INSTRUCTION
stringlengths
1
46.3k
RESPONSE
stringlengths
75
80.2k
Perform update of param_arrays from grad_arrays on NCCL kvstore.
def _update_params_on_kvstore_nccl(param_arrays, grad_arrays, kvstore, param_names): """Perform update of param_arrays from grad_arrays on NCCL kvstore.""" valid_indices = [index for index, grad_list in enumerate(grad_arrays) if grad_list[0] is not None] valid_grad_arrays = [grad_arrays[i] for i in valid_indices] valid_param_arrays = [param_arrays[i] for i in valid_indices] valid_param_names = [param_names[i] for i in valid_indices] size = len(valid_grad_arrays) start = 0 # Use aggregation by default only with NCCL default_batch = '16' batch = int(os.getenv('MXNET_UPDATE_AGGREGATION_SIZE', default_batch)) while start < size: end = start + batch if start + batch < size else size # push gradient, priority is negative index kvstore.push(valid_param_names[start:end], valid_grad_arrays[start:end], priority=-start) # pull back the weights kvstore.pull(valid_param_names[start:end], valid_param_arrays[start:end], priority=-start) start = end
Perform update of param_arrays from grad_arrays on kvstore.
def _update_params_on_kvstore(param_arrays, grad_arrays, kvstore, param_names): """Perform update of param_arrays from grad_arrays on kvstore.""" for index, pair in enumerate(zip(param_arrays, grad_arrays)): arg_list, grad_list = pair if grad_list[0] is None: continue name = param_names[index] # push gradient, priority is negative index kvstore.push(name, grad_list, priority=-index) # pull back the weights kvstore.pull(name, arg_list, priority=-index)
Perform update of param_arrays from grad_arrays not on kvstore.
def _update_params(param_arrays, grad_arrays, updater, num_device, kvstore=None, param_names=None): """Perform update of param_arrays from grad_arrays not on kvstore.""" updates = [[] for _ in range(num_device)] for i, pair in enumerate(zip(param_arrays, grad_arrays)): arg_list, grad_list = pair if grad_list[0] is None: continue index = i if kvstore: name = param_names[index] # push gradient, priority is negative index kvstore.push(name, grad_list, priority=-index) # pull back the sum gradients, to the same locations. kvstore.pull(name, grad_list, priority=-index) for k, p in enumerate(zip(arg_list, grad_list)): # faked an index here, to make optimizer create diff # state for the same index but on diff devs, TODO(mli) # use a better solution later w, g = p updates[k].append((index*num_device+k, g, w)) for dev_updates in updates: # update params if param_arrays and grad_arrays are not empty if dev_updates: i, w, g = zip(*dev_updates) updater(i, w, g)
Sends args and kwargs to any configured callbacks. This handles the cases where the 'callbacks' variable is ``None``, a single function, or a list.
def _multiple_callbacks(callbacks, *args, **kwargs): """Sends args and kwargs to any configured callbacks. This handles the cases where the 'callbacks' variable is ``None``, a single function, or a list. """ if isinstance(callbacks, list): for cb in callbacks: cb(*args, **kwargs) return if callbacks: callbacks(*args, **kwargs)
Internal training function on multiple devices. This function will also work for single device as well. Parameters ---------- symbol : Symbol The network configuration. ctx : list of Context The training devices. arg_names: list of str Name of all arguments of the network. param_names: list of str Name of all trainable parameters of the network. aux_names: list of str Name of all auxiliary states of the network. arg_params : dict of str to NDArray Model parameter, dict of name to NDArray of net's weights. aux_params : dict of str to NDArray Model parameter, dict of name to NDArray of net's auxiliary states. begin_epoch : int The begining training epoch. end_epoch : int The end training epoch. epoch_size : int, optional Number of batches in a epoch. In default, it is set to ``ceil(num_train_examples / batch_size)``. optimizer : Optimizer The optimization algorithm train_data : DataIter Training data iterator. eval_data : DataIter Validation data iterator. eval_metric : EvalMetric An evaluation function or a list of evaluation functions. epoch_end_callback : callable(epoch, symbol, arg_params, aux_states) A callback that is invoked at end of each epoch. This can be used to checkpoint model each epoch. batch_end_callback : callable(BatchEndParams) A callback that is invoked at end of each batch. This can be used to measure speed, get result from evaluation metric. etc. kvstore : KVStore The KVStore. update_on_kvstore : bool Whether or not perform weight updating on kvstore. logger : logging logger When not specified, default logger will be used. work_load_list : list of float or int, optional The list of work load for different devices, in the same order as ``ctx``. monitor : Monitor, optional Monitor installed to executor, for monitoring outputs, weights, and gradients for debugging. Notes ----- - This function will inplace update the NDArrays in `arg_params` and `aux_states`.
def _train_multi_device(symbol, ctx, arg_names, param_names, aux_names, arg_params, aux_params, begin_epoch, end_epoch, epoch_size, optimizer, kvstore, update_on_kvstore, train_data, eval_data=None, eval_metric=None, epoch_end_callback=None, batch_end_callback=None, logger=None, work_load_list=None, monitor=None, eval_end_callback=None, eval_batch_end_callback=None, sym_gen=None): """Internal training function on multiple devices. This function will also work for single device as well. Parameters ---------- symbol : Symbol The network configuration. ctx : list of Context The training devices. arg_names: list of str Name of all arguments of the network. param_names: list of str Name of all trainable parameters of the network. aux_names: list of str Name of all auxiliary states of the network. arg_params : dict of str to NDArray Model parameter, dict of name to NDArray of net's weights. aux_params : dict of str to NDArray Model parameter, dict of name to NDArray of net's auxiliary states. begin_epoch : int The begining training epoch. end_epoch : int The end training epoch. epoch_size : int, optional Number of batches in a epoch. In default, it is set to ``ceil(num_train_examples / batch_size)``. optimizer : Optimizer The optimization algorithm train_data : DataIter Training data iterator. eval_data : DataIter Validation data iterator. eval_metric : EvalMetric An evaluation function or a list of evaluation functions. epoch_end_callback : callable(epoch, symbol, arg_params, aux_states) A callback that is invoked at end of each epoch. This can be used to checkpoint model each epoch. batch_end_callback : callable(BatchEndParams) A callback that is invoked at end of each batch. This can be used to measure speed, get result from evaluation metric. etc. kvstore : KVStore The KVStore. update_on_kvstore : bool Whether or not perform weight updating on kvstore. logger : logging logger When not specified, default logger will be used. work_load_list : list of float or int, optional The list of work load for different devices, in the same order as ``ctx``. monitor : Monitor, optional Monitor installed to executor, for monitoring outputs, weights, and gradients for debugging. Notes ----- - This function will inplace update the NDArrays in `arg_params` and `aux_states`. """ if logger is None: logger = logging executor_manager = DataParallelExecutorManager(symbol=symbol, sym_gen=sym_gen, ctx=ctx, train_data=train_data, param_names=param_names, arg_names=arg_names, aux_names=aux_names, work_load_list=work_load_list, logger=logger) if monitor: executor_manager.install_monitor(monitor) executor_manager.set_params(arg_params, aux_params) if not update_on_kvstore: updater = get_updater(optimizer) else: kvstore.set_optimizer(optimizer) if kvstore: _initialize_kvstore(kvstore=kvstore, param_arrays=executor_manager.param_arrays, arg_params=arg_params, param_names=executor_manager.param_names, update_on_kvstore=update_on_kvstore) # Now start training train_data.reset() for epoch in range(begin_epoch, end_epoch): # Training phase tic = time.time() eval_metric.reset() nbatch = 0 # Iterate over training data. while True: do_reset = True for data_batch in train_data: executor_manager.load_data_batch(data_batch) if monitor is not None: monitor.tic() executor_manager.forward(is_train=True) executor_manager.backward() if update_on_kvstore: if 'nccl' in kvstore.type: _update_params_on_kvstore_nccl(executor_manager.param_arrays, executor_manager.grad_arrays, kvstore, executor_manager.param_names) else: _update_params_on_kvstore(executor_manager.param_arrays, executor_manager.grad_arrays, kvstore, executor_manager.param_names) else: _update_params(executor_manager.param_arrays, executor_manager.grad_arrays, updater=updater, num_device=len(ctx), kvstore=kvstore, param_names=executor_manager.param_names) if monitor is not None: monitor.toc_print() # evaluate at end, so we can lazy copy executor_manager.update_metric(eval_metric, data_batch.label) nbatch += 1 # batch callback (for print purpose) if batch_end_callback is not None: batch_end_params = BatchEndParam(epoch=epoch, nbatch=nbatch, eval_metric=eval_metric, locals=locals()) _multiple_callbacks(batch_end_callback, batch_end_params) # this epoch is done possibly earlier if epoch_size is not None and nbatch >= epoch_size: do_reset = False break if do_reset: logger.info('Epoch[%d] Resetting Data Iterator', epoch) train_data.reset() # this epoch is done if epoch_size is None or nbatch >= epoch_size: break toc = time.time() logger.info('Epoch[%d] Time cost=%.3f', epoch, (toc - tic)) if epoch_end_callback or epoch + 1 == end_epoch: executor_manager.copy_to(arg_params, aux_params) _multiple_callbacks(epoch_end_callback, epoch, symbol, arg_params, aux_params) # evaluation if eval_data: eval_metric.reset() eval_data.reset() total_num_batch = 0 for i, eval_batch in enumerate(eval_data): executor_manager.load_data_batch(eval_batch) executor_manager.forward(is_train=False) executor_manager.update_metric(eval_metric, eval_batch.label) if eval_batch_end_callback is not None: batch_end_params = BatchEndParam(epoch=epoch, nbatch=i, eval_metric=eval_metric, locals=locals()) _multiple_callbacks(eval_batch_end_callback, batch_end_params) total_num_batch += 1 if eval_end_callback is not None: eval_end_params = BatchEndParam(epoch=epoch, nbatch=total_num_batch, eval_metric=eval_metric, locals=locals()) _multiple_callbacks(eval_end_callback, eval_end_params) eval_data.reset()
Checkpoint the model data into file. Parameters ---------- prefix : str Prefix of model name. epoch : int The epoch number of the model. symbol : Symbol The input Symbol. arg_params : dict of str to NDArray Model parameter, dict of name to NDArray of net's weights. aux_params : dict of str to NDArray Model parameter, dict of name to NDArray of net's auxiliary states. Notes ----- - ``prefix-symbol.json`` will be saved for symbol. - ``prefix-epoch.params`` will be saved for parameters.
def save_checkpoint(prefix, epoch, symbol, arg_params, aux_params): """Checkpoint the model data into file. Parameters ---------- prefix : str Prefix of model name. epoch : int The epoch number of the model. symbol : Symbol The input Symbol. arg_params : dict of str to NDArray Model parameter, dict of name to NDArray of net's weights. aux_params : dict of str to NDArray Model parameter, dict of name to NDArray of net's auxiliary states. Notes ----- - ``prefix-symbol.json`` will be saved for symbol. - ``prefix-epoch.params`` will be saved for parameters. """ if symbol is not None: symbol.save('%s-symbol.json' % prefix) save_dict = {('arg:%s' % k) : v.as_in_context(cpu()) for k, v in arg_params.items()} save_dict.update({('aux:%s' % k) : v.as_in_context(cpu()) for k, v in aux_params.items()}) param_name = '%s-%04d.params' % (prefix, epoch) nd.save(param_name, save_dict) logging.info('Saved checkpoint to \"%s\"', param_name)
Load model checkpoint from file. Parameters ---------- prefix : str Prefix of model name. epoch : int Epoch number of model we would like to load. Returns ------- symbol : Symbol The symbol configuration of computation network. arg_params : dict of str to NDArray Model parameter, dict of name to NDArray of net's weights. aux_params : dict of str to NDArray Model parameter, dict of name to NDArray of net's auxiliary states. Notes ----- - Symbol will be loaded from ``prefix-symbol.json``. - Parameters will be loaded from ``prefix-epoch.params``.
def load_checkpoint(prefix, epoch): """Load model checkpoint from file. Parameters ---------- prefix : str Prefix of model name. epoch : int Epoch number of model we would like to load. Returns ------- symbol : Symbol The symbol configuration of computation network. arg_params : dict of str to NDArray Model parameter, dict of name to NDArray of net's weights. aux_params : dict of str to NDArray Model parameter, dict of name to NDArray of net's auxiliary states. Notes ----- - Symbol will be loaded from ``prefix-symbol.json``. - Parameters will be loaded from ``prefix-epoch.params``. """ symbol = sym.load('%s-symbol.json' % prefix) save_dict = nd.load('%s-%04d.params' % (prefix, epoch)) arg_params = {} aux_params = {} for k, v in save_dict.items(): tp, name = k.split(':', 1) if tp == 'arg': arg_params[name] = v if tp == 'aux': aux_params[name] = v return (symbol, arg_params, aux_params)
verify the argument of the default symbol and user provided parameters
def _check_arguments(self): """verify the argument of the default symbol and user provided parameters""" if self.argument_checked: return assert(self.symbol is not None) self.argument_checked = True # check if symbol contain duplicated names. _check_arguments(self.symbol) # rematch parameters to delete useless ones if self.allow_extra_params: if self.arg_params: arg_names = set(self.symbol.list_arguments()) self.arg_params = {k : v for k, v in self.arg_params.items() if k in arg_names} if self.aux_params: aux_names = set(self.symbol.list_auxiliary_states()) self.aux_params = {k : v for k, v in self.aux_params.items() if k in aux_names}
Initialize weight parameters and auxiliary states.
def _init_params(self, inputs, overwrite=False): """Initialize weight parameters and auxiliary states.""" inputs = [x if isinstance(x, DataDesc) else DataDesc(*x) for x in inputs] input_shapes = {item.name: item.shape for item in inputs} arg_shapes, _, aux_shapes = self.symbol.infer_shape(**input_shapes) assert arg_shapes is not None input_dtypes = {item.name: item.dtype for item in inputs} arg_dtypes, _, aux_dtypes = self.symbol.infer_type(**input_dtypes) assert arg_dtypes is not None arg_names = self.symbol.list_arguments() input_names = input_shapes.keys() param_names = [key for key in arg_names if key not in input_names] aux_names = self.symbol.list_auxiliary_states() param_name_attrs = [x for x in zip(arg_names, arg_shapes, arg_dtypes) if x[0] in param_names] arg_params = {k : nd.zeros(shape=s, dtype=t) for k, s, t in param_name_attrs} aux_name_attrs = [x for x in zip(aux_names, aux_shapes, aux_dtypes) if x[0] in aux_names] aux_params = {k : nd.zeros(shape=s, dtype=t) for k, s, t in aux_name_attrs} for k, v in arg_params.items(): if self.arg_params and k in self.arg_params and (not overwrite): arg_params[k][:] = self.arg_params[k][:] else: self.initializer(k, v) for k, v in aux_params.items(): if self.aux_params and k in self.aux_params and (not overwrite): aux_params[k][:] = self.aux_params[k][:] else: self.initializer(k, v) self.arg_params = arg_params self.aux_params = aux_params return (arg_names, list(param_names), aux_names)
Initialize the predictor module for running prediction.
def _init_predictor(self, input_shapes, type_dict=None): """Initialize the predictor module for running prediction.""" shapes = {name: self.arg_params[name].shape for name in self.arg_params} shapes.update(dict(input_shapes)) if self._pred_exec is not None: arg_shapes, _, _ = self.symbol.infer_shape(**shapes) assert arg_shapes is not None, "Incomplete input shapes" pred_shapes = [x.shape for x in self._pred_exec.arg_arrays] if arg_shapes == pred_shapes: return # for now only use the first device pred_exec = self.symbol.simple_bind( self.ctx[0], grad_req='null', type_dict=type_dict, **shapes) pred_exec.copy_params_from(self.arg_params, self.aux_params) _check_arguments(self.symbol) self._pred_exec = pred_exec
Initialize the iterator given input.
def _init_iter(self, X, y, is_train): """Initialize the iterator given input.""" if isinstance(X, (np.ndarray, nd.NDArray)): if y is None: if is_train: raise ValueError('y must be specified when X is numpy.ndarray') else: y = np.zeros(X.shape[0]) if not isinstance(y, (np.ndarray, nd.NDArray)): raise TypeError('y must be ndarray when X is numpy.ndarray') if X.shape[0] != y.shape[0]: raise ValueError("The numbers of data points and labels not equal") if y.ndim == 2 and y.shape[1] == 1: y = y.flatten() if y.ndim != 1: raise ValueError("Label must be 1D or 2D (with 2nd dimension being 1)") if is_train: return io.NDArrayIter(X, y, min(X.shape[0], self.numpy_batch_size), shuffle=is_train, last_batch_handle='roll_over') else: return io.NDArrayIter(X, y, min(X.shape[0], self.numpy_batch_size), shuffle=False) if not isinstance(X, io.DataIter): raise TypeError('X must be DataIter, NDArray or numpy.ndarray') return X
Initialize the iterator given eval_data.
def _init_eval_iter(self, eval_data): """Initialize the iterator given eval_data.""" if eval_data is None: return eval_data if isinstance(eval_data, (tuple, list)) and len(eval_data) == 2: if eval_data[0] is not None: if eval_data[1] is None and isinstance(eval_data[0], io.DataIter): return eval_data[0] input_data = (np.array(eval_data[0]) if isinstance(eval_data[0], list) else eval_data[0]) input_label = (np.array(eval_data[1]) if isinstance(eval_data[1], list) else eval_data[1]) return self._init_iter(input_data, input_label, is_train=True) else: raise ValueError("Eval data is NONE") if not isinstance(eval_data, io.DataIter): raise TypeError('Eval data must be DataIter, or ' \ 'NDArray/numpy.ndarray/list pair (i.e. tuple/list of length 2)') return eval_data
Run the prediction, always only use one device. Parameters ---------- X : mxnet.DataIter num_batch : int or None The number of batch to run. Go though all batches if ``None``. Returns ------- y : numpy.ndarray or a list of numpy.ndarray if the network has multiple outputs. The predicted value of the output.
