add flops profiler, support for Zipformer encoder and Conformer encoder (#1093)

* add flops profiler, support for Zipformer encoder and Conformer encoder

* support for reworked conformer and old zipformer

* skip black check

egs/librispeech/ASR/pruned_transducer_stateless/profile.py

@@ -0,0 +1,94 @@
+#!/usr/bin/env python3
+#
+# Copyright 2023 Xiaomi Corporation     (Author: Zengwei Yao)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# 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.
+
+"""
+Usage: ./pruned_transducer_stateless/profile.py
+"""
+
+import argparse
+import logging
+import sentencepiece as spm
+import torch
+
+from icefall.profiler import get_model_profile
+from train import get_encoder_model, add_model_arguments, get_params
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--bpe-model",
+        type=str,
+        default="data/lang_bpe_500/bpe.model",
+        help="Path to the BPE model",
+    )
+
+    add_model_arguments(parser)
+
+    return parser
+
+
+@torch.no_grad()
+def main():
+    parser = get_parser()
+    args = parser.parse_args()
+
+    params = get_params()
+    params.update(vars(args))
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", 0)
+    logging.info(f"Device: {device}")
+
+    sp = spm.SentencePieceProcessor()
+    sp.load(params.bpe_model)
+
+    # <blk> is defined in local/train_bpe_model.py
+    params.blank_id = sp.piece_to_id("<blk>")
+    params.vocab_size = sp.get_piece_size()
+
+    logging.info(params)
+
+    logging.info("About to create model")
+
+    # We only profile the encoder part
+    model = get_encoder_model(params)
+    model.eval()
+    model.to(device)
+
+    num_param = sum([p.numel() for p in model.parameters()])
+    logging.info(f"Number of model parameters: {num_param}")
+
+    # for 30-second input
+    B, T, D = 1, 3000, 80
+    feature = torch.ones(B, T, D, dtype=torch.float32).to(device)
+    feature_lens = torch.full((B,), T, dtype=torch.int64).to(device)
+
+    flops, params = get_model_profile(model=model, args=(feature, feature_lens))
+    logging.info(f"For the encoder part, params: {params}, flops: {flops}")
+
+
+if __name__ == "__main__":
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+    logging.basicConfig(format=formatter, level=logging.INFO)
+
+    main()

egs/librispeech/ASR/pruned_transducer_stateless4/profile.py

@@ -0,0 +1,146 @@
+#!/usr/bin/env python3
+#
+# Copyright 2023 Xiaomi Corporation     (Author: Zengwei Yao)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# 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.
+
+"""
+Usage: ./pruned_transducer_stateless4/profile.py
+"""
+
+import argparse
+import logging
+import sentencepiece as spm
+import torch
+
+from typing import Tuple
+from torch import Tensor, nn
+
+from icefall.profiler import get_model_profile
+from scaling import BasicNorm, DoubleSwish
+from train import get_encoder_model, get_joiner_model, add_model_arguments, get_params
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--bpe-model",
+        type=str,
+        default="data/lang_bpe_500/bpe.model",
+        help="Path to the BPE model",
+    )
+
+    add_model_arguments(parser)
+
+    return parser
+
+
+def _basic_norm_flops_compute(module, input, output):
+    assert len(input) == 1, len(input)
+    # estimate as layer_norm, see icefall/profiler.py
+    flops = input[0].numel() * 5
+    module.__flops__ += int(flops)
+
+
+def _doubleswish_module_flops_compute(module, input, output):
+    # For DoubleSwish
+    assert len(input) == 1, len(input)
+    # estimate as swish/silu, see icefall/profiler.py
+    flops = input[0].numel()
+    module.__flops__ += int(flops)
+
+
+MODULE_HOOK_MAPPING = {
+    BasicNorm: _basic_norm_flops_compute,
+    DoubleSwish: _doubleswish_module_flops_compute,
+}
+
+
+class Model(nn.Module):
+    """A Wrapper for encoder and encoder_proj"""
+
+    def __init__(
+        self,
+        encoder: nn.Module,
+        encoder_proj: nn.Module,
+    ) -> None:
+        super().__init__()
+        self.encoder = encoder
+        self.encoder_proj = encoder_proj
+
+    def forward(self, feature: Tensor, feature_lens: Tensor) -> Tuple[Tensor, Tensor]:
+        encoder_out, encoder_out_lens = self.encoder(feature, feature_lens)
+
+        logits = self.encoder_proj(encoder_out)
+
+        return logits, encoder_out_lens
+
+
+@torch.no_grad()
+def main():
+    parser = get_parser()
+    args = parser.parse_args()
+
+    params = get_params()
+    params.update(vars(args))
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", 0)
+    logging.info(f"Device: {device}")
+
+    sp = spm.SentencePieceProcessor()
+    sp.load(params.bpe_model)
+
+    # <blk> is defined in local/train_bpe_model.py
+    params.blank_id = sp.piece_to_id("<blk>")
+    params.vocab_size = sp.get_piece_size()
+
+    logging.info(params)
+
+    logging.info("About to create model")
+
+    # We only profile the encoder part
+    model = Model(
+        encoder=get_encoder_model(params),
+        encoder_proj=get_joiner_model(params).encoder_proj,
+    )
+    model.eval()
+    model.to(device)
+
+    num_param = sum([p.numel() for p in model.parameters()])
+    logging.info(f"Number of model parameters: {num_param}")
+
+    # for 30-second input
+    B, T, D = 1, 3000, 80
+    feature = torch.ones(B, T, D, dtype=torch.float32).to(device)
+    feature_lens = torch.full((B,), T, dtype=torch.int64).to(device)
+
+    flops, params = get_model_profile(
+        model=model,
+        args=(feature, feature_lens),
+        module_hoop_mapping=MODULE_HOOK_MAPPING,
+    )
+    logging.info(f"For the encoder part, params: {params}, flops: {flops}")
+
+
+if __name__ == "__main__":
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+    logging.basicConfig(format=formatter, level=logging.INFO)
+
+    main()

