id
int64
0
190k
prompt
stringlengths
21
13.4M
docstring
stringlengths
1
12k
23,226
import os import shutil import json import sys import numpy as np from .log import log_or_print from nndct_shared.base import NNDCT_KEYS, GLOBAL_MAP def print_mat(mat, name="tmp mat", col=20, t=0, channel=sys.stdout): max_val = mat.max() min_val = mat.min() sum_val = mat.sum() if isinstance(max_val, float) or isinstance(max_val, int): pass else: max_val = max_val[0] min_val = min_val[0] sum_val = sum_val[0] if len(mat.shape) == 2: channel.write("showing " + str(t) + " th " + name + str(mat.shape) + "\n") channel.write("sum:" + str(sum_val) + "; max:" + str(max_val) + " min:" + str(min_val) + "\n") for r in range(mat.shape[0]): for c in range(col): channel.write(str(mat[r, c]) + " ") channel.write("\n") elif len(mat.shape) == 1: channel.write("showing " + str(t) + " th " + name + "\n") channel.write("sum:" + str(sum_val) + "; max:" + str(max_val) + " min:" + str(min_val) + "\n") for r in range(col): channel.write(str(mat[r]) + " ") channel.write("\n") else: raise Exception("wrong mode!")
null
23,227
import os import shutil import json import sys import numpy as np from .log import log_or_print from nndct_shared.base import NNDCT_KEYS, GLOBAL_MAP def latest_file(folder, file_checker=None): lists = os.listdir(folder) lists.sort(key=lambda fn: os.path.getmtime(os.path.join(folder, fn))) for file_name in lists[::-1]: abs_file = os.path.join(folder, file_name) if not file_checker or file_checker(abs_file): return abs_file
null
23,228
import sys import os import logging import io from nndct_shared.base import NNDCT_KEYS, GLOBAL_MAP, NNDCT_DEBUG_LVL GLOBAL_MAP = GlobalMap() def nndct_info_print(string): logger = GLOBAL_MAP.get_ele(NNDCT_KEYS.LOGGER) if logger: logger.info("[NNDCT_INFO] {}".format(string)) else: print("[NNDCT_INFO] {}".format(string))
null
23,229
import sys import os import logging import io from nndct_shared.base import NNDCT_KEYS, GLOBAL_MAP, NNDCT_DEBUG_LVL GLOBAL_MAP = GlobalMap() def nndct_warn_print(string): if True == GLOBAL_MAP.get_ele(NNDCT_KEYS.WARN_FLAG): logger = GLOBAL_MAP.get_ele(NNDCT_KEYS.LOGGER) if logger: logger.warning("[NNDCT_WARN] {}".format(string)) else: print("[NNDCT_WARN] {}".format(string))
null
23,230
import sys import os import logging import io from nndct_shared.base import NNDCT_KEYS, GLOBAL_MAP, NNDCT_DEBUG_LVL GLOBAL_MAP = GlobalMap() def nndct_error_print(string): if True == GLOBAL_MAP.get_ele(NNDCT_KEYS.ERROR_FLAG): logger = GLOBAL_MAP.get_ele(NNDCT_KEYS.LOGGER) if logger: logger.error("[NNDCT_ERROR] {}".format(string)) else: print("[NNDCT_ERROR] {}".format(string)) sys.exit(1)
null
23,231
import sys import os import logging import io from nndct_shared.base import NNDCT_KEYS, GLOBAL_MAP, NNDCT_DEBUG_LVL def obj_to_str(obj): if isinstance(obj, list): string = '\n'.join(["{}".format(n) for n in obj]) elif isinstance(obj, dict): string = '\n'.join(["{} : {}".format(k, v) for k, v in obj.items()]) elif isinstance(obj, str): string = obj else: raise Exception("nndct_details_debug only support list and dictionary") return string def nndct_debug_print(string, title='', level=1): if True == GLOBAL_MAP.get_ele( NNDCT_KEYS.DEBUG_FLAG) and level <= GLOBAL_MAP.get_ele( NNDCT_KEYS.VERBOSE_LEVEL): logger = GLOBAL_MAP.get_ele(NNDCT_KEYS.LOGGER) if title == 'Start': string = "\n********************* <{} : {}> *********************".format( title, string) elif title == 'End': string = "\n********************* <{} : {}> *********************\n".format( title, string) if logger: logger.debug("[NNDCT_DEBUG_Lv_{}] {}".format(level, string)) else: print("[NNDCT_DEBUG_Lv_{}] {}".format(level, string)) def nndct_details_debug(obj, title, level=NNDCT_DEBUG_LVL.DETAILS): nndct_debug_print( "\n********************* <Start : {}> *********************\n{}".format( title, obj_to_str(obj)), level=level) nndct_debug_print(title, title='End', level=level)
null
23,232
import sys import os import logging import io from nndct_shared.base import NNDCT_KEYS, GLOBAL_MAP, NNDCT_DEBUG_LVL GLOBAL_MAP = GlobalMap() def nndct_info(func): def wrapper(*args, **kwargs): info_flag = GLOBAL_MAP.get_ele(NNDCT_KEYS.INFO_FLAG) if info_flag == True: print("[NNDCT_INFO]", end='') return func(*args, **kwargs) return wrapper
null
23,233
import sys import os import logging import io from nndct_shared.base import NNDCT_KEYS, GLOBAL_MAP, NNDCT_DEBUG_LVL GLOBAL_MAP = GlobalMap() def nndct_warn(func): def wrapper(*args, **kwargs): warn_flag = GLOBAL_MAP.get_ele(NNDCT_KEYS.WARN_FLAG) if warn_flag == True: print("[NNDCT_WARN]", end='') return func(*args, **kwargs) return wrapper
null
23,234
import sys import os import logging import io from nndct_shared.base import NNDCT_KEYS, GLOBAL_MAP, NNDCT_DEBUG_LVL GLOBAL_MAP = GlobalMap() def nndct_debug(func): def wrapper(*args, **kwargs): debug_flag = GLOBAL_MAP.get_ele(NNDCT_KEYS.DEBUG_FLAG) if debug_flag == True: print("[NNDCT_DEBUG]", end='') return func(*args, **kwargs) return wrapper
null
23,235
import sys import os import logging import io from nndct_shared.base import NNDCT_KEYS, GLOBAL_MAP, NNDCT_DEBUG_LVL GLOBAL_MAP = GlobalMap() def nndct_error(func): def wrapper(*args, **kwargs): error_flag = GLOBAL_MAP.get_ele(NNDCT_KEYS.ERROR_FLAG) if error_flag == True: print("[NNDCT_ERROR]", end='') return func(*args, **kwargs) if error_flag == True: exit(1) return wrapper
null
23,236
import sys import os import logging import io from nndct_shared.base import NNDCT_KEYS, GLOBAL_MAP, NNDCT_DEBUG_LVL import logging as _logging from logging import DEBUG from logging import ERROR from logging import FATAL from logging import INFO from logging import WARN from logging import NOTSET def get_nndct_logger(filename='NndctGen_log'): log_dir = os.path.dirname(filename) if log_dir and not os.path.exists(log_dir): os.makedirs(log_dir) logger = logging.getLogger(filename.replace("/", 'SPL')) logger.setLevel(logging.INFO) formatter = logging.Formatter( '%(asctime)s %(filename)s[line:%(lineno)d] %(levelname)s %(message)s') sh = logging.StreamHandler() fh = logging.FileHandler(filename, mode='w', encoding=None, delay=False) for h in [sh, fh]: h.setLevel(logging.INFO) h.setFormatter(formatter) logger.addHandler(h) return logger
null
23,237
import sys import os import logging import io from nndct_shared.base import NNDCT_KEYS, GLOBAL_MAP, NNDCT_DEBUG_LVL def get_config_str(obj, title, ignore_prefix=[], ignore_suffix=[], ignore_keys=[]): assert hasattr( obj, 'default_kwargs' ), 'object {} has no default_kwargs, failed to generate configuration string'.format( obj) config_str = '\n' + ">> <{}>".format(title) + '\n>> ' for key in obj.default_kwargs: value = getattr(obj, key, None) if value and not any(key.endswith(s) for s in ignore_suffix) and \ not any(key.startswith(p) for p in ignore_prefix) and \ key not in ignore_keys: if isinstance(value, dict): config_str += '\n>> {}: \n>> {}'.format( key, '\n>> '.join( ['{} : {}'.format(k, v) for k, v in value.items()])) else: config_str += '\n>> {} : {}'.format(key, value) return config_str
null
23,238
import math from nndct_shared.base import NNDCT_OP def calculate_op_scale(rec, node): scale = 1.0 if node.op.type in [NNDCT_OP.MEAN]: max_factor = math.ceil(math.log(rec * 128,2)) diff = 1.0 multi_factor = 0.0 shift_factor = 0.0 for shift_factor_ in range(max_factor): factor = round((2 ** shift_factor_)/rec) diff_ = abs(factor / (2 ** shift_factor_) - 1/rec) if diff_ < diff: multi_factor = factor diff = diff_ shift_factor = shift_factor_ scale = rec * multi_factor / (2 ** shift_factor) return scale
null
23,241
from nndct_shared.utils.tensor_util import DataFormatMap from typing import List def generate_indices_group(indices: List[int], dim_size: int, groups: int) -> List[List[int]]: indices_set = set(indices) interval: int = dim_size // groups start_idx = 0 end_idx = interval ret: List[List[int]] = [] while start_idx < dim_size: idx_group: List[int] = [] for i in range(start_idx, end_idx): if i in indices_set: idx_group.append(i - start_idx) ret.append(idx_group) start_idx = end_idx end_idx += interval return ret
null
23,242
from __future__ import absolute_import from __future__ import division from __future__ import print_function import abc import copy import json import numpy as np import os from typing import List from nndct_shared.base.key_names import FrameworkType from nndct_shared.pruning import errors from nndct_shared.pruning import logging from nndct_shared.pruning import pruning_lib from nndct_shared.pruning import utils from nndct_shared.pruning.pruning_lib import is_depthwise_conv from nndct_shared.pruning.pruning_lib import is_grouped_conv from nndct_shared.pruning.utils import generate_indices_group from nndct_shared.utils import io from nndct_shared.utils.tensor_util import param_layout_transformer class ModulePruningInfoGenerator(object): def __init__(self, nodename_to_modulename=None, nodes_pruning_info=None, module_pruning_info=None): self._nodename_to_modulename = nodename_to_modulename self._nodes_pruning_info = nodes_pruning_info self._module_pruning_info = module_pruning_info if not self._module_pruning_info: self._module_pruning_info = dict() self._generate_module_pruning_info() def _generate_module_pruning_info(self): for nodename, pruning_info in self._nodes_pruning_info.items(): if nodename in self._nodename_to_modulename: modulename = self._nodename_to_modulename[nodename] self._module_pruning_info[modulename] = pruning_info def nodename_to_modulename(self): return self._nodename_to_modulename def nodename_to_modulename(self, nodename_to_modulename): self._nodename_to_modulename = nodename_to_modulename def nodes_pruning_info(self): return self._nodes_pruning_info def nodes_pruning_info(self, nodes_pruning_info): self._nodes_pruning_info = nodes_pruning_info def module_pruning_info(self): return self._module_pruning_info def serialize(self): content = dict() for key, value in self._module_pruning_info.items(): content[key] = value.serialize() return content def deserialize(cls, data): module_pruning_info = {} for key, value in data.items(): module_pruning_info[key] = pruning_lib.StructuredPruning.deserialize( value) return cls(module_pruning_info=module_pruning_info) def to_json(self): return json.dumps(self.serialize(), indent=2) import os if not os.path.exists(BASE_DIR): os.makedirs(BASE_DIR) def save_pruning_info(nodename_to_modulename, nodes_pruning_info, filepath): io.create_work_dir(os.path.dirname(filepath)) with open(filepath, 'w') as f: f.write( ModulePruningInfoGenerator(nodename_to_modulename, nodes_pruning_info).to_json())
null
23,243
from __future__ import absolute_import from __future__ import division from __future__ import print_function import abc import copy import json import numpy as np import os from typing import List from nndct_shared.base.key_names import FrameworkType from nndct_shared.pruning import errors from nndct_shared.pruning import logging from nndct_shared.pruning import pruning_lib from nndct_shared.pruning import utils from nndct_shared.pruning.pruning_lib import is_depthwise_conv from nndct_shared.pruning.pruning_lib import is_grouped_conv from nndct_shared.pruning.utils import generate_indices_group from nndct_shared.utils import io from nndct_shared.utils.tensor_util import param_layout_transformer class ModulePruningInfoGenerator(object): def __init__(self, nodename_to_modulename=None, nodes_pruning_info=None, module_pruning_info=None): def _generate_module_pruning_info(self): def nodename_to_modulename(self): def nodename_to_modulename(self, nodename_to_modulename): def nodes_pruning_info(self): def nodes_pruning_info(self, nodes_pruning_info): def module_pruning_info(self): def serialize(self): def deserialize(cls, data): def to_json(self): def load_pruning_info(filepath): with open(filepath, 'r') as f: content = json.load(f) return ModulePruningInfoGenerator.deserialize(content)
null
23,247
from __future__ import absolute_import from __future__ import division from __future__ import print_function import logging as _logging import os as _os import sys as _sys import time as _time import traceback as _traceback from logging import DEBUG from logging import ERROR from logging import FATAL from logging import INFO from logging import WARN from logging import NOTSET import threading def _log_prefix(level, timestamp=None, file_and_line=None): """Generate a nndct logline prefix.""" # pylint: disable=global-variable-not-assigned global _level_names # pylint: enable=global-variable-not-assigned # Record current time now = timestamp or _time.time() now_tuple = _time.localtime(now) now_microsecond = int(1e6 * (now % 1.0)) (filename, line) = file_and_line or _get_file_and_line() basename = _os.path.basename(filename) # Severity string severity = 'I' if level in _level_names: severity = _level_names[level][0] s = '%c%02d%02d %02d:%02d:%02d.%06d %s]' % ( severity, now_tuple[1], # month now_tuple[2], # day now_tuple[3], # hour now_tuple[4], # min now_tuple[5], # sec now_microsecond, basename) return s def get_logger(): """Return logger instance.""" global _logger # Use double-checked locking to avoid taking lock unnecessarily. if _logger: return _logger _logger_lock.acquire() try: if _logger: return _logger # Scope the TensorFlow logger to not conflict with users' loggers. logger = _logging.getLogger('nndct') logger.setLevel(1) # Override findCaller on the logger to skip internal helper functions logger.findCaller = _logger_find_caller # Don't further configure the TensorFlow logger if the root logger is # already configured. This prevents double logging in those cases. if not _logging.getLogger().handlers: # Determine whether we are in an interactive environment _interactive = False try: # This is only defined in interactive shells. if _sys.ps1: _interactive = True except AttributeError: # Even now, we may be in an interactive shell with `python -i`. _interactive = _sys.flags.interactive # If we are in an interactive environment (like Jupyter), set loglevel # to INFO and pipe the output to stdout. if _interactive: #logger.setLevel(INFO) _logging_target = _sys.stdout else: _logging_target = _sys.stderr # Add the output handler. _handler = _logging.StreamHandler(_logging_target) _handler.setFormatter(_logging.Formatter(_logging_fmt, None)) logger.addHandler(_handler) _logger = logger return _logger finally: _logger_lock.release() def info(msg, *args, **kwargs): extra = {'nndct_prefix': _log_prefix(INFO)} get_logger().info(msg, *args, extra=extra, **kwargs)
null
23,249
from __future__ import absolute_import from __future__ import division from __future__ import print_function import logging as _logging import os as _os import sys as _sys import time as _time import traceback as _traceback from logging import DEBUG from logging import ERROR from logging import FATAL from logging import INFO from logging import WARN from logging import NOTSET import threading def _log_prefix(level, timestamp=None, file_and_line=None): """Generate a nndct logline prefix.""" # pylint: disable=global-variable-not-assigned global _level_names # pylint: enable=global-variable-not-assigned # Record current time now = timestamp or _time.time() now_tuple = _time.localtime(now) now_microsecond = int(1e6 * (now % 1.0)) (filename, line) = file_and_line or _get_file_and_line() basename = _os.path.basename(filename) # Severity string severity = 'I' if level in _level_names: severity = _level_names[level][0] s = '%c%02d%02d %02d:%02d:%02d.%06d %s]' % ( severity, now_tuple[1], # month now_tuple[2], # day now_tuple[3], # hour now_tuple[4], # min now_tuple[5], # sec now_microsecond, basename) return s def get_logger(): """Return logger instance.""" global _logger # Use double-checked locking to avoid taking lock unnecessarily. if _logger: return _logger _logger_lock.acquire() try: if _logger: return _logger # Scope the TensorFlow logger to not conflict with users' loggers. logger = _logging.getLogger('nndct') logger.setLevel(1) # Override findCaller on the logger to skip internal helper functions logger.findCaller = _logger_find_caller # Don't further configure the TensorFlow logger if the root logger is # already configured. This prevents double logging in those cases. if not _logging.getLogger().handlers: # Determine whether we are in an interactive environment _interactive = False try: # This is only defined in interactive shells. if _sys.ps1: _interactive = True except AttributeError: # Even now, we may be in an interactive shell with `python -i`. _interactive = _sys.flags.interactive # If we are in an interactive environment (like Jupyter), set loglevel # to INFO and pipe the output to stdout. if _interactive: #logger.setLevel(INFO) _logging_target = _sys.stdout else: _logging_target = _sys.stderr # Add the output handler. _handler = _logging.StreamHandler(_logging_target) _handler.setFormatter(_logging.Formatter(_logging_fmt, None)) logger.addHandler(_handler) _logger = logger return _logger finally: _logger_lock.release() def error(msg, *args, **kwargs): extra = {'nndct_prefix': _log_prefix(ERROR)} get_logger().error(msg, *args, extra=extra, **kwargs)
null
23,250
from __future__ import absolute_import from __future__ import division from __future__ import print_function import logging as _logging import os as _os import sys as _sys import time as _time import traceback as _traceback from logging import DEBUG from logging import ERROR from logging import FATAL from logging import INFO from logging import WARN from logging import NOTSET import threading def _log_prefix(level, timestamp=None, file_and_line=None): def get_logger(): def fatal(msg, *args, **kwargs): extra = {'nndct_prefix': _log_prefix(FATAL)} get_logger().fatal(msg, *args, extra=extra, **kwargs)
null
23,253
from __future__ import absolute_import from __future__ import division from __future__ import print_function import logging as _logging import os as _os import sys as _sys import time as _time import traceback as _traceback from logging import DEBUG from logging import ERROR from logging import FATAL from logging import INFO from logging import WARN from logging import NOTSET import threading def log(level, msg, *args, **kwargs): extra = {'nndct_prefix': _log_prefix(level)} get_logger().log(level, msg, *args, extra=extra, **kwargs) def min_vlog_level(): global _min_vlog_level if _min_vlog_level is None: try: _min_vlog_level = int(_os.getenv('VAI_MIN_VLOG_LEVEL', 0)) except ValueError: _min_vlog_level = 0 return _min_vlog_level def vlog(level, msg, *args, **kwargs): if level <= min_vlog_level(): log(level, msg, *args, **kwargs)
null
23,255
import collections import json import os from nndct_shared.pruning.pruning_lib import PruningSpec, NodeGroup from nndct_shared.pruning import errors from nndct_shared.utils import io from typing import List class SubnetSearcher(object): def __init__(self, groups: List[NodeGroup]): self._groups = groups self._supernet = None self._subnets = [] def set_supernet(self, score, macs=None): self._supernet = SubnetConfig(None, score, macs) def add_subnet(self, ratios, score, macs=None): self._subnets.append(SubnetConfig(ratios, score, macs)) def _sorted_subnet(self): return sorted(self._subnets, key=lambda x: x.score) def subnet(self, index): if not self._subnets: raise errors.OptimizerSubnetError('No subnet candidates.') if index and index >= len(self._subnets): raise errors.OptimizerInvalidArgumentError( 'Subnet index is out of range [0, {}]'.format(len(self._subnets) - 1)) return self._subnets[index] def best_subnet(self): subnets = self._sorted_subnet() base_score = self._supernet.score b0 = abs(base_score - subnets[0].score) b1 = abs(base_score - subnets[-1].score) index = 0 if b0 < b1 else -1 return subnets[index] def spec(self, ratios): return PruningSpec.from_node_groups(self._groups, ratios) def groups(self): return self._groups def config(self): return self._config def supernet(self): return self._supernet def supernet(self, supernet): self._supernet = supernet def serialize(self): subnets = self._sorted_subnet() ratios = [] macs = [] scores = {} for index, subnet in enumerate(subnets): config = subnet.serialize() ratios.append(','.join([str(ratio) for ratio in subnet.ratios])) macs.append(subnet.macs) scores[index] = subnet.score return { 'groups': [g.serialize() for g in self._groups], 'supernet': self._supernet.serialize(), 'ratios': ratios, 'macs': macs, 'scores': scores } def deserialize(cls, data): instance = cls([NodeGroup.deserialize(item) for item in data['groups']]) scores = data['scores'] macs = data['macs'] for index, ratio_str in enumerate(data['ratios']): ratio_strs = ratio_str.split(',') ratios = [float(ratio) for ratio in ratio_strs] instance.add_subnet(ratios, scores[str(index)], macs[index]) instance.supernet = SubnetConfig.deserialize(data['supernet']) return instance def to_json(self): return json.dumps(self.serialize(), indent=2) def load_searcher(filepath): with open(filepath, 'r') as f: return SubnetSearcher.deserialize(json.load(f))
null
23,256
from __future__ import absolute_import from __future__ import division from __future__ import print_function from typing import List, Mapping, Any, Union, Tuple import collections from nndct_shared.base.key_names import NNDCT_OP as OpTypes from nndct_shared.nndct_graph.base_node import Node from nndct_shared.metaclass import Singleton from nndct_shared.nndct_graph.base_graph import Graph from nndct_shared.pruning import errors from nndct_shared.pruning import logging from nndct_shared.pruning import node_group as node_group_lib from nndct_shared.utils import registry TRANSPOSECONV_OPS = [OpTypes.CONVTRANSPOSE2D, OpTypes.CONVTRANSPOSE3D] def is_transpose_conv(op): if op.type in TRANSPOSECONV_OPS: return True return False
null
23,269
from __future__ import absolute_import from __future__ import division from __future__ import print_function from typing import List, Mapping, Any, Union, Tuple import collections from nndct_shared.base.key_names import NNDCT_OP as OpTypes from nndct_shared.nndct_graph.base_node import Node from nndct_shared.metaclass import Singleton from nndct_shared.nndct_graph.base_graph import Graph from nndct_shared.pruning import errors from nndct_shared.pruning import logging from nndct_shared.pruning import node_group as node_group_lib from nndct_shared.utils import registry def find_prunable_ancestor(graph, node, target_ops=CONV_OPS): return find_ancestor(graph, node, target_ops, [OpTypes.CONCAT]) def modify_concat(graph, node, pruning_res): out_dim_missing = False for tensor in node.in_tensors: input_node = tensor.node input_pruning = pruning_res[input_node.name] if not input_pruning.has_out_dim(): out_dim_missing = True break node_pruning = pruning_res[node.name] cur_offset = 0 out_dim = 0 removed_outputs = [] for tensor in node.in_tensors: input_node = tensor.node input_pruning = pruning_res[input_node.name] if input_pruning.removed_outputs and out_dim_missing: upstream_conv = find_prunable_ancestor(graph, input_node) raise errors.OptimizerNotExcludeNodeError( 'Must exclude node from pruning: {}.'.format(upstream_conv.name)) if not out_dim_missing: for ro in input_pruning.removed_outputs: removed_outputs.append(ro + cur_offset) out_dim += input_pruning.out_dim cur_offset += (len(input_pruning.removed_outputs) + input_pruning.out_dim) node_pruning.removed_outputs = removed_outputs node_pruning.out_dim = out_dim # update removed_inputs & in_dim node_pruning.removed_inputs = node_pruning.removed_outputs node_pruning.in_dim = node_pruning.out_dim
null
23,271
from collections import deque def graph_search_handler(start_node, generator, frontier, handler=None, gen_params={}): frontier.append(start_node) explored = set() while frontier: node = frontier.pop() explored.add(node) if handler: handler(node) frontier.extend([ n for n in generator(node, **gen_params) if n not in explored and n not in frontier ]) return None class FIFOQueue(Queue): """A First-In-First-Out Queue implemented with collections.deque MODIFIED FROM AIMA VERSION - Use deque - Use an additional dict to track membership """ def __init__(self): self.A = deque() self.__keys = set() def append(self, item): self.A.append(item) self.__keys.add(item) def __len__(self): return len(self.A) def pop(self): key = self.A.popleft() self.__keys.discard(key) return key def __contains__(self, item): return item in self.__keys def breadth_first_search_handler(start_node, generator, handler=None, gen_params={}): return graph_search_handler( start_node, generator, frontier=FIFOQueue(), handler=handler, gen_params=gen_params)
null
23,276
import sys from typing import List, Optional from nndct_shared.base import NNDCT_OP from nndct_shared.nndct_graph import Graph, Operation, Node, Block from nndct_shared.nndct_graph import operator_definition as base_op from nndct_shared.utils import NndctOption, NndctScreenLogger def convert_graph_to_block_node(top_graph, graph): op = base_op.CustomOp(NNDCT_OP.BLOCK) block_node = Node(name=graph.name, op=op, in_quant_part=True) block_node.owning_block = top_graph.block block_node.owning_graph = top_graph block_node.add_block(graph.block) for block in graph.all_blocks(): block.owning_graph = top_graph for tensor in graph.tensors: top_graph.add_tensor(tensor) for node in graph.all_nodes(): node._idx = -1 node.owning_graph = top_graph for param_name in graph.param_names(): top_graph.add_param_name(param_name) return block_node def merge_multi_graphs_to_single_graph(graphs, graph_name="Nndctgraph"): top_graph = Graph(graph_name) op = base_op.CustomOp(NNDCT_OP.PLACEHOLDER) input_node = Node(name="input_placeholder", op=op, in_quant_part=False) input_node.owning_graph = top_graph op = base_op.CustomOp(NNDCT_OP.PLACEHOLDER) return_node = Node(name="return_placeholder", op=op, in_quant_part=False) return_node.owning_graph = top_graph top_block = Block(top_graph, None, input_node, return_node) top_graph.set_top_block(top_block) for graph in graphs: block_node = convert_graph_to_block_node(top_graph, graph) if not block_node.in_node_list(): top_graph.append_node(block_node) return top_graph
null
23,280