def predict(self, X, num_batch=None, return_data=False, reset=True): """Run the prediction, always only use one device. Parameters ---------- X : mxnet.DataIter num_batch : int or None The number of batch to run. Go though all batches if ``None``. Returns ------- y : numpy.ndarray or a list of numpy.ndarray if the network has multiple outputs. The predicted value of the output. """ X = self._init_iter(X, None, is_train=False) if reset: X.reset() data_shapes = X.provide_data data_names = [x[0] for x in data_shapes] type_dict = dict((key, value.dtype) for (key, value) in self.arg_params.items()) for x in X.provide_data: if isinstance(x, DataDesc): type_dict[x.name] = x.dtype else: type_dict[x[0]] = mx_real_t self._init_predictor(data_shapes, type_dict) batch_size = X.batch_size data_arrays = [self._pred_exec.arg_dict[name] for name in data_names] output_list = [[] for _ in range(len(self._pred_exec.outputs))] if return_data: data_list = [[] for _ in X.provide_data] label_list = [[] for _ in X.provide_label] i = 0 for batch in X: _load_data(batch, data_arrays) self._pred_exec.forward(is_train=False) padded = batch.pad real_size = batch_size - padded for o_list, o_nd in zip(output_list, self._pred_exec.outputs): o_list.append(o_nd[0:real_size].asnumpy()) if return_data: for j, x in enumerate(batch.data): data_list[j].append(x[0:real_size].asnumpy()) for j, x in enumerate(batch.label): label_list[j].append(x[0:real_size].asnumpy()) i += 1 if num_batch is not None and i == num_batch: break outputs = [np.concatenate(x) for x in output_list] if len(outputs) == 1: outputs = outputs[0] if return_data: data = [np.concatenate(x) for x in data_list] label = [np.concatenate(x) for x in label_list] if len(data) == 1: data = data[0] if len(label) == 1: label = label[0] return outputs, data, label else: return outputs
Run the model given an input and calculate the score as assessed by an evaluation metric. Parameters ---------- X : mxnet.DataIter eval_metric : metric.metric The metric for calculating score. num_batch : int or None The number of batches to run. Go though all batches if ``None``. Returns ------- s : float The final score.
def score(self, X, eval_metric='acc', num_batch=None, batch_end_callback=None, reset=True): """Run the model given an input and calculate the score as assessed by an evaluation metric. Parameters ---------- X : mxnet.DataIter eval_metric : metric.metric The metric for calculating score. num_batch : int or None The number of batches to run. Go though all batches if ``None``. Returns ------- s : float The final score. """ # setup metric if not isinstance(eval_metric, metric.EvalMetric): eval_metric = metric.create(eval_metric) X = self._init_iter(X, None, is_train=False) if reset: X.reset() data_shapes = X.provide_data data_names = [x[0] for x in data_shapes] type_dict = dict((key, value.dtype) for (key, value) in self.arg_params.items()) for x in X.provide_data: if isinstance(x, DataDesc): type_dict[x.name] = x.dtype else: type_dict[x[0]] = mx_real_t self._init_predictor(data_shapes, type_dict) data_arrays = [self._pred_exec.arg_dict[name] for name in data_names] for i, batch in enumerate(X): if num_batch is not None and i == num_batch: break _load_data(batch, data_arrays) self._pred_exec.forward(is_train=False) eval_metric.update(batch.label, self._pred_exec.outputs) if batch_end_callback is not None: batch_end_params = BatchEndParam(epoch=0, nbatch=i, eval_metric=eval_metric, locals=locals()) _multiple_callbacks(batch_end_callback, batch_end_params) return eval_metric.get()[1]
Fit the model. Parameters ---------- X : DataIter, or numpy.ndarray/NDArray Training data. If `X` is a `DataIter`, the name or (if name not available) the position of its outputs should match the corresponding variable names defined in the symbolic graph. y : numpy.ndarray/NDArray, optional Training set label. If X is ``numpy.ndarray`` or `NDArray`, `y` is required to be set. While y can be 1D or 2D (with 2nd dimension as 1), its first dimension must be the same as `X`, i.e. the number of data points and labels should be equal. eval_data : DataIter or numpy.ndarray/list/NDArray pair If eval_data is numpy.ndarray/list/NDArray pair, it should be ``(valid_data, valid_label)``. eval_metric : metric.EvalMetric or str or callable The evaluation metric. This could be the name of evaluation metric or a custom evaluation function that returns statistics based on a minibatch. epoch_end_callback : callable(epoch, symbol, arg_params, aux_states) A callback that is invoked at end of each epoch. This can be used to checkpoint model each epoch. batch_end_callback: callable(epoch) A callback that is invoked at end of each batch for purposes of printing. kvstore: KVStore or str, optional The KVStore or a string kvstore type: 'local', 'dist_sync', 'dist_async' In default uses 'local', often no need to change for single machiine. logger : logging logger, optional When not specified, default logger will be used. work_load_list : float or int, optional The list of work load for different devices, in the same order as `ctx`. Note ---- KVStore behavior - 'local', multi-devices on a single machine, will automatically choose best type. - 'dist_sync', multiple machines communicating via BSP. - 'dist_async', multiple machines with asynchronous communication.
def fit(self, X, y=None, eval_data=None, eval_metric='acc', epoch_end_callback=None, batch_end_callback=None, kvstore='local', logger=None, work_load_list=None, monitor=None, eval_end_callback=LogValidationMetricsCallback(), eval_batch_end_callback=None): """Fit the model. Parameters ---------- X : DataIter, or numpy.ndarray/NDArray Training data. If `X` is a `DataIter`, the name or (if name not available) the position of its outputs should match the corresponding variable names defined in the symbolic graph. y : numpy.ndarray/NDArray, optional Training set label. If X is ``numpy.ndarray`` or `NDArray`, `y` is required to be set. While y can be 1D or 2D (with 2nd dimension as 1), its first dimension must be the same as `X`, i.e. the number of data points and labels should be equal. eval_data : DataIter or numpy.ndarray/list/NDArray pair If eval_data is numpy.ndarray/list/NDArray pair, it should be ``(valid_data, valid_label)``. eval_metric : metric.EvalMetric or str or callable The evaluation metric. This could be the name of evaluation metric or a custom evaluation function that returns statistics based on a minibatch. epoch_end_callback : callable(epoch, symbol, arg_params, aux_states) A callback that is invoked at end of each epoch. This can be used to checkpoint model each epoch. batch_end_callback: callable(epoch) A callback that is invoked at end of each batch for purposes of printing. kvstore: KVStore or str, optional The KVStore or a string kvstore type: 'local', 'dist_sync', 'dist_async' In default uses 'local', often no need to change for single machiine. logger : logging logger, optional When not specified, default logger will be used. work_load_list : float or int, optional The list of work load for different devices, in the same order as `ctx`. Note ---- KVStore behavior - 'local', multi-devices on a single machine, will automatically choose best type. - 'dist_sync', multiple machines communicating via BSP. - 'dist_async', multiple machines with asynchronous communication. """ data = self._init_iter(X, y, is_train=True) eval_data = self._init_eval_iter(eval_data) if self.sym_gen: self.symbol = self.sym_gen(data.default_bucket_key) # pylint: disable=no-member self._check_arguments() self.kwargs["sym"] = self.symbol arg_names, param_names, aux_names = \ self._init_params(data.provide_data+data.provide_label) # setup metric if not isinstance(eval_metric, metric.EvalMetric): eval_metric = metric.create(eval_metric) # create kvstore (kvstore, update_on_kvstore) = _create_kvstore( kvstore, len(self.ctx), self.arg_params) param_idx2name = {} if update_on_kvstore: param_idx2name.update(enumerate(param_names)) else: for i, n in enumerate(param_names): for k in range(len(self.ctx)): param_idx2name[i*len(self.ctx)+k] = n self.kwargs["param_idx2name"] = param_idx2name # init optmizer if isinstance(self.optimizer, str): batch_size = data.batch_size if kvstore and 'dist' in kvstore.type and '_async' not in kvstore.type: batch_size *= kvstore.num_workers optimizer = opt.create(self.optimizer, rescale_grad=(1.0/batch_size), **(self.kwargs)) elif isinstance(self.optimizer, opt.Optimizer): if not optimizer.idx2name: optimizer.idx2name = param_idx2name.copy() optimizer = self.optimizer # do training _train_multi_device(self.symbol, self.ctx, arg_names, param_names, aux_names, self.arg_params, self.aux_params, begin_epoch=self.begin_epoch, end_epoch=self.num_epoch, epoch_size=self.epoch_size, optimizer=optimizer, train_data=data, eval_data=eval_data, eval_metric=eval_metric, epoch_end_callback=epoch_end_callback, batch_end_callback=batch_end_callback, kvstore=kvstore, update_on_kvstore=update_on_kvstore, logger=logger, work_load_list=work_load_list, monitor=monitor, eval_end_callback=eval_end_callback, eval_batch_end_callback=eval_batch_end_callback, sym_gen=self.sym_gen)
Load model checkpoint from file. Parameters ---------- prefix : str Prefix of model name. epoch : int epoch number of model we would like to load. ctx : Context or list of Context, optional The device context of training and prediction. kwargs : dict Other parameters for model, including `num_epoch`, optimizer and `numpy_batch_size`. Returns ------- model : FeedForward The loaded model that can be used for prediction. Notes ----- - ``prefix-symbol.json`` will be saved for symbol. - ``prefix-epoch.params`` will be saved for parameters.
def load(prefix, epoch, ctx=None, **kwargs): """Load model checkpoint from file. Parameters ---------- prefix : str Prefix of model name. epoch : int epoch number of model we would like to load. ctx : Context or list of Context, optional The device context of training and prediction. kwargs : dict Other parameters for model, including `num_epoch`, optimizer and `numpy_batch_size`. Returns ------- model : FeedForward The loaded model that can be used for prediction. Notes ----- - ``prefix-symbol.json`` will be saved for symbol. - ``prefix-epoch.params`` will be saved for parameters. """ symbol, arg_params, aux_params = load_checkpoint(prefix, epoch) return FeedForward(symbol, ctx=ctx, arg_params=arg_params, aux_params=aux_params, begin_epoch=epoch, **kwargs)
Checkpoint the model checkpoint into file. You can also use `pickle` to do the job if you only work on Python. The advantage of `load` and `save` (as compared to `pickle`) is that the resulting file can be loaded from other MXNet language bindings. One can also directly `load`/`save` from/to cloud storage(S3, HDFS) Parameters ---------- prefix : str Prefix of model name. Notes ----- - ``prefix-symbol.json`` will be saved for symbol. - ``prefix-epoch.params`` will be saved for parameters.
def save(self, prefix, epoch=None): """Checkpoint the model checkpoint into file. You can also use `pickle` to do the job if you only work on Python. The advantage of `load` and `save` (as compared to `pickle`) is that the resulting file can be loaded from other MXNet language bindings. One can also directly `load`/`save` from/to cloud storage(S3, HDFS) Parameters ---------- prefix : str Prefix of model name. Notes ----- - ``prefix-symbol.json`` will be saved for symbol. - ``prefix-epoch.params`` will be saved for parameters. """ if epoch is None: epoch = self.num_epoch assert epoch is not None save_checkpoint(prefix, epoch, self.symbol, self.arg_params, self.aux_params)
Functional style to create a model. This function is more consistent with functional languages such as R, where mutation is not allowed. Parameters ---------- symbol : Symbol The symbol configuration of a computation network. X : DataIter Training data. y : numpy.ndarray, optional If `X` is a ``numpy.ndarray``, `y` must be set. ctx : Context or list of Context, optional The device context of training and prediction. To use multi-GPU training, pass in a list of GPU contexts. num_epoch : int, optional The number of training epochs(epochs). epoch_size : int, optional Number of batches in a epoch. In default, it is set to ``ceil(num_train_examples / batch_size)``. optimizer : str or Optimizer, optional The name of the chosen optimizer, or an optimizer object, used for training. initializer : initializer function, optional The initialization scheme used. eval_data : DataIter or numpy.ndarray pair If `eval_set` is ``numpy.ndarray`` pair, it should be (`valid_data`, `valid_label`). eval_metric : metric.EvalMetric or str or callable The evaluation metric. Can be the name of an evaluation metric or a custom evaluation function that returns statistics based on a minibatch. epoch_end_callback : callable(epoch, symbol, arg_params, aux_states) A callback that is invoked at end of each epoch. This can be used to checkpoint model each epoch. batch_end_callback: callable(epoch) A callback that is invoked at end of each batch for print purposes. kvstore: KVStore or str, optional The KVStore or a string kvstore type: 'local', 'dist_sync', 'dis_async'. Defaults to 'local', often no need to change for single machine. logger : logging logger, optional When not specified, default logger will be used. work_load_list : list of float or int, optional The list of work load for different devices, in the same order as `ctx`.
def create(symbol, X, y=None, ctx=None, num_epoch=None, epoch_size=None, optimizer='sgd', initializer=Uniform(0.01), eval_data=None, eval_metric='acc', epoch_end_callback=None, batch_end_callback=None, kvstore='local', logger=None, work_load_list=None, eval_end_callback=LogValidationMetricsCallback(), eval_batch_end_callback=None, **kwargs): """Functional style to create a model. This function is more consistent with functional languages such as R, where mutation is not allowed. Parameters ---------- symbol : Symbol The symbol configuration of a computation network. X : DataIter Training data. y : numpy.ndarray, optional If `X` is a ``numpy.ndarray``, `y` must be set. ctx : Context or list of Context, optional The device context of training and prediction. To use multi-GPU training, pass in a list of GPU contexts. num_epoch : int, optional The number of training epochs(epochs). epoch_size : int, optional Number of batches in a epoch. In default, it is set to ``ceil(num_train_examples / batch_size)``. optimizer : str or Optimizer, optional The name of the chosen optimizer, or an optimizer object, used for training. initializer : initializer function, optional The initialization scheme used. eval_data : DataIter or numpy.ndarray pair If `eval_set` is ``numpy.ndarray`` pair, it should be (`valid_data`, `valid_label`). eval_metric : metric.EvalMetric or str or callable The evaluation metric. Can be the name of an evaluation metric or a custom evaluation function that returns statistics based on a minibatch. epoch_end_callback : callable(epoch, symbol, arg_params, aux_states) A callback that is invoked at end of each epoch. This can be used to checkpoint model each epoch. batch_end_callback: callable(epoch) A callback that is invoked at end of each batch for print purposes. kvstore: KVStore or str, optional The KVStore or a string kvstore type: 'local', 'dist_sync', 'dis_async'. Defaults to 'local', often no need to change for single machine. logger : logging logger, optional When not specified, default logger will be used. work_load_list : list of float or int, optional The list of work load for different devices, in the same order as `ctx`. """ model = FeedForward(symbol, ctx=ctx, num_epoch=num_epoch, epoch_size=epoch_size, optimizer=optimizer, initializer=initializer, **kwargs) model.fit(X, y, eval_data=eval_data, eval_metric=eval_metric, epoch_end_callback=epoch_end_callback, batch_end_callback=batch_end_callback, kvstore=kvstore, logger=logger, work_load_list=work_load_list, eval_end_callback=eval_end_callback, eval_batch_end_callback=eval_batch_end_callback) return model
Entry point to build and upload all built dockerimages in parallel :param platforms: List of platforms :param registry: Docker registry name :param load_cache: Load cache before building :return: 1 if error occurred, 0 otherwise
def build_save_containers(platforms, registry, load_cache) -> int: """ Entry point to build and upload all built dockerimages in parallel :param platforms: List of platforms :param registry: Docker registry name :param load_cache: Load cache before building :return: 1 if error occurred, 0 otherwise """ from joblib import Parallel, delayed if len(platforms) == 0: return 0 platform_results = Parallel(n_jobs=PARALLEL_BUILDS, backend="multiprocessing")( delayed(_build_save_container)(platform, registry, load_cache) for platform in platforms) is_error = False for platform_result in platform_results: if platform_result is not None: logging.error('Failed to generate %s', platform_result) is_error = True return 1 if is_error else 0
Build image for passed platform and upload the cache to the specified S3 bucket :param platform: Platform :param registry: Docker registry name :param load_cache: Load cache before building :return: Platform if failed, None otherwise
def _build_save_container(platform, registry, load_cache) -> Optional[str]: """ Build image for passed platform and upload the cache to the specified S3 bucket :param platform: Platform :param registry: Docker registry name :param load_cache: Load cache before building :return: Platform if failed, None otherwise """ docker_tag = build_util.get_docker_tag(platform=platform, registry=registry) # Preload cache if load_cache: load_docker_cache(registry=registry, docker_tag=docker_tag) # Start building logging.debug('Building %s as %s', platform, docker_tag) try: # Increase the number of retries for building the cache. image_id = build_util.build_docker(docker_binary='docker', platform=platform, registry=registry, num_retries=10, no_cache=False) logging.info('Built %s as %s', docker_tag, image_id) # Push cache to registry _upload_image(registry=registry, docker_tag=docker_tag, image_id=image_id) return None except Exception: logging.exception('Unexpected exception during build of %s', docker_tag) return platform
Upload the passed image by id, tag it with docker tag and upload to S3 bucket :param registry: Docker registry name :param docker_tag: Docker tag :param image_id: Image id :return: None
def _upload_image(registry, docker_tag, image_id) -> None: """ Upload the passed image by id, tag it with docker tag and upload to S3 bucket :param registry: Docker registry name :param docker_tag: Docker tag :param image_id: Image id :return: None """ # We don't have to retag the image since it is already in the right format logging.info('Uploading %s (%s) to %s', docker_tag, image_id, registry) push_cmd = ['docker', 'push', docker_tag] subprocess.check_call(push_cmd)
Login to the Docker Hub account :return: None
def _login_dockerhub(): """ Login to the Docker Hub account :return: None """ dockerhub_credentials = _get_dockerhub_credentials() logging.info('Logging in to DockerHub') # We use password-stdin instead of --password to avoid leaking passwords in case of an error. # This method will produce the following output: # > WARNING! Your password will be stored unencrypted in /home/jenkins_slave/.docker/config.json. # > Configure a credential helper to remove this warning. See # > https://docs.docker.com/engine/reference/commandline/login/#credentials-store # Since we consider the restricted slaves a secure environment, that's fine. Also, using this will require # third party applications which would need a review first as well. p = subprocess.run(['docker', 'login', '--username', dockerhub_credentials['username'], '--password-stdin'], stdout=subprocess.PIPE, input=str.encode(dockerhub_credentials['password'])) logging.info(p.stdout) logging.info('Successfully logged in to DockerHub')
Load the precompiled docker cache from the registry :param registry: Docker registry name :param docker_tag: Docker tag to load :return: None
def load_docker_cache(registry, docker_tag) -> None: """ Load the precompiled docker cache from the registry :param registry: Docker registry name :param docker_tag: Docker tag to load :return: None """ # We don't have to retag the image since it's already in the right format if not registry: return assert docker_tag logging.info('Loading Docker cache for %s from %s', docker_tag, registry) pull_cmd = ['docker', 'pull', docker_tag] # Don't throw an error if the image does not exist subprocess.run(pull_cmd, timeout=DOCKER_CACHE_TIMEOUT_MINS*60) logging.info('Successfully pulled docker cache')
Delete the local docker cache for the entire docker image chain :param docker_tag: Docker tag :return: None
def delete_local_docker_cache(docker_tag): """ Delete the local docker cache for the entire docker image chain :param docker_tag: Docker tag :return: None """ history_cmd = ['docker', 'history', '-q', docker_tag] try: image_ids_b = subprocess.check_output(history_cmd) image_ids_str = image_ids_b.decode('utf-8').strip() layer_ids = [id.strip() for id in image_ids_str.split('\n') if id != '<missing>'] delete_cmd = ['docker', 'image', 'rm', '--force'] delete_cmd.extend(layer_ids) subprocess.check_call(delete_cmd) except subprocess.CalledProcessError as error: # Could be caused by the image not being present logging.debug('Error during local cache deletion %s', error)
Utility to create and publish the Docker cache to Docker Hub :return:
def main() -> int: """ Utility to create and publish the Docker cache to Docker Hub :return: """ # We need to be in the same directory than the script so the commands in the dockerfiles work as # expected. But the script can be invoked from a different path base = os.path.split(os.path.realpath(__file__))[0] os.chdir(base) logging.getLogger().setLevel(logging.DEBUG) logging.getLogger('botocore').setLevel(logging.INFO) logging.getLogger('boto3').setLevel(logging.INFO) logging.getLogger('urllib3').setLevel(logging.INFO) logging.getLogger('s3transfer').setLevel(logging.INFO) def script_name() -> str: return os.path.split(sys.argv[0])[1] logging.basicConfig(format='{}: %(asctime)-15s %(message)s'.format(script_name())) parser = argparse.ArgumentParser(description="Utility for preserving and loading Docker cache", epilog="") parser.add_argument("--docker-registry", help="Docker hub registry name", type=str, required=True) args = parser.parse_args() platforms = build_util.get_platforms() try: _login_dockerhub() return build_save_containers(platforms=platforms, registry=args.docker_registry, load_cache=True) finally: _logout_dockerhub()
Download the chinese_text dataset and unzip it
def get_chinese_text(): """Download the chinese_text dataset and unzip it""" if not os.path.isdir("data/"): os.system("mkdir data/") if (not os.path.exists('data/pos.txt')) or \ (not os.path.exists('data/neg')): os.system("wget -q https://raw.githubusercontent.com/dmlc/web-data/master/mxnet/example/chinese_text.zip " "-P data/") os.chdir("./data") os.system("unzip -u chinese_text.zip") os.chdir("..")