egs/librispeech/ASR/pruned_transducer_stateless7/profile.py

@@ -0,0 +1,146 @@
+#!/usr/bin/env python3
+#
+# Copyright 2023 Xiaomi Corporation     (Author: Zengwei Yao)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# 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.
+
+"""
+Usage: ./pruned_transducer_stateless7/profile.py
+"""
+
+import argparse
+import logging
+import sentencepiece as spm
+import torch
+
+from typing import Tuple
+from torch import Tensor, nn
+
+from icefall.profiler import get_model_profile
+from scaling import BasicNorm, DoubleSwish
+from train import get_encoder_model, get_joiner_model, add_model_arguments, get_params
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--bpe-model",
+        type=str,
+        default="data/lang_bpe_500/bpe.model",
+        help="Path to the BPE model",
+    )
+
+    add_model_arguments(parser)
+
+    return parser
+
+
+def _basic_norm_flops_compute(module, input, output):
+    assert len(input) == 1, len(input)
+    # estimate as layer_norm, see icefall/profiler.py
+    flops = input[0].numel() * 5
+    module.__flops__ += int(flops)
+
+
+def _doubleswish_module_flops_compute(module, input, output):
+    # For DoubleSwish
+    assert len(input) == 1, len(input)
+    # estimate as swish/silu, see icefall/profiler.py
+    flops = input[0].numel()
+    module.__flops__ += int(flops)
+
+
+MODULE_HOOK_MAPPING = {
+    BasicNorm: _basic_norm_flops_compute,
+    DoubleSwish: _doubleswish_module_flops_compute,
+}
+
+
+class Model(nn.Module):
+    """A Wrapper for encoder and encoder_proj"""
+
+    def __init__(
+        self,
+        encoder: nn.Module,
+        encoder_proj: nn.Module,
+    ) -> None:
+        super().__init__()
+        self.encoder = encoder
+        self.encoder_proj = encoder_proj
+
+    def forward(self, feature: Tensor, feature_lens: Tensor) -> Tuple[Tensor, Tensor]:
+        encoder_out, encoder_out_lens = self.encoder(feature, feature_lens)
+
+        logits = self.encoder_proj(encoder_out)
+
+        return logits, encoder_out_lens
+
+
+@torch.no_grad()
+def main():
+    parser = get_parser()
+    args = parser.parse_args()
+
+    params = get_params()
+    params.update(vars(args))
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", 0)
+    logging.info(f"Device: {device}")
+
+    sp = spm.SentencePieceProcessor()
+    sp.load(params.bpe_model)
+
+    # <blk> is defined in local/train_bpe_model.py
+    params.blank_id = sp.piece_to_id("<blk>")
+    params.vocab_size = sp.get_piece_size()
+
+    logging.info(params)
+
+    logging.info("About to create model")
+
+    # We only profile the encoder part
+    model = Model(
+        encoder=get_encoder_model(params),
+        encoder_proj=get_joiner_model(params).encoder_proj,
+    )
+    model.eval()
+    model.to(device)
+
+    num_param = sum([p.numel() for p in model.parameters()])
+    logging.info(f"Number of model parameters: {num_param}")
+
+    # for 30-second input
+    B, T, D = 1, 3000, 80
+    feature = torch.ones(B, T, D, dtype=torch.float32).to(device)
+    feature_lens = torch.full((B,), T, dtype=torch.int64).to(device)
+
+    flops, params = get_model_profile(
+        model=model,
+        args=(feature, feature_lens),
+        module_hoop_mapping=MODULE_HOOK_MAPPING,
+    )
+    logging.info(f"For the encoder part, params: {params}, flops: {flops}")
+
+
+if __name__ == "__main__":
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+    logging.basicConfig(format=formatter, level=logging.INFO)
+
+    main()

egs/librispeech/ASR/zipformer/profile.py

@@ -0,0 +1,176 @@
+#!/usr/bin/env python3
+#
+# Copyright 2023 Xiaomi Corporation     (Author: Zengwei Yao)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# 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.
+
+"""
+Usage: ./zipformer/profile.py
+"""
+
+import argparse
+import logging
+import sentencepiece as spm
+import torch
+
+from typing import Tuple
+from torch import Tensor, nn
+
+from icefall.utils import make_pad_mask
+from icefall.profiler import get_model_profile
+from scaling import BiasNorm
+from train import (
+    get_encoder_embed,
+    get_encoder_model,
+    get_joiner_model,
+    add_model_arguments,
+    get_params,
+)
+from zipformer import BypassModule
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--bpe-model",
+        type=str,
+        default="data/lang_bpe_500/bpe.model",
+        help="Path to the BPE model",
+    )
+
+    add_model_arguments(parser)
+
+    return parser
+
+
+def _bias_norm_flops_compute(module, input, output):
+    assert len(input) == 1, len(input)
+    # estimate as layer_norm, see icefall/profiler.py
+    flops = input[0].numel() * 5
+    module.__flops__ += int(flops)
+
+
+def _swoosh_module_flops_compute(module, input, output):
+    # For SwooshL and SwooshR modules
+    assert len(input) == 1, len(input)
+    # estimate as swish/silu, see icefall/profiler.py
+    flops = input[0].numel()
+    module.__flops__ += int(flops)
+
+
+def _bypass_module_flops_compute(module, input, output):
+    # For Bypass module
+    assert len(input) == 2, len(input)
+    flops = input[0].numel() * 2
+    module.__flops__ += int(flops)
+
+
+MODULE_HOOK_MAPPING = {
+    BiasNorm: _bias_norm_flops_compute,
+    BypassModule: _bypass_module_flops_compute,
+}
+
+
+class Model(nn.Module):
+    """A Wrapper for encoder, encoder_embed, and encoder_proj"""
+
+    def __init__(
+        self,
+        encoder: nn.Module,
+        encoder_embed: nn.Module,
+        encoder_proj: nn.Module,
+    ) -> None:
+        super().__init__()
+        self.encoder = encoder
+        self.encoder_embed = encoder_embed
+        self.encoder_proj = encoder_proj
+
+    def forward(
+        self, feature: Tensor, feature_lens: Tensor
+    ) -> Tuple[Tensor, Tensor]:
+        x, x_lens = self.encoder_embed(feature, feature_lens)
+
+        src_key_padding_mask = make_pad_mask(x_lens)
+        x = x.permute(1, 0, 2)  # (N, T, C) -> (T, N, C)
+
+        encoder_out, encoder_out_lens = self.encoder(
+            x, x_lens, src_key_padding_mask
+        )
+
+        encoder_out = encoder_out.permute(1, 0, 2)  # (N, T, C) -> (T, N, C)
+        logits = self.encoder_proj(encoder_out)
+
+        return logits, encoder_out_lens
+
+
+@torch.no_grad()
+def main():
+    parser = get_parser()
+    args = parser.parse_args()
+
+    params = get_params()
+    params.update(vars(args))
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", 0)
+    logging.info(f"Device: {device}")
+
+    sp = spm.SentencePieceProcessor()
+    sp.load(params.bpe_model)
+
+    # <blk> is defined in local/train_bpe_model.py
+    params.blank_id = sp.piece_to_id("<blk>")
+    params.vocab_size = sp.get_piece_size()
+
+    logging.info(params)
+
+    logging.info("About to create model")
+
+    # We only profile the encoder part
+    model = Model(
+        encoder=get_encoder_model(params),
+        encoder_embed=get_encoder_embed(params),
+        encoder_proj=get_joiner_model(params).encoder_proj,
+    )
+    model.eval()
+    model.to(device)
+
+    num_param = sum([p.numel() for p in model.parameters()])
+    logging.info(f"Number of model parameters: {num_param}")
+
+    # for 30-second input
+    B, T, D = 1, 3000, 80
+    feature = torch.ones(B, T, D, dtype=torch.float32).to(device)
+    feature_lens = torch.full((B,), T, dtype=torch.int64).to(device)
+
+    flops, params = get_model_profile(
+        model=model,
+        args=(feature, feature_lens),
+        module_hoop_mapping=MODULE_HOOK_MAPPING,
+    )
+    logging.info(f"For the encoder part, params: {params}, flops: {flops}")
+
+
+if __name__ == "__main__":
+    formatter = (
+        "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+    )
+    logging.basicConfig(format=formatter, level=logging.INFO)
+
+    main()

