repo
stringlengths
1
99
file
stringlengths
13
215
code
stringlengths
12
59.2M
file_length
int64
12
59.2M
avg_line_length
float64
3.82
1.48M
max_line_length
int64
12
2.51M
extension_type
stringclasses
1 value
nnsvs
nnsvs-master/tests/test_model.py
import torch from nnsvs.base import PredictionType from nnsvs.model import ( FFN, LSTMRNN, LSTMRNNSAR, MDN, RMDN, Conv1dResnet, Conv1dResnetMDN, Conv1dResnetSAR, FFConvLSTM, MDNv2, VariancePredictor, ) from nnsvs.util import init_seed def test_deprecated_imports(): from nnsvs.model import ResF0Conv1dResnet # noqa: F401 from nnsvs.model import ResF0Conv1dResnetMDN # noqa: F401 from nnsvs.model import ResF0VariancePredictor # noqa: F401 from nnsvs.model import ResSkipF0FFConvLSTM # noqa: F401 def _test_model_impl(model, in_dim, out_dim): B = 4 T = 100 init_seed(B * T) x = torch.rand(B, T, in_dim) lengths = torch.Tensor([T] * B).long() # warmup forward pass with torch.no_grad(): y = model(x, lengths) y_inf = model.inference(x, lengths) # MDN case if model.prediction_type() == PredictionType.PROBABILISTIC: log_pi, log_sigma, mu = y num_gaussian = log_pi.shape[2] assert mu.shape == (B, T, num_gaussian, out_dim) assert log_sigma.shape == (B, T, num_gaussian, out_dim) # NOTE: infernece output shouldn't have num_gaussian axis mu_inf, sigma_inf = y_inf assert mu_inf.shape == (B, T, out_dim) assert sigma_inf.shape == (B, T, out_dim) else: assert y.shape == (B, T, out_dim) assert y.shape == y_inf.shape def test_ffn(): params = { "in_dim": 300, "hidden_dim": 8, "out_dim": 20, "num_layers": 2, "dropout": 0.1, "init_type": "none", } model = FFN(**params) assert model.prediction_type() == PredictionType.DETERMINISTIC _test_model_impl(model, params["in_dim"], params["out_dim"]) def test_lstmrnn(): params = { "in_dim": 300, "hidden_dim": 8, "out_dim": 20, "num_layers": 2, "dropout": 0.1, "init_type": "none", } model = LSTMRNN(**params) assert model.prediction_type() == PredictionType.DETERMINISTIC _test_model_impl(model, params["in_dim"], params["out_dim"]) def test_conv1d_resnet(): params = { "in_dim": 300, "hidden_dim": 8, "out_dim": 180, "num_layers": 2, "num_gaussians": 2, "dim_wise": True, "init_type": "none", } model = Conv1dResnet(**{**params, "use_mdn": False}) assert model.prediction_type() == PredictionType.DETERMINISTIC _test_model_impl(model, params["in_dim"], params["out_dim"]) model = Conv1dResnet(**{**params, "use_mdn": True}) assert model.prediction_type() == PredictionType.PROBABILISTIC _test_model_impl(model, params["in_dim"], params["out_dim"]) # Deprecated model = Conv1dResnetMDN(**params) assert model.prediction_type() == PredictionType.PROBABILISTIC _test_model_impl(model, params["in_dim"], params["out_dim"]) def test_conv1d_resnet_sar(): params = { "in_dim": 300, "hidden_dim": 8, "out_dim": 20, "num_layers": 2, "stream_sizes": [10, 10], "ar_orders": [2, 2], "init_type": "none", } model = Conv1dResnetSAR(**params) assert model.prediction_type() == PredictionType.DETERMINISTIC _test_model_impl(model, params["in_dim"], params["out_dim"]) y = torch.rand(8, 100, params["out_dim"]) y_hat = model.preprocess_target(y) assert y.shape == y_hat.shape def test_lstmrnn_sar(): params = { "in_dim": 300, "hidden_dim": 8, "out_dim": 20, "num_layers": 2, "dropout": 0.1, "stream_sizes": [10, 10], "ar_orders": [2, 2], "init_type": "none", } model = LSTMRNNSAR(**params) assert model.prediction_type() == PredictionType.DETERMINISTIC _test_model_impl(model, params["in_dim"], params["out_dim"]) y = torch.rand(8, 100, params["out_dim"]) y_hat = model.preprocess_target(y) assert y.shape == y_hat.shape def test_mdn(): params = { "in_dim": 300, "hidden_dim": 8, "out_dim": 180, "num_layers": 2, "num_gaussians": 2, "dim_wise": True, "init_type": "none", } model = MDN(**params) assert model.prediction_type() == PredictionType.PROBABILISTIC _test_model_impl(model, params["in_dim"], params["out_dim"]) def test_mdnv2(): params = { "in_dim": 300, "hidden_dim": 8, "out_dim": 180, "num_layers": 2, "dropout": 0.5, "num_gaussians": 2, "dim_wise": True, "init_type": "none", } model = MDNv2(**params) assert model.prediction_type() == PredictionType.PROBABILISTIC _test_model_impl(model, params["in_dim"], params["out_dim"]) def test_rmdn(): params = { "in_dim": 300, "hidden_dim": 8, "out_dim": 180, "num_layers": 2, "num_gaussians": 2, "dim_wise": True, "init_type": "none", } model = RMDN(**params) assert model.prediction_type() == PredictionType.PROBABILISTIC _test_model_impl(model, params["in_dim"], params["out_dim"]) def test_ff_conv_lstm(): params = { "in_dim": 300, "ff_hidden_dim": 8, "conv_hidden_dim": 8, "lstm_hidden_dim": 8, "dropout": 0.1, "num_lstm_layers": 2, "bidirectional": True, "out_dim": 180, "init_type": "none", } model = FFConvLSTM(**params) assert model.prediction_type() == PredictionType.DETERMINISTIC _test_model_impl(model, params["in_dim"], params["out_dim"]) def test_variance_predictor(): params = { "in_dim": 300, "out_dim": 180, "num_layers": 2, "hidden_dim": 8, "kernel_size": 5, "dropout": 0.5, "init_type": "none", } model = VariancePredictor(**{**params, "use_mdn": False}) assert model.prediction_type() == PredictionType.DETERMINISTIC _test_model_impl(model, params["in_dim"], params["out_dim"]) model = VariancePredictor(**{**params, "use_mdn": True}) assert model.prediction_type() == PredictionType.PROBABILISTIC _test_model_impl(model, params["in_dim"], params["out_dim"])
6,229
26.688889
66
py
nnsvs
nnsvs-master/tests/test_model_configs.py
from pathlib import Path import hydra import nnsvs.bin.train import nnsvs.bin.train_acoustic import nnsvs.bin.train_postfilter import pytest import torch from nnsvs.util import init_seed from omegaconf import OmegaConf from .util import _test_model_impl RECIPE_DIR = Path(__file__).parent.parent / "recipes" def _test_postfilter_impl(model, model_config): B = 4 T = 100 init_seed(B * T) in_dim = sum(model_config.stream_sizes) x = torch.rand(B, T, in_dim) lengths = torch.Tensor([T] * B).long() # warmup forward pass with torch.no_grad(): y = model(x, lengths) y_inf = model.inference(x, lengths) assert x.shape == y.shape assert y_inf.shape == y.shape @pytest.mark.parametrize( "model_config", (Path(nnsvs.bin.train.__file__).parent / "conf" / "train" / "model").glob("*.yaml"), ) def test_model_config(model_config): model_config = OmegaConf.load(model_config) model = hydra.utils.instantiate(model_config.netG) _test_model_impl(model, model_config.netG.in_dim, model_config.netG.out_dim) @pytest.mark.parametrize( "model_config", ( Path(nnsvs.bin.train_acoustic.__file__).parent / "conf" / "train_acoustic" / "model" ).glob("*.yaml"), ) def test_acoustic_model_config(model_config): print(model_config) model_config = OmegaConf.load(model_config) # Dummy model_config.netG.in_lf0_idx = 10 model_config.netG.in_lf0_min = 5.3936276 model_config.netG.in_lf0_max = 6.491111 model_config.netG.out_lf0_idx = 60 model_config.netG.out_lf0_mean = 5.953093881972361 model_config.netG.out_lf0_scale = 0.23435173188961034 model = hydra.utils.instantiate(model_config.netG) _test_model_impl(model, model_config.netG.in_dim, model_config.netG.out_dim) @pytest.mark.parametrize( "model_config", ( Path(nnsvs.bin.train_postfilter.__file__).parent / "conf" / "train_postfilter" / "model" ).glob("*.yaml"), ) def test_postfilter_model_config(model_config): model_config = OmegaConf.load(model_config) if "stream_sizes" in model_config.netG: model_config.netG.stream_sizes = model_config.stream_sizes # Post-filter config should have netD hydra.utils.instantiate(model_config.netD) model = hydra.utils.instantiate(model_config.netG) _test_postfilter_impl(model, model_config) @pytest.mark.parametrize( "model_config", RECIPE_DIR.glob("**/_common/conf/**/train/timelag/model/*.yaml") ) def test_timelag_model_config_recipes(model_config): model_config = OmegaConf.load(model_config) model = hydra.utils.instantiate(model_config.netG) _test_model_impl(model, model_config.netG.in_dim, model_config.netG.out_dim) @pytest.mark.parametrize( "model_config", RECIPE_DIR.glob("**/_common/conf/**/train/duration/model/*.yaml") ) def test_duration_model_config_recipes(model_config): model_config = OmegaConf.load(model_config) model = hydra.utils.instantiate(model_config.netG) _test_model_impl(model, model_config.netG.in_dim, model_config.netG.out_dim) @pytest.mark.parametrize( "model_config", RECIPE_DIR.glob("**/_common/conf/**/train_acoustic/model/*.yaml") ) def test_train_acoustic_model_config_recipes(model_config): print(model_config) model_config = OmegaConf.load(model_config) # Dummy model_config.netG.in_lf0_idx = 10 model_config.netG.in_lf0_min = 5.3936276 model_config.netG.in_lf0_max = 6.491111 model_config.netG.out_lf0_idx = 60 model_config.netG.out_lf0_mean = 5.953093881972361 model_config.netG.out_lf0_scale = 0.23435173188961034 # NOTE: phoneme embedding impl requires careful specifications of # in_ph_start_idx and in_ph_end_idx, which is diffifult to set property # without hed files. So we skip phoneme embedding impl for now. if "embed_dim" in model_config.netG: model_config.netG.embed_dim = None keys = ["lf0_model", "mgc_model", "bap_model", "vuv_model", "encoder", "mel_model"] for key in keys: if key in model_config.netG and "embed_dim" in model_config.netG[key]: model_config.netG[key].embed_dim = None # For GaussianDiffusion that includes encoder in its implementation if ( key in model_config.netG and "encoder" in model_config.netG[key] and "embed_dim" in model_config.netG[key]["encoder"] ): model_config.netG[key]["encoder"].embed_dim = None model = hydra.utils.instantiate(model_config.netG) _test_model_impl(model, model_config.netG.in_dim, model_config.netG.out_dim) @pytest.mark.parametrize( "model_config", RECIPE_DIR.glob("**/icassp2023-24k-world/conf/train_acoustic/model/*.yaml"), ) def test_train_acoustic_model_config_recipes_icassp2023(model_config): model_config = OmegaConf.load(model_config) # Dummy model_config.netG.in_lf0_idx = 10 model_config.netG.in_lf0_min = 5.3936276 model_config.netG.in_lf0_max = 6.491111 model_config.netG.out_lf0_idx = 60 model_config.netG.out_lf0_mean = 5.953093881972361 model_config.netG.out_lf0_scale = 0.23435173188961034 # NOTE: phoneme embedding impl requires careful specifications of # in_ph_start_idx and in_ph_end_idx, which is diffifult to set property # without hed files. So we skip phoneme embedding impl for now. if "embed_dim" in model_config.netG: model_config.netG.embed_dim = None keys = ["lf0_model", "mgc_model", "bap_model", "vuv_model", "encoder", "mel_model"] for key in keys: if key in model_config.netG and "embed_dim" in model_config.netG[key]: model_config.netG[key].embed_dim = None model = hydra.utils.instantiate(model_config.netG) _test_model_impl(model, model_config.netG.in_dim, model_config.netG.out_dim) @pytest.mark.parametrize( "model_config", RECIPE_DIR.glob("**/_common/conf/**/train_postfilter/model/*.yaml") ) def test_postfilter_config_recipes(model_config): model_config = OmegaConf.load(model_config) if "stream_sizes" in model_config.netG: model_config.netG.stream_sizes = model_config.stream_sizes # Post-filter config should have netD hydra.utils.instantiate(model_config.netD) model = hydra.utils.instantiate(model_config.netG) _test_postfilter_impl(model, model_config)
6,395
33.203209
88
py
nnsvs
nnsvs-master/docs/conf.py
# -*- coding: utf-8 -*- # # Configuration file for the Sphinx documentation builder. # # This file does only contain a selection of the most common options. For a # full list see the documentation: # http://www.sphinx-doc.org/en/master/config # -- Path setup -------------------------------------------------------------- # If extensions (or modules to document with autodoc) are in another directory, # add these directories to sys.path here. If the directory is relative to the # documentation root, use os.path.abspath to make it absolute, like shown here. # # import os # import sys # sys.path.insert(0, os.path.abspath('.')) # -- Project information ----------------------------------------------------- import os import pkg_resources __version__ = pkg_resources.get_distribution("nnsvs").version ON_RTD = os.environ.get("READTHEDOCS", None) == "True" project = "nnsvs" copyright = "2020, Ryuichi Yamamoto" author = "Ryuichi Yamamoto" # The short X.Y version version = __version__ # The full version, including alpha/beta/rc tags release = __version__ # -- General configuration --------------------------------------------------- # If your documentation needs a minimal Sphinx version, state it here. # # needs_sphinx = '1.0' # Add any Sphinx extension module names here, as strings. They can be # extensions coming with Sphinx (named 'sphinx.ext.*') or your custom # ones. extensions = [ "sphinx.ext.autodoc", "sphinx.ext.autosummary", "sphinx.ext.doctest", "sphinx.ext.intersphinx", "sphinx.ext.mathjax", "sphinx.ext.todo", "sphinx.ext.viewcode", "sphinx.ext.githubpages", "sphinx.ext.napoleon", "nbsphinx", "matplotlib.sphinxext.plot_directive", "IPython.sphinxext.ipython_console_highlighting", "IPython.sphinxext.ipython_directive", "sphinxcontrib.bibtex", "sphinxcontrib.youtube", ] if ON_RTD: # Remove extensions not currently supported on RTD extensions.remove("matplotlib.sphinxext.plot_directive") bibtex_bibfiles = ["refs.bib"] autodoc_member_order = "bysource" autosummary_generate = True numpydoc_show_class_members = False # ------------------------------------------------------------------------------ # Plot # ------------------------------------------------------------------------------ doctest_global_setup = """ import numpy as np import scipy import librosa np.random.seed(123) np.set_printoptions(precision=3, linewidth=64, edgeitems=2, threshold=200) """ plot_pre_code = ( doctest_global_setup + """ import matplotlib import librosa import librosa.display matplotlib.rcParams['figure.constrained_layout.use'] = librosa.__version__ >= '0.8' """ ) plot_include_source = True plot_html_show_source_link = False plot_formats = [("png", 100), ("pdf", 100)] plot_html_show_formats = False plot_rcparams = {} # Add any paths that contain templates here, relative to this directory. templates_path = ["_templates"] # List of patterns, relative to source directory, that match files and # directories to ignore when looking for source files. # This pattern also affects html_static_path and html_extra_path. exclude_patterns = ["_build", "Thumbs.db", ".DS_Store"] # The name of the Pygments (syntax highlighting) style to use. pygments_style = "default" # -- Options for HTML output ------------------------------------------------- # The theme to use for HTML and HTML Help pages. See the documentation for # a list of builtin themes. # html_theme = "sphinx_rtd_theme" # Add any paths that contain custom static files (such as style sheets) here, # relative to this directory. They are copied after the builtin static files, # so a file named "default.css" will overwrite the builtin "default.css". html_static_path = ["_static"] html_css_files = [] # -- Options for intersphinx extension --------------------------------------- # Example configuration for intersphinx: refer to the Python standard library. intersphinx_mapping = { "python": ("https://docs.python.org/dev", None), "scipy": ("https://docs.scipy.org/doc/scipy/reference", None), "matplotlib": ("https://matplotlib.org/stable", None), "pytest": ("https://docs.pytest.org/en/stable", None), } # -- Options for todo extension ---------------------------------------------- # If true, `todo` and `todoList` produce output, else they produce nothing. todo_include_todos = False
4,386
29.255172
83
py
nnsvs
nnsvs-master/neutrino_compat/server.py
"""Web server implementation for singing voice synthesis NOTE: validation is not implemented. Expect 500 errors for unexpected inputs. """ import tarfile from os import listdir, rmdir from pathlib import Path from shutil import move import numpy as np import pyworld import torch from fastapi import FastAPI, UploadFile from nnmnkwii.io import hts from nnsvs.io.hts import full_to_mono from nnsvs.svs import NEUTRINO from omegaconf import OmegaConf from scipy.io import wavfile from starlette.responses import RedirectResponse, StreamingResponse from utaupy.utils import ust2hts SCORE_DIR = Path("./score") MUSICXML_DIR = SCORE_DIR / "musicxml" FULL_LAB_DIR = SCORE_DIR / "label" / "full" MONO_LAB_DIR = SCORE_DIR / "label" / "mono" UST_DIR = SCORE_DIR / "ust" TIMING_LAB_DIR = SCORE_DIR / "label" / "timing" OUTPUT_DIR = Path("./output") MODEL_DIR = Path("./model") for d in [ SCORE_DIR, MUSICXML_DIR, FULL_LAB_DIR, MONO_LAB_DIR, UST_DIR, TIMING_LAB_DIR, OUTPUT_DIR, MODEL_DIR, ]: d.mkdir(exist_ok=True, parents=True) app = FastAPI() _models = {} def _instantiate_model(model_id): global _models if model_id in _models: return _models[model_id] model = NEUTRINO( MODEL_DIR / model_id, device="cuda" if torch.cuda.is_available() else "cpu", verbose=100, ) _models[model_id] = model return model def _finalize(): if torch.cuda.is_available(): torch.cuda.empty_cache() @app.get("/healthcheck") def perform_healthcheck(): return {"healthcheck": "OK"} @app.get("/") async def docs_redirect(): return RedirectResponse(url="/docs") @app.get("/models/list") async def model_list(): model_ids = listdir(MODEL_DIR) return {"model_ids": model_ids} @app.get("/models/{model_id}") async def model_info(model_id: str): model = NEUTRINO( MODEL_DIR / model_id, device="cuda" if torch.cuda.is_available() else "cpu", verbose=100, ) return { "config": OmegaConf.to_container(model.config), "repr": repr(model), } @app.post("/models/create") def create_model(model: UploadFile, model_id: str): filename = Path(model.filename) if str(filename).endswith(".tar.gz"): model_dir = MODEL_DIR / model_id model_dir.mkdir(exist_ok=True) with tarfile.open(fileobj=model.file, mode="r|gz") as f: f.extractall(path=model_dir) tar_dir_path = model_dir / filename.name.replace(".tar.gz", "") # Move all contents to model_dir / model_id if (tar_dir_path).exists(): for name in listdir(tar_dir_path): move(tar_dir_path / name, model_dir / name) rmdir(tar_dir_path) else: raise NotImplementedError() return {"filename": model.filename} @app.post("/score/full/upload") async def upload_full_lab(full_lab: UploadFile): with open(f"{FULL_LAB_DIR}/{full_lab.filename}", "wb") as f: f.write(full_lab.file.read()) return {"filename": full_lab.filename} @app.post("/score/timing/upload") async def upload_timing_lab(timing_lab: UploadFile): with open(f"{TIMING_LAB_DIR}/{timing_lab.filename}", "wb") as f: f.write(timing_lab.file.read()) return {"filename": timing_lab.filename} @app.post("/score/musicxml/upload") async def upload_musicxml(musicxml: UploadFile): filename = musicxml.filename musicxml_path = MUSICXML_DIR / filename with open(musicxml_path, "wb") as f: f.write(musicxml.file.read()) full_labels, mono_labels = NEUTRINO.musicxml2label(str(musicxml_path)) full_lab_path = FULL_LAB_DIR / ( filename.replace(".musicxml", "").replace(".xml", "") + ".lab" ) mono_lab_path = MONO_LAB_DIR / ( filename.replace(".musicxml", "").replace(".xml", "") + ".lab" ) with open(full_lab_path, "w") as f: f.write(str(full_labels)) with open(mono_lab_path, "w") as f: f.write(str(mono_labels)) return {"filename": filename} @app.post("/score/ust/upload") async def upload_ust(ust: UploadFile, model_id: str): ust_path = UST_DIR / ust.filename with open(ust_path, "wb") as f: f.write(ust.file.read()) model_dir = MODEL_DIR / model_id table_path = model_dir / "kana2phonemes.table" assert table_path.exists() full_lab = FULL_LAB_DIR / ust.filename.replace(".ust", ".lab") ust2hts( str(ust_path), full_lab, table_path, strict_sinsy_style=False, as_mono=False ) return {"filename": ust.filename} @app.get("/run/timing") async def run_timing(name: str, model_id: str): model = _instantiate_model(model_id) model.set_device("cuda" if torch.cuda.is_available() else "cpu") full_lab = FULL_LAB_DIR / (name + ".lab") assert full_lab.exists(), f"{full_lab} does not exist" full_labels = hts.load(full_lab) timing_labels = full_to_mono(model.predict_timing(full_labels)) # TODO: Do we want to save timing for each model? # timing_lab = TIMING_LAB_DIR / model_id / (name + ".lab") timing_lab = TIMING_LAB_DIR / (name + ".lab") with open(timing_lab, "w") as f: f.write(str(timing_labels)) _finalize() return {"timing": str(timing_labels)} @app.get("/run/phrases") async def run_phrases(name: str, model_id: str): model = _instantiate_model(model_id) model.set_device("cuda" if torch.cuda.is_available() else "cpu") full_lab = FULL_LAB_DIR / (name + ".lab") assert full_lab.exists(), f"{full_lab} does not exist" full_labels = hts.load(full_lab) timing_lab = TIMING_LAB_DIR / (name + ".lab") assert timing_lab.exists(), "Timing labels not found. " timing_labels = hts.load(timing_lab) model_output_dir = OUTPUT_DIR / model_id model_output_dir.mkdir(exist_ok=True) # Dump phraselist phraselist_path = model_output_dir / (name + "-phraselist.txt") phraselist = model.get_phraselist(full_labels, timing_labels) with open(phraselist_path, "w") as f: f.write(str(phraselist)) # Dump num_phrases for convenience num_phrases = model.get_num_phrases(full_labels) return {"phraselist": phraselist, "num_phrases": num_phrases} @app.get("/run/acoustic") async def run_acoustic(name: str, model_id: str, phrase_num: int = -1): model = _instantiate_model(model_id) model.set_device("cuda" if torch.cuda.is_available() else "cpu") full_lab = FULL_LAB_DIR / (name + ".lab") assert full_lab.exists(), f"{full_lab} does not exist" full_labels = hts.load(full_lab) timing_lab = TIMING_LAB_DIR / (name + ".lab") assert timing_lab.exists(), "Timing labels not found. " timing_labels = hts.load(timing_lab) model_output_dir = OUTPUT_DIR / model_id model_output_dir.mkdir(exist_ok=True) f0, mgc, bap = model.predict_acoustic( full_labels, timing_labels, phrase_num=phrase_num, ) _finalize() if phrase_num > 0: name = f"{name}-{phrase_num}" f0.tofile(model_output_dir / (name + ".f0")) mgc.tofile(model_output_dir / (name + ".mgc")) bap.tofile(model_output_dir / (name + ".bap")) # NOTE: pack into a single file for convenience feats = np.concatenate([f0, mgc, bap], axis=1).astype(np.float64) path = model_output_dir / (name + ".bin") feats.tofile(path) def iterfile(): with open(path, mode="rb") as file_like: yield from file_like return StreamingResponse(iterfile()) @app.get("/run/vocoder") async def run_vocoder( name: str, model_id: str, vocoder_type: str = "world", phrase_num: int = -1, loudness_norm: bool = False, dtype: str = "int16", ): model = _instantiate_model(model_id) model.set_device("cuda" if torch.cuda.is_available() else "cpu") if phrase_num > 0: name = f"{name}-{phrase_num}" f0_path = OUTPUT_DIR / model_id / (name + ".f0") mgc_path = OUTPUT_DIR / model_id / (name + ".mgc") bap_path = OUTPUT_DIR / model_id / (name + ".bap") f0 = np.fromfile(f0_path, dtype=np.float64).reshape(-1, 1) mgc = np.fromfile(mgc_path, dtype=np.float64).reshape(-1, 60) bap = np.fromfile(bap_path, dtype=np.float64).reshape( -1, pyworld.get_num_aperiodicities(model.sample_rate) ) wav = model.predict_waveform( f0, mgc, bap, vocoder_type=vocoder_type, loudness_norm=loudness_norm, dtype=dtype, ) _finalize() if vocoder_type == "world": suffix = "_syn.wav" else: suffix = "_nsf.wav" wav_path = OUTPUT_DIR / model_id / (name + suffix) wavfile.write(wav_path, model.sample_rate, wav) path = OUTPUT_DIR / model_id / (name + suffix.replace(".wav", ".raw")) wav.tofile(path) def iterfile(): with open(path, mode="rb") as file_like: yield from file_like return StreamingResponse(iterfile())
8,930
27.352381
84
py
nnsvs
nnsvs-master/neutrino_compat/bin/NEUTRINO.py
"""Predict acoustic features by NNSVS with NEUTRINO-compatible file IO NOTE: options are not yet fully implemented NEUTRINO - NEURAL SINGING SYNTHESIZER (Electron v1.2.0-Stable) Copyright (c) 2020-2022 STUDIO NEUTRINO All rights reserved. usage: NEUTRINO full.lab timing.lab output.f0 output.mgc output.bap model_directory [option] options : description [default] -n i : number of threads (CPU) [MAX] -k i : style shift [ 0] -s : skip timing prediction [off] -a : skip acoustic features prediction [off] -p i : single phrase prediction [ -1] -i filename : trace phrase information [off] -t : view information [off] """ import argparse import logging import sys import tempfile import time from pathlib import Path import numpy as np import requests def get_parser(): parser = argparse.ArgumentParser( description="Pretend as if the script is NEUTRINO", ) parser.add_argument("input_file", type=str, help="Input file (.ust or .lab)") parser.add_argument("timing_lab", type=str, help="Path of timing labels") parser.add_argument("output_f0", type=str, help="Path of output F0") parser.add_argument("output_mgc", type=str, help="Path of output MGC") parser.add_argument("output_bap", type=str, help="Path of output BAP") parser.add_argument("model_dir", type=str, help="model_dir") parser.add_argument( "-i", "--phraselist", type=str, default=None, help="Path of phraselist" ), parser.add_argument( "-p", "--phrase_num", type=int, default=-1, help="Phrase number" ) parser.add_argument("--use_api", action="store_true", help="Use web API") parser.add_argument( "--url", type=str, default="http://127.0.0.1:8001", help="URL of the server" ) return parser def run_local(args, _): import torch from nnmnkwii.io import hts from nnsvs.io.hts import full_to_mono from nnsvs.svs import NEUTRINO from utaupy.utils import ust2hts model_dir = Path(args.model_dir) # NOTE: this is needed to be compatible with NEUTRINO's Run.bat if not model_dir.exists(): model_dir = "model" / model_dir engine = NEUTRINO( model_dir, device="cuda" if torch.cuda.is_available() else "cpu", verbose=100 ) input_file = Path(args.input_file) assert input_file.exists() if input_file.suffix == ".ust": table_path = model_dir / "kana2phonemes.table" assert table_path.exists() with tempfile.NamedTemporaryFile(suffix=".lab") as tf: ust2hts( str(input_file), tf.name, table_path, strict_sinsy_style=False, as_mono=False, ) full_labels = hts.HTSLabelFile() with open(tf.name) as f: for label in f.readlines(): full_labels.append(label.split(), strict=False) elif input_file.suffix == ".lab": full_labels = hts.load(input_file) else: raise ValueError(f"Not supported file type: {input_file.suffix}") timing_lab = Path(args.timing_lab) if not timing_lab.exists(): timing_labels = full_to_mono(engine.predict_timing(full_labels)) with open(timing_lab, "w") as f: f.write(str(timing_labels)) else: timing_labels = hts.load(timing_lab) if args.phraselist is not None: phraselist = Path(args.phraselist) if not phraselist.exists(): phraselist_str = engine.get_phraselist(full_labels, timing_labels) with open(phraselist, "w") as f: f.write(phraselist_str) num_phrases = engine.get_num_phrases(full_labels) if args.phrase_num < 0 or args.phrase_num >= num_phrases: raise ValueError(f"phrase_num must be in [0, {num_phrases - 1}]") f0, mgc, bap = engine.predict_acoustic( full_labels, timing_labels, phrase_num=args.phrase_num ) return f0, mgc, bap def run_api(args, logger): input_file = Path(args.input_file) url = args.url[:-1] if args.url[-1] == "/" else args.url name = input_file.stem # pretend as if the model_dir is model_id model_id = args.model_dir # Upload full-context labels or UST if input_file.suffix == ".ust": logger.info(f"Uploading UST: {input_file}") res = requests.post( url + "/score/ust/upload", params={ "model_id": model_id, }, files={ "ust": open(input_file, "rb"), }, ) elif input_file.suffix == ".lab": logger.info(f"Uploading full_lab: {input_file}") res = requests.post( url + "/score/full/upload", files={ "full_lab": open(input_file, "rb"), }, ) if res.status_code != 200: raise RuntimeError(f"Failed to upload file: {res.status_code}") # Upload (possibly modified) timing labels if present if Path(args.timing_lab).exists(): logger.info(f"Uploading timing_lab: {args.timing_lab}") res = requests.post( url + "/score/timing/upload", files={ "timing_lab": open(args.timing_lab, "rb"), }, ) if res.status_code != 200: raise RuntimeError(f"Failed to upload file: {res.status_code}") else: # Predict timing logger.info("Predicting timing") res = requests.get( url + "/run/timing", params={ "name": name, "model_id": model_id, }, ) if res.status_code != 200: raise RuntimeError(f"Failed to predict timing: {res.status_code}") timing_str = res.json()["timing"] logger.info(timing_str) with open(args.timing_lab, "w") as f: f.write(timing_str) # Phraselist if args.phraselist is not None: logger.info("Predicting phraselist") res = requests.get( url + "/run/phrases", params={ "name": name, "model_id": model_id, }, ) if res.status_code != 200: raise RuntimeError(f"Failed to predict phraselist: {res.status_code}") phraselist_str = res.json()["phraselist"] logger.info(phraselist_str) with open(args.phraselist, "w") as f: f.write(phraselist_str) # Predict acoustic features logger.info("Predicting acoustic features") res = requests.get( url + "/run/acoustic", params={ "name": name, "model_id": model_id, "phrase_num": args.phrase_num, }, ) if res.status_code != 200: raise RuntimeError(f"Failed to predict acoustic features: {res.status_code}") feats = np.frombuffer(res.content, dtype=np.float64).reshape(-1, 66) f0 = feats[:, :1] mgc = feats[:, 1:61] bap = feats[:, 61:] logger.info(f"f0: {f0.shape}") logger.info(f"mgc: {mgc.shape}") logger.info(f"bap: {bap.shape}") return f0, mgc, bap def main(): args = get_parser().parse_args(sys.argv[1:]) start_time = time.time() format = "%(asctime)s (%(module)s:%(lineno)d) %(levelname)s: %(message)s" logger = logging.getLogger("NEUTRINO") logger.setLevel(logging.INFO) if logger.hasHandlers(): logger.handlers.clear() stream_handler = logging.StreamHandler() stream_handler.setFormatter(logging.Formatter(format)) logger.addHandler(stream_handler) if args.use_api: logger.info(f"Using webapi: {args.url} for infernce") f0, mgc, bap = run_api(args, logger) else: logger.info("Using local machine for inference") f0, mgc, bap = run_local(args, logger) # Save to file f0.tofile(args.output_f0) mgc.tofile(args.output_mgc) bap.tofile(args.output_bap) logger.info(f"Elapsed time: {time.time() - start_time:.2f} sec") if __name__ == "__main__": main()
8,240
32.5
89
py
nnsvs
nnsvs-master/neutrino_compat/bin/NSF.py
"""Predict waveform by neural vocoders with NEUTRINO-compatible file IO NOTE: options are not yet fully implemented NSF - Neural Source Filter (v1.2.0-Stable) Copyright (c) 2020-2022 STUDIO NEUTRINO All rights reserved. usage: NSF input.f0 input.mgc input.bap model_name output_wav [option] options : description [default] -s i : sampling rate (kHz) [ 48] -n i : number of parallel [ MAX] -p i : number of parallel in session [ 1] -l file name : multi phrase prediction [ none] -g : use gpu [ off] -i i : gpu id [ 0] -t : view information [ off] """ import argparse import logging import sys import time from pathlib import Path import numpy as np import requests import soundfile as sf def get_parser(): parser = argparse.ArgumentParser( description="Pretend as if the script is NSF of NEUTRINO", ) parser.add_argument("input_f0", type=str, help="Path of input F0") parser.add_argument("input_mgc", type=str, help="Path of input MGC") parser.add_argument("input_bap", type=str, help="Path of input BAP") parser.add_argument("model_dir", type=str, help="model_dir") parser.add_argument("output_wav", type=str, help="Path of output wav") parser.add_argument("--use_api", action="store_true", help="Use web API") parser.add_argument( "--url", type=str, default="http://127.0.0.1:8001", help="URL of the server" ) return parser def run_local(args, _): import pyworld import torch from nnsvs.svs import NEUTRINO model_dir = Path(args.model_dir) # NOTE: this is needed to be compatible with NEUTRINO's Run.bat if not model_dir.exists(): model_dir = "model" / model_dir engine = NEUTRINO(model_dir, device="cuda" if torch.cuda.is_available() else "cpu") f0 = np.fromfile(args.input_f0, dtype=np.float64).reshape(-1, 1) mgc = np.fromfile(args.input_mgc, dtype=np.float64).reshape(-1, 60) bap = np.fromfile(args.input_bap, dtype=np.float64).reshape( -1, pyworld.get_num_aperiodicities(engine.sample_rate) ) # NOTE: `auto` will run uSFGAN or PWG if a trained one is in the model_dir # and fallback to WORLD it doesn't exist. wav = engine.predict_waveform(f0, mgc, bap, vocoder_type="auto", dtype=np.int16) return wav, engine.sample_rate def run_api(args, logger): url = args.url[:-1] if args.url[-1] == "/" else args.url # NOTE: for webAPI, these local files are not used for inference # files on the remote server are used instead. input_f0 = Path(args.input_f0) input_mgc = Path(args.input_mgc) input_bap = Path(args.input_bap) assert input_f0.exists() assert input_mgc.exists() assert input_bap.exists() name = input_f0.stem # TODO: better way to handle phrase-based synthesis # At the moment, we need to tell the server which phrase we are synthesizing if "-" in name: # NOTE: this code does not work for complicated filenames phrase_num = int(name.split("-")[-1]) name = name.split("-")[0] else: phrase_num = -1 # pretend as if the model_dir is model_id model_id = args.model_dir # Get sampling rate res = requests.get( url + f"/models/{model_id}", ) if res.status_code != 200: raise RuntimeError(f"Failed to fetch model info: {res.status_code}") res = res.json() sample_rate = res["config"]["sample_rate"] # Run vocoder logger.info("Predicting waveform") dtype = "int16" res = requests.get( url + "/run/vocoder", params={ "name": name, "model_id": model_id, "phrase_num": phrase_num, "vocoder_type": "auto", "dtype": dtype, "loudness_norm": False, }, ) if res.status_code != 200: raise RuntimeError(f"Failed to generate waveform: {res.status_code}") wav = np.frombuffer(res.content, dtype=dtype).reshape(-1) return wav, sample_rate def main(): args = get_parser().parse_args(sys.argv[1:]) start_time = time.time() format = "%(asctime)s (%(module)s:%(lineno)d) %(levelname)s: %(message)s" logger = logging.getLogger("NSF") logger.setLevel(logging.INFO) if logger.hasHandlers(): logger.handlers.clear() stream_handler = logging.StreamHandler() stream_handler.setFormatter(logging.Formatter(format)) logger.addHandler(stream_handler) if args.use_api: logger.info(f"Using webapi: {args.url} for inference") wav, sr = run_api(args, logger) else: logger.info("Using local machine for inference") wav, sr = run_local(args, logger) sf.write(args.output_wav, wav, sr) logger.info(f"Elapsed time: {time.time() - start_time:.2f} sec") if __name__ == "__main__": main()
5,014
31.993421
87
py
nnsvs
nnsvs-master/utils/enunu2nnsvs.py
"""Convert ENUNU's packed model to NNSVS's style """ import argparse import os import shutil import sys from pathlib import Path import joblib import numpy as np import torch from nnsvs.logger import getLogger from nnsvs.util import StandardScaler as NNSVSStandardScaler from omegaconf import OmegaConf from sklearn.preprocessing import MinMaxScaler, StandardScaler def get_parser(): parser = argparse.ArgumentParser( description="Convert ENUNU's packed model to NNSVS's style", ) parser.add_argument("enunu_dir", type=str, help="ENUNU's model dir") parser.add_argument("out_dir", type=str, help="Output dir") parser.add_argument("--verbose", type=int, default=100, help="Verbose level") return parser def _scaler2numpy(input_file, out_dir, logger): scaler = joblib.load(input_file) if isinstance(scaler, StandardScaler) or isinstance(scaler, NNSVSStandardScaler): logger.info(f"Converting {input_file} mean/scale npy files") mean_path = out_dir / (input_file.stem + "_mean.npy") scale_path = out_dir / (input_file.stem + "_scale.npy") var_path = out_dir / (input_file.stem + "_var.npy") np.save(mean_path, scaler.mean_, allow_pickle=False) np.save(scale_path, scaler.scale_, allow_pickle=False) np.save(var_path, scaler.var_, allow_pickle=False) elif isinstance(scaler, MinMaxScaler): logger.info(f"Converting {input_file} min/max npy files") min_path = out_dir / (input_file.stem + "_min.npy") scale_path = out_dir / (input_file.stem + "_scale.npy") np.save(min_path, scaler.min_, allow_pickle=False) np.save(scale_path, scaler.scale_, allow_pickle=False) else: raise ValueError(f"Unknown scaler type: {type(scaler)}") def _save_checkpoint(input_file, output_file, logger): checkpoint = torch.load( input_file, map_location=torch.device("cpu") # pylint: disable='no-member' ) size = os.path.getsize(input_file) logger.info(f"Processisng: {input_file}") logger.info(f"File size (before): {size / 1024/1024:.3f} MB") for k in ["optimizer_state", "lr_scheduler_state"]: if k in checkpoint.keys(): del checkpoint[k] # For https://github.com/kan-bayashi/ParallelWaveGAN for k in ["optimizer", "lr_scheduler"]: if k in checkpoint.keys(): del checkpoint[k] if "model" in checkpoint and "discriminator" in checkpoint["model"]: del checkpoint["model"]["discriminator"] torch.save(checkpoint, output_file) size = os.path.getsize(output_file) logger.info(f"File size (after): {size / 1024/1024:.3f} MB") def main(enunu_dir, out_dir, verbose=100): """Run the main function NOTE: This function is used by https://github.com/oatsu-gh/SimpleEnunu. So we need to be careful about the changes. It would be probably better to move this functionality to under the nnsvs directory. """ logger = getLogger(verbose=verbose) enunu_dir = Path(enunu_dir) out_dir = Path(out_dir) out_dir.mkdir(exist_ok=True, parents=True) enuconfig = OmegaConf.load(enunu_dir / "enuconfig.yaml") # Hed qst_path = enunu_dir / enuconfig.question_path shutil.copyfile(qst_path, out_dir / "qst.hed") # Table table_path = enunu_dir / enuconfig.table_path shutil.copyfile(table_path, out_dir / "kana2phonemes.table") # Models model_dir = enunu_dir / enuconfig.model_dir assert model_dir.exists() for typ in ["timelag", "duration", "acoustic"]: model_config = model_dir / typ / "model.yaml" assert model_config.exists() checkpoint = model_dir / typ / enuconfig[typ]["checkpoint"] assert checkpoint.exists() shutil.copyfile(model_config, out_dir / f"{typ}_model.yaml") _save_checkpoint(checkpoint, out_dir / f"{typ}_model.pth", logger) for inout in ["in", "out"]: scaler_path = ( enunu_dir / enuconfig.stats_dir / f"{inout}_{typ}_scaler.joblib" ) _scaler2numpy(scaler_path, out_dir, logger) # Config s = f"""# Global configs sample_rate: {enuconfig.sample_rate} frame_period: 5 log_f0_conditioning: {enuconfig.log_f0_conditioning} use_world_codec: false # Model-specific synthesis configs timelag: allowed_range: {enuconfig.timelag.allowed_range} allowed_range_rest: {enuconfig.timelag.allowed_range_rest} force_clip_input_features: true duration: force_clip_input_features: true acoustic: subphone_features: "coarse_coding" force_clip_input_features: true relative_f0: {enuconfig.acoustic.relative_f0} """ with open(out_dir / "config.yaml", "w") as f: f.write(s) logger.info(f"Contents of config.yaml: \n{s}") logger.warning( """Assuming `use_world_codec: false` since the most of released ENUNU models were trained with `use_world_codec: false`. If you use the default feature extarction settings in newer NNSVS (> 0.0.3), please set `use_world_codec: true` in the config.yaml `use_world_codec` must be the same during the feature extraction and synthesis time. """ ) if __name__ == "__main__": args = get_parser().parse_args(sys.argv[1:]) main(args.enunu_dir, args.out_dir, args.verbose)
5,291
34.756757
85
py
nnsvs
nnsvs-master/utils/merge_postfilters.py
import argparse import os import sys from pathlib import Path import torch from omegaconf import OmegaConf def get_parser(): parser = argparse.ArgumentParser( description="Merge post-filters", ) parser.add_argument("mgc_checkpoint", type=str, help="mgc checkpoint") parser.add_argument("bap_checkpoint", type=str, help="bap checkpoint") parser.add_argument("output_dir", type=str, help="out_dir") return parser if __name__ == "__main__": args = get_parser().parse_args(sys.argv[1:]) mgc_checkpoint = torch.load(args.mgc_checkpoint, map_location="cpu") bap_checkpoint = torch.load(args.bap_checkpoint, map_location="cpu") for path in [args.mgc_checkpoint, args.bap_checkpoint]: size = os.path.getsize(path) print("Processisng:", path) print(f"File size: {size / 1024/1024:.3f} MB") mgc_model = OmegaConf.load(Path(args.mgc_checkpoint).parent / "model.yaml") bap_model = OmegaConf.load(Path(args.bap_checkpoint).parent / "model.yaml") if "postfilters.MultistreamPostFilter" not in mgc_model.netG._target_: raise ValueError("Only MultistreamPostFilter is supported for now") checkpoint = mgc_checkpoint checkpoint["state_dict"].update(bap_checkpoint["state_dict"]) for k in ["optimizer_state", "lr_scheduler_state"]: if k in checkpoint.keys(): del checkpoint[k] Path(args.output_dir).mkdir(exist_ok=True, parents=True) # Model definition yaml_path = Path(args.output_dir) / "model.yaml" mgc_model.netG.bap_postfilter = bap_model.netG.bap_postfilter OmegaConf.save(mgc_model, yaml_path) # Checkpoint checkpoint_path = Path(args.output_dir) / "latest.pth" torch.save(checkpoint, checkpoint_path) size = os.path.getsize(checkpoint_path) print(f"File size (after): {size / 1024/1024:.3f} MB")
1,864
32.303571
79
py
nnsvs
nnsvs-master/nnsvs/dsp.py
import torch from scipy import signal from torch import nn from torch.nn import functional as F # Part of code was adapted from: # https://github.com/nii-yamagishilab/project-NN-Pytorch-scripts def lowpass_filter(x, fs, cutoff=5, N=5): """Lowpass filter Args: x (np.ndarray): input signal fs (int): sampling rate cutoff (int): cutoff frequency Returns: np.ndarray: filtered signal """ nyquist = fs // 2 norm_cutoff = cutoff / nyquist Wn = [norm_cutoff] b, a = signal.butter(N, Wn, "lowpass") if len(x) <= max(len(a), len(b)) * (N // 2 + 1): # NOTE: input signal is too short return x # NOTE: use zero-phase filter y = signal.filtfilt(b, a, x) return y def bandpass_filter(x, sr, cutoff=70): """Band-pass filter Args: x (np.ndarray): input signal fs (int): sampling rate cutoff (int): cutoff frequency Returns: np.ndarray: filtered signal """ nyquist = sr // 2 norm_cutoff = cutoff / nyquist Wn = [norm_cutoff, 0.999] b, a = signal.butter(5, Wn, "bandpass") y = signal.filtfilt(b, a, x) return y class TimeInvFIRFilter(nn.Conv1d): """Time-invatiant FIR filter implementation Args: channels (int): input channels filt_coef (torch.Tensor): FIR filter coefficients causal (bool): causal requires_grad (bool): trainable kernel or not """ def __init__(self, channels, filt_coef, causal=True, requires_grad=False): # assuming 1-D filter coef vector and odd num taps assert len(filt_coef.shape) == 1 # assert len(filt_coef) % 2 == 1 kernel_size = len(filt_coef) self.causal = causal if causal: padding = (kernel_size - 1) * 1 else: padding = (kernel_size - 1) // 2 * 1 # channel-wise filtering (groups=channels) super(TimeInvFIRFilter, self).__init__( channels, channels, kernel_size, padding=padding, groups=channels, bias=None ) self.weight.data[:, :, :] = filt_coef.flip(-1) self.weight.requires_grad = requires_grad def forward(self, x): out = super(TimeInvFIRFilter, self).forward(x) out = out[:, :, : -self.padding[0]] if self.causal else out return out class TrTimeInvFIRFilter(nn.Conv1d): """Trainable Time-invatiant FIR filter implementation H(z) = \\sigma_{k=0}^{filt_dim} b_{k}z_{-k} Note that b_{0} is fixed to 1 if fixed_0th is True. Args: channels (int): input channels filt_dim (int): FIR filter dimension causal (bool): causal tanh (bool): apply tanh to filter coef or not. fixed_0th (bool): fix the first filt coef to 1 or not. """ def __init__(self, channels, filt_dim, causal=True, tanh=True, fixed_0th=True): # Initialize filt coef with small random values init_filt_coef = torch.randn(filt_dim) * (1 / filt_dim) # assert len(filt_coef) % 2 == 1 kernel_size = len(init_filt_coef) self.causal = causal if causal: padding = (kernel_size - 1) * 1 else: padding = (kernel_size - 1) // 2 * 1 # channel-wise filtering (groups=channels) super(TrTimeInvFIRFilter, self).__init__( channels, channels, kernel_size, padding=padding, groups=channels, bias=None ) self.weight.data[:, :, :] = init_filt_coef.flip(-1) self.weight.requires_grad = True self.tanh = tanh self.fixed_0th = fixed_0th def get_filt_coefs(self): # apply tanh for filtter stability b = torch.tanh(self.weight) if self.tanh else self.weight b = b.clone() if self.fixed_0th: b[:, :, -1] = 1 return b def forward(self, x): b = self.get_filt_coefs() out = F.conv1d( x, b, self.bias, self.stride, self.padding, self.dilation, self.groups ) if self.padding[0] > 0: out = out[:, :, : -self.padding[0]] if self.causal else out return out
4,140
28.578571
88
py
nnsvs
nnsvs-master/nnsvs/base.py