Loads MR polarity data from files, splits the data into words and generates labels. Returns split sentences and labels.
def load_data_and_labels(): """Loads MR polarity data from files, splits the data into words and generates labels. Returns split sentences and labels. """ # download dataset get_chinese_text() # Load data from files positive_examples = list(codecs.open("./data/pos.txt", "r", "utf-8").readlines()) positive_examples = [s.strip() for s in positive_examples] positive_examples = [pe for pe in positive_examples if len(pe) < 100] negative_examples = list(codecs.open("./data/neg.txt", "r", "utf-8").readlines()) negative_examples = [s.strip() for s in negative_examples] negative_examples = [ne for ne in negative_examples if len(ne) < 100] # Split by words x_text = positive_examples + negative_examples # x_text = [clean_str(sent) for sent in x_text] x_text = [list(s) for s in x_text] # Generate labels positive_labels = [[0, 1] for _ in positive_examples] negative_labels = [[1, 0] for _ in negative_examples] y = np.concatenate([positive_labels, negative_labels], 0) return [x_text, y]
override reset behavior
def reset(self): """ override reset behavior """ if getattr(self, 'num', None) is None: self.num_inst = 0 self.sum_metric = 0.0 else: self.num_inst = [0] * self.num self.sum_metric = [0.0] * self.num
override reset behavior
def reset_local(self): """ override reset behavior """ if getattr(self, 'num', None) is None: self.num_inst = 0 self.sum_metric = 0.0 else: self.num_inst = [0] * self.num self.sum_metric = [0.0] * self.num
Implementation of updating metrics
def update(self, labels, preds): """ Implementation of updating metrics """ # get generated multi label from network cls_prob = preds[0].asnumpy() loc_loss = preds[1].asnumpy() cls_label = preds[2].asnumpy() valid_count = np.sum(cls_label >= 0) # overall accuracy & object accuracy label = cls_label.flatten() mask = np.where(label >= 0)[0] indices = np.int64(label[mask]) prob = cls_prob.transpose((0, 2, 1)).reshape((-1, cls_prob.shape[1])) prob = prob[mask, indices] self.sum_metric[0] += (-np.log(prob + self.eps)).sum() self.num_inst[0] += valid_count # smoothl1loss self.sum_metric[1] += np.sum(loc_loss) self.num_inst[1] += valid_count
Get the current evaluation result. Override the default behavior Returns ------- name : str Name of the metric. value : float Value of the evaluation.
def get(self): """Get the current evaluation result. Override the default behavior Returns ------- name : str Name of the metric. value : float Value of the evaluation. """ if self.num is None: if self.num_inst == 0: return (self.name, float('nan')) else: return (self.name, self.sum_metric / self.num_inst) else: names = ['%s'%(self.name[i]) for i in range(self.num)] values = [x / y if y != 0 else float('nan') \ for x, y in zip(self.sum_metric, self.num_inst)] return (names, values)
Structure of the Deep Q Network in the NIPS 2013 workshop paper: Playing Atari with Deep Reinforcement Learning (https://www.cs.toronto.edu/~vmnih/docs/dqn.pdf) Parameters ---------- action_num : int data : mxnet.sym.Symbol, optional name : str, optional
def dqn_sym_nips(action_num, data=None, name='dqn'): """Structure of the Deep Q Network in the NIPS 2013 workshop paper: Playing Atari with Deep Reinforcement Learning (https://www.cs.toronto.edu/~vmnih/docs/dqn.pdf) Parameters ---------- action_num : int data : mxnet.sym.Symbol, optional name : str, optional """ if data is None: net = mx.symbol.Variable('data') else: net = data net = mx.symbol.Convolution(data=net, name='conv1', kernel=(8, 8), stride=(4, 4), num_filter=16) net = mx.symbol.Activation(data=net, name='relu1', act_type="relu") net = mx.symbol.Convolution(data=net, name='conv2', kernel=(4, 4), stride=(2, 2), num_filter=32) net = mx.symbol.Activation(data=net, name='relu2', act_type="relu") net = mx.symbol.Flatten(data=net) net = mx.symbol.FullyConnected(data=net, name='fc3', num_hidden=256) net = mx.symbol.Activation(data=net, name='relu3', act_type="relu") net = mx.symbol.FullyConnected(data=net, name='fc4', num_hidden=action_num) net = mx.symbol.Custom(data=net, name=name, op_type='DQNOutput') return net
A wrapper for the user-defined handle.
def _monitor_callback_wrapper(callback): """A wrapper for the user-defined handle.""" def callback_handle(name, array, _): """ ctypes function """ callback(name, array) return callback_handle
Get the dictionary given name and ndarray pairs.
def _get_dict(names, ndarrays): """Get the dictionary given name and ndarray pairs.""" nset = set() for nm in names: if nm in nset: raise ValueError('Duplicate names detected, %s' % str(names)) nset.add(nm) return dict(zip(names, ndarrays))
List all the output NDArray. Returns ------- A list of ndarray bound to the heads of executor.
def _get_outputs(self): """List all the output NDArray. Returns ------- A list of ndarray bound to the heads of executor. """ out_size = mx_uint() handles = ctypes.POINTER(NDArrayHandle)() check_call(_LIB.MXExecutorOutputs(self.handle, ctypes.byref(out_size), ctypes.byref(handles))) num_output = out_size.value outputs = [_ndarray_cls(NDArrayHandle(handles[i])) for i in range(num_output)] return outputs
Calculate the outputs specified by the bound symbol. Parameters ---------- is_train: bool, optional Whether this forward is for evaluation purpose. If True, a backward call is expected to follow. **kwargs Additional specification of input arguments. Examples -------- >>> # doing forward by specifying data >>> texec.forward(is_train=True, data=mydata) >>> # doing forward by not specifying things, but copy to the executor before hand >>> mydata.copyto(texec.arg_dict['data']) >>> texec.forward(is_train=True) >>> # doing forward by specifying data and get outputs >>> outputs = texec.forward(is_train=True, data=mydata) >>> print(outputs[0].asnumpy())
def forward(self, is_train=False, **kwargs): """Calculate the outputs specified by the bound symbol. Parameters ---------- is_train: bool, optional Whether this forward is for evaluation purpose. If True, a backward call is expected to follow. **kwargs Additional specification of input arguments. Examples -------- >>> # doing forward by specifying data >>> texec.forward(is_train=True, data=mydata) >>> # doing forward by not specifying things, but copy to the executor before hand >>> mydata.copyto(texec.arg_dict['data']) >>> texec.forward(is_train=True) >>> # doing forward by specifying data and get outputs >>> outputs = texec.forward(is_train=True, data=mydata) >>> print(outputs[0].asnumpy()) """ if len(kwargs) != 0: arg_dict = self.arg_dict for name, array in kwargs.items(): if not isinstance(array, (NDArray, np.ndarray)): raise ValueError('only accept keyword argument of NDArrays and numpy.ndarray') if name not in arg_dict: raise TypeError('Unknown argument %s' % name) if arg_dict[name].shape != array.shape: raise ValueError('Shape not match! Argument %s, need: %s, received: %s' %(name, str(arg_dict[name].shape), str(array.shape))) arg_dict[name][:] = array check_call(_LIB.MXExecutorForward( self.handle, ctypes.c_int(int(is_train)))) return self.outputs
Do backward pass to get the gradient of arguments. Parameters ---------- out_grads : NDArray or list of NDArray or dict of str to NDArray, optional Gradient on the outputs to be propagated back. This parameter is only needed when bind is called on outputs that are not a loss function. is_train : bool, default True Whether this backward is for training or inference. Note that in rare cases you want to call backward with is_train=False to get gradient during inference. Examples -------- >>> # Example for binding on loss function symbol, which gives the loss value of the model. >>> # Equivalently it gives the head gradient for backward pass. >>> # In this example the built-in SoftmaxOutput is used as loss function. >>> # MakeLoss can be used to define customized loss function symbol. >>> net = mx.sym.Variable('data') >>> net = mx.sym.FullyConnected(net, name='fc', num_hidden=6) >>> net = mx.sym.Activation(net, name='relu', act_type="relu") >>> net = mx.sym.SoftmaxOutput(net, name='softmax') >>> args = {'data': mx.nd.ones((1, 4)), 'fc_weight': mx.nd.ones((6, 4)), >>> 'fc_bias': mx.nd.array((1, 4, 4, 4, 5, 6)), 'softmax_label': mx.nd.ones((1))} >>> args_grad = {'fc_weight': mx.nd.zeros((6, 4)), 'fc_bias': mx.nd.zeros((6))} >>> texec = net.bind(ctx=mx.cpu(), args=args, args_grad=args_grad) >>> out = texec.forward(is_train=True)[0].copy() >>> print out.asnumpy() [[ 0.00378404 0.07600445 0.07600445 0.07600445 0.20660152 0.5616011 ]] >>> texec.backward() >>> print(texec.grad_arrays[1].asnumpy()) [[ 0.00378404 0.00378404 0.00378404 0.00378404] [-0.92399555 -0.92399555 -0.92399555 -0.92399555] [ 0.07600445 0.07600445 0.07600445 0.07600445] [ 0.07600445 0.07600445 0.07600445 0.07600445] [ 0.20660152 0.20660152 0.20660152 0.20660152] [ 0.5616011 0.5616011 0.5616011 0.5616011 ]] >>> >>> # Example for binding on non-loss function symbol. >>> # Here the binding symbol is neither built-in loss function >>> # nor customized loss created by MakeLoss. >>> # As a result the head gradient is not automatically provided. >>> a = mx.sym.Variable('a') >>> b = mx.sym.Variable('b') >>> # c is not a loss function symbol >>> c = 2 * a + b >>> args = {'a': mx.nd.array([1,2]), 'b':mx.nd.array([2,3])} >>> args_grad = {'a': mx.nd.zeros((2)), 'b': mx.nd.zeros((2))} >>> texec = c.bind(ctx=mx.cpu(), args=args, args_grad=args_grad) >>> out = texec.forward(is_train=True)[0].copy() >>> print(out.asnumpy()) [ 4. 7.] >>> # out_grads is the head gradient in backward pass. >>> # Here we define 'c' as loss function. >>> # Then 'out' is passed as head gradient of backward pass. >>> texec.backward(out) >>> print(texec.grad_arrays[0].asnumpy()) [ 8. 14.] >>> print(texec.grad_arrays[1].asnumpy()) [ 4. 7.]
def backward(self, out_grads=None, is_train=True): """Do backward pass to get the gradient of arguments. Parameters ---------- out_grads : NDArray or list of NDArray or dict of str to NDArray, optional Gradient on the outputs to be propagated back. This parameter is only needed when bind is called on outputs that are not a loss function. is_train : bool, default True Whether this backward is for training or inference. Note that in rare cases you want to call backward with is_train=False to get gradient during inference. Examples -------- >>> # Example for binding on loss function symbol, which gives the loss value of the model. >>> # Equivalently it gives the head gradient for backward pass. >>> # In this example the built-in SoftmaxOutput is used as loss function. >>> # MakeLoss can be used to define customized loss function symbol. >>> net = mx.sym.Variable('data') >>> net = mx.sym.FullyConnected(net, name='fc', num_hidden=6) >>> net = mx.sym.Activation(net, name='relu', act_type="relu") >>> net = mx.sym.SoftmaxOutput(net, name='softmax') >>> args = {'data': mx.nd.ones((1, 4)), 'fc_weight': mx.nd.ones((6, 4)), >>> 'fc_bias': mx.nd.array((1, 4, 4, 4, 5, 6)), 'softmax_label': mx.nd.ones((1))} >>> args_grad = {'fc_weight': mx.nd.zeros((6, 4)), 'fc_bias': mx.nd.zeros((6))} >>> texec = net.bind(ctx=mx.cpu(), args=args, args_grad=args_grad) >>> out = texec.forward(is_train=True)[0].copy() >>> print out.asnumpy() [[ 0.00378404 0.07600445 0.07600445 0.07600445 0.20660152 0.5616011 ]] >>> texec.backward() >>> print(texec.grad_arrays[1].asnumpy()) [[ 0.00378404 0.00378404 0.00378404 0.00378404] [-0.92399555 -0.92399555 -0.92399555 -0.92399555] [ 0.07600445 0.07600445 0.07600445 0.07600445] [ 0.07600445 0.07600445 0.07600445 0.07600445] [ 0.20660152 0.20660152 0.20660152 0.20660152] [ 0.5616011 0.5616011 0.5616011 0.5616011 ]] >>> >>> # Example for binding on non-loss function symbol. >>> # Here the binding symbol is neither built-in loss function >>> # nor customized loss created by MakeLoss. >>> # As a result the head gradient is not automatically provided. >>> a = mx.sym.Variable('a') >>> b = mx.sym.Variable('b') >>> # c is not a loss function symbol >>> c = 2 * a + b >>> args = {'a': mx.nd.array([1,2]), 'b':mx.nd.array([2,3])} >>> args_grad = {'a': mx.nd.zeros((2)), 'b': mx.nd.zeros((2))} >>> texec = c.bind(ctx=mx.cpu(), args=args, args_grad=args_grad) >>> out = texec.forward(is_train=True)[0].copy() >>> print(out.asnumpy()) [ 4. 7.] >>> # out_grads is the head gradient in backward pass. >>> # Here we define 'c' as loss function. >>> # Then 'out' is passed as head gradient of backward pass. >>> texec.backward(out) >>> print(texec.grad_arrays[0].asnumpy()) [ 8. 14.] >>> print(texec.grad_arrays[1].asnumpy()) [ 4. 7.] """ if out_grads is None: out_grads = [] elif isinstance(out_grads, NDArray): out_grads = [out_grads] elif isinstance(out_grads, dict): out_grads = [out_grads[k] for k in self._symbol.list_outputs()] for obj in out_grads: if not isinstance(obj, NDArray): raise TypeError("inputs must be NDArray") ndarray = c_handle_array(out_grads) check_call(_LIB.MXExecutorBackwardEx( self.handle, mx_uint(len(out_grads)), ndarray, ctypes.c_int(is_train)))
Install callback for monitor. Parameters ---------- callback : function Takes a string and an NDArrayHandle. monitor_all : bool, default False If true, monitor both input and output, otherwise monitor output only. Examples -------- >>> def mon_callback(*args, **kwargs): >>> print("Do your stuff here.") >>> >>> texe.set_monitor_callback(mon_callback)
def set_monitor_callback(self, callback, monitor_all=False): """Install callback for monitor. Parameters ---------- callback : function Takes a string and an NDArrayHandle. monitor_all : bool, default False If true, monitor both input and output, otherwise monitor output only. Examples -------- >>> def mon_callback(*args, **kwargs): >>> print("Do your stuff here.") >>> >>> texe.set_monitor_callback(mon_callback) """ cb_type = ctypes.CFUNCTYPE(None, ctypes.c_char_p, NDArrayHandle, ctypes.c_void_p) self._monitor_callback = cb_type(_monitor_callback_wrapper(callback)) check_call(_LIB.MXExecutorSetMonitorCallbackEX( self.handle, self._monitor_callback, None, ctypes.c_int(monitor_all)))
Get dictionary representation of argument arrrays. Returns ------- arg_dict : dict of str to NDArray The dictionary that maps the names of arguments to NDArrays. Raises ------ ValueError : if there are duplicated names in the arguments.
def arg_dict(self): """Get dictionary representation of argument arrrays. Returns ------- arg_dict : dict of str to NDArray The dictionary that maps the names of arguments to NDArrays. Raises ------ ValueError : if there are duplicated names in the arguments. """ if self._arg_dict is None: self._arg_dict = Executor._get_dict( self._symbol.list_arguments(), self.arg_arrays) return self._arg_dict
Get dictionary representation of gradient arrays. Returns ------- grad_dict : dict of str to NDArray The dictionary that maps name of arguments to gradient arrays.
def grad_dict(self): """Get dictionary representation of gradient arrays. Returns ------- grad_dict : dict of str to NDArray The dictionary that maps name of arguments to gradient arrays. """ if self._grad_dict is None: self._grad_dict = Executor._get_dict( self._symbol.list_arguments(), self.grad_arrays) return self._grad_dict
Get dictionary representation of auxiliary states arrays. Returns ------- aux_dict : dict of str to NDArray The dictionary that maps name of auxiliary states to NDArrays. Raises ------ ValueError : if there are duplicated names in the auxiliary states.