icefall/profiler.py

@@ -0,0 +1,941 @@
+# Copyright (c) Microsoft Corporation.
+# SPDX-License-Identifier: Apache-2.0
+
+# DeepSpeed Team
+
+# This is modified from https://github.com/microsoft/DeepSpeed/blob/master/deepspeed/profiling/flops_profiler/profiler.py
+
+import k2
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+from functools import partial
+from typing import List, Optional
+from collections import OrderedDict
+import numpy as np
+
+Tensor = torch.Tensor
+
+module_flop_count = []
+old_functions = {}
+
+
+class FlopsProfiler(object):
+    """Measures the latency, number of estimated floating-point operations and parameters of each module in a PyTorch model.
+
+    The flops-profiler profiles the forward pass of a PyTorch model and prints the model graph with the measured profile attached to each module. It shows how latency, flops and parameters are spent in the model and which modules or layers could be the bottleneck. It also outputs the names of the top k modules in terms of aggregated latency, flops, and parameters at depth l with k and l specified by the user. The output profile is computed for each batch of input.
+
+    To profile a trained model in inference, use the `get_model_profile` API.
+
+    Args:
+        object (torch.nn.Module): The PyTorch model to profile.
+    """
+
+    def __init__(self, model, module_hoop_mapping=None):
+        self.model = model
+        self.started = False
+        self.func_patched = False
+        self.module_hoop_mapping = (
+            module_hoop_mapping
+            if module_hoop_mapping is not None
+            else MODULE_HOOK_MAPPING
+        )
+
+    def start_profile(self, ignore_list=None):
+        """Starts profiling.
+
+        Extra attributes are added recursively to all the modules and the profiled torch.nn.functionals are monkey patched.
+
+        Args:
+            ignore_list (list, optional): the list of modules to ignore while profiling. Defaults to None.
+        """
+        self.reset_profile()
+        _patch_functionals()
+        _patch_tensor_methods()
+
+        def register_module_hooks(module, ignore_list):
+            if ignore_list and type(module) in ignore_list:
+                return
+
+            # if computing the flops of a module directly
+            if type(module) in self.module_hoop_mapping:
+                if not hasattr(module, "__flops_handle__"):
+                    module.__flops_handle__ = module.register_forward_hook(
+                        self.module_hoop_mapping[type(module)]
+                    )
+                return
+
+            # if computing the flops of the functionals in a module
+            def pre_hook(module, input):
+                module_flop_count.append([])
+
+            if not hasattr(module, "__pre_hook_handle__"):
+                module.__pre_hook_handle__ = module.register_forward_pre_hook(
+                    pre_hook
+                )
+
+            def post_hook(module, input, output):
+                if module_flop_count:
+                    module.__flops__ += sum(
+                        [elem[1] for elem in module_flop_count[-1]]
+                    )
+                    module_flop_count.pop()
+
+            if not hasattr(module, "__post_hook_handle__"):
+                module.__post_hook_handle__ = module.register_forward_hook(
+                    post_hook
+                )
+
+        self.model.apply(
+            partial(register_module_hooks, ignore_list=ignore_list)
+        )
+        self.started = True
+        self.func_patched = True
+
+    def stop_profile(self):
+        """Stop profiling.
+
+        All torch.nn.functionals are restored to their originals.
+        """
+        if self.started and self.func_patched:
+            _reload_functionals()
+            _reload_tensor_methods()
+            self.func_patched = False
+
+        def remove_profile_attrs(module):
+            if hasattr(module, "__pre_hook_handle__"):
+                module.__pre_hook_handle__.remove()
+                del module.__pre_hook_handle__
+            if hasattr(module, "__post_hook_handle__"):
+                module.__post_hook_handle__.remove()
+                del module.__post_hook_handle__
+            if hasattr(module, "__flops_handle__"):
+                module.__flops_handle__.remove()
+                del module.__flops_handle__
+
+        self.model.apply(remove_profile_attrs)
+
+    def reset_profile(self):
+        """Resets the profiling.
+
+        Adds or resets the extra attributes.
+        """
+
+        def add_or_reset_attrs(module):
+            module.__flops__ = 0
+            module.__params__ = sum(p.numel() for p in module.parameters())
+
+        self.model.apply(add_or_reset_attrs)
+
+    def end_profile(self):
+        """Ends profiling.
+
+        The added attributes and handles are removed recursively on all the modules.
+        """
+        if not self.started:
+            return
+        self.stop_profile()
+        self.started = False
+
+        def remove_profile_attrs(module):
+            if hasattr(module, "__flops__"):
+                del module.__flops__
+            if hasattr(module, "__params__"):
+                del module.__params__
+
+        self.model.apply(remove_profile_attrs)
+
+    def get_total_flops(self, as_string=False):