from enum import Enum from torch import nn class PredictionType(Enum): """Prediction types""" DETERMINISTIC = 1 """Deterministic prediction Non-MDN single-stream models should use this type. Pseudo code: .. code-block:: # training y = model(x) # inference y = model.inference(x) """ PROBABILISTIC = 2 """Probabilistic prediction with mixture density networks MDN-based models should use this type. Pseudo code: .. code-block:: # training mdn_params = model(x) # inference mu, sigma = model.inference(x) """ MULTISTREAM_HYBRID = 3 """Multi-stream preodictions where each prediction can be detereministic or probabilistic Multi-stream models should use this type. Pseudo code: .. code-block:: # training feature_streams = model(x) # e.g. (mgc, lf0, vuv, bap) or (mel, lf0, vuv) # inference y = model.inference(x) Note that concatenated features are assumed to be returned during inference. """ DIFFUSION = 4 """Diffusion model's prediction NOTE: may subject to change in the future Pseudo code: .. code-block:: # training noise, x_recon = model(x) # inference y = model.inference(x) """ class BaseModel(nn.Module): """Base class for all models If you want to implement your custom model, you should inherit from this class. You must need to implement the forward method. Other methods are optional. """ def forward(self, x, lengths=None, y=None): """Forward pass Args: x (tensor): input features lengths (tensor): lengths of the input features y (tensor): optional target features Returns: tensor: output features """ pass def inference(self, x, lengths=None): """Inference method If you want to implement custom inference method such as autoregressive sampling, please override this method. Defaults to call the forward method. Args: x (tensor): input features lengths (tensor): lengths of the input features Returns: tensor: output features """ return self(x, lengths) def preprocess_target(self, y): """Preprocess target signals at training time This is useful for shallow AR models in which a FIR filter is used for the target signals. For other types of model, you don't need to implement this method. Defaults to do nothing. Args: y (tensor): target features Returns: tensor: preprocessed target features """ return y def prediction_type(self): """Prediction type. If your model has a MDN layer, please return ``PredictionType.PROBABILISTIC``. Returns: PredictionType: Determisitic or probabilistic. Default is deterministic. """ return PredictionType.DETERMINISTIC def is_autoregressive(self): """Is autoregressive or not If your custom model is an autoregressive model, please return ``True``. In that case, you would need to implement autoregressive sampling in :py:meth:`inference`. Returns: bool: True if autoregressive. Default is False. """ return False def has_residual_lf0_prediction(self): """Whether the model has residual log-F0 prediction or not. This should only be used for acoustic models. Returns: bool: True if the model has residual log-F0 prediction. Default is False. """ return False
3,784
22.955696
94
py
nnsvs
nnsvs-master/nnsvs/multistream.py
# Utils for multi-stream features import numpy as np import torch from nnmnkwii import paramgen def get_windows(num_window=1): """Get windows for MLPG. Args: num_window (int): number of windows Returns: list: list of windows """ windows = [(0, 0, np.array([1.0]))] if num_window >= 2: windows.append((1, 1, np.array([-0.5, 0.0, 0.5]))) if num_window >= 3: windows.append((1, 1, np.array([1.0, -2.0, 1.0]))) if num_window >= 4: raise ValueError(f"Not supported num windows: {num_window}") return windows def select_streams( inputs, stream_sizes=None, streams=None, concat=True, ): """Select streams from multi-stream features Args: inputs (array like): input 3-d or 2-d array stream_sizes (list): stream sizes streams (list): Streams of interests. Returns all streams if streams is None. concat (bool): Concatenate streams. Defaults to True. Returns: array like: selected streams """ if stream_sizes is None: stream_sizes = [60, 1, 1, 1] if streams is None: streams = [True] * len(stream_sizes) ret = [] start_indices = np.hstack(([0], np.cumsum(stream_sizes)[:-1])) for start_idx, size, enabled in zip(start_indices, stream_sizes, streams): if not enabled: continue if len(inputs.shape) == 3: s = inputs[:, :, start_idx : start_idx + size] else: s = inputs[:, start_idx : start_idx + size] ret.append(s) if not concat: return ret if isinstance(inputs, torch.Tensor): return torch.cat(ret, dim=-1) else: return np.concatenate(ret, -1) def split_streams(inputs, stream_sizes=None): """Split streams from multi-stream features Args: inputs (array like): input 3-d array stream_sizes (list): sizes for each stream Returns: list: list of stream features """ if stream_sizes is None: stream_sizes = [60, 1, 1, 1] ret = [] start_indices = np.hstack(([0], np.cumsum(stream_sizes)[:-1])) for start_idx, size in zip(start_indices, stream_sizes): if len(inputs.shape) == 3: s = inputs[:, :, start_idx : start_idx + size] else: s = inputs[:, start_idx : start_idx + size] ret.append(s) return ret def get_static_stream_sizes(stream_sizes, has_dynamic_features, num_windows): """Get stream sizes for static features Args: inputs (array like): input 3-d or 2-d array num_windows (int): number of windows stream_sizes (list): stream sizes has_dynamic_features (list): binary flags that indicates if steams have dynamic features streams (list, optional): Streams of interests. Returns all streams if streams is None. Defaults to None. Returns: list: stream sizes """ static_stream_sizes = np.array(stream_sizes) static_stream_sizes[has_dynamic_features] = ( static_stream_sizes[has_dynamic_features] / num_windows ) return static_stream_sizes def get_static_features( inputs, num_windows, stream_sizes=None, has_dynamic_features=None, streams=None, ): """Get static features from static+dynamic features Args: inputs (array like): input 3-d or 2-d array num_windows (int): number of windows stream_sizes (list): stream sizes has_dynamic_features (list): binary flags that indicates if steams have dynamic features streams (list, optional): Streams of interests. Returns all streams if streams is None. Defaults to None. Returns: list: list of static features """ if stream_sizes is None: stream_sizes = [180, 3, 1, 15] if has_dynamic_features is None: has_dynamic_features = [True, True, False, True] if streams is None: streams = [True] * len(stream_sizes) _, _, D = inputs.shape if stream_sizes is None or (len(stream_sizes) == 1 and has_dynamic_features[0]): return inputs[:, :, : D // num_windows] if len(stream_sizes) == 1 and not has_dynamic_features[0]: return inputs # Multi stream case ret = [] start_indices = np.hstack(([0], np.cumsum(stream_sizes)[:-1])) for start_idx, size, v, enabled in zip( start_indices, stream_sizes, has_dynamic_features, streams ): start_idx = int(start_idx) size = int(size) if not enabled: continue if v: static_features = inputs[:, :, start_idx : start_idx + size // num_windows] else: static_features = inputs[:, :, start_idx : start_idx + size] ret.append(static_features) return ret def multi_stream_mlpg( inputs, variances, windows, stream_sizes=None, has_dynamic_features=None, streams=None, ): """Split streams and do apply MLPG if stream has dynamic features Args: inputs (array like): input 3-d or 2-d array variances (array like): variances of input features windows (list): windows for parameter generation stream_sizes (list): stream sizes has_dynamic_features (list): binary flags that indicates if steams have dynamic features streams (list, optional): Streams of interests. Returns all streams if streams is None. Defaults to None. Raises: RuntimeError: if stream sizes are wrong Returns: array like: generated static features """ if stream_sizes is None: stream_sizes = [180, 3, 1, 3] if has_dynamic_features is None: has_dynamic_features = [True, True, False, True] if streams is None: streams = [True] * len(stream_sizes) T, D = inputs.shape if D != sum(stream_sizes): raise RuntimeError("You probably have specified wrong dimension params.") # Straem indices for static+delta features # [0, 180, 183, 184] start_indices = np.hstack(([0], np.cumsum(stream_sizes)[:-1])) # [180, 183, 184, 199] end_indices = np.cumsum(stream_sizes) ret = [] for in_start_idx, in_end_idx, v, enabled in zip( start_indices, end_indices, has_dynamic_features, streams, ): if not enabled: continue x = inputs[:, in_start_idx:in_end_idx] if inputs.shape == variances.shape: var_ = variances[:, in_start_idx:in_end_idx] else: var_ = np.tile(variances[in_start_idx:in_end_idx], (T, 1)) y = paramgen.mlpg(x, var_, windows) if v else x ret.append(y) return np.concatenate(ret, -1)
6,737
28.946667
96
py
nnsvs
nnsvs-master/nnsvs/discriminators.py
"""Discriminator implementations mostly used for GAN-based post-filters. All the discriminators must returns list of tensors. The last tensor of the list is regarded as the output of the discrminator. The others are used as intermedieate feature maps. """ import numpy as np import torch from nnsvs.util import init_weights from torch import nn class Conv2dD(nn.Module): """Conv2d-based discriminator The implementation follows the discrimiantor of the GAN-based post-filters in :cite:t:`Kaneko2017Interspeech`. Args: in_dim (int): Input feature dim channels (int): Number of channels kernel_size (tuple): Kernel size for 2d-convolution padding (tuple): Padding for 2d-convolution last_sigmoid (bool): If True, apply sigmoid on the output init_type (str): Initialization type padding_mode (str): Padding mode """ def __init__( self, in_dim=None, channels=64, kernel_size=(5, 3), padding=(0, 0), last_sigmoid=False, init_type="kaiming_normal", padding_mode="zeros", ): super().__init__() self.last_sigmoid = last_sigmoid C = channels ks = np.asarray(list(kernel_size)) if padding is None: padding = (ks - 1) // 2 self.convs = nn.ModuleList() self.convs.append( nn.Sequential( nn.Conv2d( 1, C, kernel_size=ks, padding=padding, stride=(1, 1), padding_mode=padding_mode, ), nn.LeakyReLU(0.2), ) ) self.convs.append( nn.Sequential( nn.Conv2d( C, 2 * C, kernel_size=ks, padding=padding, stride=(2, 1), padding_mode=padding_mode, ), nn.LeakyReLU(0.2), ) ) self.convs.append( nn.Sequential( nn.Conv2d( 2 * C, 4 * C, kernel_size=ks, padding=padding, stride=(2, 1), padding_mode=padding_mode, ), nn.LeakyReLU(0.2), ) ) self.convs.append( nn.Sequential( nn.Conv2d( 4 * C, 2 * C, kernel_size=ks, padding=padding, stride=(2, 1), padding_mode=padding_mode, ), nn.LeakyReLU(0.2), ) ) self.last_conv = nn.Conv2d( 2 * C, 1, kernel_size=ks, padding=padding, stride=(1, 1), padding_mode=padding_mode, ) init_weights(self, init_type) def forward(self, x, c=None, lengths=None): """Forward step Args: x (torch.Tensor): Input tensor c (torch.Tensor): Optional conditional features lengths (torch.Tensor): Optional lengths of the input Returns: list: List of output tensors """ outs = [] # (B, T, C) -> (B, 1, T, C): x = x.unsqueeze(1) for conv in self.convs: x = conv(x) outs.append(x) y = self.last_conv(x) y = torch.sigmoid(y) if self.last_sigmoid else y # (B, 1, T, C) -> (B, T, C) y = y.squeeze(1) outs.append(y) return [outs]
3,746
26.755556
78
py
nnsvs
nnsvs-master/nnsvs/mdn.py
import torch import torch.nn.functional as F from torch import nn class MDNLayer(nn.Module): """Mixture Density Network layer The input maps to the parameters of a Mixture of Gaussians (MoG) probability distribution, where each Gaussian has out_dim dimensions and diagonal covariance. If dim_wise is True, features for each dimension are modeld by independent 1-D GMMs instead of modeling jointly. This would workaround training difficulty especially for high dimensional data. Implementation references: 1. Mixture Density Networks by Mike Dusenberry https://mikedusenberry.com/mixture-density-networks 2. PRML book https://www.microsoft.com/en-us/research/people/cmbishop/prml-book/ 3. sagelywizard/pytorch-mdn https://github.com/sagelywizard/pytorch-mdn 4. sksq96/pytorch-mdn https://github.com/sksq96/pytorch-mdn Attributes: in_dim (int): the number of dimensions in the input out_dim (int): the number of dimensions in the output num_gaussians (int): the number of mixture component dim_wise (bool): whether to model data for each dimension separately """ def __init__(self, in_dim, out_dim, num_gaussians=30, dim_wise=False): super(MDNLayer, self).__init__() self.in_dim = in_dim self.out_dim = out_dim self.num_gaussians = num_gaussians self.dim_wise = dim_wise odim_log_pi = out_dim * num_gaussians if dim_wise else num_gaussians self.log_pi = nn.Linear(in_dim, odim_log_pi) self.log_sigma = nn.Linear(in_dim, out_dim * num_gaussians) self.mu = nn.Linear(in_dim, out_dim * num_gaussians) def forward(self, minibatch): """Forward for MDN Args: minibatch (torch.Tensor): tensor of shape (B, T, D_in) B is the batch size and T is data lengths of this batch, and D_in is in_dim. Returns: torch.Tensor: Tensor of shape (B, T, G) or (B, T, G, D_out) Log of mixture weights. G is num_gaussians and D_out is out_dim. torch.Tensor: Tensor of shape (B, T, G, D_out) the log of standard deviation of each Gaussians. torch.Tensor: Tensor of shape (B, T, G, D_out) mean of each Gaussians """ B = len(minibatch) if self.dim_wise: # (B, T, G, D_out) log_pi = self.log_pi(minibatch).view( B, -1, self.num_gaussians, self.out_dim ) log_pi = F.log_softmax(log_pi, dim=2) else: # (B, T, G) log_pi = F.log_softmax(self.log_pi(minibatch), dim=2) log_sigma = self.log_sigma(minibatch) log_sigma = log_sigma.view(B, -1, self.num_gaussians, self.out_dim) mu = self.mu(minibatch) mu = mu.view(B, -1, self.num_gaussians, self.out_dim) return log_pi, log_sigma, mu def mdn_loss( log_pi, log_sigma, mu, target, log_pi_min=-7.0, log_sigma_min=-7.0, reduce=True ): """Calculates the error, given the MoG parameters and the target. The loss is the negative log likelihood of the data given the MoG parameters. Args: log_pi (torch.Tensor): Tensor of shape (B, T, G) or (B, T, G, D_out) The log of multinomial distribution of the Gaussians. B is the batch size, T is data length of this batch, and G is num_gaussians of class MDNLayer. log_sigma (torch.Tensor): Tensor of shape (B, T, G ,D_out) The log standard deviation of the Gaussians. D_out is out_dim of class MDNLayer. mu (torch.Tensor): Tensor of shape (B, T, G, D_out) The means of the Gaussians. target (torch.Tensor): Tensor of shape (B, T, D_out) The target variables. log_pi_min (float): Minimum value of log_pi (for numerical stability) log_sigma_min (float): Minimum value of log_sigma (for numerical stability) reduce: If True, the losses are averaged for each batch. Returns: loss (B) or (B, T): Negative Log Likelihood of Mixture Density Networks. """ dim_wise = len(log_pi.shape) == 4 # Clip log_sigma and log_pi with log_clamp_min for numerical stability log_sigma = torch.clamp(log_sigma, min=log_sigma_min) log_pi = torch.clamp(log_pi, min=log_pi_min) # Expand the dim of target as (B, T, D_out) -> (B, T, 1, D_out) -> (B, T,G, D_out) target = target.unsqueeze(2).expand_as(log_sigma) # Center target variables and clamp them within +/- 5SD for numerical stability. centered_target = target - mu scale = torch.exp(log_sigma) edge = 5 * scale centered_target = torch.where(centered_target > edge, edge, centered_target) centered_target = torch.where(centered_target < -edge, -edge, centered_target) # Create gaussians with mean=0 and variance=torch.exp(log_sigma)^2 dist = torch.distributions.Normal(loc=0, scale=scale) log_prob = dist.log_prob(centered_target) if dim_wise: # (B, T, D_out. D_out) loss = log_prob + log_pi else: # Here we assume that the covariance matrix of multivariate Gaussian # distribution is diagonal to handle the mean and the variance in each # dimension separately. # Reference: # https://markusthill.github.io/gaussian-distribution-with-a-diagonal-covariance-matrix/ # log pi(x)N(y|mu(x),sigma(x)) = log pi(x) + log N(y|mu(x),sigma(x)) # log N(y_1,y_2,...,y_{D_out}|mu(x),sigma(x)) # = log N(y_1|mu(x),sigma(x))...N(y_{D_out}|mu(x),sigma(x)) # = \sum_{i=1}^{D_out} log N(y_i|mu(x),sigma(x)) # (B, T, G, D_out) -> (B, T, G) loss = torch.sum(log_prob, dim=3) + log_pi # Calculate negative log likelihood. # Use torch.log_sum_exp instead of the combination of torch.sum and torch.log # (Reference: https://github.com/r9y9/nnsvs/pull/20#discussion_r495514563) # if dim_wise is True: (B, T, G, D_out) -> (B, T, D_out) # else (B, T, G) -> (B, T) loss = -torch.logsumexp(loss, dim=2) if reduce: # (B, T) -> (B) return torch.mean(loss, dim=1) else: # not averaged (for applying mask later) # (B, T) return loss return # from r9y9/wavenet_vocoder/wavenet_vocoder/mixture.py def to_one_hot(tensor, n, fill_with=1.0): # we perform one hot encore with respect to the last axis one_hot = torch.FloatTensor(tensor.size() + (n,)).zero_() if tensor.is_cuda: one_hot = one_hot.cuda() one_hot.scatter_(len(tensor.size()), tensor.unsqueeze(-1), fill_with) return one_hot def mdn_get_most_probable_sigma_and_mu(log_pi, log_sigma, mu): """Return the mean and standard deviation of the Gaussian component whose weight coefficient is the largest as the most probable predictions. Args: log_pi (torch.Tensor): Tensor of shape (B, T, G) or (B, T, G, D_out) The log of multinomial distribution of the Gaussians. B is the batch size, T is data length of this batch, G is num_gaussians of class MDNLayer. log_sigma (torch.Tensor): Tensor of shape (B, T, G, D_out) The standard deviation of the Gaussians. D_out is out_dim of class MDNLayer. mu (torch.Tensor): Tensor of shape (B, T, G, D_out) The means of the Gaussians. D_out is out_dim of class MDNLayer. Returns: tuple: tuple of torch.Tensor torch.Tensor of shape (B, T, D_out). The standardd deviations of the most probable Gaussian component. torch.Tensor of shape (B, T, D_out). Means of the Gaussians. """ dim_wise = len(log_pi.shape) == 4 _, _, num_gaussians, _ = mu.shape # Get the indexes of the largest log_pi _, max_component = torch.max(log_pi, dim=2) # (B, T) or (B, T, C_out) # Convert max_component to one_hot manner # if dim_wise: (B, T, D_out) -> (B, T, D_out, G) # else: (B, T) -> (B, T, G) one_hot = to_one_hot(max_component, num_gaussians) if dim_wise: # (B, T, G, D_out) one_hot = one_hot.transpose(2, 3) assert one_hot.shape == mu.shape else: # Expand the dim of one_hot as (B, T, G) -> (B, T, G, d_out) one_hot = one_hot.unsqueeze(3).expand_as(mu) # Multiply one_hot and sum to get mean(mu) and standard deviation(sigma) # of the Gaussians whose weight coefficient(log_pi) is the largest. # (B, T, G, d_out) -> (B, T, d_out) max_mu = torch.sum(mu * one_hot, dim=2) max_sigma = torch.exp(torch.sum(log_sigma * one_hot, dim=2)) return max_sigma, max_mu def mdn_get_sample(log_pi, log_sigma, mu): """Sample from mixture of the Gaussian component whose weight coefficient is the largest as the most probable predictions. Args: log_pi (torch.Tensor): Tensor of shape (B, T, G) or (B, T, G, D_out) The log of multinomial distribution of the Gaussians. B is the batch size, T is data length of this batch, G is num_gaussians of class MDNLayer. log_sigma (torch.Tensor): Tensor of shape (B, T, G, D_out) The log of standard deviation of the Gaussians. D_out is out_dim of class MDNLayer. mu (torch.Tensor): Tensor of shape (B, T, G, D_out) The means of the Gaussians. D_out is out_dim of class MDNLayer. Returns: torch.Tensor: Tensor of shape (B, T, D_out) Sample from the mixture of the Gaussian component. """ max_sigma, max_mu = mdn_get_most_probable_sigma_and_mu(log_pi, log_sigma, mu) # Create gaussians with mean=max_mu and variance=max_log_sigma^2 dist = torch.distributions.Normal(loc=max_mu, scale=max_sigma) # Sample from normal distribution sample = dist.sample() return sample
9,902
39.753086
96
py
nnsvs
nnsvs-master/nnsvs/svs.py
import json import time from copy import deepcopy from pathlib import Path import numpy as np import torch from hydra.utils import instantiate from nnmnkwii.io import hts from nnmnkwii.preprocessing.f0 import interp1d from nnsvs.gen import ( postprocess_acoustic, postprocess_duration, postprocess_waveform, predict_acoustic, predict_duration, predict_timelag, predict_waveform, ) from nnsvs.io.hts import ( full_to_mono, get_note_indices, get_pitch_index, get_pitch_indices, label2phrases, label2phrases_str, segment_labels, ) from nnsvs.logger import getLogger from nnsvs.usfgan import USFGANWrapper from nnsvs.util import MinMaxScaler, StandardScaler, extract_static_scaler, load_vocoder from omegaconf import OmegaConf class BaseSVS(object): """Base class for singing voice synthesis (SVS) inference All SVS engines should inherit from this class. The input of the SVS engine uses the HTS-style full-context labels. The output should be a tuple of raw waveform and sampling rate. To allow language-independent SVS, this base class does not define the interface for the frontend functionality such as converting musicXML/UST to HTS labels. The frontend processing should be done externally (e.g., using pysinsy or utaupy) or can be implemented with an optional method. """ def svs(self, labels, *args, **kwargs): """Run SVS inference and returns the synthesized waveform Args: labels (nnmnkwii.io.hts.HTSLabelFile): HTS labels Returns: tuple: (waveform, sampling rate) """ pass class SPSVS(BaseSVS): """Statistical parametric singing voice synthesis (SPSVS) Use the ``svs`` method for the simplest inference, or use the separated methods (e.g.,``predict_acoustic`` and ``predict_waveform``) to control each components of the parametric SVS system. Args: model_dir (str): directory of the model device (str): cpu or cuda verbose (int): verbosity level Examples: Synthesize wavefrom from a musicxml file .. plot:: import numpy as np import pysinsy from nnmnkwii.io import hts from nnsvs.pretrained import retrieve_pretrained_model from nnsvs.svs import SPSVS from nnsvs.util import example_xml_file import matplotlib.pyplot as plt # Instantiate the SVS engine model_dir = retrieve_pretrained_model("r9y9/yoko_latest") engine = SPSVS(model_dir) # Extract HTS labels from a MusicXML file contexts = pysinsy.extract_fullcontext(example_xml_file(key="get_over")) labels = hts.HTSLabelFile.create_from_contexts(contexts) # Run inference wav, sr = engine.svs(labels) # Plot the result fig, ax = plt.subplots(figsize=(8,2)) librosa.display.waveshow(wav.astype(np.float32), sr=sr, ax=ax) With WORLD vocoder: >>> wav, sr = engine.svs(labels, vocoder_type="world") With a uSFGAN or SiFiGAN vocoder: >>> wav, sr = engine.svs(labels, vocoder_type="usfgan") """ def __init__(self, model_dir, device="cpu", verbose=0): self.device = device # NOTE: assuming that the logger is instantiated without hydra # needs to add stream handler to the logger explicitly self.logger = getLogger(verbose=verbose, add_stream_handler=True) if isinstance(model_dir, str): model_dir = Path(model_dir) # search for config.yaml assert model_dir / "config.yaml" self.config = OmegaConf.load(model_dir / "config.yaml") self.feature_type = self.config.get("feature_type", "world") self.sample_rate = self.config.get("sample_rate", 48000) # qst self.binary_dict, self.numeric_dict = hts.load_question_set( model_dir / "qst.hed" ) self.pitch_idx = get_pitch_index(self.binary_dict, self.numeric_dict) self.pitch_indices = get_pitch_indices(self.binary_dict, self.numeric_dict) # Time-lag model self.timelag_config = OmegaConf.load(model_dir / "timelag_model.yaml") self.timelag_model = instantiate(self.timelag_config.netG).to(device) checkpoint = torch.load( model_dir / "timelag_model.pth", map_location=device, ) self.timelag_model.load_state_dict(checkpoint["state_dict"]) self.timelag_in_scaler = MinMaxScaler( np.load(model_dir / "in_timelag_scaler_min.npy"), np.load(model_dir / "in_timelag_scaler_scale.npy"), ) self.timelag_out_scaler = StandardScaler( np.load(model_dir / "out_timelag_scaler_mean.npy"), np.load(model_dir / "out_timelag_scaler_var.npy"), np.load(model_dir / "out_timelag_scaler_scale.npy"), ) self.timelag_model.eval() # Duration model self.duration_config = OmegaConf.load(model_dir / "duration_model.yaml") self.duration_model = instantiate(self.duration_config.netG).to(device) checkpoint = torch.load( model_dir / "duration_model.pth", map_location=device, ) self.duration_model.load_state_dict(checkpoint["state_dict"]) self.duration_in_scaler = MinMaxScaler( np.load(model_dir / "in_duration_scaler_min.npy"), np.load(model_dir / "in_duration_scaler_scale.npy"), ) self.duration_out_scaler = StandardScaler( np.load(model_dir / "out_duration_scaler_mean.npy"), np.load(model_dir / "out_duration_scaler_var.npy"), np.load(model_dir / "out_duration_scaler_scale.npy"), ) self.duration_model.eval() # Acoustic model self.acoustic_config = OmegaConf.load(model_dir / "acoustic_model.yaml") self.acoustic_model = instantiate(self.acoustic_config.netG).to(device) checkpoint = torch.load( model_dir / "acoustic_model.pth", map_location=device, ) self.acoustic_model.load_state_dict(checkpoint["state_dict"]) self.acoustic_in_scaler = MinMaxScaler( np.load(model_dir / "in_acoustic_scaler_min.npy"), np.load(model_dir / "in_acoustic_scaler_scale.npy"), ) self.acoustic_out_scaler = StandardScaler( np.load(model_dir / "out_acoustic_scaler_mean.npy"), np.load(model_dir / "out_acoustic_scaler_var.npy"), np.load(model_dir / "out_acoustic_scaler_scale.npy"), ) # NOTE: this is used for GV post-filtering self.acoustic_out_static_scaler = extract_static_scaler( self.acoustic_out_scaler, self.acoustic_config ) # (Optional) lf0 model if (model_dir / "lf0_model.pth").exists(): assert hasattr(self.acoustic_model, "lf0_model") self.logger.info("Loading an external lf0 model.") checkpoint = torch.load( model_dir / "lf0_model.pth", map_location=device, ) self.acoustic_model.lf0_model.load_state_dict(checkpoint["state_dict"]) self.acoustic_model.eval() # Post-filter if (model_dir / "postfilter_model.yaml").exists(): self.postfilter_config = OmegaConf.load(model_dir / "postfilter_model.yaml") self.postfilter_model = instantiate(self.postfilter_config.netG).to(device) checkpoint = torch.load( model_dir / "postfilter_model.pth", map_location=device, ) self.postfilter_model.load_state_dict(checkpoint["state_dict"]) self.postfilter_model.eval() self.postfilter_out_scaler = StandardScaler( np.load(model_dir / "out_postfilter_scaler_mean.npy"), np.load(model_dir / "out_postfilter_scaler_var.npy"), np.load(model_dir / "out_postfilter_scaler_scale.npy"), ) else: self.postfilter_model = None self.postfilter_config = None self.postfilter_out_scaler = None # Vocoder model if (model_dir / "vocoder_model.pth").exists(): self.vocoder, self.vocoder_in_scaler, self.vocoder_config = load_vocoder( model_dir / "vocoder_model.pth", device, self.acoustic_config ) else: self.logger.info( "No trained vocoder model found. WORLD vocoder will be used." ) self.vocoder = None self.vocoder_config = None self.vocoder_in_scaler = None def __repr__(self): timelag_str = json.dumps( OmegaConf.to_container(self.timelag_config.netG), sort_keys=False, indent=4, ) duration_str = json.dumps( OmegaConf.to_container(self.duration_config.netG), sort_keys=False, indent=4, ) acoustic_str = json.dumps( OmegaConf.to_container(self.acoustic_config.netG), sort_keys=False, indent=4, ) repr = f"""Statistical parametric SVS (sampling rate: {self.sample_rate}) Time-lag model: {timelag_str} Duration model: {duration_str} Acoustic model: {acoustic_str} """ if self.postfilter_model is not None: postfilter_str = json.dumps( OmegaConf.to_container(self.postfilter_config.netG), sort_keys=False, indent=4, ) repr += f"Post-filter model: {postfilter_str}\n" else: repr += "Post-filter model: None\n" if self.vocoder is not None: if ( "generator" in self.vocoder_config and "discriminator" in self.vocoder_config ): # usfgan vocoder_params = OmegaConf.to_container( self.vocoder_config["generator"], throw_on_missing=True ) else: vocoder_params = { "generator_type": self.vocoder_config.get( "generator_type", "ParallelWaveGANGenerator" # type: ignore ), "generator_params": OmegaConf.to_container( self.vocoder_config.generator_params ), } vocoder_str = json.dumps( vocoder_params, sort_keys=False, indent=4, ) repr += f"Vocoder model: {vocoder_str}\n" else: repr += "Vocoder model: WORLD\n" return repr def set_device(self, device): """Set device for the SVS model Args: device (str): cpu or cuda. """ self.logger.info(f"Set device to {device}") self.device = device self.timelag_model.to(device) self.duration_model.to(device) self.acoustic_model.to(device) self.postfilter_model.to(device) if self.postfilter_model is not None else None self.vocoder.to(device) if self.vocoder is not None else None def predict_timelag(self, labels): """Predict time-ag from HTS labels Args: labels (nnmnkwii.io.hts.HTSLabelFile): HTS labels. Returns: ndarray: Predicted time-lag. """ start_time = time.time() lag = predict_timelag( self.device, labels, timelag_model=self.timelag_model, timelag_config=self.timelag_config, timelag_in_scaler=self.timelag_in_scaler, timelag_out_scaler=self.timelag_out_scaler, binary_dict=self.binary_dict, numeric_dict=self.numeric_dict, pitch_indices=self.pitch_indices, log_f0_conditioning=self.config.log_f0_conditioning, allowed_range=self.config.timelag.allowed_range, allowed_range_rest=self.config.timelag.allowed_range_rest, force_clip_input_features=self.config.timelag.force_clip_input_features, frame_period=self.config.frame_period, ) self.logger.info( f"Elapsed time for time-lag prediction: {time.time() - start_time:.3f} sec" ) return lag def predict_duration(self, labels): """Predict durations from HTS labels Args: labels (nnmnkwii.io.hts.HTSLabelFile): HTS labels. Returns: ndarray: Predicted durations. """ start_time = time.time() durations = predict_duration( self.device, labels, duration_model=self.duration_model, duration_config=self.duration_config, duration_in_scaler=self.duration_in_scaler, duration_out_scaler=self.duration_out_scaler, binary_dict=self.binary_dict, numeric_dict=self.numeric_dict, pitch_indices=self.pitch_indices, log_f0_conditioning=self.config.log_f0_conditioning, force_clip_input_features=self.config.duration.force_clip_input_features, frame_period=self.config.frame_period, ) self.logger.info( f"Elapsed time for duration prediction: {time.time() - start_time:.3f} sec" ) return durations def postprocess_duration(self, labels, pred_durations, lag): """Post-process durations Args: labels (nnmnkwii.io.hts.HTSLabelFile): HTS labels. pred_durations (ndarray): Predicted durations. lag (ndarray): Predicted time-lag. Returns: nnmnkwii.io.hts.HTSLabelFile: duration modified HTS labels. """ start_time = time.time() duration_modified_labels = postprocess_duration( labels, pred_durations, lag, frame_period=self.config.frame_period ) self.logger.info( f"Elapsed time for duration post-processing: {time.time() - start_time:.3f} sec" ) return duration_modified_labels def predict_timing(self, labels): """Predict timing from HTS labels Args: labels (nnmnkwii.io.hts.HTSLabelFile): HTS labels. Returns: nnmnkwii.io.hts.HTSLabelFile: duration modified HTS labels. """ lag = self.predict_timelag(labels) durations = self.predict_duration(labels) duration_modified_full_labels = self.postprocess_duration( labels, durations, lag ) return duration_modified_full_labels def predict_acoustic(self, duration_modified_labels, f0_shift_in_cent=0): """Predict acoustic features from HTS labels Args: duration_modified_labels (nnmnkwii.io.hts.HTSLabelFile): HTS labels. f0_shift_in_cent (float): F0 shift in cent. Returns: ndarray: Predicted acoustic features. """ start_time = time.time() acoustic_features = predict_acoustic( device=self.device, labels=duration_modified_labels, acoustic_model=self.acoustic_model, acoustic_config=self.acoustic_config, acoustic_in_scaler=self.acoustic_in_scaler, acoustic_out_scaler=self.acoustic_out_scaler, binary_dict=self.binary_dict, numeric_dict=self.numeric_dict, subphone_features=self.acoustic_config.get( "subphone_features", "coarse_coding" ), pitch_indices=self.pitch_indices, log_f0_conditioning=self.config.log_f0_conditioning, force_clip_input_features=self.acoustic_config.get( "force_clip_input_features", True ), frame_period=self.config.frame_period, f0_shift_in_cent=f0_shift_in_cent, ) self.logger.info( f"Elapsed time for acoustic feature prediction: {time.time() - start_time:.3f} sec" ) # log real-time factor (RT) RT = (time.time() - start_time) / ( acoustic_features.shape[0] * self.config.frame_period / 1000 ) self.logger.info(f"Real-time factor for acoustic feature prediction: {RT:.3f}") return acoustic_features def postprocess_acoustic( self, duration_modified_labels, acoustic_features, post_filter_type="gv", trajectory_smoothing=True, trajectory_smoothing_cutoff=50, trajectory_smoothing_cutoff_f0=20, vuv_threshold=0.5, force_fix_vuv=False, fill_silence_to_rest=False, f0_shift_in_cent=0, ): """Post-process acoustic features The function converts acoustic features in single ndarray to tuple of multi-stream acoustic features. e.g., array -> (mgc, lf0, vuv, bap) If post_filter_type=``nnsvs`` is specified, learned post-filter is applied. However, it is recommended to use ``gv`` in general. Args: duration_modified_labels (nnmnkwii.io.hts.HTSLabelFile): HTS labels. acoustic_features (ndarray): Predicted acoustic features. post_filter_type (str): Post-filter type. One of ``gv``, ``merlin`` or ``nnsvs``. Recommended to use ``gv`` for general purpose. trajectory_smoothing (bool): Whether to apply trajectory smoothing. trajectory_smoothing_cutoff (float): Cutoff frequency for trajectory smoothing of spectral features. trajectory_smoothing_cutoff_f0 (float): Cutoff frequency for trajectory smoothing of f0. vuv_threshold (float): V/UV threshold. force_fix_vuv (bool): Force fix V/UV. fill_silence_to_rest (bool): Fill silence to rest frames. f0_shift_in_cent (float): F0 shift in cent. Returns: tuple: Post-processed multi-stream acoustic features. """ start_time = time.time() multistream_features = postprocess_acoustic( device=self.device, duration_modified_labels=duration_modified_labels, acoustic_features=acoustic_features, binary_dict=self.binary_dict, numeric_dict=self.numeric_dict, acoustic_config=self.acoustic_config, acoustic_out_static_scaler=self.acoustic_out_static_scaler, postfilter_model=self.postfilter_model, postfilter_config=self.postfilter_config, postfilter_out_scaler=self.postfilter_out_scaler, sample_rate=self.sample_rate, frame_period=self.config.frame_period, relative_f0=self.config.acoustic.relative_f0, feature_type=self.feature_type, post_filter_type=post_filter_type, trajectory_smoothing=trajectory_smoothing, trajectory_smoothing_cutoff=trajectory_smoothing_cutoff, trajectory_smoothing_cutoff_f0=trajectory_smoothing_cutoff_f0, vuv_threshold=vuv_threshold, f0_shift_in_cent=f0_shift_in_cent, vibrato_scale=1.0, # only valid for Sinsy-like models force_fix_vuv=force_fix_vuv, fill_silence_to_rest=fill_silence_to_rest, ) self.logger.info( f"Elapsed time for acoustic post-processing: {time.time() - start_time:.3f} sec" ) return multistream_features def predict_waveform( self, multistream_features, vocoder_type="world", vuv_threshold=0.5, ): """Predict waveform from acoustic features Args: multistream_features (tuple): Multi-stream acoustic features. vocoder_type (str): Vocoder type. One of ``world``, ``pwg`` or ``usfgan``. If ``auto`` is specified, the vocoder is automatically selected. vuv_threshold (float): V/UV threshold. Returns: ndarray: Predicted waveform. """ start_time = time.time() if vocoder_type in ["pwg", "usfgan"] and self.vocoder is None: raise ValueError( """Pre-trained vocodr model is not found. WORLD is only supported for waveform generation""" ) if vocoder_type == "auto": if self.feature_type == "melf0": assert self.vocoder is not None vocoder_type = ( "usfgan" if isinstance(self.vocoder, USFGANWrapper) else "pwg" ) elif self.feature_type == "world": if self.vocoder is None: vocoder_type = "world" else: vocoder_type = ( "usfgan" if isinstance(self.vocoder, USFGANWrapper) else "pwg" ) wav = predict_waveform( device=self.device, multistream_features=multistream_features, vocoder=self.vocoder, vocoder_config=self.vocoder_config, vocoder_in_scaler=self.vocoder_in_scaler, sample_rate=self.sample_rate, frame_period=self.config.frame_period, use_world_codec=self.config.get("use_world_codec", False), feature_type=self.feature_type, vocoder_type=vocoder_type, vuv_threshold=vuv_threshold, ) self.logger.info( f"Elapsed time for waveform generation: {time.time() - start_time:.3f} sec" ) RT = (time.time() - start_time) / (len(wav) / self.sample_rate) self.logger.info(f"Real-time factor for waveform generation: {RT:.3f}") return wav def postprocess_waveform( self, wav, dtype=np.int16, peak_norm=False, loudness_norm=False, target_loudness=-20, ): """Post-process waveform Args: wav (ndarray): Waveform. dtype (dtype): Data type of waveform. peak_norm (bool): Whether to apply peak normalization. loudness_norm (bool): Whether to apply loudness normalization. target_loudness (float): Target loudness in dB. Returns: ndarray: Post-processed waveform. """ start_time = time.time() wav = postprocess_waveform( wav=wav, sample_rate=self.sample_rate, dtype=dtype, peak_norm=peak_norm, loudness_norm=loudness_norm, target_loudness=target_loudness, ) self.logger.info( f"Elapsed time for waveform post-processing: {time.time() - start_time:.3f} sec" ) return wav def svs( self, labels, vocoder_type="world", post_filter_type="gv", trajectory_smoothing=True, trajectory_smoothing_cutoff=50, trajectory_smoothing_cutoff_f0=20, vuv_threshold=0.5, style_shift=0, force_fix_vuv=False, fill_silence_to_rest=False, dtype=np.int16, peak_norm=False, loudness_norm=False, target_loudness=-20, segmented_synthesis=False, ): """Synthesize waveform from HTS labels. Args: labels (nnmnkwii.io.hts.HTSLabelFile): HTS labels vocoder_type (str): Vocoder type. One of ``world``, ``pwg`` or ``usfgan``. If ``auto`` is specified, the vocoder is automatically selected. post_filter_type (str): Post-filter type. ``merlin``, ``gv`` or ``nnsvs`` is supported. trajectory_smoothing (bool): Whether to smooth acoustic feature trajectory. trajectory_smoothing_cutoff (int): Cutoff frequency for trajectory smoothing. trajectory_smoothing_cutoff_f0 (int): Cutoff frequency for trajectory smoothing of f0. vuv_threshold (float): Threshold for VUV. style_shift (int): style shift parameter force_fix_vuv (bool): Whether to correct VUV. fill_silence_to_rest (bool): Fill silence to rest frames. dtype (np.dtype): Data type of the output waveform. peak_norm (bool): Whether to normalize the waveform by peak value. loudness_norm (bool): Whether to normalize the waveform by loudness. target_loudness (float): Target loudness in dB. segmneted_synthesis (bool): Whether to use segmented synthesis. """ start_time = time.time() vocoder_type = vocoder_type.lower() if vocoder_type not in ["world", "pwg", "usfgan", "auto"]: raise ValueError(f"Unknown vocoder type: {vocoder_type}") if post_filter_type not in ["merlin", "nnsvs", "gv", "none"]: raise ValueError(f"Unknown post-filter type: {post_filter_type}") # Predict timinigs duration_modified_labels = self.predict_timing(labels) # NOTE: segmented synthesis is not well tested. There MUST be better ways # to do this. if segmented_synthesis: self.logger.warning( "Segmented synthesis is not well tested. Use it on your own risk." ) duration_modified_labels_segs = segment_labels( duration_modified_labels, # the following parameters are based on experiments in the NNSVS's paper # tuned with Namine Ritsu's database silence_threshold=0.1, min_duration=5.0, force_split_threshold=5.0, ) from tqdm.auto import tqdm else: duration_modified_labels_segs = [duration_modified_labels] def tqdm(x, **kwargs): return x # Run acoustic model and vocoder hts_frame_shift = int(self.config.frame_period * 1e4) wavs = [] self.logger.info(f"Number of segments: {len(duration_modified_labels_segs)}") for duration_modified_labels_seg in tqdm( duration_modified_labels_segs, desc="[segment]", total=len(duration_modified_labels_segs), ): duration_modified_labels_seg.frame_shift = hts_frame_shift # Predict acoustic features # NOTE: if non-zero pre_f0_shift_in_cent is specified, the input pitch # will be shifted before running the acoustic model acoustic_features = self.predict_acoustic( duration_modified_labels_seg, f0_shift_in_cent=style_shift * 100, ) # Post-processing for acoustic features # NOTE: if non-zero post_f0_shift_in_cent is specified, the output pitch # will be shifted as a part of post-processing multistream_features = self.postprocess_acoustic( acoustic_features=acoustic_features, duration_modified_labels=duration_modified_labels_seg, trajectory_smoothing=trajectory_smoothing, trajectory_smoothing_cutoff=trajectory_smoothing_cutoff, trajectory_smoothing_cutoff_f0=trajectory_smoothing_cutoff_f0, force_fix_vuv=force_fix_vuv, fill_silence_to_rest=fill_silence_to_rest, f0_shift_in_cent=-style_shift * 100, ) # Generate waveform by vocoder wav = self.predict_waveform( multistream_features=multistream_features, vocoder_type=vocoder_type, vuv_threshold=vuv_threshold, ) wavs.append(wav) # Concatenate segmented waveforms wav = np.concatenate(wavs, axis=0).reshape(-1) # Post-processing for the output waveform wav = self.postprocess_waveform( wav, dtype=dtype, peak_norm=peak_norm, loudness_norm=loudness_norm, target_loudness=target_loudness, ) self.logger.info(f"Total time: {time.time() - start_time:.3f} sec") RT = (time.time() - start_time) / (len(wav) / self.sample_rate) self.logger.info(f"Total real-time factor: {RT:.3f}") return wav, self.sample_rate def _warn_if_model_is_old(logger): logger.warning( """It is likely you have trained you model with old NNSVS. It is recommended to retrain your model with the latest version of NNSVS.""" ) class NEUTRINO(SPSVS): """NEUTRINO-like interface for singing voice synthesis Args: model_dir (str): model directory device (str): device name verbose (int): verbose level """ def __init__(self, model_dir, device="cpu", verbose=0): super().__init__(model_dir, device=device, verbose=verbose) if self.feature_type != "world": raise RuntimeError(f"Unsupported feature type: {self.feature_type}") if not self.config.get("use_world_codec", False): self.logger.warning( "WORLD coded is required to output NEUTRIN-compatible features" ) _warn_if_model_is_old(self.logger) @classmethod def musicxml2label(cls, input_file): """Convert musicXML to full and mono HTS labels Args: input_file (str): musicXML file """ import pysinsy contexts = pysinsy.extract_fullcontext(input_file) full_labels = hts.HTSLabelFile.create_from_contexts(contexts) mono_labels = full_to_mono(full_labels) return full_labels, mono_labels def get_num_phrases(self, labels): """Get number of phrases Args: labels (nnmnkwii.io.hts.HTSLabelFile): HTS label Returns: int: number of phrases """ phrases = label2phrases(labels) return len(phrases) def get_phraselist(self, full_labels, timing_labels): """Get phraselit from full and timing HTS labels Args: full_labels (nnmnkwii.io.hts.HTSLabelFile): full HTS label timing_labels (nnmnkwii.io.hts.HTSLabelFile): timing HTS label Returns: str: phraselist """ note_indices = get_note_indices(full_labels) phraselist = label2phrases_str(timing_labels, note_indices) return phraselist def predict_acoustic( self, full_labels, timing_labels=None, style_shift=0, phrase_num=-1, trajectory_smoothing=True, trajectory_smoothing_cutoff=50, trajectory_smoothing_cutoff_f0=20, vuv_threshold=0.5, force_fix_vuv=False, fill_silence_to_rest=False, ): """Main inference of timing and acoustic predictions Args: full_labels (nnmnkwii.io.hts.HTSLabelFile): full HTS label timing_labels (nnmnkwii.io.hts.HTSLabelFile): timing HTS label style_shift (int): style shift parameter phrase_num (int): phrase number to use for inference trajectory_smoothing (bool): whether to apply trajectory smoothing trajectory_smoothing_cutoff (float): cutoff frequency for trajectory smoothing trajectory_smoothing_cutoff_f0 (float): cutoff frequency for trajectory smoothing for f0 vuv_threshold (float): V/UV threshold force_fix_vuv (bool): whether to force fix V/UV fill_silence_to_rest (bool): Fill silence to rest frames. Returns: tuple: (f0, mgc, bap) """ if timing_labels is None: self.logger.warning("'timing_labels' is not provided.") # Run timing prediction duration_modified_full_labels = self.predict_timing(full_labels) timing_labels = full_to_mono(duration_modified_full_labels) else: # Load pre-estimated timing duration_modified_full_labels = deepcopy(full_labels) duration_modified_full_labels.start_times = timing_labels.start_times.copy() duration_modified_full_labels.end_times = timing_labels.end_times.copy() if phrase_num >= 0: phrases = label2phrases(duration_modified_full_labels) if phrase_num > len(phrases): raise RuntimeError( f"phrase_num is too large: {phrase_num} > {len(phrases)}" ) # Use the specified phrase for inference duration_modified_full_labels = phrases[phrase_num] self.logger.info(f"Using phrase {phrase_num}/{len(phrases)} for inference") # Predict acoustic features # NOTE: if non-zero pre_f0_shift_in_cent is specified, the input pitch # will be shifted before running the acoustic model acoustic_features = super().predict_acoustic( duration_modified_full_labels, f0_shift_in_cent=style_shift * 100, ) # Post-processing for acoustic features # NOTE: if non-zero post_f0_shift_in_cent is specified, the output pitch # will be shifted as a part of post-processing multistream_features = super().postprocess_acoustic( acoustic_features=acoustic_features, duration_modified_labels=duration_modified_full_labels, trajectory_smoothing=trajectory_smoothing, trajectory_smoothing_cutoff=trajectory_smoothing_cutoff, trajectory_smoothing_cutoff_f0=trajectory_smoothing_cutoff_f0, vuv_threshold=vuv_threshold, force_fix_vuv=force_fix_vuv, fill_silence_to_rest=fill_silence_to_rest, f0_shift_in_cent=-style_shift * 100, ) assert len(multistream_features) == 4 mgc, lf0, vuv, bap = multistream_features if not self.config.get("use_world_codec", False): self.logger.warning( """use_world_codec is not set. WORLD (NEUTRINO edition) does not work with the output of this model. """ ) # Convert lf0 to f0 f0 = np.exp(lf0.copy()) f0[vuv < vuv_threshold] = 0 # NOTE: Neutrino-compatible MGC should have negative values at the 0-th coefficient. if mgc[:, 0].mean() > 0: self.logger.warning("MGC 0-th coefficient is positive.") _warn_if_model_is_old(self.logger) # Make sure to have correct array layout and dtype # These parameters can be used to generate waveform by WORLD f0 = np.ascontiguousarray(f0).astype(np.float64) mgc = np.ascontiguousarray(mgc).astype(np.float64) bap = np.ascontiguousarray(bap).astype(np.float64) return f0, mgc, bap def predict_waveform( self, f0, mgc, bap, vocoder_type="world", vuv_threshold=0.5, dtype=np.int16, peak_norm=False, loudness_norm=False, target_loudness=-20, ): """Generate waveform from acoustic features Args: f0 (ndarray): f0 mgc (ndarray): mel-cepstrum bap (ndarray): band-aperiodicity vocoder_type (str): vocoder type vuv_threshold (float): V/UV threshold dtype (np.dtype): Data type of the output waveform. peak_norm (bool): Whether to normalize the waveform by peak value. loudness_norm (bool): Whether to normalize the waveform by loudness. target_loudness (float): Target loudness in dB. Returns: ndarray: waveform """ # Convert NEUTRINO-like features to NNSVS's one # (f0, mgc, bap) -> (mgc, lf0, vuv, bap) vuv = (f0 > 0).astype(np.float64).reshape(-1, 1) lf0 = f0.copy() lf0[np.nonzero(lf0)] = np.log(f0[np.nonzero(lf0)]) lf0 = interp1d(lf0, kind="slinear") multistream_features = (mgc, lf0, vuv, bap) wav = super().predict_waveform( multistream_features=multistream_features, vocoder_type=vocoder_type, vuv_threshold=vuv_threshold, ) wav = self.postprocess_waveform( wav, dtype=dtype, peak_norm=peak_norm, loudness_norm=loudness_norm, target_loudness=target_loudness, ) return wav def svs(self, labels): """Synthesize wavefrom from HTS labels Args: labels (nnmnkwii.io.hts.HTSLabelFile): HTS labels Returns: tuple: (waveform, sample_rate) """ self.logger.warning( "Use `predict_acoustic` and `predict_waveform` methods instead." ) f0, mgc, bap = self.predict_acoustic(labels) wav = self.predict_waveform(f0, mgc, bap) return wav, self.sample_rate
