *args, **kwargs): return { str(i): trainer.test_step(*args, **kwargs) for i, trainer in enumerate(self._trainers) } def test_epoch_end(self, outputs): for i, trainer in enumerate(self._trainers): trainer.test_epoch_end([o[str(i)] for o in outputs]) def optimizer_step( self, epoch: int, batch_idx: int, optimizer, optimizer_idx: int, optimizer_closure, on_tpu: int = False, using_native_amp: int = False, using_lbfgs: int = False, ): assert epoch == self.trainer.current_epoch epoch_trainer_idx = self._get_trainer_idx_from_epoch() optimizer_trainer_idx, offset = self._optimizer_step_to_trainer_idx[ optimizer_idx ] if epoch_trainer_idx == optimizer_trainer_idx: # FIXME: epoch argument is not really correct # Trainer will see the total epochs, including those epochs they # are inactive. self._trainers[epoch_trainer_idx].optimizer_step( epoch, batch_idx, optimizer, optimizer_idx - offset, optimizer_closure, on_tpu=on_tpu, using_native_amp=using_native_amp, using_lbfgs=using_lbfgs, ) # FIXME: this is a hack around https://github.com/PyTorchLightning/pytorch-lightning/pull/9360 # which assumes that the optimizer closure will be consumed per training step invocation # however this is not true in the multi-stage trainer as the training step is called for *all* of the # optimizers configured under `trainers` even though only one lightning module is active at a given time # A more robust solution would be to use manual optimization, where the lightning trainer does no inspection # of the optimization closure for further processing elif hasattr(optimizer_closure, "_result"): optimizer_closure._result = ClosureResult(closure_loss=None)
7,726
36.509709
116
py
ReAgent
ReAgent-master/reagent/training/ppo_trainer.py
#!/usr/bin/env python3 # Copyright (c) Facebook, Inc. and its affiliates. All rights reserved. import inspect import logging from dataclasses import field from typing import Dict, List, Optional, Union import reagent.core.types as rlt import torch import torch.optim from reagent.core.configuration import resolve_defaults from reagent.gym.policies.policy import Policy from reagent.models.base import ModelBase from reagent.optimizer.union import Optimizer__Union from reagent.training.reagent_lightning_module import ReAgentLightningModule from reagent.training.utils import discounted_returns, whiten logger = logging.getLogger(__name__) class PPOTrainer(ReAgentLightningModule): """ Proximal Policy Optimization (PPO). See https://arxiv.org/pdf/1707.06347.pdf This is the "clip" version of PPO. It does not include: - KL divergence - Bootstrapping with a critic model (our approach only works if full trajectories up to terminal state are fed in) Optionally, a value network can be trained and used as a baseline for rewards. """ @resolve_defaults def __init__( self, policy: Policy, gamma: float = 0.9, optimizer: Optimizer__Union = field( # noqa: B008 default_factory=Optimizer__Union.default ), optimizer_value_net: Optimizer__Union = field( # noqa: B008 default_factory=Optimizer__Union.default ), actions: List[str] = field(default_factory=list), # noqa: B008 reward_clip: float = 1e6, normalize: bool = True, subtract_mean: bool = True, offset_clamp_min: bool = False, update_freq: int = 1, # how many env steps between updates update_epochs: int = 1, # how many epochs to run when updating (for PPO) ppo_batch_size: int = 1, # batch size (number of trajectories) used for PPO updates ppo_epsilon: float = 0.2, # clamp importance weights between 1-epsilon and 1+epsilon entropy_weight: float = 0.0, # weight of the entropy term in the PPO loss value_net: Optional[ModelBase] = None, ): # PPO relies on customized update schemas, achieved by manual_backward() super().__init__(automatic_optimization=False) self.scorer = policy.scorer self.sampler = policy.sampler self.gamma = gamma self.optimizer_value_net = optimizer_value_net self.actions = actions self.reward_clip = reward_clip self.normalize = normalize self.subtract_mean = subtract_mean self.offset_clamp_min = offset_clamp_min self.update_freq = update_freq self.update_epochs = update_epochs self.ppo_batch_size = ppo_batch_size self.ppo_epsilon = ppo_epsilon self.entropy_weight = entropy_weight self.optimizer = optimizer self.value_net = value_net if value_net is not None: self.value_loss_fn = torch.nn.MSELoss(reduction="mean") assert ( not self.normalize ), "Can't apply a value baseline and normalize rewards simultaneously" assert (ppo_epsilon >= 0) and ( ppo_epsilon <= 1 ), "ppo_epslion has to be in [0;1]" self.traj_buffer = [] def _trajectory_to_losses( self, trajectory: rlt.PolicyGradientInput ) -> Dict[str, torch.Tensor]: """ Get a dict of losses for the trajectory. Dict always includes PPO loss. If a value baseline is trained, a loss for the value network is also included. """ losses = {} actions = trajectory.action rewards = trajectory.reward.detach() scorer_inputs = [] if inspect.getattr_static(trajectory, "graph", None) is not None: # TODO: can this line be hit currently in ReAgent? # GNN scorer_inputs.append(trajectory.graph) else: scorer_inputs.append(trajectory.state) if trajectory.possible_actions_mask is not None: scorer_inputs.append(trajectory.possible_actions_mask) scores = self.scorer(*scorer_inputs) offset_reinforcement = discounted_returns( torch.clamp(rewards, max=self.reward_clip).clone(), self.gamma ) if self.normalize: offset_reinforcement = whiten( offset_reinforcement, subtract_mean=self.subtract_mean ) if self.offset_clamp_min: offset_reinforcement = offset_reinforcement.clamp(min=0) if self.value_net is not None: # subtract learned value function baselines from rewards baselines = self.value_net(trajectory.state).squeeze() # use reward-to-go as label for training the value function losses["value_net_loss"] = self.value_loss_fn( baselines, offset_reinforcement ) # detach bcs we want PPO to tweak policy, not baseline offset_reinforcement = offset_reinforcement - baselines.detach() target_propensity = self.sampler.log_prob(scores, actions).float() characteristic_eligibility = torch.exp( target_propensity - trajectory.log_prob.detach() ).float() losses["ppo_loss"] = -torch.min( offset_reinforcement.float() @ characteristic_eligibility, offset_reinforcement.float() @ torch.clamp( characteristic_eligibility, 1 - self.ppo_epsilon, 1 + self.ppo_epsilon, ), ) if self.entropy_weight != 0: entropy = self.sampler.entropy(scores) # "-" bcs minimizing, not maximizing losses["ppo_loss"] = losses["ppo_loss"] - self.entropy_weight * entropy return losses def configure_optimizers(self): optimizers = [] # value net optimizer if self.value_net is not None: optimizers.append( self.optimizer_value_net.make_optimizer_scheduler( self.value_net.parameters() ) ) # policy optimizer optimizers.append( self.optimizer.make_optimizer_scheduler(self.scorer.parameters()) ) return optimizers def get_optimizers(self): opts = self.optimizers() if self.value_net is not None: return opts[0], opts[1] return None, opts[0] # pyre-fixme[14]: `training_step` overrides method defined in # `ReAgentLightningModule` inconsistently. def training_step( self, training_batch: Union[rlt.PolicyGradientInput, Dict[str, torch.Tensor]], batch_idx: int, ): if isinstance(training_batch, dict): training_batch = rlt.PolicyGradientInput.from_dict(training_batch) self.traj_buffer.append(training_batch) if len(self.traj_buffer) == self.update_freq: self.update_model() def update_model(self): assert ( len(self.traj_buffer) == self.update_freq ), "trajectory buffer does not have sufficient samples for model_update" for _ in range(self.update_epochs): # iterate through minibatches of PPO updates in random order random_order = torch.randperm(len(self.traj_buffer)) for i in range(0, len(self.traj_buffer), self.ppo_batch_size): idx = random_order[i : i + self.ppo_batch_size] training_batch_list = [self.traj_buffer[i] for i in idx] self._update_model(training_batch_list) self.traj_buffer = [] # empty the buffer def _update_model(self, training_batch_list: List[rlt.PolicyGradientInput]): losses = { "ppo_loss": [], "value_net_loss": [], } value_net_opt, ppo_opt = self.get_optimizers() for traj in training_batch_list: loss = self._trajectory_to_losses(traj) for k, v in loss.items(): losses[k].append(v) if self.value_net is not None: # TD loss for the baseline value network value_net_loss = torch.stack(losses["value_net_loss"]).sum() value_net_opt.zero_grad() self.manual_backward(value_net_loss) value_net_opt.step() # PPO "loss" for the policy network ppo_loss = torch.stack(losses["ppo_loss"]).sum() ppo_opt.zero_grad() self.manual_backward(ppo_loss) ppo_opt.step()
8,551
38.410138
118
py
ReAgent
ReAgent-master/reagent/training/dqn_trainer_base.py
