# Copyright (c) 2021-2023, InterDigital Communications, Inc
# All rights reserved.
# Redistribution and use in source and binary forms, with or without
# modification, are permitted (subject to the limitations in the disclaimer
# below) provided that the following conditions are met:
# * Redistributions of source code must retain the above copyright notice,
# this list of conditions and the following disclaimer.
# * Redistributions in binary form must reproduce the above copyright notice,
# this list of conditions and the following disclaimer in the documentation
# and/or other materials provided with the distribution.
# * Neither the name of InterDigital Communications, Inc nor the names of its
# contributors may be used to endorse or promote products derived from this
# software without specific prior written permission.
# NO EXPRESS OR IMPLIED LICENSES TO ANY PARTY'S PATENT RIGHTS ARE GRANTED BY
# THIS LICENSE. THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND
# CONTRIBUTORS "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT
# NOT LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A
# PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT HOLDER OR
# CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL,
# EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
# PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS;
# OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY,
# WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR
# OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF
# ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
r"""
Evaluate a model.
Evaluates a model on the configured ``dataset.infer`` (i.e. test set).
Saves bitstreams and reconstructed images to ``paths.output_dir``.
Computes metrics and saves per-file results and averaged results to
JSON and TSV files.
To evaluate models trained using CompressAI Trainer:
.. code-block:: bash
compressai-eval \
--config-path="$HOME/data/runs/e4e6d4d5e5c59c69f3bd7be2/configs" \
--config-path="$HOME/data/runs/d4d5e5c5e4e6bd7be29c69f3/configs" \
...
To evaluate models from the CompressAI zoo:
.. code-block:: bash
compressai-eval \
--config-name="eval_zoo" ++model.name="bmshj2018-factorized" ++model.quality=1 ++criterion.lmbda=0.0018 \
--config-name="eval_zoo" ++model.name="bmshj2018-factorized" ++model.quality=2 ++criterion.lmbda=0.0035 \
--config-name="eval_zoo" ++model.name="bmshj2018-factorized" ++model.quality=3 ++criterion.lmbda=0.0067 \
--config-name="eval_zoo" ++model.name="bmshj2018-factorized" ++model.quality=4 ++criterion.lmbda=0.0130 \
--config-name="eval_zoo" ++model.name="bmshj2018-factorized" ++model.quality=5 ++criterion.lmbda=0.0250 \
--config-name="eval_zoo" ++model.name="bmshj2018-factorized" ++model.quality=6 ++criterion.lmbda=0.0483 \
--config-name="eval_zoo" ++model.name="bmshj2018-factorized" ++model.quality=7 ++criterion.lmbda=0.0932 \
--config-name="eval_zoo" ++model.name="bmshj2018-factorized" ++model.quality=8 ++criterion.lmbda=0.1800
The above can be written more compactly by prepending a "default" override
``++model.name="bmshj2018-factorized"`` that applies to all configs:
.. code-block:: bash
compressai-eval \
++model.name="bmshj2018-factorized" \
--config-name="eval_zoo" ++model.quality=1 ++criterion.lmbda=0.0018 \
--config-name="eval_zoo" ++model.quality=2 ++criterion.lmbda=0.0035 \
--config-name="eval_zoo" ++model.quality=3 ++criterion.lmbda=0.0067 \
--config-name="eval_zoo" ++model.quality=4 ++criterion.lmbda=0.0130 \
--config-name="eval_zoo" ++model.quality=5 ++criterion.lmbda=0.0250 \
--config-name="eval_zoo" ++model.quality=6 ++criterion.lmbda=0.0483 \
--config-name="eval_zoo" ++model.quality=7 ++criterion.lmbda=0.0932 \
--config-name="eval_zoo" ++model.quality=8 ++criterion.lmbda=0.1800
By default, the following options are used, if not specified:
.. code-block:: bash
--config-path="conf"
--config-name="config"
++model.source="config"
# if model.source == "config":
++paths.output_dir="outputs/${model.source}-${env.aim.run_hash}-${model.name}"
++paths.model_checkpoint='${paths.checkpoints}/runner.last.pth'
# if model.source == "from_state_dict":
++paths.output_dir="outputs/${model.name}-${Path(paths.model_checkpoint).stem}"
# if model.source == "zoo":
++paths.output_dir="outputs/${model.source}-${model.name}-${model.metric}-${model.quality}"
The model is evaluated on ``dataset.infer``, which may be configured as follows:
.. code-block:: yaml
dataset:
infer:
type: "ImageFolder"
config:
root: "/path/to/directory/containing/images"
split: ""
loader:
shuffle: False
batch_size: 1
num_workers: 2
settings:
transforms:
- "ToTensor": {}
meta:
name: "Custom dataset"
identifier: "image/custom"
num_samples: 0 # ignored during eval
To evaluate a model on a custom directory of samples, use the above
config and override ``dataset.infer.config.root``.