37,016
36.657172
95
py
nnsvs
nnsvs-master/nnsvs/model.py
from warnings import warn import torch from nnsvs.base import BaseModel, PredictionType from nnsvs.dsp import TrTimeInvFIRFilter from nnsvs.layers.conv import ResnetBlock, WNConv1d from nnsvs.layers.layer_norm import LayerNorm from nnsvs.mdn import MDNLayer, mdn_get_most_probable_sigma_and_mu from nnsvs.multistream import split_streams from nnsvs.transformer.attentions import sequence_mask from nnsvs.transformer.encoder import Encoder as _TransformerEncoder from nnsvs.util import init_weights from torch import nn from torch.nn.utils.rnn import pack_padded_sequence, pad_packed_sequence __all__ = [ "FFN", "LSTMRNN", "LSTMRNNSAR", "MDN", "MDNv2", "RMDN", "Conv1dResnet", "Conv1dResnetMDN", "Conv1dResnetSAR", "FFConvLSTM", "LSTMEncoder", "VariancePredictor", "TransformerEncoder", ] class Conv1dResnet(BaseModel): """Conv1d + Resnet The model is inspired by the MelGAN's model architecture (:cite:t:`kumar2019melgan`). MDN layer is added if use_mdn is True. Args: in_dim (int): the dimension of the input hidden_dim (int): the dimension of the hidden state out_dim (int): the dimension of the output num_layers (int): the number of layers init_type (str): the type of weight initialization use_mdn (bool): whether to use MDN or not num_gaussians (int): the number of gaussians in MDN dim_wise (bool): whether to use dim-wise or not """ def __init__( self, in_dim, hidden_dim, out_dim, num_layers=4, init_type="none", use_mdn=False, num_gaussians=8, dim_wise=False, in_ph_start_idx: int = 1, in_ph_end_idx: int = 50, embed_dim=None, **kwargs, ): super().__init__() self.in_dim = in_dim self.out_dim = out_dim self.use_mdn = use_mdn self.in_ph_start_idx = in_ph_start_idx self.in_ph_end_idx = in_ph_end_idx self.num_vocab = in_ph_end_idx - in_ph_start_idx self.embed_dim = embed_dim if "dropout" in kwargs: warn( "dropout argument in Conv1dResnet is deprecated" " and will be removed in future versions" ) if self.embed_dim is not None: assert in_dim > self.num_vocab self.emb = nn.Embedding(self.num_vocab, embed_dim) self.fc_in = nn.Linear(in_dim - self.num_vocab, embed_dim) conv_in_dim = embed_dim else: conv_in_dim = in_dim model = [ nn.ReflectionPad1d(3), WNConv1d(conv_in_dim, hidden_dim, kernel_size=7, padding=0), ] for n in range(num_layers): model.append(ResnetBlock(hidden_dim, dilation=2 ** n)) last_conv_out_dim = hidden_dim if use_mdn else out_dim model += [ nn.LeakyReLU(0.2), nn.ReflectionPad1d(3), WNConv1d(hidden_dim, last_conv_out_dim, kernel_size=7, padding=0), ] self.model = nn.Sequential(*model) if self.use_mdn: self.mdn_layer = MDNLayer( in_dim=hidden_dim, out_dim=out_dim, num_gaussians=num_gaussians, dim_wise=dim_wise, ) else: self.mdn_layer = None init_weights(self, init_type) def prediction_type(self): return ( PredictionType.PROBABILISTIC if self.use_mdn else PredictionType.DETERMINISTIC ) def forward(self, x, lengths=None, y=None): """Forward step Args: x (torch.Tensor): the input tensor lengths (torch.Tensor): the lengths of the input tensor y (torch.Tensor): the optional target tensor Returns: torch.Tensor: the output tensor """ if self.embed_dim is not None: x_first, x_ph_onehot, x_last = torch.split( x, [ self.in_ph_start_idx, self.num_vocab, self.in_dim - self.num_vocab - self.in_ph_start_idx, ], dim=-1, ) x_ph = torch.argmax(x_ph_onehot, dim=-1) # Make sure to have one-hot vector assert (x_ph_onehot.sum(-1) <= 1).all() x = self.emb(x_ph) + self.fc_in(torch.cat([x_first, x_last], dim=-1)) out = self.model(x.transpose(1, 2)).transpose(1, 2) if self.use_mdn: return self.mdn_layer(out) else: return out def inference(self, x, lengths=None): """Inference step Find the most likely mean and variance if use_mdn is True Args: x (torch.Tensor): the input tensor lengths (torch.Tensor): the lengths of the input tensor Returns: tuple: mean and variance of the output features """ if self.use_mdn: log_pi, log_sigma, mu = self(x, lengths) sigma, mu = mdn_get_most_probable_sigma_and_mu(log_pi, log_sigma, mu) return mu, sigma else: return self(x, lengths) @torch.no_grad() def _shallow_ar_inference(out, stream_sizes, analysis_filts): from torchaudio.functional import lfilter out_streams = split_streams(out, stream_sizes) # back to conv1d friendly (B, C, T) format out_streams = map(lambda x: x.transpose(1, 2), out_streams) out_syn = [] for sidx, os in enumerate(out_streams): out_stream_syn = torch.zeros_like(os) a = analysis_filts[sidx].get_filt_coefs() # apply IIR filter for each dimiesion for idx in range(os.shape[1]): # NOTE: scipy.signal.lfilter accespts b, a in order, # but torchaudio expect the oppsite; a, b in order ai = a[idx].view(-1).flip(0) bi = torch.zeros_like(ai) bi[0] = 1 out_stream_syn[:, idx, :] = lfilter(os[:, idx, :], ai, bi, clamp=False) out_syn += [out_stream_syn] out_syn = torch.cat(out_syn, 1) return out_syn.transpose(1, 2) class Conv1dResnetSAR(Conv1dResnet): """Conv1dResnet with shallow AR structure Args: in_dim (int): the dimension of the input hidden_dim (int): the dimension of the hidden state out_dim (int): the dimension of the output num_layers (int): the number of layers stream_sizes (list): Stream sizes ar_orders (list): Filter dimensions for each stream. init_type (str): the type of weight initialization """ def __init__( self, in_dim, hidden_dim, out_dim, num_layers=4, stream_sizes=None, ar_orders=None, init_type="none", **kwargs, ): super().__init__( in_dim=in_dim, hidden_dim=hidden_dim, out_dim=out_dim, num_layers=num_layers ) if "dropout" in kwargs: warn( "dropout argument in Conv1dResnetSAR is deprecated" " and will be removed in future versions" ) if stream_sizes is None: stream_sizes = [180, 3, 1, 15] if ar_orders is None: ar_orders = [20, 200, 20, 20] self.stream_sizes = stream_sizes init_weights(self, init_type) self.analysis_filts = nn.ModuleList() for s, K in zip(stream_sizes, ar_orders): self.analysis_filts += [TrTimeInvFIRFilter(s, K + 1)] def preprocess_target(self, y): assert sum(self.stream_sizes) == y.shape[-1] ys = split_streams(y, self.stream_sizes) for idx, yi in enumerate(ys): ys[idx] = self.analysis_filts[idx](yi.transpose(1, 2)).transpose(1, 2) return torch.cat(ys, -1) def inference(self, x, lengths=None, y=None): out = self.model(x.transpose(1, 2)).transpose(1, 2) return _shallow_ar_inference(out, self.stream_sizes, self.analysis_filts) class FFN(BaseModel): """Feed-forward network Args: in_dim (int): the dimension of the input hidden_dim (int): the dimension of the hidden state out_dim (int): the dimension of the output num_layers (int): the number of layers dropout (float): dropout rate init_type (str): the type of weight initialization last_sigmoid (bool): whether to apply sigmoid on the output """ def __init__( self, in_dim, hidden_dim, out_dim, num_layers=2, dropout=0.0, init_type="none", last_sigmoid=False, ): super(FFN, self).__init__() self.first_linear = nn.Linear(in_dim, hidden_dim) self.hidden_layers = nn.ModuleList( [nn.Linear(hidden_dim, hidden_dim) for _ in range(num_layers)] ) self.last_linear = nn.Linear(hidden_dim, out_dim) self.relu = nn.ReLU() self.dropout = nn.Dropout(dropout) self.last_sigmoid = last_sigmoid init_weights(self, init_type) def forward(self, x, lengths=None, y=None): """Forward step Args: x (torch.Tensor): the input tensor lengths (torch.Tensor): the lengths of the input tensor y (torch.Tensor): the optional target tensor Returns: torch.Tensor: the output tensor """ h = self.relu(self.first_linear(x)) for hl in self.hidden_layers: h = self.dropout(self.relu(hl(h))) out = self.last_linear(h) out = torch.sigmoid(out) if self.last_sigmoid else out return out # For compatibility FeedForwardNet = FFN class LSTMRNN(BaseModel): """LSTM-based recurrent neural network Args: in_dim (int): the dimension of the input hidden_dim (int): the dimension of the hidden state out_dim (int): the dimension of the output num_layers (int): the number of layers bidirectional (bool): whether to use bidirectional LSTM dropout (float): dropout rate init_type (str): the type of weight initialization """ def __init__( self, in_dim, hidden_dim, out_dim, num_layers=1, bidirectional=True, dropout=0.0, init_type="none", ): super(LSTMRNN, self).__init__() self.hidden_dim = hidden_dim self.num_layers = num_layers self.num_direction = 2 if bidirectional else 1 self.lstm = nn.LSTM( in_dim, hidden_dim, num_layers, bidirectional=bidirectional, batch_first=True, dropout=dropout, ) self.hidden2out = nn.Linear(self.num_direction * self.hidden_dim, out_dim) init_weights(self, init_type) def forward(self, x, lengths, y=None): """Forward step Args: x (torch.Tensor): the input tensor lengths (torch.Tensor): the lengths of the input tensor y (torch.Tensor): the optional target tensor Returns: torch.Tensor: the output tensor """ if isinstance(lengths, torch.Tensor): lengths = lengths.to("cpu") x = pack_padded_sequence(x, lengths, batch_first=True) out, _ = self.lstm(x) out, _ = pad_packed_sequence(out, batch_first=True) out = self.hidden2out(out) return out class LSTMRNNSAR(LSTMRNN): """LSTM-RNN with shallow AR structure Args: in_dim (int): the dimension of the input hidden_dim (int): the dimension of the hidden state out_dim (int): the dimension of the output num_layers (int): the number of layers bidirectional (bool): whether to use bidirectional LSTM dropout (float): dropout rate stream_sizes (list): Stream sizes ar_orders (list): Filter dimensions for each stream. init_type (str): the type of weight initialization """ def __init__( self, in_dim, hidden_dim, out_dim, num_layers=1, bidirectional=True, dropout=0.0, stream_sizes=None, ar_orders=None, init_type="none", ): super().__init__( in_dim, hidden_dim, out_dim, num_layers, bidirectional, dropout, init_type ) if stream_sizes is None: stream_sizes = [180, 3, 1, 15] if ar_orders is None: ar_orders = [20, 200, 20, 20] self.stream_sizes = stream_sizes self.analysis_filts = nn.ModuleList() for s, K in zip(stream_sizes, ar_orders): self.analysis_filts += [TrTimeInvFIRFilter(s, K + 1)] def preprocess_target(self, y): assert sum(self.stream_sizes) == y.shape[-1] ys = split_streams(y, self.stream_sizes) for idx, yi in enumerate(ys): ys[idx] = self.analysis_filts[idx](yi.transpose(1, 2)).transpose(1, 2) return torch.cat(ys, -1) def inference(self, x, lengths=None, y=None): out = self.forward(x, lengths) return _shallow_ar_inference(out, self.stream_sizes, self.analysis_filts) class RMDN(BaseModel): """RNN-based mixture density networks (MDN) Args: in_dim (int): the dimension of the input hidden_dim (int): the dimension of the hidden state out_dim (int): the dimension of the output num_layers (int): the number of layers bidirectional (bool): whether to use bidirectional LSTM dropout (float): dropout rate num_gaussians (int): the number of gaussians dim_wise (bool): whether to use dimension-wise or not init_type (str): the type of weight initialization """ def __init__( self, in_dim, hidden_dim, out_dim, num_layers=1, bidirectional=True, dropout=0.0, num_gaussians=8, dim_wise=False, init_type="none", ): super(RMDN, self).__init__() self.linear = nn.Linear(in_dim, hidden_dim) self.relu = nn.ReLU() self.num_direction = 2 if bidirectional else 1 self.lstm = nn.LSTM( hidden_dim, hidden_dim, num_layers, bidirectional=bidirectional, batch_first=True, dropout=dropout, ) self.mdn = MDNLayer( in_dim=self.num_direction * hidden_dim, out_dim=out_dim, num_gaussians=num_gaussians, dim_wise=dim_wise, ) init_weights(self, init_type) def prediction_type(self): return PredictionType.PROBABILISTIC def forward(self, x, lengths, y=None): """Forward step Args: x (torch.Tensor): the input tensor lengths (torch.Tensor): the lengths of the input tensor y (torch.Tensor): the optional target tensor Returns: torch.Tensor: the output tensor """ if isinstance(lengths, torch.Tensor): lengths = lengths.to("cpu") out = self.linear(x) sequence = pack_padded_sequence(self.relu(out), lengths, batch_first=True) out, _ = self.lstm(sequence) out, _ = pad_packed_sequence(out, batch_first=True) out = self.mdn(out) return out def inference(self, x, lengths=None): """Inference step Find the most likely mean and variance Args: x (torch.Tensor): the input tensor lengths (torch.Tensor): the lengths of the input tensor Returns: tuple: mean and variance of the output features """ log_pi, log_sigma, mu = self.forward(x, lengths) sigma, mu = mdn_get_most_probable_sigma_and_mu(log_pi, log_sigma, mu) return mu, sigma class MDN(BaseModel): """Mixture density networks (MDN) with FFN .. warning:: It is recommended to use MDNv2 instead, unless you want to fine-turn from a old checkpoint of MDN. Args: in_dim (int): the dimension of the input hidden_dim (int): the dimension of the hidden state out_dim (int): the dimension of the output num_layers (int): the number of layers num_gaussians (int): the number of gaussians dim_wise (bool): whether to use dimension-wise or not init_type (str): the type of weight initialization """ def __init__( self, in_dim, hidden_dim, out_dim, num_layers=1, num_gaussians=8, dim_wise=False, init_type="none", **kwargs, ): super(MDN, self).__init__() if "dropout" in kwargs: warn( "dropout argument in MDN is deprecated" " and will be removed in future versions" ) model = [nn.Linear(in_dim, hidden_dim), nn.ReLU()] if num_layers > 1: for _ in range(num_layers - 1): model += [nn.Linear(hidden_dim, hidden_dim), nn.ReLU()] model += [ MDNLayer( in_dim=hidden_dim, out_dim=out_dim, num_gaussians=num_gaussians, dim_wise=dim_wise, ) ] self.model = nn.Sequential(*model) init_weights(self, init_type) def prediction_type(self): return PredictionType.PROBABILISTIC def forward(self, x, lengths=None, y=None): """Forward step Args: x (torch.Tensor): the input tensor lengths (torch.Tensor): the lengths of the input tensor y (torch.Tensor): the optional target tensor Returns: torch.Tensor: the output tensor """ return self.model(x) def inference(self, x, lengths=None): """Inference step Find the most likely mean and variance Args: x (torch.Tensor): the input tensor lengths (torch.Tensor): the lengths of the input tensor Returns: tuple: mean and variance of the output features """ log_pi, log_sigma, mu = self.forward(x, lengths) sigma, mu = mdn_get_most_probable_sigma_and_mu(log_pi, log_sigma, mu) return mu, sigma class MDNv2(BaseModel): """Mixture density networks (MDN) with FFN MDN (v1) + Dropout Args: in_dim (int): the dimension of the input hidden_dim (int): the dimension of the hidden state out_dim (int): the dimension of the output num_layers (int): the number of layers dropout (float): dropout rate num_gaussians (int): the number of gaussians dim_wise (bool): whether to use dimension-wise or not init_type (str): the type of weight initialization """ def __init__( self, in_dim, hidden_dim, out_dim, num_layers=1, dropout=0.5, num_gaussians=8, dim_wise=False, init_type="none", ): super(MDNv2, self).__init__() model = [nn.Linear(in_dim, hidden_dim), nn.ReLU(), nn.Dropout(dropout)] if num_layers > 1: for _ in range(num_layers - 1): model += [ nn.Linear(hidden_dim, hidden_dim), nn.ReLU(), nn.Dropout(dropout), ] model += [ MDNLayer( in_dim=hidden_dim, out_dim=out_dim, num_gaussians=num_gaussians, dim_wise=dim_wise, ) ] self.model = nn.Sequential(*model) init_weights(self, init_type) def prediction_type(self): return PredictionType.PROBABILISTIC def forward(self, x, lengths=None, y=None): """Forward step Args: x (torch.Tensor): the input tensor lengths (torch.Tensor): the lengths of the input tensor y (torch.Tensor): the optional target tensor Returns: torch.Tensor: the output tensor """ return self.model(x) def inference(self, x, lengths=None): """Inference step Find the most likely mean and variance Args: x (torch.Tensor): the input tensor lengths (torch.Tensor): the lengths of the input tensor Returns: tuple: mean and variance of the output features """ log_pi, log_sigma, mu = self.forward(x, lengths) sigma, mu = mdn_get_most_probable_sigma_and_mu(log_pi, log_sigma, mu) return mu, sigma class Conv1dResnetMDN(BaseModel): """Conv1dResnet with MDN output layer .. warning:: Will be removed in v0.1.0. Use Conv1dResNet with ``use_mdn=True`` instead. """ def __init__( self, in_dim, hidden_dim, out_dim, num_layers=4, num_gaussians=8, dim_wise=False, init_type="none", **kwargs, ): super().__init__() if "dropout" in kwargs: warn( "dropout argument in Conv1dResnet is deprecated" " and will be removed in future versions" ) model = [ Conv1dResnet( in_dim=in_dim, hidden_dim=hidden_dim, out_dim=hidden_dim, num_layers=num_layers, ), nn.ReLU(), MDNLayer( in_dim=hidden_dim, out_dim=out_dim, num_gaussians=num_gaussians, dim_wise=dim_wise, ), ] self.model = nn.Sequential(*model) init_weights(self, init_type) def prediction_type(self): return PredictionType.PROBABILISTIC def forward(self, x, lengths=None, y=None): """Forward step Args: x (torch.Tensor): the input tensor lengths (torch.Tensor): the lengths of the input tensor y (torch.Tensor): the optional target tensor Returns: torch.Tensor: the output tensor """ return self.model(x) def inference(self, x, lengths=None): """Inference step Find the most likely mean and variance Args: x (torch.Tensor): the input tensor lengths (torch.Tensor): the lengths of the input tensor Returns: tuple: mean and variance of the output features """ log_pi, log_sigma, mu = self.forward(x, lengths) sigma, mu = mdn_get_most_probable_sigma_and_mu(log_pi, log_sigma, mu) return mu, sigma class FFConvLSTM(BaseModel): """FFN + Conv1d + LSTM A model proposed in :cite:t:`hono2021sinsy` without residual F0 prediction. Args: in_dim (int): the dimension of the input ff_hidden_dim (int): the dimension of the hidden state of the FFN conv_hidden_dim (int): the dimension of the hidden state of the conv1d lstm_hidden_dim (int): the dimension of the hidden state of the LSTM out_dim (int): the dimension of the output dropout (float): dropout rate num_lstm_layers (int): the number of layers of the LSTM bidirectional (bool): whether to use bidirectional LSTM init_type (str): the type of weight initialization use_mdn (bool): whether to use MDN or not dim_wise (bool): whether to use dimension-wise or not num_gaussians (int): the number of gaussians in_ph_start_idx (int): the start index of phoneme identity in a hed file in_ph_end_idx (int): the end index of phoneme identity in a hed file embed_dim (int): the dimension of the phoneme embedding """ def __init__( self, in_dim, ff_hidden_dim=2048, conv_hidden_dim=1024, lstm_hidden_dim=256, out_dim=67, dropout=0.0, num_lstm_layers=2, bidirectional=True, init_type="none", use_mdn=False, dim_wise=True, num_gaussians=4, in_ph_start_idx: int = 1, in_ph_end_idx: int = 50, embed_dim=None, ): super().__init__() self.in_dim = in_dim self.out_dim = out_dim self.in_ph_start_idx = in_ph_start_idx self.in_ph_end_idx = in_ph_end_idx self.num_vocab = in_ph_end_idx - in_ph_start_idx self.embed_dim = embed_dim self.use_mdn = use_mdn if self.embed_dim is not None: assert in_dim > self.num_vocab self.emb = nn.Embedding(self.num_vocab, embed_dim) self.fc_in = nn.Linear(in_dim - self.num_vocab, embed_dim) ff_in_dim = embed_dim else: ff_in_dim = in_dim self.ff = nn.Sequential( nn.Linear(ff_in_dim, ff_hidden_dim), nn.ReLU(), nn.Linear(ff_hidden_dim, ff_hidden_dim), nn.ReLU(), nn.Linear(ff_hidden_dim, ff_hidden_dim), nn.ReLU(), ) self.conv = nn.Sequential( nn.ReflectionPad1d(3), nn.Conv1d(ff_hidden_dim, conv_hidden_dim, kernel_size=7, padding=0), nn.BatchNorm1d(conv_hidden_dim), nn.ReLU(), nn.ReflectionPad1d(3), nn.Conv1d(conv_hidden_dim, conv_hidden_dim, kernel_size=7, padding=0), nn.BatchNorm1d(conv_hidden_dim), nn.ReLU(), nn.ReflectionPad1d(3), nn.Conv1d(conv_hidden_dim, conv_hidden_dim, kernel_size=7, padding=0), nn.BatchNorm1d(conv_hidden_dim), nn.ReLU(), ) num_direction = 2 if bidirectional else 1 self.lstm = nn.LSTM( conv_hidden_dim, lstm_hidden_dim, num_lstm_layers, bidirectional=True, batch_first=True, dropout=dropout, ) last_in_dim = num_direction * lstm_hidden_dim if self.use_mdn: assert dim_wise self.fc = MDNLayer( in_dim=last_in_dim, out_dim=out_dim, num_gaussians=num_gaussians, dim_wise=dim_wise, ) else: self.fc = nn.Linear(last_in_dim, out_dim) init_weights(self, init_type) def prediction_type(self): return ( PredictionType.PROBABILISTIC if self.use_mdn else PredictionType.DETERMINISTIC ) def forward(self, x, lengths=None, y=None): if isinstance(lengths, torch.Tensor): lengths = lengths.to("cpu") if self.embed_dim is not None: x_first, x_ph_onehot, x_last = torch.split( x, [ self.in_ph_start_idx, self.num_vocab, self.in_dim - self.num_vocab - self.in_ph_start_idx, ], dim=-1, ) x_ph = torch.argmax(x_ph_onehot, dim=-1) # Make sure to have one-hot vector assert (x_ph_onehot.sum(-1) <= 1).all() x = self.emb(x_ph) + self.fc_in(torch.cat([x_first, x_last], dim=-1)) out = self.ff(x) out = self.conv(out.transpose(1, 2)).transpose(1, 2) sequence = pack_padded_sequence(out, lengths, batch_first=True) out, _ = self.lstm(sequence) out, _ = pad_packed_sequence(out, batch_first=True) return self.fc(out) def inference(self, x, lengths=None): if self.use_mdn: log_pi, log_sigma, mu = self(x, lengths) sigma, mu = mdn_get_most_probable_sigma_and_mu(log_pi, log_sigma, mu) return mu, sigma else: return self(x, lengths) class VariancePredictor(BaseModel): """Variance predictor in :cite:t:`ren2020fastspeech`. The model is composed of stacks of Conv1d + ReLU + LayerNorm layers. The model can be used for duration or pitch prediction. Args: in_dim (int): the input dimension out_dim (int): the output dimension num_layers (int): the number of layers hidden_dim (int): the hidden dimension kernel_size (int): the kernel size dropout (float): the dropout rate init_type (str): the initialization type use_mdn (bool): whether to use MDN or not num_gaussians (int): the number of gaussians dim_wise (bool): whether to use dim-wise or not in_ph_start_idx (int): the start index of phoneme identity in a hed file in_ph_end_idx (int): the end index of phoneme identity in a hed file embed_dim (int): the dimension of the phoneme embedding mask_indices (list): the input feature indices to be masked. e.g., specify pitch_idx to mask pitch features. """ def __init__( self, in_dim, out_dim, num_layers=5, hidden_dim=256, kernel_size=5, dropout=0.5, init_type="none", use_mdn=False, num_gaussians=1, dim_wise=False, in_ph_start_idx: int = 1, in_ph_end_idx: int = 50, embed_dim=None, mask_indices=None, ): super().__init__() self.in_dim = in_dim self.out_dim = out_dim self.use_mdn = use_mdn self.in_ph_start_idx = in_ph_start_idx self.in_ph_end_idx = in_ph_end_idx self.num_vocab = in_ph_end_idx - in_ph_start_idx self.embed_dim = embed_dim self.use_mdn = use_mdn self.mask_indices = mask_indices if self.embed_dim is not None: assert in_dim > self.num_vocab self.emb = nn.Embedding(self.num_vocab, embed_dim) self.fc_in = nn.Linear(in_dim - self.num_vocab, embed_dim) in_dim = embed_dim conv = nn.ModuleList() for idx in range(num_layers): in_channels = in_dim if idx == 0 else hidden_dim conv += [ nn.Sequential( nn.Conv1d( in_channels, hidden_dim, kernel_size, stride=1, padding=(kernel_size - 1) // 2, ), nn.ReLU(), LayerNorm(hidden_dim, dim=1), nn.Dropout(dropout), ) ] self.conv = nn.Sequential(*conv) if self.use_mdn: self.mdn_layer = MDNLayer( hidden_dim, out_dim, num_gaussians=num_gaussians, dim_wise=dim_wise ) else: self.fc = nn.Linear(hidden_dim, out_dim) init_weights(self, init_type) def prediction_type(self): return ( PredictionType.PROBABILISTIC if self.use_mdn else PredictionType.DETERMINISTIC ) def forward(self, x, lengths=None, y=None): """Forward step Args: x (torch.Tensor): the input tensor lengths (torch.Tensor): the lengths of the input tensor y (torch.Tensor): the optional target tensor Returns: torch.Tensor: the output tensor """ # Masking specified features if self.mask_indices is not None: for idx in self.mask_indices: x[:, :, idx] *= 0.0 if self.embed_dim is not None: x_first, x_ph_onehot, x_last = torch.split( x, [ self.in_ph_start_idx, self.num_vocab, self.in_dim - self.num_vocab - self.in_ph_start_idx, ], dim=-1, ) x_ph = torch.argmax(x_ph_onehot, dim=-1) # Make sure to have one-hot vector assert (x_ph_onehot.sum(-1) <= 1).all() x = self.emb(x_ph) + self.fc_in(torch.cat([x_first, x_last], dim=-1)) out = self.conv(x.transpose(1, 2)).transpose(1, 2) if self.use_mdn: return self.mdn_layer(out) else: return self.fc(out) def inference(self, x, lengths=None): """Inference step Find the most likely mean and variance if use_mdn is True Args: x (torch.Tensor): the input tensor lengths (torch.Tensor): the lengths of the input tensor Returns: tuple: mean and variance of the output features """ if self.use_mdn: log_pi, log_sigma, mu = self(x, lengths) sigma, mu = mdn_get_most_probable_sigma_and_mu(log_pi, log_sigma, mu) return mu, sigma else: return self(x, lengths) class LSTMEncoder(BaseModel): """LSTM encoder A simple LSTM-based encoder Args: in_dim (int): the input dimension hidden_dim (int): the hidden dimension out_dim (int): the output dimension num_layers (int): the number of layers bidirectional (bool): whether to use bidirectional or not dropout (float): the dropout rate init_type (str): the initialization type in_ph_start_idx (int): the start index of phonetic context in a hed file in_ph_end_idx (int): the end index of phonetic context in a hed file embed_dim (int): the embedding dimension """ def __init__( self, in_dim: int, hidden_dim: int, out_dim: int, num_layers: int = 1, bidirectional: bool = True, dropout: float = 0.0, init_type: str = "none", in_ph_start_idx: int = 1, in_ph_end_idx: int = 50, embed_dim=None, ): super(LSTMEncoder, self).__init__() self.in_dim = in_dim self.in_ph_start_idx = in_ph_start_idx self.in_ph_end_idx = in_ph_end_idx self.num_vocab = in_ph_end_idx - in_ph_start_idx self.embed_dim = embed_dim if self.embed_dim is not None: assert in_dim > self.num_vocab self.emb = nn.Embedding(self.num_vocab, embed_dim) self.fc_in = nn.Linear(in_dim - self.num_vocab, embed_dim) lstm_in_dim = embed_dim else: lstm_in_dim = in_dim self.num_layers = num_layers num_direction = 2 if bidirectional else 1 self.lstm = nn.LSTM( lstm_in_dim, hidden_dim, num_layers, bidirectional=bidirectional, batch_first=True, dropout=dropout, ) self.hidden2out = nn.Linear(num_direction * hidden_dim, out_dim) init_weights(self, init_type) def forward(self, x, lengths, y=None): if self.embed_dim is not None: x_first, x_ph_onehot, x_last = torch.split( x, [ self.in_ph_start_idx, self.num_vocab, self.in_dim - self.num_vocab - self.in_ph_start_idx, ], dim=-1, ) x_ph = torch.argmax(x_ph_onehot, dim=-1) # Make sure to have one-hot vector assert (x_ph_onehot.sum(-1) <= 1).all() x = self.emb(x_ph) + self.fc_in(torch.cat([x_first, x_last], dim=-1)) if isinstance(lengths, torch.Tensor): lengths = lengths.to("cpu") x = pack_padded_sequence(x, lengths, batch_first=True) out, _ = self.lstm(x) out, _ = pad_packed_sequence(out, batch_first=True) out = self.hidden2out(out) return out class TransformerEncoder(BaseModel): """Transformer encoder .. warning:: So far this is not well tested. Maybe be removed in the future. Args: in_dim (int): the input dimension out_dim (int): the output dimension hidden_dim (int): the hidden dimension attention_dim (int): the attention dimension num_heads (int): the number of heads num_layers (int): the number of layers kernel_size (int): the kernel size dropout (float): the dropout rate reduction_factor (int): the reduction factor init_type (str): the initialization type downsample_by_conv (bool): whether to use convolutional downsampling or not in_ph_start_idx (int): the start index of phonetic context in a hed file in_ph_end_idx (int): the end index of phonetic context in a hed file embed_dim (int): the embedding dimension """ def __init__( self, in_dim, out_dim, hidden_dim, attention_dim, num_heads=2, num_layers=2, kernel_size=3, dropout=0.1, reduction_factor=1, init_type="none", downsample_by_conv=False, in_ph_start_idx: int = 1, in_ph_end_idx: int = 50, embed_dim=None, ): super().__init__() self.in_dim = in_dim self.out_dim = out_dim self.in_ph_start_idx = in_ph_start_idx self.in_ph_end_idx = in_ph_end_idx self.num_vocab = in_ph_end_idx - in_ph_start_idx self.embed_dim = embed_dim if self.embed_dim is not None: assert in_dim > self.num_vocab self.emb = nn.Embedding(self.num_vocab, embed_dim) self.fc_in = nn.Linear(in_dim - self.num_vocab, embed_dim) self.fc = nn.Linear(embed_dim, hidden_dim) else: self.emb = None self.fc_in = None self.fc = nn.Linear(in_dim, hidden_dim) self.reduction_factor = reduction_factor self.encoder = _TransformerEncoder( hidden_channels=hidden_dim, filter_channels=attention_dim, n_heads=num_heads, n_layers=num_layers, kernel_size=kernel_size, p_dropout=dropout, ) self.fc_out = nn.Linear(hidden_dim, out_dim * reduction_factor) if reduction_factor > 1 and downsample_by_conv: self.conv_downsample = nn.Conv1d( in_dim, in_dim, kernel_size=reduction_factor, stride=reduction_factor, groups=in_dim, ) else: self.conv_downsample = None for f in [self.fc_in, self.emb, self.fc, self.fc_out]: if f is not None: init_weights(f, init_type) def forward(self, x, lengths=None, y=None): """Forward pass Args: x (torch.Tensor): input tensor lengths (torch.Tensor): input sequence lengths y (torch.Tensor): target tensor (optional) Returns: torch.Tensor: output tensor """ if isinstance(lengths, list): lengths = torch.tensor(lengths).to(x.device) if self.embed_dim is not None: x_first, x_ph_onehot, x_last = torch.split( x, [ self.in_ph_start_idx, self.num_vocab, self.in_dim - self.num_vocab - self.in_ph_start_idx, ], dim=-1, ) x_ph = torch.argmax(x_ph_onehot, dim=-1) # Make sure to have one-hot vector assert (x_ph_onehot.sum(-1) <= 1).all() x = self.emb(x_ph) + self.fc_in(torch.cat([x_first, x_last], dim=-1)) # Adjust lengths based on the reduction factor if self.reduction_factor > 1: lengths = (lengths / self.reduction_factor).long() if self.conv_downsample is not None: x = self.conv_downsample(x.transpose(1, 2)).transpose(1, 2) else: x = x[:, self.reduction_factor - 1 :: self.reduction_factor] x = self.fc(x) # (B, T, C) -> (B, C, T) x = x.transpose(1, 2) x_mask = sequence_mask(lengths, x.shape[2]).unsqueeze(1).to(x.device) x = self.encoder(x * x_mask, x_mask) # (B, C, T) -> (B, T, C) x = self.fc_out(x.transpose(1, 2)).view(x.shape[0], -1, self.out_dim) return x # For backward compatibility # Will be removed in v0.1.0 def ResF0Conv1dResnet(*args, **kwargs): from nnsvs.acoustic_models import ResF0Conv1dResnet return ResF0Conv1dResnet(*args, **kwargs) def ResF0Conv1dResnetMDN(*args, **kwargs): from nnsvs.acoustic_models import ResF0Conv1dResnetMDN return ResF0Conv1dResnetMDN(*args, **kwargs) def ResF0VariancePredictor(*args, **kwargs): from nnsvs.acoustic_models import ResF0VariancePredictor return ResF0VariancePredictor(*args, **kwargs) def ResSkipF0FFConvLSTM(*args, **kwargs): from nnsvs.acoustic_models import ResSkipF0FFConvLSTM return ResSkipF0FFConvLSTM(*args, **kwargs)
40,986
30.577042
89
py
nnsvs
nnsvs-master/nnsvs/gen.py
from warnings import warn import librosa import numpy as np import pyloudnorm as pyln import pysptk import pyworld import scipy import torch from nnmnkwii.frontend import merlin as fe from nnmnkwii.io import hts from nnmnkwii.postfilters import merlin_post_filter from nnmnkwii.preprocessing.f0 import interp1d from nnsvs.base import PredictionType from nnsvs.dsp import bandpass_filter from nnsvs.io.hts import ( get_note_frame_indices, get_note_indices, get_pitch_index, get_pitch_indices, ) from nnsvs.multistream import ( get_static_stream_sizes, get_windows, multi_stream_mlpg, split_streams, ) from nnsvs.pitch import gen_sine_vibrato, lowpass_filter from nnsvs.postfilters import variance_scaling from sklearn.preprocessing import MinMaxScaler def _midi_to_hz(x, idx, log_f0=False): z = np.zeros(len(x)) indices = x[:, idx] > 0 z[indices] = librosa.midi_to_hz(x[indices, idx]) if log_f0: z[indices] = np.log(z[indices]) return z def _is_silence(label): is_full_context = "@" in label if is_full_context: is_silence = "-sil" in label or "-pau" in label else: is_silence = label == "sil" or label == "pau" return is_silence @torch.no_grad() def predict_timelag( device, labels, timelag_model, timelag_config, timelag_in_scaler, timelag_out_scaler, binary_dict, numeric_dict, pitch_indices=None, log_f0_conditioning=True, allowed_range=None, allowed_range_rest=None, force_clip_input_features=False, frame_period=5, ): """Predict time-lag from HTS labels Args: device (torch.device): device labels (nnmnkwii.io.hts.HTSLabelFile): HTS-style labels timelag_model (nn.Module): time-lag model timelag_config (dict): time-lag model config timelag_in_scaler (sklearn.preprocessing.MinMaxScaler): input scaler timelag_out_scaler (sklearn.preprocessing.MinMaxScaler): output scaler binary_dict (dict): binary feature dict numeric_dict (dict): numeric feature dict pitch_indices (list): indices of pitch features log_f0_conditioning (bool): whether to condition on log f0 allowed_range (list): allowed range of time-lag allowed_range_rest (list): allowed range of time-lag for rest force_clip_input_features (bool): whether to clip input features Returns; ndarray: time-lag predictions """ hts_frame_shift = int(frame_period * 1e4) # make sure to set frame shift properly before calling round_ method labels.frame_shift = hts_frame_shift if pitch_indices is None: pitch_indices = get_pitch_indices(binary_dict, numeric_dict) if allowed_range is None: allowed_range = [-20, 20] if allowed_range_rest is None: allowed_range_rest = [-40, 40] # round start/end times just in case. labels.round_() # Extract note-level labels note_indices = get_note_indices(labels) note_labels = labels[note_indices] # Extract musical/linguistic context timelag_linguistic_features = fe.linguistic_features( note_labels, binary_dict, numeric_dict, add_frame_features=False, subphone_features=None, frame_shift=hts_frame_shift, ).astype(np.float32) # Adjust input features if we use log-f0 conditioning if log_f0_conditioning: if pitch_indices is None: raise ValueError("Pitch feature indices must be specified!") for idx in pitch_indices: timelag_linguistic_features[:, idx] = interp1d( _midi_to_hz(timelag_linguistic_features, idx, log_f0_conditioning), kind="slinear", ) # Normalization timelag_linguistic_features = timelag_in_scaler.transform( timelag_linguistic_features ) if force_clip_input_features and isinstance(timelag_in_scaler, MinMaxScaler): # clip to feature range (except for pitch-related features) non_pitch_indices = [ idx for idx in range(timelag_linguistic_features.shape[1]) if idx not in pitch_indices ] timelag_linguistic_features[:, non_pitch_indices] = np.clip( timelag_linguistic_features[:, non_pitch_indices], timelag_in_scaler.feature_range[0], timelag_in_scaler.feature_range[1], ) # Run model x = torch.from_numpy(timelag_linguistic_features).unsqueeze(0).to(device) # Run model if timelag_model.prediction_type() == PredictionType.PROBABILISTIC: # (B, T, D_out) max_mu, max_sigma = timelag_model.inference(x, [x.shape[1]]) if np.any(timelag_config.has_dynamic_features): # Apply denormalization # (B, T, D_out) -> (T, D_out) max_sigma_sq = ( max_sigma.squeeze(0).cpu().data.numpy() ** 2 * timelag_out_scaler.var_ ) max_sigma_sq = np.maximum(max_sigma_sq, 1e-14) max_mu = timelag_out_scaler.inverse_transform( max_mu.squeeze(0).cpu().data.numpy() ) # (T, D_out) -> (T, static_dim) pred_timelag = multi_stream_mlpg( max_mu, max_sigma_sq, get_windows(timelag_config.num_windows), timelag_config.stream_sizes, timelag_config.has_dynamic_features, ) else: # Apply denormalization pred_timelag = timelag_out_scaler.inverse_transform( max_mu.squeeze(0).cpu().data.numpy() ) else: # (T, D_out) pred_timelag = ( timelag_model.inference(x, [x.shape[1]]).squeeze(0).cpu().data.numpy() ) # Apply denormalization pred_timelag = timelag_out_scaler.inverse_transform(pred_timelag) if np.any(timelag_config.has_dynamic_features): # (T, D_out) -> (T, static_dim) pred_timelag = multi_stream_mlpg( pred_timelag, timelag_out_scaler.var_, get_windows(timelag_config.num_windows), timelag_config.stream_sizes, timelag_config.has_dynamic_features, ) # Rounding pred_timelag = np.round(pred_timelag) # Clip to the allowed range for idx in range(len(pred_timelag)): if _is_silence(note_labels.contexts[idx]): pred_timelag[idx] = np.clip( pred_timelag[idx], allowed_range_rest[0], allowed_range_rest[1] ) else: pred_timelag[idx] = np.clip( pred_timelag[idx], allowed_range[0], allowed_range[1] ) # frames -> 100 ns pred_timelag *= hts_frame_shift return pred_timelag @torch.no_grad() def predict_duration( device, labels, duration_model, duration_config, duration_in_scaler, duration_out_scaler, binary_dict, numeric_dict, pitch_indices=None, log_f0_conditioning=True, force_clip_input_features=False, frame_period=5, ): """Predict phoneme durations from HTS labels Args: device (torch.device): device to run the model on labels (nnmnkwii.io.hts.HTSLabelFile): labels duration_model (nn.Module): duration model duration_config (dict): duration config duration_in_scaler (sklearn.preprocessing.MinMaxScaler): duration input scaler duration_out_scaler (sklearn.preprocessing.MinMaxScaler): duration output scaler binary_dict (dict): binary feature dictionary numeric_dict (dict): numeric feature dictionary pitch_indices (list): indices of pitch features log_f0_conditioning (bool): whether to use log-f0 conditioning force_clip_input_features (bool): whether to clip input features Returns: np.ndarray: predicted durations """ hts_frame_shift = int(frame_period * 1e4) if pitch_indices is None: pitch_indices = get_pitch_indices(binary_dict, numeric_dict) # Extract musical/linguistic features duration_linguistic_features = fe.linguistic_features( labels, binary_dict, numeric_dict, add_frame_features=False, subphone_features=None, frame_shift=hts_frame_shift, ).astype(np.float32) if log_f0_conditioning: for idx in pitch_indices: duration_linguistic_features[:, idx] = interp1d( _midi_to_hz(duration_linguistic_features, idx, log_f0_conditioning), kind="slinear", ) # Apply normalization duration_linguistic_features = duration_in_scaler.transform( duration_linguistic_features ) if force_clip_input_features and isinstance(duration_in_scaler, MinMaxScaler): # clip to feature range (except for pitch-related features) non_pitch_indices = [ idx for idx in range(duration_linguistic_features.shape[1]) if idx not in pitch_indices ] duration_linguistic_features[:, non_pitch_indices] = np.clip( duration_linguistic_features[:, non_pitch_indices], duration_in_scaler.feature_range[0], duration_in_scaler.feature_range[1], ) # Apply model x = torch.from_numpy(duration_linguistic_features).float().to(device) x = x.view(1, -1, x.size(-1)) if duration_model.prediction_type() == PredictionType.PROBABILISTIC: # (B, T, D_out) max_mu, max_sigma = duration_model.inference(x, [x.shape[1]]) if np.any(duration_config.has_dynamic_features): raise RuntimeError( "Dynamic features are not supported for duration modeling" ) # Apply denormalization max_sigma_sq = ( max_sigma.squeeze(0).cpu().data.numpy() ** 2 * duration_out_scaler.var_ ) max_sigma_sq = np.maximum(max_sigma_sq, 1e-14) max_mu = duration_out_scaler.inverse_transform( max_mu.squeeze(0).cpu().data.numpy() ) return max_mu, max_sigma_sq else: # (T, D_out) pred_durations = ( duration_model.inference(x, [x.shape[1]]).squeeze(0).cpu().data.numpy() ) # Apply denormalization pred_durations = duration_out_scaler.inverse_transform(pred_durations) if np.any(duration_config.has_dynamic_features): # (T, D_out) -> (T, static_dim) pred_durations = multi_stream_mlpg( pred_durations, duration_out_scaler.var_, get_windows(duration_config.num_windows), duration_config.stream_sizes, duration_config.has_dynamic_features, ) pred_durations[pred_durations <= 0] = 1 pred_durations = np.round(pred_durations) return pred_durations def postprocess_duration(labels, pred_durations, lag, frame_period=5): """Post-process durations based on predicted time-lag Ref : https://arxiv.org/abs/2108.02776 Args: labels (HTSLabelFile): HTS labels pred_durations (array or tuple): predicted durations for non-MDN, mean and variance for MDN lag (array): predicted time-lag Returns: HTSLabelFile: labels with adjusted durations """ hts_frame_shift = int(frame_period * 1e4) note_indices = get_note_indices(labels) # append the end of note note_indices.append(len(labels)) is_mdn = isinstance(pred_durations, tuple) and len(pred_durations) == 2 output_labels = hts.HTSLabelFile() for i in range(1, len(note_indices)): p = labels[note_indices[i - 1] : note_indices[i]] # Compute note duration with time-lag # eq (11) L = int(fe.duration_features(p)[0]) if i < len(note_indices) - 1: L_hat = L - (lag[i - 1] - lag[i]) / hts_frame_shift else: L_hat = L - (lag[i - 1]) / hts_frame_shift # Prevent negative duration L_hat = max(L_hat, 1) # adjust the start time of the note p.start_times = np.minimum( np.asarray(p.start_times) + lag[i - 1].reshape(-1), np.asarray(p.end_times) - hts_frame_shift * len(p), ) p.start_times = np.maximum(p.start_times, 0) if len(output_labels) > 0: p.start_times = np.maximum( p.start_times, output_labels.start_times[-1] + hts_frame_shift ) # Compute normalized phoneme durations if is_mdn: mu = pred_durations[0][note_indices[i - 1] : note_indices[i]] sigma_sq = pred_durations[1][note_indices[i - 1] : note_indices[i]] # eq (17) rho = (L_hat - mu.sum()) / sigma_sq.sum() # eq (16) d_norm = mu + rho * sigma_sq if np.any(d_norm <= 0): # eq (12) (using mu as d_hat) s = frame_period * 0.001 print( f"Negative phoneme durations are predicted at {i}-th note. " "The note duration: ", f"{round(float(L)*s,3)} sec -> {round(float(L_hat)*s,3)} sec", ) print( "It's likely that the model couldn't predict correct durations " "for short notes." ) print( f"Variance scaling based durations (in frame):\n{(mu + rho * sigma_sq)}" ) print( f"Fallback to uniform scaling (in frame):\n{(L_hat * mu / mu.sum())}" ) d_norm = L_hat * mu / mu.sum() else: # eq (12) d_hat = pred_durations[note_indices[i - 1] : note_indices[i]] d_norm = L_hat * d_hat / d_hat.sum() d_norm = np.round(d_norm) d_norm[d_norm <= 0] = 1 p.set_durations(d_norm) if len(output_labels) > 0: output_labels.end_times[-1] = p.start_times[0] for n in p: output_labels.append(n) return output_labels @torch.no_grad() def predict_timing( device, labels, binary_dict, numeric_dict, timelag_model, timelag_config, timelag_in_scaler, timelag_out_scaler, duration_model, duration_config, duration_in_scaler, duration_out_scaler, log_f0_conditioning=True, allowed_range=None, allowed_range_rest=None, force_clip_input_features=True, frame_period=5, ): """Predict timinigs from HTS labels This is equivalent to ``predict_timelag + predict_duration + postprocess_duration``. Args: device (torch.device): device to run the model on labels (nnmnkwii.io.hts.HTSLabelFile): labels binary_dict (dict): binary feature dictionary numeric_dict (dict): numeric feature dictionary timelag_model (nn.Module): timelag model timelag_config (dict): timelag config timelag_in_scaler (sklearn.preprocessing.MinMaxScaler): timelag input scaler timelag_out_scaler (sklearn.preprocessing.MinMaxScaler): timelag output scaler duration_model (nn.Module): duration model duration_config (dict): duration config duration_in_scaler (sklearn.preprocessing.MinMaxScaler): duration input scaler duration_out_scaler (sklearn.preprocessing.MinMaxScaler): duration output scaler log_f0_conditioning (bool): whether to condition on log f0 allowed_range (list): allowed range of time-lag allowed_range_rest (list): allowed range of time-lag for rest force_clip_input_features (bool): whether to clip input features frame_period (int): frame period in milliseconds Returns: nnmnkwii.io.hts.HTSLabelFile: duration modified labels """ hts_frame_shift = int(frame_period * 1e4) labels.frame_shift = hts_frame_shift pitch_indices = get_pitch_indices(binary_dict, numeric_dict) # Time-lag lag = predict_timelag( device=device, labels=labels, timelag_model=timelag_model, timelag_config=timelag_config, timelag_in_scaler=timelag_in_scaler, timelag_out_scaler=timelag_out_scaler, binary_dict=binary_dict, numeric_dict=numeric_dict, pitch_indices=pitch_indices, log_f0_conditioning=log_f0_conditioning, allowed_range=allowed_range, allowed_range_rest=allowed_range_rest, force_clip_input_features=force_clip_input_features, frame_period=frame_period, ) # Duration predictions durations = predict_duration( device=device, labels=labels, duration_model=duration_model, duration_config=duration_config, duration_in_scaler=duration_in_scaler, duration_out_scaler=duration_out_scaler, binary_dict=binary_dict, numeric_dict=numeric_dict, pitch_indices=pitch_indices, log_f0_conditioning=log_f0_conditioning, force_clip_input_features=force_clip_input_features, frame_period=frame_period, ) # Normalize phoneme durations duration_modified_labels = postprocess_duration(labels, durations, lag) return duration_modified_labels @torch.no_grad() def predict_acoustic( device, labels, acoustic_model, acoustic_config, acoustic_in_scaler, acoustic_out_scaler, binary_dict, numeric_dict, subphone_features="coarse_coding", pitch_indices=None, log_f0_conditioning=True, force_clip_input_features=False, frame_period=5, f0_shift_in_cent=0, ): """Predict