+        """Returns the total flops of the model.
+
+        Args:
+            as_string (bool, optional): whether to output the flops as string. Defaults to False.
+
+        Returns:
+            The number of multiply-accumulate operations of the model forward pass.
+        """
+        total_flops = get_module_flops(self.model)
+        return num_to_string(total_flops) if as_string else total_flops
+
+    def get_total_params(self, as_string=False):
+        """Returns the total parameters of the model.
+
+        Args:
+            as_string (bool, optional): whether to output the parameters as string. Defaults to False.
+
+        Returns:
+            The number of parameters in the model.
+        """
+        return (
+            params_to_string(self.model.__params__)
+            if as_string
+            else self.model.__params__
+        )
+
+
+def _prod(dims):
+    p = 1
+    for v in dims:
+        p *= v
+    return p
+
+
+def _linear_flops_compute(input, weight, bias=None):
+    out_features = weight.shape[0]
+    macs = input.numel() * out_features
+    return 2 * macs
+
+
+def _relu_flops_compute(input, inplace=False):
+    return input.numel()
+
+
+def _prelu_flops_compute(input: Tensor, weight: Tensor):
+    return input.numel()
+
+
+def _elu_flops_compute(
+    input: Tensor, alpha: float = 1.0, inplace: bool = False
+):
+    return input.numel()
+
+
+def _leaky_relu_flops_compute(
+    input: Tensor, negative_slope: float = 0.01, inplace: bool = False
+):
+    return input.numel()
+
+
+def _relu6_flops_compute(input: Tensor, inplace: bool = False):
+    return input.numel()
+
+
+def _silu_flops_compute(input: Tensor, inplace: bool = False):
+    return input.numel()
+
+
+def _gelu_flops_compute(input, **kwargs):
+    return input.numel()
+
+
+def _pool_flops_compute(
+    input,
+    kernel_size,
+    stride=None,
+    padding=0,
+    dilation=None,
+    ceil_mode=False,
+    count_include_pad=True,
+    divisor_override=None,
+    return_indices=None,
+):
+    return input.numel()
+
+
+def _conv_flops_compute(
+    input, weight, bias=None, stride=1, padding=0, dilation=1, groups=1
+):
+    assert weight.shape[1] * groups == input.shape[1]
+
+    batch_size = input.shape[0]
+    in_channels = input.shape[1]
+    out_channels = weight.shape[0]
+    kernel_dims = list(weight.shape[2:])
+    input_dims = list(input.shape[2:])
+
+    length = len(input_dims)
+
+    paddings = padding if type(padding) is tuple else (padding,) * length
+    strides = stride if type(stride) is tuple else (stride,) * length
+    dilations = dilation if type(dilation) is tuple else (dilation,) * length
+
+    output_dims = []
+    for idx, input_dim in enumerate(input_dims):
+        output_dim = (
+            input_dim
+            + 2 * paddings[idx]
+            - (dilations[idx] * (kernel_dims[idx] - 1) + 1)
+        ) // strides[idx] + 1
+        output_dims.append(output_dim)
+
+    filters_per_channel = out_channels // groups
+    conv_per_position_macs = (
+        int(_prod(kernel_dims)) * in_channels * filters_per_channel
+    )
+    active_elements_count = batch_size * int(_prod(output_dims))
+    overall_conv_macs = conv_per_position_macs * active_elements_count
+    overall_conv_flops = 2 * overall_conv_macs
+
+    bias_flops = 0
+    if bias is not None:
+        bias_flops = out_channels * active_elements_count
+
+    return int(overall_conv_flops + bias_flops)
+
+
+def _conv_trans_flops_compute(
+    input,
+    weight,
+    bias=None,
+    stride=1,
+    padding=0,
+    output_padding=0,
+    groups=1,
+    dilation=1,
+):
+    batch_size = input.shape[0]
+    in_channels = input.shape[1]
+    out_channels = weight.shape[0]
+    kernel_dims = list(weight.shape[2:])
+    input_dims = list(input.shape[2:])
+
+    length = len(input_dims)
+
+    paddings = padding if type(padding) is tuple else (padding,) * length
+    strides = stride if type(stride) is tuple else (stride,) * length
+    dilations = dilation if type(dilation) is tuple else (dilation,) * length
+
+    output_dims = []
+    for idx, input_dim in enumerate(input_dims):
+
+        output_dim = (
+            input_dim
+            + 2 * paddings[idx]
+            - (dilations[idx] * (kernel_dims[idx] - 1) + 1)
+        ) // strides[idx] + 1
+        output_dims.append(output_dim)
+
+    paddings = padding if type(padding) is tuple else (padding, padding)
+    strides = stride if type(stride) is tuple else (stride, stride)
+    dilations = dilation if type(dilation) is tuple else (dilation, dilation)
+
+    filters_per_channel = out_channels // groups
+    conv_per_position_macs = (
+        int(_prod(kernel_dims)) * in_channels * filters_per_channel
+    )
+    active_elements_count = batch_size * int(_prod(input_dims))
+    overall_conv_macs = conv_per_position_macs * active_elements_count
+    overall_conv_flops = 2 * overall_conv_macs
+
+    bias_flops = 0
+    if bias is not None:
+        bias_flops = out_channels * batch_size * int(_prod(output_dims))
+
+    return int(overall_conv_flops + bias_flops)
+
+
+def _batch_norm_flops_compute(
+    input,
+    running_mean,
+    running_var,
+    weight=None,
+    bias=None,
+    training=False,
+    momentum=0.1,
+    eps=1e-05,
+):
+    has_affine = weight is not None
+    if training:
+        # estimation
+        return input.numel() * (5 if has_affine else 4), 0