import math import itertools from typing import List, Dict, Any, NoReturn, Tuple import numpy as np from functools import partial from nndct_shared.base import NNDCT_OP, NNDCT_KEYS from nndct_shared.nndct_graph import Tensor, Node from .xgraph import XGraph from nndct_shared.utils import calculate_op_scale, DataXopError NndctQuantInfo = Dict[str, Dict[str, List[int]]] class _Converter: _nndct2xir_type = {np.float32: "FLOAT32", np.float64: "FLOAT64", np.int64: "INT64", np.int32: "INT32", } _nndct2numpy_type = { "float32": np.float32, "float64": np.float64, "int32": np.int32, "int64": np.int64 } _pad_mode = {"pad_mode": {0: "FLOOR", 1: "CEIL", 2: "SAME", 3: "VALID"} } _nndct2xir_value = {NNDCT_OP.CONV2D: _pad_mode, NNDCT_OP.DEPTHWISE_CONV2D: _pad_mode, NNDCT_OP.CONVTRANSPOSE2D: _pad_mode, NNDCT_OP.DEPTHWISE_CONVTRANSPOSE2D: _pad_mode, NNDCT_OP.MAX_POOL: _pad_mode, NNDCT_OP.MAX_POOL1D: _pad_mode, NNDCT_OP.AVG_POOL: _pad_mode, NNDCT_OP.PAD: {"mode": {0: "CONSTANT", 1: "REFLECT", 2: "SYMMETRIC"}} } def to_xir_dtype(cls, numpy_dtype): return cls._nndct2xir_type[numpy_dtype] def to_xir_dtype_by_string(cls, dtype): return { "float32" : "FLOAT32", "float64": "FLOAT64", "int32" : "INT32", "int64" : "INT64" }.get(dtype, dtype) def to_xir_attr_value(cls, node_op_type, nndct_attr_name: str, nndct_attr_value: Any): if node_op_type not in cls._nndct2xir_value or nndct_attr_name not in cls._nndct2xir_value[node_op_type]: return nndct_attr_value else: return cls._nndct2xir_value[node_op_type][nndct_attr_name][nndct_attr_value] def to_numpy_dtype(cls, nndct_dtype): return cls._nndct2numpy_type[nndct_dtype] def shape(xgraph: XGraph, node: Node, quant_config: NndctQuantInfo) -> NoReturn: r""" nndct shape is a macro operator, including shape, stridedslice """ # raise NotImplementedError("shape") input_list = [] shape_input_ops: Dict[str, List["xir.Op"]] = {} for input in node.in_nodes: input_op = xgraph.get_op_by_name(input) input_list.append(input_op) shape_input_ops["input"] = input_list sub_op_shape = xgraph.create_fixed_normal_op( node.name + "_i0", "shape", quant_config, input_ops=shape_input_ops) attrs: Dict[str, Any] = {} strided_slice_input_ops: Dict[str, List["xir.Op"]] = {} strided_slice_input_ops["input"] = [sub_op_shape] dim = node.node_attr(node.op.AttrName.AXIS) attrs["begin"] = [dim] attrs["end"] = [dim + 1] xgraph.create_fixed_normal_op( node.name, "strided_slice", quant_config, attrs=attrs, input_ops=strided_slice_input_ops) def zeros(xgraph: XGraph, node: Node, quant_config: NndctQuantInfo) -> NoReturn: shape = node.node_attr(node.op.AttrName.SHAPE) data = np.zeros(shape, dtype=_Converter.to_numpy_dtype(node.out_tensors[0].dtype)) xgraph.create_fixed_const_op(name=node.name, data=data, quant_info=quant_config) class XGraph(object): def __init__(self, name: str): self._graph = Graph(name) self._const_ops: Dict[str, Op] = {} self._ops: Dict[str, Op] = {} def _check_inputs(self, input_ops): if any([ip is None for ip in input_ops]): raise RuntimeError('The input op is `None`, please check graph.') def create_const_op(self, name: str, data: Optional[np.ndarray]) -> NoReturn: if data is not None and data.ndim == 0: data = np.array([data], data.dtype) const_op = self._graph.create_const_op(name, data) if name in self._const_ops: raise RuntimeError('The const op {} has already in graph'.format(name)) return const_op def create_input_transpose_ops(self, input_list: List[Op], input_tensors: 'List[base_tensor::Tensor]'): t_ops = [] for i, (input, tensor) in enumerate(zip(input_list, input_tensors)): if tensor.ndim in [4, 5]: attrs: Dict[str, Any] = {} attrs['order'] = [0, 3, 1, 2] if tensor.ndim == 4 else [0, 4, 3, 1, 2] input_ops: Dict[str, List[Op]] = {} input_ops['input'] = [input] op_name = input.get_name() + NNDCT_KEYS.TRANSPOSE_OP_SUFFIX t_op = self.get_op_by_name(op_name) if t_op is None: t_op = self.create_normal_op( op_name, 'transpose', attrs=attrs, input_ops=input_ops) t_ops.append(t_op) else: t_ops.append(input) return t_ops def create_normal_op(self, name: str, kind: str, tensor: Optional[np.ndarray] = None, attrs: Optional[Dict[str, Any]] = None, input_ops: Optional[List[Op]] = None) -> Op: if input_ops is not None: self._check_inputs(input_ops['input']) op = self._graph.create_op(name, kind, attrs, input_ops) return op def create_fixed_const_op(self, name: str, data: np.ndarray, quant_info: NndctQuantInfo) -> Op: formal_name = re.sub(_XMODEL_NAME_PATTERN, "_", name) const_op = self.create_const_op(formal_name, data) # print(const_op.get_name(), "const", const_op.get_output_tensor().dims) fixed_const_op = self.create_fix_op(const_op, name, quant_info) return fixed_const_op if fixed_const_op else const_op def create_fixed_normal_op(self, name: str, kind: str, quant_info: NndctQuantInfo, tensor: Optional[np.ndarray] = None, attrs: Optional[Dict[str, Any]] = None, input_ops: Optional[List[Op]] = None) -> Op: formal_name = re.sub(_XMODEL_NAME_PATTERN, "_", name) op = self.create_normal_op( name=formal_name, kind=kind, tensor=tensor, attrs=attrs, input_ops=input_ops) # print(op.get_name(), kind, op.get_output_tensor().dims) post_fixed_op = self.create_fix_op(op, name, quant_info) return post_fixed_op if post_fixed_op else op def create_input_fix_ops(self, input_list: List[Op], key_name: str, quant_info: NndctQuantInfo): pre_fix_ops = [] for i, op in enumerate(input_list): pre_fix_op = self.create_fix_op(op, key_name, quant_info, id=i, post_fix=False) if pre_fix_op: pre_fix_ops.append(pre_fix_op) else: pre_fix_ops.append(op) return pre_fix_ops def create_fix_op(self, input: Op, key_name: str, quant_info: NndctQuantInfo, id: Optional[int] = None, post_fix: bool = True) -> Optional[Op]: def _get_fix_info(name: str, quant_info: NndctQuantInfo) -> Sequence[int]: if post_fix: combinded_fix_infos = ChainMap(dict(quant_info['param']), dict(quant_info['output'])) else: combinded_fix_infos = quant_info['input'] if name in combinded_fix_infos.keys(): return combinded_fix_infos[name] else: return None # if NNDCT_KEYS.FIX_OP_SUFFIX in input.get_name(): # raise RuntimeError("The consecutive fix ops in graph is forbidden!") if not isinstance(quant_info, dict): return None # bit_width, fix_point = _get_fix_info(key_name, quant_info)[0] # if bit_width is None or fix_point is None: # return None bit_info = _get_fix_info(key_name, quant_info) if bit_info is None or bit_info[0] is None: return None bit_width, fix_point = bit_info[0] if bit_width is None or fix_point is None: return None attrs: Dict[str, Any] = {} attrs['fix_point'] = fix_point attrs['bit_width'] = bit_width attrs['round_mode'] = "DPU_ROUND" attrs['if_signed'] = True input_ops: Dict[str, List[Op]] = {} input_ops['input'] = [input] if post_fix: op_name = input.get_name() + NNDCT_KEYS.FIX_OP_SUFFIX else: formal_name = re.sub(_XMODEL_NAME_PATTERN, "_", key_name) if id is not None: op_name = formal_name + NNDCT_KEYS.PRE_FIX_OP_SUFFIX + f"_i{id}" else: op_name = formal_name + NNDCT_KEYS.PRE_FIX_OP_SUFFIX fix_op = self.create_normal_op( op_name, 'fix', attrs=attrs, input_ops=input_ops) return fix_op def get_op_by_name(self, name: str) -> Op: formal_name = re.sub(_XMODEL_NAME_PATTERN, "_", name) op = self._graph.get_op(formal_name + NNDCT_KEYS.FIX_OP_SUFFIX) if op is None: op = self._graph.get_op(formal_name) return op def get_op_output_shape(self, name: str) -> List[int]: op = self.get_op_by_name(name) if op: return op.get_output_tensor().dims else: NndctScreenLogger().warning2user(QWarning.EXPORT_XMODEL, f"{name} is not in xmodel. Please check it.") def export_to_xmodel(self, fname: str) -> NoReturn: fname += NNDCT_KEYS.XMODEL_SUFFIX try: self._graph.serialize(fname) except Exception: raise ExportXmodelError(self._graph.get_name()) else: NndctScreenLogger().info(f"=>Successfully convert '{self._graph.get_name()}' to xmodel.({fname})") def export_to_img(self, fname: str) -> NoReturn: fname += NNDCT_KEYS.XMODEL_IMAGE_SUFFIX try: shell_command = "which dot" proc = subprocess.Popen(shell_command, stdout=subprocess.PIPE, shell=True) try: outs, errs = proc.communicate(timeout=2) except subprocess.TimeoutExpired: proc.kill() outs, errs = proc.communicate() NndctScreenLogger().error2user(QError.PROC_TIMEOUT, f"{errs}") raise if outs: self._graph.save_as_dot(fname) else: NndctScreenLogger().warning2user(QWarning.EXPORT_XMODEL, f"Can't find dot command in the system, please install it." \ " Otherwise, the xmodel image will not be generated.") except Exception as e: NndctScreenLogger().warning2user(QWarning.EXPORT_XMODEL, f"Failed to generate xmodel image!({str(e)})") def graph(self): return self._graph def data_xop(xgraph: XGraph, node: Node, quant_config: NndctQuantInfo) -> NoReturn: shape = node.out_tensors[0].shape if not shape: shape = [1] if shape[0] == 0: raise DataXopError("data", shape) # shape = permute_axes(shape, node.transpose_out_order) try: out_tensor = np.zeros(shape, dtype=np.float32) attrs: Dict[str, Any] = {} attrs["shape"] = shape attrs["data_type"] = _Converter.to_xir_dtype(out_tensor.dtype.type) xgraph.create_fixed_normal_op( node.name, "data", quant_config, tensor=out_tensor, attrs=attrs) except: raise DataXopError("data", shape)
null
23,285
import math import itertools from typing import List, Dict, Any, NoReturn, Tuple import numpy as np from functools import partial from nndct_shared.base import NNDCT_OP, NNDCT_KEYS from nndct_shared.nndct_graph import Tensor, Node from .xgraph import XGraph from nndct_shared.utils import calculate_op_scale, DataXopError NndctQuantInfo = Dict[str, Dict[str, List[int]]] class XGraph(object): def __init__(self, name: str): def _check_inputs(self, input_ops): def create_const_op(self, name: str, data: Optional[np.ndarray]) -> NoReturn: def create_input_transpose_ops(self, input_list: List[Op], input_tensors: 'List[base_tensor::Tensor]'): def create_normal_op(self, name: str, kind: str, tensor: Optional[np.ndarray] = None, attrs: Optional[Dict[str, Any]] = None, input_ops: Optional[List[Op]] = None) -> Op: def create_fixed_const_op(self, name: str, data: np.ndarray, quant_info: NndctQuantInfo) -> Op: def create_fixed_normal_op(self, name: str, kind: str, quant_info: NndctQuantInfo, tensor: Optional[np.ndarray] = None, attrs: Optional[Dict[str, Any]] = None, input_ops: Optional[List[Op]] = None) -> Op: def create_input_fix_ops(self, input_list: List[Op], key_name: str, quant_info: NndctQuantInfo): def create_fix_op(self, input: Op, key_name: str, quant_info: NndctQuantInfo, id: Optional[int] = None, post_fix: bool = True) -> Optional[Op]: def _get_fix_info(name: str, quant_info: NndctQuantInfo) -> Sequence[int]: def get_op_by_name(self, name: str) -> Op: def get_op_output_shape(self, name: str) -> List[int]: def export_to_xmodel(self, fname: str) -> NoReturn: def export_to_img(self, fname: str) -> NoReturn: def graph(self): def dense(xgraph: XGraph, node: Node, quant_config: NndctQuantInfo) -> NoReturn: input_ops: Dict[str, List["xir.Op"]] = {} for param_name, param_tensor in node.op.params.items(): if param_name == node.op.ParamName.WEIGHTS: weights = xgraph.get_op_by_name(param_tensor.name) else: bias = xgraph.get_op_by_name(param_tensor.name) input_ops["bias"] = [bias] input_list = [] for input in node.in_nodes: input_op = xgraph.get_op_by_name(input) input_list.append(input_op) input_ops["input"] = xgraph.create_input_fix_ops(input_list, node.name, quant_config) input_ops["input"].append(weights) attrs: Dict[str, Any] = {} attrs["transpose_a"] = False attrs["transpose_b"] = True xgraph.create_fixed_normal_op( node.name, "matmul", quant_config, attrs=attrs, input_ops=input_ops)
null
23,289
import math import itertools from typing import List, Dict, Any, NoReturn, Tuple import numpy as np from functools import partial from nndct_shared.base import NNDCT_OP, NNDCT_KEYS from nndct_shared.nndct_graph import Tensor, Node from .xgraph import XGraph from nndct_shared.utils import calculate_op_scale, DataXopError NndctQuantInfo = Dict[str, Dict[str, List[int]]] def _get_xir_attr_from_node(node: Node): def scale(xgraph, node, quant_config): class XGraph(object): def __init__(self, name: str): def _check_inputs(self, input_ops): def create_const_op(self, name: str, data: Optional[np.ndarray]) -> NoReturn: def create_input_transpose_ops(self, input_list: List[Op], input_tensors: 'List[base_tensor::Tensor]'): def create_normal_op(self, name: str, kind: str, tensor: Optional[np.ndarray] = None, attrs: Optional[Dict[str, Any]] = None, input_ops: Optional[List[Op]] = None) -> Op: def create_fixed_const_op(self, name: str, data: np.ndarray, quant_info: NndctQuantInfo) -> Op: def create_fixed_normal_op(self, name: str, kind: str, quant_info: NndctQuantInfo, tensor: Optional[np.ndarray] = None, attrs: Optional[Dict[str, Any]] = None, input_ops: Optional[List[Op]] = None) -> Op: def create_input_fix_ops(self, input_list: List[Op], key_name: str, quant_info: NndctQuantInfo): def create_fix_op(self, input: Op, key_name: str, quant_info: NndctQuantInfo, id: Optional[int] = None, post_fix: bool = True) -> Optional[Op]: def _get_fix_info(name: str, quant_info: NndctQuantInfo) -> Sequence[int]: def get_op_by_name(self, name: str) -> Op: def get_op_output_shape(self, name: str) -> List[int]: def export_to_xmodel(self, fname: str) -> NoReturn: def export_to_img(self, fname: str) -> NoReturn: def graph(self): def hsigmoid(xgraph: XGraph, node: Node, quant_config: NndctQuantInfo) -> NoReturn: scale = 6.0 * 2731.0 / 16384.0 attrs = _get_xir_attr_from_node(node) input_ops: Dict[str, List["xir.Op"]] = {} input_ops["input"] = [xgraph.get_op_by_name(node.in_nodes[0])] input_ops["input"] = xgraph.create_input_fix_ops(input_ops["input"], node.name, quant_config) xgraph.create_fixed_normal_op( node.name + "_i0", "hard-sigmoid", quant_config, attrs=attrs, input_ops=input_ops) scale = [scale] xgraph.create_fixed_const_op(name=node.name + "_i1", data=np.array(scale, dtype=np.float32), quant_info=quant_config) input_ops: Dict[str, List["xir.Op"]] = {} input_ops["input"] = [xgraph.get_op_by_name(node.name + "_i0"), xgraph.get_op_by_name(node.name + "_i1")] xgraph.create_fixed_normal_op( node.name, "mul", quant_config, input_ops=input_ops)
null
23,300
from collections import defaultdict from nndct_shared.utils import NndctDebugLogger, NndctOption def log_debug_info(msg): if NndctOption.nndct_inspect_debug.value: NndctDebugLogger.write(f"{msg}\n")
null
23,301
import copy import networkx as nx from networkx.algorithms import is_isomorphic from nndct_shared.base import NNDCT_OP from nndct_shared.inspector.utils import build_xir_nndct_op_map, log_debug_info from nndct_shared.compile.xir_helper import XIRHelper from .graph import Graph, Node _SIMULATION_PATTERNS = [ {"name": "conv2d_fix_with_hardwish", "nodes":[("weights", Node({"fix2float"})), ("bias", Node({"fix2float"})), ("input", Node({"fix2float"})), ("conv2d", Node({"conv2d", "matmul", "scale"})), ("conv2d_out", Node({"float2fix"})), ("hsigmoid_in", Node({"fix2float"})), ("hsigmoid", Node({"hard-sigmoid"})), ("mul", Node({"mul"})), ("hsigmoid_out", Node({"float2fix"})), ("hswish_i0", Node({"fix2float"})), ("hswish_i1", Node({"fix2float"})), ("hswish", Node({"mul"})), ("output", Node({"float2fix"})), ], "edges": [("weights", "conv2d"), ("bias", "conv2d"), ("input", "conv2d"), ("conv2d", "conv2d_out"), ("conv2d_out", "hsigmoid_in"), ("hsigmoid_in", "hsigmoid"), ("hsigmoid", "mul"), ("mul", "hsigmoid_out"), ("conv2d_out", "hswish_i0"), ("hsigmoid_out", "hswish_i1"), ("hswish_i0", "hswish"), ("hswish_i1", "hswish"), ("hswish", "output"), ] }, {"name": "conv2d_fix_with_hardsigmoid", "nodes":[("weights", Node({"fix2float"})), ("bias", Node({"fix2float"})), ("input", Node({"fix2float"})), ("conv2d", Node({"conv2d", "matmul", "scale"})), ("conv2d_out", Node({"float2fix"})), ("hsigmoid_in", Node({"fix2float"})), ("hsigmoid", Node({"hard-sigmoid"})), ("mul", Node({"mul"})), ("output", Node({"float2fix"})) ], "edges": [("weights", "conv2d"), ("bias", "conv2d"), ("input", "conv2d"), ("conv2d", "conv2d_out"), ("conv2d_out", "hsigmoid_in"), ("hsigmoid_in", "hsigmoid"), ("hsigmoid", "mul"), ("mul", "output")] }, {"name": "conv2d_fix_with_relu", "nodes":[("weights", Node({"fix2float"})), ("bias", Node({"fix2float"})), ("input", Node({"fix2float"})), # ("conv2d", Node({"conv2d", "matmul", "depthwise-conv2d", "transposed-conv2d", "transposed-depthwise-conv2d", "scale"})), ("conv2d", Node({"matmul", "scale"})), ("relu", Node({"relu", "prelu", "leaky-relu", "relu6"})), ("output", Node({"float2fix"})) ], "edges": [("weights", "conv2d"), ("bias", "conv2d"), ("input", "conv2d"), ("conv2d", "relu"), ("relu", "output")] }, {"name": "conv2d_fix_without_relu", "nodes": [("weights", Node({"fix2float"})), ("bias", Node({"fix2float"})), ("input", Node({"fix2float"})), # ("conv2d", Node({"conv2d", "matmul", "depthwise-conv2d", "transposed-conv2d", "scale"})), ("conv2d", Node({"matmul", "scale"})), ("output", Node({"float2fix"})) ], "edges": [("weights", "conv2d"), ("bias", "conv2d"), ("input", "conv2d"), ("conv2d", "output")] }, { "name": "reduction_mean_with_mul_relu", "nodes": [ ("input", Node({"fix2float"})), ("reduction_mean", Node({"reduction_mean"})), ("const", Node({"const"})), ("mul", Node({"mul"})), ("relu", Node({"relu","prelu", "leaky-relu", "relu6"})), ("output", Node({"float2fix"})), ], "edges": [("input", "reduction_mean"), ("reduction_mean", "mul"), ("const", "mul"), ("mul", "relu"), ("relu", "output")] }, { "name": "reduction_mean_with_mul", "nodes": [ ("input", Node({"fix2float"})), ("reduction_mean", Node({"reduction_mean"})), ("const", Node({"const"})), ("mul", Node({"mul"})), ("output", Node({"float2fix"})), ], "edges": [("input", "reduction_mean"), ("reduction_mean", "mul"), ("const", "mul"), ("mul", "output")] }, # { # "name": "reduction_mean_with_relu", # "nodes": [ # ("input", Node({"fix2float"})), # ("reduction_mean", Node({"reduction_mean"})), # ("relu", Node({"relu","prelu", "leaky-relu", "relu6"})), # ("output", Node({"float2fix"})), # ], # "edges": [("input", "reduction_mean"), ("reduction_mean", "relu"), ("relu", "output")] # }, # {"name": "pool_with_mul", # "nodes": [("input", Node({"fix2float"})), # ("pool", Node({"avgpool2d"})), # ("const", Node({"const"})), # ("mul", Node({"mul"})), # ("output", Node({"float2fix"})) # ], # "edges": [("input", "pool"), ("pool", "mul"), ("const", "mul"), ("mul", "output")] # }, # {"name": "pool_fix", # "nodes": [("input", Node({"fix2float"})), # ("pool", Node({"avgpool2d", "maxpool2d"})), # ("output", Node({"float2fix"})) # ], # "edges": [("input", "pool"), ("pool", "output")] # }, # { "name": "eltwise_fix_with_relu", # "nodes": [("add", Node({"add", "mul"})), # ("relu", Node({"relu"})), # ("output", Node({"float2fix"})) # ], # "edges": [("add", "relu"), ("relu", "output")] # }, # { "name": "eltwise_fix", # "nodes": [("add", Node({"add", "mul"})), # ("output", Node({"float2fix"})) # ], # "edges": [("add", "output")] # }, # { "name": "concat_fix", # "nodes": [("concat", Node({"concat"})), # ("output", Node({"float2fix"})) # ], # "edges": [("concat", "output")] # }, # { # "name": "resize_fix", # "nodes": [("input", Node({"fix2float"})), # ("resize", Node({"resize"})), # ("output", Node({"float2fix"})) # ], # "edges": [("input", "resize"), ("resize", "output")] # }, # { # "name": "pad_fix", # "nodes": [("input", Node({"fix2float"})), # ("pad", Node({"pad"})), # ("output", Node({"float2fix"})) # ], # "edges": [("input", "pad"), ("pad", "output")] # }, # { # "name": "reduction_max_fix", # "nodes": [("input", Node({"fix2float"})), # ("reduction_max", Node({"reduction_max"})), # ("output", Node({"float2fix"})) # ], # "edges": [("input", "reduction_max"), ("reduction_max", "output")], # }, # { # "name": "reshape_fix", # "nodes": [("input", Node({"fix2float"})), # ("reshape", Node({"reshape"})), # ("output", Node({"float2fix"})) # ], # "edges": [("input", "reshape"), ("reshape", "output")], # }, # { # "name": "hsigmoid_fix", # "nodes": [("input", Node({"fix2float"})), # ("hsigmoid", Node({"hard-sigmoid"})), # ("output", Node({"float2fix"})) # ], # "edges": [("input", "hsigmoid"), ("hsigmoid", "output")], # }, # { # "name": "reduction_mean", # "nodes": [("input", Node({"fix"})), # ("reduction_mean", Node({"reduction_mean"})), # ("output", Node({"fix"})) # ], # "edges": [("input", "reduction_mean"), ("reduction_mean", "output")], # }, # { # "name": "reduce_max", # "nodes": [("input", Node({"fix"})), # ("poollikeop", Node({"reduction_max"})), # ("output", Node({"fix"})) # ], # "edges": [("input", "poollikeop"), ("poollikeop", "output")], # }, ] class Graph(object): def __init__(self, name): def __str__(self): def __eq__(self, other): def node_match(node_1, node_2): def add_node(self, id, obj): def add_edge(self, u, v): def children(self, n): def parents(self, n): def get_node_types(self, id): def set_node_types(self, id, types): def remove_node(self, n): def remove_one_node(self, n): def remove_edge(self, u, v): def visualize(self): def copy(cls, original_graph): def nodes(self): def op_types(self): def graph(self): def name(self): def _gen_pattern_from_sim_pattern(): pattern_graphs = [] patterns = copy.deepcopy(_SIMULATION_PATTERNS) for pattern_info in patterns: pattern_graph = Graph(pattern_info["name"]) for id, attr in pattern_info["nodes"]: pattern_graph.add_node(id, attr) for u, v in pattern_info["edges"]: pattern_graph.add_edge(u, v) pattern_graphs.append(pattern_graph) return pattern_graphs
null
23,303
import copy import networkx as nx from networkx.algorithms import is_isomorphic from nndct_shared.base import NNDCT_OP from nndct_shared.inspector.utils import build_xir_nndct_op_map, log_debug_info from nndct_shared.compile.xir_helper import XIRHelper from .graph import Graph, Node def get_templates_from_dpu_compiler(): def is_valid_pattern(pattern): def reorder_patterns(patterns): def create_pattern_graph(name: str, ops: "List[xir.op_template]"): def convert_xir_type_to_nndct_type(pattern_graph): def transform_pattern_graph(pattern_graph): def log_debug_info(msg): class XIRHelper(object): def find_xops_from_nndct_node(cls, nndct_node, xmodel): def get_xop_device_type(xop): def get_xop_name(xop): def get_xop_template_name(op_template): def get_xop_template_types(op_template): def get_xmodel_ops(xmodel): def get_xop_type(xop): def get_input_xops(xop): def get_op_partition_msg(xop): def is_dpu_pattern(cls, xmodel): def get_pattern_partition_msg(cls, xmodel): def is_valid_compiled_pattern(cls, xmodel): def build_patterns_from_dpu_templates(): templates = get_templates_from_dpu_compiler() log_debug_info("\nAll patterns from xcompiler:") for id, (name, ops) in enumerate(templates): log_debug_info(f"pattern id:{id}") for op in ops: log_debug_info(f"op name:{XIRHelper.get_xop_template_name(op)} type:{XIRHelper.get_xop_template_types(op)}") patterns = [] pattern_graphs = [] for id, (name, ops) in enumerate(templates): pattern_graph = create_pattern_graph(f"{name}_{id}", ops) ret, msg = is_valid_pattern(pattern_graph) if ret: pattern_graphs.append(pattern_graph) else: log_debug_info(f"{pattern_graph.name} is filtered.({msg}).") # pattern_graphs = pattern_graphs + _gen_pattern_from_sim_pattern() log_debug_info("\nPattern Transformation:") for pattern_graph in pattern_graphs: log_debug_info(f"{pattern_graph.name} pattern") log_debug_info("================Before transformation====================") log_debug_info(str(pattern_graph)) transform_pattern_graph(pattern_graph) if convert_xir_type_to_nndct_type(pattern_graph): patterns.append(pattern_graph) else: log_debug_info(f"{pattern_graph.name} is ignored for there is at least one unknown op in the pattern.") log_debug_info("================After transformation====================") log_debug_info(str(pattern_graph)) patterns = reorder_patterns(patterns) return patterns
null
23,304
import copy import networkx as nx from networkx.algorithms import is_isomorphic from nndct_shared.base import NNDCT_OP from nndct_shared.inspector.utils import build_xir_nndct_op_map, log_debug_info from nndct_shared.compile.xir_helper import XIRHelper from .graph import Graph, Node class Graph(object): def __init__(self, name): def __str__(self): def __eq__(self, other): def node_match(node_1, node_2): def add_node(self, id, obj): def add_edge(self, u, v): def children(self, n): def parents(self, n): def get_node_types(self, id): def set_node_types(self, id, types): def remove_node(self, n): def remove_one_node(self, n): def remove_edge(self, u, v): def visualize(self): def copy(cls, original_graph): def nodes(self): def op_types(self): def graph(self): def name(self): def drop_fix_in_pattern(pattern_graph): fix = {NNDCT_OP.FIX} removed_node = [] pattern_without_fix = Graph.copy(pattern_graph) for node in pattern_graph.nodes: if pattern_without_fix.get_node_types(node) == fix: removed_node.append(node) for node in removed_node: pattern_without_fix.remove_one_node(node) return pattern_without_fix
null
23,305