acoustic features from HTS labels MLPG is applied to the predicted features if the output features have dynamic features. Args: device (torch.device): device to use labels (HTSLabelFile): HTS labels acoustic_model (nn.Module): acoustic model acoustic_config (AcousticConfig): acoustic configuration acoustic_in_scaler (sklearn.preprocessing.StandardScaler): input scaler acoustic_out_scaler (sklearn.preprocessing.StandardScaler): output scaler binary_dict (dict): binary feature dictionary numeric_dict (dict): numeric feature dictionary subphone_features (str): subphone feature type pitch_indices (list): indices of pitch features log_f0_conditioning (bool): whether to use log f0 conditioning force_clip_input_features (bool): whether to force clip input features frame_period (float): frame period in msec f0_shift_in_cent (float): F0 shift in cent-scale before the inference Returns: ndarray: predicted acoustic features """ hts_frame_shift = int(frame_period * 1e4) if pitch_indices is None: pitch_indices = get_pitch_indices(binary_dict, numeric_dict) # Musical/linguistic features linguistic_features = fe.linguistic_features( labels, binary_dict, numeric_dict, add_frame_features=True, subphone_features=subphone_features, frame_shift=hts_frame_shift, ) if log_f0_conditioning: for idx in pitch_indices: linguistic_features[:, idx] = interp1d( _midi_to_hz(linguistic_features, idx, log_f0_conditioning), kind="slinear", ) if f0_shift_in_cent != 0: lf0_offset = f0_shift_in_cent * np.log(2) / 1200 linguistic_features[:, idx] += lf0_offset # Apply normalization linguistic_features = acoustic_in_scaler.transform(linguistic_features) if force_clip_input_features and isinstance(acoustic_in_scaler, MinMaxScaler): # clip to feature range (except for pitch-related features) non_pitch_indices = [ idx for idx in range(linguistic_features.shape[1]) if idx not in pitch_indices ] linguistic_features[:, non_pitch_indices] = np.clip( linguistic_features[:, non_pitch_indices], acoustic_in_scaler.feature_range[0], acoustic_in_scaler.feature_range[1], ) # Predict acoustic features x = torch.from_numpy(linguistic_features).float().to(device) x = x.view(1, -1, x.size(-1)) if acoustic_model.prediction_type() in [ PredictionType.PROBABILISTIC, PredictionType.MULTISTREAM_HYBRID, ]: # (B, T, D_out) max_mu, max_sigma = acoustic_model.inference(x, [x.shape[1]]) if np.any(acoustic_config.has_dynamic_features): # Apply denormalization # (B, T, D_out) -> (T, D_out) max_sigma_sq = ( max_sigma.squeeze(0).cpu().data.numpy() ** 2 * acoustic_out_scaler.var_ ) max_sigma_sq = np.maximum(max_sigma_sq, 1e-14) max_mu = acoustic_out_scaler.inverse_transform( max_mu.squeeze(0).cpu().data.numpy() ) # (T, D_out) -> (T, static_dim) pred_acoustic = multi_stream_mlpg( max_mu, max_sigma_sq, get_windows(acoustic_config.num_windows), acoustic_config.stream_sizes, acoustic_config.has_dynamic_features, ) else: # Apply denormalization pred_acoustic = acoustic_out_scaler.inverse_transform( max_mu.squeeze(0).cpu().data.numpy() ) else: # (T, D_out) pred_acoustic = ( acoustic_model.inference(x, [x.shape[1]]).squeeze(0).cpu().data.numpy() ) # Apply denormalization pred_acoustic = acoustic_out_scaler.inverse_transform(pred_acoustic) if np.any(acoustic_config.has_dynamic_features): # (T, D_out) -> (T, static_dim) pred_acoustic = multi_stream_mlpg( pred_acoustic, acoustic_out_scaler.var_, get_windows(acoustic_config.num_windows), acoustic_config.stream_sizes, acoustic_config.has_dynamic_features, ) return pred_acoustic @torch.no_grad() def postprocess_acoustic( device, acoustic_features, duration_modified_labels, binary_dict, numeric_dict, acoustic_config, acoustic_out_static_scaler, postfilter_model=None, postfilter_config=None, postfilter_out_scaler=None, sample_rate=48000, frame_period=5, relative_f0=False, feature_type="world", post_filter_type="gv", trajectory_smoothing=True, trajectory_smoothing_cutoff=50, trajectory_smoothing_cutoff_f0=20, vuv_threshold=0.5, f0_shift_in_cent=0, vibrato_scale=1.0, force_fix_vuv=False, fill_silence_to_rest=False, ): """Post-process acoustic features The function converts acoustic features in single ndarray to tuple of multi-stream acoustic features. e.g., array -> (mgc, lf0, vuv, bap) Args: device (torch.device): Device. duration_modified_labels (nnmnkwii.io.hts.HTSLabelFile): HTS label file. binary_dict (dict): Dictionary of binary features. numeric_dict (dict): Dictionary of numeric features. acoustic_config (dict): Acoustic model configuration. acoustic_features (np.ndarray): Acoustic features. acoustic_out_static_scaler (sklearn.preprocessing.StandardScaler): Scaler for acoustic features. postfilter_model (nn.Module): Post-filter model. postfilter_config (dict): Post-filter model configuration. postfilter_out_scaler (sklearn.preprocessing.StandardScaler): Scaler for post-filter. sample_rate (int): Sampling rate. frame_period (float): Frame period in milliseconds. relative_f0 (bool): If True, use relative f0. feature_type (str): Feature type. post_filter_type (str): Post-filter type. One of ``gv``, ``merlin`` or ``nnsvs``. Recommended to use ``gv`` for general purpose. trajectory_smoothing (bool): Whether to apply trajectory smoothing. trajectory_smoothing_cutoff (float): Cutoff frequency for trajectory smoothing of spectral features. trajectory_smoothing_cutoff_f0 (float): Cutoff frequency for trajectory smoothing of f0. vuv_threshold (float): V/UV threshold. f0_shift_in_cent (float): F0 shift in cents. vibrato_scale (float): Vibrato scale. force_fix_vuv (bool): If True, force to fix V/UV. fill_silence_to_rest (bool): Fill silence to rest frames. Returns: tuple: Post-processed acoustic features. """ hts_frame_shift = int(frame_period * 1e4) pitch_idx = get_pitch_index(binary_dict, numeric_dict) static_stream_sizes = get_static_stream_sizes( acoustic_config.stream_sizes, acoustic_config.has_dynamic_features, acoustic_config.num_windows, ) linguistic_features = fe.linguistic_features( duration_modified_labels, binary_dict, numeric_dict, add_frame_features=True, frame_shift=hts_frame_shift, ) # GV post-filter if post_filter_type == "gv" or ( post_filter_type == "nnsvs" and feature_type == "world" ): note_frame_indices = get_note_frame_indices( binary_dict, numeric_dict, linguistic_features ) if feature_type == "world": offset = 2 elif feature_type == "melf0": # NOTE: set offset so that post-filter does not affect F0 mel_freq = librosa.mel_frequencies( n_mels=80, fmin=63, fmax=sample_rate // 2 ) # NOTE: the threshold could be tuned for better performance offset = np.argmax(mel_freq > 1200) # NOTE: apply the post-filter for note frames only mgc_end_dim = static_stream_sizes[0] acoustic_features[:, :mgc_end_dim] = variance_scaling( acoustic_out_static_scaler.var_.reshape(-1)[:mgc_end_dim], acoustic_features[:, :mgc_end_dim], offset=offset, note_frame_indices=note_frame_indices, ) # Learned post-filter using nnsvs if post_filter_type == "nnsvs" and postfilter_model is not None: # (1) Raw spectrogram or (2) mgc rawsp_output = postfilter_config.stream_sizes[0] >= 128 # If the post-filter output is raw spectrogrma, convert mgc to log spectrogram if rawsp_output: outs = split_streams(acoustic_features, static_stream_sizes) assert len(outs) == 4 mgc, lf0, vuv, bap = outs fft_size = pyworld.get_cheaptrick_fft_size(sample_rate) sp = pyworld.decode_spectral_envelope( mgc.astype(np.float64), sample_rate, fft_size ).astype(np.float32) sp = np.log(sp) acoustic_features = np.concatenate([sp, lf0, vuv, bap], axis=-1) in_feats = torch.from_numpy(acoustic_features).float().unsqueeze(0) in_feats = postfilter_out_scaler.transform(in_feats).float().to(device) # Run inference out_feats = postfilter_model.inference(in_feats, [in_feats.shape[1]]) acoustic_features = ( postfilter_out_scaler.inverse_transform(out_feats.cpu()).squeeze(0).numpy() ) # Convert log spectrogram to mgc # NOTE: mgc is used to reduce possible artifacts # Ref: https://bit.ly/3AHjstU if rawsp_output: sp, lf0, vuv, bap = split_streams( acoustic_features, postfilter_config.stream_sizes ) sp = np.exp(sp) mgc = pyworld.code_spectral_envelope( sp.astype(np.float64), sample_rate, 60 ).astype(np.float32) acoustic_features = np.concatenate([mgc, lf0, vuv, bap], axis=-1) # Generate WORLD parameters if feature_type == "world": mgc, lf0, vuv, bap = gen_spsvs_static_features( labels=duration_modified_labels, acoustic_features=acoustic_features, binary_dict=binary_dict, numeric_dict=numeric_dict, stream_sizes=acoustic_config.stream_sizes, has_dynamic_features=acoustic_config.has_dynamic_features, pitch_idx=pitch_idx, num_windows=acoustic_config.num_windows, frame_period=frame_period, relative_f0=relative_f0, vibrato_scale=vibrato_scale, vuv_threshold=vuv_threshold, force_fix_vuv=force_fix_vuv, ) elif feature_type == "melf0": mel, lf0, vuv = split_streams(acoustic_features, [80, 1, 1]) else: raise ValueError(f"Unknown feature type: {feature_type}") if fill_silence_to_rest: mask = _get_nonrest_frame_soft_mask( binary_dict, numeric_dict, linguistic_features ) if feature_type == "world": mgc, lf0, vuv, bap = _fill_silence_to_world_params(mgc, lf0, vuv, bap, mask) elif feature_type == "melf0": mel, lf0, vuv = _fill_silence_to_mel_params(mel, lf0, vuv, mask) if f0_shift_in_cent != 0: lf0_offset = f0_shift_in_cent * np.log(2) / 1200 lf0 = lf0 + lf0_offset # NOTE: spectral enhancement based on the Merlin's post-filter implementation if feature_type == "world" and post_filter_type == "merlin": alpha = pysptk.util.mcepalpha(sample_rate) mgc = merlin_post_filter(mgc, alpha) # Remove high-frequency components of lf0/mgc/bap # NOTE: Useful to reduce high-frequency artifacts if trajectory_smoothing: modfs = int(1 / (frame_period * 0.001)) lf0[:, 0] = lowpass_filter( lf0[:, 0], modfs, cutoff=trajectory_smoothing_cutoff_f0 ) if feature_type == "world": for d in range(mgc.shape[1]): mgc[:, d] = lowpass_filter( mgc[:, d], modfs, cutoff=trajectory_smoothing_cutoff ) for d in range(bap.shape[1]): bap[:, d] = lowpass_filter( bap[:, d], modfs, cutoff=trajectory_smoothing_cutoff ) elif feature_type == "melf0": for d in range(mel.shape[1]): mel[:, d] = lowpass_filter( mel[:, d], modfs, cutoff=trajectory_smoothing_cutoff ) if feature_type == "world": use_mcep_aperiodicity = bap.shape[-1] > 5 if feature_type == "world" and not use_mcep_aperiodicity: bap = np.clip(bap, a_min=-60, a_max=0) if feature_type == "world": return mgc, lf0, vuv, bap elif feature_type == "melf0": return mel, lf0, vuv @torch.no_grad() def predict_waveform( device, multistream_features, vocoder=None, vocoder_config=None, vocoder_in_scaler=None, sample_rate=48000, frame_period=5, use_world_codec=True, feature_type="world", vocoder_type="world", vuv_threshold=0.5, ): """Predict waveform from multi-stream acoustic features Vocoders can be 1) WORLD, 2) PWG or 3) uSFGAN. Args: device (torch.device): Device to run inference features (tuple): Acoustic features vocoder (nn.Module): Vocoder model vocoder_config (dict): Vocoder config vocoder_in_scaler (StandardScaler): Vocoder input scaler sample_rate (int,): Sampling rate. frame_period (float): Frame period in msec. use_world_codec (bool): Whether to use WORLD codec for decoding. feature_type (str): Feature type. ``world`` ``world_org``, ``melf0`` or ``neutrino``. vocoder_type (str): Vocoder type. ``world`` or ``pwg`` or ``usfgan`` vuv_threshold (float): VUV threshold. Returns: np.ndarray: Predicted waveform """ if feature_type == "world": mgc, lf0, vuv, bap = multistream_features elif feature_type == "world_org": f0, spectrogram, aperiodicity = multistream_features elif feature_type == "neutrino": mgc, f0, bap = multistream_features # prepare (mgc, lf0, vuv, bap) to be compatible with NNSVS lf0 = f0.copy() lf0[np.nonzero(f0)] = np.log(f0[np.nonzero(f0)]) vuv = (f0 > 0).astype(np.float32) elif feature_type == "melf0": mel, lf0, vuv = multistream_features else: raise ValueError(f"Unknown feature type: {feature_type}") # NOTE: `use_mcep_aperiodicity` was used for experimental purpose but didn't make # significant difference. Please just ignore or ping @r9y9 for details. if feature_type in ["world", "neutrino"]: use_mcep_aperiodicity = bap.shape[-1] > 5 if feature_type == "neutrino" and not use_world_codec: raise ValueError("use_world_codec must be True when feature_type is neutrino") # Waveform generation by WORLD if vocoder_type == "world": if feature_type not in ["world", "world_org", "neutrino"]: raise ValueError(f"Invalid feature type for WORLD vocoder: {feature_type}") if feature_type == "world_org": # NOTE: WORLD-based features are already converted to raw WORLD parameters pass else: f0, spectrogram, aperiodicity = gen_world_params( mgc, lf0, vuv, bap, sample_rate, vuv_threshold=vuv_threshold, use_world_codec=use_world_codec, ) # make sure to have float64 typed parameters wav = pyworld.synthesize( f0.astype(np.float64), spectrogram.astype(np.float64), aperiodicity.astype(np.float64), sample_rate, frame_period, ) elif vocoder_type == "pwg": # NOTE: So far vocoder models are trained on binary V/UV features vuv = (vuv > vuv_threshold).astype(np.float32) if feature_type == "world": voc_inp = ( torch.from_numpy( vocoder_in_scaler.transform( np.concatenate([mgc, lf0, vuv, bap], axis=-1) ) ) .float() .to(device) ) elif feature_type == "melf0": voc_inp = ( torch.from_numpy( vocoder_in_scaler.transform( np.concatenate([mel, lf0, vuv], axis=-1) ) ) .float() .to(device) ) wav = vocoder.inference(voc_inp).view(-1).to("cpu").numpy() elif vocoder_type == "usfgan": if feature_type in ["world", "neutrino"]: fftlen = pyworld.get_cheaptrick_fft_size(sample_rate) if use_mcep_aperiodicity: # Convert mel-cepstrum-based aperiodicity to WORLD's aperiodicity aperiodicity_order = bap.shape[-1] - 1 alpha = pysptk.util.mcepalpha(sample_rate) aperiodicity = pysptk.mc2sp( np.ascontiguousarray(bap).astype(np.float64), fftlen=fftlen, alpha=alpha, ) else: aperiodicity = pyworld.decode_aperiodicity( np.ascontiguousarray(bap).astype(np.float64), sample_rate, fftlen, ) # fill aperiodicity with ones for unvoiced regions aperiodicity[vuv.reshape(-1) < vuv_threshold, 0] = 1.0 # WORLD fails catastrophically for out of range aperiodicity aperiodicity = np.clip(aperiodicity, 0.0, 1.0) # Convert aperiodicity back to BAP if use_mcep_aperiodicity: bap = pysptk.sp2mc( aperiodicity, order=aperiodicity_order, alpha=alpha, ) else: bap = pyworld.code_aperiodicity(aperiodicity, sample_rate).astype( np.float32 ) aux_feats = [mgc, bap] elif feature_type == "melf0": aux_feats = [mel] elif feature_type == "world_org": # it is possible to implement here but I suppose nobody wants to use raise NotImplementedError() aux_feats = ( torch.from_numpy( vocoder_in_scaler.transform(np.concatenate(aux_feats, axis=-1)) ) .float() .to(device) ) contf0 = np.exp(lf0) if vocoder_config.data.sine_f0_type in ["contf0", "cf0"]: f0_inp = contf0 elif vocoder_config.data.sine_f0_type == "f0": f0_inp = contf0 f0_inp[vuv < vuv_threshold] = 0 # NOTE: uSFGAN internally performs normalization # so we don't need to normalize inputs here wav = vocoder.inference(f0_inp, aux_feats).view(-1).to("cpu").numpy() return wav def postprocess_waveform( wav, sample_rate, dtype=np.int16, peak_norm=False, loudness_norm=False, target_loudness=-20.0, ): """Perform post-processing for synthesized waveform Args: wav (ndarray): The input waveform sample_rate (int): The sampling rate dtype (np.dtype): The dtype of output waveform. Default is np.int16. peak_norm (bool): Whether to perform peak normalization loudness_norm (bool): Whether to perform loudness normalization target_loudness (float): Target loudness in dB Returns: ndarray: The post-processed waveform """ wav = bandpass_filter(wav, sample_rate) # Peak normalize audio to 0 dB if peak_norm: wav = pyln.normalize.peak(wav, 0.0) # Normalize loudnes # NOTE: -20 dB is roughly the same as the NEURINO (NSF ver.) if loudness_norm: meter = pyln.Meter(sample_rate) loudness = meter.integrated_loudness(wav) wav = pyln.normalize.loudness(wav, loudness, target_loudness) # NOTE: use np.int16 to save disk space if dtype in [np.int16, "int16"]: # NOTE: NNSVS (>=0.0.3) uses waveforms normalized in [-1, 1] for training. # so the following code shouldn't be used but in case for models trained # with earlier NNSVS if np.max(np.abs(wav)) > 10: # data is likely already in [-32768, 32767] wav = wav.astype(np.int16) elif np.max(np.abs(wav)) <= 1: wav = (wav * 32767.0).astype(np.int16) else: # may need to handle int32 data (if any) warn("Unexpected waveform range: {} - {}".format(np.min(wav), np.max(wav))) warn("Failed to convert to int16. Returning waveform with floating point.") elif dtype is None: pass else: wav = wav.astype(dtype) return wav def _get_nonrest_frame_soft_mask( binary_dict, numeric_dict, linguistic_features, win_length=200, duration_threshold=1.0, ): """Get mask for non-rest frames Args: binary_dict (dict): Dictionary for binary features numeric_dict (dict): Dictionary for numeric features linguistic_features (ndarray): Linguistic features win_length (int): Window length Returns: ndarray: Soft mask for non-rest frames. 1 for non-rest frames and 0 for otherwise. """ mask = np.ones(len(linguistic_features)) in_sil_indices = [] for k, v in binary_dict.items(): name, _ = v if "C-Phone_sil" in name or "C-Phone_pau" in name: in_sil_indices.append(k) if len(in_sil_indices) == 0: return mask in_note_dur_idx = None for k, v in numeric_dict.items(): name, _ = v if "e7" in name: in_note_dur_idx = k break dur = linguistic_features[:, len(binary_dict) + in_note_dur_idx] dur_in_sec = dur * 0.01 for in_sil_idx in in_sil_indices: # Only mask out sil/pau segments over ${silence_threshold} sec. such as long pause mask[ (linguistic_features[:, in_sil_idx] > 0) & (dur_in_sec > duration_threshold) ] = 0 # make a smoothed mask with ${win_length} * 5ms window length mask = scipy.signal.convolve(mask, np.ones(win_length) / win_length, mode="same") # make sure that we don't mask out frames where notes are assigned pitch_idx = get_pitch_index(binary_dict, numeric_dict) score_f0 = linguistic_features[:, pitch_idx] mask[score_f0 > 0] = 1.0 return mask.reshape(-1, 1) def _fill_silence_to_world_params(mgc, lf0, vuv, bap, mask): mgc_sil = np.zeros((1, mgc.shape[1])) # NOTE: mgc_sil is a VERY ROUGH estimate of mgc for silence regions # the speech signal is assumed to be in [-1, 1]. # sr = 48000 # noise = np.random.randn(sr * 10) * 1e-5 # f0, timeaxis = pyworld.harvest(noise, sr, frame_period=5) # f0[:] = 0 # spectrogram = pyworld.cheaptrick(noise, f0, timeaxis, sr) # mgc = pyworld.code_spectral_envelope(spectrogram, sr, 60) # print(mgc.mean(0)) mgc_sil[0, 0] = -23.3 mgc_sil[0, 1] = 0.0679 mgc_sil[0, 2] = 0.00640 mgc_sil[0, 3:] = 1e-3 bap_sil = np.zeros_like(bap) + 1e-11 mgc = mgc * mask + (1 - mask) * mgc_sil bap = bap * mask + (1 - mask) * bap_sil return mgc, lf0, vuv, bap def _fill_silence_to_mel_params(mel, lf0, vuv, mask): # NOTE: -5.5 is also a very rough estimate of log-melspectrogram # for silence regions mel_sil = np.zeros((1, mel.shape[1])) - 5.5 mel = mel * mask + (1 - mask) * mel_sil return mel, lf0, vuv def correct_vuv_by_phone(vuv, binary_dict, linguistic_features): """Correct V/UV by phone-related flags in a hed file This function allows us to control V/UV explicitly by ``C-VUV_Voiced`` and ``C-VUV_Unvoied`` flags in a hed file. This is useful when you see your trained acoustic model have lots of V/UV errors. Note that manually controlling V/UV means we are ignoring the acoustic model's prediction. It would have negative impact in some cases, but most cases it would help workaround V/UV errors. Args: vuv (ndarray): V/UV flags binary_dict (dict): binary feature dictionary linguistic_features (ndarray): linguistic features Returns: ndarray: corrected V/UV flags """ vuv = vuv.copy() # Set V/UV to 1 based on the C-VUV_Voiced flag in_voiced_idx = -1 for k, v in binary_dict.items(): name, _ = v if "C-VUV_Voiced" in name: in_voiced_idx = k break if in_voiced_idx > 0: indices = linguistic_features[:, in_voiced_idx : in_voiced_idx + 1] > 0 vuv[indices] = 1.0 # Set V/UV to 0 based on the C-VUV_Unvoiced flag in_unvoiced_indices = [] for k, v in binary_dict.items(): name, _ = v if "C-VUV_Unvoiced" in name: in_unvoiced_indices.append(k) if len(in_unvoiced_indices) > 0: for in_unvoiced_idx in in_unvoiced_indices: indices = linguistic_features[:, in_unvoiced_idx : in_unvoiced_idx + 1] > 0 vuv[indices] = 0.0 # Set V/UV to 0 for sil/pau/br in_sil_indices = [] for k, v in binary_dict.items(): name, _ = v if "C-Phone_sil" in name or "C-Phone_pau" in name or "C-Phone_br" in name: in_sil_indices.append(k) if len(in_sil_indices) > 0: for in_sil_idx in in_sil_indices: indices = linguistic_features[:, in_sil_idx : in_sil_idx + 1] > 0 vuv[indices] = 0.0 return vuv def gen_spsvs_static_features( labels, acoustic_features, binary_dict, numeric_dict, stream_sizes, has_dynamic_features, pitch_idx=None, num_windows=3, frame_period=5, relative_f0=True, vibrato_scale=1.0, vuv_threshold=0.3, force_fix_vuv=True, ): """Generate static features from predicted acoustic features Args: labels (HTSLabelFile): HTS labels acoustic_features (ndarray): predicted acoustic features binary_dict (dict): binary feature dictionary numeric_dict (dict): numeric feature dictionary stream_sizes (list): stream sizes has_dynamic_features (list): whether each stream has dynamic features pitch_idx (int): index of pitch features num_windows (int): number of windows frame_period (float): frame period relative_f0 (bool): whether to use relative f0 vibrato_scale (float): vibrato scale vuv_threshold (float): vuv threshold force_fix_vuv (bool): whether to use post-processing to fix VUV. Returns: tuple: tuple of mgc, lf0, vuv and bap. """ hts_frame_shift = int(frame_period * 1e4) if pitch_idx is None: pitch_idx = get_pitch_index(binary_dict, numeric_dict) if np.any(has_dynamic_features): static_stream_sizes = get_static_stream_sizes( stream_sizes, has_dynamic_features, num_windows ) else: static_stream_sizes = stream_sizes # Copy here to avoid inplace operations on input acoustic features acoustic_features = acoustic_features.copy() # Split multi-stream features streams = split_streams(acoustic_features, static_stream_sizes) if len(streams) == 4: mgc, target_f0, vuv, bap = streams vib, vib_flags = None, None elif len(streams) == 5: # Assuming diff-based vibrato parameters mgc, target_f0, vuv, bap, vib = streams vib_flags = None elif len(streams) == 6: # Assuming sine-based vibrato parameters mgc, target_f0, vuv, bap, vib, vib_flags = streams else: raise RuntimeError("Not supported streams") linguistic_features = fe.linguistic_features( labels, binary_dict, numeric_dict, add_frame_features=True, frame_shift=hts_frame_shift, ) # Correct V/UV based on special phone flags if force_fix_vuv: vuv = correct_vuv_by_phone(vuv, binary_dict, linguistic_features) # F0 if relative_f0: diff_lf0 = target_f0 f0_score = _midi_to_hz(linguistic_features, pitch_idx, False)[:, None] lf0_score = f0_score.copy() nonzero_indices = np.nonzero(lf0_score) lf0_score[nonzero_indices] = np.log(f0_score[nonzero_indices]) lf0_score = interp1d(lf0_score, kind="slinear") f0 = diff_lf0 + lf0_score f0[vuv < vuv_threshold] = 0 f0[np.nonzero(f0)] = np.exp(f0[np.nonzero(f0)]) else: f0 = target_f0 f0[vuv < vuv_threshold] = 0 f0[np.nonzero(f0)] = np.exp(f0[np.nonzero(f0)]) if vib is not None: if vib_flags is not None: # Generate sine-based vibrato vib_flags = vib_flags.flatten() m_a, m_f = vib[:, 0], vib[:, 1] # Fill zeros for non-vibrato frames m_a[vib_flags < 0.5] = 0 m_f[vib_flags < 0.5] = 0 # Gen vibrato sr_f0 = int(1 / (frame_period * 0.001)) f0 = gen_sine_vibrato(f0.flatten(), sr_f0, m_a, m_f, vibrato_scale) else: # Generate diff-based vibrato f0 = f0.flatten() + vibrato_scale * vib.flatten() # NOTE: Back to log-domain for convenience lf0 = f0.copy() lf0[np.nonzero(lf0)] = np.log(f0[np.nonzero(lf0)]) # NOTE: interpolation is necessary lf0 = interp1d(lf0, kind="slinear") lf0 = lf0[:, None] if len(lf0.shape) == 1 else lf0 vuv = vuv[:, None] if len(vuv.shape) == 1 else vuv return mgc, lf0, vuv, bap def gen_world_params( mgc, lf0, vuv, bap, sample_rate, vuv_threshold=0.3, use_world_codec=False, ): """Generate WORLD parameters from mgc, lf0, vuv and bap. Args: mgc (ndarray): mgc lf0 (ndarray): lf0 vuv (ndarray): vuv bap (ndarray): bap sample_rate (int): sample rate vuv_threshold (float): threshold for VUV use_world_codec (bool): whether to use WORLD codec for spectral envelope Returns: tuple: tuple of f0, spectrogram and aperiodicity """ fftlen = pyworld.get_cheaptrick_fft_size(sample_rate) alpha = pysptk.util.mcepalpha(sample_rate) use_mcep_aperiodicity = bap.shape[-1] > 5 if use_world_codec: spectrogram = pyworld.decode_spectral_envelope( np.ascontiguousarray(mgc).astype(np.float64), sample_rate, fftlen ) else: spectrogram = pysptk.mc2sp( np.ascontiguousarray(mgc), fftlen=fftlen, alpha=alpha ) if use_mcep_aperiodicity: aperiodicity = pysptk.mc2sp( np.ascontiguousarray(bap), fftlen=fftlen, alpha=alpha ) else: aperiodicity = pyworld.decode_aperiodicity( np.ascontiguousarray(bap).astype(np.float64), sample_rate, fftlen ) # fill aperiodicity with ones for unvoiced regions aperiodicity[vuv.reshape(-1) < vuv_threshold, 0] = 1.0 # WORLD fails catastrophically for out of range aperiodicity aperiodicity = np.clip(aperiodicity, 0.0, 1.0) f0 = lf0.copy() f0[np.nonzero(f0)] = np.exp(f0[np.nonzero(f0)]) f0[vuv < vuv_threshold] = 0 f0 = f0.flatten().astype(np.float64) spectrogram = spectrogram.astype(np.float64) aperiodicity = aperiodicity.astype(np.float64) return f0, spectrogram, aperiodicity
49,738
34.275887
96
py
nnsvs
nnsvs-master/nnsvs/postfilters.py
import numpy as np import torch from nnsvs.base import BaseModel from nnsvs.multistream import split_streams from nnsvs.util import init_weights from torch import nn def variance_scaling(gv, feats, offset=2, note_frame_indices=None): """Variance scaling method to enhance synthetic speech quality Method proposed in :cite:t:`silen2012ways`. Args: gv (tensor): global variance computed over training data feats (tensor): input features offset (int): offset note_frame_indices (tensor): indices of note frames Returns: tensor: scaled features """ if note_frame_indices is not None: if len(note_frame_indices) == 0: return feats utt_gv = feats[note_frame_indices].var(0) utt_mu = feats[note_frame_indices].mean(0) else: utt_gv = feats.var(0) utt_mu = feats.mean(0) out = feats.copy() if note_frame_indices is not None: out[note_frame_indices, offset:] = ( np.sqrt(gv[offset:] / utt_gv[offset:]) * (feats[note_frame_indices, offset:] - utt_mu[offset:]) + utt_mu[offset:] ) else: out[:, offset:] = ( np.sqrt(gv[offset:] / utt_gv[offset:]) * (feats[:, offset:] - utt_mu[offset:]) + utt_mu[offset:] ) return out class MovingAverage1d(nn.Conv1d): """Moving average filter on 1-D signals Args: in_channels (int): number of input channels out_channels (int): number of output channels kernel_size (int): kernel size padding_mode (str): padding mode """ def __init__(self, in_channels, out_channels, kernel_size, padding_mode="reflect"): # NOTE: process each channel independently by setting groups=in_channels super().__init__( in_channels, out_channels, kernel_size, groups=in_channels, bias=False, padding="same", padding_mode=padding_mode, ) nn.init.constant_(self.weight, 1 / kernel_size) for p in self.parameters(): p.requires_grad = False class Conv2dPostFilter(BaseModel): """A post-filter based on Conv2d A model proposed in :cite:t:`kaneko2017generative`. Args: channels (int): number of channels kernel_size (tuple): kernel sizes for Conv2d init_type (str): type of initialization noise_scale (float): scale of noise noise_type (str): type of noise. "frame_wise" or "bin_wise" padding_mode (str): padding mode smoothing_width (int): Width of smoothing window. The larger the smoother. Only used at inference time. """ def __init__( self, in_dim=None, channels=128, kernel_size=(5, 5), init_type="kaiming_normal", noise_scale=1.0, noise_type="bin_wise", padding_mode="zeros", smoothing_width=-1, ): super().__init__() self.in_dim = in_dim self.noise_type = noise_type self.noise_scale = noise_scale C = channels self.smoothing_width = smoothing_width assert len(kernel_size) == 2 ks = np.asarray(list(kernel_size)) padding = (ks - 1) // 2 self.conv1 = nn.Sequential( nn.Conv2d( 2, C, kernel_size=ks, padding=padding, padding_mode=padding_mode, ), nn.ReLU(), ) self.conv2 = nn.Sequential( nn.Conv2d( C + 1, C * 2, kernel_size=ks, padding=padding, padding_mode=padding_mode ), nn.ReLU(), ) self.conv3 = nn.Sequential( nn.Conv2d( C * 2 + 1, C, kernel_size=ks, padding=padding, padding_mode=padding_mode ), nn.ReLU(), ) self.conv4 = nn.Conv2d( C + 1, 1, kernel_size=ks, padding=padding, padding_mode=padding_mode ) if self.noise_type == "frame_wise": # noise: (B, T, 1) self.fc = nn.Linear(1, in_dim) elif self.noise_type == "bin_wise": # noise: (B, T, C) self.fc = None else: raise ValueError("Unknown noise type: {}".format(self.noise_type)) init_weights(self, init_type) def forward(self, x, lengths=None, y=None, is_inference=False): """Forward step Args: x (torch.Tensor): input tensor of shape (B, T, C) lengths (torch.Tensor): lengths of shape (B,) Returns: torch.Tensor: output tensor of shape (B, T, C) """ # (B, T, C) -> (B, 1, T, C): x = x.unsqueeze(1) if self.noise_type == "bin_wise": # (B, C, T) z = torch.randn_like(x).squeeze(1).transpose(1, 2) * self.noise_scale # Apply moving average filter at inference time only if is_inference and self.smoothing_width > 0: ave_filt = MovingAverage1d( self.in_dim, self.in_dim, self.smoothing_width ).to(x.device) z = ave_filt(z) # (B, 1, T, C) z = z.transpose(1, 2).unsqueeze(1) elif self.noise_type == "frame_wise": # (B, 1, T) z = torch.randn(x.shape[0], 1, x.shape[2]).to(x.device) * self.noise_scale # Apply moving average filter at inference time only if is_inference and self.smoothing_width > 0: ave_filt = MovingAverage1d(1, 1, self.smoothing_width).to(x.device) z = ave_filt(z) # (B, 1, T, 1) z = z.unsqueeze(-1) # (B, 1, T, C) z = self.fc(z) x_syn = x y = self.conv1(torch.cat([x_syn, z], dim=1)) y = self.conv2(torch.cat([x_syn, y], dim=1)) y = self.conv3(torch.cat([x_syn, y], dim=1)) residual = self.conv4(torch.cat([x_syn, y], dim=1)) out = x_syn + residual # (B, 1, T, C) -> (B, T, C) out = out.squeeze(1) return out def inference(self, x, lengths=None): return self(x, lengths, is_inference=True) class MultistreamPostFilter(BaseModel): """A multi-stream post-filter that applies post-filtering for each feature stream Currently, post-filtering for MGC, BAP and log-F0 are supported. Note that it doesn't make much sense to apply post-filtering for other features. Args: mgc_postfilter (nn.Module): post-filter for MGC bap_postfilter (nn.Module): post-filter for BAP lf0_postfilter (nn.Module): post-filter for log-F0 stream_sizes (list): sizes of each feature stream mgc_offset (int): offset for MGC. Defaults to 2. bap_offset (int): offset for BAP. Defaults to 0. """ def __init__( self, mgc_postfilter: nn.Module, bap_postfilter: nn.Module, lf0_postfilter: nn.Module, stream_sizes: list, mgc_offset: int = 2, bap_offset: int = 0, ): super().__init__() self.mgc_postfilter = mgc_postfilter self.bap_postfilter = bap_postfilter self.lf0_postfilter = lf0_postfilter self.stream_sizes = stream_sizes self.mgc_offset = mgc_offset self.bap_offset = bap_offset def forward(self, x, lengths=None, y=None, is_inference=False): """Forward step Each feature stream is processed independently. Args: x (torch.Tensor): input tensor of shape (B, T, C) lengths (torch.Tensor): lengths of shape (B,) Returns: torch.Tensor: output tensor of shape (B, T, C) """ streams = split_streams(x, self.stream_sizes) if len(streams) == 4: mgc, lf0, vuv, bap = streams elif len(streams) == 5: mgc, lf0, vuv, bap, vuv = streams elif len(streams) == 6: mgc, lf0, vuv, bap, vib, vib_flags = streams else: raise ValueError("Invalid number of streams") if self.mgc_postfilter is not None: if self.mgc_offset > 0: # keep unchanged for the 0-to-${mgc_offset}-th dim of mgc mgc0 = mgc[:, :, : self.mgc_offset] if is_inference: mgc_pf = self.mgc_postfilter.inference( mgc[:, :, self.mgc_offset :], lengths ) else: mgc_pf = self.mgc_postfilter(mgc[:, :, self.mgc_offset :], lengths) mgc_pf = torch.cat([mgc0, mgc_pf], dim=-1) else: if is_inference: mgc_pf = self.mgc_postfilter.inference(mgc, lengths) else: mgc_pf = self.mgc_postfilter(mgc, lengths) mgc = mgc_pf if self.bap_postfilter is not None: if self.bap_offset > 0: # keep unchanged for the 0-to-${bap_offset}-th dim of bap bap0 = bap[:, :, : self.bap_offset] if is_inference: bap_pf = self.bap_postfilter.inference( bap[:, :, self.bap_offset :], lengths ) else: bap_pf = self.bap_postfilter(bap[:, :, self.bap_offset :], lengths) bap_pf = torch.cat([bap0, bap_pf], dim=-1) else: if is_inference: bap_pf = self.bap_postfilter.inference(bap, lengths) else: bap_pf = self.bap_postfilter(bap, lengths) bap = bap_pf if self.lf0_postfilter is not None: if is_inference: lf0 = self.lf0_postfilter.inference(lf0, lengths) else: lf0 = self.lf0_postfilter(lf0, lengths) if len(streams) == 4: out = torch.cat([mgc, lf0, vuv, bap], dim=-1) elif len(streams) == 5: out = torch.cat([mgc, lf0, vuv, bap, vib], dim=-1) elif len(streams) == 6: out = torch.cat([mgc, lf0, vuv, bap, vib, vib_flags], dim=-1) return out def inference(self, x, lengths): return self(x, lengths, is_inference=True) class MelF0MultistreamPostFilter(BaseModel): def __init__( self, mel_postfilter: nn.Module, lf0_postfilter: nn.Module, stream_sizes: list, mel_offset: int = 0, ): super().__init__() self.mel_postfilter = mel_postfilter self.lf0_postfilter = lf0_postfilter self.stream_sizes = stream_sizes self.mel_offset = mel_offset def forward(self, x, lengths=None, y=None, is_inference=False): """Forward step Each feature stream is processed independently. Args: x (torch.Tensor): input tensor of shape (B, T, C) lengths (torch.Tensor): lengths of shape (B,) Returns: torch.Tensor: output tensor of shape (B, T, C) """ streams = split_streams(x, self.stream_sizes) assert len(streams) == 3 mel, lf0, vuv = streams if self.mel_postfilter is not None: if self.mel_offset > 0: # keep unchanged for the 0-to-${mgc_offset}-th dim of mgc mel0 = mel[:, :, : self.mel_offset] if is_inference: mel_pf = self.mel_postfilter.inference( mel[:, :, self.mel_offset :], lengths ) else: mel_pf = self.mel_postfilter(mel[:, :, self.mel_offset :], lengths) mel_pf = torch.cat([mel0, mel_pf], dim=-1) else: if is_inference: mel_pf = self.mel_postfilter.inference(mel, lengths) else: mel_pf = self.mel_postfilter(mel, lengths) mel = mel_pf if self.lf0_postfilter is not None: if is_inference: lf0 = self.lf0_postfilter.inference(lf0, lengths) else: lf0 = self.lf0_postfilter(lf0, lengths) out = torch.cat([mel, lf0, vuv], dim=-1) return out def inference(self, x, lengths): return self(x, lengths, is_inference=True) class _PadConv2dPostFilter(BaseModel): def __init__( self, in_dim=None, channels=128, kernel_size=5, init_type="kaiming_normal", padding_side="left", ): super().__init__() assert not isinstance(kernel_size, list) C = channels ks = kernel_size padding = (ks - 1) // 2 self.padding = padding # Treat padding for the feature-axis carefully # use normal padding for the time-axis (i.e., (padding, padding)) self.padding_side = padding_side if padding_side == "left": self.pad = nn.ReflectionPad2d((padding, 0, padding, padding)) elif padding_side == "none": self.pad = nn.ReflectionPad2d((0, 0, padding, padding)) elif padding_side == "right": self.pad = nn.ReflectionPad2d((0, padding, padding, padding)) else: raise ValueError("Invalid padding side") self.conv1 = nn.Sequential( nn.Conv2d(2, C, kernel_size=(ks, ks)), nn.ReLU(), ) # NOTE: for the subsequent layers, use fixed kernel_size 3 for feature-axis self.conv2 = nn.Sequential( nn.Conv2d( C + 1, C * 2, kernel_size=(ks, 3), padding=(padding, 1), padding_mode="reflect", ), nn.ReLU(), ) self.conv3 = nn.Sequential( nn.Conv2d( C * 2 + 1, C, kernel_size=(ks, 3), padding=(padding, 1), padding_mode="reflect", ), nn.ReLU(), ) self.conv4 = nn.Conv2d( C + 1, 1, kernel_size=(ks, 1), padding=(padding, 0), padding_mode="reflect" ) self.fc = nn.Linear(1, in_dim) init_weights(self, init_type) def forward(self, x, z, lengths=None): # (B, T, C) -> (B, 1, T, C): x = x.unsqueeze(1) z = z.unsqueeze(1) z = self.fc(z) x_syn = x y = self.conv1(torch.cat([self.pad(x_syn), self.pad(z)], dim=1)) if self.padding_side == "left": x_syn = x[:, :, :, : -self.padding] elif self.padding_side == "none": x_syn = x[:, :, :, self.padding : -self.padding] elif self.padding_side == "right": x_syn = x[:, :, :, self.padding :] y = self.conv2(torch.cat([x_syn, y], dim=1)) y = self.conv3(torch.cat([x_syn, y], dim=1)) residual = self.conv4(torch.cat([x_syn, y], dim=1)) out = x_syn + residual # (B, 1, T, C) -> (B, T, C) out = out.squeeze(1) return out class MultistreamConv2dPostFilter(nn.Module): """Conv2d-based multi-stream post-filter designed for MGC Divide the MGC transformation into low/mid/high dim transfomations with small overlaps. Overlap is determined by the kernel size. """ def __init__( self, in_dim=None, channels=128, kernel_size=5, init_type="kaiming_normal", noise_scale=1.0, stream_sizes=(8, 20, 30), ): super().__init__() assert len(stream_sizes) == 3 self.padding = (kernel_size - 1) // 2 self.noise_scale = noise_scale self.stream_sizes = stream_sizes self.low_postfilter = _PadConv2dPostFilter( stream_sizes[0] + self.padding, channels=channels, kernel_size=kernel_size, init_type=init_type, padding_side="left", ) self.mid_postfilter = _PadConv2dPostFilter( stream_sizes[1] + 2 * self.padding, channels=channels, kernel_size=kernel_size, init_type=init_type, padding_side="none", ) self.high_postfilter = _PadConv2dPostFilter( stream_sizes[2] + self.padding, channels=channels, kernel_size=kernel_size, init_type=init_type, padding_side="right", ) def forward(self, x, lengths=None, y=None): assert x.shape[-1] == sum(self.stream_sizes) # (B, T, C) z = torch.randn(x.shape[0], x.shape[1], 1).to(x.device) * self.noise_scale # Process three streams separately with a overlap width of padding out1 = self.low_postfilter(x[:, :, : self.stream_sizes[0] + self.padding], z) out2 = self.mid_postfilter( x[ :, :, self.stream_sizes[0] - self.padding : sum(self.stream_sizes[:2]) + self.padding, ], z, ) out3 = self.high_postfilter( x[:, :, sum(self.stream_sizes[:2]) - self.padding :], z ) # Merge the three outputs out = torch.cat([out1, out2, out3], dim=-1) return out
17,384
31.801887
88
py
nnsvs
nnsvs-master/nnsvs/util.py
import importlib import random from os.path import join from pathlib import Path from typing import Any import numpy as np import pkg_resources import pyworld import torch from hydra.utils import instantiate from nnsvs.multistream import get_static_features, get_static_stream_sizes from nnsvs.usfgan import USFGANWrapper from omegaconf import OmegaConf from torch import nn try: from parallel_wavegan.utils import load_model _pwg_available = True except ImportError: _pwg_available = False # mask-related functions were adapted from https://github.com/espnet/espnet EXAMPLE_DIR = "_example_data" # Adapted from https://github.com/junyanz/pytorch-CycleGAN-and-pix2pix def init_weights(net, init_type="normal", init_gain=0.02): """Initialize network weights. Args: net (torch.nn.Module): network to initialize init_type (str): the name of an initialization method: normal | xavier | kaiming | orthogonal | none. init_gain (float): scaling factor for normal, xavier and orthogonal. """ if init_type == "none": return def init_func(m): classname = m.__class__.__name__ if hasattr(m, "weight") and ( classname.find("Conv") != -1 or classname.find("Linear") != -1 ): if init_type == "normal": nn.init.normal_(m.weight.data, 0.0, init_gain) elif init_type == "xavier_normal": nn.init.xavier_normal_(m.weight.data, gain=init_gain) elif init_type == "kaiming_normal": nn.init.kaiming_normal_(m.weight.data, a=0, mode="fan_in") elif init_type == "orthogonal": nn.init.orthogonal_(m.weight.data, gain=init_gain) else: raise NotImplementedError( "initialization method [%s] is not implemented" % init_type ) if hasattr(m, "bias") and m.bias is not None: nn.init.constant_(m.bias.data, 0.0) elif classname.find("BatchNorm2d") != -1: # BatchNorm Layer's weight is not a matrix; only normal distribution applies. nn.init.normal_(m.weight.data, 1.0, init_gain) nn.init.constant_(m.bias.data, 0.0) net.apply(init_func) def get_world_stream_info( sr: int, mgc_order: int, num_windows: int = 3, vibrato_mode: str = "none", use_mcep_aperiodicity: bool = False, mcep_aperiodicity_order: int = 24, ): """Get stream sizes for WORLD-based acoustic features Args: sr (int): sampling rate mgc_order (int): order of mel-generalized cepstrum num_windows (int): number of windows vibrato_mode (str): vibrato analysis mode Returns: list: stream sizes """ # [mgc, lf0, vuv, bap] stream_sizes = [ (mgc_order + 1) * num_windows, num_windows, 1, pyworld.get_num_aperiodicities(sr) * num_windows if not use_mcep_aperiodicity else mcep_aperiodicity_order + 1, ] if vibrato_mode == "diff": # vib stream_sizes.append(num_windows) elif vibrato_mode == "sine": # vib + vib_flags stream_sizes.append(3 * num_windows) stream_sizes.append(1) elif vibrato_mode == "none": pass else: raise RuntimeError("Unknown vibrato mode: {}".format(vibrato_mode)) return stream_sizes def load_utt_list(utt_list): """Load a list of utterances. Args: utt_list (str): path to a file containing a list of utterances Returns: List[str]: list of utterances """ with open(utt_list) as f: utt_ids = f.readlines() utt_ids = map(lambda utt_id: utt_id.strip(), utt_ids) utt_ids = filter(lambda utt_id: len(utt_id) > 0, utt_ids) return list(utt_ids) def example_xml_file(key="haruga_kita"): """Get the path to an included xml file. Args: key (str): key of the file Returns: str: path to an example xml file Raises: FileNotFoundError: if the file is not found """ return pkg_resources.resource_filename(__name__, join(EXAMPLE_DIR, f"{key}.xml")) def init_seed(seed): """Initialize random seed. Args: seed (int): random seed """ random.seed(seed) np.random.seed(seed) torch.manual_seed(seed) if torch.cuda.is_available(): torch.cuda.manual_seed_all(seed) def dynamic_import(name: str) -> Any: """Dynamic import Args: name (str): module_name + ":" + class_name Returns: Any: class object """ mod_name, class_name = name.split(":") mod = importlib.import_module(mod_name) return getattr(mod, class_name) def pad_2d(x, max_len, constant_values=0): """Pad a 2d-tensor. Args: x (torch.Tensor): tensor to pad max_len (int): maximum length of the tensor constant_values (int, optional): value to pad with. Default: 0 Returns: torch.Tensor: padded tensor """ x = np.pad( x, [(0, max_len - len(x)), (0, 0)], mode="constant", constant_values=constant_values, ) return x def make_pad_mask(lengths, xs=None, length_dim=-1, maxlen=None): """Make mask tensor containing indices of padded part. Args: lengths (LongTensor or List): Batch of lengths (B,). xs (Tensor, optional): The reference tensor. If set, masks will be the same shape as this tensor. length_dim (int, optional): Dimension indicator of the above tensor. Returns: Tensor: Mask tensor containing indices of padded part. dtype=torch.uint8 in PyTorch 1.2- dtype=torch.bool in PyTorch 1.2+ (including 1.2) """ if length_dim == 0: raise ValueError("length_dim cannot be 0: {}".format(length_dim)) if not isinstance(lengths, list): lengths = lengths.tolist() bs = int(len(lengths)) if maxlen is None: if xs is None: maxlen = int(max(lengths)) else: maxlen = xs.size(length_dim) seq_range = torch.arange(0, maxlen, dtype=torch.int64) seq_range_expand = seq_range.unsqueeze(0).expand(bs, maxlen) seq_length_expand = seq_range_expand.new(lengths).unsqueeze(-1) mask = seq_range_expand >= seq_length_expand if xs is not None: assert xs.size(0) == bs, (xs.size(0), bs) if length_dim < 0: length_dim = xs.dim() + length_dim # ind = (:, None, ..., None, :, , None, ..., None) ind = tuple( slice(None) if i in (0, length_dim) else None for i in range(xs.dim()) ) mask = mask[ind].expand_as(xs).to(xs.device) return mask def make_non_pad_mask(lengths, xs=None, length_dim=-1, maxlen=None): """Make mask tensor containing indices of non-padded part. Args: lengths (LongTensor or List): Batch of lengths (B,). xs (Tensor, optional): The reference tensor. If set, masks will be the same shape as this tensor. length_dim (int, optional): Dimension indicator of the above tensor. Returns: ByteTensor: mask tensor containing indices of padded part. dtype=torch.uint8 in PyTorch 1.2- dtype=torch.bool in PyTorch 1.2+ (including 1.2) """ return ~make_pad_mask(lengths, xs, length_dim, maxlen) class PyTorchStandardScaler(nn.Module): """PyTorch module for standardization. Args: mean (torch.Tensor): mean scale (torch.Tensor): scale """ def __init__(self, mean, scale): super().__init__() self.mean_ = nn.Parameter(mean, requires_grad=False) self.scale_ = nn.Parameter(scale, requires_grad=False) def transform(self, x): return (x - self.mean_) / self.scale_ def inverse_transform(self, x): return x * self.scale_ + self.mean_ class StandardScaler: """sklearn.preprocess.StandardScaler like class with only transform functionality Args: mean (np.ndarray): mean var (np.ndarray): variance scale (np.ndarray): scale """ def __init__(self, mean, var, scale): self.mean_ = mean self.var_ = var # NOTE: scale may not exactly same as np.sqrt(var) self.scale_ = scale def transform(self, x): return (x - self.mean_) / self.scale_ def inverse_transform(self, x): return x * self.scale_ + self.mean_ class MinMaxScaler: """sklearn.preprocess.MinMaxScaler like class with only transform functionality Args: min (np.ndarray): minimum scale (np.ndarray): scale data_min (np.ndarray): minimum of input data data_max (np.ndarray): maximum of input data feature_range (tuple): (min, max) """ def __init__(self, min, scale, data_min=None, data_max=None, feature_range=(0, 1)): self.min_ = min self.scale_ = scale self.data_min_ = data_min self.data_max_ = data_max self.feature_range = feature_range def transform(self, x): return self.scale_ * x + self.min_ def inverse_transform(self, x): return (x - self.min_) / self.scale_ def extract_static_scaler(out_scaler, model_config): """Extract scaler for static features Args: out_scaler (StandardScaler or MinMaxScaler): target scaler model_config (dict): model config that contain stream information Returns: StandardScaler or MinMaxScaler: scaler for static features """ mean_ = get_static_features( out_scaler.mean_.reshape(1, 1, out_scaler.mean_.shape[-1]), model_config.num_windows, model_config.stream_sizes, model_config.has_dynamic_features, ) mean_ = np.concatenate(mean_, -1).reshape(1, -1) var_ = get_static_features( out_scaler.var_.reshape(1, 1, out_scaler.var_.shape[-1]), model_config.num_windows, model_config.stream_sizes, model_config.has_dynamic_features, ) var_ = np.concatenate(var_, -1).reshape(1, -1) scale_ = get_static_features( out_scaler.scale_.reshape(1, 1, out_scaler.scale_.shape[-1]), model_config.num_windows, model_config.stream_sizes, model_config.has_dynamic_features, ) scale_ = np.concatenate(scale_, -1).reshape(1, -1) static_scaler = StandardScaler(mean_, var_, scale_) return static_scaler def load_vocoder(path, device, acoustic_config): """Load vocoder model from a given checkpoint path Note that the path needs to be a checkpoint of PWG or USFGAN. Args: path (str or Path): Path to the vocoder model device (str): Device to load the model acoustic_config (dict): Acoustic model config Returns: tuple: (vocoder, vocoder_in_scaler, vocoder_config) """ if not _pwg_available: raise RuntimeError( "parallel_wavegan is required to load pre-trained checkpoint." ) path = Path(path) if isinstance(path, str) else path model_dir = path.parent if (model_dir / "vocoder_model.yaml").exists(): # packed model vocoder_config = OmegaConf.load(model_dir / "vocoder_model.yaml") elif (model_dir / "config.yml").exists(): # PWG checkpoint vocoder_config = OmegaConf.load(model_dir / "config.yml") else: # usfgan vocoder_config = OmegaConf.load(model_dir / "config.yaml") if "generator" in vocoder_config and "discriminator" in vocoder_config: # usfgan checkpoint = torch.load( path, map_location=lambda storage, loc: storage, ) vocoder = instantiate(vocoder_config.generator).to(device) vocoder.load_state_dict(checkpoint["model"]["generator"]) vocoder.remove_weight_norm() vocoder = USFGANWrapper(vocoder_config, vocoder) stream_sizes = get_static_stream_sizes( acoustic_config.stream_sizes, acoustic_config.has_dynamic_features, acoustic_config.num_windows, ) # Extract scaler params for [mgc, bap] if vocoder_config.data.aux_feats == ["mcep", "codeap"]: # streams: (mgc, lf0, vuv, bap) mean_ = np.load(model_dir / "in_vocoder_scaler_mean.npy") var_ = np.load(model_dir / "in_vocoder_scaler_var.npy") scale_ = np.load(model_dir / "in_vocoder_scaler_scale.npy") mgc_end_dim = stream_sizes[0] bap_start_dim = sum(stream_sizes[:3]) bap_end_dim = sum(stream_sizes[:4]) vocoder_in_scaler = StandardScaler( np.concatenate([mean_[:mgc_end_dim], mean_[bap_start_dim:bap_end_dim]]), np.concatenate([var_[:mgc_end_dim], var_[bap_start_dim:bap_end_dim]]), np.concatenate( [scale_[:mgc_end_dim], scale_[bap_start_dim:bap_end_dim]] ), ) else: # streams: (mel, lf0, vuv) mel_dim = stream_sizes[0] vocoder_in_scaler = StandardScaler( np.load(model_dir / "in_vocoder_scaler_mean.npy")[:mel_dim], np.load(model_dir / "in_vocoder_scaler_var.npy")[:mel_dim], np.load(model_dir / "in_vocoder_scaler_scale.npy")[:mel_dim], ) else: # PWG vocoder = load_model(path, config=vocoder_config).to(device) vocoder.remove_weight_norm() vocoder_in_scaler = StandardScaler( np.load(model_dir / "in_vocoder_scaler_mean.npy"), np.load(model_dir / "in_vocoder_scaler_var.npy"), np.load(model_dir / "in_vocoder_scaler_scale.npy"), ) vocoder.eval() return vocoder, vocoder_in_scaler, vocoder_config