+    flops = input.numel() * (2 if has_affine else 1)
+    return flops
+
+
+def _layer_norm_flops_compute(
+    input: Tensor,
+    normalized_shape: List[int],
+    weight: Optional[Tensor] = None,
+    bias: Optional[Tensor] = None,
+    eps: float = 1e-5,
+):
+    has_affine = weight is not None
+    # estimation
+    return input.numel() * (5 if has_affine else 4)
+
+
+def _group_norm_flops_compute(
+    input: Tensor,
+    num_groups: int,
+    weight: Optional[Tensor] = None,
+    bias: Optional[Tensor] = None,
+    eps: float = 1e-5,
+):
+    has_affine = weight is not None
+    # estimation
+    return input.numel() * (5 if has_affine else 4)
+
+
+def _instance_norm_flops_compute(
+    input: Tensor,
+    running_mean: Optional[Tensor] = None,
+    running_var: Optional[Tensor] = None,
+    weight: Optional[Tensor] = None,
+    bias: Optional[Tensor] = None,
+    use_input_stats: bool = True,
+    momentum: float = 0.1,
+    eps: float = 1e-5,
+):
+    has_affine = weight is not None
+    # estimation
+    return input.numel() * (5 if has_affine else 4)
+
+
+def _upsample_flops_compute(input, **kwargs):
+    size = kwargs.get("size", None)
+    if size is not None:
+        if isinstance(size, tuple) or isinstance(size, list):
+            return int(_prod(size)), 0
+        else:
+            return int(size), 0
+    scale_factor = kwargs.get("scale_factor", None)
+    assert (
+        scale_factor is not None
+    ), "either size or scale_factor should be defined"
+    flops = input.numel()
+    if isinstance(scale_factor, tuple) and len(scale_factor) == len(input):
+        flops * int(_prod(scale_factor))
+    else:
+        flops * scale_factor ** len(input)
+    return flops
+
+
+def _softmax_flops_compute(input, dim=None, _stacklevel=3, dtype=None):
+    return input.numel()
+
+
+def _sigmoid_flops_compute(input):
+    return input.numel()
+
+
+def _embedding_flops_compute(
+    input,
+    weight,
+    padding_idx=None,
+    max_norm=None,
+    norm_type=2.0,
+    scale_grad_by_freq=False,
+    sparse=False,
+):
+    return 0
+
+
+def _dropout_flops_compute(input, p=0.5, training=True, inplace=False):
+    return 0
+
+
+def _matmul_flops_compute(input, other, *, out=None):
+    """
+    Count flops for the matmul operation.
+    """
+    macs = _prod(input.shape) * other.shape[-1]
+    return 2 * macs
+
+
+def _addmm_flops_compute(input, mat1, mat2, *, beta=1, alpha=1, out=None):
+    """
+    Count flops for the addmm operation.
+    """
+    macs = _prod(mat1.shape) * mat2.shape[-1]
+    return 2 * macs + _prod(input.shape)
+
+
+def _einsum_flops_compute(equation, *operands):
+    """
+    Count flops for the einsum operation.
+    """
+    equation = equation.replace(" ", "")
+    input_shapes = [o.shape for o in operands]
+
+    # Re-map equation so that same equation with different alphabet
+    # representations will look the same.
+    letter_order = OrderedDict((k, 0) for k in equation if k.isalpha()).keys()
+    mapping = {ord(x): 97 + i for i, x in enumerate(letter_order)}
+    equation = equation.translate(mapping)
+
+    np_arrs = [np.zeros(s) for s in input_shapes]
+    optim = np.einsum_path(equation, *np_arrs, optimize="optimal")[1]
+    for line in optim.split("\n"):
+        if "optimized flop" in line.lower():
+            flop = int(float(line.split(":")[-1]))
+            return flop
+    raise NotImplementedError("Unsupported einsum operation.")
+
+
+def _tensor_addmm_flops_compute(self, mat1, mat2, *, beta=1, alpha=1, out=None):
+    """
+    Count flops for the tensor addmm operation.
+    """
+    macs = _prod(mat1.shape) * mat2.shape[-1]
+    return 2 * macs + _prod(self.shape)
+
+
+def _mul_flops_compute(input, other, *, out=None):
+    print("mul")
+    return _elementwise_flops_compute(input, other)
+
+
+def _add_flops_compute(input, other, *, alpha=1, out=None):
+    print("add")
+    return _elementwise_flops_compute(input, other)
+
+
+def _sum_flops_compute(input, dim, keepdim=False):
+    return input.numel()
+
+
+def _elementwise_flops_compute(input, other):
+    if not torch.is_tensor(input):
+        if torch.is_tensor(other):
+            return _prod(other.shape)
+        else:
+            return 1
+    elif not torch.is_tensor(other):
+        return _prod(input.shape)
+    else:
+        dim_input = len(input.shape)
+        dim_other = len(other.shape)
+        max_dim = max(dim_input, dim_other)
+
+        final_shape = []
+        for i in range(max_dim):
+            in_i = input.shape[i] if i < dim_input else 1
+            ot_i = other.shape[i] if i < dim_other else 1
+            if in_i > ot_i:
+                final_shape.append(in_i)
+            else:
+                final_shape.append(ot_i)
+        flops = _prod(final_shape)
+        return flops
+
+
+def _tanh_flops_compute(input):
+    return input.numel()
+
+
+def _k2_swoosh_flops_compute(input):
+    # For SwooshLForward and SwooshRForward
+    # estimate as swish/silu
+    return input.numel()
+
+
+def wrapFunc(func, funcFlopCompute):
+    oldFunc = func
+    name = func.__str__
+    old_functions[name] = oldFunc
+
+    def newFunc(*args, **kwds):
+        flops = funcFlopCompute(*args, **kwds)
+        if module_flop_count:
+            module_flop_count[-1].append((name, flops))
+        return oldFunc(*args, **kwds)
+