from typing import Mapping from nndct_shared.nndct_graph.base_graph import Graph from nndct_shared.nndct_graph.base_node import Node from nndct_shared.expanding.op_modifier import op_modifier from nndct_shared.expanding.spec import DataInsert, GenericStructuredExpanding, StructuredExpanding from nndct_shared.expanding.op_modifier import op_modifier def propagate_node_expanding(node, node_expand_desc: Mapping[str, StructuredExpanding]): node_expanding = node_expand_desc[node.name] assert isinstance(node_expanding, GenericStructuredExpanding), \ "Variable node_expanding here has to be instance of GenericStructuredExpanding" input_expanding = node_expand_desc[node.in_nodes[0]] node_expanding.in_dim = input_expanding.out_dim node_expanding.out_dim = node_expanding.in_dim for insert in input_expanding.out_inserts: node_expanding.add_insert(insert) class Graph(GraphBase): """ Graph object of NNDCT, contain list of NndctNodes. That will be used for topology or export to XGraph""" def __init__(self, graph_name=None): super(Graph, self).__init__() self._name = graph_name or 'NndctGraph' self._nodes_by_name = {} self._nodes_by_id = {} self._end_tensors = [] self._copy_tensor_map = {} self._tensors = {} self._blocks = [] self._param_names = [] self._top_block = None def __contains__(self, node_or_name: Union[str, Node]) -> bool: if isinstance(node_or_name, str): return node_or_name in self._nodes_by_name else: return node_or_name.name in self._nodes_by_name def __deepcopy__(self, memo): raise NotImplementedError("Deep copy is prohibited, use `clone_from` instead.") def clone(self): graph = self.__class__(self.name) graph.clone_from(self) return graph def clone_from(self, src_graph): local_map = {} converted_nodes = [] head_node = self.create_node_from(src_graph.head_node, local_map, converted_nodes) return_node = self.create_node_from(src_graph.return_node, local_map, converted_nodes) top_block = Block(self, None, head_node, return_node) self.set_top_block(top_block) self._top_block.clone_from(src_graph.block, local_map, converted_nodes) def create_node_from(self, src_node, local_map, converted_nodes): node = Node(src_node.name, dtype=src_node.dtype, in_quant_part=src_node.in_quant_part) node.owning_graph = self node.idx = src_node.idx node.scope_name = src_node.scope_name node.source_range = src_node.source_range node.target_device = src_node.target_device node.normalized_name = src_node.normalized_name converted_nodes.append(src_node.name) for out in src_node.out_tensors: if out.name in local_map: node.add_out_tensor(local_map[out.name]) else: tensor = Tensor(name=out.name) tensor.clone_from(out) local_map[out.name] = tensor node.add_out_tensor(tensor) for inp in src_node.in_tensors: if inp.name in local_map: node.add_in_tensor(local_map[inp.name]) else: tensor = Tensor(name=inp.name) tensor.clone_from(inp) local_map[inp.name] = tensor node.add_in_tensor(tensor) node.clone_from(src_node, local_map) for src_block in src_node.blocks: head_node = self.create_node_from(src_block.input_node, local_map, converted_nodes) return_node = self.create_node_from(src_block.return_node, local_map, converted_nodes) block = Block(self, node, head_node, return_node) block.clone_from(src_block, local_map, converted_nodes) node.add_block(block) return node def node(self, name): """Return node with the specified name""" return self._nodes_by_name.get(name, None) def get_node_by_idx(self, idx): node = self._nodes_by_id.get(idx, None) assert node is not None return node def get_input_nodes(self): input_nodes = [] for node in self.nodes: if (len(self.parents(node)) == 0) and \ (node.op.type==NNDCT_OP.INPUT or node.op.type==NNDCT_OP.TUPLE_INPUT): input_nodes.append(node) return input_nodes def get_input_tensors(self, input_args): input_tensors = [] graph_name = self.name input_nodes = self.get_input_nodes() for idx in range(len(input_args)): #input_node_name = graph_name + "::input_" + str(idx) #input_node = self.node(input_node_name) input_node = input_nodes[idx] input_tensor = input_node.out_tensors[0] if input_node.op.type == NNDCT_OP.INPUT: input_tensors.append(input_tensor.name) elif input_node.op.type == NNDCT_OP.TUPLE_INPUT: for index in range(len(input_args[idx])): input_tensor_name = input_tensor.name + '.' + str(index) input_tensors.append(input_tensor_name) return input_tensors def get_return_tensors(self): return_tensors = [] for tensor in self.return_node.in_tensors: return_tensors.append(tensor.name) return return_tensors def add_node(self, node: Node) -> None: if node.name in self._nodes_by_name: return if node.idx in self._nodes_by_id and node is not self._nodes_by_id[node.idx]: raise RuntimeError(f"The id `{node.idx}` of {node.name} has been added into graph") if node.idx == -1: # if not self._nodes_by_id: # node._idx = 0 # else: # node._idx = max([node.idx for node in self.all_nodes()]) + 1 node._idx = -sys.maxsize + len(list(self.all_nodes())) self._nodes_by_name[node.name] = node self._nodes_by_id[node.idx] = node def free_node(self, node): node.owning_graph = None self._nodes_by_name.pop(node.name) self._nodes_by_id.pop(node.idx) def remove_node(self, node): assert node.in_tensors assert len(node.out_tensors) == 1 out_tensor = node.out_tensors[0] inp_tensor = node.in_tensors[0] out_tensor.replace_uses_with(inp_tensor) node.destroy() def remove_node_by_types(self, node_types: List[str]) -> Dict[str, str]: if any([node_type in self.op_types for node_type in node_types]): nodes_to_remove = [] for node in self.nodes: if node.op.type in node_types: nodes_to_remove.append(node) for node in nodes_to_remove: self.remove_node(node) def find_nodes_by_types(self, node_types: List[NNDCT_OP]): conv_nodes = [] for node in self.nodes: if node.op.type in node_types: conv_nodes.append(node) else: continue return conv_nodes def reconnect_nodes(self): self._nodes_by_id.clear() for idx, node in enumerate(self.nodes): node.idx = idx self._nodes_by_id[idx] = node node.clean_connections() self.connect_nodes() def connect_nodes(self): for nodeA in self.nodes: for input_tensor in nodeA.in_tensors: for nodeB in self.nodes: if nodeB is not nodeA and input_tensor in nodeB.out_tensors: #nodeB.outputs.add(input_tensor.node.name) nodeB.add_out_node(nodeA.name) nodeA.add_in_node(input_tensor.node.name) def parents(self, node: Union[Node, str]) -> List[Node]: if isinstance(node, str): node = self.node(node) return [self.node(node_name) for node_name in node.in_nodes] def children(self, node: Union[Node, str]) -> List[Node]: if isinstance(node, str): node = self.node(node) return [self.node(node_name) for node_name in node.out_nodes] def add_tensor(self, tensor): self._tensors[tensor.name] = tensor def tensor(self, name): return self._tensors.get(name, None) def param_tensor(self, name): for node in self.all_nodes(): for _, tensor in node.op.params.items(): if tensor.name == name: return tensor def add_end_tensor(self, tensor): self._end_tensors.append(tensor) def __repr__(self): return f"Graph(name={self.name})" def __str__(self): return json.dumps(self.description(), indent=2, separators=(',', ': ')) def description(self): graph_des = {} graph_des['graph_name'] = f"{self.__class__.__name__}" graph_des['nodes'] = [] for n in sorted(self.nodes, key=lambda n: n.idx): graph_des['nodes'].append(n.description()) return graph_des def set_node_id(self, index, node): node.idx = index self._nodes_by_id[index] = node def create_subgraph_from_nodeset(cls, origin_graph, nodeset, graph_name): """ create a subgraph from nodeset belong to origin graph """ assert len(nodeset) >= 2 sorted_nodeset = origin_graph.top_sort_nodeset(nodeset) for node in sorted_nodeset: node.remove_from_list() subgraph = cls(graph_name) sorted_nodeset[0].owning_graph = subgraph sorted_nodeset[-1].owning_graph = subgraph block = Block(subgraph, None, sorted_nodeset[0], sorted_nodeset[-1]) subgraph.set_top_block(block) if len(sorted_nodeset) > 2: for node in sorted_nodeset[1:-1]: node.owning_graph = subgraph subgraph.append_node(node) return subgraph def top_sort_nodeset(nodeset: Sequence[Node]) -> List[Node]: sorted_nodeset = sorted(nodeset, key=lambda n: n.topo_position) return sorted_nodeset def get_topological_graph_nodes_list(self): nodes_list = [node for node in self.nodes] return Graph.top_sort_nodeset(nodes_list) def name(self): return self._name def name(self, name): self._name = name def nodes(self): return self._top_block.nodes def reverse_nodes(self): return self._top_block.reverse_nodes def tensors(self): for tensor in self._tensors.values(): yield tensor # TODO: Remove def end_tensors(self): return [tensor for tensor in self.return_node.in_tensors] def inputs(self): return [node for node in self.all_nodes() if not node.in_nodes] def outputs(self): return [node for node in self.all_nodes() if not node.out_nodes] def op_types(self): return {node.op.type for node in self.all_nodes()} def append_node(self, node): self._top_block.append_node(node) def add_param_name(self, param_name): if param_name not in self._param_names: self._param_names.append(param_name) def param_names(self): return list(self._param_names) def block(self): return self._top_block def is_tensor_in_graph(self, tensor_name): return True if tensor_name in self._tensors else False def update_node_idx(self, node, index): self._nodes_by_id[index] = node def clear_node_id_map(self): self._nodes_by_id.clear() def remove_tensor(self, tensor): self._tensors.pop(tensor.name) if tensor.name in self._param_names: self._param_names.remove(tensor.name) def insert_node_between_nodes(self, new_node, parent_node, child_node): assert parent_node.in_node_list() and child_node.in_node_list() assert (parent_node.owning_graph == child_node.owning_graph and parent_node.owning_block == child_node.owning_block) new_node.owning_block = parent_node.owning_block new_node.owning_graph = parent_node.owning_graph tensor = Tensor(name=new_node.name, node=new_node) new_node.add_out_tensor(tensor) out_tensor = None offset = None for out in parent_node.out_tensors: for use in out.uses: if use.user is child_node: out_tensor = out offset = use.offset break #out_tensor.replace_uses_with(new_node.out_tensors[0]) child_node.replace_input_at(offset, new_node.out_tensors[0]) new_node.add_in_tensor(out_tensor) new_node.insert_after(parent_node) def set_top_block(self, block): self._top_block = block def add_block(self, block): self._blocks.append(block) def all_blocks(self): return self._blocks def all_nodes(self): for _, node in self._nodes_by_name.items(): yield node def head_node(self): return self._top_block.input_node def return_node(self): return self._top_block.return_node def clean_tensors_data(self): for tensor in self.tensors: tensor.clean_data() def assign_node_topological_name(self, prefix="", suffix=""): count = itertools.count(0) illegal_char_regex = re.compile('[^0-9a-zA-Z_]+') def _assgin_nodes(nodes): sorted_nodes = self.top_sort_nodeset(list(nodes)) for n in sorted_nodes: if n.blocks: for block in n.blocks: _assgin_nodes(block.nodes) else: candidate = illegal_char_regex.sub("_", n.op.type) n.normalized_name = f"{prefix}{candidate}_{next(count)}{suffix}" _assgin_nodes(self.nodes) def simple_description(self): """ Only describe op type topological info. """ def get_node_simple_info(node): node_des = {} node_des['op'] = node.op.type node_des["input_ops"] = [node.owning_graph.node(inode).op.type for inode in node.in_nodes] node_des["output_ops"] = [node.owning_graph.node(onode).op.type for onode in node.out_nodes] if node.blocks: for i, block in enumerate(node.blocks): node_des[f'block_{i}'] = [] for n in self.top_sort_nodeset(list(block.nodes)): node_des[f'block_{i}'].append(get_node_simple_info(n)) return node_des graph_des = {} graph_des['nodes'] = [] for n in self.top_sort_nodeset(list(self.nodes)): graph_des['nodes'].append(get_node_simple_info(n)) graph_str = json.dumps(graph_des, indent=2, separators=(',', ': ')) return graph_str def get_md5(self): import hashlib graph_str = self.simple_description() md = hashlib.md5() md.update(graph_str.encode("utf-8")) return md.hexdigest() class Node(NodeBase): """A node contains an op and its input and output tensor. """ def __init__(self, name: str, op: Optional[str] = None, dtype: Optional[str] = None, in_quant_part: Optional[bool] = False): super().__init__() self._name = name self._op = op self._dtype = dtype self._idx = -1 self._scope_name = "" self._source_range = "" self._normalized_name = "" self._in_tensors = [] self._out_tensors = [] self._in_nodes = [] self._out_nodes = [] self._blocks = [] self._is_quantizable = in_quant_part self._is_merged = False self._transpose_in_order = None self._transpose_out_order = None self._topo_position = 0 self._block = None self._graph = None self._neighbor_nodes = [None, None] self._target_device = None def __repr__(self): return f"Node(name={self.name}, id={self.idx}, op_type={self.op.type}, quant_state={self.in_quant_part})" def __str__(self): return json.dumps(self.description(), indent=2, separators=(',', ': ')) def __deepcopy__(self, memo): raise NotImplementedError("Deep copy is prohibited, use `clone_from` instead.") def clone_from(self, src_node, local_map): tmp_attrs = src_node.op._attrs tmp_params = src_node.op._params tmp_configs = src_node.op._configs src_node.op._params = copy.copy(tmp_params) src_node.op._attrs = copy.copy(tmp_attrs) src_node.op._configs = copy.copy(tmp_configs) self.op = copy.copy(src_node.op) self.op._export_attr_and_param() src_node.op._attrs = tmp_attrs src_node.op._params = tmp_params src_node.op._configs = tmp_configs self.op.clone_from(src_node.op, local_map) def scope_name(self): return self._scope_name def scope_name(self, name): self._scope_name = name def description(self): node_des = {} node_des['name'] = self._name node_des['scope_name'] = self._scope_name node_des['idx'] = self._idx node_des['dtype'] = self._dtype node_des['enable_quant'] = self._is_quantizable node_des['in_nodes'] = [i for i in self.in_nodes] node_des['out_nodes'] = [o for o in self.out_nodes] node_des['in_tensors'] = [it.description() for it in self.in_tensors] node_des['out_tensors'] = [ot.description() for ot in self.out_tensors] node_des['op'] = self._op.description() if self._blocks: for i, block in enumerate(self._blocks): node_des[f'block_{i}'] = [] for n in sorted(block.nodes, key=lambda n: n.idx): node_des[f'block_{i}'].append(n.description()) return node_des def clean_connections(self): self._in_nodes = [] self._out_nodes = [] def add_in_node(self, node_name: str): if node_name not in self._in_nodes: self._in_nodes.append(node_name) def add_out_node(self, node_name: str): if node_name not in self._out_nodes: self._out_nodes.append(node_name) def in_tensors(self): return self._in_tensors def out_tensors(self): return self._out_tensors def in_nodes(self): nodes = [] for tensor in self.in_tensors: if tensor.node is not None: nodes.append(tensor.node.name) return nodes def out_nodes(self): nodes = [] for out in self.out_tensors: for use in out.uses: nodes.append(use.user.name) return nodes def node_attr(self, key): return self._op.get_attr(key) def set_node_attr(self, key, value): if all([val is None for val in self._op._attr_value_mem[key]]): self._op.set_attr(key, value) else: self._op.update_attr(key, value) def node_config(self, key): return self._op.get_config(key) def set_node_config(self, key, value): self._op.set_config(key, value) def has_bound_params(self): return self._op.has_native_params() def op_type(self): return self.op.type def name(self): return self._name def name(self, value): self._name = value def idx(self): return self._idx def idx(self, index): self._idx = index self.owning_graph.update_node_idx(self, index) def op(self): return self._op def op(self, op): self._op = op def dtype(self): return self._dtype # @property # def alias(self): # return self._alias def in_quant_part(self) -> bool: return self._is_quantizable def in_quant_part(self, quant_state: bool) -> None: self._is_quantizable = quant_state def module(self): return self._module() def module(self, module): self._module = weakref.ref(module) def blocks(self): return self._blocks def add_block(self, block): self._blocks.append(block) def has_custom_op(self): return isinstance(self.op, CustomOp) def get_attr_val(self, attr_name): attr = self.node_attr(attr_name) return attr.data if isinstance(attr, Tensor) else attr def merged(self): return self._is_merged def merged(self, flag): self._is_merged = flag def transpose_in_order(self): return self._transpose_in_order def transpose_in_order(self, order): self._transpose_in_order = order def transpose_out_order(self): return self._transpose_out_order def transpose_out_order(self, order): self._transpose_out_order = order def set_node_attr_tensor_value(self, old_tensor, new_tensor): for attr_name, attr_value in self.op.attrs.items(): if attr_value.value is old_tensor: self.set_node_attr(attr_name, new_tensor) def destroy(self): if len(self.blocks) > 0: raise RuntimeError("Can't destroy if or loop node.") while len(self.out_tensors) > 0: self.remove_output(len(self.out_tensors) - 1) self.remove_all_inputs() if self.in_node_list(): self.remove_from_list() self.owning_graph.free_node(self) def remove_output(self, i): assert i < len(self.out_tensors) assert len(self.out_tensors[i].uses) == 0 output = self.out_tensors.pop(i) self.owning_graph.remove_tensor(output) for output_offset in range(i, len(self.out_tensors)): self.out_tensors[output_offset].offset -= 1 def replace_input_at(self, i, new_tensor): old_tensor = self.in_tensors[i] if old_tensor is new_tensor: return self.in_tensors[i] = new_tensor uses = [u for u in old_tensor.uses] attr_uses = [attr_u for attr_u in old_tensor.attr_uses] for u in uses: if u.user is self: new_tensor.uses.append(u) old_tensor.uses.remove(u) for attr_u in attr_uses: if attr_u.user is self.op: old_tensor.replace_attr_with_new_tensor_v2(attr_u, new_tensor) def remove_input(self, i): self.drop_input(i) for j in range(i + 1, len(self._in_tensors)): it = self.find_use_for_input(j) it.offset -= 1 self._in_tensors.pop(i) def remove_all_inputs(self): for i in range(len(self.in_tensors)): self.drop_input(i) self.in_tensors.clear() def drop_input(self, i): assert i < len(self.in_tensors) input_value = self.in_tensors[i] use_it = self.find_use_for_input(i) input_value.uses.remove(use_it) self.in_tensors[i] = None return input_value def find_use_for_input(self, i): use_it = None for use in self.in_tensors[i].uses: if use.offset == i and use.user is self: use_it = use assert use_it is not None return use_it def owning_block(self): return self._block def owning_block(self, block): self._block = block def owning_graph(self): return self._graph def owning_graph(self, graph): self._graph = graph if self._graph: self._graph.add_node(self) def topo_position(self): return self._topo_position def topo_position(self, pos): self._topo_position = pos def insert_before(self, node): assert node.in_node_list() self.insert_after(node.prev_node) def insert_after(self, node): assert not self.in_node_list() and node.in_node_list() assert node.owning_block is not None self._block = node.owning_block next_node = node.next_node node.next_node = self self.prev_node = node self.next_node = next_node next_node.prev_node = self self.update_topo_position() def update_topo_position(self): is_first_node = self.prev_node is self.owning_block.input_node is_last_node = self.next_node is self.owning_block.return_node prev_pos = self.prev_node.topo_position next_pos = self.next_node.topo_position if is_last_node: if is_first_node: self.topo_position = MID_POSITION return if prev_pos >= (POSITION_UPPER_BOUND - APPEND_INTERVAL): self.owning_block.reindex_topo() return self.topo_position = prev_pos + APPEND_INTERVAL elif is_first_node: if next_pos <= (POSITION_LOWER_BOUND + APPEND_INTERVAL): self.owning_block.reindex_topo() return self.topo_position = next_pos - APPEND_INTERVAL else: pos_between = prev_pos + (next_pos - prev_pos) / 2 if pos_between == prev_pos: self.owning_block.reindex_topo() return self.topo_position = pos_between def next_node(self): return self._neighbor_nodes[1] def next_node(self, node): self._neighbor_nodes[1] = node def prev_node(self): return self._neighbor_nodes[0] def prev_node(self, node): self._neighbor_nodes[0] = node def in_node_list(self): if self.next_node is None: assert self.prev_node is None return self.next_node is not None def remove_from_list(self): assert self.in_node_list() if self.owning_block.input_node is self: self.owning_block.input_node = self.next_node self.owning_block = None next_node = self.next_node prev_node = self.prev_node prev_node.next_node = next_node next_node.prev_node = prev_node self.next_node = None self.prev_node = None def add_in_tensor(self, tensor): tensor.uses.append(Use(self, len(self.in_tensors))) self._in_tensors.append(tensor) self.owning_graph.add_tensor(tensor) def add_out_tensor(self, tensor): tensor.offset = len(self.out_tensors) self._out_tensors.append(tensor) tensor.node = self self.owning_graph.add_tensor(tensor) def target_device(self): return self._target_device def target_device(self, device): self._target_device = device def scope_name(self): return self._scope_name def scope_name(self, scope_name): self._scope_name = scope_name def source_range(self): return self._source_range def source_range(self, source_range): self._source_range = source_range def normalized_name(self): return self._normalized_name def normalized_name(self, name): self._normalized_name = name op_modifier = registry.Registry("expanding Modifier Functions") ]) ]) class StructuredExpanding(object): def __init__(self, node_name: str) -> None: self._node_name: str = node_name self._in_dim: int = 0 self._out_dim: int = 0 def node_name(self) -> str: return self._node_name def in_dim(self) -> int: return self._in_dim def in_dim(self, v: int) -> None: self._in_dim = v def out_dim(self) -> int: return self._out_dim def out_dim(self, v: int) -> None: self._out_dim = v def added_out_channel(self) -> int: raise NotImplementedError("method added_out_channel is not implemented") def added_in_channel(self) -> int: raise NotImplementedError("method added_in_channel is not implemented") def out_inserts(self) -> List[DataInsert]: raise NotImplementedError("method out_inserts is not implemented") def update_node_by_expanding(graph: Graph, node: Node, node_expand_desc: Mapping[str, StructuredExpanding]): op_type = node.op.type if op_type in op_modifier: mod_func = op_modifier.lookup(op_type) mod_func(graph, node, node_expand_desc) else: propagate_node_expanding(node, node_expand_desc)
null
23,306
from typing import Mapping from nndct_shared.expanding.spec import BatchNormStructuredExpanding, InstanceNormStructuredExpanding, \ DataInsert, GenericStructuredExpanding, StructuredExpanding, WeightedNodeStructuredExpanding from nndct_shared.nndct_graph.base_graph import Graph from nndct_shared.nndct_graph.base_node import Node from nndct_shared.nndct_graph.base_tensor import Tensor from nndct_shared.utils import registry from nndct_shared.base.key_names import NNDCT_OP as OpTypes from nndct_shared.pruning.pruning_lib import is_depthwise_conv, _DISALLOW_PRUNED_INPUT_OPS, find_prunable_ancestor, CONV_OPS import numpy as np def _modify_depthwise(graph: Graph, node: Node, expanding_desc: Mapping[str, StructuredExpanding]): def modify_depthwise_conv(graph, node, pruning_res): _modify_depthwise(graph, node, pruning_res)
null
23,312
from typing import Mapping from nndct_shared.expanding.spec import BatchNormStructuredExpanding, InstanceNormStructuredExpanding, \ DataInsert, GenericStructuredExpanding, StructuredExpanding, WeightedNodeStructuredExpanding from nndct_shared.nndct_graph.base_graph import Graph from nndct_shared.nndct_graph.base_node import Node from nndct_shared.nndct_graph.base_tensor import Tensor from nndct_shared.utils import registry from nndct_shared.base.key_names import NNDCT_OP as OpTypes from nndct_shared.pruning.pruning_lib import is_depthwise_conv, _DISALLOW_PRUNED_INPUT_OPS, find_prunable_ancestor, CONV_OPS import numpy as np class StructuredExpanding(object): def __init__(self, node_name: str) -> None: self._node_name: str = node_name self._in_dim: int = 0 self._out_dim: int = 0 def node_name(self) -> str: return self._node_name def in_dim(self) -> int: return self._in_dim def in_dim(self, v: int) -> None: self._in_dim = v def out_dim(self) -> int: return self._out_dim def out_dim(self, v: int) -> None: self._out_dim = v def added_out_channel(self) -> int: raise NotImplementedError("method added_out_channel is not implemented") def added_in_channel(self) -> int: raise NotImplementedError("method added_in_channel is not implemented") def out_inserts(self) -> List[DataInsert]: raise NotImplementedError("method out_inserts is not implemented") class Graph(GraphBase): """ Graph object of NNDCT, contain list of NndctNodes. That will be used for topology or export to XGraph""" def __init__(self, graph_name=None): super(Graph, self).__init__() self._name = graph_name or 'NndctGraph' self._nodes_by_name = {} self._nodes_by_id = {} self._end_tensors = [] self._copy_tensor_map = {} self._tensors = {} self._blocks = [] self._param_names = [] self._top_block = None def __contains__(self, node_or_name: Union[str, Node]) -> bool: if isinstance(node_or_name, str): return node_or_name in self._nodes_by_name else: return node_or_name.name in self._nodes_by_name def __deepcopy__(self, memo): raise NotImplementedError("Deep copy is prohibited, use `clone_from` instead.") def clone(self): graph = self.__class__(self.name) graph.clone_from(self) return graph def clone_from(self, src_graph): local_map = {} converted_nodes = [] head_node = self.create_node_from(src_graph.head_node, local_map, converted_nodes) return_node = self.create_node_from(src_graph.return_node, local_map, converted_nodes) top_block = Block(self, None, head_node, return_node) self.set_top_block(top_block) self._top_block.clone_from(src_graph.block, local_map, converted_nodes) def create_node_from(self, src_node, local_map, converted_nodes): node = Node(src_node.name, dtype=src_node.dtype, in_quant_part=src_node.in_quant_part) node.owning_graph = self node.idx = src_node.idx node.scope_name = src_node.scope_name node.source_range = src_node.source_range node.target_device = src_node.target_device node.normalized_name = src_node.normalized_name converted_nodes.append(src_node.name) for out in src_node.out_tensors: if out.name in local_map: node.add_out_tensor(local_map[out.name]) else: tensor = Tensor(name=out.name) tensor.clone_from(out) local_map[out.name] = tensor node.add_out_tensor(tensor) for inp in src_node.in_tensors: if inp.name in local_map: node.add_in_tensor(local_map[inp.name]) else: tensor = Tensor(name=inp.name) tensor.clone_from(inp) local_map[inp.name] = tensor node.add_in_tensor(tensor) node.clone_from(src_node, local_map) for src_block in src_node.blocks: head_node = self.create_node_from(src_block.input_node, local_map, converted_nodes) return_node = self.create_node_from(src_block.return_node, local_map, converted_nodes) block = Block(self, node, head_node, return_node) block.clone_from(src_block, local_map, converted_nodes) node.add_block(block) return node def node(self, name): """Return node with the specified name""" return self._nodes_by_name.get(name, None) def get_node_by_idx(self, idx): node = self._nodes_by_id.get(idx, None) assert node is not None return node def get_input_nodes(self): input_nodes = [] for node in self.nodes: if (len(self.parents(node)) == 0) and \ (node.op.type==NNDCT_OP.INPUT or node.op.type==NNDCT_OP.TUPLE_INPUT): input_nodes.append(node) return input_nodes def get_input_tensors(self, input_args): input_tensors = [] graph_name = self.name input_nodes = self.get_input_nodes() for idx in range(len(input_args)): #input_node_name = graph_name + "::input_" + str(idx) #input_node = self.node(input_node_name) input_node = input_nodes[idx] input_tensor = input_node.out_tensors[0] if input_node.op.type == NNDCT_OP.INPUT: input_tensors.append(input_tensor.name) elif input_node.op.type == NNDCT_OP.TUPLE_INPUT: for index in range(len(input_args[idx])): input_tensor_name = input_tensor.name + '.' + str(index) input_tensors.append(input_tensor_name) return input_tensors def get_return_tensors(self): return_tensors = [] for tensor in self.return_node.in_tensors: return_tensors.append(tensor.name) return return_tensors def add_node(self, node: Node) -> None: if node.name in self._nodes_by_name: return if node.idx in self._nodes_by_id and node is not self._nodes_by_id[node.idx]: raise RuntimeError(f"The id `{node.idx}` of {node.name} has been added into graph") if node.idx == -1: # if not self._nodes_by_id: # node._idx = 0 # else: # node._idx = max([node.idx for node in self.all_nodes()]) + 1 node._idx = -sys.maxsize + len(list(self.all_nodes())) self._nodes_by_name[node.name] = node self._nodes_by_id[node.idx] = node def free_node(self, node): node.owning_graph = None self._nodes_by_name.pop(node.name) self._nodes_by_id.pop(node.idx) def remove_node(self, node): assert node.in_tensors assert len(node.out_tensors) == 1 out_tensor = node.out_tensors[0] inp_tensor = node.in_tensors[0] out_tensor.replace_uses_with(inp_tensor) node.destroy() def remove_node_by_types(self, node_types: List[str]) -> Dict[str, str]: if any([node_type in self.op_types for node_type in node_types]): nodes_to_remove = [] for node in self.nodes: if node.op.type in node_types: nodes_to_remove.append(node) for node in nodes_to_remove: self.remove_node(node) def find_nodes_by_types(self, node_types: List[NNDCT_OP]): conv_nodes = [] for node in self.nodes: if node.op.type in node_types: conv_nodes.append(node) else: continue return conv_nodes def reconnect_nodes(self): self._nodes_by_id.clear() for idx, node in enumerate(self.nodes): node.idx = idx self._nodes_by_id[idx] = node node.clean_connections() self.connect_nodes() def connect_nodes(self): for nodeA in self.nodes: for input_tensor in nodeA.in_tensors: for nodeB in self.nodes: if nodeB is not nodeA and input_tensor in nodeB.out_tensors: #nodeB.outputs.add(input_tensor.node.name) nodeB.add_out_node(nodeA.name) nodeA.add_in_node(input_tensor.node.name) def parents(self, node: Union[Node, str]) -> List[Node]: if isinstance(node, str): node = self.node(node) return [self.node(node_name) for node_name in node.in_nodes] def children(self, node: Union[Node, str]) -> List[Node]: if isinstance(node, str): node = self.node(node) return [self.node(node_name) for node_name in node.out_nodes] def add_tensor(self, tensor): self._tensors[tensor.name] = tensor def tensor(self, name): return self._tensors.get(name, None) def param_tensor(self, name): for node in self.all_nodes(): for _, tensor in node.op.params.items(): if tensor.name == name: return tensor def add_end_tensor(self, tensor): self._end_tensors.append(tensor) def __repr__(self): return f"Graph(name={self.name})" def __str__(self): return json.dumps(self.description(), indent=2, separators=(',', ': ')) def description(self): graph_des = {} graph_des['graph_name'] = f"{self.