def aux_dict(self): """Get dictionary representation of auxiliary states arrays. Returns ------- aux_dict : dict of str to NDArray The dictionary that maps name of auxiliary states to NDArrays. Raises ------ ValueError : if there are duplicated names in the auxiliary states. """ if self._aux_dict is None: self._aux_dict = Executor._get_dict( self._symbol.list_auxiliary_states(), self.aux_arrays) return self._aux_dict
Get dictionary representation of output arrays. Returns ------- output_dict : dict of str to NDArray The dictionary that maps name of output names to NDArrays. Raises ------ ValueError : if there are duplicated names in the outputs.
def output_dict(self): """Get dictionary representation of output arrays. Returns ------- output_dict : dict of str to NDArray The dictionary that maps name of output names to NDArrays. Raises ------ ValueError : if there are duplicated names in the outputs. """ if self._output_dict is None: self._output_dict = Executor._get_dict( self._symbol.list_outputs(), self.outputs) return self._output_dict
Copy parameters from arg_params, aux_params into executor's internal array. Parameters ---------- arg_params : dict of str to NDArray Parameters, dict of name to NDArray of arguments. aux_params : dict of str to NDArray, optional Parameters, dict of name to NDArray of auxiliary states. allow_extra_params : boolean, optional Whether allow extra parameters that are not needed by symbol. If this is True, no error will be thrown when arg_params or aux_params contain extra parameters that is not needed by the executor. Raises ------ ValueError If there is additional parameters in the dict but ``allow_extra_params=False``. Examples -------- >>> # set parameters with existing model checkpoint >>> model_prefix = 'mx_mlp' >>> sym, arg_params, aux_params = mx.model.load_checkpoint(model_prefix, 0) >>> texec.copy_params_from(arg_params, aux_params)
def copy_params_from(self, arg_params, aux_params=None, allow_extra_params=False): """Copy parameters from arg_params, aux_params into executor's internal array. Parameters ---------- arg_params : dict of str to NDArray Parameters, dict of name to NDArray of arguments. aux_params : dict of str to NDArray, optional Parameters, dict of name to NDArray of auxiliary states. allow_extra_params : boolean, optional Whether allow extra parameters that are not needed by symbol. If this is True, no error will be thrown when arg_params or aux_params contain extra parameters that is not needed by the executor. Raises ------ ValueError If there is additional parameters in the dict but ``allow_extra_params=False``. Examples -------- >>> # set parameters with existing model checkpoint >>> model_prefix = 'mx_mlp' >>> sym, arg_params, aux_params = mx.model.load_checkpoint(model_prefix, 0) >>> texec.copy_params_from(arg_params, aux_params) """ for name, array in arg_params.items(): if name in self.arg_dict: dst = self.arg_dict[name] array.astype(dst.dtype).copyto(dst) elif not allow_extra_params: raise ValueError('Find name \"%s\" that is not in the arguments' % name) if aux_params is None: return for name, array in aux_params.items(): if name in self.aux_dict: dst = self.aux_dict[name] array.astype(dst.dtype).copyto(dst) elif not allow_extra_params: raise ValueError('Find name %s that is not in the auxiliary states' % name)
Return a new executor with the same symbol and shared memory, but different input/output shapes. For runtime reshaping, variable length sequences, etc. The returned executor shares state with the current one, and cannot be used in parallel with it. Parameters ---------- partial_shaping : bool Whether to allow changing the shape of unspecified arguments. allow_up_sizing : bool Whether to allow allocating new ndarrays that's larger than the original. kwargs : dict of string to tuple of int New shape for arguments. Returns ------- exec : Executor A new executor that shares memory with self. Examples -------- >>> a = mx.sym.Variable('a') >>> b = mx.sym.Variable('b') >>> c = 2 * a + b >>> texec = c.bind(mx.cpu(), {'a': mx.nd.zeros((2, 1)), 'b': mx.nd.ones((2,1))}) >>> new_shape = {'a': (4, 2), 'b': (4, 2)} >>> texec.reshape(allow_up_sizing=True, **new_shape)
def reshape(self, partial_shaping=False, allow_up_sizing=False, **kwargs): """Return a new executor with the same symbol and shared memory, but different input/output shapes. For runtime reshaping, variable length sequences, etc. The returned executor shares state with the current one, and cannot be used in parallel with it. Parameters ---------- partial_shaping : bool Whether to allow changing the shape of unspecified arguments. allow_up_sizing : bool Whether to allow allocating new ndarrays that's larger than the original. kwargs : dict of string to tuple of int New shape for arguments. Returns ------- exec : Executor A new executor that shares memory with self. Examples -------- >>> a = mx.sym.Variable('a') >>> b = mx.sym.Variable('b') >>> c = 2 * a + b >>> texec = c.bind(mx.cpu(), {'a': mx.nd.zeros((2, 1)), 'b': mx.nd.ones((2,1))}) >>> new_shape = {'a': (4, 2), 'b': (4, 2)} >>> texec.reshape(allow_up_sizing=True, **new_shape) """ # pylint: disable=too-many-branches provided_arg_shape_data = [] # shape data # argument shape index in sdata, # e.g. [sdata[indptr[0]], sdata[indptr[1]]) is the shape of the first arg provided_arg_shape_idx = [0] provided_arg_shape_names = [] # provided argument names for k, v in kwargs.items(): if isinstance(v, tuple): provided_arg_shape_names.append(k) provided_arg_shape_data.extend(v) provided_arg_shape_idx.append(len(provided_arg_shape_data)) ctx_map_keys = [] ctx_map_dev_types = [] ctx_map_dev_ids = [] if self._group2ctx: for key, val in self._group2ctx.items(): ctx_map_keys.append(key) ctx_map_dev_types.append(val.device_typeid) ctx_map_dev_ids.append(val.device_id) handle = ExecutorHandle() shared_handle = self.handle num_in_args = ctypes.c_uint() in_arg_handles = ctypes.POINTER(NDArrayHandle)() arg_grad_handles = ctypes.POINTER(NDArrayHandle)() num_aux_states = ctypes.c_uint() aux_state_handles = ctypes.POINTER(NDArrayHandle)() check_call(_LIB.MXExecutorReshapeEx(ctypes.c_int(int(partial_shaping)), ctypes.c_int(int(allow_up_sizing)), ctypes.c_int(self._ctx.device_typeid), ctypes.c_int(self._ctx.device_id), mx_uint(len(ctx_map_keys)), c_str_array(ctx_map_keys), c_array_buf(ctypes.c_int, py_array('i', ctx_map_dev_types)), c_array_buf(ctypes.c_int, py_array('i', ctx_map_dev_ids)), mx_uint(len(provided_arg_shape_names)), c_str_array(provided_arg_shape_names), c_array_buf(mx_int, py_array('i', provided_arg_shape_data)), c_array_buf(mx_uint, py_array('I', provided_arg_shape_idx)), ctypes.byref(num_in_args), ctypes.byref(in_arg_handles), ctypes.byref(arg_grad_handles), ctypes.byref(num_aux_states), ctypes.byref(aux_state_handles), shared_handle, ctypes.byref(handle))) arg_arrays = [_ndarray_cls(NDArrayHandle(in_arg_handles[i])) for i in range(num_in_args.value)] grad_arrays = [_ndarray_cls(NDArrayHandle(arg_grad_handles[i])) if arg_grad_handles[i] is not None else None for i in range(num_in_args.value)] aux_arrays = [_ndarray_cls(NDArrayHandle(aux_state_handles[i])) for i in range(num_aux_states.value)] executor = Executor(handle, self._symbol, self._ctx, self._grad_req, self._group2ctx) executor.arg_arrays = arg_arrays executor.grad_arrays = grad_arrays executor.aux_arrays = aux_arrays return executor
Get a debug string about internal execution plan. Returns ------- debug_str : string Debug string of the executor. Examples -------- >>> a = mx.sym.Variable('a') >>> b = mx.sym.sin(a) >>> c = 2 * a + b >>> texec = c.bind(mx.cpu(), {'a': mx.nd.array([1,2]), 'b':mx.nd.array([2,3])}) >>> print(texec.debug_str()) Symbol Outputs: output[0]=_plus0(0) Variable:a -------------------- Op:_mul_scalar, Name=_mulscalar0 Inputs: arg[0]=a(0) version=0 Attrs: scalar=2 -------------------- Op:sin, Name=sin0 Inputs: arg[0]=a(0) version=0 -------------------- Op:elemwise_add, Name=_plus0 Inputs: arg[0]=_mulscalar0(0) arg[1]=sin0(0) Total 0 MB allocated Total 11 TempSpace resource requested
def debug_str(self): """Get a debug string about internal execution plan. Returns ------- debug_str : string Debug string of the executor. Examples -------- >>> a = mx.sym.Variable('a') >>> b = mx.sym.sin(a) >>> c = 2 * a + b >>> texec = c.bind(mx.cpu(), {'a': mx.nd.array([1,2]), 'b':mx.nd.array([2,3])}) >>> print(texec.debug_str()) Symbol Outputs: output[0]=_plus0(0) Variable:a -------------------- Op:_mul_scalar, Name=_mulscalar0 Inputs: arg[0]=a(0) version=0 Attrs: scalar=2 -------------------- Op:sin, Name=sin0 Inputs: arg[0]=a(0) version=0 -------------------- Op:elemwise_add, Name=_plus0 Inputs: arg[0]=_mulscalar0(0) arg[1]=sin0(0) Total 0 MB allocated Total 11 TempSpace resource requested """ debug_str = ctypes.c_char_p() check_call(_LIB.MXExecutorPrint( self.handle, ctypes.byref(debug_str))) return py_str(debug_str.value)
parse pascal voc record into a dictionary :param filename: xml file path :return: list of dict
def parse_voc_rec(filename): """ parse pascal voc record into a dictionary :param filename: xml file path :return: list of dict """ import xml.etree.ElementTree as ET tree = ET.parse(filename) objects = [] for obj in tree.findall('object'): obj_dict = dict() obj_dict['name'] = obj.find('name').text obj_dict['difficult'] = int(obj.find('difficult').text) bbox = obj.find('bndbox') obj_dict['bbox'] = [int(bbox.find('xmin').text), int(bbox.find('ymin').text), int(bbox.find('xmax').text), int(bbox.find('ymax').text)] objects.append(obj_dict) return objects
pascal voc evaluation :param detpath: detection results detpath.format(classname) :param annopath: annotations annopath.format(classname) :param imageset_file: text file containing list of images :param classname: category name :param cache_dir: caching annotations :param ovthresh: overlap threshold :param use_07_metric: whether to use voc07's 11 point ap computation :return: rec, prec, ap
def voc_eval(detpath, annopath, imageset_file, classname, cache_dir, ovthresh=0.5, use_07_metric=False): """ pascal voc evaluation :param detpath: detection results detpath.format(classname) :param annopath: annotations annopath.format(classname) :param imageset_file: text file containing list of images :param classname: category name :param cache_dir: caching annotations :param ovthresh: overlap threshold :param use_07_metric: whether to use voc07's 11 point ap computation :return: rec, prec, ap """ if not os.path.isdir(cache_dir): os.mkdir(cache_dir) cache_file = os.path.join(cache_dir, 'annotations.pkl') with open(imageset_file, 'r') as f: lines = f.readlines() image_filenames = [x.strip() for x in lines] # load annotations from cache if not os.path.isfile(cache_file): recs = {} for ind, image_filename in enumerate(image_filenames): recs[image_filename] = parse_voc_rec(annopath.format(image_filename)) if ind % 100 == 0: print('reading annotations for {:d}/{:d}'.format(ind + 1, len(image_filenames))) print('saving annotations cache to {:s}'.format(cache_file)) with open(cache_file, 'wb') as f: pickle.dump(recs, f) else: with open(cache_file, 'rb') as f: recs = pickle.load(f) # extract objects in :param classname: class_recs = {} npos = 0 for image_filename in image_filenames: objects = [obj for obj in recs[image_filename] if obj['name'] == classname] bbox = np.array([x['bbox'] for x in objects]) difficult = np.array([x['difficult'] for x in objects]).astype(np.bool) det = [False] * len(objects) # stand for detected npos = npos + sum(~difficult) class_recs[image_filename] = {'bbox': bbox, 'difficult': difficult, 'det': det} # read detections detfile = detpath.format(classname) with open(detfile, 'r') as f: lines = f.readlines() splitlines = [x.strip().split(' ') for x in lines] image_ids = [x[0] for x in splitlines] confidence = np.array([float(x[1]) for x in splitlines]) bbox = np.array([[float(z) for z in x[2:]] for x in splitlines]) # sort by confidence sorted_inds = np.argsort(-confidence) sorted_scores = np.sort(-confidence) bbox = bbox[sorted_inds, :] image_ids = [image_ids[x] for x in sorted_inds] # go down detections and mark true positives and false positives nd = len(image_ids) tp = np.zeros(nd) fp = np.zeros(nd) for d in range(nd): r = class_recs[image_ids[d]] bb = bbox[d, :].astype(float) ovmax = -np.inf bbgt = r['bbox'].astype(float) if bbgt.size > 0: # compute overlaps # intersection ixmin = np.maximum(bbgt[:, 0], bb[0]) iymin = np.maximum(bbgt[:, 1], bb[1]) ixmax = np.minimum(bbgt[:, 2], bb[2]) iymax = np.minimum(bbgt[:, 3], bb[3]) iw = np.maximum(ixmax - ixmin + 1., 0.) ih = np.maximum(iymax - iymin + 1., 0.) inters = iw * ih # union uni = ((bb[2] - bb[0] + 1.) * (bb[3] - bb[1] + 1.) + (bbgt[:, 2] - bbgt[:, 0] + 1.) * (bbgt[:, 3] - bbgt[:, 1] + 1.) - inters) overlaps = inters / uni ovmax = np.max(overlaps) jmax = np.argmax(overlaps) if ovmax > ovthresh: if not r['difficult'][jmax]: if not r['det'][jmax]: tp[d] = 1. r['det'][jmax] = 1 else: fp[d] = 1. else: fp[d] = 1. # compute precision recall fp = np.cumsum(fp) tp = np.cumsum(tp) rec = tp / float(npos) # avoid division by zero in case first detection matches a difficult ground ruth prec = tp / np.maximum(tp + fp, np.finfo(np.float64).eps) ap = voc_ap(rec, prec, use_07_metric) return rec, prec, ap
Register operators
def register(op_name): """Register operators""" def wrapper(func): """Helper function to map functions""" try: import onnx as _ MXNetGraph.registry_[op_name] = func except ImportError: pass return func return wrapper
Convert MXNet layer to ONNX
def convert_layer(node, **kwargs): """Convert MXNet layer to ONNX""" op = str(node["op"]) if op not in MXNetGraph.registry_: raise AttributeError("No conversion function registered for op type %s yet." % op) convert_func = MXNetGraph.registry_[op] return convert_func(node, **kwargs)
Helper function to split params dictionary into args and aux params Parameters ---------- sym : :class:`~mxnet.symbol.Symbol` MXNet symbol object params : dict of ``str`` to :class:`~mxnet.ndarray.NDArray` Dict of converted parameters stored in ``mxnet.ndarray.NDArray`` format Returns ------- arg_params : dict of ``str`` to :class:`~mxnet.ndarray.NDArray` Dict of converted parameters stored in ``mxnet.ndarray.NDArray`` format aux_params : dict of ``str`` to :class:`~mxnet.ndarray.NDArray` Dict of converted parameters stored in ``mxnet.ndarray.NDArray`` format
def split_params(sym, params): """Helper function to split params dictionary into args and aux params Parameters ---------- sym : :class:`~mxnet.symbol.Symbol` MXNet symbol object params : dict of ``str`` to :class:`~mxnet.ndarray.NDArray` Dict of converted parameters stored in ``mxnet.ndarray.NDArray`` format Returns ------- arg_params : dict of ``str`` to :class:`~mxnet.ndarray.NDArray` Dict of converted parameters stored in ``mxnet.ndarray.NDArray`` format aux_params : dict of ``str`` to :class:`~mxnet.ndarray.NDArray` Dict of converted parameters stored in ``mxnet.ndarray.NDArray`` format """ arg_params = {} aux_params = {} for args in sym.list_arguments(): if args in params: arg_params.update({args: nd.array(params[args])}) for aux in sym.list_auxiliary_states(): if aux in params: aux_params.update({aux: nd.array(params[aux])}) return arg_params, aux_params
Infer output shapes and return dictionary of output name to shape :param :class:`~mxnet.symbol.Symbol` sym: symbol to perform infer shape on :param dic of (str, nd.NDArray) params: :param list of tuple(int, ...) in_shape: list of all input shapes :param in_label: name of label typically used in loss that may be left in graph. This name is removed from list of inputs required by symbol :return: dictionary of output name to shape :rtype: dict of (str, tuple(int, ...))