13,862
30.578588
89
py
nnsvs
nnsvs-master/nnsvs/pitch.py
"""This module provides functionality for pitch analysis. References: Nakano et al, "An Automatic Singing Skill Evaluation Method for Unknown Melodies Using Pitch Interval Accuracy and Vibrato Features" Proc. Interspeech 2006. 山田 et al, "HMM に基づく歌声合成のためのビブラートモデル化" IPSJ SIG Tech. Report 2009. Note that vibrato extraction method in this module is exerimental. Because details of the vibrato extraction method are not described in the above papers and not trivial to implement (in my opinion), my implementation may not work well compared to the original author's one. Also note that there are a lot of tunable parameters (threshold, window size, min/max extent, cut-off frequency, etc.). If you want to get maximum performance, you might want to tune these parameters with your dataset. I tested this code with kiritan_singing and nit-song070 database. """ import librosa import numpy as np import torch from nnsvs.dsp import lowpass_filter from scipy.signal import argrelmax, argrelmin _c4_hz = 440 * 2 ** (3 / 12 - 1) _c4_cent = 4800 def hz_to_cent_based_c4(hz): """Convert Hz to cent based on C4 Args: hz (np.ndarray): array of Hz Returns: np.ndarray: array of cent """ out = hz.copy() nonzero_indices = np.where(hz > 0)[0] out[nonzero_indices] = ( 1200 * np.log(hz[nonzero_indices] / _c4_hz) / np.log(2) + _c4_cent ) return out def cent_to_hz_based_c4(cent): """Convert cent to Hz based on C4 Args: cent (np.ndarray): array of cent Returns: np.ndarray: array of Hz """ out = cent.copy() nonzero_indices = np.where(cent > 0)[0] out[nonzero_indices] = ( np.exp((cent[nonzero_indices] - _c4_cent) * np.log(2) / 1200) * _c4_hz ) return out def nonzero_segments(f0): """Find nonzero segments Args: f0 (np.ndarray): array of f0 Returns: list: list of (start, end) """ vuv = f0 > 0 started = False s, e = 0, 0 segments = [] for idx in range(len(f0)): if vuv[idx] > 0 and not started: started = True s = idx elif started and (vuv[idx] <= 0): e = idx started = False segments.append((s, e)) else: pass if started and vuv[-1] > 0: segments.append((s, len(vuv) - 1)) return segments def note_segments(lf0_score_denorm): """Compute note segments (start and end indices) from log-F0 Note that unvoiced frames must be set to 0 in advance. Args: lf0_score_denorm (Tensor): (B, T) Returns: list: list of note (start, end) indices """ segments = [] for s, e in nonzero_segments(lf0_score_denorm): out = torch.sign(torch.abs(torch.diff(lf0_score_denorm[s : e + 1]))) transitions = torch.where(out > 0)[0] note_start, note_end = s, -1 for pos in transitions: note_end = int(s + pos) segments.append((note_start, note_end)) note_start = note_end + 1 # Handle last note while ( note_start < len(lf0_score_denorm) - 1 and lf0_score_denorm[note_start] <= 0 ): note_start += 1 note_end = note_start + 1 while note_end < len(lf0_score_denorm) - 1 and lf0_score_denorm[note_end] > 0: note_end += 1 if note_end != note_start + 1: segments.append((note_start, note_end)) return segments def compute_f0_correction_ratio( f0, f0_score, edges_to_be_excluded=50, out_of_tune_threshold=200, correction_threshold=100, ): """Compute f0 correction ratio Args: f0 (np.ndarray): array of f0 f0_score (np.ndarray): array of f0 score Returns: float: correction ratio to multiplied to F0 (i.e. f0 * ratio) """ segments = note_segments(torch.from_numpy(f0_score)) center_f0s = [] center_score_f0s = [] # edges_to_be_excluded = 50 # 0.25 sec for excluding overshoot/preparation for s, e in segments: L = e - s if L > edges_to_be_excluded * 2: center_f0s.append(f0[s + edges_to_be_excluded : e - edges_to_be_excluded]) center_score_f0s.append( f0_score[s + edges_to_be_excluded : e - edges_to_be_excluded] ) center_f0s = np.concatenate(center_f0s) center_score_f0s = np.concatenate(center_score_f0s) # Compute pitch ratio to be multiplied nonzero_indices = (center_f0s > 0) & (center_score_f0s > 0) ratio = center_score_f0s[nonzero_indices] / center_f0s[nonzero_indices] # Exclude too out-of-tune frames (over 2 semitone) up_threshold = np.exp(out_of_tune_threshold * np.log(2) / 1200) low_threshold = np.exp(-out_of_tune_threshold * np.log(2) / 1200) ratio = ratio[(ratio < up_threshold) & (ratio > low_threshold)] global_offset = ratio.mean() # Avoid corrections over semi-tone # If more than semi-tone pitch correction is needed, it is better to correct # data by hand or fix musicxml or UST instead. up_threshold = np.exp(correction_threshold * np.log(2) / 1200) low_threshold = np.exp(-correction_threshold * np.log(2) / 1200) if global_offset > up_threshold or global_offset < low_threshold: print( f"""warn: more than 1 semitone pitch correction is needed. global_offset: {global_offset} cent. It is likely that manual pitch corrections are preferable.""" ) global_offset = np.clip(global_offset, low_threshold, up_threshold) return global_offset def extract_vibrato_parameters_impl(pitch_seg, sr): """Extract vibrato parameters for a single pitch segment Nakano et al, "An Automatic Singing Skill Evaluation Method for Unknown Melodies Using Pitch Interval Accuracy and Vibrato Features" Proc. Interspeech 2006. 山田 et al, "HMM に基づく歌声合成のためのビブラートモデル化" IPSJ SIG Tech. Report 2009. Args: pitch_seg (np.ndarray): array of pitch sr (int): sampling rate Returns: tuple: (R, E, m_a, m_f) """ peak_high_pos = argrelmax(pitch_seg)[0] peak_low_pos = argrelmin(pitch_seg)[0] m_a = np.zeros(len(pitch_seg)) m_f = np.zeros(len(pitch_seg)) if len(peak_high_pos) != len(peak_low_pos) + 1: print("Warning! Probably a bug...T.T") print(peak_high_pos, peak_low_pos) return None, None, None, None peak_high_pos_diff = np.diff(peak_high_pos) peak_low_pos_diff = np.diff(peak_low_pos) R = np.zeros(len(peak_high_pos_diff) + len(peak_low_pos_diff)) R[0::2] = peak_high_pos_diff R[1::2] = peak_low_pos_diff m_f_ind = np.zeros(len(R), dtype=int) m_f_ind[0::2] = peak_high_pos[:-1] m_f_ind[1::2] = peak_low_pos[:-1] m_f[m_f_ind] = (1 / R) * sr peak_high_pitch = pitch_seg[peak_high_pos] peak_low_pitch = pitch_seg[peak_low_pos] E = np.zeros(len(R)) E[0::2] = (peak_high_pitch[1:] + peak_high_pitch[:-1]) / 2 - peak_low_pitch E[1::2] = peak_high_pitch[1:-1] - (peak_low_pitch[1:] + peak_low_pitch[:-1]) / 2 m_a_ind = np.zeros(len(R), dtype=int) m_a_ind[0::2] = peak_low_pos m_a_ind[1::2] = peak_high_pos[1:-1] m_a[m_a_ind] = 0.5 * E rate = 1 / R.mean() * sr extent = 0.5 * E.mean() print(f"Rate: {rate}, Extent: {extent}") return R, E, m_a, m_f def compute_extent(pitch_seg): """Compute extent of a pitch segment Args: pitch_seg (np.ndarray): array of pitch Returns: np.ndarray: array of extent """ peak_high_pos = argrelmax(pitch_seg)[0] peak_low_pos = argrelmin(pitch_seg)[0] if len(peak_high_pos) == 1 or len(peak_low_pos) == 1: return np.array([-1]) if len(peak_high_pos) < len(peak_low_pos): peak_low_pos = peak_low_pos[:-2] elif len(peak_high_pos) == len(peak_low_pos): peak_low_pos = peak_low_pos[:-1] peak_high_pitch = pitch_seg[peak_high_pos] peak_low_pitch = pitch_seg[peak_low_pos] peak_high_pos_diff = np.diff(peak_high_pos) peak_low_pos_diff = np.diff(peak_low_pos) # TODO: would probably be a bug... if len(peak_high_pitch) != len(peak_low_pitch) + 1: return np.array([-1]) E = np.zeros(len(peak_high_pos_diff) + len(peak_low_pos_diff)) E[0::2] = (peak_high_pitch[1:] + peak_high_pitch[:-1]) / 2 - peak_low_pitch E[1::2] = peak_high_pitch[1:-1] - (peak_low_pitch[1:] + peak_low_pitch[:-1]) / 2 return E def extract_smoothed_f0(f0, sr, cutoff=8): """Extract smoothed f0 by low-pass filtering Note that the low-pass filter is only applied to voiced segments. Args: f0 (np.ndarray): array of f0 sr (int): sampling rate cutoff (float): cutoff frequency Returns: np.ndarray: array of smoothed f0 """ segments = nonzero_segments(f0) f0_smooth = f0.copy() for s, e in segments: f0_smooth[s:e] = lowpass_filter(f0[s:e], sr, cutoff=cutoff) return f0_smooth def extract_smoothed_continuous_f0(f0, sr, cutoff=20): """Extract smoothed continuous f0 by low-pass filtering Note that the input must be continuous F0 or log-F0. Args: f0 (np.ndarray): array of continuous f0 sr (int): sampling rate cutoff (float): initial cutoff frequency Returns: np.ndarray: array of smoothed continuous f0 """ is_2d = len(f0.shape) == 2 f0 = f0.reshape(-1) if is_2d else f0 # Ref: https://bit.ly/3SOePFw f0_smooth = lowpass_filter(f0, sr, cutoff=cutoff) # Fallback case: shound't happen I believe # NOTE: hard-coded for now next_cutoff = 50 while (f0_smooth < 0).any(): f0_smooth = lowpass_filter(f0, sr, cutoff=next_cutoff) next_cutoff *= 2 f0_smooth = f0_smooth.reshape(len(f0), 1) if is_2d else f0_smooth return f0_smooth def extract_vibrato_likelihood( f0_smooth, sr, win_length=32, n_fft=128, min_freq=3, max_freq=8 ): """Extract vibrato likelihood Args: f0_smooth (np.ndarray): array of smoothed f0 sr (int): sampling rate win_length (int): window length n_fft (int): FFT size min_freq (float): minimum frequency of the vibrato max_freq (float): maximum frequency of the vibrato Returns: np.ndarray: array of vibrato likelihood """ # STFT on 1st order diffference of F0 X = np.abs( librosa.stft( np.diff(f0_smooth), hop_length=1, win_length=win_length, n_fft=n_fft, window="hann", ) ) X_norm = X / (X.sum(0) + 1e-7) freq_per_bin = sr / n_fft min_freq_bin = int(min_freq / freq_per_bin) max_freq_bin = int(max_freq / freq_per_bin) # Compute vibrato likelhiood St = np.abs(np.diff(X_norm, axis=0)).sum(0) Ft = X_norm[min_freq_bin:max_freq_bin, :].sum(0) vibrato_likelihood = St * Ft return vibrato_likelihood def interp_vibrato(m_f): """Interpolate a sequence of vibrato parameter by linear interpolation Args: m_f (np.ndarray): array of vibrato parameter Returns: np.ndarray: array of vibrato parameter """ nonzero_indices = np.where(m_f > 0)[0] nonzero_indices = [0] + list(nonzero_indices) + [len(m_f) - 1] out = np.interp(np.arange(len(m_f)), nonzero_indices, m_f[nonzero_indices]) return out def extract_vibrato_parameters( pitch, vibrato_likelihood, sr=100, threshold=0.12, min_cross_count=5, min_extent=30, max_extent=150, interp_params=True, smooth_params=False, smooth_width=15, clip_extent=True, ): """Extract vibrato parameters Args: pitch (np.ndarray): array of pitch (smoothed f0) vibrato_likelihood (np.ndarray): array of vibrato likelihood sr (int): sampling rate threshold (float): threshold of vibrato likelihood min_cross_count (int): minimum number of cross points min_extent (int): minimum extent of vibrato (cent) max_extent (int): maximum extent of vibrato (cent) interp_params (bool): whether to interpolate vibrato parameters smooth_params (bool): whether to smooth vibrato parameters smooth_width (int): width of smoothing window clip_extent (bool): whether to clip extent Returns: tuple: tuple of vibrato parameters """ T = len(vibrato_likelihood) vibrato_flags = np.zeros(T, dtype=int) m_a = np.zeros(T) m_f = np.zeros(T) peak_high_pos = argrelmax(pitch)[0] peak_low_pos = argrelmin(pitch)[0] # iterate over every peak position peak_high_idx = 0 while peak_high_idx < len(peak_high_pos): peak_frame_idx = peak_high_pos[peak_high_idx] found = False if vibrato_likelihood[peak_frame_idx] > threshold: # Initial positions for vibrato section start_index = peak_frame_idx peaks = peak_low_pos[peak_low_pos > peak_frame_idx] if len(peaks) > 0: end_index = peaks[0] else: peak_high_idx += 1 continue next_start_peak_high_idx = -1 # Find a peak position that is close to the next non-speech segment # assuming that there's a non-speech segment right after vibrato # NOTE: we may want to remove this constraint peak_high_pos_rest = peak_high_pos[peak_high_pos > peak_frame_idx] for frame_idx in range(end_index, T): if pitch[frame_idx] <= 0: peaks = peak_high_pos_rest[peak_high_pos_rest < frame_idx] if len(peaks) > 0: end_index = peaks[-1] next_start_peak_high_idx = ( len(peak_high_pos[peak_high_pos < end_index]) + 1 ) break # Set the search width (backward) search_width_backward = 0 for frame_idx in range(start_index, 0, -1): if pitch[frame_idx] <= 0: peaks_backward = peak_high_pos[ (peak_high_pos < peak_frame_idx) & (peak_high_pos > frame_idx) ] if len(peaks_backward) > 0: backward = peaks_backward[0] search_width_backward = len( peak_high_pos[ (peak_high_pos > backward) & (peak_high_pos <= peak_frame_idx) ] ) break # Find a peak position that satisfies the following vibrato constraints # 1) more than 5 times crossing # 2) 30 ~ 150 cent oscillation estimate_start_index = start_index rate = 0 for peak_idx in range( max(peak_high_idx - search_width_backward, 0), peak_high_idx ): if peak_high_pos[peak_idx] >= T: break f0_seg = pitch[peak_high_pos[peak_idx] : end_index] # Check if the segment satisfies vibrato constraints m = f0_seg.mean() cross_count = len(np.where(np.diff(np.sign(f0_seg - m)))[0]) # Find the start_index so that the vibrato section has more than 5 crossing E = compute_extent(f0_seg) extent = 0.5 * E.mean() having_large_deviation = ((0.5 * E) > max_extent * 2).any() if ( cross_count >= min_cross_count and cross_count >= rate and extent >= min_extent and extent <= max_extent and not having_large_deviation and (E > 0).all() ): rate = cross_count estimate_start_index = peak_high_pos[peak_idx] start_index = estimate_start_index if rate >= min_cross_count: R, E, m_a_seg, m_f_seg = extract_vibrato_parameters_impl( pitch[start_index - 1 : end_index + 2], sr ) if m_a_seg is None: found = False break found = True vibrato_flags[start_index:end_index] = 1 if interp_params: m_a_seg = interp_vibrato(m_a_seg) m_f_seg = np.clip(interp_vibrato(m_f_seg), 3, 8) if smooth_params: m_a_seg = np.convolve( m_a_seg, np.ones(smooth_width) / smooth_width, mode="same" ) m_f_seg = np.convolve( m_f_seg, np.ones(smooth_width) / smooth_width, mode="same" ) if clip_extent: m_a_seg = np.clip(m_a_seg, min_extent, max_extent) m_a[start_index:end_index] = m_a_seg[1:-2] m_f[start_index:end_index] = m_f_seg[1:-2] assert next_start_peak_high_idx > peak_high_idx peak_high_idx = next_start_peak_high_idx if not found: peak_high_idx += 1 return vibrato_flags, m_a, m_f def gen_sine_vibrato(f0, sr, m_a, m_f, scale=1.0): """Generate F0 with sine-based vibrato Args: f0 (ndarray): fundamental frequency sr (int): sampling rate m_a (ndarray): amplitude of vibrato m_f (ndarray): frequency of vibrato scale (float): scale factor Returns: ndarray: F0 with sine-based vibrato """ f0_gen = f0.copy() voiced_end_indices = np.asarray([e for _, e in nonzero_segments(f0)]) for s, e in nonzero_segments(m_a): # limit vibrato rate to [3, 8] Hz m_f_seg = np.clip(m_f[s:e], 3, 8) # limit vibrato extent to [30, 150] cent m_a_seg = np.clip(m_a[s:e], 30, 150) cent = scale * m_a_seg * np.sin(2 * np.pi / sr * m_f_seg * np.arange(0, e - s)) new_f0 = f0[s:e] * np.exp(cent * np.log(2) / 1200) f0_gen[s:e] = new_f0 # NOTE: this is a hack to avoid discontinuity at the end of vibrato voiced_ends_next_to_vibrato = voiced_end_indices[voiced_end_indices > e] if len(voiced_ends_next_to_vibrato) > 0: voiced_end = voiced_ends_next_to_vibrato[0] f0_gen[s:voiced_end] = lowpass_filter(f0_gen[s:voiced_end], sr, cutoff=12) return f0_gen
18,639
30.863248
91
py
nnsvs
nnsvs-master/nnsvs/train_util.py
import os import random import shutil import sys import types from glob import glob from multiprocessing import Manager from os.path import join from pathlib import Path import hydra import joblib import librosa import librosa.display import matplotlib.pyplot as plt import mlflow import numpy as np import pysptk import pyworld import torch import torch.distributed as dist from hydra.utils import get_original_cwd, to_absolute_path from nnmnkwii import metrics from nnsvs.base import PredictionType from nnsvs.gen import gen_world_params from nnsvs.logger import getLogger from nnsvs.mdn import mdn_get_most_probable_sigma_and_mu from nnsvs.multistream import ( get_static_features, get_static_stream_sizes, get_windows, multi_stream_mlpg, select_streams, split_streams, ) from nnsvs.pitch import lowpass_filter, note_segments from nnsvs.util import MinMaxScaler, StandardScaler, init_seed, pad_2d from omegaconf import DictConfig, ListConfig, OmegaConf from sklearn.preprocessing import MinMaxScaler as SKMinMaxScaler from torch import nn, optim from torch.cuda.amp import GradScaler from torch.nn.parallel import DistributedDataParallel as DDP from torch.utils import data as data_utils from torch.utils.data.sampler import BatchSampler from torch.utils.tensorboard import SummaryWriter plt.style.use("seaborn-whitegrid") class ShuffleBatchSampler(BatchSampler): def __init__( self, batches, drop_last=False, shuffle=True, ): self.shuffle = shuffle self.batches = batches self.drop_last = drop_last def __iter__(self): batches = self.batches if self.shuffle: random.shuffle(batches) return iter(batches) def __len__(self): return len(self.batches) def log_params_from_omegaconf_dict(params): for param_name, element in params.items(): _explore_recursive(param_name, element) def _explore_recursive(parent_name, element): if isinstance(element, DictConfig): for k, v in element.items(): if isinstance(v, DictConfig) or isinstance(v, ListConfig): _explore_recursive(f"{parent_name}.{k}", v) else: mlflow.log_param(f"{parent_name}.{k}", v) elif isinstance(element, ListConfig): for i, v in enumerate(element): mlflow.log_param(f"{parent_name}.{i}", v) def num_trainable_params(model): """Count the number of trainable parameters in the model. Args: model (torch.nn.Module): Model to count the number of trainable parameters. Returns: int: Number of trainable parameters. """ parameters = filter(lambda p: p.requires_grad, model.parameters()) return sum([np.prod(p.size()) for p in parameters]) def get_filtered_files( data_root, logger, filter_long_segments=False, filter_num_frames=6000, filter_min_num_frames=0, ): files = sorted(glob(join(data_root, "*-feats.npy"))) if filter_long_segments: valid_files = [] num_filtered = 0 for path in files: length = len(np.load(path)) if length < filter_num_frames and length > filter_min_num_frames: valid_files.append(path) else: if logger is not None: logger.info(f"Filtered: {path} is too long or short: {length}") num_filtered += 1 if num_filtered > 0 and logger is not None: logger.info(f"Filtered {num_filtered} files") # Print stats of lengths if logger is not None: lengths = [len(np.load(f)) for f in files] logger.debug(f"[before] Size of dataset: {len(files)}") logger.debug(f"[before] maximum length: {max(lengths)}") logger.debug(f"[before] minimum length: {min(lengths)}") logger.debug(f"[before] mean length: {np.mean(lengths)}") logger.debug(f"[before] std length: {np.std(lengths)}") logger.debug(f"[before] median length: {np.median(lengths)}") files = valid_files lengths = [len(np.load(f)) for f in files] if logger is not None: logger.debug(f"[after] Size of dataset: {len(files)}") logger.debug(f"[after] maximum length: {max(lengths)}") logger.debug(f"[after] minimum length: {min(lengths)}") logger.debug(f"[after] mean length: {np.mean(lengths)}") logger.debug(f"[after] std length: {np.std(lengths)}") logger.debug(f"[after] median length: {np.median(lengths)}") else: lengths = [len(np.load(f)) for f in files] return files, lengths def _is_batch_full(batch, num_tokens, max_tokens, max_sentences): if len(batch) == 0: return 0 if len(batch) == max_sentences: return 1 if num_tokens > max_tokens: return 1 return 0 def batch_by_size( indices, num_tokens_fn, max_tokens=None, max_sentences=None, required_batch_size_multiple=1, ): """ Yield mini-batches of indices bucketed by size. Batches may contain sequences of different lengths. Args: indices (List[int]): ordered list of dataset indices num_tokens_fn (callable): function that returns the number of tokens at a given index max_tokens (int, optional): max number of tokens in each batch (default: None). max_sentences (int, optional): max number of sentences in each batch (default: None). required_batch_size_multiple (int, optional): require batch size to be a multiple of N (default: 1). """ max_tokens = max_tokens if max_tokens is not None else sys.maxsize max_sentences = max_sentences if max_sentences is not None else sys.maxsize bsz_mult = required_batch_size_multiple if isinstance(indices, types.GeneratorType): indices = np.fromiter(indices, dtype=np.int64, count=-1) sample_len = 0 sample_lens = [] batch = [] batches = [] for i in range(len(indices)): idx = indices[i] num_tokens = num_tokens_fn(idx) sample_lens.append(num_tokens) sample_len = max(sample_len, num_tokens) assert ( sample_len <= max_tokens ), "sentence at index {} of size {} exceeds max_tokens " "limit of {}!".format( idx, sample_len, max_tokens ) num_tokens = (len(batch) + 1) * sample_len if _is_batch_full(batch, num_tokens, max_tokens, max_sentences): mod_len = max( bsz_mult * (len(batch) // bsz_mult), len(batch) % bsz_mult, ) batches.append(batch[:mod_len]) batch = batch[mod_len:] sample_lens = sample_lens[mod_len:] sample_len = max(sample_lens) if len(sample_lens) > 0 else 0 batch.append(idx) if len(batch) > 0: batches.append(batch) return batches class Dataset(data_utils.Dataset): # type: ignore """Dataset for numpy files Args: in_paths (list): List of paths to input files out_paths (list): List of paths to output files """ def __init__(self, in_paths, out_paths, lengths, shuffle=False, allow_cache=True): self.in_paths = in_paths self.out_paths = out_paths self.lengths = lengths self.sort_by_len = True self.shuffle = shuffle self.allow_cache = allow_cache if allow_cache: self.manager = Manager() self.caches = self.manager.list() self.caches += [() for _ in range(len(in_paths))] def __getitem__(self, idx): """Get a pair of input and target Args: idx (int): index of the pair Returns: tuple: input and target in numpy format """ if self.allow_cache and len(self.caches[idx]) != 0: return self.caches[idx] x, y = np.load(self.in_paths[idx]), np.load(self.out_paths[idx]) if self.allow_cache: self.caches[idx] = (x, y) return x, y def num_tokens(self, index): return self.lengths[index] def ordered_indices(self): """Return an ordered list of indices. Batches will be constructed based on this order. """ if self.shuffle: indices = np.random.permutation(len(self)) if self.sort_by_len: indices = indices[ np.argsort(np.array(self.lengths)[indices], kind="mergesort") ] else: indices = np.arange(len(self)) return indices def __len__(self): """Returns the size of the dataset Returns: int: size of the dataset """ return len(self.in_paths) def ensure_divisible_by(feats, N): """Ensure that the number of frames is divisible by N. Args: feats (np.ndarray): Input features. N (int): Target number of frames. Returns: np.ndarray: Input features with number of frames divisible by N. """ if N == 1: return feats mod = len(feats) % N if mod != 0: feats = feats[: len(feats) - mod] return feats def collate_fn_default(batch, reduction_factor=1, stream_sizes=None, streams=None): """Create batch Args: batch(tuple): List of tuples - x[0] (ndarray,int) : list of (T, D_in) - x[1] (ndarray,int) : list of (T, D_out) reduction_factor (int): Reduction factor. Returns: tuple: Tuple of batch - x (FloatTensor) : Network inputs (B, max(T), D_in) - y (FloatTensor) : Network targets (B, max(T), D_out) - lengths (LongTensor): Input lengths """ lengths = [len(ensure_divisible_by(x[0], reduction_factor)) for x in batch] max_len = max(lengths) x_batch = torch.stack( [ torch.from_numpy( pad_2d(ensure_divisible_by(x[0], reduction_factor), max_len) ) for x in batch ] ) if stream_sizes is not None: assert streams is not None y_batch = torch.stack( [ torch.from_numpy( pad_2d( ensure_divisible_by( select_streams(x[1], stream_sizes, streams), reduction_factor, ), max_len, ) ) for x in batch ] ) else: y_batch = torch.stack( [ torch.from_numpy( pad_2d(ensure_divisible_by(x[1], reduction_factor), max_len) ) for x in batch ] ) l_batch = torch.tensor(lengths, dtype=torch.long) return x_batch, y_batch, l_batch def collate_fn_random_segments(batch, max_time_frames=256): """Collate function with random segments Use segmented frames instead of padded entire frames. No padding is performed. .. warning:: max_time_frames must be larger than the shortest sequence in the training data. Args: batch (tuple): tupls of lit - x[0] (ndarray,int) : list of (T, D_in) - x[1] (ndarray,int) : list of (T, D_out) max_time_frames (int, optional): Number of time frames. Defaults to 256. Returns: tuple: Tuple of batch - x (FloatTensor) : Network inputs (B, max(T), D_in) - y (FloatTensor) : Network targets (B, max(T), D_out) - lengths (LongTensor): Input lengths """ xs, ys = [b[0] for b in batch], [b[1] for b in batch] lengths = [len(x[0]) for x in batch] start_frames = np.array( [np.random.randint(0, xl - max_time_frames) for xl in lengths] ) starts = start_frames ends = starts + max_time_frames x_cut = [torch.from_numpy(x[s:e]) for x, s, e in zip(xs, starts, ends)] y_cut = [torch.from_numpy(y[s:e]) for y, s, e in zip(ys, starts, ends)] x_batch = torch.stack(x_cut).float() y_batch = torch.stack(y_cut).float() # NOTE: we don't actually need lengths since we don't perform padding # but just for consistency with collate_fn_default l_batch = torch.tensor([max_time_frames] * len(lengths), dtype=torch.long) return x_batch, y_batch, l_batch def get_data_loaders(data_config, collate_fn, logger): """Get data loaders for training and validation. Args: data_config (dict): Data configuration. collate_fn (callable): Collate function. logger (logging.Logger): Logger. Returns: dict: Data loaders. """ if "filter_long_segments" not in data_config: logger.warning( "filter_long_segments is not found in the data config. Consider set it explicitly." ) logger.info("Disable filtering for long segments.") filter_long_segments = False else: filter_long_segments = data_config.filter_long_segments if "filter_num_frames" not in data_config: logger.warning( "filter_num_frames is not found in the data config. Consider set it explicitly." ) filter_num_frames = 6000 filter_min_num_frames = 0 else: filter_num_frames = data_config.filter_num_frames filter_min_num_frames = data_config.filter_min_num_frames data_loaders = {} samplers = {} for phase in ["train_no_dev", "dev"]: in_dir = to_absolute_path(data_config[phase].in_dir) out_dir = to_absolute_path(data_config[phase].out_dir) train = phase.startswith("train") in_files, lengths = get_filtered_files( in_dir, logger, filter_long_segments=filter_long_segments, filter_num_frames=filter_num_frames, filter_min_num_frames=filter_min_num_frames, ) out_files, _ = get_filtered_files( out_dir, None, filter_long_segments=filter_long_segments, filter_num_frames=filter_num_frames, filter_min_num_frames=filter_min_num_frames, ) # Dynamic batch size if data_config.batch_max_frames > 0: logger.debug( f"Dynamic batch size with batch_max_frames={data_config.batch_max_frames}" ) dataset = Dataset( in_files, out_files, lengths, shuffle=train, allow_cache=data_config.get("allow_cache", False), ) if dist.is_initialized(): required_batch_size_multiple = dist.get_world_size() else: required_batch_size_multiple = 1 indices = dataset.ordered_indices() batches = batch_by_size( indices, dataset.num_tokens, max_tokens=data_config.batch_max_frames, required_batch_size_multiple=required_batch_size_multiple, ) # Split mini-batches for each rank manually if dist.is_initialized(): num_replicas = dist.get_world_size() rank = dist.get_rank() logger.debug(f"Splitting mini-batches for rank {rank}") batches = [ x[rank::num_replicas] for x in batches if len(x) % num_replicas == 0 ] logger.info(f"Num mini-batches: {len(batches)}") for batch in batches: sizes = [dataset.num_tokens(i) for i in batch] logger.debug(f"Batch-size: {len(batch)}, Lens: {sizes}") logger.info(f"Average batch size: {np.mean([len(b) for b in batches])}") data_loader_extra_kwargs = { "batch_sampler": ShuffleBatchSampler(batches) if train else batches, } sampler = None else: logger.debug(f"Fixed batch size: {data_config.batch_size}") dataset = Dataset(in_files, out_files, lengths) if dist.is_initialized(): sampler = torch.utils.data.distributed.DistributedSampler( dataset, shuffle=train ) shuffle = False else: sampler = None shuffle = train data_loader_extra_kwargs = { "batch_size": data_config.batch_size, "sampler": sampler, "shuffle": shuffle, } data_loaders[phase] = data_utils.DataLoader( dataset, collate_fn=collate_fn, pin_memory=data_config.pin_memory, num_workers=data_config.num_workers, **data_loader_extra_kwargs, ) samplers[phase] = sampler return data_loaders, samplers def set_epochs_based_on_max_steps_(train_config, steps_per_epoch, logger): """Set epochs based on max steps. Args: train_config (TrainConfig): Train config. steps_per_epoch (int): Number of steps per epoch. logger (logging.Logger): Logger. """ if "max_train_steps" not in train_config: logger.warning("max_train_steps is not found in the train config.") return logger.info(f"Number of iterations per epoch: {steps_per_epoch}") if train_config.max_train_steps < 0: # Set max_train_steps based on nepochs max_train_steps = train_config.nepochs * steps_per_epoch train_config.max_train_steps = max_train_steps logger.info( "Number of max_train_steps is set based on nepochs: {}".format( max_train_steps ) ) else: # Set nepochs based on max_train_steps max_train_steps = train_config.max_train_steps epochs = int(np.ceil(max_train_steps / steps_per_epoch)) train_config.nepochs = epochs logger.info( "Number of epochs is set based on max_train_steps: {}".format(epochs) ) logger.info(f"Number of epochs: {train_config.nepochs}") logger.info(f"Number of iterations: {train_config.max_train_steps}") def save_checkpoint( logger, out_dir, model, optimizer, lr_scheduler, epoch, is_best=False, postfix="", ): """Save a checkpoint. Args: logger (logging.Logger): Logger. out_dir (str): Output directory. model (nn.Module): Model. optimizer (Optimizer): Optimizer. lr_scheduler (LRScheduler): Learning rate scheduler. epoch (int): Current epoch. is_best (bool, optional): Whether or not the current model is the best. Defaults to False. postfix (str, optional): Postfix. Defaults to "". """ if dist.is_initialized() and dist.get_rank() != 0: return if isinstance(model, nn.DataParallel) or isinstance(model, DDP): model = model.module out_dir.mkdir(parents=True, exist_ok=True) if is_best: path = out_dir / f"best_loss{postfix}.pth" else: path = out_dir / "epoch{:04d}{}.pth".format(epoch, postfix) torch.save( { "state_dict": model.state_dict(), "optimizer_state": optimizer.state_dict(), "lr_scheduler_state": lr_scheduler.state_dict(), }, path, ) logger.info(f"Saved checkpoint at {path}") if not is_best: shutil.copyfile(path, out_dir / f"latest{postfix}.pth") def get_stream_weight(stream_weights, stream_sizes): if stream_weights is not None: assert len(stream_weights) == len(stream_sizes) return torch.tensor(stream_weights) S = sum(stream_sizes) w = torch.tensor(stream_sizes).float() / S return w def _instantiate_optim(optim_config, model): # Optimizer optimizer_class = getattr(optim, optim_config.optimizer.name) optimizer = optimizer_class(model.parameters(), **optim_config.optimizer.params) # Scheduler lr_scheduler_class = getattr(optim.lr_scheduler, optim_config.lr_scheduler.name) lr_scheduler = lr_scheduler_class(optimizer, **optim_config.lr_scheduler.params) return optimizer, lr_scheduler def _resume(logger, resume_config, model, optimizer, lr_scheduler): if resume_config.checkpoint is not None and len(resume_config.checkpoint) > 0: logger.info("Load weights from %s", resume_config.checkpoint) checkpoint = torch.load(to_absolute_path(resume_config.checkpoint)) state_dict = checkpoint["state_dict"] model_dict = model.state_dict() valid_state_dict = { k: v for k, v in state_dict.items() if (k in model_dict) and (v.shape == model_dict[k].shape) } non_valid_state_dict = { k: v for k, v in state_dict.items() if k not in valid_state_dict } if len(non_valid_state_dict) > 0: for k, _ in non_valid_state_dict.items(): logger.warning(f"Skip loading {k} from checkpoint") model_dict.update(valid_state_dict) model.load_state_dict(model_dict) if resume_config.load_optimizer: logger.info("Load optimizer state") optimizer.load_state_dict(checkpoint["optimizer_state"]) lr_scheduler.load_state_dict(checkpoint["lr_scheduler_state"]) def setup(config, device, collate_fn=collate_fn_default): """Setup for training Args: config (dict): configuration for training device (torch.device): device to use for training collate_fn (callable, optional): collate function. Defaults to collate_fn_default. Returns: (tuple): tuple containing model, optimizer, learning rate scheduler, data loaders, tensorboard writer, logger, and scalers. """ if dist.is_initialized(): rank = dist.get_rank() logger = getLogger(config.verbose) if rank == 0 else getLogger(0) sys.stdout = open(os.devnull, "w") if rank != 0 else sys.stdout else: logger = getLogger(config.verbose) rank = 0 logger.info(OmegaConf.to_yaml(config)) logger.info(f"PyTorch version: {torch.__version__}") if torch.cuda.is_available(): from torch.backends import cudnn cudnn.benchmark = config.train.cudnn.benchmark cudnn.deterministic = config.train.cudnn.deterministic logger.info(f"cudnn.deterministic: {cudnn.deterministic}") logger.info(f"cudnn.benchmark: {cudnn.benchmark}") if torch.backends.cudnn.version() is not None: logger.info(f"cuDNN version: {torch.backends.cudnn.version()}") logger.info(f"Random seed: {config.seed}") init_seed(config.seed) if config.train.use_detect_anomaly: torch.autograd.set_detect_anomaly(True) logger.info("Set to use torch.autograd.detect_anomaly") if "use_amp" in config.train and config.train.use_amp: logger.info("Use mixed precision training") grad_scaler = GradScaler() else: grad_scaler = None # Model model = hydra.utils.instantiate(config.model.netG).to(device) logger.info(model) logger.info( "Number of trainable params: {:.3f} million".format( num_trainable_params(model) / 1000000.0 ) ) # Distributed training if dist.is_initialized(): device_id = rank % torch.cuda.device_count() model = DDP(model, device_ids=[device_id]) # Optimizer optimizer_class = getattr(optim, config.train.optim.optimizer.name) optimizer = optimizer_class( model.parameters(), **config.train.optim.optimizer.params ) # Scheduler lr_scheduler_class = getattr( optim.lr_scheduler, config.train.optim.lr_scheduler.name ) lr_scheduler = lr_scheduler_class( optimizer, **config.train.optim.lr_scheduler.params ) # DataLoader data_loaders, samplers = get_data_loaders(config.data, collate_fn, logger) set_epochs_based_on_max_steps_( config.train, len(data_loaders["train_no_dev"]), logger ) # Resume _resume(logger, config.train.resume, model, optimizer, lr_scheduler) if config.data_parallel: model = nn.DataParallel(model) # Mlflow if config.mlflow.enabled: mlflow.set_tracking_uri("file://" + get_original_cwd() + "/mlruns") mlflow.set_experiment(config.mlflow.experiment) # NOTE: disable tensorboard if mlflow is enabled writer = None logger.info("Using mlflow instead of tensorboard") else: # Tensorboard if rank == 0: writer = SummaryWriter(to_absolute_path(config.train.log_dir)) else: writer = None # Scalers if "in_scaler_path" in config.data and config.data.in_scaler_path is not None: in_scaler = joblib.load(to_absolute_path(config.data.in_scaler_path)) in_scaler = MinMaxScaler( in_scaler.min_, in_scaler.scale_, in_scaler.data_min_, in_scaler.data_max_ ) else: in_scaler = None if "out_scaler_path" in config.data and config.data.out_scaler_path is not None: out_scaler = joblib.load(to_absolute_path(config.data.out_scaler_path)) out_scaler = StandardScaler( out_scaler.mean_, out_scaler.var_, out_scaler.scale_ ) else: out_scaler = None return ( model, optimizer, lr_scheduler, grad_scaler, data_loaders, samplers, writer, logger, in_scaler, out_scaler, ) def setup_gan(config, device, collate_fn=collate_fn_default): """Setup for training GAN Args: config (dict): configuration for training device (torch.device): device to use for training collate_fn (callable, optional): collate function. Defaults to collate_fn_default. Returns: (tuple): tuple containing model, optimizer, learning rate scheduler, data loaders, tensorboard writer, logger, and scalers. """ if dist.is_initialized(): rank = dist.get_rank() logger = getLogger(config.verbose) if rank == 0 else getLogger(0) sys.stdout = open(os.devnull, "w") if rank != 0 else sys.stdout else: logger = getLogger(config.verbose) rank = 0 logger.info(OmegaConf.to_yaml(config)) logger.info(f"PyTorch version: {torch.