__class__.__name__}" graph_des['nodes'] = [] for n in sorted(self.nodes, key=lambda n: n.idx): graph_des['nodes'].append(n.description()) return graph_des def set_node_id(self, index, node): node.idx = index self._nodes_by_id[index] = node def create_subgraph_from_nodeset(cls, origin_graph, nodeset, graph_name): """ create a subgraph from nodeset belong to origin graph """ assert len(nodeset) >= 2 sorted_nodeset = origin_graph.top_sort_nodeset(nodeset) for node in sorted_nodeset: node.remove_from_list() subgraph = cls(graph_name) sorted_nodeset[0].owning_graph = subgraph sorted_nodeset[-1].owning_graph = subgraph block = Block(subgraph, None, sorted_nodeset[0], sorted_nodeset[-1]) subgraph.set_top_block(block) if len(sorted_nodeset) > 2: for node in sorted_nodeset[1:-1]: node.owning_graph = subgraph subgraph.append_node(node) return subgraph def top_sort_nodeset(nodeset: Sequence[Node]) -> List[Node]: sorted_nodeset = sorted(nodeset, key=lambda n: n.topo_position) return sorted_nodeset def get_topological_graph_nodes_list(self): nodes_list = [node for node in self.nodes] return Graph.top_sort_nodeset(nodes_list) def name(self): return self._name def name(self, name): self._name = name def nodes(self): return self._top_block.nodes def reverse_nodes(self): return self._top_block.reverse_nodes def tensors(self): for tensor in self._tensors.values(): yield tensor # TODO: Remove def end_tensors(self): return [tensor for tensor in self.return_node.in_tensors] def inputs(self): return [node for node in self.all_nodes() if not node.in_nodes] def outputs(self): return [node for node in self.all_nodes() if not node.out_nodes] def op_types(self): return {node.op.type for node in self.all_nodes()} def append_node(self, node): self._top_block.append_node(node) def add_param_name(self, param_name): if param_name not in self._param_names: self._param_names.append(param_name) def param_names(self): return list(self._param_names) def block(self): return self._top_block def is_tensor_in_graph(self, tensor_name): return True if tensor_name in self._tensors else False def update_node_idx(self, node, index): self._nodes_by_id[index] = node def clear_node_id_map(self): self._nodes_by_id.clear() def remove_tensor(self, tensor): self._tensors.pop(tensor.name) if tensor.name in self._param_names: self._param_names.remove(tensor.name) def insert_node_between_nodes(self, new_node, parent_node, child_node): assert parent_node.in_node_list() and child_node.in_node_list() assert (parent_node.owning_graph == child_node.owning_graph and parent_node.owning_block == child_node.owning_block) new_node.owning_block = parent_node.owning_block new_node.owning_graph = parent_node.owning_graph tensor = Tensor(name=new_node.name, node=new_node) new_node.add_out_tensor(tensor) out_tensor = None offset = None for out in parent_node.out_tensors: for use in out.uses: if use.user is child_node: out_tensor = out offset = use.offset break #out_tensor.replace_uses_with(new_node.out_tensors[0]) child_node.replace_input_at(offset, new_node.out_tensors[0]) new_node.add_in_tensor(out_tensor) new_node.insert_after(parent_node) def set_top_block(self, block): self._top_block = block def add_block(self, block): self._blocks.append(block) def all_blocks(self): return self._blocks def all_nodes(self): for _, node in self._nodes_by_name.items(): yield node def head_node(self): return self._top_block.input_node def return_node(self): return self._top_block.return_node def clean_tensors_data(self): for tensor in self.tensors: tensor.clean_data() def assign_node_topological_name(self, prefix="", suffix=""): count = itertools.count(0) illegal_char_regex = re.compile('[^0-9a-zA-Z_]+') def _assgin_nodes(nodes): sorted_nodes = self.top_sort_nodeset(list(nodes)) for n in sorted_nodes: if n.blocks: for block in n.blocks: _assgin_nodes(block.nodes) else: candidate = illegal_char_regex.sub("_", n.op.type) n.normalized_name = f"{prefix}{candidate}_{next(count)}{suffix}" _assgin_nodes(self.nodes) def simple_description(self): """ Only describe op type topological info. """ def get_node_simple_info(node): node_des = {} node_des['op'] = node.op.type node_des["input_ops"] = [node.owning_graph.node(inode).op.type for inode in node.in_nodes] node_des["output_ops"] = [node.owning_graph.node(onode).op.type for onode in node.out_nodes] if node.blocks: for i, block in enumerate(node.blocks): node_des[f'block_{i}'] = [] for n in self.top_sort_nodeset(list(block.nodes)): node_des[f'block_{i}'].append(get_node_simple_info(n)) return node_des graph_des = {} graph_des['nodes'] = [] for n in self.top_sort_nodeset(list(self.nodes)): graph_des['nodes'].append(get_node_simple_info(n)) graph_str = json.dumps(graph_des, indent=2, separators=(',', ': ')) return graph_str def get_md5(self): import hashlib graph_str = self.simple_description() md = hashlib.md5() md.update(graph_str.encode("utf-8")) return md.hexdigest() class Node(NodeBase): """A node contains an op and its input and output tensor. """ def __init__(self, name: str, op: Optional[str] = None, dtype: Optional[str] = None, in_quant_part: Optional[bool] = False): super().__init__() self._name = name self._op = op self._dtype = dtype self._idx = -1 self._scope_name = "" self._source_range = "" self._normalized_name = "" self._in_tensors = [] self._out_tensors = [] self._in_nodes = [] self._out_nodes = [] self._blocks = [] self._is_quantizable = in_quant_part self._is_merged = False self._transpose_in_order = None self._transpose_out_order = None self._topo_position = 0 self._block = None self._graph = None self._neighbor_nodes = [None, None] self._target_device = None def __repr__(self): return f"Node(name={self.name}, id={self.idx}, op_type={self.op.type}, quant_state={self.in_quant_part})" def __str__(self): return json.dumps(self.description(), indent=2, separators=(',', ': ')) def __deepcopy__(self, memo): raise NotImplementedError("Deep copy is prohibited, use `clone_from` instead.") def clone_from(self, src_node, local_map): tmp_attrs = src_node.op._attrs tmp_params = src_node.op._params tmp_configs = src_node.op._configs src_node.op._params = copy.copy(tmp_params) src_node.op._attrs = copy.copy(tmp_attrs) src_node.op._configs = copy.copy(tmp_configs) self.op = copy.copy(src_node.op) self.op._export_attr_and_param() src_node.op._attrs = tmp_attrs src_node.op._params = tmp_params src_node.op._configs = tmp_configs self.op.clone_from(src_node.op, local_map) def scope_name(self): return self._scope_name def scope_name(self, name): self._scope_name = name def description(self): node_des = {} node_des['name'] = self._name node_des['scope_name'] = self._scope_name node_des['idx'] = self._idx node_des['dtype'] = self._dtype node_des['enable_quant'] = self._is_quantizable node_des['in_nodes'] = [i for i in self.in_nodes] node_des['out_nodes'] = [o for o in self.out_nodes] node_des['in_tensors'] = [it.description() for it in self.in_tensors] node_des['out_tensors'] = [ot.description() for ot in self.out_tensors] node_des['op'] = self._op.description() if self._blocks: for i, block in enumerate(self._blocks): node_des[f'block_{i}'] = [] for n in sorted(block.nodes, key=lambda n: n.idx): node_des[f'block_{i}'].append(n.description()) return node_des def clean_connections(self): self._in_nodes = [] self._out_nodes = [] def add_in_node(self, node_name: str): if node_name not in self._in_nodes: self._in_nodes.append(node_name) def add_out_node(self, node_name: str): if node_name not in self._out_nodes: self._out_nodes.append(node_name) def in_tensors(self): return self._in_tensors def out_tensors(self): return self._out_tensors def in_nodes(self): nodes = [] for tensor in self.in_tensors: if tensor.node is not None: nodes.append(tensor.node.name) return nodes def out_nodes(self): nodes = [] for out in self.out_tensors: for use in out.uses: nodes.append(use.user.name) return nodes def node_attr(self, key): return self._op.get_attr(key) def set_node_attr(self, key, value): if all([val is None for val in self._op._attr_value_mem[key]]): self._op.set_attr(key, value) else: self._op.update_attr(key, value) def node_config(self, key): return self._op.get_config(key) def set_node_config(self, key, value): self._op.set_config(key, value) def has_bound_params(self): return self._op.has_native_params() def op_type(self): return self.op.type def name(self): return self._name def name(self, value): self._name = value def idx(self): return self._idx def idx(self, index): self._idx = index self.owning_graph.update_node_idx(self, index) def op(self): return self._op def op(self, op): self._op = op def dtype(self): return self._dtype # @property # def alias(self): # return self._alias def in_quant_part(self) -> bool: return self._is_quantizable def in_quant_part(self, quant_state: bool) -> None: self._is_quantizable = quant_state def module(self): return self._module() def module(self, module): self._module = weakref.ref(module) def blocks(self): return self._blocks def add_block(self, block): self._blocks.append(block) def has_custom_op(self): return isinstance(self.op, CustomOp) def get_attr_val(self, attr_name): attr = self.node_attr(attr_name) return attr.data if isinstance(attr, Tensor) else attr def merged(self): return self._is_merged def merged(self, flag): self._is_merged = flag def transpose_in_order(self): return self._transpose_in_order def transpose_in_order(self, order): self._transpose_in_order = order def transpose_out_order(self): return self._transpose_out_order def transpose_out_order(self, order): self._transpose_out_order = order def set_node_attr_tensor_value(self, old_tensor, new_tensor): for attr_name, attr_value in self.op.attrs.items(): if attr_value.value is old_tensor: self.set_node_attr(attr_name, new_tensor) def destroy(self): if len(self.blocks) > 0: raise RuntimeError("Can't destroy if or loop node.") while len(self.out_tensors) > 0: self.remove_output(len(self.out_tensors) - 1) self.remove_all_inputs() if self.in_node_list(): self.remove_from_list() self.owning_graph.free_node(self) def remove_output(self, i): assert i < len(self.out_tensors) assert len(self.out_tensors[i].uses) == 0 output = self.out_tensors.pop(i) self.owning_graph.remove_tensor(output) for output_offset in range(i, len(self.out_tensors)): self.out_tensors[output_offset].offset -= 1 def replace_input_at(self, i, new_tensor): old_tensor = self.in_tensors[i] if old_tensor is new_tensor: return self.in_tensors[i] = new_tensor uses = [u for u in old_tensor.uses] attr_uses = [attr_u for attr_u in old_tensor.attr_uses] for u in uses: if u.user is self: new_tensor.uses.append(u) old_tensor.uses.remove(u) for attr_u in attr_uses: if attr_u.user is self.op: old_tensor.replace_attr_with_new_tensor_v2(attr_u, new_tensor) def remove_input(self, i): self.drop_input(i) for j in range(i + 1, len(self._in_tensors)): it = self.find_use_for_input(j) it.offset -= 1 self._in_tensors.pop(i) def remove_all_inputs(self): for i in range(len(self.in_tensors)): self.drop_input(i) self.in_tensors.clear() def drop_input(self, i): assert i < len(self.in_tensors) input_value = self.in_tensors[i] use_it = self.find_use_for_input(i) input_value.uses.remove(use_it) self.in_tensors[i] = None return input_value def find_use_for_input(self, i): use_it = None for use in self.in_tensors[i].uses: if use.offset == i and use.user is self: use_it = use assert use_it is not None return use_it def owning_block(self): return self._block def owning_block(self, block): self._block = block def owning_graph(self): return self._graph def owning_graph(self, graph): self._graph = graph if self._graph: self._graph.add_node(self) def topo_position(self): return self._topo_position def topo_position(self, pos): self._topo_position = pos def insert_before(self, node): assert node.in_node_list() self.insert_after(node.prev_node) def insert_after(self, node): assert not self.in_node_list() and node.in_node_list() assert node.owning_block is not None self._block = node.owning_block next_node = node.next_node node.next_node = self self.prev_node = node self.next_node = next_node next_node.prev_node = self self.update_topo_position() def update_topo_position(self): is_first_node = self.prev_node is self.owning_block.input_node is_last_node = self.next_node is self.owning_block.return_node prev_pos = self.prev_node.topo_position next_pos = self.next_node.topo_position if is_last_node: if is_first_node: self.topo_position = MID_POSITION return if prev_pos >= (POSITION_UPPER_BOUND - APPEND_INTERVAL): self.owning_block.reindex_topo() return self.topo_position = prev_pos + APPEND_INTERVAL elif is_first_node: if next_pos <= (POSITION_LOWER_BOUND + APPEND_INTERVAL): self.owning_block.reindex_topo() return self.topo_position = next_pos - APPEND_INTERVAL else: pos_between = prev_pos + (next_pos - prev_pos) / 2 if pos_between == prev_pos: self.owning_block.reindex_topo() return self.topo_position = pos_between def next_node(self): return self._neighbor_nodes[1] def next_node(self, node): self._neighbor_nodes[1] = node def prev_node(self): return self._neighbor_nodes[0] def prev_node(self, node): self._neighbor_nodes[0] = node def in_node_list(self): if self.next_node is None: assert self.prev_node is None return self.next_node is not None def remove_from_list(self): assert self.in_node_list() if self.owning_block.input_node is self: self.owning_block.input_node = self.next_node self.owning_block = None next_node = self.next_node prev_node = self.prev_node prev_node.next_node = next_node next_node.prev_node = prev_node self.next_node = None self.prev_node = None def add_in_tensor(self, tensor): tensor.uses.append(Use(self, len(self.in_tensors))) self._in_tensors.append(tensor) self.owning_graph.add_tensor(tensor) def add_out_tensor(self, tensor): tensor.offset = len(self.out_tensors) self._out_tensors.append(tensor) tensor.node = self self.owning_graph.add_tensor(tensor) def target_device(self): return self._target_device def target_device(self, device): self._target_device = device def scope_name(self): return self._scope_name def scope_name(self, scope_name): self._scope_name = scope_name def source_range(self): return self._source_range def source_range(self, source_range): self._source_range = source_range def normalized_name(self): return self._normalized_name def normalized_name(self, name): self._normalized_name = name CONV_OPS = [ OpTypes.CONV2D, OpTypes.CONVTRANSPOSE2D, OpTypes.CONV3D, OpTypes.CONVTRANSPOSE3D, OpTypes.SEPARABLECONV2D ] def find_prunable_ancestor(graph, node, target_ops=CONV_OPS): return find_ancestor(graph, node, target_ops, [OpTypes.CONCAT]) ]) def raise_if_has_pruned_input(graph: Graph, node: Node, expanding_desc: Mapping[str, StructuredExpanding]): for node_name in node.in_nodes: input_pruning = expanding_desc[node_name] if input_pruning.added_out_channel > 0: input_node = graph.node(node_name) if input_node.op.type in CONV_OPS: prunable_node = input_node else: prunable_node = find_prunable_ancestor(graph, input_node) raise RuntimeError(('Operation "{}" cannot take expanded tensor as input, ' 'please exclude node "{}" from pruning.').format( node.op.type, prunable_node.name))
null
23,314
import numpy as np import math from nndct_shared.base import NNDCT_KEYS, GLOBAL_MAP from nndct_shared.utils import NndctOption from nndct_shared.algorithms import breadth_first_search_handler from .quant_ops import normal_quant_neuron def normal_quant_neuron(data, maxamps=[[32768], [2048]], strides=[-1], round_method=2, keep_scale=True, name='', quantizer=None, on_gpu=True, as_int=False): #integer need not keep scale as precondition if as_int: keep_scale = False if len(strides) == 1: data = __amplify_data( data, maxamps[0][0], maxamps[1][0], method=round_method) if keep_scale: data = data / maxamps[1][0] else: org_shape = data.shape flatten_data = data.flatten() pos = 0 for idx, s in enumerate(strides): flatten_data[pos:pos + s] = __amplify_data( flatten_data[pos:pos + s], maxamps[0][idx], maxamps[1][idx], method=round_method) if keep_scale: flatten_data[pos:pos + s] = flatten_data[pos:pos + s] / maxamps[1][idx] pos += s data = flatten_data.reshape(org_shape) #return integer or origional dtype if as_int: assert all(m == maxamps[0][0] for m in maxamps[0]), "all max limitation should be the same" if maxamps[0][0] == 2**7 or maxamps[0][0] == 2**3: return data.astype(np.int8) elif maxamps[0][0] == 2**15: return data.astype(np.int16) else: raise TypeError("unexpected max found " + str(maxamps[0][0])) else: return data def quantize_data2int(data, bn, fp, method=2): return normal_quant_neuron( data, maxamps=[[2**(bn - 1)], [2**fp]], round_method=method, as_int=True)
null
23,319
import numpy as np import math from nndct_shared.base import NNDCT_KEYS, GLOBAL_MAP from nndct_shared.utils import NndctOption from nndct_shared.algorithms import breadth_first_search_handler from .quant_ops import normal_quant_neuron def maybe_get_quantizer(quantizer=None): def quant_channel_scale_params(node, channel_scale): quant_mode, quantizer = maybe_get_quantizer() # ignore parameters quantization if the node is not to be quantized #print('---- quant o: {}, in quant part:{}'.format(node.name, node.in_quant_part)) if not node.in_quant_part or quantizer is None: return channel_scale if quantizer.need_quantize_tensor(node.name, 'output'): #print('---- quant o: {}'.format(node.name)) output_name = node.name #print('qmode = %d, q_end: %d activation: %s' % # (quant_mode, is_quant_end, output_name)) if quant_mode == 2: datatype = 'int' if NndctOption.nndct_only_int_quant.value is False: datatype = quantizer.get_quant_dtype(node.name, tensor_type='output') channel_scale = quantizer.quantize( channel_scale, output_name, node, tensor_type='output', datatype=datatype) return channel_scale
null
23,328
from __future__ import absolute_import from __future__ import division from __future__ import print_function import numpy as np from nndct_shared.base.key_names import NNDCT_OP from nndct_shared.base.key_names import FrameworkType from nndct_shared.nndct_graph import base_tensor class DataFormatMap(object): """A dict mapping of framework and op type to its data format. """ _blob_format_map = { FrameworkType.NNDCT: { 2: "NH", 3: "NLC", 4: "NHWC", 5: "NHWDC" }, FrameworkType.TORCH: { 2: "NH", 3: "NCL", 4: "NCHW", 5: "NCDHW" }, # TF format generated in runtime. } _parameter_format_map = { FrameworkType.NNDCT: { 2: "OI", 3: "OLI", 4: "OHWI", 5: "OHWDI" }, FrameworkType.TORCH: { 2: "OI", 3: "OIL", 4: "OIHW", 5: "OIDHW" }, FrameworkType.TENSORFLOW: { 2: "IO", 3: "LIO", 4: "HWIO", 5: "DHWIO", } } def blob_format(cls, framework_type, ndim): if framework_type not in cls._blob_format_map: raise KeyError( "Framework type '{}' not supported now.".format(framework_type)) return cls._blob_format_map[framework_type][ndim] def param_format(cls, framework_type, ndim): if framework_type not in cls._parameter_format_map: raise KeyError( "Framework type '{}' not supported now.".format(framework_type)) return cls._parameter_format_map[framework_type][ndim] def layout_transformer(src_layout, dst_layout): assert len(src_layout) == len(dst_layout) axes = [] for axis in dst_layout: axes.append(src_layout.index(axis)) return tuple(axes) def convert_parameter_tensor_format(tensor: base_tensor.Tensor, src_framework: str, dst_framework: str) -> base_tensor.Tensor: if not isinstance(tensor, base_tensor.Tensor): raise TypeError("'tensor' must be Tensor, but given {}".format( type(tensor))) if not tensor.is_complete_tensor(): return tensor if src_framework == dst_framework: return tensor if tensor.ndim not in DataFormatMap._parameter_format_map[src_framework].keys(): return tensor src_format = DataFormatMap.param_format(src_framework, tensor.ndim) dst_format = DataFormatMap.param_format(dst_framework, tensor.ndim) if src_format == dst_format: return tensor tensor.transpose(layout_transformer(src_format, dst_format)) return tensor
null
23,333
from __future__ import absolute_import from __future__ import division from __future__ import print_function import logging as _logging import os as _os import sys as _sys import time as _time import traceback as _traceback from logging import DEBUG from logging import ERROR from logging import FATAL from logging import INFO from logging import WARN from logging import NOTSET import threading from .option_list import NndctOption from nndct_shared.base import SingletonMeta from .msg_code import QError, QWarning def _log_prefix(level, timestamp=None, file_and_line=None): def get_logger(name=None, level=None, file_name=None, only2file=False): def debug(msg, *args, **kwargs): extra = {'nndct_prefix': _log_prefix(DEBUG)} get_logger().debug(msg, extra=extra, *args, **kwargs)
null
23,334
from __future__ import absolute_import from __future__ import division from __future__ import print_function import logging as _logging import os as _os import sys as _sys import time as _time import traceback as _traceback from logging import DEBUG from logging import ERROR from logging import FATAL from logging import INFO from logging import WARN from logging import NOTSET import threading from .option_list import NndctOption from nndct_shared.base import SingletonMeta from .msg_code import QError, QWarning def _log_prefix(level, timestamp=None, file_and_line=None): """Generate a nndct logline prefix.""" # pylint: disable=global-variable-not-assigned global _level_names # pylint: enable=global-variable-not-assigned # Record current time now = timestamp or _time.time() now_tuple = _time.localtime(now) now_microsecond = int(1e6 * (now % 1.0)) (filename, line) = file_and_line or _get_file_and_line() basename = _os.path.basename(filename) # Severity string severity = 'I' if level in _level_names: severity = _level_names[level][0] s = '%c%02d%02d %02d:%02d:%02d.%06d %s:%d]' % ( severity, now_tuple[1], # month now_tuple[2], # day now_tuple[3], # hour now_tuple[4], # min now_tuple[5], # sec now_microsecond, basename, line) return s def get_logger(name=None, level=None, file_name=None, only2file=False): """Return logger instance.""" # global _logger # Use double-checked locking to avoid taking lock unnecessarily. # if _logger: # return _logger _logger_lock.acquire() try: # Scope the TensorFlow logger to not conflict with users' loggers. logger = _logging.getLogger(name) if level: logger.setLevel(_logging.INFO) else: logger.setLevel(1) # Override findCaller on the logger to skip internal helper functions logger.findCaller = _logger_find_caller # Don't further configure the TensorFlow logger if the root logger is # already configured. This prevents double logging in those cases. if not _logging.getLogger().handlers: # Determine whether we are in an interactive environment _interactive = False try: # This is only defined in interactive shells. if _sys.ps1: _interactive = True except AttributeError: # Even now, we may be in an interactive shell with `python -i`. _interactive = _sys.flags.interactive # If we are in an interactive environment (like Jupyter), set loglevel # to INFO and pipe the output to stdout. if _interactive: #logger.setLevel(INFO) _logging_target = _sys.stdout else: _logging_target = _sys.stderr if not only2file and all([not isinstance(hdler, _logging.StreamHandler) for hdler in logger.handlers]): # Add the output handler. _handler = _logging.StreamHandler(_logging_target) # _handler.setFormatter(_logging.Formatter(_logging_fmt, None)) logger.addHandler(_handler) if file_name is not None: _file_handler = _logging.FileHandler(file_name) logger.addHandler(_file_handler) return logger finally: _logger_lock.release() def info(msg, *args, **kwargs): extra = {'nndct_prefix': _log_prefix(INFO)} get_logger().info(msg, *args, extra=extra, **kwargs)
null
23,336
from __future__ import absolute_import from __future__ import division from __future__ import print_function import logging as _logging import os as _os import sys as _sys import time as _time import traceback as _traceback from logging import DEBUG from logging import ERROR from logging import FATAL from logging import INFO from logging import WARN from logging import NOTSET import threading from .option_list import NndctOption from nndct_shared.base import SingletonMeta from .msg_code import QError, QWarning def _log_prefix(level, timestamp=None, file_and_line=None): """Generate a nndct logline prefix.""" # pylint: disable=global-variable-not-assigned global _level_names # pylint: enable=global-variable-not-assigned # Record current time now = timestamp or _time.time() now_tuple = _time.localtime(now) now_microsecond = int(1e6 * (now % 1.0)) (filename, line) = file_and_line or _get_file_and_line() basename = _os.path.basename(filename) # Severity string severity = 'I' if level in _level_names: severity = _level_names[level][0] s = '%c%02d%02d %02d:%02d:%02d.%06d %s:%d]' % ( severity, now_tuple[1], # month now_tuple[2], # day now_tuple[3], # hour now_tuple[4], # min now_tuple[5], # sec now_microsecond, basename, line) return s def get_logger(name=None, level=None, file_name=None, only2file=False): """Return logger instance.""" # global _logger # Use double-checked locking to avoid taking lock unnecessarily. # if _logger: # return _logger _logger_lock.acquire() try: # Scope the TensorFlow logger to not conflict with users' loggers. logger = _logging.getLogger(name) if level: logger.setLevel(_logging.INFO) else: logger.setLevel(1) # Override findCaller on the logger to skip internal helper functions logger.findCaller = _logger_find_caller # Don't further configure the TensorFlow logger if the root logger is # already configured. This prevents double logging in those cases. if not _logging.getLogger().handlers: # Determine whether we are in an interactive environment _interactive = False try: # This is only defined in interactive shells. if _sys.ps1: _interactive = True except AttributeError: # Even now, we may be in an interactive shell with `python -i`. _interactive = _sys.flags.interactive # If we are in an interactive environment (like Jupyter), set loglevel # to INFO and pipe the output to stdout. if _interactive: #logger.setLevel(INFO) _logging_target = _sys.stdout else: _logging_target = _sys.stderr if not only2file and all([not isinstance(hdler, _logging.StreamHandler) for hdler in logger.handlers]): # Add the output handler. _handler = _logging.StreamHandler(_logging_target) # _handler.setFormatter(_logging.Formatter(_logging_fmt, None)) logger.addHandler(_handler) if file_name is not None: _file_handler = _logging.FileHandler(file_name) logger.addHandler(_file_handler) return logger finally: _logger_lock.release() def error(msg, *args, **kwargs): extra = {'nndct_prefix': _log_prefix(ERROR)} get_logger().error(msg, *args, extra=extra, **kwargs)