#!/usr/bin/env python3 # Copyright (c) Facebook, Inc. and its affiliates. All rights reserved. import logging from typing import List, Optional import reagent.core.types as rlt import torch import torch.nn.functional as F from reagent.core.parameters import EvaluationParameters, RLParameters from reagent.core.torch_utils import masked_softmax from reagent.evaluation.evaluation_data_page import EvaluationDataPage from reagent.evaluation.evaluator import Evaluator from reagent.optimizer import Optimizer__Union from reagent.training.reagent_lightning_module import ReAgentLightningModule from reagent.training.rl_trainer_pytorch import RLTrainerMixin logger = logging.getLogger(__name__) class DQNTrainerMixin: # Q-value for action that is not possible. Guaranteed to be worse than any # legitimate action ACTION_NOT_POSSIBLE_VAL = -1e9 def get_max_q_values(self, q_values, possible_actions_mask): return self.get_max_q_values_with_target( q_values, q_values, possible_actions_mask ) def get_max_q_values_with_target( self, q_values, q_values_target, possible_actions_mask ): """ Used in Q-learning update. :param q_values: PyTorch tensor with shape (batch_size, action_dim). Each row contains the list of Q-values for each possible action in this state. :param q_values_target: PyTorch tensor with shape (batch_size, action_dim). Each row contains the list of Q-values from the target network for each possible action in this state. :param possible_actions_mask: PyTorch tensor with shape (batch_size, action_dim). possible_actions[i][j] = 1 iff the agent can take action j from state i. Returns a tensor of maximum Q-values for every state in the batch and also the index of the corresponding action (which is used in evaluation_data_page.py, in create_from_tensors_dqn()). """ # The parametric DQN can create flattened q values so we reshape here. q_values = q_values.reshape(possible_actions_mask.shape) q_values_target = q_values_target.reshape(possible_actions_mask.shape) # Set q-values of impossible actions to a very large negative number. inverse_pna = 1 - possible_actions_mask impossible_action_penalty = self.ACTION_NOT_POSSIBLE_VAL * inverse_pna q_values = q_values + impossible_action_penalty q_values_target = q_values_target + impossible_action_penalty if self.double_q_learning: # Use indices of the max q_values from the online network to select q-values # from the target network. This prevents overestimation of q-values. # The torch.gather function selects the entry from each row that corresponds # to the max_index in that row. max_q_values, max_indicies = torch.max(q_values, dim=1, keepdim=True) max_q_values_target = torch.gather(q_values_target, 1, max_indicies) else: max_q_values_target, max_indicies = torch.max( q_values_target, dim=1, keepdim=True ) return max_q_values_target, max_indicies class DQNTrainerBaseLightning(DQNTrainerMixin, RLTrainerMixin, ReAgentLightningModule): def __init__( self, rl_parameters: RLParameters, metrics_to_score=None, actions: Optional[List[str]] = None, evaluation_parameters: Optional[EvaluationParameters] = None, ): super().__init__() self.rl_parameters = rl_parameters self.time_diff_unit_length = rl_parameters.time_diff_unit_length self.tensorboard_logging_freq = rl_parameters.tensorboard_logging_freq self.calc_cpe_in_training = ( evaluation_parameters and evaluation_parameters.calc_cpe_in_training ) assert actions is not None self._actions: List[str] = actions if rl_parameters.q_network_loss == "mse": self.q_network_loss = F.mse_loss elif rl_parameters.q_network_loss == "huber": self.q_network_loss = F.smooth_l1_loss else: raise Exception( "Q-Network loss type {} not valid loss.".format( rl_parameters.q_network_loss ) ) if metrics_to_score: self.metrics_to_score = metrics_to_score + ["reward"] else: self.metrics_to_score = ["reward"] def _check_input(self, training_batch: rlt.DiscreteDqnInput): assert isinstance(training_batch, rlt.DiscreteDqnInput) assert training_batch.not_terminal.dim() == training_batch.reward.dim() == 2 assert ( training_batch.not_terminal.shape[1] == training_batch.reward.shape[1] == 1 ) assert training_batch.action.dim() == training_batch.next_action.dim() == 2 assert ( training_batch.action.shape[1] == training_batch.next_action.shape[1] == self.num_actions ) if torch.logical_and( training_batch.possible_next_actions_mask.float().sum(dim=1) == 0, training_batch.not_terminal.squeeze().bool(), ).any(): # make sure there's no non-terminal state with no possible next actions raise ValueError( "No possible next actions. Should the environment have terminated?" ) @property def num_actions(self) -> int: assert self._actions is not None, "Not a discrete action DQN" return len(self._actions) # pyre-fixme[56]: Decorator `torch.no_grad(...)` could not be called, because # its type `no_grad` is not callable. @torch.no_grad() def boost_rewards( self, rewards: torch.Tensor, actions: torch.Tensor ) -> torch.Tensor: # Apply reward boost if specified reward_boosts = torch.sum( actions.float() * self.reward_boosts, dim=1, keepdim=True, ) return rewards + reward_boosts def _initialize_cpe( self, reward_network, q_network_cpe, q_network_cpe_target, optimizer: Optimizer__Union, ) -> None: if not self.calc_cpe_in_training: # pyre-fixme[16]: `DQNTrainerBase` has no attribute `reward_network`. self.reward_network = None return assert reward_network is not None, "reward_network is required for CPE" self.reward_network = reward_network # pyre-fixme[16]: `DQNTrainerBase` has no attribute `reward_network_optimizer`. self.reward_network_optimizer = optimizer assert ( q_network_cpe is not None and q_network_cpe_target is not None ), "q_network_cpe and q_network_cpe_target are required for CPE" # pyre-fixme[16]: `DQNTrainerBase` has no attribute `q_network_cpe`. self.q_network_cpe = q_network_cpe # pyre-fixme[16]: `DQNTrainerBase` has no attribute `q_network_cpe_target`. self.q_network_cpe_target = q_network_cpe_target # pyre-fixme[16]: `DQNTrainerBase` has no attribute `q_network_cpe_optimizer`. self.q_network_cpe_optimizer = optimizer num_output_nodes = len(self.metrics_to_score) * self.num_actions reward_idx_offsets = torch.arange( 0, num_output_nodes, self.num_actions, dtype=torch.long, ) self.register_buffer("reward_idx_offsets", reward_idx_offsets) reward_stripped_metrics_to_score = ( self.metrics_to_score[:-1] if len(self.metrics_to_score) > 1 else None ) # pyre-fixme[16]: `DQNTrainerBase` has no attribute `evaluator`. self.evaluator = Evaluator( self._actions, self.rl_parameters.gamma, self, metrics_to_score=reward_stripped_metrics_to_score, ) def _configure_cpe_optimizers(self): target_params = list(self.q_network_cpe_target.parameters()) source_params = list(self.q_network_cpe.parameters()) # TODO: why is reward net commented out? # source_params += list(self.reward_network.parameters()) optimizers = [] optimizers.append( self.reward_network_optimizer.make_optimizer_scheduler( self.reward_network.parameters() ) ) optimizers.append( self.q_network_cpe_optimizer.make_optimizer_scheduler( self.q_network_cpe.parameters() ) ) return target_params, source_params, optimizers def _calculate_cpes( self, training_batch, states, next_states, all_action_scores, all_next_action_scores, logged_action_idxs, discount_tensor, not_done_mask, ): if not self.calc_cpe_in_training: return if training_batch.extras.metrics is None: metrics_reward_concat_real_vals = training_batch.reward else: metrics_reward_concat_real_vals = torch.cat( (training_batch.reward, training_batch.extras.metrics), dim=1 ) model_propensities_next_states = masked_softmax( all_next_action_scores, training_batch.possible_next_actions_mask if self.maxq_learning else training_batch.next_action, self.rl_temperature, ) ######### Train separate reward network for CPE evaluation ############# reward_estimates = self.reward_network(states) reward_estimates_for_logged_actions = reward_estimates.gather( 1, self.reward_idx_offsets + logged_action_idxs ) reward_loss = F.mse_loss( reward_estimates_for_logged_actions, metrics_reward_concat_real_vals ) yield reward_loss ######### Train separate q-network for CPE evaluation ############# metric_q_values = self.q_network_cpe(states).gather( 1, self.reward_idx_offsets + logged_action_idxs ) all_metrics_target_q_values = torch.chunk( self.q_network_cpe_target(next_states).detach(), len(self.metrics_to_score), dim=1, ) target_metric_q_values = [] for i, per_metric_target_q_values in enumerate(all_metrics_target_q_values): per_metric_next_q_values = torch.sum( per_metric_target_q_values * model_propensities_next_states, 1, keepdim=True, ) per_metric_next_q_values = per_metric_next_q_values * not_done_mask per_metric_target_q_values = metrics_reward_concat_real_vals[ :, i : i + 1 ] + (discount_tensor * per_metric_next_q_values) target_metric_q_values.append(per_metric_target_q_values) target_metric_q_values = torch.cat(target_metric_q_values, dim=1) metric_q_value_loss = self.q_network_loss( metric_q_values, target_metric_q_values ) # The model_propensities computed below are not used right now. The CPE graphs in the Outputs # tab use model_propensities computed in the function create_from_tensors_dqn() in evaluation_data_page.py, # which is called on the eval_table_sample in the gather_eval_data() function below. model_propensities = masked_softmax( all_action_scores, training_batch.possible_actions_mask if self.maxq_learning else training_batch.action, self.rl_temperature, ) # Extract rewards predicted by the reward_network. The other columns will # give predicted values for other metrics, if such were specified. model_rewards = reward_estimates[ :, torch.arange( self.reward_idx_offsets[0], self.reward_idx_offsets[0] + self.num_actions, ), ] self.reporter.log( reward_loss=reward_loss, model_propensities=model_propensities, model_rewards=model_rewards, ) yield metric_q_value_loss def gather_eval_data(self, validation_step_outputs): was_on_gpu = self.on_gpu self.cpu() eval_data = None for edp in validation_step_outputs: if eval_data is None: eval_data = edp else: eval_data = eval_data.append(edp) if eval_data and eval_data.mdp_id is not None: eval_data = eval_data.sort() eval_data = eval_data.compute_values(self.gamma) eval_data.validate() if was_on_gpu: self.cuda() return eval_data def validation_step(self, batch, batch_idx): if isinstance(batch, dict): batch = rlt.DiscreteDqnInput.from_dict(batch) # HACK: Move to cpu in order to hold more batches in memory # This is only needed when trainers need in-memory # EvaluationDataPages of the full evaluation dataset return EvaluationDataPage.create_from_training_batch(batch, self).cpu() def validation_epoch_end(self, valid_step_outputs): # As explained in the comments to the validation_step function in # pytorch_lightning/core/lightning.py, this function is generally used as follows: # val_outs = [] # for val_batch in val_data: # out = validation_step(val_batch) # val_outs.append(out) # validation_epoch_end(val_outs) # The input arguments of validation_epoch_end() is a list of EvaluationDataPages, # which matches the way it is used in gather_eval_data() above. eval_data = self.gather_eval_data(valid_step_outputs) if eval_data and eval_data.mdp_id is not None: cpe_details = self.evaluator.evaluate_post_training(eval_data) self.reporter.log(cpe_details=cpe_details)