__version__}") if torch.cuda.is_available(): from torch.backends import cudnn cudnn.benchmark = config.train.cudnn.benchmark cudnn.deterministic = config.train.cudnn.deterministic logger.info(f"cudnn.deterministic: {cudnn.deterministic}") logger.info(f"cudnn.benchmark: {cudnn.benchmark}") if torch.backends.cudnn.version() is not None: logger.info(f"cuDNN version: {torch.backends.cudnn.version()}") logger.info(f"Random seed: {config.seed}") init_seed(config.seed) if config.train.use_detect_anomaly: torch.autograd.set_detect_anomaly(True) logger.info("Set to use torch.autograd.detect_anomaly") if "use_amp" in config.train and config.train.use_amp: logger.info("Use mixed precision training") grad_scaler = GradScaler() else: grad_scaler = None # Model G netG = hydra.utils.instantiate(config.model.netG).to(device) logger.info(netG) logger.info( "[Generator] Number of trainable params: {:.3f} million".format( num_trainable_params(netG) / 1000000.0 ) ) if dist.is_initialized(): device_id = rank % torch.cuda.device_count() netG = DDP(netG, device_ids=[device_id]) # Optimizer and LR scheduler for G optG, schedulerG = _instantiate_optim(config.train.optim.netG, netG) # Model D netD = hydra.utils.instantiate(config.model.netD).to(device) logger.info(netD) logger.info( "[Discriminator] Number of trainable params: {:.3f} million".format( num_trainable_params(netD) / 1000000.0 ) ) if dist.is_initialized(): device_id = rank % torch.cuda.device_count() netD = DDP(netD, device_ids=[device_id]) # Optimizer and LR scheduler for D optD, schedulerD = _instantiate_optim(config.train.optim.netD, netD) # DataLoader data_loaders, samplers = get_data_loaders(config.data, collate_fn, logger) set_epochs_based_on_max_steps_( config.train, len(data_loaders["train_no_dev"]), logger ) # Resume _resume(logger, config.train.resume.netG, netG, optG, schedulerG) _resume(logger, config.train.resume.netD, netD, optD, schedulerD) if config.data_parallel: netG = nn.DataParallel(netG) netD = nn.DataParallel(netD) # Mlflow if config.mlflow.enabled: mlflow.set_tracking_uri("file://" + get_original_cwd() + "/mlruns") mlflow.set_experiment(config.mlflow.experiment) # NOTE: disable tensorboard if mlflow is enabled writer = None logger.info("Using mlflow instead of tensorboard") else: # Tensorboard writer = SummaryWriter(to_absolute_path(config.train.log_dir)) # Scalers if "in_scaler_path" in config.data and config.data.in_scaler_path is not None: in_scaler = joblib.load(to_absolute_path(config.data.in_scaler_path)) if isinstance(in_scaler, SKMinMaxScaler): in_scaler = MinMaxScaler( in_scaler.min_, in_scaler.scale_, in_scaler.data_min_, in_scaler.data_max_, ) else: in_scaler = None if "out_scaler_path" in config.data and config.data.out_scaler_path is not None: out_scaler = joblib.load(to_absolute_path(config.data.out_scaler_path)) out_scaler = StandardScaler( out_scaler.mean_, out_scaler.var_, out_scaler.scale_ ) else: out_scaler = None return ( (netG, optG, schedulerG), (netD, optD, schedulerD), grad_scaler, data_loaders, samplers, writer, logger, in_scaler, out_scaler, ) def save_configs(config): out_dir = Path(to_absolute_path(config.train.out_dir)) out_dir.mkdir(parents=True, exist_ok=True) with open(out_dir / "model.yaml", "w") as f: OmegaConf.save(config.model, f) with open(out_dir / "config.yaml", "w") as f: OmegaConf.save(config, f) def check_resf0_config(logger, model, config, in_scaler, out_scaler): logger.info("Checking model configs for residual F0 prediction") if in_scaler is None or out_scaler is None: raise ValueError("in_scaler and out_scaler must be specified") if isinstance(model, nn.DataParallel) or isinstance(model, DDP): model = model.module in_lf0_idx = config.data.in_lf0_idx in_rest_idx = config.data.in_rest_idx out_lf0_idx = config.data.out_lf0_idx if in_lf0_idx is None or in_rest_idx is None or out_lf0_idx is None: raise ValueError("in_lf0_idx, in_rest_idx and out_lf0_idx must be specified") logger.info("in_lf0_idx: %s", in_lf0_idx) logger.info("in_rest_idx: %s", in_rest_idx) logger.info("out_lf0_idx: %s", out_lf0_idx) ok = True if hasattr(model, "in_lf0_idx"): if model.in_lf0_idx != in_lf0_idx: logger.warning( "in_lf0_idx in model and data config must be same", model.in_lf0_idx, in_lf0_idx, ) ok = False if hasattr(model, "out_lf0_idx"): if model.out_lf0_idx != out_lf0_idx: logger.warning( "out_lf0_idx in model and data config must be same", model.out_lf0_idx, out_lf0_idx, ) ok = False if hasattr(model, "in_lf0_min") and hasattr(model, "in_lf0_max"): # Inject values from the input scaler if model.in_lf0_min is None or model.in_lf0_max is None: model.in_lf0_min = in_scaler.data_min_[in_lf0_idx] model.in_lf0_max = in_scaler.data_max_[in_lf0_idx] logger.info("in_lf0_min: %s", model.in_lf0_min) logger.info("in_lf0_max: %s", model.in_lf0_max) if not np.allclose(model.in_lf0_min, in_scaler.data_min_[model.in_lf0_idx]): logger.warning( f"in_lf0_min is set to {model.in_lf0_min}, " f"but should be {in_scaler.data_min_[model.in_lf0_idx]}" ) ok = False if not np.allclose(model.in_lf0_max, in_scaler.data_max_[model.in_lf0_idx]): logger.warning( f"in_lf0_max is set to {model.in_lf0_max}, " f"but should be {in_scaler.data_max_[model.in_lf0_idx]}" ) ok = False if hasattr(model, "out_lf0_mean") and hasattr(model, "out_lf0_scale"): # Inject values from the output scaler if model.out_lf0_mean is None or model.out_lf0_scale is None: model.out_lf0_mean = float(out_scaler.mean_[out_lf0_idx]) model.out_lf0_scale = float(out_scaler.scale_[out_lf0_idx]) logger.info("model.out_lf0_mean: %s", model.out_lf0_mean) logger.info("model.out_lf0_scale: %s", model.out_lf0_scale) if not np.allclose(model.out_lf0_mean, out_scaler.mean_[model.out_lf0_idx]): logger.warning( f"out_lf0_mean is set to {model.out_lf0_mean}, " f"but should be {out_scaler.mean_[model.out_lf0_idx]}" ) ok = False if not np.allclose(model.out_lf0_scale, out_scaler.scale_[model.out_lf0_idx]): logger.warning( f"out_lf0_scale is set to {model.out_lf0_scale}, " f"but should be {out_scaler.scale_[model.out_lf0_idx]}" ) ok = False if not ok: if ( model.in_lf0_idx == in_lf0_idx and hasattr(model, "in_lf0_min") and hasattr(model, "out_lf0_mean") ): logger.info( f""" If you are 100% sure that you set model.in_lf0_idx and model.out_lf0_idx correctly, Please consider the following parameters in your model config: in_lf0_idx: {model.in_lf0_idx} out_lf0_idx: {model.out_lf0_idx} in_lf0_min: {in_scaler.data_min_[model.in_lf0_idx]} in_lf0_max: {in_scaler.data_max_[model.in_lf0_idx]} out_lf0_mean: {out_scaler.mean_[model.out_lf0_idx]} out_lf0_scale: {out_scaler.scale_[model.out_lf0_idx]} """ ) raise ValueError("The model config has wrong configurations.") # Overwrite the parameters to the config for key in ["in_lf0_min", "in_lf0_max", "out_lf0_mean", "out_lf0_scale"]: if hasattr(model, key): config.model.netG[key] = float(getattr(model, key)) def compute_pitch_regularization_weight(segments, N, decay_size=25, max_w=0.5): """Compute pitch regularization weight given note segments Args: segments (list): list of note (start, end) indices N (int): number of frames decay_size (int): size of the decay window max_w (float): maximum weight Returns: Tensor: weights of shape (N,) """ w = torch.zeros(N) for s, e in segments: L = e - s w[s:e] = max_w if L > decay_size * 2: w[s : s + decay_size] *= torch.arange(decay_size) / decay_size w[e - decay_size : e] *= torch.arange(decay_size - 1, -1, -1) / decay_size else: # For shote notes (less than decay_size*0.01 sec) we don't use pitch regularization w[s:e] = 0.0 return w def compute_batch_pitch_regularization_weight(lf0_score_denorm, decay_size): """Batch version of computing pitch regularization weight Args: lf0_score_denorm (Tensor): (B, T) Returns: Tensor: weights of shape (B, N, 1) """ B, T = lf0_score_denorm.shape w = torch.zeros_like(lf0_score_denorm) for idx in range(len(lf0_score_denorm)): segments = note_segments(lf0_score_denorm[idx]) w[idx, :] = compute_pitch_regularization_weight( segments, N=T, decay_size=decay_size ).to(w.device) return w.unsqueeze(-1) @torch.no_grad() def compute_distortions(pred_out_feats, out_feats, lengths, out_scaler, model_config): """Compute distortion measures between predicted and ground-truth acoustic features Args: pred_out_feats (nn.Tensor): predicted acoustic features out_feats (nn.Tensor): ground-truth acoustic features lengths (nn.Tensor): lengths of the sequences out_scaler (nn.Module): scaler to denormalize features model_config (dict): model configuration Returns: dict: a dict that includes MCD for mgc/bap, V/UV error and F0 RMSE """ out_feats = out_scaler.inverse_transform(out_feats) pred_out_feats = out_scaler.inverse_transform(pred_out_feats) out_streams = get_static_features( out_feats, model_config.num_windows, model_config.stream_sizes, model_config.has_dynamic_features, ) pred_out_streams = get_static_features( pred_out_feats, model_config.num_windows, model_config.stream_sizes, model_config.has_dynamic_features, ) if len(out_streams) >= 4: mgc, lf0, vuv, bap = ( out_streams[0], out_streams[1], out_streams[2], out_streams[3], ) pred_mgc, pred_lf0, pred_vuv, pred_bap = ( pred_out_streams[0], pred_out_streams[1], pred_out_streams[2], pred_out_streams[3], ) elif len(out_streams) == 3: mgc, lf0, vuv = out_streams[0], out_streams[1], out_streams[2] pred_mgc, pred_lf0, pred_vuv = ( pred_out_streams[0], pred_out_streams[1], pred_out_streams[2], ) bap = None pred_bap = None # binarize vuv vuv, pred_vuv = (vuv > 0.5).float(), (pred_vuv > 0.5).float() dist = { "ObjEval_MGC_MCD": metrics.melcd( mgc[:, :, 1:], pred_mgc[:, :, 1:], lengths=lengths ), "ObjEval_VUV_ERR": metrics.vuv_error(vuv, pred_vuv, lengths=lengths), } if bap is not None: dist["ObjEval_BAP_MCD"] = metrics.melcd(bap, pred_bap, lengths=lengths) / 10.0 try: f0_mse = metrics.lf0_mean_squared_error( lf0, vuv, pred_lf0, pred_vuv, lengths=lengths, linear_domain=True ) dist["ObjEval_F0_RMSE"] = np.sqrt(f0_mse) except ZeroDivisionError: pass return dist @torch.no_grad() def eval_pitch_model( phase, step, netG, in_feats, out_feats, lengths, model_config, out_scaler, writer, sr, lf0_score_denorm, max_num_eval_utts=10, ): if dist.is_initialized() and dist.get_rank() != 0: return if writer is None: return # make sure to be in eval mode netG.eval() prediction_type = ( netG.module.prediction_type() if isinstance(netG, nn.DataParallel) or isinstance(netG, DDP) else netG.prediction_type() ) utt_indices = list(range(max_num_eval_utts)) utt_indices = utt_indices[: min(len(utt_indices), len(in_feats))] assert not np.any(model_config.has_dynamic_features) for utt_idx in utt_indices: out_feats_denorm_ = out_scaler.inverse_transform( out_feats[utt_idx, : lengths[utt_idx]].unsqueeze(0) ) lf0 = out_feats_denorm_.squeeze(0).cpu().numpy().reshape(-1) lf0_score_denorm_ = ( lf0_score_denorm[utt_idx, : lengths[utt_idx]].cpu().numpy().reshape(-1) ) # Run forward outs = netG( in_feats[utt_idx, : lengths[utt_idx]].unsqueeze(0), [lengths[utt_idx]], out_feats[utt_idx, : lengths[utt_idx]].unsqueeze(0), ) # ResF0 case if netG.has_residual_lf0_prediction(): outs, _ = outs if prediction_type == PredictionType.PROBABILISTIC: pi, sigma, mu = outs pred_out_feats = mdn_get_most_probable_sigma_and_mu(pi, sigma, mu)[1] else: pred_out_feats = outs # NOTE: multiple outputs if isinstance(pred_out_feats, list): pred_out_feats = pred_out_feats[-1] if isinstance(pred_out_feats, tuple): pred_out_feats = pred_out_feats[0] if not isinstance(pred_out_feats, list): pred_out_feats = [pred_out_feats] # Run inference if prediction_type == PredictionType.PROBABILISTIC: if isinstance(netG, nn.DataParallel) or isinstance(netG, DDP): inference_out_feats, _ = netG.module.inference( in_feats[utt_idx, : lengths[utt_idx]].unsqueeze(0), [lengths[utt_idx]], ) else: inference_out_feats, _ = netG.inference( in_feats[utt_idx, : lengths[utt_idx]].unsqueeze(0), [lengths[utt_idx]], ) else: if isinstance(netG, nn.DataParallel) or isinstance(netG, DDP): inference_out_feats = netG.module.inference( in_feats[utt_idx, : lengths[utt_idx]].unsqueeze(0), [lengths[utt_idx]], ) else: inference_out_feats = netG.inference( in_feats[utt_idx, : lengths[utt_idx]].unsqueeze(0), [lengths[utt_idx]], ) pred_out_feats.append(inference_out_feats) assert len(pred_out_feats) == 2 for idx, pred_out_feats_ in enumerate(pred_out_feats): pred_out_feats_ = pred_out_feats_.squeeze(0).cpu().numpy() pred_lf0 = ( out_scaler.inverse_transform( torch.from_numpy(pred_out_feats_).to(in_feats.device) ) .cpu() .numpy() ).reshape(-1) if idx == 1: group = f"{phase}_utt{np.abs(utt_idx)}_inference" else: group = f"{phase}_utt{np.abs(utt_idx)}_forward" # Continuous log-F0 fig, ax = plt.subplots(1, 1, figsize=(8, 3)) timeaxis = np.arange(len(lf0)) * 0.005 ax.plot( timeaxis, lf0, linewidth=1.5, color="tab:blue", label="Target log-F0" ) ax.plot( timeaxis, pred_lf0, linewidth=1.5, color="tab:orange", label="Predicted log-F0", ) ax.plot( timeaxis, lf0_score_denorm_, "--", color="gray", linewidth=1.3, label="Note log-F0", ) ax.set_xlabel("Time [sec]") ax.set_ylabel("Log-frequency [Hz]") ax.set_xlim(timeaxis[0], timeaxis[-1]) ax.set_ylim( min( min(lf0_score_denorm_[lf0_score_denorm_ > 0]), min(lf0), min(pred_lf0), ) - 0.1, max(max(lf0_score_denorm_), max(lf0), max(pred_lf0)) + 0.1, ) plt.legend(loc="upper right", borderaxespad=0, ncol=3) plt.tight_layout() writer.add_figure(f"{group}/ContinuousLogF0", fig, step) plt.close() # F0 lf0_score = lf0_score_denorm_.copy() note_indices = lf0_score > 0 lf0_score[note_indices] = np.exp(lf0_score[note_indices]) fig, ax = plt.subplots(1, 1, figsize=(8, 3)) timeaxis = np.arange(len(lf0)) * 0.005 ax.plot( timeaxis, np.exp(lf0) * np.sign(lf0_score), linewidth=1.5, color="tab:blue", label="Target F0", ) ax.plot( timeaxis, np.exp(pred_lf0) * np.sign(lf0_score), linewidth=1.5, color="tab:orange", label="Predicted F0", ) ax.plot( timeaxis, lf0_score, "--", linewidth=1.3, color="gray", label="Note F0" ) ax.set_xlabel("Time [sec]") ax.set_ylabel("Frequency [Hz]") ax.set_xlim(timeaxis[0], timeaxis[-1]) ax.set_ylim( min( min(lf0_score[lf0_score > 0]), min(np.exp(lf0)), min(np.exp(pred_lf0)), ) - 10, max(max(lf0_score), max(np.exp(lf0)), max(np.exp(pred_lf0))) + 10, ) plt.legend(loc="upper right", borderaxespad=0, ncol=3) plt.tight_layout() writer.add_figure(f"{group}/F0", fig, step) plt.close() def synthesize( device, mgc, lf0, vuv, bap, sr, use_world_codec=False, vuv_threshold=0.3, vocoder=None, vocoder_in_scaler=None, vocoder_config=None, ): if vocoder is not None: is_usfgan = "generator" in vocoder_config and "discriminator" in vocoder_config assert vocoder_in_scaler is not None if not is_usfgan: # NOTE: So far vocoder models are trained on binary V/UV features vuv = (vuv > vuv_threshold).astype(np.float32) voc_inp = ( torch.from_numpy( vocoder_in_scaler.transform( np.concatenate([mgc, lf0, vuv, bap], axis=-1) ) ) .float() .to(device) ) wav = vocoder.inference(voc_inp).view(-1).to("cpu").numpy() else: fftlen = pyworld.get_cheaptrick_fft_size(sr) use_mcep_aperiodicity = bap.shape[-1] > 5 if use_mcep_aperiodicity: mcep_aperiodicity_order = bap.shape[-1] - 1 alpha = pysptk.util.mcepalpha(sr) aperiodicity = pysptk.mc2sp( np.ascontiguousarray(bap).astype(np.float64), fftlen=fftlen, alpha=alpha, ) else: aperiodicity = pyworld.decode_aperiodicity( np.ascontiguousarray(bap).astype(np.float64), sr, fftlen ) # fill aperiodicity with ones for unvoiced regions aperiodicity[vuv.reshape(-1) < vuv_threshold, 0] = 1.0 # WORLD fails catastrophically for out of range aperiodicity aperiodicity = np.clip(aperiodicity, 0.0, 1.0) # back to bap if use_mcep_aperiodicity: bap = pysptk.sp2mc( aperiodicity, order=mcep_aperiodicity_order, alpha=alpha, ) else: bap = pyworld.code_aperiodicity(aperiodicity, sr).astype(np.float32) aux_feats = ( torch.from_numpy( vocoder_in_scaler.transform(np.concatenate([mgc, bap], axis=-1)) ) .float() .to(device) ) contf0 = np.exp(lf0) if vocoder_config.data.sine_f0_type in ["contf0", "cf0"]: f0_inp = contf0 elif vocoder_config.data.sine_f0_type == "f0": f0_inp = contf0 f0_inp[vuv < vuv_threshold] = 0 wav = vocoder.inference(f0_inp, aux_feats).view(-1).to("cpu").numpy() else: # Fallback to WORLD f0, spectrogram, aperiodicity = gen_world_params( mgc, lf0, vuv, bap, sr, use_world_codec=use_world_codec, ) wav = pyworld.synthesize(f0, spectrogram, aperiodicity, sr, 5) return wav def synthesize_from_mel( device, logmel, lf0, vuv, sr, vuv_threshold=0.3, vocoder=None, vocoder_in_scaler=None, vocoder_config=None, ): if vocoder is not None: is_usfgan = "generator" in vocoder_config and "discriminator" in vocoder_config assert vocoder_in_scaler is not None if not is_usfgan: # NOTE: So far vocoder models are trained on binary V/UV features vuv = (vuv > vuv_threshold).astype(np.float32) voc_inp = ( torch.from_numpy( vocoder_in_scaler.transform( np.concatenate([logmel, lf0, vuv], axis=-1) ) ) .float() .to(device) ) wav = vocoder.inference(voc_inp).view(-1).to("cpu").numpy() else: # NOTE: So far vocoder models are trained on binary V/UV features vuv = (vuv > vuv_threshold).astype(np.float32) aux_feats = ( torch.from_numpy(vocoder_in_scaler.transform(logmel)).float().to(device) ) contf0 = np.exp(lf0) if vocoder_config.data.sine_f0_type in ["contf0", "cf0"]: f0_inp = contf0 elif vocoder_config.data.sine_f0_type == "f0": f0_inp = contf0 f0_inp[vuv < vuv_threshold] = 0 wav = vocoder.inference(f0_inp, aux_feats).view(-1).to("cpu").numpy() else: raise RuntimeError("Not supported") return wav @torch.no_grad() def eval_model( phase, step, netG, in_feats, out_feats, lengths, model_config, out_scaler, writer, sr, lf0_score_denorm=None, trajectory_smoothing=True, trajectory_smoothing_cutoff=50, trajectory_smoothing_cutoff_f0=20, use_world_codec=False, vocoder=None, vocoder_in_scaler=None, vocoder_config=None, vuv_threshold=0.3, max_num_eval_utts=10, ): if len(model_config.stream_sizes) >= 4: return eval_spss_model( phase=phase, step=step, netG=netG, in_feats=in_feats, out_feats=out_feats, lengths=lengths, model_config=model_config, out_scaler=out_scaler, writer=writer, sr=sr, lf0_score_denorm=lf0_score_denorm, trajectory_smoothing=trajectory_smoothing, trajectory_smoothing_cutoff=trajectory_smoothing_cutoff, trajectory_smoothing_cutoff_f0=trajectory_smoothing_cutoff_f0, use_world_codec=use_world_codec, vocoder=vocoder, vocoder_in_scaler=vocoder_in_scaler, vocoder_config=vocoder_config, vuv_threshold=vuv_threshold, max_num_eval_utts=max_num_eval_utts, ) else: return eval_mel_model( phase=phase, step=step, netG=netG, in_feats=in_feats, out_feats=out_feats, lengths=lengths, model_config=model_config, out_scaler=out_scaler, writer=writer, sr=sr, lf0_score_denorm=lf0_score_denorm, vocoder=vocoder, vocoder_in_scaler=vocoder_in_scaler, vocoder_config=vocoder_config, vuv_threshold=vuv_threshold, max_num_eval_utts=max_num_eval_utts, ) @torch.no_grad() def eval_spss_model( phase, step, netG, in_feats, out_feats, lengths, model_config, out_scaler, writer, sr, lf0_score_denorm=None, trajectory_smoothing=True, trajectory_smoothing_cutoff=50, trajectory_smoothing_cutoff_f0=20, use_world_codec=False, vocoder=None, vocoder_in_scaler=None, vocoder_config=None, vuv_threshold=0.3, max_num_eval_utts=10, ): if dist.is_initialized() and dist.get_rank() != 0: return if writer is None: return # make sure to be in eval mode netG.eval() prediction_type = ( netG.module.prediction_type() if isinstance(netG, nn.DataParallel) or isinstance(netG, DDP) else netG.prediction_type() ) utt_indices = list(range(max_num_eval_utts)) utt_indices = utt_indices[: min(len(utt_indices), len(in_feats))] if np.any(model_config.has_dynamic_features): static_stream_sizes = get_static_stream_sizes( model_config.stream_sizes, model_config.has_dynamic_features, model_config.num_windows, ) else: static_stream_sizes = model_config.stream_sizes rawsp_output = False for utt_idx in utt_indices: out_feats_denorm_ = out_scaler.inverse_transform( out_feats[utt_idx, : lengths[utt_idx]].unsqueeze(0) ) mgc, lf0, vuv, bap = get_static_features( out_feats_denorm_, model_config.num_windows, model_config.stream_sizes, model_config.has_dynamic_features, )[:4] mgc = mgc.squeeze(0).cpu().numpy() lf0 = lf0.squeeze(0).cpu().numpy() vuv = vuv.squeeze(0).cpu().numpy() bap = bap.squeeze(0).cpu().numpy() if lf0_score_denorm is not None: lf0_score_denorm_ = ( lf0_score_denorm[utt_idx, : lengths[utt_idx]].cpu().numpy().reshape(-1) ) else: lf0_score_denorm_ = None # log spectrogram case rawsp_output = mgc.shape[1] >= 128 if rawsp_output: sp = np.exp(mgc) # NOTE: 60-dim mgc is asummed mgc = pyworld.code_spectral_envelope(sp, sr, 60) assert use_world_codec else: sp = None wav = synthesize( device=in_feats.device, mgc=mgc, lf0=lf0, vuv=vuv, bap=bap, sr=sr, use_world_codec=use_world_codec, vuv_threshold=vuv_threshold, vocoder=vocoder, vocoder_in_scaler=vocoder_in_scaler, vocoder_config=vocoder_config, ) group = f"{phase}_utt{np.abs(utt_idx)}_reference" wav = wav / np.abs(wav).max() if np.abs(wav).max() > 1.0 else wav writer.add_audio(group, wav, step, sr) # Run forward outs = netG( in_feats[utt_idx, : lengths[utt_idx]].unsqueeze(0), [lengths[utt_idx]], out_feats[utt_idx, : lengths[utt_idx]].unsqueeze(0), ) # ResF0 case if netG.has_residual_lf0_prediction(): outs, _ = outs # Hybrid if prediction_type == PredictionType.MULTISTREAM_HYBRID: pred_mgc, pred_lf0, pred_vuv, pred_bap = outs if isinstance(pred_lf0, tuple) and len(pred_lf0) == 3: pred_lf0 = mdn_get_most_probable_sigma_and_mu(*pred_lf0)[1] elif isinstance(pred_lf0, tuple) and len(pred_lf0) == 2: pred_lf0 = pred_lf0[1] if isinstance(pred_mgc, tuple) and len(pred_mgc) == 3: pred_mgc = mdn_get_most_probable_sigma_and_mu(*pred_mgc)[1] elif isinstance(pred_mgc, tuple) and len(pred_mgc) == 2: pred_mgc = pred_mgc[1] if isinstance(pred_bap, tuple) and len(pred_bap) == 3: pred_bap = mdn_get_most_probable_sigma_and_mu(*pred_bap)[1] elif isinstance(pred_bap, tuple) and len(pred_bap) == 2: pred_bap = pred_bap[1] pred_out_feats = torch.cat([pred_mgc, pred_lf0, pred_vuv, pred_bap], dim=-1) elif prediction_type == PredictionType.PROBABILISTIC: pi, sigma, mu = outs pred_out_feats = mdn_get_most_probable_sigma_and_mu(pi, sigma, mu)[1] else: pred_out_feats = outs # NOTE: multiple outputs if isinstance(pred_out_feats, list): pred_out_feats = pred_out_feats[-1] if isinstance(pred_out_feats, tuple): pred_out_feats = pred_out_feats[0] if not isinstance(pred_out_feats, list): pred_out_feats = [pred_out_feats] # Run inference if prediction_type in [ PredictionType.PROBABILISTIC, PredictionType.MULTISTREAM_HYBRID, ]: if isinstance(netG, nn.DataParallel) or isinstance(netG, DDP): inference_out_feats, _ = netG.module.inference( in_feats[utt_idx, : lengths[utt_idx]].unsqueeze(0), [lengths[utt_idx]], ) else: inference_out_feats, _ = netG.inference( in_feats[utt_idx, : lengths[utt_idx]].unsqueeze(0), [lengths[utt_idx]], ) else: if isinstance(netG, nn.DataParallel) or isinstance(netG, DDP): inference_out_feats = netG.module.inference( in_feats[utt_idx, : lengths[utt_idx]].unsqueeze(0), [lengths[utt_idx]], ) else: inference_out_feats = netG.inference( in_feats[utt_idx, : lengths[utt_idx]].unsqueeze(0), [lengths[utt_idx]], ) pred_out_feats.append(inference_out_feats) assert len(pred_out_feats) == 2 for idx, pred_out_feats_ in enumerate(pred_out_feats): pred_out_feats_ = pred_out_feats_.squeeze(0).cpu().numpy() pred_out_feats_denorm = ( out_scaler.inverse_transform( torch.from_numpy(pred_out_feats_).to(in_feats.device) ) .cpu() .numpy() ) if np.any(model_config.has_dynamic_features): # (T, D_out) -> (T, static_dim) pred_out_feats_denorm = multi_stream_mlpg( pred_out_feats_denorm, (out_scaler.scale_ ** 2).cpu().numpy(), get_windows(model_config.num_windows), model_config.stream_sizes, model_config.has_dynamic_features, ) pred_mgc, pred_lf0, pred_vuv, pred_bap = split_streams( pred_out_feats_denorm, static_stream_sizes )[:4] # log spectrogram case if rawsp_output: pred_sp = np.exp(pred_mgc) # NOTE: 60-dim mgc is asummed pred_mgc = pyworld.code_spectral_envelope(pred_sp, sr, 60) else: pred_sp = None # Remove high-frequency components of lf0/mgc/bap # NOTE: Useful to reduce high-frequency artifacts if trajectory_smoothing: modfs = int(1 / 0.005) pred_lf0[:, 0] = lowpass_filter( pred_lf0[:, 0], modfs, cutoff=trajectory_smoothing_cutoff_f0 ) for d in range(pred_mgc.shape[1]): pred_mgc[:, d] = lowpass_filter( pred_mgc[:, d], modfs, cutoff=trajectory_smoothing_cutoff ) for d in range(pred_bap.shape[1]): pred_bap[:, d] = lowpass_filter( pred_bap[:, d], modfs, cutoff=trajectory_smoothing_cutoff ) # Generated sample wav = synthesize( device=in_feats.device, mgc=pred_mgc, lf0=pred_lf0, vuv=pred_vuv, bap=pred_bap, sr=sr, use_world_codec=use_world_codec, vuv_threshold=vuv_threshold, vocoder=vocoder, vocoder_in_scaler=vocoder_in_scaler, vocoder_config=vocoder_config, ) wav = wav / np.abs(wav).max() if np.abs(wav).max() > 1.0 else wav if idx == 1: group = f"{phase}_utt{np.abs(utt_idx)}_inference" else: group = f"{phase}_utt{np.abs(utt_idx)}_forward" writer.add_audio(group, wav, step, sr) try: plot_spsvs_params( step, writer, mgc, lf0, vuv, bap, pred_mgc, pred_lf0, pred_vuv, pred_bap, lf0_score=lf0_score_denorm_, group=group, sr=sr, use_world_codec=use_world_codec, sp=sp, pred_sp=pred_sp, ) except IndexError as e: # In _quantile_ureduce_func: # IndexError: index -1 is out of bounds for axis 0 with size 0 print(str(e)) @torch.no_grad() def eval_mel_model( phase, step, netG, in_feats, out_feats, lengths, model_config, out_scaler, writer, sr, lf0_score_denorm=None, vocoder=None, vocoder_in_scaler=None, vocoder_config=None, vuv_threshold=0.3, max_num_eval_utts=10, ): if dist.is_initialized() and dist.get_rank() != 0: return if writer is None: return # make sure to be in eval mode netG.eval() prediction_type = ( netG.module.prediction_type() if isinstance(netG, nn.DataParallel) or isinstance(netG, DDP) else netG.prediction_type() ) utt_indices = list(range(max_num_eval_utts)) utt_indices = utt_indices[: min(len(utt_indices), len(in_feats))] assert not np.any(model_config.has_dynamic_features) static_stream_sizes = model_config.stream_sizes for utt_idx in utt_indices: out_feats_denorm_ = out_scaler.inverse_transform( out_feats[utt_idx, : lengths[utt_idx]].unsqueeze(0) ) logmel, lf0, vuv = get_static_features( out_feats_denorm_, model_config.num_windows, model_config.stream_sizes, model_config.has_dynamic_features, ) logmel = logmel.squeeze(0).cpu().numpy() lf0 = lf0.squeeze(0).cpu().numpy() vuv = vuv.squeeze(0).cpu().numpy() if lf0_score_denorm is not None: lf0_score_denorm_ = ( lf0_score_denorm[utt_idx, : lengths[utt_idx]].cpu().numpy().reshape(-1) ) else: lf0_score_denorm_ = None group = f"{phase}_utt{np.abs(utt_idx)}_reference" if vocoder is not None: wav = synthesize_from_mel( device=in_feats.device, logmel=logmel, lf0=lf0, vuv=vuv, sr=sr, vuv_threshold=vuv_threshold, vocoder=vocoder, vocoder_in_scaler=vocoder_in_scaler, vocoder_config=vocoder_config, ) wav = wav / np.abs(wav).max() if np.abs(wav).max() > 1.0 else wav writer.add_audio(group, wav, step, sr) # Run forward outs = netG( in_feats[utt_idx, : lengths[utt_idx]].unsqueeze(0), [lengths[utt_idx]], out_feats[utt_idx, : lengths[utt_idx]].unsqueeze(0), ) # ResF0 case if netG.has_residual_lf0_prediction(): outs, _ = outs # Hybrid if prediction_type == PredictionType.MULTISTREAM_HYBRID: pred_logmel, pred_lf0, pred_vuv = outs if isinstance(pred_lf0, tuple) and len(pred_lf0) == 3: pred_lf0 = mdn_get_most_probable_sigma_and_mu(*pred_lf0)[1] elif isinstance(pred_lf0, tuple) and len(pred_lf0) == 2: pred_lf0 = pred_lf0[1] if isinstance(pred_logmel, tuple) and len(pred_logmel) == 3: pred_logmel = mdn_get_most_probable_sigma_and_mu(*pred_logmel)[1] elif isinstance(pred_logmel, tuple) and len(pred_logmel) == 2: pred_logmel = pred_logmel[1] pred_out_feats = torch.cat([pred_logmel, pred_lf0, pred_vuv], dim=-1) elif prediction_type == PredictionType.PROBABILISTIC: pi, sigma, mu = outs pred_out_feats = mdn_get_most_probable_sigma_and_mu(pi, sigma, mu)[1] else: pred_out_feats = outs # NOTE: multiple outputs if isinstance(pred_out_feats, list): pred_out_feats = pred_out_feats[-1] if isinstance(pred_out_feats, tuple): pred_out_feats = pred_out_feats[0] if not isinstance(pred_out_feats, list): pred_out_feats = [pred_out_feats] # Run inference if prediction_type in [ PredictionType.PROBABILISTIC, PredictionType.MULTISTREAM_HYBRID, ]: if isinstance(netG, nn.DataParallel) or isinstance(netG, DDP): inference_out_feats, _ = netG.module.inference( in_feats[utt_idx, : lengths[utt_idx]].unsqueeze(0), [lengths[utt_idx]], ) else: inference_out_feats, _ = netG.inference( in_feats[utt_idx, : lengths[utt_idx]].unsqueeze(0), [lengths[utt_idx]], ) else: if isinstance(netG, nn.DataParallel) or isinstance(netG, DDP): inference_out_feats = netG.module.inference( in_feats[utt_idx, : lengths[utt_idx]].unsqueeze(0), [lengths[utt_idx]], ) else: inference_out_feats = netG.inference( in_feats[utt_idx, : lengths[utt_idx]].unsqueeze(0), [lengths[utt_idx]], ) pred_out_feats.append(inference_out_feats) assert len(pred_out_feats) == 2 for idx, pred_out_feats_ in enumerate(pred_out_feats): pred_out_feats_ = pred_out_feats_.squeeze(0).cpu().numpy() pred_out_feats_denorm = ( out_scaler.inverse_transform( torch.from_numpy(pred_out_feats_).to(in_feats.device) ) .cpu() .numpy() ) pred_logmel, pred_lf0, pred_vuv = split_streams( pred_out_feats_denorm, static_stream_sizes ) if idx == 1: group = f"{phase}_utt{np.abs(utt_idx)}_inference" else: group = f"{phase}_utt{np.abs(utt_idx)}_forward" # Generated sample if vocoder is not None: wav = synthesize_from_mel( device=in_feats.device, logmel=pred_logmel, lf0=pred_lf0, vuv=pred_vuv, sr=sr, vuv_threshold=vuv_threshold, vocoder=vocoder, vocoder_in_scaler=vocoder_in_scaler, vocoder_config=vocoder_config, ) wav = wav / np.abs(wav).max() if np.abs(wav).max() > 1.0 else wav writer.add_audio(group, wav, step, sr) try: plot_mel_params( step, writer, logmel, lf0, vuv, pred_logmel, pred_lf0, pred_vuv, lf0_score=lf0_score_denorm_, group=group, sr=sr, ) except IndexError as e: # In _quantile_ureduce_func: # IndexError: index -1 is out of bounds for axis 0 with size 0 print(str(e)) def _colorbar_wrap(fig, mesh, ax, format="%+2.f dB"): try: fig.colorbar(mesh, ax=ax, format=format) except IndexError as e: # In _quantile_ureduce_func: # IndexError: index -1 is out of bounds for axis 0 with size 0 print(str(e)) @torch.no_grad() def plot_spsvs_params( step, writer, mgc, lf0, vuv, bap, pred_mgc, pred_lf0, pred_vuv, pred_bap, lf0_score, group, sr, use_world_codec=False, sp=None, pred_sp=None, ): """Plot acoustic parameters of parametric SVS Args: step (int): step of the current iteration writer (tensorboard.SummaryWriter): tensorboard writer mgc (np.ndarray): mgc lf0 (np.ndarray): lf0 vuv (np.ndarray): vuv bap (np.ndarray): bap pred_mgc (np.ndarray): predicted mgc pred_lf0 (np.ndarray): predicted lf0 pred_vuv (np.ndarray): predicted vuv pred_bap (np.ndarray): predicted bap f0_score (np.ndarray): lf0 score group (str): group name sr (int): sampling rate use_world_codec (bool): use world codec for spectral envelope or not """ if dist.is_initialized() and dist.get_rank() != 0: return assert writer is not None fftlen = pyworld.get_cheaptrick_fft_size(sr) alpha = pysptk.util.mcepalpha(sr) hop_length = int(sr * 0.005) use_mcep_aperiodicity = bap.shape[-1] > 5 # Log-F0 if lf0_score is not None: fig, ax = plt.subplots(1, 1, figsize=(8, 3)) timeaxis = np.arange(len(lf0)) * 0.005 ax.plot(timeaxis, lf0, linewidth=1.5, color="tab:blue", label="Target log-F0") ax.plot( timeaxis, pred_lf0, linewidth=1.5, color="tab:orange", label="Predicted log-F0", ) ax.plot( timeaxis, lf0_score, "--", color="gray", linewidth=1.3, label="Note log-F0", ) ax.set_xlabel("Time [sec]") ax.set_ylabel("Log-frequency [Hz]") ax.set_xlim(timeaxis[0], timeaxis[-1]) ax.set_ylim( min(min(lf0_score[lf0_score > 0]), min(lf0), min(pred_lf0)) - 0.1, max(max(lf0_score), max(lf0), max(pred_lf0)) + 0.1, ) plt.legend(loc="upper right", borderaxespad=0, ncol=3) plt.tight_layout() writer.add_figure(f"{group}/ContinuousLogF0", fig, step) plt.close() f0_score = lf0_score.copy() note_indices = f0_score > 0 f0_score[note_indices] = np.exp(lf0_score[note_indices]) # F0 fig, ax = plt.subplots(1, 1, figsize=(8, 3)) timeaxis = np.arange(len(lf0)) * 0.005 f0 = np.exp(lf0) f0[vuv < 0.5] = 0 pred_f0 = np.exp(pred_lf0) pred_f0[pred_vuv < 0.5] = 0 ax.plot( timeaxis, f0, linewidth=1.5, color="tab:blue", label="Target F0", ) ax.plot( timeaxis, pred_f0, linewidth=1.5, color="tab:orange", label="Predicted F0", ) ax.plot(timeaxis, f0_score, "--", linewidth=1.3, color="gray", label="Note F0") ax.set_xlabel("Time [sec]") ax.set_ylabel("Frequency [Hz]") ax.set_xlim(timeaxis[0], timeaxis[-1]) ax.set_ylim( min(min(f0_score[f0_score > 0]), min(np.exp(lf0)), min(np.exp(pred_lf0))) - 10, max(max(f0_score), max(np.exp(lf0)), max(np.exp(pred_lf0))) + 10, ) plt.legend(loc="upper right", borderaxespad=0, ncol=3) plt.tight_layout() writer.add_figure(f"{group}/F0", fig, step) plt.close() # V/UV fig, ax = plt.subplots(1, 1, figsize=(8, 3)) timeaxis = np.arange(len(lf0)) * 0.005 ax.plot(timeaxis, vuv, linewidth=2, label="Target V/UV") ax.plot(timeaxis, pred_vuv, "--", linewidth=2, label="Predicted V/UV") ax.set_xlabel("Time [sec]") ax.set_ylabel("V/UV") ax.set_xlim(timeaxis[0], timeaxis[-1]) plt.legend(loc="upper right", borderaxespad=0, ncol=2) plt.tight_layout() writer.add_figure(f"{group}/VUV", fig, step) plt.close() # Spectrogram fig, ax = plt.subplots(2, 1, figsize=(8, 6)) ax[0].set_title("Reference spectrogram") ax[1].set_title("Predicted spectrogram") if use_world_codec: if sp is not None: spectrogram = sp.T else: spectrogram = pyworld.decode_spectral_envelope( np.ascontiguousarray(mgc), sr, fftlen ).T else: spectrogram = pysptk.mc2sp(mgc, fftlen=fftlen, alpha=alpha).T mesh = librosa.display.specshow( librosa.power_to_db(np.abs(spectrogram), ref=np.max), sr=sr, hop_length=hop_length, x_axis="time", y_axis="hz", cmap="viridis", ax=ax[0], ) _colorbar_wrap(fig, mesh, ax[0]) if use_world_codec: if pred_sp is not None: pred_spectrogram = pred_sp.T else: pred_spectrogram = pyworld.decode_spectral_envelope( np.ascontiguousarray(pred_mgc), sr, fftlen ).T else: pred_spectrogram = pysptk.mc2sp( np.ascontiguousarray(pred_mgc), fftlen=fftlen, alpha=alpha ).T mesh = librosa.display.specshow( librosa.power_to_db(np.abs(pred_spectrogram), ref=np.max), sr=sr, hop_length=hop_length, x_axis="time", y_axis="hz", cmap="viridis", ax=ax[1], ) _colorbar_wrap(fig, mesh, ax[1]) for a in ax: a.set_ylim(0, sr // 2) plt.tight_layout() writer.add_figure(f"{group}/Spectrogram", fig, step) plt.close() # Aperiodicity fig, ax = plt.subplots(2, 1, figsize=(8, 6)) ax[0].set_title("Reference aperiodicity") ax[1].set_title("Predicted aperiodicity") if use_mcep_aperiodicity: aperiodicity = pysptk.mc2sp(bap, fftlen=fftlen, alpha=alpha).T else: aperiodicity = pyworld.decode_aperiodicity(bap.astype(np.float64), sr, fftlen).T mesh = librosa.display.specshow( 20 * np.log10(aperiodicity), sr=sr, hop_length=hop_length, x_axis="time", y_axis="linear", cmap="viridis", ax=ax[0], ) _colorbar_wrap(fig, mesh, ax[0]) if use_mcep_aperiodicity: pred_aperiodicity = pysptk.mc2sp( np.ascontiguousarray(pred_bap), fftlen=fftlen, alpha=alpha ).T else: pred_aperiodicity = pyworld.decode_aperiodicity( np.ascontiguousarray(pred_bap).astype(np.float64), sr, fftlen ).T mesh = librosa.display.specshow( 20 * np.log10(pred_aperiodicity), sr=sr, hop_length=hop_length, x_axis="time", y_axis="linear", cmap="viridis", ax=ax[1], ) _colorbar_wrap(fig, mesh, ax[1]) for a in ax: a.set_ylim(0, sr // 2) plt.tight_layout() writer.add_figure(f"{group}/Aperiodicity", fig, step) plt.close() # GV for mgc fig, ax = plt.subplots(1, 1, figsize=(8, 3)) ax.plot(np.var(mgc, axis=0), "--", linewidth=2, label="Natural: global variances") ax.plot(np.var(pred_mgc, axis=0), linewidth=2, label="Generated: global variances") ax.legend() ax.set_yscale("log") ax.set_xlabel("Dimension of mgc") min_ = min(np.var(mgc, axis=0).min(), np.var(pred_mgc, axis=0).min(), 1e-4) ax.set_ylim(min_) plt.tight_layout() writer.add_figure(f"{group}/GV_mgc", fig, step) plt.close() # GV for bap fig, ax = plt.subplots(1, 1, figsize=(8, 3)) ax.plot(np.var(bap, axis=0), "--", linewidth=2, label="Natural: global variances") ax.plot(np.var(pred_bap, axis=0), linewidth=2, label="Generated: global variances") ax.legend() ax.set_yscale("log") ax.set_xlabel("Dimension of bap") min_ = min(np.var(bap, axis=0).min(), np.var(pred_bap, axis=0).min(), 10) ax.set_ylim(min_) plt.tight_layout() writer.add_figure(f"{group}/GV_bap", fig, step) plt.close() @torch.no_grad() def plot_mel_params( step, writer, logmel, lf0, vuv, pred_logmel, pred_lf0, pred_vuv, lf0_score, group, sr, ): if dist.is_initialized() and dist.get_rank() != 0: return assert writer is not None hop_length = int(sr * 0.005) # Log-F0 if lf0_score is not None: fig, ax = plt.subplots(1, 1, figsize=(8, 3)) timeaxis = np.arange(len(lf0)) * 0.005 ax.plot(timeaxis, lf0, linewidth=1.5, color="tab:blue", label="Target log-F0") ax.plot( timeaxis, pred_lf0, linewidth=1.5, color="tab:orange", label="Predicted log-F0", ) ax.plot( timeaxis, lf0_score, "--", color="gray", linewidth=1.3, label="Note log-F0", ) ax.set_xlabel("Time [sec]") ax.set_ylabel("Log-frequency [Hz]") ax.set_xlim(timeaxis[0], timeaxis[-1]) ax.set_ylim( min(min(lf0_score[lf0_score > 0]), min(lf0), min(pred_lf0)) - 0.1, max(max(lf0_score), max(lf0), max(pred_lf0)) + 0.1, ) plt.legend(loc="upper right", borderaxespad=0, ncol=3) plt.tight_layout() writer.add_figure(f"{group}/ContinuousLogF0", fig, step) plt.close() f0_score = lf0_score.copy() note_indices = f0_score > 0 f0_score[note_indices] = np.exp(lf0_score[note_indices]) # F0 fig, ax = plt.subplots(1, 1, figsize=(8, 3)) timeaxis = np.arange(len(lf0)) * 0.005 f0 = np.exp(lf0) f0[vuv < 0.5] = 0 pred_f0 = np.exp(pred_lf0) pred_f0[pred_vuv < 0.5] = 0 ax.plot( timeaxis, f0, linewidth=1.5, color="tab:blue", label="Target F0", ) ax.plot( timeaxis, pred_f0, linewidth=1.5, color="tab:orange", label="Predicted F0", ) ax.plot(timeaxis, f0_score, "--", linewidth=1.3, color="gray", label="Note F0") ax.set_xlabel("Time [sec]") ax.set_ylabel("Frequency [Hz]") ax.set_xlim(timeaxis[0], timeaxis[-1]) ax.set_ylim( min(min(f0_score[f0_score > 0]), min(np.exp(lf0)), min(np.exp(pred_lf0))) - 10, max(max(f0_score), max(np.exp(lf0)), max(np.exp(pred_lf0))) + 10, ) plt.legend(loc="upper right", borderaxespad=0, ncol=3) plt.tight_layout() writer.add_figure(f"{group}/F0", fig, step) plt.close() # V/UV fig, ax = plt.subplots(1, 1, figsize=(8, 3)) timeaxis = np.arange(len(lf0)) * 0.005 ax.plot(timeaxis, vuv, linewidth=2, label="Target V/UV") ax.plot(timeaxis, pred_vuv, "--", linewidth=2, label="Predicted V/UV") ax.set_xlabel("Time [sec]") ax.set_ylabel("V/UV") ax.set_xlim(timeaxis[0], timeaxis[-1]) plt.legend(loc="upper right", borderaxespad=0, ncol=2) plt.tight_layout() writer.add_figure(f"{group}/VUV", fig, step) plt.close() # Mel-spectrogram fig, ax = plt.subplots(2, 1, figsize=(8, 6)) ax[0].set_title("Reference spectrogram") ax[1].set_title("Predicted spectrogram") mesh = librosa.display.specshow( logmel.T, sr=sr, hop_length=hop_length, x_axis="time", y_axis="hz", cmap="viridis", ax=ax[0], ) _colorbar_wrap(fig, mesh, ax[0]) mesh = librosa.display.specshow( pred_logmel.T, sr=sr, hop_length=hop_length, x_axis="time", y_axis="hz", cmap="viridis", ax=ax[1], ) _colorbar_wrap(fig, mesh, ax[1]) for a in ax: a.set_ylim(0, sr // 2) plt.tight_layout() writer.add_figure(f"{group}/Spectrogram", fig, step) plt.close()
80,222
32.538043
95
py
nnsvs
nnsvs-master/nnsvs/acoustic_models/multistream.py
import torch from nnsvs.acoustic_models.util import pad_inference from nnsvs.base import BaseModel, PredictionType from nnsvs.multistream import split_streams from torch import nn __all__ = [ "MultistreamSeparateF0ParametricModel", "NPSSMultistreamParametricModel", "NPSSMDNMultistreamParametricModel", "MultistreamSeparateF0MelModel", "MDNMultistreamSeparateF0MelModel", ] class MultistreamSeparateF0ParametricModel(BaseModel): """Multi-stream model with a separate F0 prediction model acoustic features: [MGC, LF0, VUV, BAP] vib_model and vib_flags_model are optional and will be likely to be removed. Conditional dependency: p(MGC, LF0, VUV, BAP |C) = p(LF0|C) p(MGC|LF0, C) p(BAP|LF0, C) p(VUV|LF0, C) Args: in_dim (int): Input dimension. out_dim (int): Output dimension. stream_sizes (list): List of stream sizes. reduction_factor (int): Reduction factor. encoder (nn.Module): A shared encoder. mgc_model (nn.Module): MGC prediction model. lf0_model (nn.Module): log-F0 prediction model. vuv_model (nn.Module): V/UV prediction model. bap_model (nn.Module): BAP prediction model. in_rest_idx (int): Index of the rest symbol in the input features. in_lf0_idx (int): index of lf0 in input features in_lf0_min (float): minimum value of lf0 in the training data of input features in_lf0_max (float): maximum value of lf0 in the training data of input features out_lf0_idx (int): index of lf0 in output features. Typically 180. out_lf0_mean (float): mean of lf0 in the training data of output features out_lf0_scale (float): scale of lf0 in the training data of output features lf0_teacher_forcing (bool): Whether to use teacher forcing for F0 prediction. """ def __init__( self, in_dim: int, out_dim: int, stream_sizes: list, reduction_factor: int, encoder: nn.Module, mgc_model: nn.Module, lf0_model: nn.Module, vuv_model: nn.Module, bap_model: nn.Module, vib_model: nn.Module = None, # kept as is for compatibility vib_flags_model: nn.Module = None, # kept as is for compatibility # NOTE: you must carefully set the following parameters in_rest_idx=1, in_lf0_idx=300, in_lf0_min=5.3936276, in_lf0_max=6.491111, out_lf0_idx=180, out_lf0_mean=5.953093881972361, out_lf0_scale=0.23435173188961034, lf0_teacher_forcing=True, ): super().__init__() self.in_dim = in_dim self.out_dim = out_dim self.stream_sizes = stream_sizes self.reduction_factor = reduction_factor self.lf0_teacher_forcing = lf0_teacher_forcing assert len(stream_sizes) in [4] self.encoder = encoder if self.encoder is not None: assert not encoder.is_autoregressive() self.mgc_model = mgc_model self.lf0_model = lf0_model self.vuv_model = vuv_model self.bap_model = bap_model self.in_rest_idx = in_rest_idx self.in_lf0_idx = in_lf0_idx self.in_lf0_min = in_lf0_min self.in_lf0_max = in_lf0_max self.out_lf0_idx = out_lf0_idx self.out_lf0_mean = out_lf0_mean self.out_lf0_scale = out_lf0_scale def has_residual_lf0_prediction(self): return True def _set_lf0_params(self): # Special care for residual F0 prediction models # NOTE: don't overwrite out_lf0_idx and in_lf0_idx if hasattr(self.lf0_model, "out_lf0_mean"): self.lf0_model.in_lf0_min = self.in_lf0_min self.lf0_model.in_lf0_max = self.in_lf0_max self.lf0_model.out_lf0_mean = self.out_lf0_mean self.lf0_model.out_lf0_scale = self.out_lf0_scale def is_autoregressive(self): return ( self.mgc_model.is_autoregressive() or self.lf0_model.is_autoregressive() or self.vuv_model.is_autoregressive() or self.bap_model.is_autoregressive() ) def forward(self, x, lengths=None, y=None): self._set_lf0_params() assert x.shape[-1] == self.in_dim if y is not None: # Teacher-forcing y_mgc, y_lf0, y_vuv, y_bap = split_streams(y, self.stream_sizes) else: # Inference y_mgc, y_lf0, y_vuv, y_bap = None, None, None, None # Predict continuous log-F0 first lf0, lf0_residual = self.lf0_model(x, lengths, y_lf0) if self.encoder is not None: encoder_outs = self.encoder(x, lengths) # Concat log-F0, rest flags and the outputs of the encoder # This may make the decoder to be aware of the input F0 rest_flags = x[:, :, self.in_rest_idx].unsqueeze(-1) if self.lf0_teacher_forcing and y is not None: encoder_outs = torch.cat([encoder_outs, rest_flags, y_lf0], dim=-1) else: encoder_outs = torch.cat([encoder_outs, rest_flags, lf0], dim=-1) else: encoder_outs = x # Decoders for each stream mgc = self.mgc_model(encoder_outs, lengths, y_mgc) vuv = self.vuv_model(encoder_outs, lengths, y_vuv) bap = self.bap_model(encoder_outs, lengths, y_bap) # make a concatenated stream has_postnet_output = ( isinstance(mgc, list) or isinstance(lf0, list) or isinstance(vuv, list) or isinstance(bap, list) ) if has_postnet_output: outs = [] for idx in range(len(mgc)): mgc_ = mgc[idx] if isinstance(mgc, list) else mgc lf0_ = lf0[idx] if isinstance(lf0, list) else lf0 vuv_ = vuv[idx] if isinstance(vuv, list) else vuv bap_ = bap[idx] if isinstance(bap, list) else bap out = torch.cat([mgc_, lf0_, vuv_, bap_], dim=-1) assert out.shape[-1] == self.out_dim outs.append(out) return outs, lf0_residual else: out = torch.cat([mgc, lf0, vuv, bap], dim=-1) assert out.shape[-1] == self.out_dim return out, lf0_residual def inference(self, x, lengths=None): return pad_inference( model=self, x=x, lengths=lengths, reduction_factor=self.reduction_factor ) class NPSSMultistreamParametricModel(BaseModel): """NPSS-like cascaded multi-stream model with no mixture density networks. NPSS: :cite:t:`blaauw2017neural` Different from the original NPSS, we don't use spectral parameters for the inputs of aperiodicity and V/UV prediction models. This is because (1) D4C does not use spectral parameters as input for aperiodicity estimation. (2) V/UV detection is done from aperiodicity at 0-3 kHz in WORLD. In addition, f0 and VUV models dont use MDNs. Empirically, we found the above configuration works better than the original one. Conditional dependency: p(MGC, LF0, VUV, BAP |C) = p(LF0|C) p(MGC|LF0, C) p(BAP|LF0, C) p(VUV|LF0, BAP, C) Args: in_dim (int): Input dimension. out_dim (int): Output dimension. stream_sizes (list): List of stream sizes. lf0_model (BaseModel): Model for predicting log-F0. mgc_model (BaseModel): Model for predicting MGC. bap_model (BaseModel): Model for predicting BAP. vuv_model (BaseModel): Model for predicting V/UV. in_rest_idx (int): Index of the rest symbol in the input features. in_lf0_idx (int): index of lf0 in input features in_lf0_min (float): minimum value of lf0 in the training data of input features in_lf0_max (float): maximum value of lf0 in the training data of input features out_lf0_idx (int): index of lf0 in output features. Typically 180. out_lf0_mean (float): mean of lf0 in the training data of output features out_lf0_scale (float): scale of lf0 in the training data of output features vuv_model_bap_conditioning (bool): If True, use BAP features for V/UV prediction. vuv_model_bap0_conditioning (bool): If True, use only 0-th coef. of BAP for V/UV prediction. vuv_model_lf0_conditioning (bool): If True, use log-F0 features for V/UV prediction. vuv_model_mgc_conditioning (bool): If True, use MGC features for V/UV prediction. """ def __init__( self, in_dim: int, out_dim: int, stream_sizes: list, reduction_factor: int, lf0_model: nn.Module, mgc_model: nn.Module, bap_model: nn.Module, vuv_model: nn.Module, # NOTE: you must carefully set the following parameters in_rest_idx=0, in_lf0_idx=51, in_lf0_min=5.3936276, in_lf0_max=6.491111, out_lf0_idx=60, out_lf0_mean=5.953093881972361, out_lf0_scale=0.23435173188961034, npss_style_conditioning=False, vuv_model_bap_conditioning=True, vuv_model_bap0_conditioning=False, vuv_model_lf0_conditioning=True, vuv_model_mgc_conditioning=False, ): super().__init__() self.in_dim = in_dim self.out_dim = out_dim self.stream_sizes = stream_sizes self.reduction_factor = reduction_factor self.vuv_model_bap_conditioning = vuv_model_bap_conditioning self.vuv_model_bap0_conditioning = vuv_model_bap0_conditioning self.vuv_model_lf0_conditioning = vuv_model_lf0_conditioning self.vuv_model_mgc_conditioning = vuv_model_mgc_conditioning assert not npss_style_conditioning, "Not supported" assert len(stream_sizes) in [4] self.lf0_model = lf0_model self.mgc_model = mgc_model self.bap_model = bap_model self.vuv_model = vuv_model self.in_rest_idx = in_rest_idx self.in_lf0_idx = in_lf0_idx self.in_lf0_min = in_lf0_min self.in_lf0_max = in_lf0_max self.out_lf0_idx = out_lf0_idx self.out_lf0_mean = out_lf0_mean self.out_lf0_scale = out_lf0_scale def _set_lf0_params(self): # Special care for residual F0 prediction models # NOTE: don't overwrite out_lf0_idx and in_lf0_idx if hasattr(self.lf0_model, "out_lf0_mean"): self.lf0_model.in_lf0_min = self.in_lf0_min self.lf0_model.in_lf0_max = self.in_lf0_max self.lf0_model.out_lf0_mean = self.out_lf0_mean self.lf0_model.out_lf0_scale = self.out_lf0_scale def prediction_type(self): return PredictionType.DETERMINISTIC def is_autoregressive(self): return ( self.mgc_model.is_autoregressive() or self.lf0_model.is_autoregressive() or self.vuv_model.is_autoregressive() or self.bap_model.is_autoregressive() ) def has_residual_lf0_prediction(self): return True def forward(self, x, lengths=None, y=None): self._set_lf0_params() assert x.shape[-1] == self.in_dim is_inference = y is None if is_inference: y_mgc, y_lf0, y_vuv, y_bap = ( None, None, None, None, ) else: # Teacher-forcing outs = split_streams(y, self.stream_sizes) y_mgc, y_lf0, y_vuv, y_bap = outs # Predict continuous log-F0 first if is_inference: lf0, lf0_residual = self.lf0_model.inference(x, lengths), None else: lf0, lf0_residual = self.lf0_model(x, lengths, y_lf0) # Predict spectral parameters if is_inference: mgc_inp = torch.cat([x, lf0], dim=-1) mgc = self.mgc_model.inference(mgc_inp, lengths) else: mgc_inp = torch.cat([x, y_lf0], dim=-1) mgc = self.mgc_model(mgc_inp, lengths, y_mgc) # Predict aperiodic parameters if is_inference: bap_inp = torch.cat([x, lf0], dim=-1) bap = self.bap_model.inference(bap_inp, lengths) else: bap_inp = torch.cat([x, y_lf0], dim=-1) bap = self.bap_model(bap_inp, lengths, y_bap) # Predict V/UV if is_inference: if self.vuv_model_bap0_conditioning: bap_cond = bap[:, :, 0:1] else: bap_cond = bap # full cond: (x, mgc, lf0, bap) vuv_inp = [x] if self.vuv_model_mgc_conditioning: vuv_inp.append(mgc) if self.vuv_model_bap_conditioning: vuv_inp.append(bap_cond) if self.vuv_model_lf0_conditioning: vuv_inp.append(lf0) vuv_inp = torch.cat(vuv_inp, dim=-1) vuv = self.vuv_model.inference(vuv_inp, lengths) else: if self.vuv_model_bap0_conditioning: y_bap_cond = y_bap[:, :, 0:1] else: y_bap_cond = y_bap vuv_inp = [x] if self.vuv_model_mgc_conditioning: vuv_inp.append(y_mgc) if self.vuv_model_bap_conditioning: vuv_inp.append(y_bap_cond) if self.vuv_model_lf0_conditioning: vuv_inp.append(y_lf0) vuv_inp = torch.cat(vuv_inp, dim=-1) vuv = self.vuv_model(vuv_inp, lengths, y_vuv) # make a concatenated stream has_postnet_output = ( isinstance(mgc, list) or isinstance(bap, list) or isinstance(vuv, list) ) if has_postnet_output: outs = [] for idx in range(len(mgc)): mgc_ = mgc[idx] if isinstance(mgc, list) else mgc lf0_ = lf0[idx] if isinstance(lf0, list) else lf0 vuv_ = vuv[idx] if isinstance(vuv, list) else vuv bap_ = bap[idx] if isinstance(bap, list) else bap out = torch.cat([mgc_, lf0_, vuv_, bap_], dim=-1) assert out.shape[-1] == self.out_dim outs.append(out) else: outs = torch.cat([mgc, lf0, vuv, bap], dim=-1) assert outs.shape[-1] == self.out_dim return outs, lf0_residual def inference(self, x, lengths=None): return pad_inference( model=self, x=x, lengths=lengths, reduction_factor=self.reduction_factor, mdn=False, ) class NPSSMDNMultistreamParametricModel(BaseModel): """NPSS-like cascaded multi-stream parametric model with mixture density networks. .. note:: This class was originally designed to be used with MDNs. However, the internal design was changed to make it work with non-MDN and diffusion models. For example, you can use non-MDN models for MGC prediction. NPSS: :cite:t:`blaauw2017neural` acoustic features: [MGC, LF0, VUV, BAP] Conditional dependency: p(MGC, LF0, VUV, BAP |C) = p(LF0|C) p(MGC|LF0, C) p(BAP|LF0, C) p(VUV|LF0, BAP, C) Args: in_dim (int): Input dimension. out_dim (int): Output dimension. stream_sizes (list): List of stream sizes. lf0_model (BaseModel): Model for predicting log-F0. mgc_model (BaseModel): Model for predicting MGC. bap_model (BaseModel): Model for predicting BAP. vuv_model (BaseModel): Model for predicting V/UV. in_rest_idx (int): Index of the rest symbol in the input features. in_lf0_idx (int): index of lf0 in input features in_lf0_min (float): minimum value of lf0 in the training data of input features in_lf0_max (float): maximum value of lf0 in the training data of input features out_lf0_idx (int): index of lf0 in output features. Typically 180. out_lf0_mean (float): mean of lf0 in the training data of output features out_lf0_scale (float): scale of lf0 in the training data of output features vuv_model_bap_conditioning (bool): If True, use BAP features for V/UV prediction. vuv_model_bap0_conditioning (bool): If True, use only 0-th coef. of BAP for V/UV prediction. vuv_model_lf0_conditioning (bool): If True, use log-F0 features for V/UV prediction. vuv_model_mgc_conditioning (bool): If True, use MGC features for V/UV prediction. """ def __init__( self, in_dim: int, out_dim: int, stream_sizes: list, reduction_factor: int, lf0_model: nn.Module, mgc_model: nn.Module, bap_model: nn.Module, vuv_model: nn.Module, # NOTE: you must carefully set the following parameters in_rest_idx=0, in_lf0_idx=51, in_lf0_min=5.3936276, in_lf0_max=6.491111, out_lf0_idx=60, out_lf0_mean=5.953093881972361, out_lf0_scale=0.23435173188961034, vuv_model_bap_conditioning=True, vuv_model_bap0_conditioning=False, vuv_model_lf0_conditioning=True, vuv_model_mgc_conditioning=False, ): super().