null
23,346
import numpy as np from nndct_shared.base import NNDCT_OP def get_in_out_channel_idx(ndim, optype, data_formats): def get_tensor_in_dim(tensor, optype, data_formats): in_idx, _ = get_in_out_channel_idx(tensor.ndim, optype, data_formats) return tensor.shape[in_idx]
null
23,350
from typing import TypeVar, NoReturn, Optional, Iterator, List from .option_list import NndctOption from .option_def import Option, T class NndctOption(object): nndct_help = Option(name="help", dtype=bool, default=False, action="store_true", help="list all api usage description") nndct_quant_off = Option(name="quant_off", dtype=bool, default=False, action="store_true", help="disable quantization flow") nndct_option_list = Option(name="option_list", dtype=bool, default=False, action="store_true", help="list all the options in nndct") nndct_parse_debug = Option(name="parse_debug", dtype=int, default=0, env="NNDCT_PARSE_DEBUG", help="logging graph, 1: torch raw graph, 2: nndct graph 3: nndct quant graph") nndct_logging_level = Option(name="logging_level", dtype=int, default=0, help="logging level") nndct_quant_mode = Option(name="quant_mode", dtype=int, default=0, help="quant mode, 1:calibration, 2:quantization") nndct_dump_float_format = Option(name="dump_float_format", dtype=int, default=0, help="deploy check data format, 0: bin, 1: txt") nndct_record_slow_mode = Option(name="record_slow_mode", dtype=bool, default=False, action="store_true", help="record outputs every iteration") nndct_quant_opt = Option(name="quant_opt", dtype=int, default=3, help="quant opt level") nndct_relu6_replace = Option(name="relu6_replace", dtype=str, default='relu', help="relu6 replace operator") nndct_equalization = Option(name="equalization", dtype=bool, default=True, action="store_true", help="enable weights equalization") # nndct_wes = Option(name="weights_equalizing_shift", dtype=bool, default=False, action="store_true", # help="enable weights equalizing shift") # nndct_wes_in_cle = Option(name="weights_equalizing_shift in cle", dtype=bool, default=False, action="store_true", # help="enable weights equalizing shift in cle") nndct_param_corr = Option(name="param_corr", dtype=bool, default=True, action="store_true", help="enable parameter correction") nndct_param_corr_rate = Option(name="param_corr_rate", dtype=float, default=0.05, help="parameter correction rate") nndct_cv_app = Option(name="cv_app", dtype=bool, default=True, action="store_true", help="cv application") nndct_finetune_lr_factor = Option(name="finetune_lr_factor", dtype=float, default=0.01, help="finetune learning rate factor") nndct_partition_mode = Option(name="partition_mode", dtype=int, default=0, help="0: quant stub controled. 1: custom op controled") nndct_stat = Option(name="stat", dtype=int, default=0, help="quantizer statistic level") nndct_jit_script_mode = Option(name="jit_script_mode", dtype=bool, default=False, action="store_true", help="enable torch script parser") nndct_diffs_mode = Option(name="diffs_mode", dtype=str, default='mse', help="diffs_mode: mse, maxmin") nndct_ft_mode = Option(name="ft_mode", dtype=int, default=1, help="1: mix mode 0: cache mode") nndct_visualize = Option(name="visualize", dtype=bool, default=False, action="store_true", help="visualize tensors") nndct_dump_no_quant_part = Option(name="dump_no_quant_part", dtype=bool, default=False, action="store_true", help="dump no quantized nodes") nndct_max_fix_position = Option(name="max_fix_position", dtype=int, default=12, help="maximum of fix position") nndct_use_torch_quantizer = Option(name="use_torch_quantizer", dtype=bool, default=False, action="store_true", help="enable torch quantizer") nndct_jit_trace = Option(name="jit_trace", dtype=bool, default=False, action="store_true", env="NNDCT_JIT_TRACE", help="parse graph from script tracing") nndct_jit_script = Option(name="jit_script", dtype=bool, default=False, action="store_true", help="parse graph from script") nndct_calib_histogram_bins = Option(name="calib_histogram_bins", dtype=int, default=2048, help="calibration histogram bins number") nndct_mse_start_bin = Option(name="mse_start_bin", dtype=int, default=1536, help="mse calibration method start bin") nndct_mse_stride = Option(name="mse_stride", dtype=int, default=16, help="mse calibration method stride") nndct_entropy_start_bin = Option(name="entropy_start_bin", dtype=int, default=1536, help="entropy calibration method start bin") nndct_entropy_stride = Option(name="entropy_stride", dtype=int, default=16, help="entropy calibration method stride") nndct_convert_relu6_to_relu = Option(name="convert_relu6_to_relu", dtype=bool, default=False, help="convert relu6 to relu") nndct_convert_sigmoid_to_hsigmoid = Option(name="convert_sigmoid_to_hsigmoid", dtype=bool, default=False, action="store_true", help="convert sigmoid to hsigmoid") nndct_convert_silu_to_hswish = Option(name="convert_silu_to_hswish", dtype=bool, default=False, action="store_true", help="convert silu to hswish") nndct_keep_first_last_layer_accuracy = Option(name="keep_first_last_layer_accuracy", dtype=bool, default=False, help="keep accuracy of first and last layer") nndct_keep_add_layer_accuracy = Option(name="keep_add_layer_accuracy", dtype=bool, default=False, help="keep accuracy of add layer") nndct_avg_pool_approximate = Option(name="avg_pool_approximate", dtype=bool, default=True, action="store_true", help="enable average pooling approximate for dpu") nndct_leaky_relu_approximate = Option(name="leaky_relu_approximate", dtype=bool, default=True, action="store_true", help="enable leaky relu approximate for dpu") nndct_conv_bn_merge = Option(name="conv_bn_merge", dtype=bool, default=True, action="store_true", help="enable conv and bn merge") nndct_input_quant_only = Option(name="input_quant_only", dtype=bool, default=False, action="store_false", help="only quantize the input") nndct_tensorrt_strategy = Option(name="tensorrt_strategy", dtype=bool, default=False, action="store_true", help="use quantization strategy as tensorrt") nndct_tensorrt_quant_algo = Option(name="tensorrt_quant_algo", dtype=bool, default=False, action="store_true", help="use tensorrt quantization algorithm") nndct_calibration_local = Option(name="calibration_local", dtype=bool, default=True, action="store_true", help="calibration in local batch data") nndct_change_concat_input_fix = Option(name="change_concat_input_fix", dtype=bool, default=False, action="store_true", help="change concat input nodes fix point to be the same as concat output node") nndct_change_pool_input_fix = Option(name="change_pool_input_fix", dtype=bool, default=False, action="store_true", help="change pooling input nodes fix point to be the same as their output node") nndct_change_add_input_fix = Option(name="change_add_input_fix", dtype=bool, default=False, action="store_true", help="change add input nodes fix point to be the identical") nndct_insert_concat_input_fix = Option(name="insert_concat_input_fix", dtype=bool, default=False, action="store_true", help="insert concat input nodes fix point to be the same as concat output node") nndct_export_jit = Option(name="export_jit", dtype=bool, default=False, action="store_true", env="NNDCT_EXPORT_JIT", help="export quant script by inserting fixneuron") nndct_deploy_check = Option(name="deploy_check", dtype=bool, default=False, action="store_true", help="dump deploy data in forward process") nndct_input_check = Option(name="input_check", dtype=bool, default=False, action="store_true", help="dump input float data in forward process") nndct_op_tanh_sigmoid_mode = Option(name="tanh_sigmoid_mode", dtype=str, default='quant_input_output', help="Tanh/sigmoid quantization mode: quant_input_output, table_look_up, simulation, aie2_lut_16bw") nndct_op_softmax_mode = Option(name="softmax_mode", dtype=str, default='quant_input_output', help="Softmax quantization mode: quant_input_output, hardware_pl, liyi, aie2_lut_16bw, bert_8bw, ipu_8bw") nndct_op_logsoftmax_mode = Option(name="logsoftmax_mode", dtype=str, default='quant_input_output', help="Logsoftmax quantization mode: quant_input_output, aie2_lut_16bw") nndct_op_gelu_mode = Option(name="gelu_mode", dtype=str, default='quant_input_output', help="GELU quantization mode: quant_input_output, dynamic_table") nndct_op_layernorm_mode = Option(name="layernorm_mode", dtype=str, default='quant_input_output', help="Layernorm quantization mode: quant_input_output, aie2_16bw, bert_8bw") nndct_ip_asr = Option(name="ip_asr", dtype=bool, default=False, action="store_true", help="asr quant method") nndct_ip_v70_bert = Option(name="ip_v70_bert", dtype=bool, default=False, action="store_true", help="bert v70 quant method") nndct_ip_v70_bert_qat = Option(name="ip_v70_bert_qat", dtype=bool, default=False, action="store_true", help="bert v70 quant method") nndct_use_old_inspector = Option(name="use_old_inspector", dtype=bool, default=False, action="store_true", env="NNDCT_USE_OLD_INSPECTOR", help="switch to old inspector") nndct_calib_before_finetune = Option(name="calib_before_finetune", dtype=bool, default=False, action="store_true", help="calibration before fast finetune") nndct_inspect_debug = Option(name="inspect_debug", dtype=bool, default=False, action="store_true", env="NNDCT_INSPECT_DEBUG", help="turn on inspector") nndct_op_instancenorm_mode = Option(name="instancenorm_mode", dtype=str, default='quant_input_output', help="Instancenorm quantization mode: quant_input_output, ipu_8bw") nndct_op_groupnorm_mode = Option(name="groupnorm_mode", dtype=str, default='quant_input_output', help="Groupnorm quantization mode: quant_input_output, ipu_8bw") nndct_native_onnx = Option(name="native_onnx", dtype=bool, default=False, action="store_true", help="export native quant-dequant onnx models") nndct_inspect_test = Option(name="inspect_test", dtype=bool, default=False, action="store_true", env="NNDCT_INSPECT_TEST", help="embed target related test in torch quantizer") nndct_target = Option(name="target", dtype=str, default="", env="NNDCT_TARGET", help="target name") nndct_traversal_graph_mode = Option(name="nndct_traversal_graph_mode", dtype=int, default=0, env="NNDCT_GRAPH_SEARCH",help="0: auto, 1:recursion, 2: iteration") nndct_op_sqrt_mode = Option(name="sqrt_mode", dtype=str, default='quant_input_output', help="sqrt quantization mode: quant_input_output, ipu_8bw") nndct_onnx_opset_version = Option(name="onnx_opset_version", dtype=int, default=-1, help="opset_version of dumped onnx graph") nndct_only_int_quant = Option(name="only_int_quant", dtype=bool, default=True, help="only int datatype quantization included") nndct_gemm88 = Option(name="gemm88", dtype=bool, default=False, action="store_true", help="only quant gemm88 and matmul88") nndct_pooling_split_mode = Option(name="nndct_pooling_split_mode", dtype=bool, default=False, help="default big pooling will split to small pooling") nndct_fx_mode = Option(name="fx_mode", dtype=bool, default=False, action="store_true", env="NNDCT_FX_MODE", help="turn on fx mode") T = TypeVar('T') def set_option_value(option_name: str, option_value: T) -> NoReturn: NndctOption.__dict__[option_name].value = option_value
null
23,360
from nndct_shared.base import NNDCT_KEYS, GLOBAL_MAP, NNDCT_DEBUG_LVL from .log import nndct_debug_print def remove_trans_scp_prefix(name, scp=None): name = remove_prefix(name, scp) if name.startswith(NNDCT_KEYS.TRANS_SCOPE): name = '/'.join(name.split('/')[1:]) return name def scoped_untrans_name(name, scp): org_name = remove_trans_scp_prefix(name, scp) return scp + org_name
null
23,361
from nndct_shared.base import NNDCT_KEYS, GLOBAL_MAP, NNDCT_DEBUG_LVL from .log import nndct_debug_print def remove_prefix(obj, prefix): if obj is None: return obj if prefix is None or prefix == '': return obj if isinstance(prefix, str): if isinstance(obj, str) and len(prefix) > 0 and obj.startswith(prefix): obj = obj[len(prefix):] elif isinstance(obj, dict): obj = {k: remove_prefix(v, prefix) for k, v in obj.items()} elif isinstance(obj, list): obj = [remove_prefix(item, prefix) for item in obj] return obj elif isinstance(prefix, list): for pre in prefix: obj = remove_prefix(obj, pre) return obj else: raise Exception('prefix {} is not string or list!'.format(prefix)) def scoped_trans_name(name, scp): org_name = remove_prefix(name, scp) if org_name.startswith(NNDCT_KEYS.TRANS_SCOPE): return scp + org_name else: return scp + NNDCT_KEYS.TRANS_SCOPE + '/' + org_name
null
23,362
from nndct_shared.base import NNDCT_KEYS, GLOBAL_MAP, NNDCT_DEBUG_LVL from .log import nndct_debug_print def remove_trans_scp_prefix(name, scp=None): name = remove_prefix(name, scp) if name.startswith(NNDCT_KEYS.TRANS_SCOPE): name = '/'.join(name.split('/')[1:]) return name GLOBAL_MAP = GlobalMap() def nndct_debug_print(string, title='', level=1): if True == GLOBAL_MAP.get_ele( NNDCT_KEYS.DEBUG_FLAG) and level <= GLOBAL_MAP.get_ele( NNDCT_KEYS.VERBOSE_LEVEL): logger = GLOBAL_MAP.get_ele(NNDCT_KEYS.LOGGER) if title == 'Start': string = "\n********************* <{} : {}> *********************".format( title, string) elif title == 'End': string = "\n********************* <{} : {}> *********************\n".format( title, string) if logger: logger.debug("[NNDCT_DEBUG_Lv_{}] {}".format(level, string)) else: print("[NNDCT_DEBUG_Lv_{}] {}".format(level, string)) def map_output_and_node(output, node_or_name, model_type): if node_or_name is None: return if isinstance(node_or_name, str): node_name = node_or_name else: node_name = node_or_name.name node_name = remove_trans_scp_prefix(node_name) def _do_map(output_name, node_name): if not output_name == node_name: if not GLOBAL_MAP.get_ele(NNDCT_KEYS.OUTPUT_TO_NODE_MAP): GLOBAL_MAP.set_map(NNDCT_KEYS.OUTPUT_TO_NODE_MAP, {}) if not GLOBAL_MAP.get_ele(NNDCT_KEYS.NODE_TO_OUTPUT_MAP): GLOBAL_MAP.set_map(NNDCT_KEYS.NODE_TO_OUTPUT_MAP, {}) #map output to node output_to_node_map = GLOBAL_MAP.get_ele(NNDCT_KEYS.OUTPUT_TO_NODE_MAP) if not output_name in output_to_node_map: nndct_debug_print( "<map_output_and_node> map out {} and node{}".format( output_name, node_name), level=NNDCT_DEBUG_LVL.BUILD_GRAPH) output_to_node_map[output_name] = node_name else: assert output_to_node_map[ output_name] == node_name, "restored node name for output_name {} is {}, meet new node name {}".format( output_name, output_to_node_map[output_name], node_name) #add output to list keyed by node_name node_to_output_map = GLOBAL_MAP.get_ele(NNDCT_KEYS.NODE_TO_OUTPUT_MAP) if not node_name in node_to_output_map: node_to_output_map[node_name] = [output_name] else: node_to_output_map[node_name].append(output_name) if isinstance(output, str): _do_map(output, node_name)
null
23,366
from enum import Enum def readable_num(number): s = '' if number < 0: s += '-' number = -number if number < 1000: s += '%d' % number elif number > 1e15: s += '%0.3G' % number else: units = 'KMGT' unit_index = 0 while number > 1000000: number /= 1000 unit_index += 1 s += '%.2f%s' % (number / 1000.0, units[unit_index]) return s
null
23,367
from enum import Enum def print_table(header, rows): if any(len(row) != len(header) for row in rows): raise ValueError('Column length must be equal to headers') column_widths = [len(field) for field in header] for row in rows: for i, field in enumerate(row): column_widths[i] = max(len(str(field)), column_widths[i]) spaces_between_columns = 1 current_pos = 0 column_positions = [] for i in range(len(column_widths)): column_positions.append(current_pos + column_widths[i] + spaces_between_columns) current_pos = column_positions[-1] line_length = column_positions[-1] def print_row(fields, positions): line = '' for i in range(len(fields)): if i > 0: line = line[:-1] + ' ' line += str(fields[i]) line = line[:positions[i]] line += ' ' * (positions[i] - len(line)) print(line) print('=' * line_length) print_row(header, column_positions) print('=' * line_length) for row in rows: print_row(row, column_positions) print('-' * line_length)
null
23,373
import os import shutil import json import sys import numpy as np from .log import log_or_print from nndct_shared.base import NNDCT_KEYS, GLOBAL_MAP def copy_folder_files(new_dir, old_dir): def force_create_dir(dir_name, copy_from_dir=None): if os.path.exists(dir_name): shutil.rmtree(dir_name) os.makedirs(dir_name) if copy_from_dir: copy_folder_files(dir_name, copy_from_dir)
null
23,374
import os import shutil import json import sys import numpy as np from .log import log_or_print from nndct_shared.base import NNDCT_KEYS, GLOBAL_MAP def create_work_dir(dir_name): if not os.path.exists(dir_name): os.makedirs(dir_name)
null
23,383
import sys import os import logging import io from nndct_shared.base import NNDCT_KEYS, GLOBAL_MAP, NNDCT_DEBUG_LVL def obj_to_str(obj): def nndct_debug_print(string, title='', level=1): def nndct_details_debug(obj, title, level=NNDCT_DEBUG_LVL.DETAILS): nndct_debug_print( "\n********************* <Start : {}> *********************\n{}".format( title, obj_to_str(obj)), level=level) nndct_debug_print(title, title='End', level=level)
null
23,391
import math from nndct_shared.base import NNDCT_OP def get_avgpool_dpu_coeff(kernel): scale = 1.0 if kernel == [3, 3]: scale = 9.0 * 7.0 / 64.0 elif kernel == [5, 5]: scale = 25.0 * 10.0 / 256.0 elif kernel in [[6, 6], [3, 6], [6, 3]]: scale = 36.0 * 7.0 / 256.0 elif kernel == [7, 7]: scale = 49.0 * 21.0 / 1024.0 elif kernel == [14, 14]: scale = 196.0 * 21.0 / 4096.0 else: rec = kernel[0] * kernel[1] max_factor = math.ceil(math.log(rec * 128,2)) diff = 1.0 multi_factor = 0.0 shift_factor = 0.0 for shift_factor_ in range(max_factor): factor = round((2 ** shift_factor_)/rec) diff_ = abs(factor / (2 ** shift_factor_) - 1/rec) if diff_ < diff: multi_factor = factor diff = diff_ shift_factor = shift_factor_ scale = rec * multi_factor / (2 ** shift_factor) return scale
null
23,392
import h5py import json from nndct_shared.nndct_graph.base_tensor import Tensor class GraphHDF5Saver(): def __init__(self, nndct_graph): def get_node_config(self, node): def get_model_config(self): def save(self, hdf5_path): def save_graph(nndct_graph, hdf5_path='graph.hdf5'): GraphHDF5Saver(nndct_graph).save(hdf5_path)
null
23,393
import copy import gc import inspect import json import numpy as np import os import random import tensorflow as tf import types from tensorflow.python.distribute import distribution_strategy_context as ds_context from nndct_shared.pruning import errors from nndct_shared.pruning import pruner as pruner_lib from nndct_shared.pruning import pruning_lib from nndct_shared.pruning import search from nndct_shared.pruning import sensitivity as sens from nndct_shared.utils import common from tf_nndct.graph import parser from tf_nndct.graph.ops import OpTypes from tf_nndct.pruning import pruning_impl from tf_nndct.utils import generic_utils from tf_nndct.utils import keras_utils as ku from tf_nndct.utils import logging from tf_nndct.utils import tensor_utils from nndct_shared.pruning.utils import generate_indices_group from nndct_shared.pruning.pruning_lib import is_grouped_conv, is_depthwise_conv import os if not os.path.exists(BASE_DIR): os.makedirs(BASE_DIR) def _is_debug(): return os.environ.get('VAI_OPTIMIZER_DEBUG', None) == '1'
null
23,394
import copy import gc import inspect import json import numpy as np import os import random import tensorflow as tf import types from tensorflow.python.distribute import distribution_strategy_context as ds_context from nndct_shared.pruning import errors from nndct_shared.pruning import pruner as pruner_lib from nndct_shared.pruning import pruning_lib from nndct_shared.pruning import search from nndct_shared.pruning import sensitivity as sens from nndct_shared.utils import common from tf_nndct.graph import parser from tf_nndct.graph.ops import OpTypes from tf_nndct.pruning import pruning_impl from tf_nndct.utils import generic_utils from tf_nndct.utils import keras_utils as ku from tf_nndct.utils import logging from tf_nndct.utils import tensor_utils from nndct_shared.pruning.utils import generate_indices_group from nndct_shared.pruning.pruning_lib import is_grouped_conv, is_depthwise_conv keras = tf.keras def add_pruning_mask(instance): def build(self, input_shape): type(self).build(self, input_shape) weight_vars, mask_vars, = [], [] # For each of the weights, add mask variables. for weight in self.weights: # res2a_branch2a/kernel:0 -> kernel weight_name = weight.name.split('/')[-1].split(':')[0] mask = self.add_weight( weight_name + '_mask', shape=weight.shape, initializer=keras.initializers.get('ones'), dtype=weight.dtype, trainable=False, aggregation=tf.VariableAggregation.MEAN) weight_vars.append(weight) mask_vars.append(mask) pruning_vars = list(zip(weight_vars, mask_vars)) # Create a pruning object self.pruning_obj = pruning_impl.Pruning(pruning_vars) def call(self, inputs, training=None, **kwargs): if training is None: training = keras.backend.learning_phase() # Always execute the op that performs weights = weights * mask # Relies on UpdatePruningStep callback to ensure the weights # are sparse after the final backpropagation. # # self.add_update does nothing during eager execution. self.add_update(self.pruning_obj.weight_mask_op()) # TODO(evcu) remove this check after dropping py2 support. In py3 getargspec # is deprecated. layer_call = type(self).call if hasattr(inspect, 'getfullargspec'): args = inspect.getfullargspec(layer_call).args else: args = inspect.getargspec(layer_call).args # Propagate the training bool to the underlying layer if it accepts # training as an arg. if 'training' in args: return layer_call(self, inputs, training=training, **kwargs) return layer_call(self, inputs, **kwargs) instance.build = types.MethodType(build, instance) instance.call = types.MethodType(call, instance) return instance
null
23,395
import copy import gc import inspect import json import numpy as np import os import random import tensorflow as tf import types from tensorflow.python.distribute import distribution_strategy_context as ds_context from nndct_shared.pruning import errors from nndct_shared.pruning import pruner as pruner_lib from nndct_shared.pruning import pruning_lib from nndct_shared.pruning import search from nndct_shared.pruning import sensitivity as sens from nndct_shared.utils import common from tf_nndct.graph import parser from tf_nndct.graph.ops import OpTypes from tf_nndct.pruning import pruning_impl from tf_nndct.utils import generic_utils from tf_nndct.utils import keras_utils as ku from tf_nndct.utils import logging from tf_nndct.utils import tensor_utils from nndct_shared.pruning.utils import generate_indices_group from nndct_shared.pruning.pruning_lib import is_grouped_conv, is_depthwise_conv keras = tf.keras def save_weights(model, filepath, overwrite=True, save_format=None, options=None): filepath = generic_utils.path_to_string(filepath) filepath_is_h5 = ( filepath.endswith('.h5') or filepath.endswith('.keras') or filepath.endswith('.hdf5')) if save_format is None: if filepath_is_h5: save_format = 'h5' else: save_format = 'tf' else: user_format = save_format.lower().strip() if user_format in ('tensorflow', 'tf'): save_format = 'tf' elif user_format in ('hdf5', 'h5', 'keras'): save_format = 'h5' else: raise errors.OptimizerDataFormatError( 'Unknown format "%s". Was expecting one of {"tf", "h5"}.' % (save_format,)) if save_format == 'h5': raise errors.OptimizerDataFormatError(( 'HDF5 format is not allowed for sparse model, please ' 'use "tf" format. See ' 'https://www.tensorflow.org/api_docs/python/tf/keras/Model#save_weights' )) keras.Model.save_weights(model, filepath, overwrite, save_format, options)
null
23,396
import copy import gc import inspect import json import numpy as np import os import random import tensorflow as tf import types from tensorflow.python.distribute import distribution_strategy_context as ds_context from nndct_shared.pruning import errors from nndct_shared.pruning import pruner as pruner_lib from nndct_shared.pruning import pruning_lib from nndct_shared.pruning import search from nndct_shared.pruning import sensitivity as sens from nndct_shared.utils import common from tf_nndct.graph import parser from tf_nndct.graph.ops import OpTypes from tf_nndct.pruning import pruning_impl from tf_nndct.utils import generic_utils from tf_nndct.utils import keras_utils as ku from tf_nndct.utils import logging from tf_nndct.utils import tensor_utils from nndct_shared.pruning.utils import generate_indices_group from nndct_shared.pruning.pruning_lib import is_grouped_conv, is_depthwise_conv def generate_indices_group(indices: List[int], dim_size: int, groups: int) -> List[List[int]]: indices_set = set(indices) interval: int = dim_size // groups start_idx = 0 end_idx = interval ret: List[List[int]] = [] while start_idx < dim_size: idx_group: List[int] = [] for i in range(start_idx, end_idx): if i in indices_set: idx_group.append(i - start_idx) ret.append(idx_group) start_idx = end_idx end_idx += interval return ret The provided code snippet includes necessary dependencies for implementing the `_sparsify_tensor` function. Write a Python function `def _sparsify_tensor(t: np.ndarray, out_channels, in_channels, groups: int = 1)` to solve the following problem: Fill 0 in removed channels. Here is the function: def _sparsify_tensor(t: np.ndarray, out_channels, in_channels, groups: int = 1): """Fill 0 in removed channels.""" tensor = copy.deepcopy(t) dim_size = len(tensor.shape) if dim_size == 0: # e.g layers.Normalization the param 'count' is a num so the shape = 0 return tensor assert dim_size in [1, 2, 4, 5] # 5: 3D conv # weight format in tensorflow: HWIO/IO/DHWIO if groups == 1: if out_channels: tensor[..., out_channels] = 0.0 if in_channels and dim_size > 1: if dim_size == 2: tensor[in_channels, :] = 0.0 else: tensor[..., in_channels, :] = 0.0 else: out_dims_group = generate_indices_group(out_channels, tensor.shape[-1], groups) in_dims_group = generate_indices_group( in_channels, tensor.shape[-2] * groups, groups) if dim_size > 1 else [[]] * groups parts = np.split(tensor, groups, axis=-1) sparse_parts: List[torch.Tensor] = [] for part, o, i in zip(parts, out_dims_group, in_dims_group): sparse_parts.append(_sparsify_tensor(part, o, i)) tensor = np.concatenate(sparse_parts, axis=-1) return tensor
Fill 0 in removed channels.