"""
from __future__ import annotations
import json
import os
import sys
import types
from pathlib import Path
import catalyst
import catalyst.utils
import numpy as np
import pandas as pd
from omegaconf import DictConfig, OmegaConf, open_dict
from PIL import Image
from compressai_trainer.config import (
configure_conf,
create_criterion,
create_dataloaders,
create_runner,
load_model,
)
from compressai_trainer.typing import TRunner
from compressai_trainer.utils.args import iter_configs
from compressai_trainer.utils.metrics import _METRICS, compute_metrics
from compressai_trainer.utils.utils import ld_to_dl, tensor_to_np_img
thisdir = Path(__file__).parent
config_path = str(thisdir.joinpath("../../conf").resolve())
DEFAULT_MODEL_SOURCE = "config"
DEFAULT_PATHS_MODEL_CHECKPOINT = "${paths.checkpoints}/runner.last.pth"
DEFAULT_PATHS_OUTPUT_DIR_ROOT = "outputs"
[docs]def setup(conf: DictConfig) -> TRunner:
catalyst.utils.set_global_seed(conf.misc.seed)
catalyst.utils.prepare_cudnn(benchmark=False, deterministic=True)
configure_conf(conf)
model = load_model(conf).eval()
criterion = create_criterion(conf.criterion)
loaders = create_dataloaders(conf)
runner = create_runner(conf.runner)
runner.model = model
runner.criterion = criterion
runner.loaders = loaders
runner.engine = catalyst.utils.get_available_engine()
runner._hparams = OmegaConf.to_container(conf, resolve=True)
return runner
def _get_filenames(runner, num_files):
dataset = runner.loaders["infer"].dataset
if type(dataset).__name__ == "ImageFolder":
return [x.stem for x in dataset.samples]
return [f"unknown_{i:06d}" for i in range(1, num_files + 1)]
[docs]def run_eval_model(runner, batches, filenames, output_dir, metrics):
runner.model_module.update(force=True)
output_dir = Path(output_dir)
os.makedirs(output_dir, exist_ok=True)
outputs = []
for batch, filename in zip(batches, filenames):
assert len(batch) == 1
x = batch.to(runner.engine.device)
out_infer = runner.predict_batch(x)
out_criterion = runner.criterion(out_infer["out_net"], x)
out_metrics = compute_metrics(x, out_infer["out_dec"]["x_hat"], metrics)
output = {
"filename": filename,
"bpp": out_infer["bpp"],
**{
k: v.item()
for k, v in out_criterion.items()
if k in runner.hparams["runner"]["meters"]["infer"]
},
**out_metrics,
"encoding_time": out_infer["encoding_time"],
"decoding_time": out_infer["decoding_time"],
}
outputs.append(output)
print(json.dumps(output, indent=2))
output_filename = (output_dir / filename).with_suffix("")
os.makedirs(output_filename.parent, exist_ok=True)
_write_bitstreams(out_infer["out_enc"]["strings"], output_filename)
_write_image(out_infer["out_dec"]["x_hat"], output_filename)
return outputs
def _write_bitstreams(strings, filename):
for i, s in enumerate(strings):
assert len(s) == 1
with open(filename.with_suffix(f".{i:02d}.bin"), "wb") as f:
f.write(s[0])
def _write_image(x, filename):
assert len(x.shape) == 4 and x.shape[0] == 1
Image.fromarray(tensor_to_np_img(x[0])).save(filename.with_suffix(".png"))
def _plot_rd(runner, results):
def log_figure(self, tag, fig, runner=runner, context=None, **kwargs):
def slugify(s):
return f"{s}".replace("/", "-")
output_dir = runner.hparams["paths"]["output_dir"]
context_str = ";".join(f"{slugify(k)}={slugify(v)}" for k, v in context.items())
fig.write_html(f"{output_dir}/{tag};{context_str}.html")
runner.log_figure = types.MethodType(log_figure, runner)
runner._log_rd_curves()
def _plot_rd_all(runner, dfs):
def log_figure(self, tag, fig, runner=runner, context=None, **kwargs):
def slugify(s):
return f"{s}".replace("/", "-")
output_dir = DEFAULT_PATHS_OUTPUT_DIR_ROOT
context_str = ";".join(f"{slugify(k)}={slugify(v)}" for k, v in context.items())
fig.write_html(f"{output_dir}/{tag};{context_str}.html")