__init__() self.in_dim = in_dim self.out_dim = out_dim self.stream_sizes = stream_sizes self.reduction_factor = reduction_factor self.vuv_model_bap_conditioning = vuv_model_bap_conditioning self.vuv_model_bap0_conditioning = vuv_model_bap0_conditioning self.vuv_model_lf0_conditioning = vuv_model_lf0_conditioning self.vuv_model_mgc_conditioning = vuv_model_mgc_conditioning assert len(stream_sizes) in [4] self.lf0_model = lf0_model self.mgc_model = mgc_model self.bap_model = bap_model self.vuv_model = vuv_model self.in_rest_idx = in_rest_idx self.in_lf0_idx = in_lf0_idx self.in_lf0_min = in_lf0_min self.in_lf0_max = in_lf0_max self.out_lf0_idx = out_lf0_idx self.out_lf0_mean = out_lf0_mean self.out_lf0_scale = out_lf0_scale def _set_lf0_params(self): # Special care for residual F0 prediction models # NOTE: don't overwrite out_lf0_idx and in_lf0_idx if hasattr(self.lf0_model, "out_lf0_mean"): self.lf0_model.in_lf0_min = self.in_lf0_min self.lf0_model.in_lf0_max = self.in_lf0_max self.lf0_model.out_lf0_mean = self.out_lf0_mean self.lf0_model.out_lf0_scale = self.out_lf0_scale def prediction_type(self): return PredictionType.MULTISTREAM_HYBRID def is_autoregressive(self): return ( self.mgc_model.is_autoregressive() or self.lf0_model.is_autoregressive() or self.vuv_model.is_autoregressive() or self.bap_model.is_autoregressive() ) def has_residual_lf0_prediction(self): return True def forward(self, x, lengths=None, y=None): self._set_lf0_params() assert x.shape[-1] == self.in_dim is_inference = y is None if is_inference: y_mgc, y_lf0, y_vuv, y_bap = ( None, None, None, None, ) else: # Teacher-forcing outs = split_streams(y, self.stream_sizes) y_mgc, y_lf0, y_vuv, y_bap = outs # Predict continuous log-F0 first if is_inference: lf0, lf0_residual = self.lf0_model.inference(x, lengths), None if self.lf0_model.prediction_type() == PredictionType.PROBABILISTIC: lf0_cond = lf0[0] else: lf0_cond = lf0 else: lf0, lf0_residual = self.lf0_model(x, lengths, y_lf0) # Predict spectral parameters if is_inference: mgc_inp = torch.cat([x, lf0_cond], dim=-1) mgc = self.mgc_model.inference(mgc_inp, lengths) else: mgc_inp = torch.cat([x, y_lf0], dim=-1) mgc = self.mgc_model(mgc_inp, lengths, y_mgc) # Predict aperiodic parameters if is_inference: bap_inp = torch.cat([x, lf0_cond], dim=-1) bap = self.bap_model.inference(bap_inp, lengths) else: bap_inp = torch.cat([x, y_lf0], dim=-1) bap = self.bap_model(bap_inp, lengths, y_bap) # Predict V/UV if is_inference: if self.bap_model.prediction_type() == PredictionType.PROBABILISTIC: bap_cond = bap[0] else: bap_cond = bap if self.mgc_model.prediction_type() == PredictionType.PROBABILISTIC: mgc_cond = mgc[0] else: mgc_cond = mgc if self.vuv_model_bap0_conditioning: bap_cond = bap_cond[:, :, 0:1] # full cond: (x, mgc, lf0, bap) vuv_inp = [x] if self.vuv_model_mgc_conditioning: vuv_inp.append(mgc_cond) if self.vuv_model_lf0_conditioning: vuv_inp.append(lf0_cond) if self.vuv_model_bap_conditioning: vuv_inp.append(bap_cond) vuv_inp = torch.cat(vuv_inp, dim=-1) vuv = self.vuv_model.inference(vuv_inp, lengths) else: if self.vuv_model_bap0_conditioning: y_bap_cond = y_bap[:, :, 0:1] else: y_bap_cond = y_bap vuv_inp = [x] if self.vuv_model_mgc_conditioning: vuv_inp.append(y_mgc) if self.vuv_model_lf0_conditioning: vuv_inp.append(y_lf0) if self.vuv_model_bap_conditioning: vuv_inp.append(y_bap_cond) vuv_inp = torch.cat(vuv_inp, dim=-1) vuv = self.vuv_model(vuv_inp, lengths, y_vuv) if is_inference: if self.lf0_model.prediction_type() == PredictionType.PROBABILISTIC: lf0_ = lf0[0] else: lf0_ = lf0 if self.bap_model.prediction_type() == PredictionType.PROBABILISTIC: bap_ = bap[0] else: bap_ = bap if self.mgc_model.prediction_type() == PredictionType.PROBABILISTIC: mgc_ = mgc[0] else: mgc_ = mgc out = torch.cat([mgc_, lf0_, vuv, bap_], dim=-1) assert out.shape[-1] == self.out_dim # TODO: better design return out, out else: return (mgc, lf0, vuv, bap), lf0_residual def inference(self, x, lengths=None): return pad_inference( model=self, x=x, lengths=lengths, reduction_factor=self.reduction_factor, mdn=True, ) class MultistreamSeparateF0MelModel(BaseModel): """Multi-stream model with a separate F0 prediction model (mel-version) Conditional dependency: p(MEL, LF0, VUV|C) = p(LF0|C) p(MEL|LF0, C) p(VUV|LF0, C) Args: in_dim (int): Input dimension. out_dim (int): Output dimension. stream_sizes (list): List of stream sizes. reduction_factor (int): Reduction factor. encoder (nn.Module): A shared encoder. mel_model (nn.Module): MEL prediction model. lf0_model (nn.Module): log-F0 prediction model. vuv_model (nn.Module): V/UV prediction model. in_rest_idx (int): Index of the rest symbol in the input features. in_lf0_idx (int): index of lf0 in input features in_lf0_min (float): minimum value of lf0 in the training data of input features in_lf0_max (float): maximum value of lf0 in the training data of input features out_lf0_idx (int): index of lf0 in output features. Typically 180. out_lf0_mean (float): mean of lf0 in the training data of output features out_lf0_scale (float): scale of lf0 in the training data of output features """ def __init__( self, in_dim: int, out_dim: int, stream_sizes: list, reduction_factor: int, encoder: nn.Module, mel_model: nn.Module, lf0_model: nn.Module, vuv_model: nn.Module, # NOTE: you must carefully set the following parameters in_rest_idx=1, in_lf0_idx=300, in_lf0_min=5.3936276, in_lf0_max=6.491111, out_lf0_idx=180, out_lf0_mean=5.953093881972361, out_lf0_scale=0.23435173188961034, ): super().__init__() self.in_dim = in_dim self.out_dim = out_dim self.stream_sizes = stream_sizes self.reduction_factor = reduction_factor assert len(stream_sizes) == 3 self.encoder = encoder if self.encoder is not None: assert not encoder.is_autoregressive() self.mel_model = mel_model self.lf0_model = lf0_model self.vuv_model = vuv_model self.in_rest_idx = in_rest_idx self.in_lf0_idx = in_lf0_idx self.in_lf0_min = in_lf0_min self.in_lf0_max = in_lf0_max self.out_lf0_idx = out_lf0_idx self.out_lf0_mean = out_lf0_mean self.out_lf0_scale = out_lf0_scale def _set_lf0_params(self): # Special care for residual F0 prediction models # NOTE: don't overwrite out_lf0_idx and in_lf0_idx if hasattr(self.lf0_model, "out_lf0_mean"): self.lf0_model.in_lf0_min = self.in_lf0_min self.lf0_model.in_lf0_max = self.in_lf0_max self.lf0_model.out_lf0_mean = self.out_lf0_mean self.lf0_model.out_lf0_scale = self.out_lf0_scale def is_autoregressive(self): return ( self.mel_model.is_autoregressive() or self.lf0_model.is_autoregressive() or self.vuv_model.is_autoregressive() ) def has_residual_lf0_prediction(self): return True def forward(self, x, lengths=None, y=None): self._set_lf0_params() assert x.shape[-1] == self.in_dim if y is not None: # Teacher-forcing outs = split_streams(y, self.stream_sizes) y_mel, y_lf0, y_vuv = outs else: # Inference y_mel, y_lf0, y_vuv = ( None, None, None, ) # Predict continuous log-F0 first lf0, lf0_residual = self.lf0_model(x, lengths, y_lf0) if self.encoder is not None: encoder_outs = self.encoder(x, lengths) # Concat log-F0, rest flags and the outputs of the encoder # This may make the decoder to be aware of the input F0 rest_flags = x[:, :, self.in_rest_idx].unsqueeze(-1) if y is not None: encoder_outs = torch.cat([encoder_outs, rest_flags, y_lf0], dim=-1) else: encoder_outs = torch.cat([encoder_outs, rest_flags, lf0], dim=-1) else: encoder_outs = x # Decoders for each stream mel = self.mel_model(encoder_outs, lengths, y_mel) vuv = self.vuv_model(encoder_outs, lengths, y_vuv) # make a concatenated stream has_postnet_output = ( isinstance(mel, list) or isinstance(lf0, list) or isinstance(vuv, list) ) if has_postnet_output: outs = [] for idx in range(len(mel)): mel_ = mel[idx] if isinstance(mel, list) else mel lf0_ = lf0[idx] if isinstance(lf0, list) else lf0 vuv_ = vuv[idx] if isinstance(vuv, list) else vuv out = torch.cat([mel_, lf0_, vuv_], dim=-1) assert out.shape[-1] == self.out_dim outs.append(out) return outs, lf0_residual else: out = torch.cat( [ mel, lf0, vuv, ], dim=-1, ) assert out.shape[-1] == self.out_dim return out, lf0_residual def inference(self, x, lengths=None): return pad_inference( model=self, x=x, lengths=lengths, reduction_factor=self.reduction_factor ) class MDNMultistreamSeparateF0MelModel(BaseModel): """Multi-stream model with a separate F0 model (mel-version) and mDN V/UV prediction is performed given a mel-spectrogram. Conditional dependency: p(MEL, LF0, VUV|C) = p(LF0|C) p(MEL|LF0, C) p(VUV|LF0, MEL, C) .. note:: This class was originally designed to be used with MDNs. However, the internal design was changed to make it work with non-MDN and diffusion models. For example, you can use non-MDN models for mel prediction. Args: in_dim (int): Input dimension. out_dim (int): Output dimension. stream_sizes (list): List of stream sizes. reduction_factor (int): Reduction factor. encoder (nn.Module): A shared encoder. mel_model (nn.Module): MEL prediction model. lf0_model (nn.Module): log-F0 prediction model. vuv_model (nn.Module): V/UV prediction model. in_rest_idx (int): Index of the rest symbol in the input features. in_lf0_idx (int): index of lf0 in input features in_lf0_min (float): minimum value of lf0 in the training data of input features in_lf0_max (float): maximum value of lf0 in the training data of input features out_lf0_idx (int): index of lf0 in output features. Typically 180. out_lf0_mean (float): mean of lf0 in the training data of output features out_lf0_scale (float): scale of lf0 in the training data of output features vuv_model_lf0_conditioning (bool): If True, use log-F0 features for V/UV prediction. vuv_model_mel_conditioning (bool): If True, use mel features for V/UV prediction. """ def __init__( self, in_dim: int, out_dim: int, stream_sizes: list, reduction_factor: int, mel_model: nn.Module, lf0_model: nn.Module, vuv_model: nn.Module, # NOTE: you must carefully set the following parameters in_rest_idx=0, in_lf0_idx=51, in_lf0_min=5.3936276, in_lf0_max=6.491111, out_lf0_idx=60, out_lf0_mean=5.953093881972361, out_lf0_scale=0.23435173188961034, vuv_model_lf0_conditioning=True, vuv_model_mel_conditioning=True, ): super().__init__() self.in_dim = in_dim self.out_dim = out_dim self.stream_sizes = stream_sizes self.reduction_factor = reduction_factor self.vuv_model_lf0_conditioning = vuv_model_lf0_conditioning self.vuv_model_mel_conditioning = vuv_model_mel_conditioning assert len(stream_sizes) in [3] self.mel_model = mel_model self.lf0_model = lf0_model self.vuv_model = vuv_model self.in_rest_idx = in_rest_idx self.in_lf0_idx = in_lf0_idx self.in_lf0_min = in_lf0_min self.in_lf0_max = in_lf0_max self.out_lf0_idx = out_lf0_idx self.out_lf0_mean = out_lf0_mean self.out_lf0_scale = out_lf0_scale def _set_lf0_params(self): # Special care for residual F0 prediction models # NOTE: don't overwrite out_lf0_idx and in_lf0_idx if hasattr(self.lf0_model, "out_lf0_mean"): self.lf0_model.in_lf0_min = self.in_lf0_min self.lf0_model.in_lf0_max = self.in_lf0_max self.lf0_model.out_lf0_mean = self.out_lf0_mean self.lf0_model.out_lf0_scale = self.out_lf0_scale def prediction_type(self): return PredictionType.MULTISTREAM_HYBRID def is_autoregressive(self): return ( self.mel_model.is_autoregressive() or self.lf0_model.is_autoregressive() or self.vuv_model.is_autoregressive() ) def has_residual_lf0_prediction(self): return True def forward(self, x, lengths=None, y=None): self._set_lf0_params() assert x.shape[-1] == self.in_dim is_inference = y is None if y is not None: # Teacher-forcing outs = split_streams(y, self.stream_sizes) y_mel, y_lf0, y_vuv = outs else: # Inference y_mel, y_lf0, y_vuv = ( None, None, None, ) # Predict continuous log-F0 first if is_inference: lf0, lf0_residual = self.lf0_model.inference(x, lengths), None if self.lf0_model.prediction_type() == PredictionType.PROBABILISTIC: lf0_cond = lf0[0] else: lf0_cond = lf0 else: lf0, lf0_residual = self.lf0_model(x, lengths, y_lf0) # Predict mel if is_inference: mel_inp = torch.cat([x, lf0_cond], dim=-1) mel = self.mel_model.inference(mel_inp, lengths) else: mel_inp = torch.cat([x, y_lf0], dim=-1) mel = self.mel_model(mel_inp, lengths, y_mel) # Predict V/UV if is_inference: if self.mel_model.prediction_type() == PredictionType.PROBABILISTIC: mel_cond = mel[0] else: mel_cond = mel # full cond: (x, lf0, mel) vuv_inp = [x] if self.vuv_model_lf0_conditioning: vuv_inp.append(lf0_cond) if self.vuv_model_mel_conditioning: vuv_inp.append(mel_cond) vuv_inp = torch.cat(vuv_inp, dim=-1) vuv = self.vuv_model.inference(vuv_inp, lengths) else: vuv_inp = [x] if self.vuv_model_lf0_conditioning: vuv_inp.append(y_lf0) if self.vuv_model_mel_conditioning: vuv_inp.append(y_mel) vuv_inp = torch.cat(vuv_inp, dim=-1) vuv = self.vuv_model(vuv_inp, lengths, y_vuv) if is_inference: if self.lf0_model.prediction_type() == PredictionType.PROBABILISTIC: lf0_ = lf0[0] else: lf0_ = lf0 if self.mel_model.prediction_type() == PredictionType.PROBABILISTIC: mel_ = mel[0] else: mel_ = mel out = torch.cat([mel_, lf0_, vuv], dim=-1) assert out.shape[-1] == self.out_dim # TODO: better design return out, out else: return (mel, lf0, vuv), lf0_residual def inference(self, x, lengths=None): return pad_inference( model=self, x=x, lengths=lengths, reduction_factor=self.reduction_factor, mdn=True, )
34,956
36.267591
92
py
nnsvs
nnsvs-master/nnsvs/acoustic_models/tacotron.py
import torch from nnsvs.acoustic_models.util import pad_inference from nnsvs.base import BaseModel from nnsvs.tacotron.decoder import MDNNonAttentiveDecoder from nnsvs.tacotron.decoder import NonAttentiveDecoder as TacotronNonAttentiveDecoder from nnsvs.tacotron.postnet import Postnet as TacotronPostnet from nnsvs.util import init_weights from torch import nn from torch.nn.utils.rnn import pack_padded_sequence, pad_packed_sequence __all__ = [ "NonAttentiveDecoder", "MDNNonAttentiveDecoder", "BiLSTMNonAttentiveDecoder", "BiLSTMMDNNonAttentiveDecoder", ] class NonAttentiveDecoder(TacotronNonAttentiveDecoder): """Non-attentive autoregresive model based on the duration-informed Tacotron Duration-informed Tacotron :cite:t:`okamoto2019tacotron`. .. note:: if the target features of the decoder is normalized to N(0, 1), consider setting the initial value carefully so that it roughly matches the value of silence. e.g., -4 to -10. ``initial_value=0`` works okay for large databases but I found that -4 or lower worked better for smaller databases such as nit-song070. Args: in_dim (int): Input dimension. out_dim (int): Output dimension. layers (int): Number of LSTM layers. hidden_dim (int): Hidden dimension of LSTM. prenet_layers (int): Number of prenet layers. prenet_hidden_dim (int): Hidden dimension of prenet. prenet_dropout (float): Dropout rate of prenet. zoneout (float): Zoneout rate. reduction_factor (int): Reduction factor. downsample_by_conv (bool): If True, downsampling is performed by convolution. postnet_layers (int): Number of postnet layers. postnet_channels (int): Number of postnet channels. postnet_kernel_size (int): Kernel size of postnet. postnet_dropout (float): Dropout rate of postnet. init_type (str): Initialization type. eval_dropout (bool): If True, dropout is applied in evaluation. initial_value (float) : initial value for the autoregressive decoder. """ def __init__( self, in_dim=512, out_dim=80, layers=2, hidden_dim=1024, prenet_layers=2, prenet_hidden_dim=256, prenet_dropout=0.5, zoneout=0.1, reduction_factor=1, downsample_by_conv=False, postnet_layers=0, postnet_channels=512, postnet_kernel_size=5, postnet_dropout=0.0, init_type="none", eval_dropout=True, prenet_noise_std=0.0, initial_value=0.0, ): super().__init__( in_dim=in_dim, out_dim=out_dim, layers=layers, hidden_dim=hidden_dim, prenet_layers=prenet_layers, prenet_hidden_dim=prenet_hidden_dim, prenet_dropout=prenet_dropout, zoneout=zoneout, reduction_factor=reduction_factor, downsample_by_conv=downsample_by_conv, eval_dropout=eval_dropout, prenet_noise_std=prenet_noise_std, initial_value=initial_value, ) if postnet_layers > 0: self.postnet = TacotronPostnet( out_dim, layers=postnet_layers, channels=postnet_channels, kernel_size=postnet_kernel_size, dropout=postnet_dropout, ) else: self.postnet = None init_weights(self, init_type) def forward(self, x, lengths=None, y=None): outs = super().forward(x, lengths, y) if self.postnet is not None: # NOTE: `outs.clone()`` is necessary to compute grad on both outs and outs_fine outs_fine = outs + self.postnet(outs.transpose(1, 2).clone()).transpose( 1, 2 ) return [outs, outs_fine] else: return outs def inference(self, x, lengths=None): return pad_inference( model=self, x=x, lengths=lengths, reduction_factor=self.reduction_factor ) class BiLSTMNonAttentiveDecoder(BaseModel): """BiLSTM-based encoder + NonAttentiveDecoder The encoder is based on the arthitecture of the Sinsy acoustic model. Args: in_dim (int): Input dimension. ff_hidden_dim (int): Hidden dimension of feed-forward layers in the encoder. conv_hidden_dim (int): Hidden dimension of convolution layers in the encoder. lstm_hidden_dim (int): Hidden dimension of LSTM layers in the encoder. num_lstm_layers (int): Number of LSTM layers in the encoder. out_dim (int): Output dimension. layers (int): Number of LSTM layers. hidden_dim (int): Hidden dimension of LSTM. prenet_layers (int): Number of prenet layers. prenet_hidden_dim (int): Hidden dimension of prenet. prenet_dropout (float): Dropout rate of prenet. zoneout (float): Zoneout rate. reduction_factor (int): Reduction factor. downsample_by_conv (bool): If True, downsampling is performed by convolution. postnet_layers (int): Number of postnet layers. postnet_channels (int): Number of postnet channels. postnet_kernel_size (int): Kernel size of postnet. postnet_dropout (float): Dropout rate of postnet. in_ph_start_idx (int): Start index of phoneme features. in_ph_end_idx (int): End index of phoneme features. embed_dim (int): Embedding dimension. init_type (str): Initialization type. eval_dropout (bool): If True, dropout is applied in evaluation. initial_value (float) : initial value for the autoregressive decoder. """ def __init__( self, in_dim=512, ff_hidden_dim=2048, conv_hidden_dim=1024, lstm_hidden_dim=256, num_lstm_layers=2, out_dim=80, decoder_layers=2, decoder_hidden_dim=1024, prenet_layers=2, prenet_hidden_dim=256, prenet_dropout=0.5, zoneout=0.1, reduction_factor=1, downsample_by_conv=False, postnet_layers=0, postnet_channels=512, postnet_kernel_size=5, postnet_dropout=0.0, in_ph_start_idx: int = 1, in_ph_end_idx: int = 50, embed_dim=None, init_type="none", eval_dropout=True, prenet_noise_std=0.0, initial_value=0.0, ): super().__init__() self.in_dim = in_dim self.out_dim = out_dim self.in_ph_start_idx = in_ph_start_idx self.in_ph_end_idx = in_ph_end_idx self.num_vocab = in_ph_end_idx - in_ph_start_idx self.embed_dim = embed_dim self.reduction_factor = reduction_factor if self.embed_dim is not None: assert in_dim > self.num_vocab self.emb = nn.Embedding(self.num_vocab, embed_dim) self.fc_in = nn.Linear(in_dim - self.num_vocab, embed_dim) ff_in_dim = embed_dim else: ff_in_dim = in_dim # Encoder # NOTE: can be simply replaced by a BiLSTM? # so far I use sinsy like architecture self.ff = nn.Sequential( nn.Linear(ff_in_dim, ff_hidden_dim), nn.ReLU(), nn.Linear(ff_hidden_dim, ff_hidden_dim), nn.ReLU(), nn.Linear(ff_hidden_dim, ff_hidden_dim), nn.ReLU(), ) self.conv = nn.Sequential( nn.ReflectionPad1d(3), nn.Conv1d(ff_hidden_dim, conv_hidden_dim, kernel_size=7, padding=0), nn.BatchNorm1d(conv_hidden_dim), nn.ReLU(), nn.ReflectionPad1d(3), nn.Conv1d(conv_hidden_dim, conv_hidden_dim, kernel_size=7, padding=0), nn.BatchNorm1d(conv_hidden_dim), nn.ReLU(), nn.ReflectionPad1d(3), nn.Conv1d(conv_hidden_dim, conv_hidden_dim, kernel_size=7, padding=0), nn.BatchNorm1d(conv_hidden_dim), nn.ReLU(), ) self.lstm = nn.LSTM( conv_hidden_dim, lstm_hidden_dim, num_lstm_layers, bidirectional=True, batch_first=True, dropout=0.0, ) # Autoregressive decoder decoder_in_dim = 2 * lstm_hidden_dim self.decoder = TacotronNonAttentiveDecoder( in_dim=decoder_in_dim, out_dim=out_dim, layers=decoder_layers, hidden_dim=decoder_hidden_dim, prenet_layers=prenet_layers, prenet_hidden_dim=prenet_hidden_dim, prenet_dropout=prenet_dropout, zoneout=zoneout, reduction_factor=reduction_factor, downsample_by_conv=downsample_by_conv, eval_dropout=eval_dropout, prenet_noise_std=prenet_noise_std, initial_value=initial_value, ) if postnet_layers > 0: self.postnet = TacotronPostnet( out_dim, layers=postnet_layers, channels=postnet_channels, kernel_size=postnet_kernel_size, dropout=postnet_dropout, ) else: self.postnet = None init_weights(self, init_type) def is_autoregressive(self): return self.decoder.is_autoregressive() def forward(self, x, lengths=None, y=None): if isinstance(lengths, torch.Tensor): lengths = lengths.to("cpu") if self.embed_dim is not None: x_first, x_ph_onehot, x_last = torch.split( x, [ self.in_ph_start_idx, self.num_vocab, self.in_dim - self.num_vocab - self.in_ph_start_idx, ], dim=-1, ) x_ph = torch.argmax(x_ph_onehot, dim=-1) # Make sure to have one-hot vector assert (x_ph_onehot.sum(-1) <= 1).all() x = self.emb(x_ph) + self.fc_in(torch.cat([x_first, x_last], dim=-1)) out = self.ff(x) out = self.conv(out.transpose(1, 2)).transpose(1, 2) sequence = pack_padded_sequence(out, lengths, batch_first=True) out, _ = self.lstm(sequence) out, _ = pad_packed_sequence(out, batch_first=True) outs = self.decoder(out, lengths, y) if self.postnet is not None: # NOTE: `outs.clone()`` is necessary to compute grad on both outs and outs_fine outs_fine = outs + self.postnet(outs.transpose(1, 2).clone()).transpose( 1, 2 ) return [outs, outs_fine] else: return outs def inference(self, x, lengths=None): return pad_inference( model=self, x=x, lengths=lengths, reduction_factor=self.reduction_factor ) class BiLSTMMDNNonAttentiveDecoder(BaseModel): """BiLSTM-based encoder + NonAttentiveDecoder (MDN version) The encoder is based on the arthitecture of the Sinsy acoustic model. Args: in_dim (int): Input dimension. ff_hidden_dim (int): Hidden dimension of feed-forward layers in the encoder. conv_hidden_dim (int): Hidden dimension of convolution layers in the encoder. lstm_hidden_dim (int): Hidden dimension of LSTM layers in the encoder. num_lstm_layers (int): Number of LSTM layers in the encoder. out_dim (int): Output dimension. layers (int): Number of LSTM layers. hidden_dim (int): Hidden dimension of LSTM. prenet_layers (int): Number of prenet layers. prenet_hidden_dim (int): Hidden dimension of prenet. prenet_dropout (float): Dropout rate of prenet. zoneout (float): Zoneout rate. reduction_factor (int): Reduction factor. downsample_by_conv (bool): If True, downsampling is performed by convolution. num_gaussians (int): Number of Gaussians. sampling_mode (str): Sampling mode. postnet_layers (int): Number of postnet layers. postnet_channels (int): Number of postnet channels. postnet_kernel_size (int): Kernel size of postnet. postnet_dropout (float): Dropout rate of postnet. in_ph_start_idx (int): Start index of phoneme features. in_ph_end_idx (int): End index of phoneme features. embed_dim (int): Embedding dimension. init_type (str): Initialization type. eval_dropout (bool): If True, dropout is applied in evaluation. initial_value (float) : initial value for the autoregressive decoder. """ def __init__( self, in_dim=512, ff_hidden_dim=2048, conv_hidden_dim=1024, lstm_hidden_dim=256, num_lstm_layers=2, out_dim=80, decoder_layers=2, decoder_hidden_dim=1024, prenet_layers=2, prenet_hidden_dim=256, prenet_dropout=0.5, zoneout=0.1, reduction_factor=1, downsample_by_conv=False, num_gaussians=8, sampling_mode="mean", in_ph_start_idx: int = 1, in_ph_end_idx: int = 50, embed_dim=None, init_type="none", eval_dropout=True, prenet_noise_std=0, initial_value=0.0, ): super().__init__() self.in_dim = in_dim self.out_dim = out_dim self.in_ph_start_idx = in_ph_start_idx self.in_ph_end_idx = in_ph_end_idx self.num_vocab = in_ph_end_idx - in_ph_start_idx self.embed_dim = embed_dim self.reduction_factor = reduction_factor if self.embed_dim is not None: assert in_dim > self.num_vocab self.emb = nn.Embedding(self.num_vocab, embed_dim) self.fc_in = nn.Linear(in_dim - self.num_vocab, embed_dim) ff_in_dim = embed_dim else: ff_in_dim = in_dim # Encoder # NOTE: can be simply replaced by a BiLSTM? # so far I use sinsy like architecture self.ff = nn.Sequential( nn.Linear(ff_in_dim, ff_hidden_dim), nn.ReLU(), nn.Linear(ff_hidden_dim, ff_hidden_dim), nn.ReLU(), nn.Linear(ff_hidden_dim, ff_hidden_dim), nn.ReLU(), ) self.conv = nn.Sequential( nn.ReflectionPad1d(3), nn.Conv1d(ff_hidden_dim, conv_hidden_dim, kernel_size=7, padding=0), nn.BatchNorm1d(conv_hidden_dim), nn.ReLU(), nn.ReflectionPad1d(3), nn.Conv1d(conv_hidden_dim, conv_hidden_dim, kernel_size=7, padding=0), nn.BatchNorm1d(conv_hidden_dim), nn.ReLU(), nn.ReflectionPad1d(3), nn.Conv1d(conv_hidden_dim, conv_hidden_dim, kernel_size=7, padding=0), nn.BatchNorm1d(conv_hidden_dim), nn.ReLU(), ) self.lstm = nn.LSTM( conv_hidden_dim, lstm_hidden_dim, num_lstm_layers, bidirectional=True, batch_first=True, dropout=0.0, ) # Autoregressive decoder decoder_in_dim = 2 * lstm_hidden_dim self.decoder = MDNNonAttentiveDecoder( in_dim=decoder_in_dim, out_dim=out_dim, layers=decoder_layers, hidden_dim=decoder_hidden_dim, prenet_layers=prenet_layers, prenet_hidden_dim=prenet_hidden_dim, prenet_dropout=prenet_dropout, zoneout=zoneout, reduction_factor=reduction_factor, downsample_by_conv=downsample_by_conv, num_gaussians=num_gaussians, sampling_mode=sampling_mode, eval_dropout=eval_dropout, prenet_noise_std=prenet_noise_std, initial_value=initial_value, ) init_weights(self, init_type) def is_autoregressive(self): return self.decoder.is_autoregressive() def prediction_type(self): return self.decoder.prediction_type() def forward(self, x, lengths=None, y=None): if isinstance(lengths, torch.Tensor): lengths = lengths.to("cpu") if self.embed_dim is not None: x_first, x_ph_onehot, x_last = torch.split( x, [ self.in_ph_start_idx, self.num_vocab, self.in_dim - self.num_vocab - self.in_ph_start_idx, ], dim=-1, ) x_ph = torch.argmax(x_ph_onehot, dim=-1) # Make sure to have one-hot vector assert (x_ph_onehot.sum(-1) <= 1).all() x = self.emb(x_ph) + self.fc_in(torch.cat([x_first, x_last], dim=-1)) out = self.ff(x) out = self.conv(out.transpose(1, 2)).transpose(1, 2) sequence = pack_padded_sequence(out, lengths, batch_first=True) out, _ = self.lstm(sequence) out, _ = pad_packed_sequence(out, batch_first=True) outs = self.decoder(out, lengths, y) return outs def inference(self, x, lengths=None): return pad_inference( model=self, x=x, lengths=lengths, reduction_factor=self.reduction_factor, mdn=True, )
17,378
35.056017
91
py
nnsvs
nnsvs-master/nnsvs/acoustic_models/tacotron_f0.py
import numpy as np import torch from nnsvs.acoustic_models.util import pad_inference from nnsvs.base import BaseModel, PredictionType from nnsvs.mdn import MDNLayer, mdn_get_most_probable_sigma_and_mu, mdn_get_sample from nnsvs.tacotron.decoder import Prenet, ZoneOutCell from nnsvs.util import init_weights from torch import nn from torch.nn import functional as F from torch.nn.utils.rnn import pack_padded_sequence, pad_packed_sequence __all__ = [ "ResF0NonAttentiveDecoder", "MDNResF0NonAttentiveDecoder", "BiLSTMResF0NonAttentiveDecoder", ] class ResF0NonAttentiveDecoder(BaseModel): """Duration-informed Tacotron with residual F0 prediction. Args: in_dim (int) : dimension of encoder hidden layer out_dim (int) : dimension of output layers (int) : number of LSTM layers hidden_dim (int) : dimension of hidden layer prenet_layers (int) : number of pre-net layers prenet_hidden_dim (int) : dimension of pre-net hidden layer prenet_dropout (float) : dropout rate of pre-net zoneout (float) : zoneout rate reduction_factor (int) : reduction factor downsample_by_conv (bool) : if True, downsample by convolution scaled_tanh (bool) : if True, use scaled tanh for residual F0 prediction in_lf0_idx (int): index of lf0 in input features in_lf0_min (float): minimum of lf0 in the training data of input features in_lf0_max (float): maximum of lf0 in the training data of input features out_lf0_idx (int): index of lf0 in output features out_lf0_mean (float): mean of lf0 in the training data of output features out_lf0_scale (float): scale of lf0 in the training data of output features init_type (str): initialization type eval_dropout (bool): if True, use dropout in evaluation """ def __init__( self, in_dim=512, out_dim=1, layers=2, hidden_dim=1024, prenet_layers=2, prenet_hidden_dim=256, prenet_dropout=0.5, zoneout=0.1, reduction_factor=1, downsample_by_conv=False, scaled_tanh=True, # NOTE: you must carefully set the following parameters in_lf0_idx=300, in_lf0_min=5.3936276, in_lf0_max=6.491111, out_lf0_idx=180, out_lf0_mean=5.953093881972361, out_lf0_scale=0.23435173188961034, init_type="none", eval_dropout=True, ): super().__init__() self.out_dim = out_dim self.reduction_factor = reduction_factor self.prenet_dropout = prenet_dropout self.scaled_tanh = scaled_tanh self.in_lf0_idx = in_lf0_idx self.in_lf0_min = in_lf0_min self.in_lf0_max = in_lf0_max self.out_lf0_idx = out_lf0_idx self.out_lf0_mean = out_lf0_mean self.out_lf0_scale = out_lf0_scale if prenet_layers > 0: self.prenet = Prenet( out_dim, prenet_layers, prenet_hidden_dim, prenet_dropout, eval_dropout=eval_dropout, ) lstm_in_dim = in_dim + prenet_hidden_dim else: self.prenet = None prenet_hidden_dim = 0 lstm_in_dim = in_dim + out_dim self.lstm = nn.ModuleList() for layer in range(layers): lstm = nn.LSTMCell( lstm_in_dim if layer == 0 else hidden_dim, hidden_dim, ) self.lstm += [ZoneOutCell(lstm, zoneout)] proj_in_dim = in_dim + hidden_dim self.feat_out = nn.Linear(proj_in_dim, out_dim * reduction_factor, bias=False) if reduction_factor > 1 and downsample_by_conv: self.conv_downsample = nn.Conv1d( in_dim, in_dim, kernel_size=reduction_factor, stride=reduction_factor, groups=in_dim, ) else: self.conv_downsample = None init_weights(self, init_type) def _zero_state(self, hs): init_hs = hs.new_zeros(hs.size(0), self.lstm[0].hidden_size) return init_hs def is_autoregressive(self): return True def has_residual_lf0_prediction(self): return True def forward(self, encoder_outs, in_lens, decoder_targets=None): """Forward step Args: encoder_outs (torch.Tensor): encoder outputs (B, T, C) in_lens (torch.Tensor): input lengths decoder_targets (torch.Tensor): decoder targets for teacher-forcing. (B, T, C) Returns: torch.Tensor: the output (B, C, T) """ is_inference = decoder_targets is None if not is_inference: assert encoder_outs.shape[1] == decoder_targets.shape[1] # Denormalize lf0 from input musical score lf0_score = encoder_outs[:, :, self.in_lf0_idx].unsqueeze(-1) lf0_score_denorm = ( lf0_score * (self.in_lf0_max - self.in_lf0_min) + self.in_lf0_min ) # (B, T, C) -> (B, C, T) lf0_score_denorm = lf0_score_denorm.transpose(1, 2) # To avoid unbounded residual f0 that would potentially cause artifacts, # let's constrain the residual F0 to be in a certain range by the scaled tanh residual_f0_max_cent = 600 max_lf0_ratio = residual_f0_max_cent * np.log(2) / 1200 # Adjust number of frames according to the reduction factor # (B, Lmax, out_dim) -> (B, Lmax/r, out_dim) if self.reduction_factor > 1 and not is_inference: decoder_targets = decoder_targets[ :, self.reduction_factor - 1 :: self.reduction_factor ] if self.reduction_factor > 1: if self.conv_downsample is not None: encoder_outs = self.conv_downsample( encoder_outs.transpose(1, 2) ).transpose(1, 2) else: encoder_outs = encoder_outs[ :, self.reduction_factor - 1 :: self.reduction_factor ] h_list, c_list = [], [] for _ in range(len(self.lstm)): h_list.append(self._zero_state(encoder_outs)) c_list.append(self._zero_state(encoder_outs)) go_frame = encoder_outs.new_zeros(encoder_outs.size(0), self.out_dim) prev_out = go_frame if not is_inference and self.prenet is not None: prenet_outs = self.prenet(decoder_targets) outs = [] lf0_residuals = [] for t in range(encoder_outs.shape[1]): # Pre-Net if self.prenet is not None: if is_inference: prenet_out = self.prenet(prev_out) else: prenet_out = prenet_outs[:, t, :] else: prenet_out = F.dropout(prev_out, self.prenet_dropout, training=True) # LSTM xs = torch.cat([encoder_outs[:, t], prenet_out], dim=1) h_list[0], c_list[0] = self.lstm[0](xs, (h_list[0], c_list[0])) for i in range(1, len(self.lstm)): h_list[i], c_list[i] = self.lstm[i]( h_list[i - 1], (h_list[i], c_list[i]) ) # Output hcs = torch.cat([h_list[-1], encoder_outs[:, t]], dim=1) out = self.feat_out(hcs).view(encoder_outs.size(0), self.out_dim, -1) # Residual F0 if self.scaled_tanh: lf0_residual = max_lf0_ratio * torch.tanh( out[:, self.out_lf0_idx, :] ).unsqueeze(1) else: lf0_residual = out[:, self.out_lf0_idx, :].unsqueeze(1) # Residual connection in the denormalized f0 domain lf0_score_denorm_t = lf0_score_denorm[ :, :, t * self.reduction_factor : (t + 1) * self.reduction_factor ] lf0_pred_denorm = lf0_score_denorm_t + lf0_residual # Back to normalized f0 lf0_pred = (lf0_pred_denorm - self.out_lf0_mean) / self.out_lf0_scale out[:, self.out_lf0_idx, :] = lf0_pred.squeeze(1) outs.append(out) lf0_residuals.append(lf0_residual) # Update decoder input for the next time step if is_inference: prev_out = outs[-1][:, :, -1] # (1, out_dim) else: # Teacher forcing prev_out = decoder_targets[:, t, :] outs = torch.cat(outs, dim=2) # (B, out_dim, Lmax) lf0_residuals = torch.cat(lf0_residuals, dim=2) # (B, out_dim, Lmax) # (B, C, T) -> (B, T, C) outs = outs.transpose(1, 2) lf0_residuals = lf0_residuals.transpose(1, 2) return outs, lf0_residuals def inference(self, x, lengths=None): return pad_inference( model=self, x=x, lengths=lengths, reduction_factor=self.reduction_factor ) class MDNResF0NonAttentiveDecoder(BaseModel): """Duration-informed Tacotron with residual F0 prediction (MDN-version) Args: in_dim (int) : dimension of encoder hidden layer out_dim (int) : dimension of output layers (int) : number of LSTM layers hidden_dim (int) : dimension of hidden layer prenet_layers (int) : number of pre-net layers prenet_hidden_dim (int) : dimension of pre-net hidden layer prenet_dropout (float) : dropout rate of pre-net zoneout (float) : zoneout rate reduction_factor (int) : reduction factor downsample_by_conv (bool) : if True, downsample by convolution scaled_tanh (bool) : if True, use scaled tanh for residual F0 prediction num_gaussians (int) : number of Gaussian sampling_mode (str) : sampling mode in_lf0_idx (int): index of lf0 in input features in_lf0_min (float): minimum of lf0 in the training data of input features in_lf0_max (float): maximum of lf0 in the training data of input features out_lf0_idx (int): index of lf0 in output features out_lf0_mean (float): mean of lf0 in the training data of output features out_lf0_scale (float): scale of lf0 in the training data of output features init_type (str): initialization type eval_dropout (bool): if True, use dropout in evaluation """ def __init__( self, in_dim=512, out_dim=80, layers=2, hidden_dim=1024, prenet_layers=2, prenet_hidden_dim=256, prenet_dropout=0.5, zoneout=0.1, reduction_factor=1, downsample_by_conv=False, scaled_tanh=True, num_gaussians=4, sampling_mode="mean", # NOTE: you must carefully set the following parameters in_lf0_idx=300, in_lf0_min=5.3936276, in_lf0_max=6.491111, out_lf0_idx=180, out_lf0_mean=5.953093881972361, out_lf0_scale=0.23435173188961034, init_type="none", eval_dropout=True, ): super().__init__() self.out_dim = out_dim self.reduction_factor = reduction_factor self.prenet_dropout = prenet_dropout self.scaled_tanh = scaled_tanh self.in_lf0_idx = in_lf0_idx self.in_lf0_min = in_lf0_min self.in_lf0_max = in_lf0_max self.out_lf0_idx = out_lf0_idx self.out_lf0_mean = out_lf0_mean self.out_lf0_scale = out_lf0_scale self.num_gaussians = num_gaussians self.sampling_mode = sampling_mode assert sampling_mode in ["mean", "random"] if prenet_layers > 0: self.prenet = Prenet( out_dim, prenet_layers, prenet_hidden_dim, prenet_dropout, eval_dropout=eval_dropout, ) lstm_in_dim = in_dim + prenet_hidden_dim else: self.prenet = None prenet_hidden_dim = 0 lstm_in_dim = in_dim + out_dim self.lstm = nn.ModuleList() for layer in range(layers): lstm = nn.LSTMCell( lstm_in_dim if layer == 0 else hidden_dim, hidden_dim, ) self.lstm += [ZoneOutCell(lstm, zoneout)] proj_in_dim = in_dim + hidden_dim self.feat_out = MDNLayer( proj_in_dim, out_dim * reduction_factor, num_gaussians=num_gaussians, dim_wise=True, ) if reduction_factor > 1 and downsample_by_conv: self.conv_downsample = nn.Conv1d( in_dim, in_dim, kernel_size=reduction_factor, stride=reduction_factor, groups=in_dim, ) else: self.conv_downsample = None init_weights(self, init_type) def _zero_state(self, hs): init_hs = hs.new_zeros(hs.size(0), self.lstm[0].hidden_size) return init_hs def prediction_type(self): return PredictionType.PROBABILISTIC def is_autoregressive(self): return True def has_residual_lf0_prediction(self): return True def forward(self, encoder_outs, in_lens, decoder_targets=None): """Forward step Args: encoder_outs (torch.Tensor): encoder outputs (B, T, C) in_lens (torch.Tensor): input lengths decoder_targets (torch.Tensor): decoder targets for teacher-forcing. (B, T, C) Returns: torch.Tensor: the output (B, C, T) """ is_inference = decoder_targets is None if not is_inference: assert encoder_outs.shape[1] == decoder_targets.shape[1] # Denormalize lf0 from input musical score lf0_score = encoder_outs[:, :, self.in_lf0_idx].unsqueeze(-1) lf0_score_denorm = ( lf0_score * (self.in_lf0_max - self.in_lf0_min) + self.in_lf0_min ) # (B, T, C) -> (B, C, T) # (B, T, C) # lf0_score_denorm = lf0_score_denorm.transpose(1, 2) # To avoid unbounded residual f0 that would potentially cause artifacts, # let's constrain the residual F0 to be in a certain range by the scaled tanh residual_f0_max_cent = 600 max_lf0_ratio = residual_f0_max_cent * np.log(2) / 1200 # Adjust number of frames according to the reduction factor # (B, Lmax, out_dim) -> (B, Lmax/r, out_dim) if self.reduction_factor > 1 and not is_inference: decoder_targets = decoder_targets[ :, self.reduction_factor - 1 :: self.reduction_factor ] if self.reduction_factor > 1: if self.conv_downsample is not None: encoder_outs = self.conv_downsample( encoder_outs.transpose(1, 2) ).transpose(1, 2) else: encoder_outs = encoder_outs[ :, self.reduction_factor - 1 :: self.reduction_factor ] h_list, c_list = [], [] for _ in range(len(self.lstm)): h_list.append(self._zero_state(encoder_outs)) c_list.append(self._zero_state(encoder_outs)) go_frame = encoder_outs.new_zeros(encoder_outs.size(0), self.out_dim) prev_out = go_frame if not is_inference and self.prenet is not None: prenet_outs = self.prenet(decoder_targets) mus = [] log_pis = [] log_sigmas = [] lf0_residuals = [] mus_inf = [] for t in range(encoder_outs.shape[1]): # Pre-Net if self.prenet is not None: if is_inference: prenet_out = self.prenet(prev_out) else: prenet_out = prenet_outs[:, t, :] else: prenet_out = F.dropout(prev_out, self.prenet_dropout, training=True) # LSTM xs = torch.cat([encoder_outs[:, t], prenet_out], dim=1) h_list[0], c_list[0] = self.lstm[0](xs, (h_list[0], c_list[0])) for i in range(1, len(self.lstm)): h_list[i], c_list[i] = self.lstm[i]( h_list[i - 1], (h_list[i], c_list[i]) ) # Output hcs = torch.cat([h_list[-1], encoder_outs[:, t]], dim=1) log_pi, log_sigma, mu = self.feat_out(hcs.unsqueeze(1)) # (B, reduction_factor, num_gaussians, out_dim) log_pi = ( log_pi.transpose(1, 2) .view(encoder_outs.size(0), self.num_gaussians, -1, self.out_dim) .transpose(1, 2) ) log_sigma = ( log_sigma.transpose(1, 2) .view(encoder_outs.size(0), self.num_gaussians, -1, self.out_dim) .transpose(1, 2) ) mu = ( mu.transpose(1, 2) .view(encoder_outs.size(0), self.num_gaussians, -1, self.out_dim) .transpose(1, 2) ) # Residual F0 # (B, reduction_factor, num_gaussians) if self.scaled_tanh: lf0_residual = max_lf0_ratio * torch.tanh(mu[:, :, :, self.out_lf0_idx]) else: lf0_residual = mu[:, :, :, self.out_lf0_idx] # Residual connection in the denormalized f0 domain lf0_score_denorm_t = lf0_score_denorm[ :, t * self.reduction_factor : (t + 1) * self.reduction_factor, : ] # NOTE: broadcast against num_gaussians axis # (B, 1, 1) + (B, 1, num_gaussians) -> (B, 1, num_gaussians) lf0_pred_denorm = lf0_score_denorm_t + lf0_residual # Back to normalized f0 lf0_pred = (lf0_pred_denorm - self.out_lf0_mean) / self.out_lf0_scale mu[:, :, :, self.out_lf0_idx] = lf0_pred mus.append(mu) log_pis.append(log_pi) log_sigmas.append(log_sigma) lf0_residuals.append(lf0_residual) # Update decoder input for the next time step if is_inference: # (B, reduction_factor, out_dim) if self.sampling_mode == "mean": _, mu = mdn_get_most_probable_sigma_and_mu(log_pi, log_sigma, mu) elif self.sampling_mode == "random": mu = mdn_get_sample(log_pi, log_sigma, mu) # Feed last sample for the feedback loop prev_out = mu[:, -1] mus_inf.append(mu) else: # Teacher forcing prev_out = decoder_targets[:, t, :] # (B, T, G, out_dim) mus = torch.cat(mus, dim=1) log_pis = torch.cat(log_pis, dim=1) log_sigmas = torch.cat(log_sigmas, dim=1) # (B, T, num_gaussians) lf0_residuals = torch.cat(lf0_residuals, dim=1) if is_inference: mu = torch.cat(mus_inf, dim=1) # TODO: may need to track sigma. For now we only use mu return mu, mu else: return (log_pis, log_sigmas, mus), lf0_residuals def inference(self, x, lengths=None): return pad_inference( model=self, x=x, lengths=lengths, reduction_factor=self.reduction_factor, mdn=True, ) class BiLSTMResF0NonAttentiveDecoder(BaseModel): """BiLSTM-based encoder + duration-informed Tacotron with residual F0 prediction. Args: in_dim (int) : dimension of encoder hidden layer ff_hidden_dim (int): Hidden dimension of feed-forward layers in the encoder. conv_hidden_dim (int): Hidden dimension of convolution layers in the encoder. lstm_hidden_dim (int): Hidden dimension of LSTM layers in the encoder. num_lstm_layers (int): Number of LSTM layers in the encoder. out_dim (int) : dimension of output layers (int) : number of LSTM layers hidden_dim (int) : dimension of hidden layer prenet_layers (int) : number of pre-net layers prenet_hidden_dim (int) : dimension of pre-net hidden layer prenet_dropout (float) : dropout rate of pre-net zoneout (float) : zoneout rate reduction_factor (int) : reduction factor downsample_by_conv (bool) : if True, downsample by convolution scaled_tanh (bool) : if True, use scaled tanh for residual F0 prediction in_lf0_idx (int): index of lf0 in input features in_lf0_min (float): minimum of lf0 in the training data of input features in_lf0_max (float): maximum of lf0 in the training data of input features out_lf0_idx (int): index of lf0 in output features out_lf0_mean (float): mean of lf0 in the training data of output features out_lf0_scale (float): scale of lf0 in the training data of output features use_mdn (bool): if True, use mixture density network for F0 prediction num_gaussians (int): number of gaussians in MDN sampling_mode (str): sampling mode in inference. "mean" or "random" in_ph_start_idx (int): Start index of phoneme features. in_ph_end_idx (int): End index of phoneme features. embed_dim (int): Embedding dimension. init_type (str): initialization type """ def __init__( self, in_dim=512, ff_hidden_dim=2048, conv_hidden_dim=1024, lstm_hidden_dim=256, num_lstm_layers=2, dropout=0.0, out_dim=80, decoder_layers=2, decoder_hidden_dim=1024, prenet_layers=2, prenet_hidden_dim=256, prenet_dropout=0.5, zoneout=0.1, reduction_factor=1, downsample_by_conv=False, scaled_tanh=True, # NOTE: you must carefully set the following parameters in_lf0_idx=300, in_lf0_min=5.3936276, in_lf0_max=6.491111, out_lf0_idx=180, out_lf0_mean=5.953093881972361, out_lf0_scale=0.23435173188961034, use_mdn=False, num_gaussians=4, sampling_mode="mean", in_ph_start_idx: int = 1, in_ph_end_idx: int = 50, embed_dim=None, init_type="none", ): super().__init__() self.reduction_factor = reduction_factor self.in_lf0_idx = in_lf0_idx self.in_lf0_min = in_lf0_min self.in_lf0_max = in_lf0_max self.out_lf0_idx = out_lf0_idx self.out_lf0_mean = out_lf0_mean self.out_lf0_scale = out_lf0_scale self.use_mdn = use_mdn self.in_dim = in_dim self.in_ph_start_idx = in_ph_start_idx self.in_ph_end_idx = in_ph_end_idx self.num_vocab = in_ph_end_idx - in_ph_start_idx self.embed_dim = embed_dim if self.embed_dim is not None: assert in_dim > self.num_vocab self.emb = nn.Embedding(self.num_vocab, embed_dim) self.fc_in = nn.Linear(in_dim - self.num_vocab, embed_dim) in_ff_dim = embed_dim else: in_ff_dim = in_dim # Encoder # NOTE: can be simply replaced by a BiLSTM? # so far I use sinsy like architecture self.ff = nn.Sequential( nn.Linear(in_ff_dim, ff_hidden_dim), nn.ReLU(), nn.Linear(ff_hidden_dim, ff_hidden_dim), nn.ReLU(), nn.Linear(ff_hidden_dim, ff_hidden_dim), nn.ReLU(), ) self.conv = nn.Sequential( nn.ReflectionPad1d(3), nn.Conv1d(ff_hidden_dim + 1, conv_hidden_dim, kernel_size=7, padding=0), nn.BatchNorm1d(conv_hidden_dim), nn.ReLU(), nn.ReflectionPad1d(3), nn.Conv1d(conv_hidden_dim, conv_hidden_dim, kernel_size=7, padding=0), nn.BatchNorm1d(conv_hidden_dim), nn.ReLU(), nn.ReflectionPad1d(3), nn.Conv1d(conv_hidden_dim, conv_hidden_dim, kernel_size=7, padding=0), nn.BatchNorm1d(conv_hidden_dim), nn.ReLU(), ) self.lstm = nn.LSTM( conv_hidden_dim, lstm_hidden_dim, num_lstm_layers, bidirectional=True, batch_first=True, dropout=dropout, ) # Autoregressive decoder decoder_in_dim = 2 * lstm_hidden_dim + 1 # NOTE: hard code in_lf0_idx to -1 cls = MDNResF0NonAttentiveDecoder if use_mdn else ResF0NonAttentiveDecoder if use_mdn: ex_kwargs = {"num_gaussians": num_gaussians, "sampling_mode": sampling_mode} else: ex_kwargs = {} self.decoder = cls( in_dim=decoder_in_dim, out_dim=out_dim, layers=decoder_layers, hidden_dim=decoder_hidden_dim, prenet_layers=prenet_layers, prenet_hidden_dim=prenet_hidden_dim, prenet_dropout=prenet_dropout, zoneout=zoneout, reduction_factor=reduction_factor, downsample_by_conv=downsample_by_conv, scaled_tanh=scaled_tanh, in_lf0_idx=-1, in_lf0_min=in_lf0_min, in_lf0_max=in_lf0_max, out_lf0_idx=out_lf0_idx, out_lf0_mean=out_lf0_mean, out_lf0_scale=out_lf0_scale, **ex_kwargs, ) init_weights(self, init_type) def _set_lf0_params(self): # Special care for residual F0 prediction models # NOTE: don't overwrite out_lf0_idx and in_lf0_idx if hasattr(self.decoder, "out_lf0_mean"): self.decoder.in_lf0_min = self.in_lf0_min self.decoder.in_lf0_max = self.in_lf0_max self.decoder.out_lf0_mean = self.out_lf0_mean self.decoder.out_lf0_scale = self.out_lf0_scale def is_autoregressive(self): return self.decoder.is_autoregressive() def prediction_type(self): return ( PredictionType.PROBABILISTIC if self.use_mdn else PredictionType.DETERMINISTIC ) def has_residual_lf0_prediction(self): return True def forward(self, x, lengths=None, y=None): self._set_lf0_params() lf0_score = x[:, :, self.in_lf0_idx].unsqueeze(-1) if self.embed_dim is not None: x_first, x_ph_onehot, x_last = torch.split( x, [ self.in_ph_start_idx, self.num_vocab, self.in_dim - self.num_vocab - self.in_ph_start_idx, ], dim=-1, ) x_ph = torch.argmax(x_ph_onehot, dim=-1) # Make sure to have one-hot vector assert (x_ph_onehot.sum(-1) <= 1).all() x = self.emb(x_ph) + self.fc_in(torch.cat([x_first, x_last], dim=-1)) if isinstance(lengths, torch.Tensor): lengths = lengths.to("cpu") out = self.ff(x) out = torch.cat([out, lf0_score], dim=-1) out = self.conv(out.transpose(1, 2)).transpose(1, 2) sequence = pack_padded_sequence(out, lengths, batch_first=True) out, _ = self.lstm(sequence) out, _ = pad_packed_sequence(out, batch_first=True) # NOTE: need to concat the lf0 score to the output of the lstm to tell # the decoder the lf0 out = torch.cat([out, lf0_score], dim=-1) outs, lf0_residual = self.decoder(out, lengths, y) return outs, lf0_residual def inference(self, x, lengths=None): return pad_inference( model=self, x=x, lengths=lengths, reduction_factor=self.reduction_factor, mdn=self.use_mdn, )
