r''' **This file is EXPERIMENTAL and is mostly used for testing purposes! Do not rely on it for anything!** ''' from torch.fx import Graph, GraphModule from torch.fx.graph import map_arg from torch.fx.proxy import Proxy import sys import torch from torch.nn.utils import fuse_conv_bn_weights import operator # can be a # module type, a builtin function, or a string to match target def _minmax_scale_zeropoint(min_val, max_val, qmin=-127, qmax=128, eps=torch.finfo(torch.float32).eps): min_val = min(0.0, min_val) max_val = max(0.0, max_val) if max_val == min_val: return 1.0, 0 else: scale = (max_val - min_val) / float(qmax - qmin) scale = max(scale, eps) zero_point = qmin - round(min_val / scale) zero_point = max(qmin, zero_point) zero_point = min(qmax, zero_point) zero_point = int(zero_point) return scale, zero_point class MinMaxObserver: def __init__(self, quantizer, node): self.min, self.max = float('inf'), float('-inf') self.all_tensors = True def observe(self, node, env): v = env[node.name] if not isinstance(v, torch.Tensor): self.all_tensors = False return self.max = max(self.max, float(v.max())) self.min = min(self.min, float(v.min())) def scale_zeropoint(self): return _minmax_scale_zeropoint(self.min, self.max, qmin=0, qmax=255) class NoObserver: def __init__(self, quantizer, node): pass def observe(self, node, env): pass DEFAULT_QUANTIZATION_PATTERNS = {} def register_pattern(pattern): def insert(fn): DEFAULT_QUANTIZATION_PATTERNS[pattern] = fn return fn return insert @register_pattern(operator.add) class Add(MinMaxObserver): def quantize(self, quantizer, node, load_arg): if not self.all_tensors: return NotImplemented scale, zeropoint = self.scale_zeropoint() return quantizer.quantized_graph.create_node( 'call_function', torch.ops.quantized.add, load_arg(node.args), {'scale': scale, 'zero_point': zeropoint}) class Relu(NoObserver): def quantize(self, quantizer, node, load_arg): return torch.relu(load_arg(node.args[0])) # torch.relu works directly on quantized tensors? # these ops have quantized equivalents that do not need any extra information @register_pattern(torch.nn.ReLU) @register_pattern(torch.nn.AvgPool2d) @register_pattern(torch.nn.MaxPool2d) @register_pattern(torch.nn.AdaptiveAvgPool2d) class CopyNode(NoObserver): def quantize(self, quantizer, node, load_arg): return quantizer.quantized_graph.node_copy(node, load_arg) class IdentityModule(torch.nn.Module): def forward(self, x): return x # handle conv, maybe followed by bn, maybe followed by relu @register_pattern(torch.nn.modules.conv.Conv2d) @register_pattern((torch.nn.ReLU, torch.nn.modules.conv.Conv2d)) @register_pattern((torch.nn.modules.batchnorm.BatchNorm2d, torch.nn.modules.conv.Conv2d)) @register_pattern((torch.nn.ReLU, (torch.nn.modules.batchnorm.BatchNorm2d, torch.nn.modules.conv.Conv2d))) class ConvNormRelu(MinMaxObserver): def __init__(self, quantizer, node): super().__init__(quantizer, node) self.relu_node, self.bn_node = None, None if isinstance(quantizer.modules[node.target], torch.nn.ReLU): self.relu_node = node node = node.args[0] if isinstance(quantizer.modules[node.target], torch.nn.BatchNorm2d): self.bn_node = node self.bn = quantizer.modules[self.bn_node.target] node = node.args[0] assert isinstance(quantizer.modules[node.target], torch.nn.modules.Conv2d) self.conv_node = node self.conv = quantizer.modules[self.conv_node.target] def quantize(self, quantizer, node, load_arg): mod = self.conv weight, bias = mod.weight, mod.bias if self.bn_node is not None: weight, bias = fuse_conv_bn_weights( weight, bias, self.bn.running_mean, self.bn.running_var, self.bn.eps, self.bn.weight, self.bn.bias) min_val, max_val = float(weight.min()), float(weight.max()) act_scale, act_zp = self.scale_zeropoint() weight_scale, weight_zp = _minmax_scale_zeropoint(min_val, max_val) qweight = torch.quantize_per_tensor(weight, weight_scale, weight_zp, torch.qint8) ctor = torch.nn.intrinsic.quantized.ConvReLU2d if self.relu_node is not None else torch.nn.quantized.Conv2d qconv = ctor(mod.in_channels, mod.out_channels, mod.kernel_size, mod.stride, mod.padding, mod.dilation, mod.groups, mod.bias is not None, mod.padding_mode) qconv.set_weight_bias(qweight, bias) qconv.scale = float(act_scale) qconv.zero_point = int(act_zp) parent_name, name = _parent_name(self.conv_node.target) setattr(quantizer.modules[parent_name], name, qconv) if self.bn_node is not None: parent_bn, bn_name = _parent_name(self.bn_node.target) # we can't just delete this because submodules's forwards (which are not longer use) # try to call it, so replace with something that does nothing. setattr(quantizer.modules[parent_name], bn_name, IdentityModule()) return quantizer.quantized_graph.create_node('call_module', self.conv_node.target, (load_arg(self.conv_node.args[0]),), {}) # turn foo.bar -> ['foo', 'bar'] def _parent_name(target): r = target.rsplit('.', 1) if len(r) == 1: return '', r[0] else: return r[0], r[1] class DefaultQuant(MinMaxObserver): def quantize(self, input): assert self.all_tensors scale, zeropoint = self.scale_zeropoint() return