# WARNING: This uses the latest runner's hparams, so the context_str may be unusual.
runner.log_figure = types.MethodType(log_figure, runner)
runner._log_rd_curves(df=pd.concat(dfs), traces=[])
def _results_dict(conf, outputs):
result_keys = list(outputs[0].keys())
result_non_avg_keys = ["filename"]
result_avg_keys = [k for k in result_keys if k not in result_non_avg_keys]
return {
"name": conf.model.name,
"description": "",
"meta": {
"dataset": conf.dataset.infer.meta.name,
"env.aim.run_hash": conf.env.aim.run_hash,
"misc.device": conf.misc.device,
"model.source": conf.model.get("source"),
"model.name": conf.model.get("name"),
"model.metric": conf.model.get("metric"),
"model.quality": conf.model.get("quality"),
"criterion.lmbda": conf.criterion.get("lmbda"),
"paths.model_checkpoint": conf.paths.get("model_checkpoint"),
},
"results_averaged": {
**{k: np.mean([out[k] for out in outputs]) for k in result_avg_keys},
},
"results_by_sample": {
**{k: [out[k] for out in outputs] for k in result_keys},
},
}
def _write_results(conf, results):
with open(f"{conf.paths.output_dir}/results.json", "w") as f:
json.dump(results, f, indent=2)
table = [
list(results["results_by_sample"].keys()),
*zip(*results["results_by_sample"].values()),
]
with open(f"{conf.paths.output_dir}/results.tsv", "w") as f:
_write_tsv(table, file=f)
def _write_results_final(results_list):
meta_common = {
key: _get_common_value(results_list, ("meta", key))
for key in results_list[0]["meta"]
if _is_common_value(results_list, ("meta", key))
}
meta_noncommon = {
key: [_get_value(results, ("meta", key)) for results in results_list]
for key in results_list[0]["meta"]
if not _is_common_value(results_list, ("meta", key))
}
results_averaged = ld_to_dl(
[results["results_averaged"] for results in results_list]
)
results = {
"name": _get_common_value(results_list, ("name",)),
"description": _get_common_value(results_list, ("description",)),
"meta": meta_common,
"meta_": meta_noncommon,
"results": results_averaged,
}
with open(f"{DEFAULT_PATHS_OUTPUT_DIR_ROOT}/results.json", "w") as f:
json.dump(results, f, indent=2)
def _write_tsv(rows, file):
for row in rows:
print("\t".join(f"{x}" for x in row), file=file)
def _is_common_value(ds, path):
value = _get_value(ds[0], path)
return all(_get_value(d, path) == value for d in ds)
def _get_common_value(ds, path):
value = _get_value(ds[0], path)
assert all(_get_value(d, path) == value for d in ds)
return value
def _get_value(d, path):
for key in path:
d = d[key]
return d
def _prepare_conf(conf):
if "source" not in conf.get("model", {}):
with open_dict(conf):
conf.model.source = DEFAULT_MODEL_SOURCE
if (conf.model.source == "config") and (
conf.get("paths", {}).get("model_checkpoint", None) is None
):
with open_dict(conf):
conf.paths.model_checkpoint = DEFAULT_PATHS_MODEL_CHECKPOINT
if "output_dir" not in conf.get("paths", {}):
with open_dict(conf):
conf.paths.output_dir = _get_output_dir(conf)
def _get_output_dir(conf):
source = conf.model.source
if conf.paths.get("output_dir") is not None:
return conf.paths.output_dir
if source == "config":
return (
f"{DEFAULT_PATHS_OUTPUT_DIR_ROOT}/"
"${model.source}-${env.aim.run_hash}-${model.name}"
)
if source == "from_state_dict":
return (
f"{DEFAULT_PATHS_OUTPUT_DIR_ROOT}/"
"${model.name}-"
f"{Path(conf.paths.model_checkpoint).stem}"
)
if source == "zoo":
return (
f"{DEFAULT_PATHS_OUTPUT_DIR_ROOT}/"
"${model.source}-${model.name}-${model.metric}-${model.quality}"
)
raise ValueError(f"Unknown model.source: {source}")
[docs]def main():
results_list = []
dfs = []
argv = [
# Prepend default overrides.
"++misc.cudnn.deterministic=True",
"++paths.checkpoint=null",
"++paths.model_checkpoint=null",
*sys.argv[1:],
]
for conf in iter_configs(argv=argv, start=thisdir):
_prepare_conf(conf)
runner = setup(conf)
batches = runner.loaders["infer"]
filenames = _get_filenames(runner, len(batches))
output_dir = Path(conf.paths.output_dir)
metrics = [x for x in conf.runner.meters.infer if x in _METRICS]
outputs = run_eval_model(runner, batches, filenames, output_dir, metrics)
results = _results_dict(conf, outputs)
results_list.append(results)
_write_results(conf, results)
runner.epoch_step = None
runner.loader_metrics = results["results_averaged"]
runner._loader_metrics = results["results_by_sample"]
dfs.append(runner._current_dataframe)
_plot_rd(runner, results)
_write_results_final(results_list)
_plot_rd_all(runner, dfs)
if __name__ == "__main__":
main()