14,123
39.469914
115
py
ReAgent
ReAgent-master/reagent/training/slate_q_trainer.py
#!/usr/bin/env python3 # Copyright (c) Facebook, Inc. and its affiliates. All rights reserved. import enum import logging from typing import Optional import reagent.core.parameters as rlp import reagent.core.types as rlt import torch import torch.nn.functional as F from reagent.core.dataclasses import field from reagent.optimizer import Optimizer__Union, SoftUpdate from reagent.training.reagent_lightning_module import ReAgentLightningModule from reagent.training.rl_trainer_pytorch import RLTrainerMixin logger = logging.getLogger(__name__) class NextSlateValueNormMethod(enum.Enum): """ The Q value of the current slate item is the sum of the item's short-term reward and the normalized sum of all item Q-values on the next slate. We can normalize the sum by either the current slate size (NORM_BY_CURRENT_SLATE_SIZE) or the next slate size (NORM_BY_NEXT_SLATE_SIZE). This enum distinguishes between these two different ways of normalizing the next slate value. """ NORM_BY_CURRENT_SLATE_SIZE = "norm_by_current_slate_size" NORM_BY_NEXT_SLATE_SIZE = "norm_by_next_slate_size" class SlateQTrainer(RLTrainerMixin, ReAgentLightningModule): def __init__( self, q_network, q_network_target, slate_size, # Start SlateQTrainerParameters rl: rlp.RLParameters = field( # noqa: B008 default_factory=lambda: rlp.RLParameters(maxq_learning=False) ), optimizer: Optimizer__Union = field( # noqa: B008 default_factory=Optimizer__Union.default ), slate_opt_parameters: Optional[rlp.SlateOptParameters] = None, discount_time_scale: Optional[float] = None, single_selection: bool = True, next_slate_value_norm_method: NextSlateValueNormMethod = NextSlateValueNormMethod.NORM_BY_CURRENT_SLATE_SIZE, minibatch_size: int = 1024, evaluation: rlp.EvaluationParameters = field( # noqa: B008 default_factory=lambda: rlp.EvaluationParameters(calc_cpe_in_training=False) ), ) -> None: """ Args: q_network: states, action -> q-value slate_size(int): a fixed slate size rl (optional): an instance of the RLParameter class, which defines relevant hyperparameters optimizer (optional): the optimizer class and optimizer hyperparameters for the q network(s) optimizer discount_time_scale (optional): use to control the discount factor (gamma) relative to the time difference (t2-t1), i.e., gamma^((t2-t1)/time_scale). If it is absent, we won't adjust the discount factor by the time difference. single_selection (optional): TBD next_slate_value_norm_method (optional): how to calculate the next slate value when single_selection is False. By default we use NORM_BY_CURRENT_SLATE_SIZE. minibatch_size (optional): the size of the minibatch evaluation (optional): TBD """ super().__init__() self.rl_parameters = rl self.discount_time_scale = discount_time_scale self.single_selection = single_selection self.next_slate_value_norm_method = next_slate_value_norm_method self.q_network = q_network self.q_network_target = q_network_target self.q_network_optimizer = optimizer self.slate_size = slate_size self.slate_opt_parameters = slate_opt_parameters def configure_optimizers(self): optimizers = [] optimizers.append( self.q_network_optimizer.make_optimizer_scheduler( self.q_network.parameters() ) ) target_params = list(self.q_network_target.parameters()) source_params = list(self.q_network.parameters()) optimizers.append( SoftUpdate.make_optimizer_scheduler( target_params, source_params, tau=self.tau ) ) return optimizers def _action_docs( self, state: rlt.FeatureData, action: torch.Tensor, terminal_mask: Optional[torch.Tensor] = None, ) -> rlt.DocList: # for invalid indices, simply set action to 0 so we can batch index still if terminal_mask is not None: assert terminal_mask.shape == ( action.shape[0], ), f"{terminal_mask.shape} != 0th dim of {action.shape}" action[terminal_mask] = torch.zeros_like(action[terminal_mask]) docs = state.candidate_docs assert docs is not None return docs.select_slate(action) def _get_unmasked_q_values( self, q_network, state: rlt.FeatureData, slate: rlt.DocList ) -> torch.Tensor: """Gets the q values from the model and target networks""" batch_size, slate_size, _ = slate.float_features.shape # TODO: Probably should create a new model type return q_network( state.repeat_interleave(slate_size, dim=0), slate.as_feature_data() ).view(batch_size, slate_size) # pyre-fixme[56]: Decorator `torch.no_grad(...)` could not be called, because # its type `no_grad` is not callable. @torch.no_grad() def _get_maxq_next_action(self, next_state: rlt.FeatureData) -> torch.Tensor: """Get the next action list based on the slate optimization strategy.""" slate_opt_parameters = self.slate_opt_parameters assert slate_opt_parameters is not None if slate_opt_parameters.method == rlp.SlateOptMethod.TOP_K: return self._get_maxq_topk(next_state) else: raise NotImplementedError( "SlateQ with optimization method other than TOP_K is not implemented." ) def _get_maxq_topk(self, next_state: rlt.FeatureData) -> torch.Tensor: candidate_docs = next_state.candidate_docs assert candidate_docs is not None batch_size, num_candidates, _ = candidate_docs.float_features.shape assert 0 < self.slate_size <= num_candidates docs = candidate_docs.select_slate( torch.arange(num_candidates).repeat(batch_size, 1) ) next_q_values = self._get_unmasked_q_values( self.q_network_target, next_state, docs ) * self._get_docs_value(docs) _, next_actions = torch.topk(next_q_values, self.slate_size, dim=1) return next_actions def _get_docs_value(self, docs: rlt.DocList) -> torch.Tensor: # Multiplying by the mask to filter out selected padding items. value = docs.value * docs.mask if self.single_selection: value = F.softmax(value, dim=1) return value def _get_slate_size(self, state: rlt.FeatureData) -> torch.Tensor: """Get the actual size (ignore all padded items) of each slate by summing item masks.""" mask = self._get_item_mask(state) return torch.minimum( mask.sum(1, keepdim=True), torch.tensor([self.slate_size], device=mask.device), ) def _get_item_mask(self, state: rlt.FeatureData) -> torch.Tensor: """Get the mask from the given state.""" candidate_docs = state.candidate_docs assert candidate_docs is not None return candidate_docs.mask def _get_avg_by_slate_size(self, batch: rlt.SlateQInput): """Get the slate_size for averaging the sum of slate value.""" if ( self.next_slate_value_norm_method == NextSlateValueNormMethod.NORM_BY_NEXT_SLATE_SIZE ): return self._get_slate_size(batch.next_state) if ( self.next_slate_value_norm_method == NextSlateValueNormMethod.NORM_BY_CURRENT_SLATE_SIZE ): return self._get_slate_size(batch.state) raise NotImplementedError( f"The next_slate_value_norm_method {self.next_slate_value_norm_method} has not been implemented" ) def train_step_gen(self, training_batch: rlt.SlateQInput, batch_idx: int): assert isinstance( training_batch, rlt.SlateQInput ), f"learning input is a {type(training_batch)}" reward = training_batch.reward reward_mask = training_batch.reward_mask discount_tensor = torch.full_like(reward, self.gamma) # Adjust the discount factor by the time_diff if the discount_time_scale is provided, # and the time_diff exists in the training_batch. if self.discount_time_scale and training_batch.time_diff is not None: discount_tensor = discount_tensor ** ( training_batch.time_diff / self.discount_time_scale ) next_action = ( self._get_maxq_next_action(training_batch.next_state) if self.rl_parameters.maxq_learning else training_batch.next_action ) terminal_mask = (training_batch.not_terminal.to(torch.bool) == False).squeeze(1) next_action_docs = self._action_docs( training_batch.next_state, next_action, terminal_mask=terminal_mask, ) next_q_values = torch.sum( self._get_unmasked_q_values( self.q_network_target, training_batch.next_state, next_action_docs, ) * self._get_docs_value(next_action_docs), dim=1, keepdim=True, ) # If not single selection, divide max-Q by the actual slate size. if not self.single_selection: next_q_values = next_q_values / self._get_avg_by_slate_size(training_batch) filtered_max_q_vals = next_q_values * training_batch.not_terminal.float() target_q_values = reward + (discount_tensor * filtered_max_q_vals) # Don't mask if not single selection if self.single_selection: target_q_values = target_q_values[reward_mask] # Get Q-value of action taken action_docs = self._action_docs(training_batch.state, training_batch.action) q_values = self._get_unmasked_q_values( self.q_network, training_batch.state, action_docs ) if self.single_selection: q_values = q_values[reward_mask] all_action_scores = q_values.detach() value_loss = F.mse_loss(q_values, target_q_values) yield value_loss if not self.single_selection: all_action_scores = all_action_scores.sum(dim=1, keepdim=True) # Logging at the end to schedule all the cuda operations first self.reporter.log( td_loss=value_loss, model_values_on_logged_actions=all_action_scores, ) # Use the soft update rule to update the target networks result = self.soft_update_result() self.log("td_loss", value_loss, prog_bar=True) yield result