def get_outputs(sym, params, in_shape, in_label): """ Infer output shapes and return dictionary of output name to shape :param :class:`~mxnet.symbol.Symbol` sym: symbol to perform infer shape on :param dic of (str, nd.NDArray) params: :param list of tuple(int, ...) in_shape: list of all input shapes :param in_label: name of label typically used in loss that may be left in graph. This name is removed from list of inputs required by symbol :return: dictionary of output name to shape :rtype: dict of (str, tuple(int, ...)) """ # remove any input listed in params from sym.list_inputs() and bind them to the input shapes provided # by user. Also remove in_label, which is the name of the label symbol that may have been used # as the label for loss during training. inputs = {n: tuple(s) for n, s in zip([n for n in sym.list_inputs() if n not in params and n != in_label], in_shape)} # Add params and their shape to list of inputs inputs.update({n: v.shape for n, v in params.items() if n in sym.list_inputs()}) # Provide input data as well as input params to infer_shape() _, out_shapes, _ = sym.infer_shape(**inputs) out_names = list() for name in sym.list_outputs(): if name.endswith('_output'): out_names.append(name[:-len('_output')]) else: logging.info("output '%s' does not end with '_output'", name) out_names.append(name) assert len(out_shapes) == len(out_names) # bind output shapes with output names graph_outputs = {n: s for n, s in zip(out_names, out_shapes)} return graph_outputs
Convert weights to numpy
def convert_weights_to_numpy(weights_dict): """Convert weights to numpy""" return dict([(k.replace("arg:", "").replace("aux:", ""), v.asnumpy()) for k, v in weights_dict.items()])
Convert MXNet graph to ONNX graph Parameters ---------- sym : :class:`~mxnet.symbol.Symbol` MXNet symbol object params : dict of ``str`` to :class:`~mxnet.ndarray.NDArray` Dict of converted parameters stored in ``mxnet.ndarray.NDArray`` format in_shape : List of tuple Input shape of the model e.g [(1,3,224,224)] in_type : data type Input data type e.g. np.float32 verbose : Boolean If true will print logs of the model conversion Returns ------- graph : GraphProto ONNX graph
def create_onnx_graph_proto(self, sym, params, in_shape, in_type, verbose=False): """Convert MXNet graph to ONNX graph Parameters ---------- sym : :class:`~mxnet.symbol.Symbol` MXNet symbol object params : dict of ``str`` to :class:`~mxnet.ndarray.NDArray` Dict of converted parameters stored in ``mxnet.ndarray.NDArray`` format in_shape : List of tuple Input shape of the model e.g [(1,3,224,224)] in_type : data type Input data type e.g. np.float32 verbose : Boolean If true will print logs of the model conversion Returns ------- graph : GraphProto ONNX graph """ try: from onnx import (checker, helper, NodeProto, ValueInfoProto, TensorProto) from onnx.helper import make_tensor_value_info except ImportError: raise ImportError("Onnx and protobuf need to be installed. " + "Instructions to install - https://github.com/onnx/onnx") # When MXNet model is saved to json file , MXNet adds a node for label. # The name of this node is, name of the last node + "_label" ( i.e if last node # name is "Softmax", this node will have a name "Softmax_label". Also, the new node # will always be second last node in the json graph. # Deriving the output_label name. output_label = sym.get_internals()[len(sym.get_internals()) - 1].name + "_label" weights = MXNetGraph.convert_weights_to_numpy(params) mx_graph = json.loads(sym.tojson())["nodes"] initializer = [] all_processed_nodes = [] onnx_processed_nodes = [] onnx_processed_inputs = [] onnx_processed_outputs = [] index_lookup = [] # Determine output shape graph_outputs = MXNetGraph.get_outputs(sym, params, in_shape, output_label) graph_input_idx = 0 for idx, node in enumerate(mx_graph): op = node["op"] name = node["name"] if verbose: logging.info("Converting idx: %d, op: %s, name: %s", idx, op, name) # A node is an input node if its op_name is "null" and is not # in params dict if op == "null" and name not in params: # Handling graph input # Skipping output_label node, as this node is not part of graph # Refer "output_label" assignment above for more details. if name == output_label: continue converted = MXNetGraph.convert_layer( node, is_input=True, mx_graph=mx_graph, weights=weights, in_shape=in_shape[graph_input_idx], in_type=in_type, proc_nodes=all_processed_nodes, initializer=initializer, index_lookup=index_lookup) graph_input_idx += 1 else: # Handling graph layers converted = MXNetGraph.convert_layer( node, is_input=False, mx_graph=mx_graph, weights=weights, in_shape=in_shape, in_type=in_type, proc_nodes=all_processed_nodes, initializer=initializer, index_lookup=index_lookup, idx=idx ) if isinstance(converted, list): # Iterate for all converted nodes for converted_node in converted: # If converted node is ValueInfoProto, add it in inputs if isinstance(converted_node, ValueInfoProto): onnx_processed_inputs.append(converted_node) # If converted node is NodeProto, add it in processed nodes list elif isinstance(converted_node, NodeProto): onnx_processed_nodes.append(converted_node) # some operators have multiple outputs, # therefore, check all output node names node_names = list(converted_node.output) for nodename in node_names: if nodename in graph_outputs: onnx_processed_outputs.append( make_tensor_value_info( name=nodename, elem_type=in_type, shape=graph_outputs[nodename] ) ) if verbose: logging.info("Output node is: %s", nodename) elif isinstance(converted_node, TensorProto): raise ValueError("Did not expect TensorProto") else: raise ValueError("node is of an unrecognized type: %s" % type(node)) all_processed_nodes.append(converted_node) if idx > 0: # Handling extra node added to the graph if the MXNet model was # saved to json file, # refer "output_label" initialization above for more details. # if extra node was added then prev_index to the last node is adjusted. if idx == (len(mx_graph) - 1) and \ mx_graph[len(mx_graph)-2]["name"] == output_label: prev_index = index_lookup[idx - 2] else: prev_index = index_lookup[idx - 1] index_lookup.append(prev_index+len(converted)) else: index_lookup.append(len(converted) - 1) else: logging.info("Operator converter function should always return a list") graph = helper.make_graph( onnx_processed_nodes, "mxnet_converted_model", onnx_processed_inputs, onnx_processed_outputs ) graph.initializer.extend(initializer) checker.check_graph(graph) return graph
Compute learning rate and refactor scheduler Parameters: --------- learning_rate : float original learning rate lr_refactor_step : comma separated str epochs to change learning rate lr_refactor_ratio : float lr *= ratio at certain steps num_example : int number of training images, used to estimate the iterations given epochs batch_size : int training batch size begin_epoch : int starting epoch Returns: --------- (learning_rate, mx.lr_scheduler) as tuple
def get_lr_scheduler(learning_rate, lr_refactor_step, lr_refactor_ratio, num_example, batch_size, begin_epoch): """ Compute learning rate and refactor scheduler Parameters: --------- learning_rate : float original learning rate lr_refactor_step : comma separated str epochs to change learning rate lr_refactor_ratio : float lr *= ratio at certain steps num_example : int number of training images, used to estimate the iterations given epochs batch_size : int training batch size begin_epoch : int starting epoch Returns: --------- (learning_rate, mx.lr_scheduler) as tuple """ assert lr_refactor_ratio > 0 iter_refactor = [int(r) for r in lr_refactor_step.split(',') if r.strip()] if lr_refactor_ratio >= 1: return (learning_rate, None) else: lr = learning_rate epoch_size = num_example // batch_size for s in iter_refactor: if begin_epoch >= s: lr *= lr_refactor_ratio if lr != learning_rate: logging.getLogger().info("Adjusted learning rate to {} for epoch {}".format(lr, begin_epoch)) steps = [epoch_size * (x - begin_epoch) for x in iter_refactor if x > begin_epoch] if not steps: return (lr, None) lr_scheduler = mx.lr_scheduler.MultiFactorScheduler(step=steps, factor=lr_refactor_ratio) return (lr, lr_scheduler)
Wrapper for training phase. Parameters: ---------- net : str symbol name for the network structure train_path : str record file path for training num_classes : int number of object classes, not including background batch_size : int training batch-size data_shape : int or tuple width/height as integer or (3, height, width) tuple mean_pixels : tuple of floats mean pixel values for red, green and blue resume : int resume from previous checkpoint if > 0 finetune : int fine-tune from previous checkpoint if > 0 pretrained : str prefix of pretrained model, including path epoch : int load epoch of either resume/finetune/pretrained model prefix : str prefix for saving checkpoints ctx : [mx.cpu()] or [mx.gpu(x)] list of mxnet contexts begin_epoch : int starting epoch for training, should be 0 if not otherwise specified end_epoch : int end epoch of training frequent : int frequency to print out training status learning_rate : float training learning rate momentum : float trainig momentum weight_decay : float training weight decay param lr_refactor_ratio : float multiplier for reducing learning rate lr_refactor_step : comma separated integers at which epoch to rescale learning rate, e.g. '30, 60, 90' freeze_layer_pattern : str regex pattern for layers need to be fixed num_example : int number of training images label_pad_width : int force padding training and validation labels to sync their label widths nms_thresh : float non-maximum suppression threshold for validation force_nms : boolean suppress overlaped objects from different classes train_list : str list file path for training, this will replace the embeded labels in record val_path : str record file path for validation val_list : str list file path for validation, this will replace the embeded labels in record iter_monitor : int monitor internal stats in networks if > 0, specified by monitor_pattern monitor_pattern : str regex pattern for monitoring network stats log_file : str log to file if enabled
def train_net(net, train_path, num_classes, batch_size, data_shape, mean_pixels, resume, finetune, pretrained, epoch, prefix, ctx, begin_epoch, end_epoch, frequent, learning_rate, momentum, weight_decay, lr_refactor_step, lr_refactor_ratio, freeze_layer_pattern='', num_example=10000, label_pad_width=350, nms_thresh=0.45, force_nms=False, ovp_thresh=0.5, use_difficult=False, class_names=None, voc07_metric=False, nms_topk=400, force_suppress=False, train_list="", val_path="", val_list="", iter_monitor=0, monitor_pattern=".*", log_file=None, kv_store=None): """ Wrapper for training phase. Parameters: ---------- net : str symbol name for the network structure train_path : str record file path for training num_classes : int number of object classes, not including background batch_size : int training batch-size data_shape : int or tuple width/height as integer or (3, height, width) tuple mean_pixels : tuple of floats mean pixel values for red, green and blue resume : int resume from previous checkpoint if > 0 finetune : int fine-tune from previous checkpoint if > 0 pretrained : str prefix of pretrained model, including path epoch : int load epoch of either resume/finetune/pretrained model prefix : str prefix for saving checkpoints ctx : [mx.cpu()] or [mx.gpu(x)] list of mxnet contexts begin_epoch : int starting epoch for training, should be 0 if not otherwise specified end_epoch : int end epoch of training frequent : int frequency to print out training status learning_rate : float training learning rate momentum : float trainig momentum weight_decay : float training weight decay param lr_refactor_ratio : float multiplier for reducing learning rate lr_refactor_step : comma separated integers at which epoch to rescale learning rate, e.g. '30, 60, 90' freeze_layer_pattern : str regex pattern for layers need to be fixed num_example : int number of training images label_pad_width : int force padding training and validation labels to sync their label widths nms_thresh : float non-maximum suppression threshold for validation force_nms : boolean suppress overlaped objects from different classes train_list : str list file path for training, this will replace the embeded labels in record val_path : str record file path for validation val_list : str list file path for validation, this will replace the embeded labels in record iter_monitor : int monitor internal stats in networks if > 0, specified by monitor_pattern monitor_pattern : str regex pattern for monitoring network stats log_file : str log to file if enabled """ # set up logger logging.basicConfig() logger = logging.getLogger() logger.setLevel(logging.INFO) if log_file: fh = logging.FileHandler(log_file) logger.addHandler(fh) # check args if isinstance(data_shape, int): data_shape = (3, data_shape, data_shape) assert len(data_shape) == 3 and data_shape[0] == 3 prefix += '_' + net + '_' + str(data_shape[1]) if isinstance(mean_pixels, (int, float)): mean_pixels = [mean_pixels, mean_pixels, mean_pixels] assert len(mean_pixels) == 3, "must provide all RGB mean values" train_iter = DetRecordIter(train_path, batch_size, data_shape, mean_pixels=mean_pixels, label_pad_width=label_pad_width, path_imglist=train_list, **cfg.train) if val_path: val_iter = DetRecordIter(val_path, batch_size, data_shape, mean_pixels=mean_pixels, label_pad_width=label_pad_width, path_imglist=val_list, **cfg.valid) else: val_iter = None # load symbol net = get_symbol_train(net, data_shape[1], num_classes=num_classes, nms_thresh=nms_thresh, force_suppress=force_suppress, nms_topk=nms_topk) # define layers with fixed weight/bias if freeze_layer_pattern.strip(): re_prog = re.compile(freeze_layer_pattern) fixed_param_names = [name for name in net.list_arguments() if re_prog.match(name)] else: fixed_param_names = None # load pretrained or resume from previous state ctx_str = '('+ ','.join([str(c) for c in ctx]) + ')' if resume > 0: logger.info("Resume training with {} from epoch {}" .format(ctx_str, resume)) _, args, auxs = mx.model.load_checkpoint(prefix, resume) begin_epoch = resume elif finetune > 0: logger.info("Start finetuning with {} from epoch {}" .format(ctx_str, finetune)) _, args, auxs = mx.model.load_checkpoint(prefix, finetune) begin_epoch = finetune # the prediction convolution layers name starts with relu, so it's fine fixed_param_names = [name for name in net.list_arguments() \ if name.startswith('conv')] elif pretrained: logger.info("Start training with {} from pretrained model {}" .format(ctx_str, pretrained)) _, args, auxs = mx.model.load_checkpoint(pretrained, epoch) args = convert_pretrained(pretrained, args) else: logger.info("Experimental: start training from scratch with {}" .format(ctx_str)) args = None auxs = None fixed_param_names = None # helper information if fixed_param_names: logger.info("Freezed parameters: [" + ','.join(fixed_param_names) + ']') # init training module mod = mx.mod.Module(net, label_names=('label',), logger=logger, context=ctx, fixed_param_names=fixed_param_names) # fit parameters batch_end_callback = mx.callback.Speedometer(train_iter.batch_size, frequent=frequent) epoch_end_callback = mx.callback.do_checkpoint(prefix) learning_rate, lr_scheduler = get_lr_scheduler(learning_rate, lr_refactor_step, lr_refactor_ratio, num_example, batch_size, begin_epoch) optimizer_params={'learning_rate':learning_rate, 'momentum':momentum, 'wd':weight_decay, 'lr_scheduler':lr_scheduler, 'clip_gradient':None, 'rescale_grad': 1.0 / len(ctx) if len(ctx) > 0 else 1.0 } monitor = mx.mon.Monitor(iter_monitor, pattern=monitor_pattern) if iter_monitor > 0 else None # run fit net, every n epochs we run evaluation network to get mAP if voc07_metric: valid_metric = VOC07MApMetric(ovp_thresh, use_difficult, class_names, pred_idx=3) else: valid_metric = MApMetric(ovp_thresh, use_difficult, class_names, pred_idx=3) # create kvstore when there are gpus kv = mx.kvstore.create(kv_store) if kv_store else None mod.fit(train_iter, val_iter, eval_metric=MultiBoxMetric(), validation_metric=valid_metric, batch_end_callback=batch_end_callback, epoch_end_callback=epoch_end_callback, optimizer='sgd', optimizer_params=optimizer_params, begin_epoch=begin_epoch, num_epoch=end_epoch, initializer=mx.init.Xavier(), arg_params=args, aux_params=auxs, allow_missing=True, monitor=monitor, kvstore=kv)
This is a set of 50 images representative of ImageNet images. This dataset was collected by randomly finding a working ImageNet link and then pasting the original ImageNet image into Google image search restricted to images licensed for reuse. A similar image (now with rights to reuse) was downloaded as a rough replacment for the original ImageNet image. The point is to have a random sample of ImageNet for use as a background distribution for explaining models trained on ImageNet data. Note that because the images are only rough replacements the labels might no longer be correct.
def imagenet50(display=False, resolution=224): """ This is a set of 50 images representative of ImageNet images. This dataset was collected by randomly finding a working ImageNet link and then pasting the original ImageNet image into Google image search restricted to images licensed for reuse. A similar image (now with rights to reuse) was downloaded as a rough replacment for the original ImageNet image. The point is to have a random sample of ImageNet for use as a background distribution for explaining models trained on ImageNet data. Note that because the images are only rough replacements the labels might no longer be correct. """ prefix = github_data_url + "imagenet50_" X = np.load(cache(prefix + "%sx%s.npy" % (resolution, resolution))).astype(np.float32) y = np.loadtxt(cache(prefix + "labels.csv")) return X, y
Return the boston housing data in a nice package.
def boston(display=False): """ Return the boston housing data in a nice package. """ d = sklearn.datasets.load_boston() df = pd.DataFrame(data=d.data, columns=d.feature_names) # pylint: disable=E1101 return df, d.target
Return the clssic IMDB sentiment analysis training data in a nice package. Full data is at: http://ai.stanford.edu/~amaas/data/sentiment/aclImdb_v1.tar.gz Paper to cite when using the data is: http://www.aclweb.org/anthology/P11-1015
def imdb(display=False): """ Return the clssic IMDB sentiment analysis training data in a nice package. Full data is at: http://ai.stanford.edu/~amaas/data/sentiment/aclImdb_v1.tar.gz Paper to cite when using the data is: http://www.aclweb.org/anthology/P11-1015 """ with open(cache(github_data_url + "imdb_train.txt")) as f: data = f.readlines() y = np.ones(25000, dtype=np.bool) y[:12500] = 0 return data, y
Predict total number of non-violent crimes per 100K popuation. This dataset is from the classic UCI Machine Learning repository: https://archive.ics.uci.edu/ml/datasets/Communities+and+Crime+Unnormalized
def communitiesandcrime(display=False): """ Predict total number of non-violent crimes per 100K popuation. This dataset is from the classic UCI Machine Learning repository: https://archive.ics.uci.edu/ml/datasets/Communities+and+Crime+Unnormalized """ raw_data = pd.read_csv( cache(github_data_url + "CommViolPredUnnormalizedData.txt"), na_values="?" ) # find the indices where the total violent crimes are known valid_inds = np.where(np.invert(np.isnan(raw_data.iloc[:,-2])))[0] y = np.array(raw_data.iloc[valid_inds,-2], dtype=np.float) # extract the predictive features and remove columns with missing values X = raw_data.iloc[valid_inds,5:-18] valid_cols = np.where(np.isnan(X.values).sum(0) == 0)[0] X = X.iloc[:,valid_cols] return X, y
Return the diabetes data in a nice package.
def diabetes(display=False): """ Return the diabetes data in a nice package. """ d = sklearn.datasets.load_diabetes() df = pd.DataFrame(data=d.data, columns=d.feature_names) # pylint: disable=E1101 return df, d.target
Return the classic iris data in a nice package.
def iris(display=False): """ Return the classic iris data in a nice package. """ d = sklearn.datasets.load_iris() df = pd.DataFrame(data=d.data, columns=d.feature_names) # pylint: disable=E1101 if display: return df, [d.target_names[v] for v in d.target] # pylint: disable=E1101 else: return df, d.target
Return the Adult census data in a nice package.
def adult(display=False): """ Return the Adult census data in a nice package. """ dtypes = [ ("Age", "float32"), ("Workclass", "category"), ("fnlwgt", "float32"), ("Education", "category"), ("Education-Num", "float32"), ("Marital Status", "category"), ("Occupation", "category"), ("Relationship", "category"), ("Race", "category"), ("Sex", "category"), ("Capital Gain", "float32"), ("Capital Loss", "float32"), ("Hours per week", "float32"), ("Country", "category"), ("Target", "category") ] raw_data = pd.read_csv( cache(github_data_url + "adult.data"), names=[d[0] for d in dtypes], na_values="?", dtype=dict(dtypes) ) data = raw_data.drop(["Education"], axis=1) # redundant with Education-Num filt_dtypes = list(filter(lambda x: not (x[0] in ["Target", "Education"]), dtypes)) data["Target"] = data["Target"] == " >50K" rcode = { "Not-in-family": 0, "Unmarried": 1, "Other-relative": 2, "Own-child": 3, "Husband": 4, "Wife": 5 } for k, dtype in filt_dtypes: if dtype == "category": if k == "Relationship": data[k] = np.array([rcode[v.strip()] for v in data[k]]) else: data[k] = data[k].cat.codes if display: return raw_data.drop(["Education", "Target", "fnlwgt"], axis=1), data["Target"].values else: return data.drop(["Target", "fnlwgt"], axis=1), data["Target"].values
A nicely packaged version of NHANES I data with surivival times as labels.
