# Copyright The Lightning AI team. # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. """Utilities related to model weights summary.""" import contextlib import logging import math from collections import OrderedDict from typing import Any, Optional, Union import torch import torch.nn as nn from torch import Tensor from torch.utils.flop_counter import FlopCounterMode from torch.utils.hooks import RemovableHandle import pytorch_lightning as pl from lightning_fabric.utilities import rank_zero_warn from lightning_fabric.utilities.distributed import _is_dtensor from lightning_fabric.utilities.imports import _TORCH_GREATER_EQUAL_2_4 from pytorch_lightning.utilities.model_helpers import _ModuleMode from pytorch_lightning.utilities.rank_zero import WarningCache log = logging.getLogger(__name__) warning_cache = WarningCache() PARAMETER_NUM_UNITS = [" ", "K", "M", "B", "T"] UNKNOWN_SIZE = "?" LEFTOVER_PARAMS_NAME = "other params" NOT_APPLICABLE = "n/a" class LayerSummary: """Summary class for a single layer in a :class:`~pytorch_lightning.core.LightningModule`. It collects the following information: - Type of the layer (e.g. Linear, BatchNorm1d, ...) - Input shape - Output shape - Number of parameters The input and output shapes are only known after the example input array was passed through the model. Example:: >>> model = torch.nn.Conv2d(3, 8, 3) >>> summary = LayerSummary(model) >>> summary.num_parameters 224 >>> summary.layer_type 'Conv2d' >>> output = model(torch.rand(1, 3, 5, 5)) >>> summary.in_size [1, 3, 5, 5] >>> summary.out_size [1, 8, 3, 3] Args: module: A module to summarize """ def __init__(self, module: nn.Module) -> None: super().__init__() self._module = module self._hook_handle = self._register_hook() self._in_size: Optional[Union[str, list]] = None self._out_size: Optional[Union[str, list]] = None def __del__(self) -> None: self.detach_hook() def _register_hook(self) -> Optional[RemovableHandle]: """Registers a hook on the module that computes the input- and output size(s) on the first forward pass. If the hook is called, it will remove itself from the from the module, meaning that recursive models will only record their input- and output shapes once. Registering hooks on :class:`~torch.jit.ScriptModule` is not supported. Return: A handle for the installed hook, or ``None`` if registering the hook is not possible. """ def hook(_: nn.Module, inp: Any, out: Any) -> None: if len(inp) == 1: inp = inp[0] self._in_size = parse_batch_shape(inp) self._out_size = parse_batch_shape(out) assert self._hook_handle is not None self._hook_handle.remove() def hook_with_kwargs(_: nn.Module, args: Any, kwargs: Any, out: Any) -> None: # We can't write them in the same function, since the forward hook # uses positional arguments. inp = (*args, *kwargs.values()) if kwargs is not None else args hook(_, inp, out) handle = None if not isinstance(self._module, torch.jit.ScriptModule): handle = self._module.register_forward_hook(hook_with_kwargs, with_kwargs=True) return handle def detach_hook(self) -> None: """Removes the forward hook if it was not already removed in the forward pass. Will be called after the summary is created. """ if self._hook_handle is not None: self._hook_handle.remove() @property def in_size(self) -> Union[str, list]: return self._in_size or UNKNOWN_SIZE @property def out_size(self) -> Union[str, list]: return self._out_size or UNKNOWN_SIZE @property def layer_type(self) -> str: """Returns the class name of the module.""" return str(self._module.__class__.__name__) @property def num_parameters(self) -> int: """Returns the number of parameters in this module.""" return sum(p.numel() if not _tensor_has_shape(p) else 0 for p in self._module.parameters()) @property def training(self) -> bool: """Returns whether the module is in training mode.""" return self._module.training class ModelSummary: """Generates a summary of all layers in a :class:`~pytorch_lightning.core.LightningModule`. Args: model: The model to summarize (also referred to as the root module). max_depth: Maximum depth of modules to show. Use -1 to show all modules or 0 to show no summary. Defaults to 1. The string representation of this summary prints a table with columns containing the name, type and number of parameters for each layer. The root module may also have an attribute ``example_input_array`` as shown in the example below. If present, the root module will be called with it as input to determine the intermediate input- and output shapes of all layers. Supported are tensors and nested lists and tuples of tensors. All other types of inputs will be skipped and show as `?