10,984
38.800725
117
py
ReAgent
ReAgent-master/reagent/training/imitator_training.py
#!/usr/bin/env python3 # Copyright (c) Facebook, Inc. and its affiliates. All rights reserved. import logging import torch logger = logging.getLogger(__name__) def get_valid_actions_from_imitator(imitator, input, drop_threshold): """Create mask for non-viable actions under the imitator.""" if isinstance(imitator, torch.nn.Module): # pytorch model imitator_outputs = imitator(input.float_features) on_policy_action_probs = torch.nn.functional.softmax(imitator_outputs, dim=1) else: # sci-kit learn model on_policy_action_probs = torch.tensor(imitator(input.float_features.cpu())) filter_values = ( on_policy_action_probs / on_policy_action_probs.max(keepdim=True, dim=1)[0] ) return (filter_values >= drop_threshold).float()
804
31.2
85
py
ReAgent
ReAgent-master/reagent/training/reward_network_trainer.py
#!/usr/bin/env python3 # Copyright (c) Facebook, Inc. and its affiliates. All rights reserved. import logging from enum import Enum from typing import Optional import numpy as np import reagent.core.types as rlt import torch from reagent.core.dataclasses import field from reagent.models.base import ModelBase from reagent.optimizer.union import Optimizer__Union from reagent.training.reagent_lightning_module import ReAgentLightningModule logger = logging.getLogger(__name__) class LossFunction(Enum): MSE = "MSE_Loss" SmoothL1Loss = "SmoothL1_Loss" L1Loss = "L1_Loss" BCELoss = "BCE_Loss" def _get_loss_function( loss_fn: LossFunction, reward_ignore_threshold: Optional[float], weighted_by_inverse_propensity: bool, ): reduction_type = "none" if loss_fn == LossFunction.MSE: torch_fn = torch.nn.MSELoss(reduction=reduction_type) elif loss_fn == LossFunction.SmoothL1Loss: torch_fn = torch.nn.SmoothL1Loss(reduction=reduction_type) elif loss_fn == LossFunction.L1Loss: torch_fn = torch.nn.L1Loss(reduction=reduction_type) elif loss_fn == LossFunction.BCELoss: torch_fn = torch.nn.BCELoss(reduction=reduction_type) def wrapper_loss_fn(pred, target, weight): loss = torch_fn(pred, target) if weighted_by_inverse_propensity: assert weight.shape == loss.shape loss = loss * weight # ignore abnormal reward only during training if pred.requires_grad and reward_ignore_threshold is not None: loss = loss[target <= reward_ignore_threshold] assert len(loss) > 0, ( f"reward ignore threshold set too small. target={target}, " f"threshold={reward_ignore_threshold}" ) return torch.mean(loss) return wrapper_loss_fn class RewardNetTrainer(ReAgentLightningModule): def __init__( self, reward_net: ModelBase, optimizer: Optimizer__Union = field( # noqa: B008 default_factory=Optimizer__Union.default ), loss_type: LossFunction = LossFunction.MSE, reward_ignore_threshold: Optional[float] = None, weighted_by_inverse_propensity: bool = False, ) -> None: super().__init__() self.reward_net = reward_net self.optimizer = optimizer self.loss_type = loss_type self.reward_ignore_threshold = reward_ignore_threshold self.weighted_by_inverse_propensity = weighted_by_inverse_propensity self.loss_fn = _get_loss_function( loss_type, reward_ignore_threshold, weighted_by_inverse_propensity ) def configure_optimizers(self): optimizers = [] optimizers.append( self.optimizer.make_optimizer_scheduler(self.reward_net.parameters()) ) return optimizers def _get_sample_weight(self, batch: rlt.PreprocessedRankingInput): weight = None if self.weighted_by_inverse_propensity: if isinstance(batch, rlt.PreprocessedRankingInput): assert batch.tgt_out_probs is not None weight = 1.0 / batch.tgt_out_probs else: raise NotImplementedError( f"Sampling weighting not implemented for {type(batch)}" ) return weight def _get_target_reward(self, batch: rlt.PreprocessedRankingInput): if isinstance(batch, rlt.PreprocessedRankingInput): target_reward = batch.slate_reward else: target_reward = batch.reward assert target_reward is not None return target_reward # pyre-fixme[56]: Decorator `torch.no_grad(...)` could not be called, because # its type `no_grad` is not callable. @torch.no_grad() def _compute_unweighted_loss( self, predicted_reward: torch.Tensor, target_reward: torch.Tensor ): return self.loss_fn( predicted_reward, target_reward, weight=torch.ones_like(predicted_reward) ) def train_step_gen( self, training_batch: rlt.PreprocessedRankingInput, batch_idx: int ): weight = self._get_sample_weight(training_batch) target_reward = self._get_target_reward(training_batch) predicted_reward = self.reward_net(training_batch).predicted_reward assert ( predicted_reward.shape == target_reward.shape and len(target_reward.shape) == 2 and target_reward.shape[1] == 1 ) loss = self.loss_fn(predicted_reward, target_reward, weight) detached_loss = loss.detach().cpu() self.reporter.log(loss=detached_loss) if weight is not None: unweighted_loss = self._compute_unweighted_loss( predicted_reward, target_reward ) self.reporter.log(unweighted_loss=unweighted_loss) if self.all_batches_processed % 10 == 0: logger.info( f"{self.all_batches_processed}-th batch: " f"{self.loss_type}={detached_loss.item()}" ) yield loss # pyre-ignore inconsistent override because lightning doesn't use types def validation_step(self, batch: rlt.PreprocessedRankingInput, batch_idx: int): reward = self._get_target_reward(batch) self.reporter.log(eval_rewards=reward.flatten().detach().cpu()) pred_reward = self.reward_net(batch).predicted_reward self.reporter.log(eval_pred_rewards=pred_reward.flatten().detach().cpu()) weight = self._get_sample_weight(batch) loss = self.loss_fn(pred_reward, reward, weight) detached_loss = loss.detach().cpu() self.reporter.log(eval_loss=detached_loss) if weight is not None: unweighted_loss = self._compute_unweighted_loss(pred_reward, reward) self.reporter.log(eval_unweighted_loss=unweighted_loss) return detached_loss.item() def validation_epoch_end(self, outputs): self.reporter.update_best_model(np.mean(outputs), self.reward_net) def warm_start_components(self): return ["reward_net"]
6,165
34.034091
85
py
ReAgent
ReAgent-master/reagent/training/dqn_trainer.py
#!/usr/bin/env python3 # Copyright (c) Facebook, Inc. and its affiliates. All rights reserved. import logging from typing import List, Optional, Tuple import reagent.core.types as rlt import torch from reagent.core.configuration import resolve_defaults from reagent.core.dataclasses import dataclass, field from reagent.core.parameters import EvaluationParameters, RLParameters from reagent.optimizer import Optimizer__Union, SoftUpdate from reagent.training.dqn_trainer_base import DQNTrainerBaseLightning from reagent.training.imitator_training import get_valid_actions_from_imitator logger = logging.getLogger(__name__) @dataclass(frozen=True) class BCQConfig: # 0 = max q-learning, 1 = imitation learning drop_threshold: float = 0.1 class DQNTrainer(DQNTrainerBaseLightning): @resolve_defaults def __init__( self, q_network, q_network_target, reward_network, q_network_cpe=None, q_network_cpe_target=None, metrics_to_score=None, evaluation: EvaluationParameters = field( # noqa: B008 default_factory=EvaluationParameters ), imitator=None, # Start DQNTrainerParameters. All parameters above should be # in the blocklist for DQNTrainerParameters in parameters.py actions: List[str] = field(default_factory=list), # noqa: B008 rl: RLParameters = field(default_factory=RLParameters), # noqa: B008 double_q_learning: bool = True, bcq: Optional[BCQConfig] = None, minibatch_size: int = 1024, minibatches_per_step: int = 1, optimizer: Optimizer__Union = field( # noqa: B008 default_factory=Optimizer__Union.default ), ) -> None: """ Args: q_network: states -> q-value for each action q_network_target: copy of q-network for training stability reward_network: states -> reward for each action q_network_cpe: q_network_cpe_target: metrics_to_score: imitator (optional): The behavior policy, used for BCQ training actions: list of action names rl: RLParameters double_q_learning: boolean flag to use double-q learning bcq: a config file for batch-constrained q-learning, defaults to normal minibatch_size: samples per minibatch minibatches_per_step: minibatch updates per step optimizer: q-network optimizer evaluation: evaluation params, primarily whether to use CPE in eval or not """ super().