+    newFunc.__str__ = func.__str__
+
+    return newFunc
+
+
+def _patch_functionals():
+    # FC
+    F.linear = wrapFunc(F.linear, _linear_flops_compute)
+
+    # convolutions
+    F.conv1d = wrapFunc(F.conv1d, _conv_flops_compute)
+    F.conv2d = wrapFunc(F.conv2d, _conv_flops_compute)
+    F.conv3d = wrapFunc(F.conv3d, _conv_flops_compute)
+
+    # conv transposed
+    F.conv_transpose1d = wrapFunc(F.conv_transpose1d, _conv_trans_flops_compute)
+    F.conv_transpose2d = wrapFunc(F.conv_transpose2d, _conv_trans_flops_compute)
+    F.conv_transpose3d = wrapFunc(F.conv_transpose3d, _conv_trans_flops_compute)
+
+    # activations
+    F.relu = wrapFunc(F.relu, _relu_flops_compute)
+    F.prelu = wrapFunc(F.prelu, _prelu_flops_compute)
+    F.elu = wrapFunc(F.elu, _elu_flops_compute)
+    F.leaky_relu = wrapFunc(F.leaky_relu, _leaky_relu_flops_compute)
+    F.relu6 = wrapFunc(F.relu6, _relu6_flops_compute)
+    if hasattr(F, "silu"):
+        F.silu = wrapFunc(F.silu, _silu_flops_compute)
+    F.gelu = wrapFunc(F.gelu, _gelu_flops_compute)
+
+    # Normalizations
+    F.batch_norm = wrapFunc(F.batch_norm, _batch_norm_flops_compute)
+    F.layer_norm = wrapFunc(F.layer_norm, _layer_norm_flops_compute)
+    F.instance_norm = wrapFunc(F.instance_norm, _instance_norm_flops_compute)
+    F.group_norm = wrapFunc(F.group_norm, _group_norm_flops_compute)
+
+    # poolings
+    F.avg_pool1d = wrapFunc(F.avg_pool1d, _pool_flops_compute)
+    F.avg_pool2d = wrapFunc(F.avg_pool2d, _pool_flops_compute)
+    F.avg_pool3d = wrapFunc(F.avg_pool3d, _pool_flops_compute)
+    F.max_pool1d = wrapFunc(F.max_pool1d, _pool_flops_compute)
+    F.max_pool2d = wrapFunc(F.max_pool2d, _pool_flops_compute)
+    F.max_pool3d = wrapFunc(F.max_pool3d, _pool_flops_compute)
+    F.adaptive_avg_pool1d = wrapFunc(F.adaptive_avg_pool1d, _pool_flops_compute)
+    F.adaptive_avg_pool2d = wrapFunc(F.adaptive_avg_pool2d, _pool_flops_compute)
+    F.adaptive_avg_pool3d = wrapFunc(F.adaptive_avg_pool3d, _pool_flops_compute)
+    F.adaptive_max_pool1d = wrapFunc(F.adaptive_max_pool1d, _pool_flops_compute)
+    F.adaptive_max_pool2d = wrapFunc(F.adaptive_max_pool2d, _pool_flops_compute)
+    F.adaptive_max_pool3d = wrapFunc(F.adaptive_max_pool3d, _pool_flops_compute)
+
+    # upsample
+    F.upsample = wrapFunc(F.upsample, _upsample_flops_compute)
+    F.interpolate = wrapFunc(F.interpolate, _upsample_flops_compute)
+
+    # softmax
+    F.softmax = wrapFunc(F.softmax, _softmax_flops_compute)
+
+    # sigmoid
+    F.sigmoid = wrapFunc(F.sigmoid, _sigmoid_flops_compute)
+
+    # embedding
+    F.embedding = wrapFunc(F.embedding, _embedding_flops_compute)
+
+    # swoosh functions in k2
+    k2.swoosh_l_forward = wrapFunc(
+        k2.swoosh_l_forward, _k2_swoosh_flops_compute
+    )
+    k2.swoosh_r_forward = wrapFunc(
+        k2.swoosh_r_forward, _k2_swoosh_flops_compute
+    )
+    k2.swoosh_l = wrapFunc(k2.swoosh_l, _k2_swoosh_flops_compute)
+    k2.swoosh_r = wrapFunc(k2.swoosh_r, _k2_swoosh_flops_compute)
+
+
+def _patch_tensor_methods():
+    torch.matmul = wrapFunc(torch.matmul, _matmul_flops_compute)
+    torch.Tensor.matmul = wrapFunc(torch.Tensor.matmul, _matmul_flops_compute)
+    torch.mm = wrapFunc(torch.mm, _matmul_flops_compute)
+    torch.Tensor.mm = wrapFunc(torch.Tensor.mm, _matmul_flops_compute)
+    torch.bmm = wrapFunc(torch.bmm, _matmul_flops_compute)
+    torch.Tensor.bmm = wrapFunc(torch.Tensor.bmm, _matmul_flops_compute)
+
+    torch.addmm = wrapFunc(torch.addmm, _addmm_flops_compute)
+    torch.Tensor.addmm = wrapFunc(
+        torch.Tensor.addmm, _tensor_addmm_flops_compute
+    )
+
+    torch.mul = wrapFunc(torch.mul, _mul_flops_compute)
+    torch.Tensor.mul = wrapFunc(torch.Tensor.mul, _mul_flops_compute)
+
+    torch.add = wrapFunc(torch.add, _add_flops_compute)
+    torch.Tensor.add = wrapFunc(torch.Tensor.add, _add_flops_compute)
+
+    torch.sum = wrapFunc(torch.sum, _sum_flops_compute)
+    torch.Tensor.sum = wrapFunc(torch.Tensor.sum, _sum_flops_compute)
+
+    torch.einsum = wrapFunc(torch.einsum, _einsum_flops_compute)
+
+    torch.baddbmm = wrapFunc(torch.baddbmm, _tensor_addmm_flops_compute)
+
+    torch.tanh = wrapFunc(torch.tanh, _tanh_flops_compute)
+
+    torch.Tensor.softmax = wrapFunc(
+        torch.Tensor.softmax, _softmax_flops_compute
+    )
+
+    torch.sigmoid = wrapFunc(torch.sigmoid, _sigmoid_flops_compute)
+    torch.Tensor.sigmoid = wrapFunc(
+        torch.Tensor.sigmoid, _sigmoid_flops_compute
+    )
+
+
+def _reload_functionals():
+    # torch.nn.functional does not support importlib.reload()
+    F.linear = old_functions[F.linear.__str__]
+    F.conv1d = old_functions[F.conv1d.__str__]
+    F.conv2d = old_functions[F.conv2d.__str__]
+    F.conv3d = old_functions[F.conv3d.__str__]
+    F.conv_transpose1d = old_functions[F.conv_transpose1d.__str__]
+    F.conv_transpose2d = old_functions[F.conv_transpose2d.__str__]
+    F.conv_transpose3d = old_functions[F.conv_transpose3d.__str__]
+    F.relu = old_functions[F.relu.__str__]
+    F.prelu = old_functions[F.prelu.__str__]
+    F.elu = old_functions[F.elu.__str__]
+    F.leaky_relu = old_functions[F.leaky_relu.__str__]
+    F.relu6 = old_functions[F.relu6.__str__]
+    if hasattr(F, "silu"):
+        F.silu = old_functions[F.silu.__str__]
+    F.gelu = old_functions[F.gelu.__str__]
+    F.batch_norm = old_functions[F.batch_norm.__str__]