` in the summary table. The summary will also display `?` for layers not used in the forward pass. If there are parameters not associated with any layers or modules, the count of those parameters will be displayed in the table under `other params`. The summary will display `n/a` for module type, in size, and out size. Example:: >>> import pytorch_lightning as pl >>> class LitModel(pl.LightningModule): ... ... def __init__(self): ... super().__init__() ... self.net = nn.Sequential(nn.Linear(256, 512), nn.BatchNorm1d(512)) ... self.example_input_array = torch.zeros(10, 256) # optional ... ... def forward(self, x): ... return self.net(x) ... >>> model = LitModel() >>> ModelSummary(model, max_depth=1) # doctest: +NORMALIZE_WHITESPACE | Name | Type | Params | Mode | FLOPs | In sizes | Out sizes ---------------------------------------------------------------------------- 0 | net | Sequential | 132 K | train | 2.6 M | [10, 256] | [10, 512] ---------------------------------------------------------------------------- 132 K Trainable params 0 Non-trainable params 132 K Total params 0.530 Total estimated model params size (MB) 3 Modules in train mode 0 Modules in eval mode 2.6 M Total Flops >>> ModelSummary(model, max_depth=-1) # doctest: +NORMALIZE_WHITESPACE | Name | Type | Params | Mode | FLOPs | In sizes | Out sizes ------------------------------------------------------------------------------ 0 | net | Sequential | 132 K | train | 2.6 M | [10, 256] | [10, 512] 1 | net.0 | Linear | 131 K | train | 2.6 M | [10, 256] | [10, 512] 2 | net.1 | BatchNorm1d | 1.0 K | train | 0 | [10, 512] | [10, 512] ------------------------------------------------------------------------------ 132 K Trainable params 0 Non-trainable params 132 K Total params 0.530 Total estimated model params size (MB) 3 Modules in train mode 0 Modules in eval mode 2.6 M Total Flops """ def __init__(self, model: "pl.LightningModule", max_depth: int = 1) -> None: self._model = model if not isinstance(max_depth, int) or max_depth < -1: raise ValueError(f"`max_depth` can be -1, 0 or > 0, got {max_depth}.") # The max-depth needs to be plus one because the root module is already counted as depth 0. self._flop_counter = FlopCounterMode( mods=None if _TORCH_GREATER_EQUAL_2_4 else self._model, display=False, depth=max_depth + 1, ) self._max_depth = max_depth self._layer_summary = self.summarize() # 1 byte -> 8 bits # TODO: how do we compute precision_megabytes in case of mixed precision? precision_to_bits = { "64": 64, "32": 32, "16": 16, "bf16": 16, "16-true": 16, "bf16-true": 16, "32-true": 32, "64-true": 64, } if self._model._trainer and self._model.trainer.precision not in precision_to_bits: rank_zero_warn( f"Precision {self._model.trainer.precision} is not supported by the model summary. " " Estimated model size in MB will not be accurate. Using 32 bits instead.", category=UserWarning, ) precision = precision_to_bits.get(self._model.trainer.precision, 32) if self._model._trainer else 32 self._precision_megabytes = (precision / 8.0) * 1e-6 @property def named_modules(self) -> list[tuple[str, nn.Module]]: mods: list[tuple[str, nn.Module]] if self._max_depth == 0: mods = [] elif self._max_depth == 1: # the children are the top-level modules mods = list(self._model.named_children()) else: mods = self._model.named_modules() mods = list(mods)[1:] # do not include root module (LightningModule) return mods @property def layer_names(self) -> list[str]: return list(self._layer_summary.keys()) @property def layer_types(self) -> list[str]: return [layer.layer_type for layer in self._layer_summary.values()] @property def in_sizes(self) -> list: return [layer.in_size for layer in self._layer_summary.values()] @property def out_sizes(self) -> list: return [layer.out_size for layer in self._layer_summary.values()] @property def param_nums(self) -> list[int]: return [layer.num_parameters for layer in self._layer_summary.values()] @property def training_modes(self) -> list[bool]: return [layer.training for layer in self._layer_summary.values()] @property def