torch.quantize_per_tensor(Proxy(input), scale, zeropoint, torch.quint8).node def matches(modules, node, pattern, max_uses=sys.maxsize): if isinstance(pattern, tuple): self_match, *arg_matches = pattern else: self_match = pattern arg_matches = None if len(node.users) > max_uses: return False if isinstance(self_match, type) and issubclass(self_match, torch.nn.Module): if node.op != 'call_module': return False if not isinstance(modules[node.target], self_match): return False elif callable(self_match): if node.op != 'call_function' or node.target is not self_match: return False elif node.target != self_match: return False if not arg_matches: return True if len(arg_matches) != len(node.args): return False return all(matches(modules, node, arg_match, max_uses=1) for node, arg_match in zip(node.args, arg_matches)) class Quantizer: def __init__(self, mod, patterns=DEFAULT_QUANTIZATION_PATTERNS, quant_ctor=DefaultQuant): self.root = mod self.graph = mod.graph self.quant_ctor = quant_ctor # cached information for observe self.state_dict = self.root.state_dict() self.modules = dict(self.root.named_modules()) # match the patterns that will get quantized self.matches = self._find_matches(patterns) # find _inputs_ to matched nodes that are not quantized, these # have to be quantized, which requires measuring stats, # initialize an quant_ctor object for each self.quants = self._find_quants(quant_ctor) def observe(self, args): # most of this function is just an interpreter for the graph # it would be possible to put this in some abstraction, but # it is pretty nice to just be able to see exactly what is happening here # and hack on it. # maybe we should just provide an example interpreter that people copy/paste # then edit. args_iter = iter(args) env = {} def load_arg(a): return map_arg(a, lambda node: env[node.name]) output_node : Optional[Node] = None for node in self.graph.nodes: if node.op == 'placeholder': result = next(args_iter) elif node.op == 'get_attr': result = self.state_dict[node.target] elif node.op == 'call_function': result = node.target(*load_arg(node.args), **load_arg(node.kwargs)) elif node.op == 'call_method': self_obj, *args = load_arg(node.args) kwargs = load_arg(node.kwargs) result = getattr(self_obj, node.target)(*args, **kwargs) elif node.op == 'call_module': result = self.modules[node.target](*load_arg(node.args), **load_arg(node.kwargs)) elif node.op == 'output': return load_arg(node.args[0]) env[node.name] = result root_node, obj = self.matches.get(node.name, (None, None)) if root_node is node: obj.observe(node, env) if node.name in self.quants: self.quants[node.name].observe(node, env) raise RuntimeError('Graph had no output node!') def quantize(self): self.quantized_graph = Graph() env = {} quant_env = {} def load_arg(n, quantized): if not quantized: if n.name not in env and n.name in quant_env: env[n.name] = Proxy(quant_env[n.name]).dequantize().node return env[n.name] else: if n.name not in quant_env and n.name in env: quant_env[n.name] = self.quants[n.name].quantize(env[n.name]) return quant_env[n.name] def copy_recursive(node): def load_or_emit(n): if n.name in env or e.name in quant_env: return load_arg(n, quantized=False) else: return copy_recusive(n) r = env[node.name] = self.quantized_graph.node_copy(node, lambda n: load_arg(n, quantized=False)) return r for node in self.graph.nodes: root_node, obj = self.matches.get(node.name, (None, None)) if root_node is None: # not quantized just copy it env[node.name] = self.quantized_graph.node_copy(node, lambda n: load_arg(n, quantized=False)) elif root_node is node: r = obj.quantize(self, node, lambda a: map_arg(a, lambda n: load_arg(n, quantized=True))) if r is NotImplemented: # quantizer choose to to quantize the node take the entire match, and just copy it over env[node.name] = copy_recursive(node) else: quant_env[node.name] = r return GraphModule(self.root, self.quantized_graph) def _find_matches(self, patterns): modules = dict(self.root.named_modules()) match_map = {} # node name -> (root_node, match_value?) def apply_match(pattern, node, match): if isinstance(pattern, tuple): s, *args = pattern apply_match(s, node, match) for subpattern, arg in zip(args, node.args): apply_match(subpattern, arg, match) else: match_map[node.name] = match for node in reversed(self.graph.nodes): if node.name not in match_map: for pattern, value in patterns.items(): if matches(modules, node, pattern): apply_match(pattern, node, (node, value(self, node))) return match_map def _find_quants(self, quant_ctor): quants = {} def visit_arg(n): # note: we have to measure quantization information # even for nodes where we might not use it because it is already # quantized. This is because each match has the option to # say NotImplemented (if for instance, it is an __add__ and the data type is not appropriate) if n.name not in quants: quants[n.name] = quant_ctor(self, n) for node in self.graph.nodes: if node.name in self.matches: map_arg(node.args, visit_arg) map_arg(node.kwargs, visit_arg) return quants