+    F.layer_norm = old_functions[F.layer_norm.__str__]
+    F.instance_norm = old_functions[F.instance_norm.__str__]
+    F.group_norm = old_functions[F.group_norm.__str__]
+    F.avg_pool1d = old_functions[F.avg_pool1d.__str__]
+    F.avg_pool2d = old_functions[F.avg_pool2d.__str__]
+    F.avg_pool3d = old_functions[F.avg_pool3d.__str__]
+    F.max_pool1d = old_functions[F.max_pool1d.__str__]
+    F.max_pool2d = old_functions[F.max_pool2d.__str__]
+    F.max_pool3d = old_functions[F.max_pool3d.__str__]
+    F.adaptive_avg_pool1d = old_functions[F.adaptive_avg_pool1d.__str__]
+    F.adaptive_avg_pool2d = old_functions[F.adaptive_avg_pool2d.__str__]
+    F.adaptive_avg_pool3d = old_functions[F.adaptive_avg_pool3d.__str__]
+    F.adaptive_max_pool1d = old_functions[F.adaptive_max_pool1d.__str__]
+    F.adaptive_max_pool2d = old_functions[F.adaptive_max_pool2d.__str__]
+    F.adaptive_max_pool3d = old_functions[F.adaptive_max_pool3d.__str__]
+    F.upsample = old_functions[F.upsample.__str__]
+    F.interpolate = old_functions[F.interpolate.__str__]
+    F.softmax = old_functions[F.softmax.__str__]
+    F.sigmoid = old_functions[F.sigmoid.__str__]
+    F.embedding = old_functions[F.embedding.__str__]
+    # swoosh functions in k2
+    k2.swoosh_l = old_functions[k2.swoosh_l.__str__]
+    k2.swoosh_r = old_functions[k2.swoosh_r.__str__]
+    k2.swoosh_l_forward = old_functions[k2.swoosh_l_forward.__str__]
+    k2.swoosh_r_forward = old_functions[k2.swoosh_r_forward.__str__]
+
+
+def _reload_tensor_methods():
+    torch.matmul = old_functions[torch.matmul.__str__]
+    torch.Tensor.matmul = old_functions[torch.Tensor.matmul.__str__]
+    torch.mm = old_functions[torch.mm.__str__]
+    torch.Tensor.mm = old_functions[torch.Tensor.mm.__str__]
+    torch.bmm = old_functions[torch.matmul.__str__]
+    torch.Tensor.bmm = old_functions[torch.Tensor.bmm.__str__]
+    torch.addmm = old_functions[torch.addmm.__str__]
+    torch.Tensor.addmm = old_functions[torch.Tensor.addmm.__str__]
+    torch.mul = old_functions[torch.mul.__str__]
+    torch.Tensor.mul = old_functions[torch.Tensor.mul.__str__]
+    torch.add = old_functions[torch.add.__str__]
+    torch.Tensor.add = old_functions[torch.Tensor.add.__str__]
+    torch.sum = old_functions[torch.sum.__str__]
+    torch.Tensor.sum = old_functions[torch.Tensor.sum.__str__]
+
+    torch.einsum = old_functions[torch.einsum.__str__]
+
+    torch.baddbmm = old_functions[torch.baddbmm.__str__]
+
+    torch.Tensor.softmax = old_functions[torch.Tensor.softmax.__str__]
+
+    torch.sigmoid = old_functions[torch.sigmoid.__str__]
+    torch.Tensor.sigmoid = old_functions[torch.Tensor.sigmoid.__str__]
+
+
+def _rnn_flops(flops, rnn_module, w_ih, w_hh, input_size):
+    # matrix matrix mult ih state and internal state
+    flops += w_ih.shape[0] * w_ih.shape[1]
+    # matrix matrix mult hh state and internal state
+    flops += w_hh.shape[0] * w_hh.shape[1]
+    if isinstance(rnn_module, (nn.RNN, nn.RNNCell)):
+        # add both operations
+        flops += rnn_module.hidden_size
+    elif isinstance(rnn_module, (nn.GRU, nn.GRUCell)):
+        # hadamard of r
+        flops += rnn_module.hidden_size
+        # adding operations from both states
+        flops += rnn_module.hidden_size * 3
+        # last two hadamard _product and add
+        flops += rnn_module.hidden_size * 3
+    elif isinstance(rnn_module, (nn.LSTM, nn.LSTMCell)):
+        # adding operations from both states
+        flops += rnn_module.hidden_size * 4
+        # two hadamard _product and add for C state
+        flops += (
+            rnn_module.hidden_size
+            + rnn_module.hidden_size
+            + rnn_module.hidden_size
+        )
+        # final hadamard
+        flops += (
+            rnn_module.hidden_size
+            + rnn_module.hidden_size
+            + rnn_module.hidden_size
+        )
+    return flops
+
+
+def _rnn_forward_hook(rnn_module, input, output):
+    flops = 0
+    # input is a tuple containing a sequence to process and (optionally) hidden state
+    inp = input[0]
+    batch_size = inp.shape[0]
+    seq_length = inp.shape[1]
+    num_layers = rnn_module.num_layers
+
+    for i in range(num_layers):
+        w_ih = rnn_module.__getattr__("weight_ih_l" + str(i))
+        w_hh = rnn_module.__getattr__("weight_hh_l" + str(i))
+        if i == 0:
+            input_size = rnn_module.input_size
+        else:
+            input_size = rnn_module.hidden_size
+        flops = _rnn_flops(flops, rnn_module, w_ih, w_hh, input_size)
+        if rnn_module.bias:
+            b_ih = rnn_module.__getattr__("bias_ih_l" + str(i))
+            b_hh = rnn_module.__getattr__("bias_hh_l" + str(i))
+            flops += b_ih.shape[0] + b_hh.shape[0]
+
+    flops *= batch_size
+    flops *= seq_length
+    if rnn_module.bidirectional:
+        flops *= 2
+    rnn_module.__flops__ += int(flops)
+
+
+def _rnn_cell_forward_hook(rnn_cell_module, input, output):
+    flops = 0
+    inp = input[0]
+    batch_size = inp.shape[0]
+    w_ih = rnn_cell_module.__getattr__("weight_ih")
+    w_hh = rnn_cell_module.__getattr__("weight_hh")
+    input_size = inp.shape[1]
+    flops = _rnn_flops(flops, rnn_cell_module, w_ih, w_hh, input_size)
+    if rnn_cell_module.bias:
+        b_ih = rnn_cell_module.__getattr__("bias_ih")
+        b_hh = rnn_cell_module.__getattr__("bias_hh")