total_training_modes(self) -> dict[str, int]: modes = [layer.training for layer in self._model.modules()] modes = modes[1:] # exclude the root module return {"train": modes.count(True), "eval": modes.count(False)} @property def total_parameters(self) -> int: return sum(p.numel() if not _tensor_has_shape(p) else 0 for p in self._model.parameters()) @property def trainable_parameters(self) -> int: return sum(p.numel() if not _tensor_has_shape(p) else 0 for p in self._model.parameters() if p.requires_grad) @property def total_layer_params(self) -> int: return sum(self.param_nums) @property def model_size(self) -> float: return self.total_parameters * self._precision_megabytes @property def total_flops(self) -> int: return self._flop_counter.get_total_flops() @property def flop_counts(self) -> dict[str, dict[Any, int]]: flop_counts = self._flop_counter.get_flop_counts() ret = { name: flop_counts.get( f"{type(self._model).__name__}.{name}", {}, ) for name in self.layer_names } return ret def summarize(self) -> dict[str, LayerSummary]: summary = OrderedDict((name, LayerSummary(module)) for name, module in self.named_modules) if self._model.example_input_array is not None: self._forward_example_input() for layer in summary.values(): layer.detach_hook() if self._max_depth >= 1: # remove summary entries with depth > max_depth for k in [k for k in summary if k.count(".") >= self._max_depth]: del summary[k] return summary def _forward_example_input(self) -> None: """Run the example input through each layer to get input- and output sizes.""" model = self._model # the summary is supported without a trainer instance so we need to use the underscore property trainer = self._model._trainer input_ = model.example_input_array input_ = model._on_before_batch_transfer(input_) input_ = model._apply_batch_transfer_handler(input_) mode = _ModuleMode() mode.capture(model) model.eval() # FlopCounterMode does not support ScriptModules before torch 2.4.0, so we use a null context flop_context = ( contextlib.nullcontext() if ( not _TORCH_GREATER_EQUAL_2_4 and any(isinstance(m, torch.jit.ScriptModule) for m in self._model.modules()) ) else self._flop_counter ) forward_context = contextlib.nullcontext() if trainer is None else trainer.precision_plugin.forward_context() with torch.no_grad(), forward_context, flop_context: # let the model hooks collect the input- and output shapes if isinstance(input_, (list, tuple)): model(*input_) elif isinstance(input_, dict): model(**input_) else: model(input_) mode.restore(model) def _get_summary_data(self) -> list[tuple[str, list[str]]]: """Makes a summary listing with: Layer Name, Layer Type, Number of Parameters, Input Sizes, Output Sizes, Model Size """ arrays = [ (" ", list(map(str, range(len(self._layer_summary))))), ("Name", self.layer_names), ("Type", self.layer_types), ("Params", list(map(get_human_readable_count, self.param_nums))), ("Mode", ["train" if mode else "eval" for mode in self.training_modes]), ("FLOPs", list(map(get_human_readable_count, (sum(x.values()) for x in self.flop_counts.values())))), ] if self._model.example_input_array is not None: arrays.append(("In sizes", [str(x) for x in self.in_sizes])) arrays.append(("Out sizes", [str(x) for x in self.out_sizes])) total_leftover_params = self.total_parameters - self.total_layer_params if total_leftover_params > 0: self._add_leftover_params_to_summary(arrays, total_leftover_params) return arrays def _add_leftover_params_to_summary(self, arrays: list[tuple[str, list[str]]], total_leftover_params: int) -> None: """Add summary of params not associated with module or layer to model summary.""" layer_summaries = dict(arrays) layer_summaries[" "].append(" ") layer_summaries["Name"].append(LEFTOVER_PARAMS_NAME) layer_summaries["Type"].append(NOT_APPLICABLE) layer_summaries["Params"].append(get_human_readable_count(total_leftover_params)) layer_summaries["Mode"].append(NOT_APPLICABLE) layer_summaries["FLOPs"].append(NOT_APPLICABLE) if "In sizes" in layer_summaries: layer_summaries["In sizes"].append(NOT_APPLICABLE) if "Out sizes" in layer_summaries: layer_summaries["Out sizes"].append(NOT_APPLICABLE) def __str__(self) -> str: arrays = self._get_summary_data() total_parameters = self.total_parameters trainable_parameters = self.trainable_parameters model_size = self.model_size total_training_modes = self.total_training_modes