__init__( rl, metrics_to_score=metrics_to_score, actions=actions, evaluation_parameters=evaluation, ) assert self._actions is not None, "Discrete-action DQN needs action names" self.double_q_learning = double_q_learning self.minibatch_size = minibatch_size self.minibatches_per_step = minibatches_per_step or 1 self.q_network = q_network self.q_network_target = q_network_target self.q_network_optimizer = optimizer self._initialize_cpe( reward_network, q_network_cpe, q_network_cpe_target, optimizer=optimizer ) reward_boosts = torch.zeros([1, len(self._actions)]) if rl.reward_boost is not None: # pyre-fixme[16]: `Optional` has no attribute `keys`. for k in rl.reward_boost.keys(): i = self._actions.index(k) # pyre-fixme[16]: `Optional` has no attribute `__getitem__`. reward_boosts[0, i] = rl.reward_boost[k] self.register_buffer("reward_boosts", reward_boosts) # Batch constrained q-learning self.bcq = bcq is not None if self.bcq: assert bcq is not None self.bcq_drop_threshold = bcq.drop_threshold self.bcq_imitator = imitator def configure_optimizers(self): optimizers = [] target_params = list(self.q_network_target.parameters()) source_params = list(self.q_network.parameters()) optimizers.append( self.q_network_optimizer.make_optimizer_scheduler( self.q_network.parameters() ) ) if self.calc_cpe_in_training: ( cpe_target_params, cpe_source_params, cpe_optimizers, ) = self._configure_cpe_optimizers() target_params += cpe_target_params source_params += cpe_source_params optimizers += cpe_optimizers optimizers.append( SoftUpdate.make_optimizer_scheduler( target_params, source_params, tau=self.tau ) ) return optimizers # pyre-fixme[56]: Decorator `torch.no_grad(...)` could not be called, because # its type `no_grad` is not callable. @torch.no_grad() def get_detached_model_outputs( self, state ) -> Tuple[torch.Tensor, Optional[torch.Tensor]]: """Gets the q values from the model and target networks""" q_values = self.q_network(state) q_values_target = self.q_network_target(state) return q_values, q_values_target def compute_discount_tensor( self, batch: rlt.DiscreteDqnInput, boosted_rewards: torch.Tensor ): discount_tensor = torch.full_like(boosted_rewards, self.gamma) if self.use_seq_num_diff_as_time_diff: assert self.multi_steps is None discount_tensor = torch.pow(self.gamma, batch.time_diff.float()) if self.multi_steps is not None: assert batch.step is not None discount_tensor = torch.pow(self.gamma, batch.step.float()) return discount_tensor def compute_td_loss( self, batch: rlt.DiscreteDqnInput, boosted_rewards: torch.Tensor, discount_tensor: torch.Tensor, ): not_done_mask = batch.not_terminal.float() all_next_q_values, all_next_q_values_target = self.get_detached_model_outputs( batch.next_state ) if self.maxq_learning: # Compute max a' Q(s', a') over all possible actions using target network possible_next_actions_mask = batch.possible_next_actions_mask.float() if self.bcq: action_on_policy = get_valid_actions_from_imitator( self.bcq_imitator, batch.next_state, self.bcq_drop_threshold, ) possible_next_actions_mask *= action_on_policy next_q_values, max_q_action_idxs = self.get_max_q_values_with_target( all_next_q_values, all_next_q_values_target, possible_next_actions_mask, ) else: # SARSA next_q_values, max_q_action_idxs = self.get_max_q_values_with_target( all_next_q_values, all_next_q_values_target, batch.next_action, ) filtered_next_q_vals = next_q_values * not_done_mask target_q_values = boosted_rewards + (discount_tensor * filtered_next_q_vals) # Get Q-value of action taken all_q_values = self.q_network(batch.state) # pyre-fixme[16]: `DQNTrainer` has no attribute `all_action_scores`. self.all_action_scores = all_q_values.detach() q_values = torch.sum(all_q_values * batch.action, 1, keepdim=True) td_loss = self.q_network_loss(q_values, target_q_values.detach()) return td_loss def train_step_gen(self, training_batch: rlt.DiscreteDqnInput, batch_idx: int): # TODO: calls to _maybe_run_optimizer removed, should be replaced with Trainer parameter self._check_input(training_batch) rewards = self.boost_rewards(training_batch.reward, training_batch.action) not_done_mask = training_batch.not_terminal.float() discount_tensor = self.compute_discount_tensor(training_batch, rewards) td_loss = self.compute_td_loss(training_batch, rewards, discount_tensor) yield td_loss td_loss = td_loss.detach() # Get Q-values of next states, used in computing cpe all_next_action_scores = self.q_network(training_batch.next_state).detach() logged_action_idxs = torch.argmax(training_batch.action, dim=1, keepdim=True) yield from self._calculate_cpes( training_batch, training_batch.state, training_batch.next_state, self.all_action_scores, all_next_action_scores, logged_action_idxs, discount_tensor, not_done_mask, ) if self.maxq_learning: possible_actions_mask = training_batch.possible_actions_mask if self.bcq: action_on_policy = get_valid_actions_from_imitator( self.bcq_imitator, training_batch.state, self.bcq_drop_threshold ) possible_actions_mask *= action_on_policy # Do we ever use model_action_idxs computed below? model_action_idxs = self.get_max_q_values( self.all_action_scores, possible_actions_mask if self.maxq_learning else training_batch.action, )[1] self._log_dqn( td_loss, logged_action_idxs, training_batch, rewards, model_action_idxs ) # Use the soft update rule to update target network yield self.soft_update_result() def _log_dqn( self, td_loss, logged_action_idxs, training_batch, rewards, model_action_idxs ): self.reporter.log( td_loss=td_loss, logged_actions=logged_action_idxs, logged_propensities=training_batch.extras.action_probability, logged_rewards=rewards, logged_values=None, # Compute at end of each epoch for CPE model_values=self.all_action_scores, model_values_on_logged_actions=None, # Compute at end of each epoch for CPE model_action_idxs=model_action_idxs, ) model_values = self._dense_to_action_dict(self.all_action_scores.mean(dim=0)) action_histogram = self._dense_to_action_dict( training_batch.action.float().mean(dim=0) ) if training_batch.extras.action_probability is None: logged_propensities = None else: logged_propensities = training_batch.extras.action_probability.mean(dim=0) model_action_idxs = self._dense_to_action_dict( torch.nn.functional.one_hot( model_action_idxs.squeeze(1), num_classes=self.num_actions ) .float() .mean(dim=0) ) self.logger.log_metrics( { "td_loss": td_loss, "logged_actions": action_histogram, "logged_propensities": logged_propensities, "logged_rewards": rewards.mean(), "model_values": model_values, "model_action_idxs": model_action_idxs, }, step=self.all_batches_processed, ) def _dense_to_action_dict(self, dense: torch.Tensor): assert dense.size() == ( self.num_actions, ), f"Invalid dense size {dense.size()} != {(self.num_actions,)}" retval = {} for i, a in enumerate(self._actions): retval[a] = dense[i] return retval def validation_step(self, batch, batch_idx): if isinstance(batch, dict): batch = rlt.DiscreteDqnInput.from_dict(batch) rewards = self.boost_rewards(batch.reward, batch.action) discount_tensor = self.compute_discount_tensor(batch, rewards) td_loss = self.compute_td_loss(batch, rewards, discount_tensor) # Show eval_td_loss in a tensorboard graph self.log("eval_td_loss", td_loss) return super().validation_step(batch, batch_idx)
12,069
37.935484
96
py
ReAgent
ReAgent-master/reagent/training/td3_trainer.py
#!/usr/bin/env python3 # Copyright (c) Facebook, Inc. and its affiliates. All rights reserved. import copy import logging import reagent.core.types as rlt import torch import torch.nn.functional as F from reagent.core.configuration import resolve_defaults from reagent.core.dataclasses import field from reagent.core.parameters import CONTINUOUS_TRAINING_ACTION_RANGE, RLParameters from reagent.optimizer import Optimizer__Union, SoftUpdate from reagent.training.reagent_lightning_module import ReAgentLightningModule from reagent.training.rl_trainer_pytorch import RLTrainerMixin logger = logging.getLogger(__name__) class TD3Trainer(RLTrainerMixin, ReAgentLightningModule): """ Twin Delayed Deep Deterministic Policy Gradient algorithm trainer as described in https://arxiv.org/pdf/1802.09477 """ @resolve_defaults def __init__( self, actor_network, q1_network, q2_network=None, # Start TD3TrainerParameters rl: RLParameters = field(default_factory=RLParameters), # noqa: B008 q_network_optimizer: Optimizer__Union = field( # noqa: B008 default_factory=Optimizer__Union.default ), actor_network_optimizer: Optimizer__Union = field( # noqa: B008 default_factory=Optimizer__Union.default ), minibatch_size: int = 64, noise_variance: float = 0.2, noise_clip: float = 0.5, delayed_policy_update: int = 2, minibatches_per_step: int = 1, ) -> None: """ Args: actor_network: states -> actions, trained to maximize value q1_network: states, action -> q-value q2_network (optional): double q-learning to stabilize training from overestimation bias rl (optional): an instance of the RLParameter class, which defines relevant hyperparameters q_network_optimizer (optional): the optimizer class and optimizer hyperparameters for the q network(s) optimizer actor_network_optimizer (optional): see q_network_optimizer minibatch_size (optional): the size of the minibatch noise_variance (optional): the variance of action noise added to smooth q-value estimates noise_clip (optional): the maximum absolute value of action noise added to smooth q-value estimates delayed_policy_update (optional): the ratio of q network updates to target and policy network updates minibatches_per_step (optional, TODO: currently unused): the number of minibatch updates per training step """ super().