23,397
import copy import gc import inspect import json import numpy as np import os import random import tensorflow as tf import types from tensorflow.python.distribute import distribution_strategy_context as ds_context from nndct_shared.pruning import errors from nndct_shared.pruning import pruner as pruner_lib from nndct_shared.pruning import pruning_lib from nndct_shared.pruning import search from nndct_shared.pruning import sensitivity as sens from nndct_shared.utils import common from tf_nndct.graph import parser from tf_nndct.graph.ops import OpTypes from tf_nndct.pruning import pruning_impl from tf_nndct.utils import generic_utils from tf_nndct.utils import keras_utils as ku from tf_nndct.utils import logging from tf_nndct.utils import tensor_utils from nndct_shared.pruning.utils import generate_indices_group from nndct_shared.pruning.pruning_lib import is_grouped_conv, is_depthwise_conv The provided code snippet includes necessary dependencies for implementing the `_prune_tensor` function. Write a Python function `def _prune_tensor(t: np.ndarray, out_channels, in_channels)` to solve the following problem: Remove dimensions by giving channels. Here is the function: def _prune_tensor(t: np.ndarray, out_channels, in_channels): """Remove dimensions by giving channels.""" tensor = copy.deepcopy(t) dim_size = len(tensor.shape) # weight format in tensorflow: HWIO/IO assert dim_size in [1, 2, 4] if dim_size == 1: out_axis, in_axis = 0, None elif dim_size == 2: out_axis, in_axis = 1, 0 else: out_axis, in_axis = 3, 2 if out_channels: tensor = np.delete(tensor, out_channels, axis=out_axis) if in_channels and in_axis is not None: tensor = np.delete(tensor, in_channels, axis=in_axis) return tensor
Remove dimensions by giving channels.
23,398
import copy import gc import inspect import json import numpy as np import os import random import tensorflow as tf import types from tensorflow.python.distribute import distribution_strategy_context as ds_context from nndct_shared.pruning import errors from nndct_shared.pruning import pruner as pruner_lib from nndct_shared.pruning import pruning_lib from nndct_shared.pruning import search from nndct_shared.pruning import sensitivity as sens from nndct_shared.utils import common from tf_nndct.graph import parser from tf_nndct.graph.ops import OpTypes from tf_nndct.pruning import pruning_impl from tf_nndct.utils import generic_utils from tf_nndct.utils import keras_utils as ku from tf_nndct.utils import logging from tf_nndct.utils import tensor_utils from nndct_shared.pruning.utils import generate_indices_group from nndct_shared.pruning.pruning_lib import is_grouped_conv, is_depthwise_conv class IterativePruningRunner(PruningRunner): def __init__(self, model: keras.Model, input_signature: tf.TensorSpec): super(IterativePruningRunner, self).__init__(model, input_signature) self._sens_path = os.path.join(_VAI_DIR, self._graph.name + '.sens') self._latest_spec = os.path.join(_VAI_DIR, 'latest_spec') def ana(self, eval_fn, excludes=None, forced=False, with_group_conv: bool = False): """Performs model analysis. The analysis result will be saved in '.vai' directory and this cached result will be used directly in subsequent calls unless 'forced' is set to True. Arguments: eval_fn: Callable object that takes a keras.Model object as its first argument and returns the evaluation score. excludes: excludes: A list of node name or torch module to be excluded from pruning. forced: When set to True, forced to run model analysis instead of using cached analysis result. """ if not forced: net_sens = self._load_analysis_result() if net_sens is not None: logging.info( 'Using cached analysis result. If you want to re-analyze the model, set forced=True' ) return excluded_nodes = self._get_exclude_nodes(excludes) if excludes else [] self._ana_pre_check(eval_fn, excluded_nodes, with_group_conv) self._ana(eval_fn, excluded_nodes, with_group_conv) def prune(self, ratio=None, threshold=None, spec_path=None, excludes=None, mode='sparse', channel_divisible=2): """Prune the baseline model and returns a sparse model. The degree of model reduction can be specified in three ways: ratio, threshold or pruning specification. The first method should be used in preference, the latter two are more suitable for experiments with manual tuning. Arguments: ratio: The expected percentage of MACs reduction of baseline model. This is just a hint value and the actual MACs reduction not strictly equals to this value. threshold: Relative proportion of model performance loss between baseline model and the pruned model. spec_path: Pruning specfication path used to prune the model. excludes: A list of layer name or layer instance to be excluded from pruning. mode: In which mode the pruned model is generated. Should be either 'sparse' or 'slim'. Must be 'sparse' in iterative pruning loop. channel_divisible: The number of remaining channels in the pruned layer can be divided by channel_divisble. Returns: A sparse or a slim model according to given 'mode'. """ if ratio or threshold: net_sens = self._load_analysis_result() if net_sens is None: raise errors.OptimizerNoAnaResultsError( "Must call ana() before model pruning.") excluded_nodes = self._get_exclude_nodes(excludes) if excludes else [] if ratio: if not isinstance(ratio, float): raise errors.OptimizerInvalidArgumentError( 'Expected "ratio" to be float, but got {}({})'.format( ratio, type(ratio))) logging.info('Pruning ratio = {}'.format(ratio)) spec = self._spec_from_ratio(net_sens, ratio, excluded_nodes) spec.channel_divisible = channel_divisible target = ('ratio', ratio) elif threshold: if not isinstance(threshold, float): raise errors.OptimizerInvalidArgumentError( 'Expected "threshold" to be float, but got {}({})'.format( threshold, type(threshold))) logging.info('Pruning threshold = {}'.format(threshold)) spec = self._spec_from_threshold(net_sens, threshold, excluded_nodes) spec.channel_divisible = channel_divisible target = ('threshold', threshold) elif spec_path: logging.info('Pruning specification = {}'.format(spec_path)) spec = self._spec_from_path(spec_path) else: raise errors.OptimizerInvalidArgumentError( 'One of [ratio, threshold, spec_path] must be given.') if ratio or threshold: filename = '{}_{}_{}.spec'.format(self._graph.name, *target) spec_path = os.path.join(_VAI_DIR, filename) json.dump(spec.serialize(), open(spec_path, 'w'), indent=2) logging.info('Pruning specification saves in {}'.format(spec_path)) with open(self._latest_spec, 'w') as f: json.dump(spec_path, f) if mode == 'slim': logging.warn( ('UserWarning: slim model can not be used for the next iteration. ' 'Set "mode=sparse" for iterative purpose.')) model = self._prune(spec, mode) if mode == 'sparse': model.save_weights = types.MethodType(save_weights, model) logging.info('Pruning summary:') self._summary(spec) return model def get_slim_model(self, spec_path=None): """Get a slim model from a sparse model. Use the latest pruning specification to do this transformation by default. If the sparse model was not generated from the latest specification, a specification path can be provided explicitly. Arguments: spec_path: Path of pruning specification used to transform a sparse model to a slim model. Returns: A shrinked slim model. """ spec_path = spec_path or json.load(open(self._latest_spec)) logging.info('Get slim model from specification {}'.format(spec_path)) return self._prune(self._spec_from_path(spec_path), 'slim') def _load_analysis_result(self): return sens.load_sens(self._sens_path) if os.path.exists( self._sens_path) else None def _ana_pre_check(self, eval_fn, excludes, with_group_conv: bool = False): """Prune model but not test it to check if all pruning steps can pass.""" logging.info('Pre-checking for analysis...') groups = pruning_lib.group_nodes(self._graph, excludes, with_group_conv) spec = pruning_lib.PruningSpec.from_node_groups(groups, 0.9) pruner = pruner_lib.ChannelPruner(self._graph) pruned_graph, net_pruning = pruner.prune(spec) model = self._get_slim_model(pruned_graph, net_pruning) eval_fn(model) def _ana(self, eval_fn, excludes, with_group_conv: bool = False): analyser = sens.ModelAnalyser(self._graph, excludes, with_group_conv) steps = analyser.steps() for step in range(steps): model = self._prune(analyser.spec(step)) eval_res = eval_fn(model) if not isinstance(eval_res, (int, float)): raise errors.OptimizerInvalidArgumentError( 'int or float expected, but got {}'.format(type(eval_res))) analyser.record(step, eval_res) logging.info('Analysis complete %d/%d' % (step + 1, steps)) del model keras.backend.clear_session() tf.compat.v1.reset_default_graph() analyser.save(self._sens_path) def _spec_from_ratio(self, net_sens, ratio, excludes): logging.info('Searching for appropriate ratio for each layer...') flops = ku.try_count_flops(self._model) target_flops = (1 - ratio) * flops flops_tolerance = 1e-2 min_th = 1e-5 max_th = 1 - min_th num_attempts, max_attempts = 0, 100 prev_spec = None cur_spec = None while num_attempts < max_attempts: prev_spec = cur_spec num_attempts += 1 threshold = (min_th + max_th) / 2 cur_spec = self._spec_from_threshold(net_sens, threshold, excludes) slim_model = self._prune(cur_spec, mode='slim') current_flops = ku.try_count_flops(slim_model) error = abs(target_flops - current_flops) / target_flops if error < flops_tolerance: break if current_flops < target_flops: max_th = threshold else: min_th = threshold return cur_spec def _spec_from_threshold(self, net_sens, threshold, excludes): groups = net_sens.prunable_groups_by_threshold(threshold, excludes) return pruning_lib.PruningSpec(groups) def _spec_from_path(self, path): return pruning_lib.PruningSpec.deserialize(json.load(open(path, 'r'))) def _summary(self, spec): orig_flops = common.readable_num(ku.try_count_flops(self._model)) orig_params = common.readable_num(ku.try_count_params(self._model)) slim_model = self._prune(spec, 'slim') current_flops = common.readable_num(ku.try_count_flops(slim_model)) current_params = common.readable_num(ku.try_count_params(slim_model)) header_fields = ['Metric', 'Baseline', 'Pruned'] flops_fields = ['FLOPs', orig_flops, current_flops] params_fields = ['Params', orig_params, current_params] common.print_table(header_fields, [flops_fields, params_fields]) class OneStepPruningRunner(PruningRunner): """Implements channel pruning at the model level.""" def __init__(self, model: keras.Model, input_signature: tf.TensorSpec): """Concrete example: ```python model = MyModel() pruner = OneStepPruningPruner(model, tf.TensorSpec(input_shape, tf.float32)) model = pruner.search_subnets(0.2, train_fn, eval_fn, 1000) ``` Arguments: model (keras.Model): Model to prune. input_signature(tuple or list): The input specifications of model. """ super(OneStepPruningRunner, self).__init__(model, input_signature) def _searcher_saved_path(self, ratio): return os.path.join(_VAI_DIR, self._graph.name + '_search_{}'.format(ratio)) def _random_ratios(self, count): return [random.random() for _ in range(count)] def search_subnets(self, ratio, train_fn, eval_fn, num_iterations, excludes=None, config=None, with_group_conv: bool = False): if not isinstance(ratio, float): raise errors.OptimizerInvalidArgumentError( 'Expect "ratio" to be float, but got {}({})'.format( ratio, type(ratio))) excluded_nodes = self._get_exclude_nodes(excludes) if excludes else [] groups = pruning_lib.group_nodes(self._graph, excluded_nodes, with_group_conv) searcher = search.SubnetSearcher(groups) score = eval_fn(self._model) base_flops = ku.try_count_flops(self._model) searcher.set_supernet(score, base_flops) searcher_saved_path = self._searcher_saved_path(ratio) for i in range(num_iterations): ratios = self._random_ratios(len(groups)) model = self._prune(searcher.spec(ratios)) current_flops = ku.try_count_flops(model) flops_ratio = 1 - current_flops / base_flops print('Iter {}: ratios={}, flops_ratio = {}'.format( i, ratios, flops_ratio)) eps = 0.05 if flops_ratio > ratio - eps and flops_ratio < ratio + eps: train_fn(model) score = eval_fn(model) searcher.add_subnet(ratios, score, current_flops) search.save_searcher(searcher, searcher_saved_path) logging.info('Found subnet: ratios={}, score={}'.format(ratios, score)) del model #gc.collect() keras.backend.clear_session() tf.compat.v1.reset_default_graph() logging.info('Search results saved in {}'.format(searcher_saved_path)) def get_subnet(self, ratio, index=None): searcher = search.load_searcher(self._searcher_saved_path(ratio)) subnet = searcher.subnet(index) if index else searcher.best_subnet() print('best_subnet:', subnet) spec = searcher.spec(subnet.ratios) logging.vlog(1, 'Get subnet from spec:\n{}'.format(spec)) return self._prune(spec) def get_pruning_runner(model, input_signature, method='one_step'): assert method in ['iterative', 'one_step'] cls = IterativePruningRunner if method == 'iterative' else OneStepPruningRunner return cls(model, input_signature)
null
23,399
import inspect import numpy as np from tensorflow import keras from nndct_shared.pruning import errors from tf_nndct.pruning import pruning_impl class PruneMaskedWeight(keras.layers.Wrapper): """This wrapper augments a keras layer so the weight tensor may be pruned. This wrapper implements magnitude-based pruning of the weight tensors. Magnitude-based pruning achieves a target sparsity (s% of zeros) for a given weight tensor by monitoring the distribution of the absolute values of the weight tensor and determining the weight value (referred to as threshold) below which s% of elements lie. For every weight tensor being pruned, the wrapper maintains an identically shaped tensor (referred to as mask) which stores 0 if the weight value lies below the threshold. The mask and thresholds are computed during the training based on the evolution of the weight values. Block sparse patterns: For certain SIMD hardware architectures, it may be beneficial to induce spatially correlated sparsity. To train models in which the weight tensors have block sparse structure, the pruning wrapper can be configured with the block_height and block_width configuration parameters set to the desired block configuration (2x2, 4x4, 4x1, 1x8, etc). This is applicable to rank-2 weight tensor only and the tensor partitioned into non-overlapping blocks of size [block_height, block_dim]. Either the average or max absolute value in this block is taken as a proxy for the entire block (set by block_pooling_function configuration parameter) while computing the distribution of the weight values and the threshold for pruning. Custom keras layers: The pruning wrapper can also be applied to a user-defined keras layer. Such a layer may contain one or more weight tensors that may be pruned. To apply pruning wrapper to such layers, the layer should be a `PrunableLayer` instance or, more directly, user should define a `get_prunable_weights` method for the layer (Check the pruning_wrapper_test.CustomLayerPrunable for more details about how to define a user-defined prunable layer). Sparsity function: The target sparsity for the weight tensors are set through the pruning_schedule parameter of the pruning wrapper. The user must create a python callable that returns a scalar tensorflow tensor and pass this callable to the sparsity_function parameter. This scalar tensor contains the target sparsity value for the weight tensors in the layer. The wrapper provides the following pre-built sparsity functions: """ def __init__(self, layer, **kwargs): """Create a pruning wrapper for a keras layer. Args: layer: The keras layer to be pruned. **kwargs: Additional keyword arguments to be passed to the keras layer. """ # An instance of the Pruning class. This class contains the logic to prune # the weights of this layer. self.pruning_obj = None # A list of all (weight,mask) tuples for this layer self.pruning_vars = [] if not isinstance(layer, tf.keras.layers.Layer): raise errors.OptimizerKerasLayerError( 'Please initialize `Prune` layer with a ' '`Layer` instance. You passed: {input}'.format(input=layer)) # TODO(pulkitb): This should be pushed up to the wrappers.py # Name the layer using the wrapper and underlying layer name. # Prune(Dense) becomes prune_dense_1 kwargs.update({'name': 'pruned_{}'.format(layer.name)}) super(PruneMaskedWeight, self).__init__(layer, **kwargs) self._track_trackable(layer, name='layer') # TODO(yunluli): Work-around to handle the first layer of Sequential model # properly. Can remove this when it is implemented in the Wrapper base # class. # # Enables end-user to prune the first layer in Sequential models, while # passing the input shape to the original layer. # # tf.keras.Sequential( # prune_masked_weight(tf.keras.layers.Dense(2, input_shape=(3,))) # ) # # as opposed to # # tf.keras.Sequential( # prune_masked_weight(tf.keras.layers.Dense(2), input_shape=(3,)) # ) # # Without this code, the pruning wrapper doesn't have an input # shape and being the first layer, this causes the model to not be # built. Being not built is confusing since the end-user has passed an # input shape. if not hasattr(self, '_batch_input_shape') and hasattr( layer, '_batch_input_shape'): self._batch_input_shape = self.layer._batch_input_shape #metrics.MonitorBoolGauge('prune_masked_weight_wrapper_usage').set( # layer.__class__.__name__) def build(self, input_shape): super(PruneMaskedWeight, self).build(input_shape) weight_vars, mask_vars, = [], [] # For each of the weights, add mask variables. for weight in self.layer.weights: # res2a_branch2a/kernel:0 -> kernel weight_name = weight.name.split('/')[-1].split(':')[0] mask = self.add_weight( weight_name + '_mask', shape=weight.shape, initializer=tf.keras.initializers.get('ones'), dtype=weight.dtype, trainable=False, aggregation=tf.VariableAggregation.MEAN) weight_vars.append(weight) mask_vars.append(mask) pruning_vars = list(zip(weight_vars, mask_vars)) # Create a pruning object self.pruning_obj = pruning_impl.Pruning(pruning_vars) def call(self, inputs, training=None, **kwargs): if training is None: training = K.learning_phase() # Always execute the op that performs weights = weights * mask # Relies on UpdatePruningStep callback to ensure the weights # are sparse after the final backpropagation. # # self.add_update does nothing during eager execution. self.add_update(self.pruning_obj.weight_mask_op()) # TODO(evcu) remove this check after dropping py2 support. In py3 getargspec # is deprecated. if hasattr(inspect, 'getfullargspec'): args = inspect.getfullargspec(self.layer.call).args else: args = inspect.getargspec(self.layer.call).args # Propagate the training bool to the underlying layer if it accepts # training as an arg. if 'training' in args: return self.layer.call(inputs, training=training, **kwargs) return self.layer.call(inputs, **kwargs) def compute_output_shape(self, input_shape): return self.layer.compute_output_shape(input_shape) def get_config(self): return super(PruneMaskedWeight, self).get_config() def from_config(cls, config): config = config.copy() deserialize_keras_object = keras.utils.deserialize_keras_object # pylint: disable=g-import-not-at-top layer = keras.layers.deserialize(config.pop('layer')) config['layer'] = layer return cls(**config) def trainable(self): return self.layer.trainable def trainable(self, value): self.layer.trainable = value def trainable_weights(self): return self.layer.trainable_weights def non_trainable_weights(self): return self.layer.non_trainable_weights + self._non_trainable_weights #@property #def updates(self): # return self.layer.updates + self._updates #@property #def losses(self): # return self.layer.losses + self._losses #def get_weights(self): # return self.layer.get_weights() #def set_weights(self, weights): # self.layer.set_weights(weights) The provided code snippet includes necessary dependencies for implementing the `collect_prunable_layers` function. Write a Python function `def collect_prunable_layers(model)` to solve the following problem: Recursively collect the prunable layers in the model. Here is the function: def collect_prunable_layers(model): """Recursively collect the prunable layers in the model.""" return [ layer for layer in model.submodules if isinstance(layer, PruneMaskedWeight) ]
Recursively collect the prunable layers in the model.
23,400
from __future__ import absolute_import from __future__ import division from __future__ import print_function import tensorflow as tf def assign(ref, value, name=None): if hasattr(tf, 'assign'): return tf.assign(ref, value, name=name) else: return ref.assign(value, name=name)
null
23,401
from __future__ import absolute_import from __future__ import division from __future__ import print_function import tensorflow as tf The provided code snippet includes necessary dependencies for implementing the `initialize_variables` function. Write a Python function `def initialize_variables(testcase)` to solve the following problem: Handle global variable initialization in TF 1.X. Arguments: testcase: instance of tf.test.TestCase Here is the function: def initialize_variables(testcase): """Handle global variable initialization in TF 1.X. Arguments: testcase: instance of tf.test.TestCase """ if hasattr(tf, 'global_variables_initializer') and not tf.executing_eagerly(): testcase.evaluate(tf.global_variables_initializer())
Handle global variable initialization in TF 1.X. Arguments: testcase: instance of tf.test.TestCase
23,402
from __future__ import absolute_import from __future__ import division from __future__ import print_function import tensorflow as tf def is_v1_apis(): return hasattr(tf, 'assign')
null
23,403
import sys from tensorflow.python.eager import context from tensorflow.python.util import nest def make_quantized(base): class Inspectable(base): class Attr: SavingOutputs = '_saving_outputs' SavedOutputs = '_saved_outputs' def __call__(self, *args, **kwargs): outputs = super(Inspectable, self).__call__(*args, **kwargs) if getattr(self, self.Attr.SavingOutputs, False) and context.executing_eagerly(): self._save(outputs) return outputs def enable_saving_outputs(self): setattr(self, self.Attr.SavingOutputs, True) def disable_saving_outputs(self): setattr(self, self.Attr.SavingOutputs, False) def _save(self, tensors): if not hasattr(self, self.Attr.SavedOutputs): setattr(self, self.Attr.SavedOutputs, []) batch_outputs = [] for tensor in nest.flatten(tensors): batch_outputs.append(tensor.numpy()) saved_outputs = getattr(self, self.Attr.SavedOutputs) saved_outputs.append(batch_outputs) def saved_outputs(self): return getattr(self, self.Attr.SavedOutputs, []) cls_name = base.__name__ quantized_cls = type(cls_name, (Inspectable,), {}) # Use base class __init__ directly so that the `KerasWriter` can inspect init args # automatically and write them out correctly. quantized_cls.__init__ = base.__init__ setattr(sys.modules[__name__], cls_name, quantized_cls) return quantized_cls
null
23,404
from tensorflow.python.ops.signal import fft_ops def rfft(input_tensor, fft_length=None, name=None): return fft_ops.rfft(input_tensor, fft_length, name)
null
23,405
from tensorflow.python.ops.signal import fft_ops def irfft(input_tensor, fft_length=None, name=None): return fft_ops.irfft(input_tensor, fft_length, name)
null
23,406
from tensorflow.python.ops.signal import fft_ops def ifft(input, name=None): return fft_ops.ifft(input, name)
null
23,407
from tensorflow.python.ops import array_ops def gather(params, indices, axis=None, batch_dims=0, name=None): return array_ops.gather_v2( params, indices, axis=axis, batch_dims=batch_dims, name=name)
null
23,408
import numpy as np import os import tensorflow as tf from collections import OrderedDict from tensorflow.keras import activations from tensorflow.keras import layers as keras_layers from tensorflow.python.framework import dtypes as tf_dtypes from tensorflow.python.util import nest from tensorflow.python.util import tf_inspect from tf_nndct import layers as nndct_layers from tf_nndct import ops as nndct_ops from tf_nndct.graph import OpTypes from tf_nndct.graph import dtypes from tf_nndct.graph import ops from tf_nndct.graph import utils from tf_nndct.ops.signal import fft_ops from tf_nndct.quantization import utils as quant_utils from tf_nndct.utils import generic_utils from tf_nndct.utils import keras_utils from tf_nndct.utils import registry from tf_nndct.utils import tf_utils def _get_module(object): module = tf_inspect.getmodule(object) module_full_name = module.__name__ if '.' in module_full_name: pkg, module_name = module_full_name.rsplit('.', 1) else: pkg = module_full_name module_name = module_full_name return (module, pkg, module_name)
null
23,409
from __future__ import absolute_import from __future__ import division from __future__ import print_function import os import tensorflow as tf from google.protobuf import text_format from tensorflow.core.framework import graph_pb2 from tf_nndct.graph import ops from tf_nndct.utils import generic_utils from tf_nndct.utils import viz def write_proto(path, message, as_text=False): dir_name = os.path.dirname(path) generic_utils.mkdir_if_not_exist(dir_name) if dir_name: os.makedirs(dir_name, exist_ok=True) if as_text: with open(path, "w") as f: f.write(text_format.MessageToString(message)) else: with open(path, "wb") as f: f.write(message.SerializeToString()) def write_text_proto(path, message): write_proto(path, message, as_text=True)
null
23,410
from __future__ import absolute_import from __future__ import division from __future__ import print_function import os import tensorflow as tf from google.protobuf import text_format from tensorflow.core.framework import graph_pb2 from tf_nndct.graph import ops from tf_nndct.utils import generic_utils from tf_nndct.utils import viz def op_param_by_name(op, name): for param in op.ParamName: if param.value == name: return param return None
null
23,411
from __future__ import absolute_import from __future__ import division from __future__ import print_function import os import tensorflow as tf from google.protobuf import text_format from tensorflow.core.framework import graph_pb2 from tf_nndct.graph import ops from tf_nndct.utils import generic_utils from tf_nndct.utils import viz def stringfy_to_write(x): if isinstance(x, str): x = "'{}'".format(x) return '{}'.format(x)
null
23,412
import json import os import tensorflow as tf from tensorflow.keras import layers from tensorflow.core.protobuf import config_pb2 from tensorflow.lite.python.util import run_graph_optimizations as _run_graph_optimizations from tf_nndct.graph import OpTypes from tf_nndct.graph import converter from tf_nndct.graph import ops from tf_nndct.graph import refiner from tf_nndct.graph import utils from tf_nndct.utils import generic_utils from tf_nndct.utils import keras_utils from tf_nndct.utils import logging from tf_nndct.utils import tf_utils keras = tf.keras def _parent_scope(scope): # Given 'model/dense/MatMul', return 'model/dense'. return scope.rsplit('/', 1)[0] The provided code snippet includes necessary dependencies for implementing the `belongs_to_keras_layer` function. Write a Python function `def belongs_to_keras_layer(op, scope_to_layer)` to solve the following problem: Get the keras layer the given op is generated from. Returns None if op does not belong to any layer. Trace back from current scope to parent scope recursively until it reaches the outermost scope. Here is the function: def belongs_to_keras_layer(op, scope_to_layer): """Get the keras layer the given op is generated from. Returns None if op does not belong to any layer. Trace back from current scope to parent scope recursively until it reaches the outermost scope. """ if not scope_to_layer: return None layer = None scope = op.name while True: if scope in scope_to_layer: layer = scope_to_layer[scope][0] break parent_scope = _parent_scope(scope) # Already to the outtest scope. if parent_scope == scope: break scope = parent_scope # Lambda layer is a wrapper, we need to parse ops in the layer individually. if type(layer) == keras.layers.Lambda or isinstance(layer, keras.Sequential): layer = None return layer
Get the keras layer the given op is generated from. Returns None if op does not belong to any layer. Trace back from current scope to parent scope recursively until it reaches the outermost scope.
23,413
import json import os import tensorflow as tf from tensorflow.keras import layers from tensorflow.core.protobuf import config_pb2 from tensorflow.lite.python.util import run_graph_optimizations as _run_graph_optimizations from tf_nndct.graph import OpTypes from tf_nndct.graph import converter from tf_nndct.graph import ops from tf_nndct.graph import refiner from tf_nndct.graph import utils from tf_nndct.utils import generic_utils from tf_nndct.utils import keras_utils from tf_nndct.utils import logging from tf_nndct.utils import tf_utils import logging as _logging from logging import DEBUG from logging import ERROR from logging import FATAL from logging import INFO from logging import WARN from logging import NOTSET The provided code snippet includes necessary dependencies for implementing the `get_layer_inbound_nodes` function. Write a Python function `def get_layer_inbound_nodes(layer_parent_pairs)` to solve the following problem: Get layer's inbound nodes. The config of a layer does not include connectivity information, nor the layer class name. These are handled by keras.Model. So we extract them from model's config and associate them to the corresponding layer. Here is the function: def get_layer_inbound_nodes(layer_parent_pairs): """Get layer's inbound nodes. The config of a layer does not include connectivity information, nor the layer class name. These are handled by keras.Model. So we extract them from model's config and associate them to the corresponding layer. """ layer_inbound_nodes = {} model = None # Get a keras model which is a top-level layer. for layer, parent_layer in layer_parent_pairs: if parent_layer is None: model = layer break if getattr(model, '_is_graph_network', None): # Only graph network has get_config. model_config = model.get_config() logging.vlog(4, 'model_config: {}'.format(model_config)) if 'layers' in model_config: layers_config = model_config['layers'] for config in layers_config: if 'inbound_nodes' in config: layer_inbound_nodes[config['name']] = config['inbound_nodes'] return layer_inbound_nodes
Get layer's inbound nodes. The config of a layer does not include connectivity information, nor the layer class name. These are handled by keras.Model. So we extract them from model's config and associate them to the corresponding layer.