def nhanesi(display=False): """ A nicely packaged version of NHANES I data with surivival times as labels. """ X = pd.read_csv(cache(github_data_url + "NHANESI_subset_X.csv")) y = pd.read_csv(cache(github_data_url + "NHANESI_subset_y.csv"))["y"] if display: X_display = X.copy() X_display["Sex"] = ["Male" if v == 1 else "Female" for v in X["Sex"]] return X_display, np.array(y) else: return X, np.array(y)
A nicely packaged version of CRIC data with progression to ESRD within 4 years as the label.
def cric(display=False): """ A nicely packaged version of CRIC data with progression to ESRD within 4 years as the label. """ X = pd.read_csv(cache(github_data_url + "CRIC_time_4yearESRD_X.csv")) y = np.loadtxt(cache(github_data_url + "CRIC_time_4yearESRD_y.csv")) if display: X_display = X.copy() return X_display, y else: return X, y
Correlated Groups 60 A simulated dataset with tight correlations among distinct groups of features.
def corrgroups60(display=False): """ Correlated Groups 60 A simulated dataset with tight correlations among distinct groups of features. """ # set a constant seed old_seed = np.random.seed() np.random.seed(0) # generate dataset with known correlation N = 1000 M = 60 # set one coefficent from each group of 3 to 1 beta = np.zeros(M) beta[0:30:3] = 1 # build a correlation matrix with groups of 3 tightly correlated features C = np.eye(M) for i in range(0,30,3): C[i,i+1] = C[i+1,i] = 0.99 C[i,i+2] = C[i+2,i] = 0.99 C[i+1,i+2] = C[i+2,i+1] = 0.99 f = lambda X: np.matmul(X, beta) # Make sure the sample correlation is a perfect match X_start = np.random.randn(N, M) X_centered = X_start - X_start.mean(0) Sigma = np.matmul(X_centered.T, X_centered) / X_centered.shape[0] W = np.linalg.cholesky(np.linalg.inv(Sigma)).T X_white = np.matmul(X_centered, W.T) assert np.linalg.norm(np.corrcoef(np.matmul(X_centered, W.T).T) - np.eye(M)) < 1e-6 # ensure this decorrelates the data # create the final data X_final = np.matmul(X_white, np.linalg.cholesky(C).T) X = X_final y = f(X) + np.random.randn(N) * 1e-2 # restore the previous numpy random seed np.random.seed(old_seed) return pd.DataFrame(X), y
A simulated dataset with tight correlations among distinct groups of features.
def independentlinear60(display=False): """ A simulated dataset with tight correlations among distinct groups of features. """ # set a constant seed old_seed = np.random.seed() np.random.seed(0) # generate dataset with known correlation N = 1000 M = 60 # set one coefficent from each group of 3 to 1 beta = np.zeros(M) beta[0:30:3] = 1 f = lambda X: np.matmul(X, beta) # Make sure the sample correlation is a perfect match X_start = np.random.randn(N, M) X = X_start - X_start.mean(0) y = f(X) + np.random.randn(N) * 1e-2 # restore the previous numpy random seed np.random.seed(old_seed) return pd.DataFrame(X), y
Ranking datasets from lightgbm repository.
def rank(): """ Ranking datasets from lightgbm repository. """ rank_data_url = 'https://raw.githubusercontent.com/Microsoft/LightGBM/master/examples/lambdarank/' x_train, y_train = sklearn.datasets.load_svmlight_file(cache(rank_data_url + 'rank.train')) x_test, y_test = sklearn.datasets.load_svmlight_file(cache(rank_data_url + 'rank.test')) q_train = np.loadtxt(cache(rank_data_url + 'rank.train.query')) q_test = np.loadtxt(cache(rank_data_url + 'rank.test.query')) return x_train, y_train, x_test, y_test, q_train, q_test
An approximation of holdout that only retraines the model once. This is alse called ROAR (RemOve And Retrain) in work by Google. It is much more computationally efficient that the holdout method because it masks the most important features in every sample and then retrains the model once, instead of retraining the model for every test sample like the holdout metric.
def batch_remove_retrain(nmask_train, nmask_test, X_train, y_train, X_test, y_test, attr_train, attr_test, model_generator, metric): """ An approximation of holdout that only retraines the model once. This is alse called ROAR (RemOve And Retrain) in work by Google. It is much more computationally efficient that the holdout method because it masks the most important features in every sample and then retrains the model once, instead of retraining the model for every test sample like the holdout metric. """ warnings.warn("The retrain based measures can incorrectly evaluate models in some cases!") X_train, X_test = to_array(X_train, X_test) # how many features to mask assert X_train.shape[1] == X_test.shape[1] # mask nmask top features for each explanation X_train_tmp = X_train.copy() X_train_mean = X_train.mean(0) tie_breaking_noise = const_rand(X_train.shape[1]) * 1e-6 for i in range(len(y_train)): if nmask_train[i] > 0: ordering = np.argsort(-attr_train[i, :] + tie_breaking_noise) X_train_tmp[i, ordering[:nmask_train[i]]] = X_train_mean[ordering[:nmask_train[i]]] X_test_tmp = X_test.copy() for i in range(len(y_test)): if nmask_test[i] > 0: ordering = np.argsort(-attr_test[i, :] + tie_breaking_noise) X_test_tmp[i, ordering[:nmask_test[i]]] = X_train_mean[ordering[:nmask_test[i]]] # train the model with all the given features masked model_masked = model_generator() model_masked.fit(X_train_tmp, y_train) yp_test_masked = model_masked.predict(X_test_tmp) return metric(y_test, yp_test_masked)
The model is retrained for each test sample with the non-important features set to a constant. If you want to know how important a set of features is you can ask how the model would be different if only those features had existed. To determine this we can mask the other features across the entire training and test datasets, then retrain the model. If we apply compare the output of this retrained model to the original model we can see the effect produced by only knowning the important features. Since for individualized explanation methods each test sample has a different set of most important features we need to retrain the model for every test sample to get the change in model performance when a specified fraction of the most important features are retained.
def keep_retrain(nkeep, X_train, y_train, X_test, y_test, attr_test, model_generator, metric, trained_model, random_state): """ The model is retrained for each test sample with the non-important features set to a constant. If you want to know how important a set of features is you can ask how the model would be different if only those features had existed. To determine this we can mask the other features across the entire training and test datasets, then retrain the model. If we apply compare the output of this retrained model to the original model we can see the effect produced by only knowning the important features. Since for individualized explanation methods each test sample has a different set of most important features we need to retrain the model for every test sample to get the change in model performance when a specified fraction of the most important features are retained. """ warnings.warn("The retrain based measures can incorrectly evaluate models in some cases!") # see if we match the last cached call global _keep_cache args = (X_train, y_train, X_test, y_test, model_generator, metric) cache_match = False if "args" in _keep_cache: if all(a is b for a,b in zip(_keep_cache["args"], args)) and np.all(_keep_cache["attr_test"] == attr_test): cache_match = True X_train, X_test = to_array(X_train, X_test) # how many features to mask assert X_train.shape[1] == X_test.shape[1] # this is the model we will retrain many times model_masked = model_generator() # keep nkeep top features and re-train the model for each test explanation X_train_tmp = np.zeros(X_train.shape) X_test_tmp = np.zeros(X_test.shape) yp_masked_test = np.zeros(y_test.shape) tie_breaking_noise = const_rand(X_train.shape[1]) * 1e-6 last_nkeep = _keep_cache.get("nkeep", None) last_yp_masked_test = _keep_cache.get("yp_masked_test", None) for i in tqdm(range(len(y_test)), "Retraining for the 'keep' metric"): if cache_match and last_nkeep[i] == nkeep[i]: yp_masked_test[i] = last_yp_masked_test[i] elif nkeep[i] == attr_test.shape[1]: yp_masked_test[i] = trained_model.predict(X_test[i:i+1])[0] else: # mask out the most important features for this test instance X_train_tmp[:] = X_train X_test_tmp[:] = X_test ordering = np.argsort(-attr_test[i,:] + tie_breaking_noise) X_train_tmp[:,ordering[nkeep[i]:]] = X_train[:,ordering[nkeep[i]:]].mean() X_test_tmp[i,ordering[nkeep[i]:]] = X_train[:,ordering[nkeep[i]:]].mean() # retrain the model and make a prediction model_masked.fit(X_train_tmp, y_train) yp_masked_test[i] = model_masked.predict(X_test_tmp[i:i+1])[0] # save our results so the next call to us can be faster when there is redundancy _keep_cache["nkeep"] = nkeep _keep_cache["yp_masked_test"] = yp_masked_test _keep_cache["attr_test"] = attr_test _keep_cache["args"] = args return metric(y_test, yp_masked_test)
The model is revaluated for each test sample with the non-important features set to their mean.
def keep_mask(nkeep, X_train, y_train, X_test, y_test, attr_test, model_generator, metric, trained_model, random_state): """ The model is revaluated for each test sample with the non-important features set to their mean. """ X_train, X_test = to_array(X_train, X_test) # how many features to mask assert X_train.shape[1] == X_test.shape[1] # keep nkeep top features for each test explanation X_test_tmp = X_test.copy() yp_masked_test = np.zeros(y_test.shape) tie_breaking_noise = const_rand(X_train.shape[1], random_state) * 1e-6 mean_vals = X_train.mean(0) for i in range(len(y_test)): if nkeep[i] < X_test.shape[1]: ordering = np.argsort(-attr_test[i,:] + tie_breaking_noise) X_test_tmp[i,ordering[nkeep[i]:]] = mean_vals[ordering[nkeep[i]:]] yp_masked_test = trained_model.predict(X_test_tmp) return metric(y_test, yp_masked_test)
The model is revaluated for each test sample with the non-important features set to an imputed value. Note that the imputation is done using a multivariate normality assumption on the dataset. This depends on being able to estimate the full data covariance matrix (and inverse) accuractly. So X_train.shape[0] should be significantly bigger than X_train.shape[1].
def keep_impute(nkeep, X_train, y_train, X_test, y_test, attr_test, model_generator, metric, trained_model, random_state): """ The model is revaluated for each test sample with the non-important features set to an imputed value. Note that the imputation is done using a multivariate normality assumption on the dataset. This depends on being able to estimate the full data covariance matrix (and inverse) accuractly. So X_train.shape[0] should be significantly bigger than X_train.shape[1]. """ X_train, X_test = to_array(X_train, X_test) # how many features to mask assert X_train.shape[1] == X_test.shape[1] # keep nkeep top features for each test explanation C = np.cov(X_train.T) C += np.eye(C.shape[0]) * 1e-6 X_test_tmp = X_test.copy() yp_masked_test = np.zeros(y_test.shape) tie_breaking_noise = const_rand(X_train.shape[1], random_state) * 1e-6 mean_vals = X_train.mean(0) for i in range(len(y_test)): if nkeep[i] < X_test.shape[1]: ordering = np.argsort(-attr_test[i,:] + tie_breaking_noise) observe_inds = ordering[:nkeep[i]] impute_inds = ordering[nkeep[i]:] # impute missing data assuming it follows a multivariate normal distribution Coo_inv = np.linalg.inv(C[observe_inds,:][:,observe_inds]) Cio = C[impute_inds,:][:,observe_inds] impute = mean_vals[impute_inds] + Cio @ Coo_inv @ (X_test[i, observe_inds] - mean_vals[observe_inds]) X_test_tmp[i, impute_inds] = impute yp_masked_test = trained_model.predict(X_test_tmp) return metric(y_test, yp_masked_test)
The model is revaluated for each test sample with the non-important features set to resample background values.
def keep_resample(nkeep, X_train, y_train, X_test, y_test, attr_test, model_generator, metric, trained_model, random_state): """ The model is revaluated for each test sample with the non-important features set to resample background values. """ # why broken? overwriting? X_train, X_test = to_array(X_train, X_test) # how many features to mask assert X_train.shape[1] == X_test.shape[1] # how many samples to take nsamples = 100 # keep nkeep top features for each test explanation N,M = X_test.shape X_test_tmp = np.tile(X_test, [1, nsamples]).reshape(nsamples * N, M) tie_breaking_noise = const_rand(M) * 1e-6 inds = sklearn.utils.resample(np.arange(N), n_samples=nsamples, random_state=random_state) for i in range(N): if nkeep[i] < M: ordering = np.argsort(-attr_test[i,:] + tie_breaking_noise) X_test_tmp[i*nsamples:(i+1)*nsamples, ordering[nkeep[i]:]] = X_train[inds, :][:, ordering[nkeep[i]:]] yp_masked_test = trained_model.predict(X_test_tmp) yp_masked_test = np.reshape(yp_masked_test, (N, nsamples)).mean(1) # take the mean output over all samples return metric(y_test, yp_masked_test)
The how well do the features plus a constant base rate sum up to the model output.
def local_accuracy(X_train, y_train, X_test, y_test, attr_test, model_generator, metric, trained_model): """ The how well do the features plus a constant base rate sum up to the model output. """ X_train, X_test = to_array(X_train, X_test) # how many features to mask assert X_train.shape[1] == X_test.shape[1] # keep nkeep top features and re-train the model for each test explanation yp_test = trained_model.predict(X_test) return metric(yp_test, strip_list(attr_test).sum(1))
Generate a random array with a fixed seed.
def const_rand(size, seed=23980): """ Generate a random array with a fixed seed. """ old_seed = np.random.seed() np.random.seed(seed) out = np.random.rand(size) np.random.seed(old_seed) return out
Shuffle an array in-place with a fixed seed.
def const_shuffle(arr, seed=23980): """ Shuffle an array in-place with a fixed seed. """ old_seed = np.random.seed() np.random.seed(seed) np.random.shuffle(arr) np.random.seed(old_seed)
Estimate the SHAP values for a set of samples. Parameters ---------- X : numpy.array or pandas.DataFrame A matrix of samples (# samples x # features) on which to explain the model's output. Returns ------- For a models with a single output this returns a matrix of SHAP values (# samples x # features + 1). The last column is the base value of the model, which is the expected value of the model applied to the background dataset. This causes each row to sum to the model output for that sample. For models with vector outputs this returns a list of such matrices, one for each output.
def shap_values(self, X, **kwargs): """ Estimate the SHAP values for a set of samples. Parameters ---------- X : numpy.array or pandas.DataFrame A matrix of samples (# samples x # features) on which to explain the model's output. Returns ------- For a models with a single output this returns a matrix of SHAP values (# samples x # features + 1). The last column is the base value of the model, which is the expected value of the model applied to the background dataset. This causes each row to sum to the model output for that sample. For models with vector outputs this returns a list of such matrices, one for each output. """ phi = None if self.mimic_model_type == "xgboost": if not str(type(X)).endswith("xgboost.core.DMatrix'>"): X = xgboost.DMatrix(X) phi = self.trees.predict(X, pred_contribs=True) if phi is not None: if len(phi.shape) == 3: return [phi[:, i, :] for i in range(phi.shape[1])] else: return phi
Plots SHAP values for image inputs.
def image_plot(shap_values, x, labels=None, show=True, width=20, aspect=0.2, hspace=0.2, labelpad=None): """ Plots SHAP values for image inputs. """ multi_output = True if type(shap_values) != list: multi_output = False shap_values = [shap_values] # make sure labels if labels is not None: assert labels.shape[0] == shap_values[0].shape[0], "Labels must have same row count as shap_values arrays!" if multi_output: assert labels.shape[1] == len(shap_values), "Labels must have a column for each output in shap_values!" else: assert len(labels.shape) == 1, "Labels must be a vector for single output shap_values." label_kwargs = {} if labelpad is None else {'pad': labelpad} # plot our explanations fig_size = np.array([3 * (len(shap_values) + 1), 2.5 * (x.shape[0] + 1)]) if fig_size[0] > width: fig_size *= width / fig_size[0] fig, axes = pl.subplots(nrows=x.shape[0], ncols=len(shap_values) + 1, figsize=fig_size) if len(axes.shape) == 1: axes = axes.reshape(1,axes.size) for row in range(x.shape[0]): x_curr = x[row].copy() # make sure if len(x_curr.shape) == 3 and x_curr.shape[2] == 1: x_curr = x_curr.reshape(x_curr.shape[:2]) if x_curr.max() > 1: x_curr /= 255. # get a grayscale version of the image if len(x_curr.shape) == 3 and x_curr.shape[2] == 3: x_curr_gray = (0.2989 * x_curr[:,:,0] + 0.5870 * x_curr[:,:,1] + 0.1140 * x_curr[:,:,2]) # rgb to gray else: x_curr_gray = x_curr axes[row,0].imshow(x_curr, cmap=pl.get_cmap('gray')) axes[row,0].axis('off') if len(shap_values[0][row].shape) == 2: abs_vals = np.stack([np.abs(shap_values[i]) for i in range(len(shap_values))], 0).flatten() else: abs_vals = np.stack([np.abs(shap_values[i].sum(-1)) for i in range(len(shap_values))], 0).flatten() max_val = np.nanpercentile(abs_vals, 99.9) for i in range(len(shap_values)): if labels is not None: axes[row,i+1].set_title(labels[row,i], **label_kwargs) sv = shap_values[i][row] if len(shap_values[i][row].shape) == 2 else shap_values[i][row].sum(-1) axes[row,i+1].imshow(x_curr_gray, cmap=pl.get_cmap('gray'), alpha=0.15, extent=(-1, sv.shape[0], sv.shape[1], -1)) im = axes[row,i+1].imshow(sv, cmap=colors.red_transparent_blue, vmin=-max_val, vmax=max_val) axes[row,i+1].axis('off') if hspace == 'auto': fig.tight_layout() else: fig.subplots_adjust(hspace=hspace) cb = fig.colorbar(im, ax=np.ravel(axes).tolist(), label="SHAP value", orientation="horizontal", aspect=fig_size[0]/aspect) cb.outline.set_visible(False) if show: pl.show()
A leaf ordering is under-defined, this picks the ordering that keeps nearby samples similar.
def hclust_ordering(X, metric="sqeuclidean"): """ A leaf ordering is under-defined, this picks the ordering that keeps nearby samples similar. """ # compute a hierarchical clustering D = sp.spatial.distance.pdist(X, metric) cluster_matrix = sp.cluster.hierarchy.complete(D) # merge clusters, rotating them to make the end points match as best we can sets = [[i] for i in range(X.shape[0])] for i in range(cluster_matrix.shape[0]): s1 = sets[int(cluster_matrix[i,0])] s2 = sets[int(cluster_matrix[i,1])] # compute distances between the end points of the lists d_s1_s2 = pdist(np.vstack([X[s1[-1],:], X[s2[0],:]]), metric)[0] d_s2_s1 = pdist(np.vstack([X[s1[0],:], X[s2[-1],:]]), metric)[0] d_s1r_s2 = pdist(np.vstack([X[s1[0],:], X[s2[0],:]]), metric)[0] d_s1_s2r = pdist(np.vstack([X[s1[-1],:], X[s2[-1],:]]), metric)[0] # concatenete the lists in the way the minimizes the difference between # the samples at the junction best = min(d_s1_s2, d_s2_s1, d_s1r_s2, d_s1_s2r) if best == d_s1_s2: sets.append(s1 + s2) elif best == d_s2_s1: sets.append(s2 + s1) elif best == d_s1r_s2: sets.append(list(reversed(s1)) + s2) else: sets.append(s1 + list(reversed(s2))) return sets[-1]
Order other features by how much interaction they seem to have with the feature at the given index. This just bins the SHAP values for a feature along that feature's value. For true Shapley interaction index values for SHAP see the interaction_contribs option implemented in XGBoost.