__init__() self.rl_parameters = rl self.minibatch_size = minibatch_size self.minibatches_per_step = minibatches_per_step or 1 self.q1_network = q1_network self.q1_network_target = copy.deepcopy(self.q1_network) self.q_network_optimizer = q_network_optimizer self.q2_network = q2_network if self.q2_network is not None: self.q2_network_target = copy.deepcopy(self.q2_network) self.actor_network = actor_network self.actor_network_target = copy.deepcopy(self.actor_network) self.actor_network_optimizer = actor_network_optimizer self.noise_variance = noise_variance self.noise_clip_range = (-noise_clip, noise_clip) self.delayed_policy_update = delayed_policy_update def configure_optimizers(self): optimizers = [] optimizers.append( self.q_network_optimizer.make_optimizer_scheduler( self.q1_network.parameters() ) ) if self.q2_network: optimizers.append( self.q_network_optimizer.make_optimizer_scheduler( self.q2_network.parameters() ) ) optimizers.append( self.actor_network_optimizer.make_optimizer_scheduler( self.actor_network.parameters() ) ) # soft-update target_params = list(self.q1_network_target.parameters()) source_params = list(self.q1_network.parameters()) if self.q2_network: target_params += list(self.q2_network_target.parameters()) source_params += list(self.q2_network.parameters()) target_params += list(self.actor_network_target.parameters()) source_params += list(self.actor_network.parameters()) optimizers.append( SoftUpdate.make_optimizer_scheduler( target_params, source_params, tau=self.tau ) ) return optimizers def train_step_gen(self, training_batch: rlt.PolicyNetworkInput, batch_idx: int): """ IMPORTANT: the input action here is assumed to be preprocessed to match the range of the output of the actor. """ assert isinstance(training_batch, rlt.PolicyNetworkInput) state = training_batch.state action = training_batch.action next_state = training_batch.next_state reward = training_batch.reward not_terminal = training_batch.not_terminal # Generate target = r + y * min (Q1(s',pi(s')), Q2(s',pi(s'))) with torch.no_grad(): next_actor = self.actor_network_target(next_state).action noise = torch.randn_like(next_actor) * self.noise_variance next_actor = (next_actor + noise.clamp(*self.noise_clip_range)).clamp( *CONTINUOUS_TRAINING_ACTION_RANGE ) next_state_actor = (next_state, rlt.FeatureData(next_actor)) next_q_value = self.q1_network_target(*next_state_actor) if self.q2_network is not None: next_q_value = torch.min( next_q_value, self.q2_network_target(*next_state_actor) ) target_q_value = reward + self.gamma * next_q_value * not_terminal.float() # Optimize Q1 and Q2 q1_value = self.q1_network(state, action) q1_loss = F.mse_loss(q1_value, target_q_value) if batch_idx % self.trainer.log_every_n_steps == 0: self.reporter.log( q1_loss=q1_loss, q1_value=q1_value, next_q_value=next_q_value, target_q_value=target_q_value, ) self.log("td_loss", q1_loss, prog_bar=True) yield q1_loss if self.q2_network: q2_value = self.q2_network(state, action) q2_loss = F.mse_loss(q2_value, target_q_value) if batch_idx % self.trainer.log_every_n_steps == 0: self.reporter.log( q2_loss=q2_loss, q2_value=q2_value, ) yield q2_loss # Only update actor and target networks after a fixed number of Q updates if batch_idx % self.delayed_policy_update == 0: actor_action = self.actor_network(state).action actor_q1_value = self.q1_network(state, rlt.FeatureData(actor_action)) actor_loss = -(actor_q1_value.mean()) if batch_idx % self.trainer.log_every_n_steps == 0: self.reporter.log( actor_loss=actor_loss, actor_q1_value=actor_q1_value, ) yield actor_loss # Use the soft update rule to update the target networks result = self.soft_update_result() yield result else: # Yielding None prevents the actor and target networks from updating yield None yield None
7,751
38.350254
100
py
ReAgent
ReAgent-master/reagent/training/discrete_crr_trainer.py
#!/usr/bin/env python3 # Copyright (c) Facebook, Inc. and its affiliates. All rights reserved. # Note: this files is modeled after td3_trainer.py import logging from typing import List, Tuple import reagent.core.types as rlt import torch import torch.nn.functional as F from reagent.core.configuration import resolve_defaults from reagent.core.dataclasses import field from reagent.core.parameters import EvaluationParameters, RLParameters from reagent.optimizer import Optimizer__Union, SoftUpdate from reagent.training.dqn_trainer_base import DQNTrainerBaseLightning from torch import distributions as pyd logger = logging.getLogger(__name__) class DiscreteCRRTrainer(DQNTrainerBaseLightning): """ Critic Regularized Regression (CRR) algorithm trainer as described in https://arxiv.org/abs/2006.15134 """ @resolve_defaults def __init__( self, actor_network, actor_network_target, q1_network, q1_network_target, reward_network, q2_network=None, q2_network_target=None, q_network_cpe=None, q_network_cpe_target=None, metrics_to_score=None, evaluation: EvaluationParameters = field( # noqa: B008 default_factory=EvaluationParameters ), # Start CRRTrainerParameters. All parameters above should be # in the blocklist for CRRTrainerParameters in parameters.py rl: RLParameters = field(default_factory=RLParameters), # noqa: B008 double_q_learning: bool = True, q_network_optimizer: Optimizer__Union = field( # noqa: B008 default_factory=Optimizer__Union.default ), actor_network_optimizer: Optimizer__Union = field( # noqa: B008 default_factory=Optimizer__Union.default ), use_target_actor: bool = False, actions: List[str] = field(default_factory=list), # noqa: B008 delayed_policy_update: int = 1, beta: float = 1.0, entropy_coeff: float = 0.0, clip_limit: float = 10.0, max_weight: float = 20.0, ) -> None: """ Args: actor_network: states -> actions, trained to maximize value actor_network_target: copy of actor network for training stability q1_network: states -> q-value for all actions q1_network_target: copy of q-network for training stability q2_network (optional): double q-learning to stabilize training from overestimation bias. The presence of q2_network is specified in discrete_crr.py using the config parameter double_q_learning q2_network_target (optional): copy of q-network for training stability rl (optional): an instance of the RLParameter class, which defines relevant hyperparameters q_network_optimizer (optional): the optimizer class and optimizer hyperparameters for the q network(s) optimizer actor_network_optimizer (optional): see q_network_optimizer use_target_actor (optional): specifies whether target actor is used delayed_policy_update (optional): the ratio of q network updates to target and policy network updates beta: coefficient for KL-divergence policy constaint regularization of CRR see eq(5) in https://arxiv.org/pdf/2006.15134.pdf. With large beta, the output policy of CRR can not leaves too far away from the logged policy entropy_coeff: coefficient for entropy regularization clip_limit: threshold for importance sampling when compute entropy regularization using offline samples max_weight: the maximum possible action weight in the actor loss Explaination of entropy regularization: Entropy regularization punishes deterministic policy and encourages "unifom" policy. Entropy regularized MDP can be viewed as add the term (-entropy_coeff * pi_ratio * log_pi_b) to each reward. For detailed formulation of entropy regularized please see eq.(9) & eq.(10) in https://arxiv.org/pdf/2007.06558.pdf """ super().__init__( rl, metrics_to_score=metrics_to_score, actions=actions, evaluation_parameters=evaluation, ) self._actions = actions assert self._actions is not None, "Discrete-action CRR needs action names" self.rl_parameters = rl self.double_q_learning = double_q_learning self.use_target_actor = use_target_actor self.q1_network = q1_network self.q1_network_target = q1_network_target self.q_network_optimizer = q_network_optimizer self.q2_network = q2_network if self.q2_network is not None: assert ( q2_network_target is not None ), "q2_network provided without a target network" self.q2_network_target = q2_network_target self.actor_network = actor_network self.actor_network_target = actor_network_target self.actor_network_optimizer = actor_network_optimizer self.delayed_policy_update = delayed_policy_update self.register_buffer("reward_boosts", None) self.reward_boosts = torch.zeros([1, len(self._actions)]) if rl.reward_boost is not None: # pyre-fixme[16]: Optional type has no attribute `keys`. for k in rl.reward_boost.keys(): i = self._actions.index(k) # pyre-fixme[16]: Optional type has no attribute `__getitem__`. self.reward_boosts[0, i] = rl.reward_boost[k] self._initialize_cpe( reward_network, q_network_cpe, q_network_cpe_target, optimizer=q_network_optimizer, ) self.beta = beta self.entropy_coeff = entropy_coeff self.clip_limit = clip_limit self.max_weight = max_weight @property def q_network(self): return self.q1_network # pyre-fixme[56]: Decorator `torch.no_grad(...)