23,414
import imp from tensorflow import keras from tensorflow.python.ops import array_ops from tensorflow.python.util import nest from nndct_shared.pruning import pruning_lib from tf_nndct.graph import OpTypes from tf_nndct.graph import parser from tf_nndct.graph import utils from tf_nndct.graph import writer as writer_lib from tf_nndct.layers import base_layer from tf_nndct.utils import keras_utils from tf_nndct.utils import logging from tf_nndct.utils import tensor_utils class KerasBuilder(object): def __init__(self, graph): self._graph = graph def build(self, filepath, quantized=False, as_layer=False): class_name = 'Model' base_class = keras.Model call_fn_name = 'call' if as_layer: if quantized: base_class = base_layer.Layer call_fn_name = '_internal_call' else: base_class = keras.layers.Layer class_spec = ClassSpec(class_name, base_class, call_fn_name, quantized) writer = writer_lib.GraphCodeGenerator(self._graph, class_spec) layer_to_node = writer.write(filepath) # TODO(yuwang): Use code below. #py_module_name = "_".join(["nndct", module_name]) #spec = importlib.util.spec_from_file_location(py_module_name, filepath) #py_module = importlib.util.module_from_spec(spec) #sys.modules[py_module_name] = py_module #spec.loader.exec_module(py_module) #rebuilt_module = py_module.__dict__[module_name]() loaded_module = imp.load_source('nndct_rebuilt_model', filepath) rebuilt_model = getattr(loaded_module, class_name)() dummy_inputs = [] for spec in nest.flatten(self._graph.input_signature): logging.vlog(1, spec) dummy_inputs.append(array_ops.ones(spec.shape, dtype=spec.dtype)) dummy_inputs = nest.pack_sequence_as(self._graph.input_signature, dummy_inputs) # Call the subclassed model once to build the model (mainly to create the weights) #input_data = dummy_inputs if len(dummy_inputs) > 1 else dummy_inputs[0] rebuilt_model(*dummy_inputs) layer_nodes = [] # Reload weights for layer_name, node in layer_to_node.items(): layer = getattr(rebuilt_model, layer_name) # If there is a ParamName definition, then map ParamName's member to # keras layer's param; If there is no ParamName, # then export params in the order they are saved in the op. weights = [] if hasattr(node.op, 'ParamName'): named_weights = keras_utils.get_named_weights(layer) for name in named_weights: param = utils.op_param_by_name(node.op, name) if not param: raise ValueError('Can not get value of "{}" in node({})'.format( name, node.name)) ndarray = tensor_utils.param_to_tf_numpy(node.op.get_param(param)) if pruning_lib.is_transpose_conv(node.op): ndarray = tensor_utils.transposeconv_weight_dim_trans(ndarray) weights.append(ndarray) logging.vlog( 2, 'Reload weights of {}: name={}, shape={}'.format( layer.name, name, ndarray.shape)) else: for name, tensor in node.op.params.items(): ndarray = tensor_utils.param_to_tf_numpy(tensor) if pruning_lib.is_transpose_conv(node.op): ndarray = tensor_utils.transposeconv_weight_dim_trans(ndarray) weights.append(ndarray) logging.vlog( 2, "Reload weights of {}: name={}, shape={}".format( layer.name, name, ndarray.shape)) layer_nodes.append((layer, node)) if weights: layer.set_weights(weights) return rebuilt_model, layer_nodes def rebuild_model(model, input_signature, path=None): graph = parser.from_keras_model(model, input_signature) builder = KerasBuilder(graph) if not path: path = '{}_rebuilt.py'.format(model.name) rebuilt_model, layer_names = builder.build(path) return rebuilt_model
null
23,415
import numpy as np from enum import Enum from tensorflow.core.framework import types_pb2 def from_numpy(dtype): return _NP_TO_NNDCT[dtype]
null
23,416
import numpy as np from enum import Enum from tensorflow.core.framework import types_pb2 _NNDCT_TO_NP = { DType.FLOAT: np.float32, DType.FLOAT16: np.float16, DType.DOUBLE: np.float64, DType.INT32: np.int32, DType.INT16: np.int16, DType.INT8: np.int8, DType.UINT8: np.uint8, DType.UINT16: np.uint16, DType.INT64: np.int64, DType.UINT64: np.uint64, DType.BOOL: bool } def to_numpy(dtype): return _NNDCT_TO_NP[dtype]
null
23,417
from __future__ import absolute_import from __future__ import division from __future__ import print_function import copy import numpy as np import tensorflow as tf from tensorflow.keras import activations from tensorflow.keras import layers from tensorflow.python.util import nest from tf_nndct.graph import OpTypes from tf_nndct.graph import dtypes from tf_nndct.graph import op_def from tf_nndct.graph import ops from tf_nndct.graph import utils from tf_nndct.utils import generic_utils from tf_nndct.utils import keras_utils from tf_nndct.utils import registry from tf_nndct.utils import tensor_utils from tf_nndct.utils import tf_utils _tf_type_to_nndct = { 'Add': op_def.TFAdd, 'AddV2': op_def.TFAdd, 'BiasAdd': op_def.TFBiasAdd, 'Identity': op_def.TFIdentity, 'NoOp': op_def.TFNoOp, 'Reshape': op_def.TFReshape, 'Sigmoid': op_def.TFSigmoid, 'Tanh': op_def.TFTanh, 'GatherV2': op_def.TFGather, 'RFFT': op_def.TFRFFT, 'ComplexAbs': op_def.TFComplexAbs, 'Angle': op_def.TFAngle, 'Exp': op_def.TFExp, 'IRFFT': op_def.TFIRFFT, 'Pad': op_def.TFPad, 'Transpose': op_def.TFTranspose, 'Sum': op_def.TFSum, 'reshape': op_def.TFReshape, 'concat': op_def.TFConcat, 'ConcatV2': op_def.TFConcat, '__operators__.add': op_def.TFAdd, } def create_node(name, op, input_names, output_names): def convert_simple_tf_op(node): op = _tf_type_to_nndct[node.type]() return create_node(node.name, op, node.input_names, node.output_names)
null
23,418
from __future__ import absolute_import from __future__ import division from __future__ import print_function import copy import numpy as np import tensorflow as tf from tensorflow.keras import activations from tensorflow.keras import layers from tensorflow.python.util import nest from tf_nndct.graph import OpTypes from tf_nndct.graph import dtypes from tf_nndct.graph import op_def from tf_nndct.graph import ops from tf_nndct.graph import utils from tf_nndct.utils import generic_utils from tf_nndct.utils import keras_utils from tf_nndct.utils import registry from tf_nndct.utils import tensor_utils from tf_nndct.utils import tf_utils if tf_utils.is_tf_version_greater_equal('2.9.0'): from keras.layers.rnn import lstm as recurrent_v2 from keras.layers.rnn import lstm_v1 as recurrent elif tf_utils.is_tf_version_greater_equal('2.6'): # Keras is seperate from tensorflow since tf 2.6 from keras.layers import recurrent from keras.layers import recurrent_v2 else: from tensorflow.python.keras.layers import recurrent from tensorflow.python.keras.layers import recurrent_v2 class OpBuilder(object): _OP_COUNT = {} # TODO(yuwang): Maybe use op object instead of op class ? def __init__(self, op_cls, config=None, params=None, *args, **kwargs): self._op = op_cls(*args, **kwargs) self._config = config if config else {} self._params = params if params else {} self._AttrName = op_cls.AttrName if hasattr(op_cls, 'AttrName') else None self._ParamName = op_cls.ParamName if hasattr(op_cls, 'ParamName') else None op_type = self._op.type if op_type not in self._OP_COUNT: self._OP_COUNT[op_type] = 0 self._OP_COUNT[op_type] = self._OP_COUNT[op_type] + 1 for name, value in self._config.items(): self._op.set_config(name, value) for name, value in self._params.items(): self.param(name, value) def config(self, name, value): self._op.set_config(name, value) return self def attr(self, name, value): if not self._AttrName: raise ValueError('Op {} does not has any attributes'.format( type(self._op))) for attr in self._AttrName: if name == attr.value: break self._op.set_attr(attr, value) return self def param(self, name, value): if not self._ParamName: param = name else: param = utils.op_param_by_name(self._op, name) if not param: raise ValueError('{} does not has a param named "{}"'.format( type(self._op), name)) op_type = self._op.type index = self._OP_COUNT[op_type] - 1 # Naming tensor like dense_0:weight name = '{}_{}:{}'.format(op_type, index, name) tensor = tensor_utils.param_from_tf_numpy(name, value) self._op.set_param(param, tensor) return self def build(self): return self._op if tf_utils.is_tf_version_greater_equal('2.6'): normalization_layer = layers.Normalization else: normalization_layer = layers.experimental.preprocessing.Normalization if tf_utils.is_tf_version_greater_equal('2.6'): rescaling_layer = layers.Rescaling else: rescaling_layer = layers.experimental.preprocessing.Rescaling def create_node(name, op, input_names, output_names): node = ops.Node(name, op) node.input_names[:] = input_names node.output_names[:] = output_names return node def convert_op_placeholder(node): config = node.get_config() shape = tf_utils.tf_shape_to_list(config['shape']) dtype = dtypes.from_tf(config['dtype']) op = ( OpBuilder(op_def.TFInput).config('shape', shape).config('dtype', dtype).build()) return create_node(node.name, op, node.input_names, node.output_names)
null
23,419
from __future__ import absolute_import from __future__ import division from __future__ import print_function import copy import numpy as np import tensorflow as tf from tensorflow.keras import activations from tensorflow.keras import layers from tensorflow.python.util import nest from tf_nndct.graph import OpTypes from tf_nndct.graph import dtypes from tf_nndct.graph import op_def from tf_nndct.graph import ops from tf_nndct.graph import utils from tf_nndct.utils import generic_utils from tf_nndct.utils import keras_utils from tf_nndct.utils import registry from tf_nndct.utils import tensor_utils from tf_nndct.utils import tf_utils class OpBuilder(object): def __init__(self, op_cls, config=None, params=None, *args, **kwargs): def config(self, name, value): def attr(self, name, value): def param(self, name, value): def build(self): def create_node(name, op, input_names, output_names): def convert_op_cast(node): config = node.get_config() op = ( OpBuilder(op_def.TFCast, None, None).config('dtype', config['DstT']).build()) return create_node(node.name, op, node.input_names, node.output_names)
null
23,420
from __future__ import absolute_import from __future__ import division from __future__ import print_function import copy import numpy as np import tensorflow as tf from tensorflow.keras import activations from tensorflow.keras import layers from tensorflow.python.util import nest from tf_nndct.graph import OpTypes from tf_nndct.graph import dtypes from tf_nndct.graph import op_def from tf_nndct.graph import ops from tf_nndct.graph import utils from tf_nndct.utils import generic_utils from tf_nndct.utils import keras_utils from tf_nndct.utils import registry from tf_nndct.utils import tensor_utils from tf_nndct.utils import tf_utils if tf_utils.is_tf_version_greater_equal('2.9.0'): from keras.layers.rnn import lstm as recurrent_v2 from keras.layers.rnn import lstm_v1 as recurrent elif tf_utils.is_tf_version_greater_equal('2.6'): # Keras is seperate from tensorflow since tf 2.6 from keras.layers import recurrent from keras.layers import recurrent_v2 else: from tensorflow.python.keras.layers import recurrent from tensorflow.python.keras.layers import recurrent_v2 class OpBuilder(object): _OP_COUNT = {} # TODO(yuwang): Maybe use op object instead of op class ? def __init__(self, op_cls, config=None, params=None, *args, **kwargs): self._op = op_cls(*args, **kwargs) self._config = config if config else {} self._params = params if params else {} self._AttrName = op_cls.AttrName if hasattr(op_cls, 'AttrName') else None self._ParamName = op_cls.ParamName if hasattr(op_cls, 'ParamName') else None op_type = self._op.type if op_type not in self._OP_COUNT: self._OP_COUNT[op_type] = 0 self._OP_COUNT[op_type] = self._OP_COUNT[op_type] + 1 for name, value in self._config.items(): self._op.set_config(name, value) for name, value in self._params.items(): self.param(name, value) def config(self, name, value): self._op.set_config(name, value) return self def attr(self, name, value): if not self._AttrName: raise ValueError('Op {} does not has any attributes'.format( type(self._op))) for attr in self._AttrName: if name == attr.value: break self._op.set_attr(attr, value) return self def param(self, name, value): if not self._ParamName: param = name else: param = utils.op_param_by_name(self._op, name) if not param: raise ValueError('{} does not has a param named "{}"'.format( type(self._op), name)) op_type = self._op.type index = self._OP_COUNT[op_type] - 1 # Naming tensor like dense_0:weight name = '{}_{}:{}'.format(op_type, index, name) tensor = tensor_utils.param_from_tf_numpy(name, value) self._op.set_param(param, tensor) return self def build(self): return self._op if tf_utils.is_tf_version_greater_equal('2.6'): normalization_layer = layers.Normalization else: normalization_layer = layers.experimental.preprocessing.Normalization if tf_utils.is_tf_version_greater_equal('2.6'): rescaling_layer = layers.Rescaling else: rescaling_layer = layers.experimental.preprocessing.Rescaling def create_node(name, op, input_names, output_names): node = ops.Node(name, op) node.input_names[:] = input_names node.output_names[:] = output_names return node def convert_op_const(node): config = node.get_config() ndarray = tf_utils.values_from_tf_const(node.op.node_def) # Save ndarray or raw tf.Tensor? config['value'] = ndarray op = ( OpBuilder(op_def.TFConst, config, None).param(node.name, ndarray).build()) return create_node(node.name, op, node.input_names, node.output_names)
null
23,421
from __future__ import absolute_import from __future__ import division from __future__ import print_function import copy import numpy as np import tensorflow as tf from tensorflow.keras import activations from tensorflow.keras import layers from tensorflow.python.util import nest from tf_nndct.graph import OpTypes from tf_nndct.graph import dtypes from tf_nndct.graph import op_def from tf_nndct.graph import ops from tf_nndct.graph import utils from tf_nndct.utils import generic_utils from tf_nndct.utils import keras_utils from tf_nndct.utils import registry from tf_nndct.utils import tensor_utils from tf_nndct.utils import tf_utils class OpBuilder(object): _OP_COUNT = {} # TODO(yuwang): Maybe use op object instead of op class ? def __init__(self, op_cls, config=None, params=None, *args, **kwargs): self._op = op_cls(*args, **kwargs) self._config = config if config else {} self._params = params if params else {} self._AttrName = op_cls.AttrName if hasattr(op_cls, 'AttrName') else None self._ParamName = op_cls.ParamName if hasattr(op_cls, 'ParamName') else None op_type = self._op.type if op_type not in self._OP_COUNT: self._OP_COUNT[op_type] = 0 self._OP_COUNT[op_type] = self._OP_COUNT[op_type] + 1 for name, value in self._config.items(): self._op.set_config(name, value) for name, value in self._params.items(): self.param(name, value) def config(self, name, value): self._op.set_config(name, value) return self def attr(self, name, value): if not self._AttrName: raise ValueError('Op {} does not has any attributes'.format( type(self._op))) for attr in self._AttrName: if name == attr.value: break self._op.set_attr(attr, value) return self def param(self, name, value): if not self._ParamName: param = name else: param = utils.op_param_by_name(self._op, name) if not param: raise ValueError('{} does not has a param named "{}"'.format( type(self._op), name)) op_type = self._op.type index = self._OP_COUNT[op_type] - 1 # Naming tensor like dense_0:weight name = '{}_{}:{}'.format(op_type, index, name) tensor = tensor_utils.param_from_tf_numpy(name, value) self._op.set_param(param, tensor) return self def build(self): return self._op def create_node(name, op, input_names, output_names): node = ops.Node(name, op) node.input_names[:] = input_names node.output_names[:] = output_names return node def convert_op_mul(node): op = OpBuilder(op_def.TFMultiply).build() return create_node(node.name, op, node.input_names, node.output_names)
null
23,422
from __future__ import absolute_import from __future__ import division from __future__ import print_function import copy import numpy as np import tensorflow as tf from tensorflow.keras import activations from tensorflow.keras import layers from tensorflow.python.util import nest from tf_nndct.graph import OpTypes from tf_nndct.graph import dtypes from tf_nndct.graph import op_def from tf_nndct.graph import ops from tf_nndct.graph import utils from tf_nndct.utils import generic_utils from tf_nndct.utils import keras_utils from tf_nndct.utils import registry from tf_nndct.utils import tensor_utils from tf_nndct.utils import tf_utils class OpBuilder(object): def __init__(self, op_cls, config=None, params=None, *args, **kwargs): def config(self, name, value): def attr(self, name, value): def param(self, name, value): def build(self): def create_node(name, op, input_names, output_names): def convert_op_strided_slice(node): op = (OpBuilder(op_def.TFStridedSlice, node.get_config()).build()) return create_node(node.name, op, node.input_names, node.output_names)
null
23,423
from __future__ import absolute_import from __future__ import division from __future__ import print_function import copy import numpy as np import tensorflow as tf from tensorflow.keras import activations from tensorflow.keras import layers from tensorflow.python.util import nest from tf_nndct.graph import OpTypes from tf_nndct.graph import dtypes from tf_nndct.graph import op_def from tf_nndct.graph import ops from tf_nndct.graph import utils from tf_nndct.utils import generic_utils from tf_nndct.utils import keras_utils from tf_nndct.utils import registry from tf_nndct.utils import tensor_utils from tf_nndct.utils import tf_utils class OpBuilder(object): _OP_COUNT = {} # TODO(yuwang): Maybe use op object instead of op class ? def __init__(self, op_cls, config=None, params=None, *args, **kwargs): self._op = op_cls(*args, **kwargs) self._config = config if config else {} self._params = params if params else {} self._AttrName = op_cls.AttrName if hasattr(op_cls, 'AttrName') else None self._ParamName = op_cls.ParamName if hasattr(op_cls, 'ParamName') else None op_type = self._op.type if op_type not in self._OP_COUNT: self._OP_COUNT[op_type] = 0 self._OP_COUNT[op_type] = self._OP_COUNT[op_type] + 1 for name, value in self._config.items(): self._op.set_config(name, value) for name, value in self._params.items(): self.param(name, value) def config(self, name, value): self._op.set_config(name, value) return self def attr(self, name, value): if not self._AttrName: raise ValueError('Op {} does not has any attributes'.format( type(self._op))) for attr in self._AttrName: if name == attr.value: break self._op.set_attr(attr, value) return self def param(self, name, value): if not self._ParamName: param = name else: param = utils.op_param_by_name(self._op, name) if not param: raise ValueError('{} does not has a param named "{}"'.format( type(self._op), name)) op_type = self._op.type index = self._OP_COUNT[op_type] - 1 # Naming tensor like dense_0:weight name = '{}_{}:{}'.format(op_type, index, name) tensor = tensor_utils.param_from_tf_numpy(name, value) self._op.set_param(param, tensor) return self def build(self): return self._op def create_node(name, op, input_names, output_names): node = ops.Node(name, op) node.input_names[:] = input_names node.output_names[:] = output_names return node def convert_op_matmul(node): # Parse MatMul to Dense without bias. op = (OpBuilder(op_def.TFDense).config('use_bias', False).build()) return create_node(node.name, op, node.input_names, node.output_names)
null
23,424
from __future__ import absolute_import from __future__ import division from __future__ import print_function import copy import numpy as np import tensorflow as tf from tensorflow.keras import activations from tensorflow.keras import layers from tensorflow.python.util import nest from tf_nndct.graph import OpTypes from tf_nndct.graph import dtypes from tf_nndct.graph import op_def from tf_nndct.graph import ops from tf_nndct.graph import utils from tf_nndct.utils import generic_utils from tf_nndct.utils import keras_utils from tf_nndct.utils import registry from tf_nndct.utils import tensor_utils from tf_nndct.utils import tf_utils class OpBuilder(object): _OP_COUNT = {} # TODO(yuwang): Maybe use op object instead of op class ? def __init__(self, op_cls, config=None, params=None, *args, **kwargs): self._op = op_cls(*args, **kwargs) self._config = config if config else {} self._params = params if params else {} self._AttrName = op_cls.AttrName if hasattr(op_cls, 'AttrName') else None self._ParamName = op_cls.ParamName if hasattr(op_cls, 'ParamName') else None op_type = self._op.type if op_type not in self._OP_COUNT: self._OP_COUNT[op_type] = 0 self._OP_COUNT[op_type] = self._OP_COUNT[op_type] + 1 for name, value in self._config.items(): self._op.set_config(name, value) for name, value in self._params.items(): self.param(name, value) def config(self, name, value): self._op.set_config(name, value) return self def attr(self, name, value): if not self._AttrName: raise ValueError('Op {} does not has any attributes'.format( type(self._op))) for attr in self._AttrName: if name == attr.value: break self._op.set_attr(attr, value) return self def param(self, name, value): if not self._ParamName: param = name else: param = utils.op_param_by_name(self._op, name) if not param: raise ValueError('{} does not has a param named "{}"'.format( type(self._op), name)) op_type = self._op.type index = self._OP_COUNT[op_type] - 1 # Naming tensor like dense_0:weight name = '{}_{}:{}'.format(op_type, index, name) tensor = tensor_utils.param_from_tf_numpy(name, value) self._op.set_param(param, tensor) return self def build(self): return self._op def create_node(name, op, input_names, output_names): node = ops.Node(name, op) node.input_names[:] = input_names node.output_names[:] = output_names return node The provided code snippet includes necessary dependencies for implementing the `convert_layer_normalization` function. Write a Python function `def convert_layer_normalization(node)` to solve the following problem: Convert layers.Normalization to GeTFNormalizationneric Here is the function: def convert_layer_normalization(node): """ Convert layers.Normalization to GeTFNormalizationneric """ params = node.get_params() if 'count' in params: params['count'] = np.array(params['count']) # del params['count'] # 0 op = OpBuilder(op_def.TFNormalization, node.get_config(), params).build() return create_node(node.name, op, node.input_names, node.output_names)
Convert layers.Normalization to GeTFNormalizationneric
23,425
from __future__ import absolute_import from __future__ import division from __future__ import print_function import copy import numpy as np import tensorflow as tf from tensorflow.keras import activations from tensorflow.keras import layers from tensorflow.python.util import nest from tf_nndct.graph import OpTypes from tf_nndct.graph import dtypes from tf_nndct.graph import op_def from tf_nndct.graph import ops from tf_nndct.graph import utils from tf_nndct.utils import generic_utils from tf_nndct.utils import keras_utils from tf_nndct.utils import registry from tf_nndct.utils import tensor_utils from tf_nndct.utils import tf_utils class OpBuilder(object): _OP_COUNT = {} # TODO(yuwang): Maybe use op object instead of op class ? def __init__(self, op_cls, config=None, params=None, *args, **kwargs): self._op = op_cls(*args, **kwargs) self._config = config if config else {} self._params = params if params else {} self._AttrName = op_cls.AttrName if hasattr(op_cls, 'AttrName') else None self._ParamName = op_cls.ParamName if hasattr(op_cls, 'ParamName') else None op_type = self._op.type if op_type not in self._OP_COUNT: self._OP_COUNT[op_type] = 0 self._OP_COUNT[op_type] = self._OP_COUNT[op_type] + 1 for name, value in self._config.items(): self._op.set_config(name, value) for name, value in self._params.items(): self.param(name, value) def config(self, name, value): self._op.set_config(name, value) return self def attr(self, name, value): if not self._AttrName: raise ValueError('Op {} does not has any attributes'.format( type(self._op))) for attr in self._AttrName: if name == attr.value: break self._op.set_attr(attr, value) return self def param(self, name, value): if not self._ParamName: param = name else: param = utils.op_param_by_name(self._op, name) if not param: raise ValueError('{} does not has a param named "{}"'.format( type(self._op), name)) op_type = self._op.type index = self._OP_COUNT[op_type] - 1 # Naming tensor like dense_0:weight name = '{}_{}:{}'.format(op_type, index, name) tensor = tensor_utils.param_from_tf_numpy(name, value) self._op.set_param(param, tensor) return self def build(self): return self._op def create_node(name, op, input_names, output_names): node = ops.Node(name, op) node.input_names[:] = input_names node.output_names[:] = output_names return node The provided code snippet includes necessary dependencies for implementing the `convert_layer_rescaling` function. Write a Python function `def convert_layer_rescaling(node)` to solve the following problem: Convert layers.Rescaling to TFRescaling Here is the function: def convert_layer_rescaling(node): """ Convert layers.Rescaling to TFRescaling """ params = node.get_params() op = OpBuilder(op_def.TFRescaling, node.get_config(), params).build() return create_node(node.name, op, node.input_names, node.output_names)
Convert layers.Rescaling to TFRescaling
23,426
from __future__ import absolute_import from __future__ import division from __future__ import print_function import copy import numpy as np import tensorflow as tf from tensorflow.keras import activations from tensorflow.keras import layers from tensorflow.python.util import nest from tf_nndct.graph import OpTypes from tf_nndct.graph import dtypes from tf_nndct.graph import op_def from tf_nndct.graph import ops from tf_nndct.graph import utils from tf_nndct.utils import generic_utils from tf_nndct.utils import keras_utils from tf_nndct.utils import registry from tf_nndct.utils import tensor_utils from tf_nndct.utils import tf_utils class OpBuilder(object): def __init__(self, op_cls, config=None, params=None, *args, **kwargs): def config(self, name, value): def attr(self, name, value): def param(self, name, value): def build(self): def create_node(name, op, input_names, output_names): def convert_layer_batchnorm(node): config = node.get_config() params = node.get_params() # See https://github.com/tensorflow/tensorflow/blob/r2.3/tensorflow/python/keras/layers/normalization.py#L358 if config['scale']: param_shape = params['gamma'].shape elif config['center']: param_shape = params['beta'].shape else: param_shape = params['moving_mean'].shape for dim in param_shape: if dim != 1: out_dim = dim break config_axis = config['axis'] axis = config_axis[0] if isinstance(config_axis, list) else config_axis op = ( OpBuilder(op_def.TFBatchNorm, config, params).attr('out_dim', out_dim).attr('axis', axis).build()) return create_node(node.name, op, node.input_names, node.output_names)
null
23,427
from __future__ import absolute_import from __future__ import division from __future__ import print_function import copy import numpy as np import tensorflow as tf from tensorflow.keras import activations from tensorflow.keras import layers from tensorflow.python.util import nest from tf_nndct.graph import OpTypes from tf_nndct.graph import dtypes from tf_nndct.graph import op_def from tf_nndct.graph import ops from tf_nndct.graph import utils from tf_nndct.utils import generic_utils from tf_nndct.utils import keras_utils from tf_nndct.utils import registry from tf_nndct.utils import tensor_utils from tf_nndct.utils import tf_utils class OpBuilder(object): def __init__(self, op_cls, config=None, params=None, *args, **kwargs): def config(self, name, value): def attr(self, name, value): def param(self, name, value): def build(self): def create_activation_node(kernel_node, activation, output_names): def convert_layer_dense(node): config = node.get_config() params = node.get_params() op = ( OpBuilder(op_def.TFDense, config, params).config('activation', None).attr( 'activation', config['activation']).attr('in_dim', params['kernel'].shape[0]).build()) dense_node = ops.Node(node.name, op) dense_node.input_names = copy.deepcopy(node.input_names) if config['activation']: actv_node = create_activation_node(dense_node, config['activation'], node.output_names) return [dense_node, actv_node] else: dense_node.output_names = copy.deepcopy(node.output_names) return dense_node
null
23,428