def approximate_interactions(index, shap_values, X, feature_names=None): """ Order other features by how much interaction they seem to have with the feature at the given index. This just bins the SHAP values for a feature along that feature's value. For true Shapley interaction index values for SHAP see the interaction_contribs option implemented in XGBoost. """ # convert from DataFrames if we got any if str(type(X)).endswith("'pandas.core.frame.DataFrame'>"): if feature_names is None: feature_names = X.columns X = X.values index = convert_name(index, shap_values, feature_names) if X.shape[0] > 10000: a = np.arange(X.shape[0]) np.random.shuffle(a) inds = a[:10000] else: inds = np.arange(X.shape[0]) x = X[inds, index] srt = np.argsort(x) shap_ref = shap_values[inds, index] shap_ref = shap_ref[srt] inc = max(min(int(len(x) / 10.0), 50), 1) interactions = [] for i in range(X.shape[1]): val_other = X[inds, i][srt].astype(np.float) v = 0.0 if not (i == index or np.sum(np.abs(val_other)) < 1e-8): for j in range(0, len(x), inc): if np.std(val_other[j:j + inc]) > 0 and np.std(shap_ref[j:j + inc]) > 0: v += abs(np.corrcoef(shap_ref[j:j + inc], val_other[j:j + inc])[0, 1]) val_v = v val_other = np.isnan(X[inds, i][srt].astype(np.float)) v = 0.0 if not (i == index or np.sum(np.abs(val_other)) < 1e-8): for j in range(0, len(x), inc): if np.std(val_other[j:j + inc]) > 0 and np.std(shap_ref[j:j + inc]) > 0: v += abs(np.corrcoef(shap_ref[j:j + inc], val_other[j:j + inc])[0, 1]) nan_v = v interactions.append(max(val_v, nan_v)) return np.argsort(-np.abs(interactions))
Converts human agreement differences to numerical scores for coloring.
def _human_score_map(human_consensus, methods_attrs): """ Converts human agreement differences to numerical scores for coloring. """ v = 1 - min(np.sum(np.abs(methods_attrs - human_consensus)) / (np.abs(human_consensus).sum() + 1), 1.0) return v
Draw the bars and separators.
def draw_bars(out_value, features, feature_type, width_separators, width_bar): """Draw the bars and separators.""" rectangle_list = [] separator_list = [] pre_val = out_value for index, features in zip(range(len(features)), features): if feature_type == 'positive': left_bound = float(features[0]) right_bound = pre_val pre_val = left_bound separator_indent = np.abs(width_separators) separator_pos = left_bound colors = ['#FF0D57', '#FFC3D5'] else: left_bound = pre_val right_bound = float(features[0]) pre_val = right_bound separator_indent = - np.abs(width_separators) separator_pos = right_bound colors = ['#1E88E5', '#D1E6FA'] # Create rectangle if index == 0: if feature_type == 'positive': points_rectangle = [[left_bound, 0], [right_bound, 0], [right_bound, width_bar], [left_bound, width_bar], [left_bound + separator_indent, (width_bar / 2)] ] else: points_rectangle = [[right_bound, 0], [left_bound, 0], [left_bound, width_bar], [right_bound, width_bar], [right_bound + separator_indent, (width_bar / 2)] ] else: points_rectangle = [[left_bound, 0], [right_bound, 0], [right_bound + separator_indent * 0.90, (width_bar / 2)], [right_bound, width_bar], [left_bound, width_bar], [left_bound + separator_indent * 0.90, (width_bar / 2)]] line = plt.Polygon(points_rectangle, closed=True, fill=True, facecolor=colors[0], linewidth=0) rectangle_list += [line] # Create seperator points_separator = [[separator_pos, 0], [separator_pos + separator_indent, (width_bar / 2)], [separator_pos, width_bar]] line = plt.Polygon(points_separator, closed=None, fill=None, edgecolor=colors[1], lw=3) separator_list += [line] return rectangle_list, separator_list
Format data.
def format_data(data): """Format data.""" # Format negative features neg_features = np.array([[data['features'][x]['effect'], data['features'][x]['value'], data['featureNames'][x]] for x in data['features'].keys() if data['features'][x]['effect'] < 0]) neg_features = np.array(sorted(neg_features, key=lambda x: float(x[0]), reverse=False)) # Format postive features pos_features = np.array([[data['features'][x]['effect'], data['features'][x]['value'], data['featureNames'][x]] for x in data['features'].keys() if data['features'][x]['effect'] >= 0]) pos_features = np.array(sorted(pos_features, key=lambda x: float(x[0]), reverse=True)) # Define link function if data['link'] == 'identity': convert_func = lambda x: x elif data['link'] == 'logit': convert_func = lambda x: 1 / (1 + np.exp(-x)) else: assert False, "ERROR: Unrecognized link function: " + str(data['link']) # Convert negative feature values to plot values neg_val = data['outValue'] for i in neg_features: val = float(i[0]) neg_val = neg_val + np.abs(val) i[0] = convert_func(neg_val) if len(neg_features) > 0: total_neg = np.max(neg_features[:, 0].astype(float)) - \ np.min(neg_features[:, 0].astype(float)) else: total_neg = 0 # Convert positive feature values to plot values pos_val = data['outValue'] for i in pos_features: val = float(i[0]) pos_val = pos_val - np.abs(val) i[0] = convert_func(pos_val) if len(pos_features) > 0: total_pos = np.max(pos_features[:, 0].astype(float)) - \ np.min(pos_features[:, 0].astype(float)) else: total_pos = 0 # Convert output value and base value data['outValue'] = convert_func(data['outValue']) data['baseValue'] = convert_func(data['baseValue']) return neg_features, total_neg, pos_features, total_pos
Draw additive plot.
def draw_additive_plot(data, figsize, show, text_rotation=0): """Draw additive plot.""" # Turn off interactive plot if show == False: plt.ioff() # Format data neg_features, total_neg, pos_features, total_pos = format_data(data) # Compute overall metrics base_value = data['baseValue'] out_value = data['outValue'] offset_text = (np.abs(total_neg) + np.abs(total_pos)) * 0.04 # Define plots fig, ax = plt.subplots(figsize=figsize) # Compute axis limit update_axis_limits(ax, total_pos, pos_features, total_neg, neg_features, base_value) # Define width of bar width_bar = 0.1 width_separators = (ax.get_xlim()[1] - ax.get_xlim()[0]) / 200 # Create bar for negative shap values rectangle_list, separator_list = draw_bars(out_value, neg_features, 'negative', width_separators, width_bar) for i in rectangle_list: ax.add_patch(i) for i in separator_list: ax.add_patch(i) # Create bar for positive shap values rectangle_list, separator_list = draw_bars(out_value, pos_features, 'positive', width_separators, width_bar) for i in rectangle_list: ax.add_patch(i) for i in separator_list: ax.add_patch(i) # Add labels total_effect = np.abs(total_neg) + total_pos fig, ax = draw_labels(fig, ax, out_value, neg_features, 'negative', offset_text, total_effect, min_perc=0.05, text_rotation=text_rotation) fig, ax = draw_labels(fig, ax, out_value, pos_features, 'positive', offset_text, total_effect, min_perc=0.05, text_rotation=text_rotation) # higher lower legend draw_higher_lower_element(out_value, offset_text) # Add label for base value draw_base_element(base_value, ax) # Add output label out_names = data['outNames'][0] draw_output_element(out_names, out_value, ax) if show: plt.show() else: return plt.gcf()
Fails gracefully when various install steps don't work.
def try_run_setup(**kwargs): """ Fails gracefully when various install steps don't work. """ try: run_setup(**kwargs) except Exception as e: print(str(e)) if "xgboost" in str(e).lower(): kwargs["test_xgboost"] = False print("Couldn't install XGBoost for testing!") try_run_setup(**kwargs) elif "lightgbm" in str(e).lower(): kwargs["test_lightgbm"] = False print("Couldn't install LightGBM for testing!") try_run_setup(**kwargs) elif kwargs["with_binary"]: kwargs["with_binary"] = False print("WARNING: The C extension could not be compiled, sklearn tree models not supported.") try_run_setup(**kwargs) else: print("ERROR: Failed to build!")
The backward hook which computes the deeplift gradient for an nn.Module
def deeplift_grad(module, grad_input, grad_output): """The backward hook which computes the deeplift gradient for an nn.Module """ # first, get the module type module_type = module.__class__.__name__ # first, check the module is supported if module_type in op_handler: if op_handler[module_type].__name__ not in ['passthrough', 'linear_1d']: return op_handler[module_type](module, grad_input, grad_output) else: print('Warning: unrecognized nn.Module: {}'.format(module_type)) return grad_input
The forward hook used to save interim tensors, detached from the graph. Used to calculate the multipliers
def add_interim_values(module, input, output): """The forward hook used to save interim tensors, detached from the graph. Used to calculate the multipliers """ try: del module.x except AttributeError: pass try: del module.y except AttributeError: pass module_type = module.__class__.__name__ if module_type in op_handler: func_name = op_handler[module_type].__name__ # First, check for cases where we don't need to save the x and y tensors if func_name == 'passthrough': pass else: # check only the 0th input varies for i in range(len(input)): if i != 0 and type(output) is tuple: assert input[i] == output[i], "Only the 0th input may vary!" # if a new method is added, it must be added here too. This ensures tensors # are only saved if necessary if func_name in ['maxpool', 'nonlinear_1d']: # only save tensors if necessary if type(input) is tuple: setattr(module, 'x', torch.nn.Parameter(input[0].detach())) else: setattr(module, 'x', torch.nn.Parameter(input.detach())) if type(output) is tuple: setattr(module, 'y', torch.nn.Parameter(output[0].detach())) else: setattr(module, 'y', torch.nn.Parameter(output.detach())) if module_type in failure_case_modules: input[0].register_hook(deeplift_tensor_grad)
A forward hook which saves the tensor - attached to its graph. Used if we want to explain the interim outputs of a model
def get_target_input(module, input, output): """A forward hook which saves the tensor - attached to its graph. Used if we want to explain the interim outputs of a model """ try: del module.target_input except AttributeError: pass setattr(module, 'target_input', input)
Add handles to all non-container layers in the model. Recursively for non-container layers
def add_handles(self, model, forward_handle, backward_handle): """ Add handles to all non-container layers in the model. Recursively for non-container layers """ handles_list = [] for child in model.children(): if 'nn.modules.container' in str(type(child)): handles_list.extend(self.add_handles(child, forward_handle, backward_handle)) else: handles_list.append(child.register_forward_hook(forward_handle)) handles_list.append(child.register_backward_hook(backward_handle)) return handles_list
Removes the x and y attributes which were added by the forward handles Recursively searches for non-container layers
def remove_attributes(self, model): """ Removes the x and y attributes which were added by the forward handles Recursively searches for non-container layers """ for child in model.children(): if 'nn.modules.container' in str(type(child)): self.remove_attributes(child) else: try: del child.x except AttributeError: pass try: del child.y except AttributeError: pass
This gets a JSON dump of an XGBoost model while ensuring the features names are their indexes.
def get_xgboost_json(model): """ This gets a JSON dump of an XGBoost model while ensuring the features names are their indexes. """ fnames = model.feature_names model.feature_names = None json_trees = model.get_dump(with_stats=True, dump_format="json") model.feature_names = fnames # this fixes a bug where XGBoost can return invalid JSON json_trees = [t.replace(": inf,", ": 1000000000000.0,") for t in json_trees] json_trees = [t.replace(": -inf,", ": -1000000000000.0,") for t in json_trees] return json_trees
This computes the expected value conditioned on the given label value.
def __dynamic_expected_value(self, y): """ This computes the expected value conditioned on the given label value. """ return self.model.predict(self.data, np.ones(self.data.shape[0]) * y, output=self.model_output).mean(0)
Estimate the SHAP values for a set of samples. Parameters ---------- X : numpy.array, pandas.DataFrame or catboost.Pool (for catboost) A matrix of samples (# samples x # features) on which to explain the model's output. y : numpy.array An array of label values for each sample. Used when explaining loss functions. tree_limit : None (default) or int Limit the number of trees used by the model. By default None means no use the limit of the original model, and -1 means no limit. approximate : bool Run fast, but only roughly approximate the Tree SHAP values. This runs a method previously proposed by Saabas which only considers a single feature ordering. Take care since this does not have the consistency guarantees of Shapley values and places too much weight on lower splits in the tree. Returns ------- For models with a single output this returns a matrix of SHAP values (# samples x # features). Each row sums to the difference between the model output for that sample and the expected value of the model output (which is stored in the expected_value attribute of the explainer when it is constant). For models with vector outputs this returns a list of such matrices, one for each output.
def shap_values(self, X, y=None, tree_limit=None, approximate=False): """ Estimate the SHAP values for a set of samples. Parameters ---------- X : numpy.array, pandas.DataFrame or catboost.Pool (for catboost) A matrix of samples (# samples x # features) on which to explain the model's output. y : numpy.array An array of label values for each sample. Used when explaining loss functions. tree_limit : None (default) or int Limit the number of trees used by the model. By default None means no use the limit of the original model, and -1 means no limit. approximate : bool Run fast, but only roughly approximate the Tree SHAP values. This runs a method previously proposed by Saabas which only considers a single feature ordering. Take care since this does not have the consistency guarantees of Shapley values and places too much weight on lower splits in the tree. Returns ------- For models with a single output this returns a matrix of SHAP values (# samples x # features). Each row sums to the difference between the model output for that sample and the expected value of the model output (which is stored in the expected_value attribute of the explainer when it is constant). For models with vector outputs this returns a list of such matrices, one for each output. """ # see if we have a default tree_limit in place. if tree_limit is None: tree_limit = -1 if self.model.tree_limit is None else self.model.tree_limit # shortcut using the C++ version of Tree SHAP in XGBoost, LightGBM, and CatBoost if self.feature_dependence == "tree_path_dependent" and self.model.model_type != "internal" and self.data is None: phi = None if self.model.model_type == "xgboost": assert_import("xgboost") if not str(type(X)).endswith("xgboost.core.DMatrix'>"): X = xgboost.DMatrix(X) if tree_limit == -1: tree_limit = 0 phi = self.model.original_model.predict( X, ntree_limit=tree_limit, pred_contribs=True, approx_contribs=approximate, validate_features=False ) elif self.model.model_type == "lightgbm": assert not approximate, "approximate=True is not supported for LightGBM models!" phi = self.model.original_model.predict(X, num_iteration=tree_limit, pred_contrib=True) if phi.shape[1] != X.shape[1] + 1: phi = phi.reshape(X.shape[0], phi.shape[1]//(X.shape[1]+1), X.shape[1]+1) elif self.model.model_type == "catboost": # thanks to the CatBoost team for implementing this... assert not approximate, "approximate=True is not supported for CatBoost models!" assert tree_limit == -1, "tree_limit is not yet supported for CatBoost models!" if type(X) != catboost.Pool: X = catboost.Pool(X) phi = self.model.original_model.get_feature_importance(data=X, fstr_type='ShapValues') # note we pull off the last column and keep it as our expected_value if phi is not None: if len(phi.shape) == 3: self.expected_value = [phi[0, i, -1] for i in range(phi.shape[1])] return [phi[:, i, :-1] for i in range(phi.shape[1])] else: self.expected_value = phi[0, -1] return phi[:, :-1] # convert dataframes orig_X = X if str(type(X)).endswith("pandas.core.series.Series'>"): X = X.values elif str(type(X)).endswith("pandas.core.frame.DataFrame'>"): X = X.values flat_output = False if len(X.shape) == 1: flat_output = True X = X.reshape(1, X.shape[0]) if X.dtype != self.model.dtype: X = X.astype(self.model.dtype) X_missing = np.isnan(X, dtype=np.bool) assert str(type(X)).endswith("'numpy.ndarray'>"), "Unknown instance type: " + str(type(X)) assert len(X.shape) == 2, "Passed input data matrix X must have 1 or 2 dimensions!" if tree_limit < 0 or tree_limit > self.model.values.shape[0]: tree_limit = self.model.values.shape[0] if self.model_output == "logloss": assert y is not None, "Both samples and labels must be provided when explaining the loss (i.e. `explainer.shap_values(X, y)`)!" assert X.shape[0] == len(y), "The number of labels (%d) does not match the number of samples to explain (%d)!" % (len(y), X.shape[0]) transform = self.model.get_transform(self.model_output) if self.feature_dependence == "tree_path_dependent": assert self.model.fully_defined_weighting, "The background dataset you provided does not cover all the leaves in the model, " \ "so TreeExplainer cannot run with the feature_dependence=\"tree_path_dependent\" option! " \ "Try providing a larger background dataset, or using feature_dependence=\"independent\"." # run the core algorithm using the C extension assert_import("cext") phi = np.zeros((X.shape[0], X.shape[1]+1, self.model.n_outputs)) if not approximate: _cext.dense_tree_shap( self.model.children_left, self.model.children_right, self.model.children_default, self.model.features, self.model.thresholds, self.model.values, self.model.node_sample_weight, self.model.max_depth, X, X_missing, y, self.data, self.data_missing, tree_limit, self.model.base_offset, phi, feature_dependence_codes[self.feature_dependence], output_transform_codes[transform], False ) else: _cext.dense_tree_saabas( self.model.children_left, self.model.children_right, self.model.children_default, self.model.features, self.model.thresholds, self.model.values, self.model.max_depth, tree_limit, self.model.base_offset, output_transform_codes[transform], X, X_missing, y, phi ) # note we pull off the last column and keep it as our expected_value if self.model.n_outputs == 1: if self.model_output != "logloss": self.expected_value = phi[0, -1, 0] if flat_output: return phi[0, :-1, 0] else: return phi[:, :-1, 0] else: if self.model_output != "logloss": self.expected_value = [phi[0, -1, i] for i in range(phi.shape[2])] if flat_output: return [phi[0, :-1, i] for i in range(self.model.n_outputs)] else: return [phi[:, :-1, i] for i in range(self.model.n_outputs)]
Estimate the SHAP interaction values for a set of samples. Parameters ---------- X : numpy.array, pandas.DataFrame or catboost.Pool (for catboost) A matrix of samples (# samples x # features) on which to explain the model's output. y : numpy.array An array of label values for each sample. Used when explaining loss functions (not yet supported). tree_limit : None (default) or int Limit the number of trees used by the model. By default None means no use the limit of the original model, and -1 means no limit. Returns ------- For models with a single output this returns a tensor of SHAP values (# samples x # features x # features). The matrix (# features x # features) for each sample sums to the difference between the model output for that sample and the expected value of the model output (which is stored in the expected_value attribute of the explainer). Each row of this matrix sums to the SHAP value for that feature for that sample. The diagonal entries of the matrix represent the "main effect" of that feature on the prediction and the symmetric off-diagonal entries represent the interaction effects between all pairs of features for that sample. For models with vector outputs this returns a list of tensors, one for each output.