total_flops = self.total_flops return _format_summary_table( total_parameters, trainable_parameters, model_size, total_training_modes, total_flops, *arrays, ) def __repr__(self) -> str: return str(self) def parse_batch_shape(batch: Any) -> Union[str, list]: if hasattr(batch, "shape"): return list(batch.shape) if isinstance(batch, (list, tuple)): return [parse_batch_shape(el) for el in batch] return UNKNOWN_SIZE def _format_summary_table( total_parameters: int, trainable_parameters: int, model_size: float, total_training_modes: dict[str, int], total_flops: int, *cols: tuple[str, list[str]], ) -> str: """Takes in a number of arrays, each specifying a column in the summary table, and combines them all into one big string defining the summary table that are nicely formatted.""" n_rows = len(cols[0][1]) n_cols = 1 + len(cols) # Get formatting width of each column col_widths = [] for c in cols: col_width = max(len(str(a)) for a in c[1]) if n_rows else 0 col_width = max(col_width, len(c[0])) # minimum length is header length col_widths.append(col_width) # Formatting s = "{:<{}}" total_width = sum(col_widths) + 3 * n_cols header = [s.format(c[0], w) for c, w in zip(cols, col_widths)] # Summary = header + divider + Rest of table summary = " | ".join(header) + "\n" + "-" * total_width for i in range(n_rows): line = [] for c, w in zip(cols, col_widths): line.append(s.format(str(c[1][i]), w)) summary += "\n" + " | ".join(line) summary += "\n" + "-" * total_width summary += "\n" + s.format(get_human_readable_count(trainable_parameters), 10) summary += "Trainable params" summary += "\n" + s.format(get_human_readable_count(total_parameters - trainable_parameters), 10) summary += "Non-trainable params" summary += "\n" + s.format(get_human_readable_count(total_parameters), 10) summary += "Total params" summary += "\n" + s.format(get_formatted_model_size(model_size), 10) summary += "Total estimated model params size (MB)" summary += "\n" + s.format(total_training_modes["train"], 10) summary += "Modules in train mode" summary += "\n" + s.format(total_training_modes["eval"], 10) summary += "Modules in eval mode" summary += "\n" + s.format(get_human_readable_count(total_flops), 10) summary += "Total Flops" return summary def get_formatted_model_size(total_model_size: float) -> str: return f"{total_model_size:,.3f}" def get_human_readable_count(number: int) -> str: """Abbreviates an integer number with K, M, B, T for thousands, millions, billions and trillions, respectively. Examples: >>> get_human_readable_count(123) '123 ' >>> get_human_readable_count(1234) # (one thousand) '1.2 K' >>> get_human_readable_count(2e6) # (two million) '2.0 M' >>> get_human_readable_count(3e9) # (three billion) '3.0 B' >>> get_human_readable_count(4e14) # (four hundred trillion) '400 T' >>> get_human_readable_count(5e15) # (more than trillion) '5,000 T' Args: number: a positive integer number Return: A string formatted according to the pattern described above. """ assert number >= 0 labels = PARAMETER_NUM_UNITS num_digits = int(math.floor(math.log10(number)) + 1 if number > 0 else 1) num_groups = int(math.ceil(num_digits / 3)) num_groups = min(num_groups, len(labels)) # don't abbreviate beyond trillions shift = -3 * (num_groups - 1) number = number * (10**shift) index = num_groups - 1 if index < 1 or number >= 100: return f"{int(number):,d} {labels[index]}" return f"{number:,.1f} {labels[index]}" def _tensor_has_shape(p: Tensor) -> bool: from torch.nn.parameter import UninitializedParameter # DTensor is a subtype of `UninitializedParameter`, but the shape is known if isinstance(p, UninitializedParameter) and not _is_dtensor(p): warning_cache.warn( "The total number of parameters detected may be inaccurate because the model contains" " an instance of `UninitializedParameter`. To get an accurate number, set `self.example_input_array`" " in your LightningModule." ) return True return False def summarize(lightning_module: "pl.LightningModule", max_depth: int = 1) -> ModelSummary: """Summarize the LightningModule specified by `lightning_module`. Args: lightning_module: `LightningModule` to summarize. max_depth: The maximum depth of layer nesting that the summary will include. A value of 0 turns the layer summary off. Default: 1. Return: The model summary object """ return ModelSummary(lightning_module, max_depth=max_depth)