+        flops += b_ih.shape[0] + b_hh.shape[0]
+
+    flops *= batch_size
+    rnn_cell_module.__flops__ += int(flops)
+
+
+MODULE_HOOK_MAPPING = {
+    # RNN
+    nn.RNN: _rnn_forward_hook,
+    nn.GRU: _rnn_forward_hook,
+    nn.LSTM: _rnn_forward_hook,
+    nn.RNNCell: _rnn_cell_forward_hook,
+    nn.LSTMCell: _rnn_cell_forward_hook,
+    nn.GRUCell: _rnn_cell_forward_hook,
+}
+
+
+def num_to_string(num, precision=2):
+    if num // 10**9 > 0:
+        return str(round(num / 10.0**9, precision)) + " G"
+    elif num // 10**6 > 0:
+        return str(round(num / 10.0**6, precision)) + " M"
+    elif num // 10**3 > 0:
+        return str(round(num / 10.0**3, precision)) + " K"
+    else:
+        return str(num)
+
+
+def number_to_string(num, units=None, precision=2):
+    if units is None:
+        if num // 10**9 > 0:
+            return str(round(num / 10.0**9, precision)) + " G"
+        elif num // 10**6 > 0:
+            return str(round(num / 10.0**6, precision)) + " M"
+        elif num // 10**3 > 0:
+            return str(round(num / 10.0**3, precision)) + " K"
+        else:
+            return str(num) + " "
+    else:
+        if units == "G":
+            return str(round(num / 10.0**9, precision)) + " " + units
+        elif units == "M":
+            return str(round(num / 10.0**6, precision)) + " " + units
+        elif units == "K":
+            return str(round(num / 10.0**3, precision)) + " " + units
+        else:
+            return str(num) + " "
+
+
+def flops_to_string(flops, units=None, precision=2):
+    if units is None:
+        if flops // 10**12 > 0:
+            return str(round(flops / 10.0**12, precision)) + " TFLOPS"
+        if flops // 10**9 > 0:
+            return str(round(flops / 10.0**9, precision)) + " GFLOPS"
+        elif flops // 10**6 > 0:
+            return str(round(flops / 10.0**6, precision)) + " MFLOPS"
+        elif flops // 10**3 > 0:
+            return str(round(flops / 10.0**3, precision)) + " KFLOPS"
+        else:
+            return str(flops) + " FLOPS"
+    else:
+        if units == "TFLOPS":
+            return str(round(flops / 10.0**12, precision)) + " " + units
+        if units == "GFLOPS":
+            return str(round(flops / 10.0**9, precision)) + " " + units
+        elif units == "MFLOPS":
+            return str(round(flops / 10.0**6, precision)) + " " + units
+        elif units == "KFLOPS":
+            return str(round(flops / 10.0**3, precision)) + " " + units
+        else:
+            return str(flops) + " FLOPS"
+
+
+def params_to_string(params_num, units=None, precision=2):
+    if units is None:
+        if params_num // 10**6 > 0:
+            return str(round(params_num / 10**6, 2)) + " M"
+        elif params_num // 10**3:
+            return str(round(params_num / 10**3, 2)) + " k"
+        else:
+            return str(params_num)
+    else:
+        if units == "M":
+            return str(round(params_num / 10.0**6, precision)) + " " + units
+        elif units == "K":
+            return str(round(params_num / 10.0**3, precision)) + " " + units
+        else:
+            return str(params_num)
+
+
+def get_module_flops(module):
+    sum = module.__flops__
+    # iterate over immediate children modules
+    for child in module.children():
+        sum += get_module_flops(child)
+    return sum
+
+
+def get_module_duration(module):
+    duration = module.__duration__
+    if duration == 0:  # e.g. ModuleList
+        for m in module.children():
+            duration += m.__duration__
+    return duration
+
+
+def get_model_profile(
+    model,
+    args=[],
+    as_string=True,
+    ignore_modules=None,
+    module_hoop_mapping=None,
+):
+    """Returns the total floating-point operations, MACs, and parameters of a model.
+
+    Example:
+
+    .. code-block:: python
+
+        model = torchvision.models.alexnet()
+        batch_size = 256
+        flops, params = get_model_profile(model=model, args=(feature, feature_lens))
+
+    Args:
+        model ([torch.nn.Module]): the PyTorch model to be profiled.
+        args (list): list of positional arguments to the model.
+        top_modules (int, optional): the number of top modules to print in the aggregated profile. Defaults to 3.
+        as_string (bool, optional): whether to print the output as string. Defaults to True.
+        ignore_modules ([type], optional): the list of modules to ignore during profiling. Defaults to None.
+
+    Returns:
+        The number of floating-point operations, multiply-accumulate operations (MACs), and parameters in the model.
+    """
+    assert isinstance(model, nn.Module), "model must be a PyTorch module"
+    prof = FlopsProfiler(model, module_hoop_mapping=module_hoop_mapping)
+    model.eval()
+
+    assert len(args) > 0, "input args must be specified"
+
+    prof.start_profile(ignore_list=ignore_modules)
+
+    _ = model(*args)
+
+    flops = prof.get_total_flops()
+    params = prof.get_total_params()
+
+    prof.end_profile()
+    if as_string:
+        return (
+            number_to_string(flops),
+            params_to_string(params),
+        )
+
+    return flops, params

pyproject.toml

@@ -12,5 +12,6 @@ exclude = '''
   | make_kn_lm.py
   | icefall\/__init__\.py
   | icefall\/diagnostics\.py
+  | icefall\/profiler\.py
   | egs\/librispeech\/ASR\/zipformer
 '''