` could not be called, because # its type `no_grad` is not callable. @torch.no_grad() def get_detached_model_outputs(self, state) -> Tuple[torch.Tensor, None]: # This function is only used in evaluation_data_page.py, in create_from_tensors_dqn(), # in order to compute model propensities. The definition of this function in # dqn_trainer.py returns two values, and so we also return two values here, for # consistency. action_scores = self.actor_network(state).action return action_scores, None def configure_optimizers(self): optimizers = [] target_params = list(self.q1_network_target.parameters()) source_params = list(self.q1_network.parameters()) optimizers.append( self.q_network_optimizer.make_optimizer_scheduler( self.q1_network.parameters() ) ) if self.q2_network: target_params += list(self.q2_network_target.parameters()) source_params += list(self.q2_network.parameters()) optimizers.append( self.q_network_optimizer.make_optimizer_scheduler( self.q2_network.parameters() ) ) target_params += list(self.actor_network_target.parameters()) source_params += list(self.actor_network.parameters()) optimizers.append( self.actor_network_optimizer.make_optimizer_scheduler( self.actor_network.parameters() ) ) if self.calc_cpe_in_training: ( cpe_target_params, cpe_source_params, cpe_optimizers, ) = self._configure_cpe_optimizers() target_params += cpe_target_params source_params += cpe_source_params optimizers += cpe_optimizers optimizers.append( SoftUpdate.make_optimizer_scheduler( target_params, source_params, tau=self.tau ) ) return optimizers def compute_target_q_values(self, next_state, rewards, not_terminal, next_q_values): if self.use_target_actor: next_state_actor_output = self.actor_network_target(next_state).action else: next_state_actor_output = self.actor_network(next_state).action next_dist = pyd.Categorical(logits=next_state_actor_output) next_V = (next_q_values * next_dist.probs).sum(dim=1, keepdim=True) if self.q2_network is not None: next_q2_values = self.q2_network_target(next_state) next_V2 = (next_q2_values * next_dist.probs).sum(dim=1, keepdim=True) next_V = torch.min(next_V, next_V2) target_q_values = rewards + self.gamma * next_V * not_terminal.float() return target_q_values def compute_td_loss(self, q_network, state, action, target_q_values): all_q_values = q_network(state) q_values = (all_q_values * action).sum(dim=1, keepdim=True) q_loss = F.mse_loss(q_values, target_q_values) return q_loss def compute_actor_loss( self, batch_idx, action, logged_action_probs, all_q_values, all_action_scores ): # Only update actor network after a fixed number of Q updates if batch_idx % self.delayed_policy_update != 0: # Yielding None prevents the actor network from updating actor_loss = None return (actor_loss, actor_loss) # dist is the distribution of actions derived from the actor's outputs (logits) dist = pyd.Categorical(logits=all_action_scores) # Note: D = dist.probs is equivalent to: # e_x = torch.exp(actor_actions) # D = e_x / e_x.sum(dim=1, keepdim=True) # That is, dist gives a softmax distribution over actor's outputs # values is the vector of state values in this batch values = (all_q_values * dist.probs).sum(dim=1, keepdim=True) advantages = all_q_values - values # Note: the above statement subtracts the "values" column vector from # every column of the all_q_values matrix, giving us the advantages # of every action in the present state weight = torch.clamp( ((1 / self.beta) * (advantages * action).sum(dim=1, keepdim=True)).exp(), 0, self.max_weight, ) # Remember: training_batch.action is in the one-hot format logged_action_idxs = torch.argmax(action, dim=1, keepdim=True) # Note: action space is assumed to be discrete with actions # belonging to the set {0, 1, ..., action_dim-1}. Therefore, # advantages.gather(1, logged_action_idxs) will select, for each data point # (row i of the Advantage matrix "advantages"), the element with index # action.float_features[i] # Note: dist.logits already gives log(p), which can be verified by # comparing dist.probs and dist.logits. # https://pytorch.org/docs/master/distributions.html#multinomial # states: logits (Tensor) – event log probabilities # log_pi_b is the log of the probability assigned by the # actor (abbreviated as pi) to the actions of the behavioral (b) policy log_pi_b = dist.log_prob(logged_action_idxs.squeeze(1)).unsqueeze(1) # entropy regularization pi_t = (dist.probs * action).sum(dim=1, keepdim=True) if self.entropy_coeff > 0: pi_b = logged_action_probs.view(pi_t.shape) assert torch.min(pi_b) > 0, "Logged action probability <= 0" pi_ratio = torch.clip(pi_t / pi_b, min=1e-4, max=self.clip_limit) entropy = (pi_ratio * log_pi_b).mean() else: # dummy value entropy = 0 # Note: the CRR loss for each datapoint (and the magnitude of the corresponding # parameter update) is proportional to log_pi_b * weight. Therefore, as mentioned # at the top of Section 3.2, the actor on the one hand has incentive to assign # larger probabilities to the actions observed in the dataset (so as to reduce # the magnitude of log_pi_b), but on the other hand it gives preference to doing # this on datapoints where weight is large (i.e., those points on which the # Q-value of the observed action is large). actor_loss_without_reg = (-log_pi_b * weight.detach()).mean() actor_loss = (-log_pi_b * weight.detach()).mean() + self.entropy_coeff * entropy return actor_loss_without_reg, actor_loss def train_step_gen(self, training_batch: rlt.DiscreteDqnInput, batch_idx: int): """ IMPORTANT: the input action here is preprocessed according to the training_batch type, which in this case is DiscreteDqnInput. Hence, the preprocessor in the DiscreteDqnInputMaker class in the trainer_preprocessor.py is used, which converts acion taken to a one-hot representation. """ self._check_input(training_batch) state = training_batch.state action = training_batch.action next_state = training_batch.next_state not_terminal = training_batch.not_terminal rewards = self.boost_rewards(training_batch.reward, training_batch.action) # Remember: training_batch.action is in the one-hot format logged_action_idxs = torch.argmax(action, dim=1, keepdim=True) discount_tensor = torch.full_like(rewards, self.gamma) next_q_values = self.q1_network_target(next_state) target_q_values = self.compute_target_q_values( next_state, rewards, not_terminal, next_q_values ) q1_loss = self.compute_td_loss(self.q1_network, state, action, target_q_values) # Show td_loss on the progress bar and in tensorboard graphs: self.log("td_loss", q1_loss, prog_bar=True) yield q1_loss if self.q2_network: q2_loss = self.compute_td_loss( self.q2_network, state, action, target_q_values ) yield q2_loss all_q_values = self.q1_network(state) # Q-values of all actions # Note: action_dim (the length of each row of the actor_action # matrix obtained below) is assumed to be > 1. all_action_scores = self.actor_network(state).action logged_action_probs = training_batch.extras.action_probability actor_loss_without_reg, actor_loss = self.compute_actor_loss( batch_idx, action, logged_action_probs, all_q_values, all_action_scores ) # self.reporter.log( # actor_loss=actor_loss, # actor_q1_value=actor_q1_values, # ) # Show actor_loss on the progress bar and also in Tensorboard graphs self.log("actor_loss_without_reg", actor_loss_without_reg, prog_bar=True) self.log("actor_loss", actor_loss, prog_bar=True) yield actor_loss yield from self._calculate_cpes( training_batch, state, next_state, all_action_scores, next_q_values.detach(), logged_action_idxs, discount_tensor, not_terminal.float(), ) # TODO: rename underlying function to get_max_possible_values_and_idxs model_action_idxs = self.get_max_q_values( all_action_scores, training_batch.possible_actions_mask if self.maxq_learning else action, )[1] self.reporter.log( logged_actions=logged_action_idxs, td_loss=q1_loss, logged_propensities=training_batch.extras.action_probability, logged_rewards=rewards, model_values=all_action_scores, model_action_idxs=model_action_idxs, ) # Use the soft update rule to update the target networks. # Note: this yield has to be the last one, since SoftUpdate is the last # optimizer added in the configure_optimizers() function. result = self.soft_update_result() yield result def validation_step(self, batch, batch_idx): # As explained in the comments to the validation_step function in # pytorch_lightning/core/lightning.py, this function operates on a # single batch of data from the validation set. For example: # val_outs = [] # for val_batch in val_data: # out = validation_step(val_batch) # val_outs.append(out) # validation_epoch_end(val_outs) # Note: the relevant validation_epoch_end() function is defined in dqn_trainer_base.py # RETURN ARGS: # The super() call at the end of this function calls the function with the same name # in dqn_trainer_base.py, which returns a EvaluationDataPage for data in that batch. # In other words, the validation_epoch_end() function will take a list of validation # EvaluationDataPages. if isinstance(batch, dict): batch = rlt.DiscreteDqnInput.from_dict(batch) # validation data state = batch.state action = batch.action next_state = batch.next_state not_terminal = batch.not_terminal rewards = self.boost_rewards(batch.reward, action) # intermediate values next_q_values = self.q1_network_target(next_state) target_q_values = self.compute_target_q_values( next_state, rewards, not_terminal, next_q_values ) all_q_values = self.q1_network(state) all_action_scores = self.actor_network(state).action logged_action_probs = batch.extras.action_probability # loss to log actor_loss_without_reg, actor_loss = self.compute_actor_loss( batch_idx, action, logged_action_probs, all_q_values, all_action_scores ) td_loss = self.compute_td_loss(self.q1_network, state, action, target_q_values) self.log("eval_actor_loss_without_reg", actor_loss_without_reg) self.log("eval_actor_loss", actor_loss) self.log("eval_td_loss", td_loss) return super().validation_step(batch, batch_idx)