def shap_interaction_values(self, X, y=None, tree_limit=None): """ Estimate the SHAP interaction values for a set of samples. Parameters ---------- X : numpy.array, pandas.DataFrame or catboost.Pool (for catboost) A matrix of samples (# samples x # features) on which to explain the model's output. y : numpy.array An array of label values for each sample. Used when explaining loss functions (not yet supported). tree_limit : None (default) or int Limit the number of trees used by the model. By default None means no use the limit of the original model, and -1 means no limit. Returns ------- For models with a single output this returns a tensor of SHAP values (# samples x # features x # features). The matrix (# features x # features) for each sample sums to the difference between the model output for that sample and the expected value of the model output (which is stored in the expected_value attribute of the explainer). Each row of this matrix sums to the SHAP value for that feature for that sample. The diagonal entries of the matrix represent the "main effect" of that feature on the prediction and the symmetric off-diagonal entries represent the interaction effects between all pairs of features for that sample. For models with vector outputs this returns a list of tensors, one for each output. """ assert self.model_output == "margin", "Only model_output = \"margin\" is supported for SHAP interaction values right now!" assert self.feature_dependence == "tree_path_dependent", "Only feature_dependence = \"tree_path_dependent\" is supported for SHAP interaction values right now!" transform = "identity" # see if we have a default tree_limit in place. if tree_limit is None: tree_limit = -1 if self.model.tree_limit is None else self.model.tree_limit # shortcut using the C++ version of Tree SHAP in XGBoost if self.model.model_type == "xgboost": assert_import("xgboost") if not str(type(X)).endswith("xgboost.core.DMatrix'>"): X = xgboost.DMatrix(X) if tree_limit == -1: tree_limit = 0 phi = self.model.original_model.predict(X, ntree_limit=tree_limit, pred_interactions=True) # note we pull off the last column and keep it as our expected_value if len(phi.shape) == 4: self.expected_value = [phi[0, i, -1, -1] for i in range(phi.shape[1])] return [phi[:, i, :-1, :-1] for i in range(phi.shape[1])] else: self.expected_value = phi[0, -1, -1] return phi[:, :-1, :-1] # convert dataframes if str(type(X)).endswith("pandas.core.series.Series'>"): X = X.values elif str(type(X)).endswith("pandas.core.frame.DataFrame'>"): X = X.values flat_output = False if len(X.shape) == 1: flat_output = True X = X.reshape(1, X.shape[0]) if X.dtype != self.model.dtype: X = X.astype(self.model.dtype) X_missing = np.isnan(X, dtype=np.bool) assert str(type(X)).endswith("'numpy.ndarray'>"), "Unknown instance type: " + str(type(X)) assert len(X.shape) == 2, "Passed input data matrix X must have 1 or 2 dimensions!" if tree_limit < 0 or tree_limit > self.model.values.shape[0]: tree_limit = self.model.values.shape[0] # run the core algorithm using the C extension assert_import("cext") phi = np.zeros((X.shape[0], X.shape[1]+1, X.shape[1]+1, self.model.n_outputs)) _cext.dense_tree_shap( self.model.children_left, self.model.children_right, self.model.children_default, self.model.features, self.model.thresholds, self.model.values, self.model.node_sample_weight, self.model.max_depth, X, X_missing, y, self.data, self.data_missing, tree_limit, self.model.base_offset, phi, feature_dependence_codes[self.feature_dependence], output_transform_codes[transform], True ) # note we pull off the last column and keep it as our expected_value if self.model.n_outputs == 1: self.expected_value = phi[0, -1, -1, 0] if flat_output: return phi[0, :-1, :-1, 0] else: return phi[:, :-1, :-1, 0] else: self.expected_value = [phi[0, -1, -1, i] for i in range(phi.shape[3])] if flat_output: return [phi[0, :-1, :-1, i] for i in range(self.model.n_outputs)] else: return [phi[:, :-1, :-1, i] for i in range(self.model.n_outputs)]
A consistent interface to make predictions from this model.
def get_transform(self, model_output): """ A consistent interface to make predictions from this model. """ if model_output == "margin": transform = "identity" elif model_output == "probability": if self.tree_output == "log_odds": transform = "logistic" elif self.tree_output == "probability": transform = "identity" else: raise Exception("model_output = \"probability\" is not yet supported when model.tree_output = \"" + self.tree_output + "\"!") elif model_output == "logloss": if self.objective == "squared_error": transform = "squared_loss" elif self.objective == "binary_crossentropy": transform = "logistic_nlogloss" else: raise Exception("model_output = \"logloss\" is not yet supported when model.objective = \"" + self.objective + "\"!") return transform
A consistent interface to make predictions from this model. Parameters ---------- tree_limit : None (default) or int Limit the number of trees used by the model. By default None means no use the limit of the original model, and -1 means no limit.
def predict(self, X, y=None, output="margin", tree_limit=None): """ A consistent interface to make predictions from this model. Parameters ---------- tree_limit : None (default) or int Limit the number of trees used by the model. By default None means no use the limit of the original model, and -1 means no limit. """ # see if we have a default tree_limit in place. if tree_limit is None: tree_limit = -1 if self.tree_limit is None else self.tree_limit # convert dataframes if str(type(X)).endswith("pandas.core.series.Series'>"): X = X.values elif str(type(X)).endswith("pandas.core.frame.DataFrame'>"): X = X.values flat_output = False if len(X.shape) == 1: flat_output = True X = X.reshape(1, X.shape[0]) if X.dtype != self.dtype: X = X.astype(self.dtype) X_missing = np.isnan(X, dtype=np.bool) assert str(type(X)).endswith("'numpy.ndarray'>"), "Unknown instance type: " + str(type(X)) assert len(X.shape) == 2, "Passed input data matrix X must have 1 or 2 dimensions!" if tree_limit < 0 or tree_limit > self.values.shape[0]: tree_limit = self.values.shape[0] if output == "logloss": assert y is not None, "Both samples and labels must be provided when explaining the loss (i.e. `explainer.shap_values(X, y)`)!" assert X.shape[0] == len(y), "The number of labels (%d) does not match the number of samples to explain (%d)!" % (len(y), X.shape[0]) transform = self.get_transform(output) if True or self.model_type == "internal": output = np.zeros((X.shape[0], self.n_outputs)) assert_import("cext") _cext.dense_tree_predict( self.children_left, self.children_right, self.children_default, self.features, self.thresholds, self.values, self.max_depth, tree_limit, self.base_offset, output_transform_codes[transform], X, X_missing, y, output ) elif self.model_type == "xgboost": assert_import("xgboost") output = self.original_model.predict(X, output_margin=True, tree_limit=tree_limit) # drop dimensions we don't need if flat_output: if self.n_outputs == 1: return output.flatten()[0] else: return output.reshape(-1, self.n_outputs) else: if self.n_outputs == 1: return output.flatten() else: return output
Return the values for the model applied to X. Parameters ---------- X : list, if framework == 'tensorflow': numpy.array, or pandas.DataFrame if framework == 'pytorch': torch.tensor A tensor (or list of tensors) of samples (where X.shape[0] == # samples) on which to explain the model's output. ranked_outputs : None or int If ranked_outputs is None then we explain all the outputs in a multi-output model. If ranked_outputs is a positive integer then we only explain that many of the top model outputs (where "top" is determined by output_rank_order). Note that this causes a pair of values to be returned (shap_values, indexes), where phi is a list of numpy arrays for each of the output ranks, and indexes is a matrix that tells for each sample which output indexes were choses as "top". output_rank_order : "max", "min", "max_abs", or "custom" How to order the model outputs when using ranked_outputs, either by maximum, minimum, or maximum absolute value. If "custom" Then "ranked_outputs" contains a list of output nodes. rseed : None or int Seeding the randomness in shap value computation (background example choice, interpolation between current and background example, smoothing). Returns ------- For a models with a single output this returns a tensor of SHAP values with the same shape as X. For a model with multiple outputs this returns a list of SHAP value tensors, each of which are the same shape as X. If ranked_outputs is None then this list of tensors matches the number of model outputs. If ranked_outputs is a positive integer a pair is returned (shap_values, indexes), where shap_values is a list of tensors with a length of ranked_outputs, and indexes is a matrix that tells for each sample which output indexes were chosen as "top".
def shap_values(self, X, nsamples=200, ranked_outputs=None, output_rank_order="max", rseed=None): """ Return the values for the model applied to X. Parameters ---------- X : list, if framework == 'tensorflow': numpy.array, or pandas.DataFrame if framework == 'pytorch': torch.tensor A tensor (or list of tensors) of samples (where X.shape[0] == # samples) on which to explain the model's output. ranked_outputs : None or int If ranked_outputs is None then we explain all the outputs in a multi-output model. If ranked_outputs is a positive integer then we only explain that many of the top model outputs (where "top" is determined by output_rank_order). Note that this causes a pair of values to be returned (shap_values, indexes), where phi is a list of numpy arrays for each of the output ranks, and indexes is a matrix that tells for each sample which output indexes were choses as "top". output_rank_order : "max", "min", "max_abs", or "custom" How to order the model outputs when using ranked_outputs, either by maximum, minimum, or maximum absolute value. If "custom" Then "ranked_outputs" contains a list of output nodes. rseed : None or int Seeding the randomness in shap value computation (background example choice, interpolation between current and background example, smoothing). Returns ------- For a models with a single output this returns a tensor of SHAP values with the same shape as X. For a model with multiple outputs this returns a list of SHAP value tensors, each of which are the same shape as X. If ranked_outputs is None then this list of tensors matches the number of model outputs. If ranked_outputs is a positive integer a pair is returned (shap_values, indexes), where shap_values is a list of tensors with a length of ranked_outputs, and indexes is a matrix that tells for each sample which output indexes were chosen as "top". """ return self.explainer.shap_values(X, nsamples, ranked_outputs, output_rank_order, rseed)
Visualize the given SHAP values with an additive force layout. Parameters ---------- base_value : float This is the reference value that the feature contributions start from. For SHAP values it should be the value of explainer.expected_value. shap_values : numpy.array Matrix of SHAP values (# features) or (# samples x # features). If this is a 1D array then a single force plot will be drawn, if it is a 2D array then a stacked force plot will be drawn. features : numpy.array Matrix of feature values (# features) or (# samples x # features). This provides the values of all the features, and should be the same shape as the shap_values argument. feature_names : list List of feature names (# features). out_names : str The name of the outout of the model (plural to support multi-output plotting in the future). link : "identity" or "logit" The transformation used when drawing the tick mark labels. Using logit will change log-odds numbers into probabilities. matplotlib : bool Whether to use the default Javascript output, or the (less developed) matplotlib output. Using matplotlib can be helpful in scenarios where rendering Javascript/HTML is inconvenient.
def force_plot(base_value, shap_values, features=None, feature_names=None, out_names=None, link="identity", plot_cmap="RdBu", matplotlib=False, show=True, figsize=(20,3), ordering_keys=None, ordering_keys_time_format=None, text_rotation=0): """ Visualize the given SHAP values with an additive force layout. Parameters ---------- base_value : float This is the reference value that the feature contributions start from. For SHAP values it should be the value of explainer.expected_value. shap_values : numpy.array Matrix of SHAP values (# features) or (# samples x # features). If this is a 1D array then a single force plot will be drawn, if it is a 2D array then a stacked force plot will be drawn. features : numpy.array Matrix of feature values (# features) or (# samples x # features). This provides the values of all the features, and should be the same shape as the shap_values argument. feature_names : list List of feature names (# features). out_names : str The name of the outout of the model (plural to support multi-output plotting in the future). link : "identity" or "logit" The transformation used when drawing the tick mark labels. Using logit will change log-odds numbers into probabilities. matplotlib : bool Whether to use the default Javascript output, or the (less developed) matplotlib output. Using matplotlib can be helpful in scenarios where rendering Javascript/HTML is inconvenient. """ # auto unwrap the base_value if type(base_value) == np.ndarray and len(base_value) == 1: base_value = base_value[0] if (type(base_value) == np.ndarray or type(base_value) == list): if type(shap_values) != list or len(shap_values) != len(base_value): raise Exception("In v0.20 force_plot now requires the base value as the first parameter! " \ "Try shap.force_plot(explainer.expected_value, shap_values) or " \ "for multi-output models try " \ "shap.force_plot(explainer.expected_value[0], shap_values[0]).") assert not type(shap_values) == list, "The shap_values arg looks looks multi output, try shap_values[i]." link = convert_to_link(link) if type(shap_values) != np.ndarray: return visualize(shap_values) # convert from a DataFrame or other types if str(type(features)) == "<class 'pandas.core.frame.DataFrame'>": if feature_names is None: feature_names = list(features.columns) features = features.values elif str(type(features)) == "<class 'pandas.core.series.Series'>": if feature_names is None: feature_names = list(features.index) features = features.values elif isinstance(features, list): if feature_names is None: feature_names = features features = None elif features is not None and len(features.shape) == 1 and feature_names is None: feature_names = features features = None if len(shap_values.shape) == 1: shap_values = np.reshape(shap_values, (1, len(shap_values))) if out_names is None: out_names = ["output value"] elif type(out_names) == str: out_names = [out_names] if shap_values.shape[0] == 1: if feature_names is None: feature_names = [labels['FEATURE'] % str(i) for i in range(shap_values.shape[1])] if features is None: features = ["" for _ in range(len(feature_names))] if type(features) == np.ndarray: features = features.flatten() # check that the shape of the shap_values and features match if len(features) != shap_values.shape[1]: msg = "Length of features is not equal to the length of shap_values!" if len(features) == shap_values.shape[1] - 1: msg += " You might be using an old format shap_values array with the base value " \ "as the last column. In this case just pass the array without the last column." raise Exception(msg) instance = Instance(np.zeros((1, len(feature_names))), features) e = AdditiveExplanation( base_value, np.sum(shap_values[0, :]) + base_value, shap_values[0, :], None, instance, link, Model(None, out_names), DenseData(np.zeros((1, len(feature_names))), list(feature_names)) ) return visualize(e, plot_cmap, matplotlib, figsize=figsize, show=show, text_rotation=text_rotation) else: if matplotlib: raise Exception("matplotlib = True is not yet supported for force plots with multiple samples!") if shap_values.shape[0] > 3000: warnings.warn("shap.force_plot is slow for many thousands of rows, try subsampling your data.") exps = [] for i in range(shap_values.shape[0]): if feature_names is None: feature_names = [labels['FEATURE'] % str(i) for i in range(shap_values.shape[1])] if features is None: display_features = ["" for i in range(len(feature_names))] else: display_features = features[i, :] instance = Instance(np.ones((1, len(feature_names))), display_features) e = AdditiveExplanation( base_value, np.sum(shap_values[i, :]) + base_value, shap_values[i, :], None, instance, link, Model(None, out_names), DenseData(np.ones((1, len(feature_names))), list(feature_names)) ) exps.append(e) return visualize( exps, plot_cmap=plot_cmap, ordering_keys=ordering_keys, ordering_keys_time_format=ordering_keys_time_format, text_rotation=text_rotation )
Save html plots to an output file.
def save_html(out_file, plot_html): """ Save html plots to an output file. """ internal_open = False if type(out_file) == str: out_file = open(out_file, "w") internal_open = True out_file.write("<html><head><script>\n") # dump the js code bundle_path = os.path.join(os.path.split(__file__)[0], "resources", "bundle.js") with io.open(bundle_path, encoding="utf-8") as f: bundle_data = f.read() out_file.write(bundle_data) out_file.write("</script></head><body>\n") out_file.write(plot_html.data) out_file.write("</body></html>\n") if internal_open: out_file.close()
Follows a set of ops assuming their value is False and find blocked Switch paths. This is used to prune away parts of the model graph that are only used during the training phase (like dropout, batch norm, etc.).
def tensors_blocked_by_false(ops): """ Follows a set of ops assuming their value is False and find blocked Switch paths. This is used to prune away parts of the model graph that are only used during the training phase (like dropout, batch norm, etc.). """ blocked = [] def recurse(op): if op.type == "Switch": blocked.append(op.outputs[1]) # the true path is blocked since we assume the ops we trace are False else: for out in op.outputs: for c in out.consumers(): recurse(c) for op in ops: recurse(op) return blocked
Just decompose softmax into its components and recurse, we can handle all of them :) We assume the 'axis' is the last dimension because the TF codebase swaps the 'axis' to the last dimension before the softmax op if 'axis' is not already the last dimension. We also don't subtract the max before tf.exp for numerical stability since that might mess up the attributions and it seems like TensorFlow doesn't define softmax that way (according to the docs)
def softmax(explainer, op, *grads): """ Just decompose softmax into its components and recurse, we can handle all of them :) We assume the 'axis' is the last dimension because the TF codebase swaps the 'axis' to the last dimension before the softmax op if 'axis' is not already the last dimension. We also don't subtract the max before tf.exp for numerical stability since that might mess up the attributions and it seems like TensorFlow doesn't define softmax that way (according to the docs) """ in0 = op.inputs[0] in0_max = tf.reduce_max(in0, axis=-1, keepdims=True, name="in0_max") in0_centered = in0 - in0_max evals = tf.exp(in0_centered, name="custom_exp") rsum = tf.reduce_sum(evals, axis=-1, keepdims=True) div = evals / rsum explainer.between_ops.extend([evals.op, rsum.op, div.op, in0_centered.op]) # mark these as in-between the inputs and outputs out = tf.gradients(div, in0_centered, grad_ys=grads[0])[0] del explainer.between_ops[-4:] # rescale to account for our shift by in0_max (which we did for numerical stability) xin0,rin0 = tf.split(in0, 2) xin0_centered,rin0_centered = tf.split(in0_centered, 2) delta_in0 = xin0 - rin0 dup0 = [2] + [1 for i in delta_in0.shape[1:]] return tf.where( tf.tile(tf.abs(delta_in0), dup0) < 1e-6, out, out * tf.tile((xin0_centered - rin0_centered) / delta_in0, dup0) )