from __future__ import absolute_import from __future__ import division from __future__ import print_function import copy import numpy as np import tensorflow as tf from tensorflow.keras import activations from tensorflow.keras import layers from tensorflow.python.util import nest from tf_nndct.graph import OpTypes from tf_nndct.graph import dtypes from tf_nndct.graph import op_def from tf_nndct.graph import ops from tf_nndct.graph import utils from tf_nndct.utils import generic_utils from tf_nndct.utils import keras_utils from tf_nndct.utils import registry from tf_nndct.utils import tensor_utils from tf_nndct.utils import tf_utils class OpBuilder(object): def __init__(self, op_cls, config=None, params=None, *args, **kwargs): def config(self, name, value): def attr(self, name, value): def param(self, name, value): def build(self): def _convert_node_to_generic(node): _tf_type_to_nndct = { 'Add': op_def.TFAdd, 'AddV2': op_def.TFAdd, 'BiasAdd': op_def.TFBiasAdd, 'Identity': op_def.TFIdentity, 'NoOp': op_def.TFNoOp, 'Reshape': op_def.TFReshape, 'Sigmoid': op_def.TFSigmoid, 'Tanh': op_def.TFTanh, 'GatherV2': op_def.TFGather, 'RFFT': op_def.TFRFFT, 'ComplexAbs': op_def.TFComplexAbs, 'Angle': op_def.TFAngle, 'Exp': op_def.TFExp, 'IRFFT': op_def.TFIRFFT, 'Pad': op_def.TFPad, 'Transpose': op_def.TFTranspose, 'Sum': op_def.TFSum, 'reshape': op_def.TFReshape, 'concat': op_def.TFConcat, 'ConcatV2': op_def.TFConcat, '__operators__.add': op_def.TFAdd, } def create_node(name, op, input_names, output_names): def convert_tf_op_lambda(node): op_name = node.get_config()["function"] if op_name in _tf_type_to_nndct: op = OpBuilder(_tf_type_to_nndct[op_name], node.get_config(), None).build() return create_node(node.name, op, node.input_names, node.output_names) else: return _convert_node_to_generic(node)
null
23,429
from __future__ import absolute_import from __future__ import division from __future__ import print_function import copy import numpy as np import tensorflow as tf from tensorflow.keras import activations from tensorflow.keras import layers from tensorflow.python.util import nest from tf_nndct.graph import OpTypes from tf_nndct.graph import dtypes from tf_nndct.graph import op_def from tf_nndct.graph import ops from tf_nndct.graph import utils from tf_nndct.utils import generic_utils from tf_nndct.utils import keras_utils from tf_nndct.utils import registry from tf_nndct.utils import tensor_utils from tf_nndct.utils import tf_utils class OpBuilder(object): _OP_COUNT = {} # TODO(yuwang): Maybe use op object instead of op class ? def __init__(self, op_cls, config=None, params=None, *args, **kwargs): self._op = op_cls(*args, **kwargs) self._config = config if config else {} self._params = params if params else {} self._AttrName = op_cls.AttrName if hasattr(op_cls, 'AttrName') else None self._ParamName = op_cls.ParamName if hasattr(op_cls, 'ParamName') else None op_type = self._op.type if op_type not in self._OP_COUNT: self._OP_COUNT[op_type] = 0 self._OP_COUNT[op_type] = self._OP_COUNT[op_type] + 1 for name, value in self._config.items(): self._op.set_config(name, value) for name, value in self._params.items(): self.param(name, value) def config(self, name, value): self._op.set_config(name, value) return self def attr(self, name, value): if not self._AttrName: raise ValueError('Op {} does not has any attributes'.format( type(self._op))) for attr in self._AttrName: if name == attr.value: break self._op.set_attr(attr, value) return self def param(self, name, value): if not self._ParamName: param = name else: param = utils.op_param_by_name(self._op, name) if not param: raise ValueError('{} does not has a param named "{}"'.format( type(self._op), name)) op_type = self._op.type index = self._OP_COUNT[op_type] - 1 # Naming tensor like dense_0:weight name = '{}_{}:{}'.format(op_type, index, name) tensor = tensor_utils.param_from_tf_numpy(name, value) self._op.set_param(param, tensor) return self def build(self): return self._op def _convert_node_to_generic(node): """Convert a node with unregistered type to a generic node by saving the node's config as-is. """ op = OpBuilder(op_def.TFGeneric, node.get_config(), node.get_params()).attr('layer_class', type(node.op)).build() return create_node(node.name, op, node.input_names, node.output_names) _tf_type_to_nndct = { 'Add': op_def.TFAdd, 'AddV2': op_def.TFAdd, 'BiasAdd': op_def.TFBiasAdd, 'Identity': op_def.TFIdentity, 'NoOp': op_def.TFNoOp, 'Reshape': op_def.TFReshape, 'Sigmoid': op_def.TFSigmoid, 'Tanh': op_def.TFTanh, 'GatherV2': op_def.TFGather, 'RFFT': op_def.TFRFFT, 'ComplexAbs': op_def.TFComplexAbs, 'Angle': op_def.TFAngle, 'Exp': op_def.TFExp, 'IRFFT': op_def.TFIRFFT, 'Pad': op_def.TFPad, 'Transpose': op_def.TFTranspose, 'Sum': op_def.TFSum, 'reshape': op_def.TFReshape, 'concat': op_def.TFConcat, 'ConcatV2': op_def.TFConcat, '__operators__.add': op_def.TFAdd, } def create_node(name, op, input_names, output_names): node = ops.Node(name, op) node.input_names[:] = input_names node.output_names[:] = output_names return node def convert_tensorflow_op_layer(node): op_name = node.get_config()['node_def']['op'] if op_name in _tf_type_to_nndct: op = OpBuilder(_tf_type_to_nndct[op_name], node.get_config()).build() return create_node(node.name, op, node.input_names, node.output_names) else: return _convert_node_to_generic(node)
null
23,430
from __future__ import absolute_import from __future__ import division from __future__ import print_function import copy import numpy as np import tensorflow as tf from tensorflow.keras import activations from tensorflow.keras import layers from tensorflow.python.util import nest from tf_nndct.graph import OpTypes from tf_nndct.graph import dtypes from tf_nndct.graph import op_def from tf_nndct.graph import ops from tf_nndct.graph import utils from tf_nndct.utils import generic_utils from tf_nndct.utils import keras_utils from tf_nndct.utils import registry from tf_nndct.utils import tensor_utils from tf_nndct.utils import tf_utils class OpBuilder(object): _OP_COUNT = {} # TODO(yuwang): Maybe use op object instead of op class ? def __init__(self, op_cls, config=None, params=None, *args, **kwargs): self._op = op_cls(*args, **kwargs) self._config = config if config else {} self._params = params if params else {} self._AttrName = op_cls.AttrName if hasattr(op_cls, 'AttrName') else None self._ParamName = op_cls.ParamName if hasattr(op_cls, 'ParamName') else None op_type = self._op.type if op_type not in self._OP_COUNT: self._OP_COUNT[op_type] = 0 self._OP_COUNT[op_type] = self._OP_COUNT[op_type] + 1 for name, value in self._config.items(): self._op.set_config(name, value) for name, value in self._params.items(): self.param(name, value) def config(self, name, value): self._op.set_config(name, value) return self def attr(self, name, value): if not self._AttrName: raise ValueError('Op {} does not has any attributes'.format( type(self._op))) for attr in self._AttrName: if name == attr.value: break self._op.set_attr(attr, value) return self def param(self, name, value): if not self._ParamName: param = name else: param = utils.op_param_by_name(self._op, name) if not param: raise ValueError('{} does not has a param named "{}"'.format( type(self._op), name)) op_type = self._op.type index = self._OP_COUNT[op_type] - 1 # Naming tensor like dense_0:weight name = '{}_{}:{}'.format(op_type, index, name) tensor = tensor_utils.param_from_tf_numpy(name, value) self._op.set_param(param, tensor) return self def build(self): return self._op def create_node(name, op, input_names, output_names): node = ops.Node(name, op) node.input_names[:] = input_names node.output_names[:] = output_names return node def convert_layer_add(node): op = OpBuilder(op_def.TFAdd, node.get_config(), None).build() return create_node(node.name, op, node.input_names, node.output_names)
null
23,431
from __future__ import absolute_import from __future__ import division from __future__ import print_function import copy import numpy as np import tensorflow as tf from tensorflow.keras import activations from tensorflow.keras import layers from tensorflow.python.util import nest from tf_nndct.graph import OpTypes from tf_nndct.graph import dtypes from tf_nndct.graph import op_def from tf_nndct.graph import ops from tf_nndct.graph import utils from tf_nndct.utils import generic_utils from tf_nndct.utils import keras_utils from tf_nndct.utils import registry from tf_nndct.utils import tensor_utils from tf_nndct.utils import tf_utils class OpBuilder(object): def __init__(self, op_cls, config=None, params=None, *args, **kwargs): def config(self, name, value): def attr(self, name, value): def param(self, name, value): def build(self): def create_node(name, op, input_names, output_names): def convert_layer_sub(node): op = OpBuilder(op_def.TFSubtract, node.get_config(), None).build() return create_node(node.name, op, node.input_names, node.output_names)
null
23,432
from __future__ import absolute_import from __future__ import division from __future__ import print_function import copy import numpy as np import tensorflow as tf from tensorflow.keras import activations from tensorflow.keras import layers from tensorflow.python.util import nest from tf_nndct.graph import OpTypes from tf_nndct.graph import dtypes from tf_nndct.graph import op_def from tf_nndct.graph import ops from tf_nndct.graph import utils from tf_nndct.utils import generic_utils from tf_nndct.utils import keras_utils from tf_nndct.utils import registry from tf_nndct.utils import tensor_utils from tf_nndct.utils import tf_utils class OpBuilder(object): _OP_COUNT = {} # TODO(yuwang): Maybe use op object instead of op class ? def __init__(self, op_cls, config=None, params=None, *args, **kwargs): self._op = op_cls(*args, **kwargs) self._config = config if config else {} self._params = params if params else {} self._AttrName = op_cls.AttrName if hasattr(op_cls, 'AttrName') else None self._ParamName = op_cls.ParamName if hasattr(op_cls, 'ParamName') else None op_type = self._op.type if op_type not in self._OP_COUNT: self._OP_COUNT[op_type] = 0 self._OP_COUNT[op_type] = self._OP_COUNT[op_type] + 1 for name, value in self._config.items(): self._op.set_config(name, value) for name, value in self._params.items(): self.param(name, value) def config(self, name, value): self._op.set_config(name, value) return self def attr(self, name, value): if not self._AttrName: raise ValueError('Op {} does not has any attributes'.format( type(self._op))) for attr in self._AttrName: if name == attr.value: break self._op.set_attr(attr, value) return self def param(self, name, value): if not self._ParamName: param = name else: param = utils.op_param_by_name(self._op, name) if not param: raise ValueError('{} does not has a param named "{}"'.format( type(self._op), name)) op_type = self._op.type index = self._OP_COUNT[op_type] - 1 # Naming tensor like dense_0:weight name = '{}_{}:{}'.format(op_type, index, name) tensor = tensor_utils.param_from_tf_numpy(name, value) self._op.set_param(param, tensor) return self def build(self): return self._op def create_node(name, op, input_names, output_names): node = ops.Node(name, op) node.input_names[:] = input_names node.output_names[:] = output_names return node def convert_layer_multiply(node): op = OpBuilder(op_def.TFMultiplyLayer, node.get_config(), None).build() return create_node(node.name, op, node.input_names, node.output_names)
null
23,433
from __future__ import absolute_import from __future__ import division from __future__ import print_function import copy import numpy as np import tensorflow as tf from tensorflow.keras import activations from tensorflow.keras import layers from tensorflow.python.util import nest from tf_nndct.graph import OpTypes from tf_nndct.graph import dtypes from tf_nndct.graph import op_def from tf_nndct.graph import ops from tf_nndct.graph import utils from tf_nndct.utils import generic_utils from tf_nndct.utils import keras_utils from tf_nndct.utils import registry from tf_nndct.utils import tensor_utils from tf_nndct.utils import tf_utils class OpBuilder(object): _OP_COUNT = {} # TODO(yuwang): Maybe use op object instead of op class ? def __init__(self, op_cls, config=None, params=None, *args, **kwargs): self._op = op_cls(*args, **kwargs) self._config = config if config else {} self._params = params if params else {} self._AttrName = op_cls.AttrName if hasattr(op_cls, 'AttrName') else None self._ParamName = op_cls.ParamName if hasattr(op_cls, 'ParamName') else None op_type = self._op.type if op_type not in self._OP_COUNT: self._OP_COUNT[op_type] = 0 self._OP_COUNT[op_type] = self._OP_COUNT[op_type] + 1 for name, value in self._config.items(): self._op.set_config(name, value) for name, value in self._params.items(): self.param(name, value) def config(self, name, value): self._op.set_config(name, value) return self def attr(self, name, value): if not self._AttrName: raise ValueError('Op {} does not has any attributes'.format( type(self._op))) for attr in self._AttrName: if name == attr.value: break self._op.set_attr(attr, value) return self def param(self, name, value): if not self._ParamName: param = name else: param = utils.op_param_by_name(self._op, name) if not param: raise ValueError('{} does not has a param named "{}"'.format( type(self._op), name)) op_type = self._op.type index = self._OP_COUNT[op_type] - 1 # Naming tensor like dense_0:weight name = '{}_{}:{}'.format(op_type, index, name) tensor = tensor_utils.param_from_tf_numpy(name, value) self._op.set_param(param, tensor) return self def build(self): return self._op def create_node(name, op, input_names, output_names): node = ops.Node(name, op) node.input_names[:] = input_names node.output_names[:] = output_names return node def convert_layer_embedding(node): op = OpBuilder(op_def.TFEmbedding, node.get_config(), node.get_params()).build() return create_node(node.name, op, node.input_names, node.output_names)
null
23,434
from __future__ import absolute_import from __future__ import division from __future__ import print_function import copy import numpy as np import tensorflow as tf from tensorflow.keras import activations from tensorflow.keras import layers from tensorflow.python.util import nest from tf_nndct.graph import OpTypes from tf_nndct.graph import dtypes from tf_nndct.graph import op_def from tf_nndct.graph import ops from tf_nndct.graph import utils from tf_nndct.utils import generic_utils from tf_nndct.utils import keras_utils from tf_nndct.utils import registry from tf_nndct.utils import tensor_utils from tf_nndct.utils import tf_utils class OpBuilder(object): _OP_COUNT = {} # TODO(yuwang): Maybe use op object instead of op class ? def __init__(self, op_cls, config=None, params=None, *args, **kwargs): self._op = op_cls(*args, **kwargs) self._config = config if config else {} self._params = params if params else {} self._AttrName = op_cls.AttrName if hasattr(op_cls, 'AttrName') else None self._ParamName = op_cls.ParamName if hasattr(op_cls, 'ParamName') else None op_type = self._op.type if op_type not in self._OP_COUNT: self._OP_COUNT[op_type] = 0 self._OP_COUNT[op_type] = self._OP_COUNT[op_type] + 1 for name, value in self._config.items(): self._op.set_config(name, value) for name, value in self._params.items(): self.param(name, value) def config(self, name, value): self._op.set_config(name, value) return self def attr(self, name, value): if not self._AttrName: raise ValueError('Op {} does not has any attributes'.format( type(self._op))) for attr in self._AttrName: if name == attr.value: break self._op.set_attr(attr, value) return self def param(self, name, value): if not self._ParamName: param = name else: param = utils.op_param_by_name(self._op, name) if not param: raise ValueError('{} does not has a param named "{}"'.format( type(self._op), name)) op_type = self._op.type index = self._OP_COUNT[op_type] - 1 # Naming tensor like dense_0:weight name = '{}_{}:{}'.format(op_type, index, name) tensor = tensor_utils.param_from_tf_numpy(name, value) self._op.set_param(param, tensor) return self def build(self): return self._op def _parse_base_rnn(layer): if keras_utils.is_stacked_rnn_cells(layer.cell): cell_op = _parse_stacked_rnn_cells(layer.cell) else: cell_op = _parse_rnn_layer(layer.cell) op = ( OpBuilder(op_def.TFRNN, layer.get_config(), None).config('cell', cell_op).build()) # Inherit params from cell op. for name, tensor in cell_op.params.items(): op.set_param(name, tensor) return op def create_node(name, op, input_names, output_names): node = ops.Node(name, op) node.input_names[:] = input_names node.output_names[:] = output_names return node def convert_wrapper_bidirectional(node): layer = node.op forward_op = _parse_base_rnn(layer.forward_layer) backward_op = _parse_base_rnn(layer.backward_layer) backward_op.set_config('go_backwards', True) op = ( OpBuilder(op_def.TFBidirectional, node.get_config(), None).config('layer', forward_op).config('backward_layer', backward_op).build()) combined_params = {} for name, tensor in forward_op.params.items(): combined_params['forward_rnn/' + name] = tensor for name, tensor in backward_op.params.items(): combined_params['backward_rnn/' + name] = tensor for name, tensor in combined_params.items(): op.set_param(name, tensor) return create_node(node.name, op, node.input_names, node.output_names)
null
23,435
from __future__ import absolute_import from __future__ import division from __future__ import print_function import copy import numpy as np import tensorflow as tf from tensorflow.keras import activations from tensorflow.keras import layers from tensorflow.python.util import nest from tf_nndct.graph import OpTypes from tf_nndct.graph import dtypes from tf_nndct.graph import op_def from tf_nndct.graph import ops from tf_nndct.graph import utils from tf_nndct.utils import generic_utils from tf_nndct.utils import keras_utils from tf_nndct.utils import registry from tf_nndct.utils import tensor_utils from tf_nndct.utils import tf_utils def _parse_base_rnn(layer): def create_node(name, op, input_names, output_names): def convert_layer_rnn(node): layer = node.op op = _parse_base_rnn(layer) return create_node(node.name, op, node.input_names, node.output_names)
null
23,436
from __future__ import absolute_import from __future__ import division from __future__ import print_function import copy import numpy as np import tensorflow as tf from tensorflow.keras import activations from tensorflow.keras import layers from tensorflow.python.util import nest from tf_nndct.graph import OpTypes from tf_nndct.graph import dtypes from tf_nndct.graph import op_def from tf_nndct.graph import ops from tf_nndct.graph import utils from tf_nndct.utils import generic_utils from tf_nndct.utils import keras_utils from tf_nndct.utils import registry from tf_nndct.utils import tensor_utils from tf_nndct.utils import tf_utils def _parse_rnn_layer(layer): config = layer.get_config() params = keras_utils.get_named_weights(layer) # Naive method to determine if kernel weights are concated together. splited_params = {} units = config['units'] if list(params.values())[0].shape[-1] == units * 4: suffix = ['_i', '_f', '_c', '_o'] for name, value in params.items(): for i in range(len(suffix)): splited_params[name + suffix[i]] = np.copy(value[..., i * units:(i + 1) * units]) else: splited_params = params op = ( OpBuilder(ops.Operation, config, splited_params, _rnn_layer_to_op[type(layer)]).build()) return op recurrent.LSTM, recurrent_v2.LSTM, recurrent.LSTMCell, recurrent_v2.LSTMCell def create_node(name, op, input_names, output_names): node = ops.Node(name, op) node.input_names[:] = input_names node.output_names[:] = output_names return node def convert_layer_lstm(node): layer = node.op if type(layer) in [recurrent.LSTM, recurrent_v2.LSTM]: cell = layer.cell cell_cls = [recurrent.LSTMCell, recurrent_v2.LSTMCell] if type(cell) not in cell_cls: raise NotImplementedError( 'Custom LSTM cell is not supported. Expected {}, but got {}'.format( cell_cls, type(cell))) if cell.recurrent_activation == activations.hard_sigmoid: raise ValueError( 'recurrent_activation="hard_sigmoid" is not allowd, use "sigmoid" instead.' ) op = _parse_rnn_layer(node.op) return create_node(node.name, op, node.input_names, node.output_names)
null
23,437
from __future__ import absolute_import from __future__ import division from __future__ import print_function import copy import numpy as np import tensorflow as tf from tensorflow.keras import activations from tensorflow.keras import layers from tensorflow.python.util import nest from tf_nndct.graph import OpTypes from tf_nndct.graph import dtypes from tf_nndct.graph import op_def from tf_nndct.graph import ops from tf_nndct.graph import utils from tf_nndct.utils import generic_utils from tf_nndct.utils import keras_utils from tf_nndct.utils import registry from tf_nndct.utils import tensor_utils from tf_nndct.utils import tf_utils class OpBuilder(object): _OP_COUNT = {} # TODO(yuwang): Maybe use op object instead of op class ? def __init__(self, op_cls, config=None, params=None, *args, **kwargs): self._op = op_cls(*args, **kwargs) self._config = config if config else {} self._params = params if params else {} self._AttrName = op_cls.AttrName if hasattr(op_cls, 'AttrName') else None self._ParamName = op_cls.ParamName if hasattr(op_cls, 'ParamName') else None op_type = self._op.type if op_type not in self._OP_COUNT: self._OP_COUNT[op_type] = 0 self._OP_COUNT[op_type] = self._OP_COUNT[op_type] + 1 for name, value in self._config.items(): self._op.set_config(name, value) for name, value in self._params.items(): self.param(name, value) def config(self, name, value): self._op.set_config(name, value) return self def attr(self, name, value): if not self._AttrName: raise ValueError('Op {} does not has any attributes'.format( type(self._op))) for attr in self._AttrName: if name == attr.value: break self._op.set_attr(attr, value) return self def param(self, name, value): if not self._ParamName: param = name else: param = utils.op_param_by_name(self._op, name) if not param: raise ValueError('{} does not has a param named "{}"'.format( type(self._op), name)) op_type = self._op.type index = self._OP_COUNT[op_type] - 1 # Naming tensor like dense_0:weight name = '{}_{}:{}'.format(op_type, index, name) tensor = tensor_utils.param_from_tf_numpy(name, value) self._op.set_param(param, tensor) return self def build(self): return self._op def create_node(name, op, input_names, output_names): node = ops.Node(name, op) node.input_names[:] = input_names node.output_names[:] = output_names return node def convert_layer_simplernn(node): op = OpBuilder(op_def.TFSimpleRNN, node.get_config(), node.get_params()).build() return create_node(node.name, op, node.input_names, node.output_names)
null
23,438
from __future__ import absolute_import from __future__ import division from __future__ import print_function import copy import numpy as np import tensorflow as tf from tensorflow.keras import activations from tensorflow.keras import layers from tensorflow.python.util import nest from tf_nndct.graph import OpTypes from tf_nndct.graph import dtypes from tf_nndct.graph import op_def from tf_nndct.graph import ops from tf_nndct.graph import utils from tf_nndct.utils import generic_utils from tf_nndct.utils import keras_utils from tf_nndct.utils import registry from tf_nndct.utils import tensor_utils from tf_nndct.utils import tf_utils class OpBuilder(object): _OP_COUNT = {} # TODO(yuwang): Maybe use op object instead of op class ? def __init__(self, op_cls, config=None, params=None, *args, **kwargs): self._op = op_cls(*args, **kwargs) self._config = config if config else {} self._params = params if params else {} self._AttrName = op_cls.AttrName if hasattr(op_cls, 'AttrName') else None self._ParamName = op_cls.ParamName if hasattr(op_cls, 'ParamName') else None op_type = self._op.type if op_type not in self._OP_COUNT: self._OP_COUNT[op_type] = 0 self._OP_COUNT[op_type] = self._OP_COUNT[op_type] + 1 for name, value in self._config.items(): self._op.set_config(name, value) for name, value in self._params.items(): self.param(name, value) def config(self, name, value): self._op.set_config(name, value) return self def attr(self, name, value): if not self._AttrName: raise ValueError('Op {} does not has any attributes'.format( type(self._op))) for attr in self._AttrName: if name == attr.value: break self._op.set_attr(attr, value) return self def param(self, name, value): if not self._ParamName: param = name else: param = utils.op_param_by_name(self._op, name) if not param: raise ValueError('{} does not has a param named "{}"'.format( type(self._op), name)) op_type = self._op.type index = self._OP_COUNT[op_type] - 1 # Naming tensor like dense_0:weight name = '{}_{}:{}'.format(op_type, index, name) tensor = tensor_utils.param_from_tf_numpy(name, value) self._op.set_param(param, tensor) return self def build(self): return self._op def create_node(name, op, input_names, output_names): node = ops.Node(name, op) node.input_names[:] = input_names node.output_names[:] = output_names return node def convert_layer_gru(node): op = OpBuilder(op_def.TFGRU, node.get_config(), node.get_params()).build() return create_node(node.name, op, node.input_names, node.output_names)
null
23,439
from __future__ import absolute_import from __future__ import division from __future__ import print_function import copy import numpy as np import tensorflow as tf from tensorflow.keras import activations from tensorflow.keras import layers from tensorflow.python.util import nest from tf_nndct.graph import OpTypes from tf_nndct.graph import dtypes from tf_nndct.graph import op_def from tf_nndct.graph import ops from tf_nndct.graph import utils from tf_nndct.utils import generic_utils from tf_nndct.utils import keras_utils from tf_nndct.utils import registry from tf_nndct.utils import tensor_utils from tf_nndct.utils import tf_utils class OpBuilder(object): _OP_COUNT = {} # TODO(yuwang): Maybe use op object instead of op class ? def __init__(self, op_cls, config=None, params=None, *args, **kwargs): self._op = op_cls(*args, **kwargs) self._config = config if config else {} self._params = params if params else {} self._AttrName = op_cls.AttrName if hasattr(op_cls, 'AttrName') else None self._ParamName = op_cls.ParamName if hasattr(op_cls, 'ParamName') else None op_type = self._op.type if op_type not in self._OP_COUNT: self._OP_COUNT[op_type] = 0 self._OP_COUNT[op_type] = self._OP_COUNT[op_type] + 1 for name, value in self._config.items(): self._op.set_config(name, value) for name, value in self._params.items(): self.param(name, value) def config(self, name, value): self._op.set_config(name, value) return self def attr(self, name, value): if not self._AttrName: raise ValueError('Op {} does not has any attributes'.format( type(self._op))) for attr in self._AttrName: if name == attr.value: break self._op.set_attr(attr, value) return self def param(self, name, value): if not self._ParamName: param = name else: param = utils.op_param_by_name(self._op, name) if not param: raise ValueError('{} does not has a param named "{}"'.format( type(self._op), name)) op_type = self._op.type index = self._OP_COUNT[op_type] - 1 # Naming tensor like dense_0:weight name = '{}_{}:{}'.format(op_type, index, name) tensor = tensor_utils.param_from_tf_numpy(name, value) self._op.set_param(param, tensor) return self def build(self): return self._op def create_activation_node(kernel_node, activation, output_names): actv = activations.get(activation) #op = ops.Operation(_activation_cvt_map[actv]) op = _activation_cvt_map[actv]() actv_name = activations.serialize(actv) actv_node = ops.Node('/'.join([kernel_node.name, actv_name]), op) # Connect the activation to the kernel kernel_out_tensor = kernel_node.name + ':0' kernel_node.output_names = [kernel_out_tensor] actv_node.input_names = [kernel_out_tensor] actv_node.output_names = copy.deepcopy(output_names) # Dettach activation node from its kernel node. # The layer name of activation node should not be set to kernel's layer name. actv_node.layer_name = _NO_LAYER_NAME return actv_node def convert_layer_conv2d(node): config = node.get_config() params = node.get_params() # [*kernel_size, input_channels / groups, filters] op = ( OpBuilder(op_def.TFConv2D, config, params).config('activation', None).attr( 'activation', config['activation']).attr( 'in_dim', params['kernel'].shape[-2]).build()) conv_node = ops.Node(node.name, op) conv_node.input_names = copy.deepcopy(node.input_names) if config['activation']: actv_node = create_activation_node(conv_node, config['activation'], node.output_names) return [conv_node, actv_node] else: conv_node.output_names = copy.deepcopy(node.output_names) return conv_node
null