18,427
41.266055
94
py
ReAgent
ReAgent-master/reagent/training/gradient_free/es_worker.py
#!/usr/bin/env python3 # Copyright (c) Facebook, Inc. and its affiliates. All rights reserved. import logging import torch import torch.distributed as distributed import torch.nn import torch.optim from reagent.core.parameters import EvolutionParameters from reagent.training.gradient_free.evolution_pool import EvolutionPool from torch.distributed import ProcessGroup logger = logging.getLogger(__name__) class EsWorker: def __init__( self, individual_pool: EvolutionPool, es_params: EvolutionParameters, process_group: ProcessGroup, num_nodes: int, ) -> None: logger.info("WORKER STARTED") self.iteration = 0 self.most_recent_avg_rewards = 0.0 self.individual_pool = individual_pool self.es_params = es_params self.process_group = process_group self.num_nodes = num_nodes def run_epoch(self) -> float: logger.info("Computing reward") rewards = self.individual_pool.compute_all_local_rewards() logger.info("Pushing reward") # Sum the rewards across all machines distributed.all_reduce(rewards, self.process_group) # Divide the rewards by the number of machines. We do this because # there is no "average" all_reduce operator. rewards /= self.num_nodes self.iteration += 1 self.individual_pool.apply_global_reward(rewards, self.iteration) most_recent_avg_rewards = float(torch.mean(rewards)) new_parent_reward = self.individual_pool.compute_local_reward( self.individual_pool.parent_tensors ) logger.info( "ITERATION: {0} MEAN REWARD: {1}, NEW PARENT REWARD: {2}".format( self.iteration, most_recent_avg_rewards, new_parent_reward ) ) return new_parent_reward
1,861
30.033333
77
py
ReAgent
ReAgent-master/reagent/training/gradient_free/ars_util.py
from operator import itemgetter import numpy as np import torch """ Utility functions for Advanced Random Search (ARS) algorithm based on the paper "Simple random search provides a competitive approach to reinforcement learning", Mania et al. https://arxiv.org/abs/1803.07055 Here, we show an example of training a data reweighting policy using ARS. The policy is learned to weight each sample for training a supervised learning model. ARS is a competitive alternative to the policy gradient method in "Data Valuation using Reinforcement Learning", Yoon, Arik, and Pfister. https://arxiv.org/abs/1909.11671 def reward_func(pos_param, neg_param): # Return rewards for positively/negatively perturbed parameters # model = a supervised learning model # X = training features # y = labels # Initialize a supervised learning model model_pos = model.init() # Sample weights are bounded within (0, 1) pos_weight = torch.sigmoid(torch.matmul(torch.column_stack((X, y)), pos_param)) model_pos.fit(X, y, sample_weight=pos_weight) r_pos = metric(model_pos.predict(X_e), y_e) model_neg = model.init() neg_weight = torch.sigmoid(torch.matmul(torch.column_stack((X, y)), neg_param)) model_neg.fit(X, y, sample_weight=neg_weight) r_neg = metric(model_neg.predict(X_e), y_e) return (r_pos, r_neg) # Training # feature_dim = feature dimension + 1 (for label) # n_pert = given number of random perturbations # alpha = step size # noise = noise level (between 0 ~ 1) added to the random perturbations ars_opt = ARSOptimizer(feature_dim, n_pert, alpha=alpha, noise=noise) for _ in range(n_generations): perturbed_params = ars_opt.sample_perturbed_params() rewards = [] for idx in range(0, len(perturbed_params)): pos_param, neg_param = params[idx] rewards.extend(reward_func(pos_param, neg_param)) ars_opt.update_ars_params(rewards) """ class ARSOptimizer: """ARSOptimizer is supposed to maximize an objective function""" def __init__( self, feature_dim, n_pert=10, rand_ars_params=False, alpha=1, noise=1, b_top=None, ): self.feature_dim = feature_dim self.ars_params = ( np.random.randn(feature_dim) if rand_ars_params else np.zeros(feature_dim) ) self.alpha = alpha self.noise = noise self.n_pert = n_pert self.b_top = b_top if b_top is not None else n_pert self.perturbations = [] def update_ars_params(self, rewards: torch.Tensor): """ reward should be something like [reward_pert1_pos, reward_pert1_neg, reward_pert2_pos, reward_pert2_neg, ...] """ assert ( len(self.perturbations) > 0 ), "must call sample_perturbed_params before this function" assert rewards.shape == ( 2 * self.n_pert, ), "rewards must have length 2 * n_pert" rank = {} rewards = rewards.numpy() for pert_idx in range(self.n_pert): reward_pos = rewards[2 * pert_idx] reward_neg = rewards[2 * pert_idx + 1] rank[pert_idx] = max(reward_pos, reward_neg) self.perturbations[pert_idx] *= reward_pos - reward_neg std_r = np.std(rewards) weight_sum = 0 for pert_idx in list( dict(sorted(rank.items(), key=itemgetter(1), reverse=True)).keys() )[: self.b_top]: weight_sum += self.perturbations[pert_idx] self.ars_params = self.ars_params + self.alpha * weight_sum / ( self.b_top * (std_r if std_r > 0 else 1) ) self.perturbations = [] def sample_perturbed_params(self): """Return tuples of (pos_param, neg_param)""" self.perturbations = [] perturbed_params = [] for _ in range(self.n_pert): pert = np.random.randn(self.feature_dim) self.perturbations.append(pert) perturbed_params.append( ( torch.from_numpy(self.ars_params + self.noise * pert).float(), torch.from_numpy(self.ars_params - self.noise * pert).float(), ) ) return perturbed_params
4,380
34.909836
87
py
ReAgent
ReAgent-master/reagent/training/gradient_free/evolution_pool.py
#!/usr/bin/env python3 # Copyright (c) Facebook, Inc. and its affiliates. All rights reserved. import logging from typing import Dict, List import torch import torch.fb.rendezvous.zeus import torch.nn import torch.optim from reagent.core.parameters import EvolutionParameters logger = logging.getLogger(__name__) MAX_RNG_SEED = 100000 class EvolutionPool: """ Handles spawning new individuals from a parent, computing an estimated gradient, and applying that gradient to mutate the parent. """ def __init__( self, seed: int, es_params: EvolutionParameters, tensor_sizes: Dict[str, List[int]], ) -> None: self.es_params = es_params self.tensor_sizes = tensor_sizes self.seed = seed assert self.seed < MAX_RNG_SEED, "The random seed must be less than " + str( MAX_RNG_SEED ) logger.info("Starting pool with RNG seed: " + str(self.seed)) # Fill the population with empty values: will populate later self.population_tensors: List[Dict[str, torch.Tensor]] = [] for _ in range(es_params.population_size): individual = {} for tensor_name, tensor_size in self.tensor_sizes.items(): individual[tensor_name] = torch.zeros(tensor_size, dtype=torch.float) self.population_tensors.append(individual) torch.manual_seed(self.seed) self.parent_tensors: Dict[str, torch.Tensor] = {} for tensor_name, tensor_size in self.tensor_sizes.items(): self.parent_tensors[tensor_name] = torch.randn( tensor_size, dtype=torch.float ) # pyre-fixme[41]: `grad` cannot be reassigned. It is a read-only property. self.parent_tensors[tensor_name].grad = torch.randn( tensor_size, dtype=torch.float ) self.optimizer = torch.optim.Adam( self.parent_tensors.values(), lr=self.es_params.learning_rate ) self.populate_children(0) def populate_children(self, iteration: int): torch.manual_seed(iteration * MAX_RNG_SEED + self.seed) for individual in self.population_tensors: for tensor_name, parent_tensor in self.parent_tensors.items(): individual_tensor = individual[tensor_name] individual_tensor.normal_(0, self.es_params.mutation_power) individual_tensor.add_(parent_tensor) def apply_global_reward(self, rewards: torch.Tensor, next_iteration: int): std_dev = torch.std(rewards) if torch.abs(std_dev) > 1e-6: normalized_rewards = (rewards - torch.mean(rewards)) / std_dev for parent_tensor in self.parent_tensors.values(): parent_tensor.grad.zero_() for i, individual in enumerate(self.population_tensors): for tensor_name, parent_tensor in self.parent_tensors.items(): individual_tensor = individual[tensor_name] # Subtract the parent to get the gradient estimate # pyre-fixme[16]: `Tensor` has no attribute `sub_`. individual_tensor.sub_(parent_tensor) # Amplify the gradient by the reward individual_tensor.mul_(normalized_rewards[i]) # Divide by a normalizing constant individual_tensor.div_( self.es_params.population_size * self.es_params.mutation_power * -1 ) # pyre-fixme[41]: `grad` cannot be reassigned. It is a read-only # property. parent_tensor.grad += individual_tensor self.optimizer.step() self.populate_children(next_iteration) def compute_all_local_rewards(self): return torch.tensor( [ self.compute_local_reward(individual) for individual in self.population_tensors ], dtype=torch.float, ) def compute_local_reward(self, individual): """ Given an individual as a list of tensors, return the reward of this policy """ raise NotImplementedError() class OneMaxEvolutionPool(EvolutionPool): """ A simple example of an evolution pool. The agent gets maximum reward as the tensor approaches [inf, -inf, inf, -inf, ...] """ def compute_local_reward(self, individual): sigmoid_params = torch.nn.Sigmoid()(individual["data"]) total_reward = torch.sum(sigmoid_params[0::2]) + torch.sum( 1 - sigmoid_params[1::2] ) return total_reward / sigmoid_params.shape[0]
4,802
35.112782
86
py