27,907
36.111702
90
py
nnsvs
nnsvs-master/nnsvs/acoustic_models/util.py
import numpy as np import torch from nnsvs.base import PredictionType from nnsvs.mdn import mdn_get_most_probable_sigma_and_mu from torch.nn import functional as F def predict_lf0_with_residual( in_feats, out_feats, in_lf0_idx=300, in_lf0_min=5.3936276, in_lf0_max=6.491111, out_lf0_idx=180, out_lf0_mean=5.953093881972361, out_lf0_scale=0.23435173188961034, residual_f0_max_cent=600, ): """Predict log-F0 with residual. Args: in_feats (np.ndarray): input features out_feats (np.ndarray): output of an acoustic model in_lf0_idx (int): index of LF0 in input features in_lf0_min (float): minimum value of LF0 in the training data of input features in_lf0_max (float): maximum value of LF0 in the training data of input features out_lf0_idx (int): index of LF0 in output features out_lf0_mean (float): mean of LF0 in the training data of output features out_lf0_scale (float): scale of LF0 in the training data of output features residual_f0_max_cent (int): maximum value of residual LF0 in cent Returns: tuple: (predicted log-F0, residual log-F0) """ # Denormalize lf0 from input musical score lf0_score = in_feats[:, :, in_lf0_idx].unsqueeze(-1) lf0_score_denorm = lf0_score * (in_lf0_max - in_lf0_min) + in_lf0_min # To avoid unbounded residual f0 that would potentially cause artifacts, # let's constrain the residual F0 to be in a certain range by the scaled tanh max_lf0_ratio = residual_f0_max_cent * np.log(2) / 1200 if len(out_feats.shape) == 4: # MDN case (B, T, num_gaussians, C) -> (B, T, num_gaussians) lf0_residual = out_feats[:, :, :, out_lf0_idx] else: # (B, T, C) -> (B, T, 1) lf0_residual = out_feats[:, :, out_lf0_idx].unsqueeze(-1) lf0_residual = max_lf0_ratio * torch.tanh(lf0_residual) # Residual connection in the denormalized f0 domain lf0_pred_denorm = lf0_score_denorm + lf0_residual # Back to normalized f0 lf0_pred = (lf0_pred_denorm - out_lf0_mean) / out_lf0_scale return lf0_pred, lf0_residual def pad_inference( model, x, lengths, reduction_factor, mode="replicate", y=None, mdn=False ): mod = max(lengths) % reduction_factor pad = reduction_factor - mod # Pad zeros to the end of the input features # so that the length of the input features is a multiple of the reduction factor if pad != 0: x_pad = F.pad(x, (0, 0, 0, pad), mode=mode) y_pad = F.pad(y, (0, 0, 0, pad), mode=mode) if y is not None else None if isinstance(lengths, torch.Tensor): lengths = lengths.clone() else: lengths = lengths.copy() lengths = [length + pad for length in lengths] else: x_pad = x y_pad = y if y is not None else None y = model(x_pad, lengths, y_pad) # Residual F0 prediction: (out, lf0) if mdn: assert isinstance(y, tuple) and len(y) == 2 # NOTE: need to parse per-stream output if ( model.prediction_type() == PredictionType.MULTISTREAM_HYBRID and y_pad is not None ): if len(y[0]) == 4: mgc, lf0, vuv, bap = y[0] if isinstance(mgc, tuple) and len(mgc) == 3: mgc = mdn_get_most_probable_sigma_and_mu(*mgc)[1] elif isinstance(mgc, tuple) and len(mgc) == 2: mgc = mgc[1] if isinstance(bap, tuple) and len(bap) == 3: bap = mdn_get_most_probable_sigma_and_mu(*bap)[1] elif isinstance(bap, tuple) and len(bap) == 2: bap = bap[1] if pad != 0: mgc = mgc[:, :-pad] if mgc.shape[1] > x.shape[1] else mgc bap = bap[:, :-pad] if bap.shape[1] > x.shape[1] else bap lf0 = lf0[:, :-pad] if lf0.shape[1] > x.shape[1] else lf0 vuv = vuv[:, :-pad] if vuv.shape[1] > x.shape[1] else vuv mu = torch.cat([mgc, lf0, vuv, bap], dim=-1) elif len(y[0]) == 3: mel, lf0, vuv = y[0] if isinstance(mel, tuple) and len(mel) == 3: mel = mdn_get_most_probable_sigma_and_mu(*mel)[1] elif isinstance(mel, tuple) and len(mel) == 2: mel = mel[1] if pad != 0: mel = mel[:, :-pad] if mel.shape[1] > x.shape[1] else mel lf0 = lf0[:, :-pad] if lf0.shape[1] > x.shape[1] else lf0 vuv = vuv[:, :-pad] if vuv.shape[1] > x.shape[1] else vuv mu = torch.cat([mel, lf0, vuv], dim=-1) else: raise ValueError("Invalid number of streams: {}".format(len(y))) sigma = mu else: mu, sigma = y if pad != 0: mu = mu[:, :-pad] sigma = sigma[:, :-pad] y = (mu, sigma) else: if model.has_residual_lf0_prediction(): y = y[0] # Multiple output: (out, out_fine) if isinstance(y, list): y = y[-1] if pad != 0: y = y[:, :-pad] return y
5,262
38.276119
87
py
nnsvs
nnsvs-master/nnsvs/acoustic_models/sinsy.py
import torch from nnsvs.acoustic_models.util import predict_lf0_with_residual from nnsvs.base import BaseModel, PredictionType from nnsvs.mdn import MDNLayer, mdn_get_most_probable_sigma_and_mu from nnsvs.util import init_weights from torch import nn from torch.nn.utils.rnn import pack_padded_sequence, pad_packed_sequence __all__ = [ "ResSkipF0FFConvLSTM", ] class ResSkipF0FFConvLSTM(BaseModel): """FFN + Conv1d + LSTM + residual/skip connections A model proposed in :cite:t:`hono2021sinsy`. Args: in_dim (int): input dimension ff_hidden_dim (int): hidden dimension of feed-forward layer conv_hidden_dim (int): hidden dimension of convolutional layer lstm_hidden_dim (int): hidden dimension of LSTM layer out_dim (int): output dimension dropout (float): dropout rate num_ls (int): number of layers of LSTM bidirectional (bool): whether to use bidirectional LSTM or not in_lf0_idx (int): index of lf0 in input features in_lf0_min (float): minimum of lf0 in the training data of input features in_lf0_max (float): maximum of lf0 in the training data of input features out_lf0_idx (int): index of lf0 in output features out_lf0_mean (float): mean of lf0 in the training data of output features out_lf0_scale (float): scale of lf0 in the training data of output features skip_inputs (bool): whether to use skip connection for the input features init_type (str): initialization type use_mdn (bool): whether to use MDN or not num_gaussians (int): number of gaussians in MDN dim_wise (bool): whether to use MDN with dim-wise or not """ def __init__( self, in_dim, ff_hidden_dim=2048, conv_hidden_dim=1024, lstm_hidden_dim=256, out_dim=199, dropout=0.0, num_lstm_layers=2, bidirectional=True, # NOTE: you must carefully set the following parameters in_lf0_idx=300, in_lf0_min=5.3936276, in_lf0_max=6.491111, out_lf0_idx=180, out_lf0_mean=5.953093881972361, out_lf0_scale=0.23435173188961034, skip_inputs=False, init_type="none", use_mdn=False, num_gaussians=8, dim_wise=False, ): super().__init__() self.in_lf0_idx = in_lf0_idx self.in_lf0_min = in_lf0_min self.in_lf0_max = in_lf0_max self.out_lf0_idx = out_lf0_idx self.out_lf0_mean = out_lf0_mean self.out_lf0_scale = out_lf0_scale self.skip_inputs = skip_inputs self.use_mdn = use_mdn self.ff = nn.Sequential( nn.Linear(in_dim, ff_hidden_dim), nn.ReLU(), nn.Linear(ff_hidden_dim, ff_hidden_dim), nn.ReLU(), nn.Linear(ff_hidden_dim, ff_hidden_dim), nn.ReLU(), ) self.conv = nn.Sequential( nn.ReflectionPad1d(3), nn.Conv1d(ff_hidden_dim + 1, conv_hidden_dim, kernel_size=7, padding=0), nn.BatchNorm1d(conv_hidden_dim), nn.ReLU(), nn.ReflectionPad1d(3), nn.Conv1d(conv_hidden_dim, conv_hidden_dim, kernel_size=7, padding=0), nn.BatchNorm1d(conv_hidden_dim), nn.ReLU(), nn.ReflectionPad1d(3), nn.Conv1d(conv_hidden_dim, conv_hidden_dim, kernel_size=7, padding=0), nn.BatchNorm1d(conv_hidden_dim), nn.ReLU(), ) num_direction = 2 if bidirectional else 1 self.lstm = nn.LSTM( conv_hidden_dim, lstm_hidden_dim, num_lstm_layers, bidirectional=True, batch_first=True, dropout=dropout, ) if self.skip_inputs: last_in_dim = num_direction * lstm_hidden_dim + in_dim else: last_in_dim = num_direction * lstm_hidden_dim if self.use_mdn: self.mdn_layer = MDNLayer( last_in_dim, out_dim, num_gaussians=num_gaussians, dim_wise=dim_wise ) else: self.fc = nn.Linear(last_in_dim, out_dim) init_weights(self, init_type) def prediction_type(self): return ( PredictionType.PROBABILISTIC if self.use_mdn else PredictionType.DETERMINISTIC ) def has_residual_lf0_prediction(self): return True def forward(self, x, lengths=None, y=None): """Forward step Args: x (torch.Tensor): the input tensor lengths (torch.Tensor): the lengths of the input tensor y (torch.Tensor): the optional target tensor Returns: torch.Tensor: the output tensor """ if isinstance(lengths, torch.Tensor): lengths = lengths.to("cpu") lf0_score = x[:, :, self.in_lf0_idx].unsqueeze(-1) out = self.ff(x) out = torch.cat([out, lf0_score], dim=-1) out = self.conv(out.transpose(1, 2)).transpose(1, 2) sequence = pack_padded_sequence(out, lengths, batch_first=True) out, _ = self.lstm(sequence) out, _ = pad_packed_sequence(out, batch_first=True) out = torch.cat([out, x], dim=-1) if self.skip_inputs else out if self.use_mdn: log_pi, log_sigma, mu = self.mdn_layer(out) else: mu = self.fc(out) lf0_pred, lf0_residual = predict_lf0_with_residual( x, mu, self.in_lf0_idx, self.in_lf0_min, self.in_lf0_max, self.out_lf0_idx, self.out_lf0_mean, self.out_lf0_scale, ) # Inject the predicted lf0 into the output features if self.use_mdn: mu[:, :, :, self.out_lf0_idx] = lf0_pred else: mu[:, :, self.out_lf0_idx] = lf0_pred.squeeze(-1) if self.use_mdn: return (log_pi, log_sigma, mu), lf0_residual else: return mu, lf0_residual def inference(self, x, lengths=None): """Inference step Args: x (torch.Tensor): input features lengths (torch.Tensor): lengths of input features Returns: tuple: (mu, sigma) if use_mdn, (output, ) otherwise """ if self.use_mdn: (log_pi, log_sigma, mu), _ = self(x, lengths) sigma, mu = mdn_get_most_probable_sigma_and_mu(log_pi, log_sigma, mu) return mu, sigma else: return self(x, lengths)[0]
6,623
32.12
84
py
nnsvs
nnsvs-master/nnsvs/acoustic_models/__init__.py
from functools import partial from nnsvs.acoustic_models.multistream import ( MDNMultistreamSeparateF0MelModel, MultistreamSeparateF0MelModel, MultistreamSeparateF0ParametricModel, NPSSMDNMultistreamParametricModel, NPSSMultistreamParametricModel, ) from nnsvs.acoustic_models.sinsy import ResSkipF0FFConvLSTM from nnsvs.acoustic_models.tacotron import ( BiLSTMMDNNonAttentiveDecoder, BiLSTMNonAttentiveDecoder, MDNNonAttentiveDecoder, NonAttentiveDecoder, ) from nnsvs.acoustic_models.tacotron_f0 import ( BiLSTMResF0NonAttentiveDecoder, MDNResF0NonAttentiveDecoder, ResF0NonAttentiveDecoder, ) from nnsvs.acoustic_models.util import predict_lf0_with_residual from nnsvs.base import BaseModel, PredictionType from nnsvs.layers.conv import ResnetBlock, WNConv1d from nnsvs.mdn import MDNLayer, mdn_get_most_probable_sigma_and_mu from nnsvs.model import TransformerEncoder, VariancePredictor from nnsvs.util import init_weights from torch import nn __all__ = [ # Non-autoregressive models "ResF0Conv1dResnet", "ResSkipF0FFConvLSTM", "ResF0VariancePredictor", "ResF0TransformerEncoder", # Autoregressive models "NonAttentiveDecoder", "MDNNonAttentiveDecoder", "BiLSTMNonAttentiveDecoder", "BiLSTMMDNNonAttentiveDecoder", "ResF0NonAttentiveDecoder", "MDNResF0NonAttentiveDecoder", "BiLSTMResF0NonAttentiveDecoder", # Multi-stream models "MultistreamSeparateF0ParametricModel", "NPSSMDNMultistreamParametricModel", "NPSSMultistreamParametricModel", "MultistreamSeparateF0MelModel", "MDNMultistreamSeparateF0MelModel", ] class ResF0Conv1dResnet(BaseModel): """Conv1d + Resnet + Residual F0 prediction Residual F0 prediction is inspired by :cite:t:`hono2021sinsy`. Args: in_dim (int): input dimension hidden_dim (int): hidden dimension out_dim (int): output dimension num_layers (int): number of layers in_lf0_idx (int): index of lf0 in input features in_lf0_min (float): minimum value of lf0 in the training data of input features in_lf0_max (float): maximum value of lf0 in the training data of input features out_lf0_idx (int): index of lf0 in output features. Typically 180. out_lf0_mean (float): mean of lf0 in the training data of output features out_lf0_scale (float): scale of lf0 in the training data of output features init_type (str): initialization type use_mdn (bool): whether to use MDN or not num_gaussians (int): number of gaussians in MDN dim_wise (bool): whether to use dimension-wise MDN or not """ def __init__( self, in_dim, hidden_dim, out_dim, num_layers=4, # NOTE: you must carefully set the following parameters in_lf0_idx=300, in_lf0_min=5.3936276, in_lf0_max=6.491111, out_lf0_idx=180, out_lf0_mean=5.953093881972361, out_lf0_scale=0.23435173188961034, init_type="none", use_mdn=False, num_gaussians=8, dim_wise=False, ): super().__init__() self.in_lf0_idx = in_lf0_idx self.in_lf0_min = in_lf0_min self.in_lf0_max = in_lf0_max self.out_lf0_idx = out_lf0_idx self.out_lf0_mean = out_lf0_mean self.out_lf0_scale = out_lf0_scale self.use_mdn = use_mdn model = [ nn.ReflectionPad1d(3), WNConv1d(in_dim, hidden_dim, kernel_size=7, padding=0), ] for n in range(num_layers): model.append(ResnetBlock(hidden_dim, dilation=2 ** n)) last_conv_out_dim = hidden_dim if use_mdn else out_dim model += [ nn.LeakyReLU(0.2), nn.ReflectionPad1d(3), WNConv1d(hidden_dim, last_conv_out_dim, kernel_size=7, padding=0), ] self.model = nn.Sequential(*model) if self.use_mdn: self.mdn_layer = MDNLayer( in_dim=hidden_dim, out_dim=out_dim, num_gaussians=num_gaussians, dim_wise=dim_wise, ) else: self.mdn_layer = None init_weights(self, init_type) def prediction_type(self): return ( PredictionType.PROBABILISTIC if self.use_mdn else PredictionType.DETERMINISTIC ) def has_residual_lf0_prediction(self): return True def forward(self, x, lengths=None, y=None): """Forward step Args: x (torch.Tensor): input features lengths (torch.Tensor): lengths of input features y (torch.Tensor): output features Returns: tuple: (output features, residual log-F0) """ out = self.model(x.transpose(1, 2)).transpose(1, 2) if self.use_mdn: log_pi, log_sigma, mu = self.mdn_layer(out) else: mu = out lf0_pred, lf0_residual = predict_lf0_with_residual( x, mu, self.in_lf0_idx, self.in_lf0_min, self.in_lf0_max, self.out_lf0_idx, self.out_lf0_mean, self.out_lf0_scale, ) # Inject the predicted lf0 into the output features if self.use_mdn: mu[:, :, :, self.out_lf0_idx] = lf0_pred else: mu[:, :, self.out_lf0_idx] = lf0_pred.squeeze(-1) if self.use_mdn: return (log_pi, log_sigma, mu), lf0_residual else: return mu, lf0_residual def inference(self, x, lengths=None): """Inference step Args: x (torch.Tensor): input features lengths (torch.Tensor): lengths of input features Returns: tuple: (mu, sigma) if use_mdn, (output, ) otherwise """ if self.use_mdn: (log_pi, log_sigma, mu), _ = self(x, lengths) sigma, mu = mdn_get_most_probable_sigma_and_mu(log_pi, log_sigma, mu) return mu, sigma else: return self(x, lengths)[0] # Will be removed in v0.1.0 ResF0Conv1dResnetMDN = partial(ResF0Conv1dResnet, use_mdn=True) class ResF0VariancePredictor(VariancePredictor): """Variance predictor in :cite:t:`ren2020fastspeech` with residual F0 prediction Args: in_dim (int): the input dimension out_dim (int): the output dimension num_layers (int): the number of layers hidden_dim (int): the hidden dimension kernel_size (int): the kernel size dropout (float): the dropout rate in_lf0_idx (int): the index of the input LF0 in_lf0_min (float): the minimum value of the input LF0 in_lf0_max (float): the maximum value of the input LF0 out_lf0_idx (int): the index of the output LF0 out_lf0_mean (float): the mean value of the output LF0 out_lf0_scale (float): the scale value of the output LF0 init_type (str): the initialization type use_mdn (bool): whether to use MDN or not num_gaussians (int): the number of gaussians dim_wise (bool): whether to use dim-wise or not """ def __init__( self, in_dim, out_dim, num_layers=5, hidden_dim=256, kernel_size=5, dropout=0.5, init_type="none", use_mdn=False, num_gaussians=1, dim_wise=False, # NOTE: you must carefully set the following parameters in_lf0_idx=300, in_lf0_min=5.3936276, in_lf0_max=6.491111, out_lf0_idx=180, out_lf0_mean=5.953093881972361, out_lf0_scale=0.23435173188961034, ): super().__init__( in_dim=in_dim, out_dim=out_dim, num_layers=num_layers, hidden_dim=hidden_dim, kernel_size=kernel_size, dropout=dropout, init_type=init_type, use_mdn=use_mdn, num_gaussians=num_gaussians, dim_wise=dim_wise, ) self.in_lf0_idx = in_lf0_idx self.in_lf0_min = in_lf0_min self.in_lf0_max = in_lf0_max self.out_lf0_idx = out_lf0_idx self.out_lf0_mean = out_lf0_mean self.out_lf0_scale = out_lf0_scale def has_residual_lf0_prediction(self): return True def forward(self, x, lengths=None, y=None): """Forward step Args: x (torch.Tensor): input features lengths (torch.Tensor): lengths of input features y (torch.Tensor): output features Returns: tuple: (output features, residual log-F0) """ out = super().forward(x, lengths, y) if self.use_mdn: log_pi, log_sigma, mu = out else: mu = out lf0_pred, lf0_residual = predict_lf0_with_residual( x, mu, self.in_lf0_idx, self.in_lf0_min, self.in_lf0_max, self.out_lf0_idx, self.out_lf0_mean, self.out_lf0_scale, ) # Inject the predicted lf0 into the output features if self.use_mdn: mu[:, :, :, self.out_lf0_idx] = lf0_pred else: mu[:, :, self.out_lf0_idx] = lf0_pred.squeeze(-1) if self.use_mdn: return (log_pi, log_sigma, mu), lf0_residual else: return mu, lf0_residual def inference(self, x, lengths=None): """Inference step Args: x (torch.Tensor): input features lengths (torch.Tensor): lengths of input features Returns: tuple: (mu, sigma) if use_mdn, (output, ) otherwise """ if self.use_mdn: (log_pi, log_sigma, mu), _ = self(x, lengths) sigma, mu = mdn_get_most_probable_sigma_and_mu(log_pi, log_sigma, mu) return mu, sigma else: return self(x, lengths)[0] class ResF0TransformerEncoder(BaseModel): """Transformer encoder with residual f0 prediction""" def __init__( self, in_dim, out_dim, hidden_dim, attention_dim, num_heads=2, num_layers=2, kernel_size=3, dropout=0.1, reduction_factor=1, init_type="none", downsample_by_conv=False, # NOTE: you must carefully set the following parameters in_lf0_idx=300, in_lf0_min=5.3936276, in_lf0_max=6.491111, out_lf0_idx=180, out_lf0_mean=5.953093881972361, out_lf0_scale=0.23435173188961034, ): super().__init__() self.in_lf0_idx = in_lf0_idx self.in_lf0_min = in_lf0_min self.in_lf0_max = in_lf0_max self.out_lf0_idx = out_lf0_idx self.out_lf0_mean = out_lf0_mean self.out_lf0_scale = out_lf0_scale self.reduction_factor = reduction_factor self.encoder = TransformerEncoder( in_dim=in_dim, out_dim=out_dim, hidden_dim=hidden_dim, attention_dim=attention_dim, num_heads=num_heads, num_layers=num_layers, kernel_size=kernel_size, dropout=dropout, reduction_factor=reduction_factor, init_type=init_type, downsample_by_conv=downsample_by_conv, ) def has_residual_lf0_prediction(self): return True def forward(self, x, lengths=None, y=None): """Forward pass Args: x (torch.Tensor): input tensor lengths (torch.Tensor): input sequence lengths y (torch.Tensor): target tensor (optional) Returns: torch.Tensor: output tensor """ outs = self.encoder(x, lengths) lf0_pred, lf0_residual = predict_lf0_with_residual( x, outs, self.in_lf0_idx, self.in_lf0_min, self.in_lf0_max, self.out_lf0_idx, self.out_lf0_mean, self.out_lf0_scale, ) outs[:, :, self.out_lf0_idx] = lf0_pred.squeeze(-1) return outs, lf0_residual def inference(self, x, lengths): return self(x, lengths)[0] def LSTMEncoder(*args, **kwargs): from nnsvs.model import LSTMEncoder return LSTMEncoder(*args, **kwargs)
12,451
29.669951
87
py
nnsvs
nnsvs-master/nnsvs/bin/train_postfilter.py
from functools import partial from pathlib import Path import hydra import mlflow import numpy as np import torch import torch.distributed as dist from hydra.utils import to_absolute_path from nnsvs.multistream import select_streams from nnsvs.train_util import ( collate_fn_default, collate_fn_random_segments, compute_distortions, eval_model, log_params_from_omegaconf_dict, save_checkpoint, save_configs, setup_gan, ) from nnsvs.util import PyTorchStandardScaler, load_vocoder, make_non_pad_mask from omegaconf import DictConfig from torch import nn from torch.cuda.amp import autocast from torch.nn import functional as F def train_step( model_config, optim_config, netG, optG, netD, optD, grad_scaler, train, in_feats, out_feats, lengths, out_scaler, mse_weight=1.0, adv_weight=1.0, adv_streams=None, fm_weight=0.0, mask_nth_mgc_for_adv_loss=0, gan_type="lsgan", vuv_mask=False, ): netG.train() if train else netG.eval() netD.train() if train else netD.eval() log_metrics = {} if vuv_mask: # NOTE: Assuming 3rd stream is the V/UV vuv_idx = np.sum(model_config.stream_sizes[:2]) is_v = torch.logical_and( out_feats[:, :, vuv_idx : vuv_idx + 1] > 0, in_feats[:, :, vuv_idx : vuv_idx + 1] > 0, ) vuv = is_v else: vuv = 1.0 # Run forward with autocast(enabled=grad_scaler is not None): pred_out_feats = netG(in_feats, lengths) real_netD_in_feats = select_streams( out_feats, model_config.stream_sizes, adv_streams ) fake_netD_in_feats = select_streams( pred_out_feats, model_config.stream_sizes, adv_streams, ) # Ref: http://sython.org/papers/ASJ/saito2017asja.pdf # 0-th mgc with adversarial trainging affects speech quality # NOTE: assuming that the first stream contains mgc if mask_nth_mgc_for_adv_loss > 0: real_netD_in_feats = real_netD_in_feats[:, :, mask_nth_mgc_for_adv_loss:] fake_netD_in_feats = fake_netD_in_feats[:, :, mask_nth_mgc_for_adv_loss:] # Real with autocast(enabled=grad_scaler is not None): D_real = netD(real_netD_in_feats * vuv, in_feats, lengths) # NOTE: must be list of list to support multi-scale discriminators assert isinstance(D_real, list) and isinstance(D_real[-1], list) # Fake D_fake_det = netD(fake_netD_in_feats.detach() * vuv, in_feats, lengths) # Mask (B, T, 1) mask = make_non_pad_mask(lengths).unsqueeze(-1).to(in_feats.device) # Update discriminator eps = 1e-14 loss_real = 0 loss_fake = 0 with autocast(enabled=grad_scaler is not None): for idx, (D_real_, D_fake_det_) in enumerate(zip(D_real, D_fake_det)): if gan_type == "lsgan": loss_real_ = (D_real_[-1] - 1) ** 2 loss_fake_ = D_fake_det_[-1] ** 2 elif gan_type == "vanilla-gan": loss_real_ = -torch.log(D_real_[-1] + eps) loss_fake_ = -torch.log(1 - D_fake_det_[-1] + eps) elif gan_type == "hinge": loss_real_ = F.relu(1 - D_real_[-1]) loss_fake_ = F.relu(1 + D_fake_det_[-1]) else: raise ValueError(f"Unknown gan type: {gan_type}") # mask for D if ( hasattr(netD, "downsample_scale") and mask.shape[1] // netD.downsample_scale == D_real_[-1].shape[1] ): D_mask = mask[:, :: netD.downsample_scale, :] else: if D_real_[-1].shape[1] == out_feats.shape[1]: D_mask = mask else: D_mask = None if D_mask is not None: loss_real_ = loss_real_.masked_select(D_mask).mean() loss_fake_ = loss_fake_.masked_select(D_mask).mean() else: loss_real_ = loss_real_.mean() loss_fake_ = loss_fake_.mean() log_metrics[f"Loss_Real_Scale{idx}"] = loss_real_.item() log_metrics[f"Loss_Fake_Scale{idx}"] = loss_fake_.item() loss_real += loss_real_ loss_fake += loss_fake_ loss_d = loss_real + loss_fake if train: optD.zero_grad() if grad_scaler is not None: grad_scaler.scale(loss_d).backward() grad_scaler.unscale_(optD) grad_norm_d = torch.nn.utils.clip_grad_norm_( netD.parameters(), optim_config.netD.clip_norm ) log_metrics["GradNorm_D"] = grad_norm_d grad_scaler.step(optD) else: loss_d.backward() grad_norm_d = torch.nn.utils.clip_grad_norm_( netD.parameters(), optim_config.netD.clip_norm ) log_metrics["GradNorm_D"] = grad_norm_d optD.step() # MSE loss with autocast(enabled=grad_scaler is not None): loss_feats = nn.MSELoss(reduction="none")( pred_out_feats.masked_select(mask), out_feats.masked_select(mask) ).mean() # adversarial loss with autocast(enabled=grad_scaler is not None): D_fake = netD(fake_netD_in_feats * vuv, in_feats, lengths) loss_adv = 0 for idx, D_fake_ in enumerate(D_fake): if gan_type == "lsgan": loss_adv_ = (1 - D_fake_[-1]) ** 2 elif gan_type == "vanilla-gan": loss_adv_ = -torch.log(D_fake_[-1] + eps) elif gan_type == "hinge": loss_adv_ = -D_fake_[-1] else: raise ValueError(f"Unknown gan type: {gan_type}") if ( hasattr(netD, "downsample_scale") and mask.shape[1] // netD.downsample_scale == D_fake_[-1].shape[1] ): D_mask = mask[:, :: netD.downsample_scale, :] else: if D_real_[-1].shape[1] == out_feats.shape[1]: D_mask = mask else: D_mask = None if D_mask is not None: loss_adv_ = loss_adv_.masked_select(D_mask).mean() else: loss_adv_ = loss_adv_.mean() log_metrics[f"Loss_Adv_Scale{idx}"] = loss_adv_.item() loss_adv += loss_adv_ # Feature matching loss loss_fm = torch.tensor(0.0).to(in_feats.device) if fm_weight > 0: for D_fake_, D_real_ in zip(D_fake, D_real): for fake_fmap, real_fmap in zip(D_fake_[:-1], D_real_[:-1]): loss_fm += F.l1_loss(fake_fmap, real_fmap.detach()) loss = mse_weight * loss_feats + adv_weight * loss_adv + fm_weight * loss_fm if train: optG.zero_grad() if grad_scaler is not None: grad_scaler.scale(loss).backward() grad_scaler.unscale_(optG) grad_norm_g = torch.nn.utils.clip_grad_norm_( netG.parameters(), optim_config.netG.clip_norm ) log_metrics["GradNorm_G"] = grad_norm_g grad_scaler.step(optG) else: loss.backward() grad_norm_g = torch.nn.utils.clip_grad_norm_( netG.parameters(), optim_config.netG.clip_norm ) log_metrics["GradNorm_G"] = grad_norm_g optG.step() # NOTE: this shouldn't be called multiple times in a training step if train and grad_scaler is not None: grad_scaler.update() # Metrics distortions = compute_distortions( pred_out_feats, out_feats, lengths, out_scaler, model_config ) log_metrics.update(distortions) log_metrics.update( { "Loss": loss.item(), "Loss_Feats": loss_feats.item(), "Loss_Adv_Total": loss_adv.item(), "Loss_Feature_Matching": loss_fm.item(), "Loss_Real_Total": loss_real.item(), "Loss_Fake_Total": loss_fake.item(), "Loss_D": loss_d.item(), } ) return loss, log_metrics def train_loop( config, logger, device, netG, optG, schedulerG, netD, optD, schedulerD, grad_scaler, data_loaders, samplers, writer, out_scaler, use_mlflow, vocoder, vocoder_in_scaler, ): out_dir = Path(to_absolute_path(config.train.out_dir)) best_dev_loss = torch.finfo(torch.float32).max last_dev_loss = torch.finfo(torch.float32).max adv_streams = config.train.adv_streams if len(adv_streams) != len(config.model.stream_sizes): raise ValueError("adv_streams must be specified for all streams") if dist.is_initialized() and dist.get_rank() != 0: def tqdm(x, **kwargs): return x else: from tqdm import tqdm train_iter = 1 for epoch in tqdm(range(1, config.train.nepochs + 1)): for phase in data_loaders.keys(): train = phase.startswith("train") # https://pytorch.org/docs/stable/data.html#torch.utils.data.distributed.DistributedSampler if dist.is_initialized() and train and samplers[phase] is not None: samplers[phase].set_epoch(epoch) running_loss = 0 running_metrics = {} evaluated = False for in_feats, out_feats, lengths in tqdm( data_loaders[phase], desc=f"{phase} iter", leave=False ): # NOTE: This is needed for pytorch's PackedSequence lengths, indices = torch.sort(lengths, dim=0, descending=True) in_feats, out_feats = ( in_feats[indices].to(device), out_feats[indices].to(device), ) if (not train) and (not evaluated): eval_model( phase, epoch, netG, in_feats, out_feats, lengths, config.model, out_scaler, writer, sr=config.data.sample_rate, use_world_codec=config.data.use_world_codec, vocoder=vocoder, vocoder_in_scaler=vocoder_in_scaler, max_num_eval_utts=config.train.max_num_eval_utts, ) evaluated = True loss, log_metrics = train_step( model_config=config.model, optim_config=config.train.optim, netG=netG, optG=optG, netD=netD, optD=optD, grad_scaler=grad_scaler, train=train, in_feats=in_feats, out_feats=out_feats, lengths=lengths, out_scaler=out_scaler, mse_weight=config.train.mse_weight, adv_weight=config.train.adv_weight, adv_streams=adv_streams, fm_weight=config.train.fm_weight, mask_nth_mgc_for_adv_loss=config.train.mask_nth_mgc_for_adv_loss, gan_type=config.train.gan_type, vuv_mask=config.train.vuv_mask, ) if train: if writer is not None: for key, val in log_metrics.items(): writer.add_scalar(f"{key}_Step/{phase}", val, train_iter) train_iter += 1 running_loss += loss.item() for k, v in log_metrics.items(): try: running_metrics[k] += float(v) except KeyError: running_metrics[k] = float(v) ave_loss = running_loss / len(data_loaders[phase]) logger.info("[%s] [Epoch %s]: loss %s", phase, epoch, ave_loss) if writer is not None: writer.add_scalar(f"Loss_Epoch/{phase}", ave_loss, epoch) if use_mlflow: mlflow.log_metric(f"{phase}_loss", ave_loss, step=epoch) for k, v in running_metrics.items(): ave_v = v / len(data_loaders[phase]) if writer is not None: writer.add_scalar(f"{k}_Epoch/{phase}", ave_v, epoch) if use_mlflow: mlflow.log_metric(f"{phase}_{k}", ave_v, step=epoch) if not train: last_dev_loss = ave_loss if not train and ave_loss < best_dev_loss: best_dev_loss = ave_loss for model, opt, scheduler, postfix in [ (netG, optG, schedulerG, ""), (netD, optD, schedulerD, "_D"), ]: save_checkpoint( logger, out_dir, model, opt, scheduler, epoch, is_best=True, postfix=postfix, ) schedulerG.step() schedulerD.step() if epoch % config.train.checkpoint_epoch_interval == 0: for model, opt, scheduler, postfix in [ (netG, optG, schedulerG, ""), (netD, optD, schedulerD, "_D"), ]: save_checkpoint( logger, out_dir, model, opt, scheduler, epoch, is_best=False, postfix=postfix, ) for model, opt, scheduler, postfix in [ (netG, optG, schedulerG, ""), (netD, optD, schedulerD, "_D"), ]: save_checkpoint( logger, out_dir, model, opt, scheduler, config.train.nepochs, postfix=postfix, ) logger.info("The best loss was %s", best_dev_loss) if use_mlflow: mlflow.log_metric("best_dev_loss", best_dev_loss, step=epoch) mlflow.log_artifacts(out_dir) return last_dev_loss @hydra.main(config_path="conf/train_postfilter", config_name="config") def my_app(config: DictConfig) -> None: # NOTE: set discriminator's in_dim automatically if config.model.netD.in_dim is None: stream_sizes = np.asarray(config.model.stream_sizes) D_in_dim = int((stream_sizes * np.asarray(config.train.adv_streams)).sum()) if config.train.mask_nth_mgc_for_adv_loss > 0: D_in_dim -= config.train.mask_nth_mgc_for_adv_loss config.model.netD.in_dim = D_in_dim if "stream_sizes" in config.model.netG: config.model.netG.stream_sizes = config.model.stream_sizes if "max_time_frames" in config.data and config.data.max_time_frames > 0: collate_fn = partial( collate_fn_random_segments, max_time_frames=config.data.max_time_frames ) else: collate_fn = collate_fn_default if config.train.use_ddp: dist.init_process_group("nccl") rank = dist.get_rank() device_id = rank % torch.cuda.device_count() torch.cuda.set_device(device_id) device = torch.device("cuda" if torch.cuda.is_available() else "cpu") ( (netG, optG, schedulerG), (netD, optD, schedulerD), grad_scaler, data_loaders, samplers, writer, logger, _, out_scaler, ) = setup_gan(config, device, collate_fn) path = config.train.pretrained_vocoder_checkpoint if path is not None and len(path) > 0: logger.info(f"Loading pretrained vocoder checkpoint from {path}") vocoder, vocoder_in_scaler = load_vocoder(path, device) else: vocoder, vocoder_in_scaler = None, None out_scaler = PyTorchStandardScaler( torch.from_numpy(out_scaler.mean_), torch.from_numpy(out_scaler.scale_) ).to(device) use_mlflow = config.mlflow.enabled if use_mlflow: with mlflow.start_run() as run: # NOTE: modify out_dir when running with mlflow config.train.out_dir = f"{config.train.out_dir}/{run.info.run_id}" save_configs(config) log_params_from_omegaconf_dict(config) last_dev_loss = train_loop( config, logger, device, netG, optG, schedulerG, netD, optD, schedulerD, grad_scaler, data_loaders, samplers, writer, out_scaler, use_mlflow, vocoder, vocoder_in_scaler, ) else: save_configs(config) last_dev_loss = train_loop( config, logger, device, netG, optG, schedulerG, netD, optD, schedulerD, grad_scaler, data_loaders, samplers, writer, out_scaler, use_mlflow, vocoder, vocoder_in_scaler, ) return last_dev_loss def entry(): my_app() if __name__ == "__main__": my_app()
17,689
31.820037
103
py
nnsvs
nnsvs-master/nnsvs/bin/generate.py
# coding: utf-8 import os from os.path import basename, join import hydra import joblib import numpy as np import torch from hydra.utils import to_absolute_path from nnmnkwii.datasets import FileSourceDataset from nnsvs.base import PredictionType from nnsvs.logger import getLogger from nnsvs.multistream import get_windows, multi_stream_mlpg from nnsvs.train_util import NpyFileSource from omegaconf import DictConfig, OmegaConf from tqdm import tqdm logger = None use_cuda = torch.cuda.is_available() @hydra.main(config_path="conf/generate", config_name="config") def my_app(config: DictConfig) -> None: global logger logger = getLogger(config.verbose) logger.info(OmegaConf.to_yaml(config)) device = torch.device("cuda" if use_cuda else "cpu") in_dir = to_absolute_path(config.in_dir) out_dir = to_absolute_path(config.out_dir) os.makedirs(out_dir, exist_ok=True) model_config = OmegaConf.load(to_absolute_path(config.model.model_yaml)) model = hydra.utils.instantiate(model_config.netG).to(device) checkpoint = torch.load( to_absolute_path(config.model.checkpoint), map_location=lambda storage, loc: storage, ) model.load_state_dict(checkpoint["state_dict"]) model.eval() scaler = joblib.load(to_absolute_path(config.out_scaler_path)) in_feats = FileSourceDataset(NpyFileSource(in_dir, logger)) with torch.no_grad(): for idx in tqdm(range(len(in_feats))): feats = torch.from_numpy(in_feats[idx]).unsqueeze(0).to(device) if model.prediction_type() == PredictionType.PROBABILISTIC: max_mu, max_sigma = model.inference(feats, [feats.shape[1]]) if np.any(model_config.has_dynamic_features): # Apply denormalization # (B, T, D_out) -> (T, D_out) max_sigma_sq = ( max_sigma.squeeze(0).cpu().data.numpy() ** 2 * scaler.var_ ) max_mu = scaler.inverse_transform( max_mu.squeeze(0).cpu().data.numpy() ) # Apply MLPG # (T, D_out) -> (T, static_dim) out = multi_stream_mlpg( max_mu, max_sigma_sq, get_windows(model_config.num_windows), model_config.stream_sizes, model_config.has_dynamic_features, ) else: # (T, D_out) out = max_mu.squeeze(0).cpu().data.numpy() out = scaler.inverse_transform(out) else: out = ( model.inference(feats, [feats.shape[1]]) .squeeze(0) .cpu() .data.numpy() ) out = scaler.inverse_transform(out) # Apply MLPG if necessary if np.any(model_config.has_dynamic_features): out = multi_stream_mlpg( out, scaler.var_, get_windows(model_config.num_windows), model_config.stream_sizes, model_config.has_dynamic_features, ) name = basename(in_feats.collected_files[idx][0]) out_path = join(out_dir, name) np.save(out_path, out, allow_pickle=False) def entry(): my_app() if __name__ == "__main__": my_app()
3,601
32.045872
82
py
nnsvs
nnsvs-master/nnsvs/bin/train_acoustic.py
from functools import partial from pathlib import Path import hydra import mlflow import torch import torch.distributed as dist from hydra.utils import to_absolute_path from nnsvs.base import PredictionType from nnsvs.mdn import mdn_get_most_probable_sigma_and_mu, mdn_loss from nnsvs.multistream import split_streams from nnsvs.svs import load_vocoder from nnsvs.train_util import ( check_resf0_config, collate_fn_default, collate_fn_random_segments, compute_batch_pitch_regularization_weight, compute_distortions, eval_model, get_stream_weight, log_params_from_omegaconf_dict, save_checkpoint, save_configs, setup, ) from nnsvs.util import PyTorchStandardScaler, make_non_pad_mask, make_pad_mask from omegaconf import DictConfig from torch import nn from torch.cuda.amp import autocast from torch.nn.parallel import DistributedDataParallel as DDP def train_step( logger, model, model_config, optim_config, optimizer, grad_scaler, train, in_feats, out_feats, lengths, out_scaler, feats_criterion="mse", pitch_reg_dyn_ws=1.0, pitch_reg_weight=1.0, stream_wise_loss=False, stream_weights=None, ): model.train() if train else model.eval() optimizer.zero_grad() log_metrics = {} if feats_criterion in ["l2", "mse"]: criterion = nn.MSELoss(reduction="none") elif feats_criterion in ["l1", "mae"]: criterion = nn.L1Loss(reduction="none") else: raise RuntimeError("not supported criterion") prediction_type = ( model.module.prediction_type() if isinstance(model, nn.DataParallel) or isinstance(model, DDP) else model.prediction_type() ) # Run forward with autocast(enabled=grad_scaler is not None): outs = model(in_feats, lengths, out_feats) if model.has_residual_lf0_prediction(): pred_out_feats, lf0_residual = outs else: pred_out_feats, lf0_residual = outs, None # Mask (B, T, 1) with torch.no_grad(): mask = make_non_pad_mask(lengths).unsqueeze(-1).to(in_feats.device) # Compute loss if prediction_type == PredictionType.MULTISTREAM_HYBRID: loss_feats = 0 streams = split_streams(out_feats, model_config.stream_sizes) assert len(streams) == len(pred_out_feats) if stream_wise_loss: weights = get_stream_weight(stream_weights, model_config.stream_sizes).to( in_feats.device ) else: weights = [None] * len(streams) N = 0 for pred_stream, stream, sw in zip(pred_out_feats, streams, weights): # DDPM if isinstance(pred_stream, tuple) and len(pred_stream) == 2: noise, x_recon = pred_stream loss_feats_ = criterion( noise.masked_select(mask), x_recon.masked_select(mask) ) if stream_wise_loss: loss_feats += sw * loss_feats_.mean() else: loss_feats += loss_feats_.sum() N += len(loss_feats_.view(-1)) # MDN elif isinstance(pred_stream, tuple) and len(pred_stream) == 3: pi, sigma, mu = pred_stream # (B, max(T)) or (B, max(T), D_out) mask_ = mask if len(pi.shape) == 4 else mask.squeeze(-1) # Compute loss and apply mask with autocast(enabled=grad_scaler is not None): loss_feats_ = mdn_loss( pi, sigma, mu, stream, reduce=False ).masked_select(mask_) if stream_wise_loss: loss_feats += sw * loss_feats_.mean() else: loss_feats += loss_feats_.sum() N += len(loss_feats_.view(-1)) else: # non-MDN with autocast(enabled=grad_scaler is not None): loss_feats_ = criterion( pred_stream.masked_select(mask), stream.masked_select(mask) ) if stream_wise_loss: loss_feats += sw * loss_feats_.mean() else: loss_feats += loss_feats_.sum() N += len(loss_feats_.view(-1)) # NOTE: Mean over batch, time and feature axis if not stream_wise_loss: loss_feats /= N elif prediction_type == PredictionType.PROBABILISTIC: pi, sigma, mu = pred_out_feats # (B, max(T)) or (B, max(T), D_out) mask_ = mask if len(pi.shape) == 4 else mask.squeeze(-1) # Compute loss and apply mask with autocast(enabled=grad_scaler is not None): loss_feats = mdn_loss(pi, sigma, mu, out_feats, reduce=False) loss_feats = loss_feats.masked_select(mask_).mean() else: with autocast(enabled=grad_scaler is not None): if not isinstance(pred_out_feats, list): # NOTE: treat as multiple predictions pred_out_feats = [pred_out_feats] loss_feats = 0 for pred_out_feats_ in pred_out_feats: if stream_wise_loss: weights = get_stream_weight( stream_weights, model_config.stream_sizes ).to(in_feats.device) streams = split_streams(out_feats, model_config.stream_sizes) pred_streams = split_streams( pred_out_feats_, model_config.stream_sizes ) else: streams = [out_feats] pred_streams = [pred_out_feats_] weights = [1.0] for pred_stream, stream, sw in zip(pred_streams, streams, weights): with autocast(enabled=grad_scaler is not None): loss_feats += ( sw * criterion( pred_stream.masked_select(mask), stream.masked_select(mask), ).mean() ) # Pitch regularization # NOTE: l1 loss seems to be better than mse loss in my experiments # we could use l2 loss as suggested in the sinsy's paper if pitch_reg_weight > 0.0 and lf0_residual is not None: with autocast(enabled=grad_scaler is not None): if isinstance(lf0_residual, list): loss_pitch = 0 for lf0_residual_ in lf0_residual: loss_pitch += ( (pitch_reg_dyn_ws * lf0_residual_.abs()) .masked_select(mask) .mean() ) else: loss_pitch = ( (pitch_reg_dyn_ws * lf0_residual.abs()).masked_select(mask).mean() ) else: loss_pitch = torch.tensor(0.0).to(in_feats.device) loss = loss_feats + pitch_reg_weight * loss_pitch if not train: with torch.no_grad(): if prediction_type == PredictionType.MULTISTREAM_HYBRID: if len(pred_out_feats) == 4: pred_mgc, pred_lf0, pred_vuv, pred_bap = pred_out_feats if isinstance(pred_lf0, tuple) and len(pred_lf0) == 3: pred_lf0 = mdn_get_most_probable_sigma_and_mu(*pred_lf0)[1] elif isinstance(pred_lf0, tuple) and len(pred_lf0) == 2: # Diffusion case: noise pred_lf0 = pred_lf0[1] if isinstance(pred_mgc, tuple) and len(pred_mgc) == 3: pred_mgc = mdn_get_most_probable_sigma_and_mu(*pred_mgc)[1] elif isinstance(pred_mgc, tuple) and len(pred_mgc) == 2: # Diffusion case: noise pred_mgc = pred_mgc[1] if isinstance(pred_bap, tuple) and len(pred_bap) == 3: pred_bap = mdn_get_most_probable_sigma_and_mu(*pred_bap)[1] elif isinstance(pred_bap, tuple) and len(pred_bap) == 2: # Diffusion case: noise pred_bap = pred_bap[1] pred_out_feats_ = torch.cat( [pred_mgc, pred_lf0, pred_vuv, pred_bap], dim=-1 ) elif len(pred_out_feats) == 3: pred_mel, pred_lf0, pred_vuv = pred_out_feats if isinstance(pred_lf0, tuple) and len(pred_lf0) == 3: pred_lf0 = mdn_get_most_probable_sigma_and_mu(*pred_lf0)[1] elif isinstance(pred_lf0, tuple) and len(pred_lf0) == 2: # Diffusion case: noise pred_lf0 = pred_lf0[1] if isinstance(pred_mel, tuple) and len(pred_mel) == 3: pred_mel = mdn_get_most_probable_sigma_and_mu(*pred_mel)[1] elif isinstance(pred_mel, tuple) and len(pred_mel) == 2: # Diffusion case: noise pred_mel = pred_mel[1] pred_out_feats_ = torch.cat([pred_mel, pred_lf0, pred_vuv], dim=-1) else: raise RuntimeError("not supported") elif prediction_type == PredictionType.PROBABILISTIC: pred_out_feats_ = mdn_get_most_probable_sigma_and_mu(pi, sigma, mu)[1] else: if isinstance(pred_out_feats, list): pred_out_feats_ = pred_out_feats[-1] else: pred_out_feats_ = pred_out_feats distortions = compute_distortions( pred_out_feats_, out_feats, lengths, out_scaler, model_config ) log_metrics.update(distortions) if train: if grad_scaler is not None: grad_scaler.scale(loss).backward() grad_scaler.unscale_(optimizer) grad_norm = torch.nn.utils.clip_grad_norm_( model.parameters(), optim_config.clip_norm ) if not torch.isfinite(grad_norm): logger.info("grad norm is NaN. Skip updating") else: log_metrics["GradNorm"] = grad_norm grad_scaler.step(optimizer) grad_scaler.update() else: loss.backward() grad_norm = torch.nn.utils.clip_grad_norm_( model.parameters(), optim_config.clip_norm ) if not torch.isfinite(grad_norm): logger.info("grad norm is NaN. Skip updating") else: log_metrics["GradNorm"] = grad_norm optimizer.step() log_metrics.update( { "Loss": loss.item(), "Loss_Feats": loss_feats.item(), "Loss_Pitch": loss_pitch.item(), } ) return loss, log_metrics def train_loop( config, logger, device, model, optimizer, lr_scheduler, grad_scaler, data_loaders, samplers, writer, in_scaler, out_scaler, use_mlflow, vocoder=None, vocoder_in_scaler=None, vocoder_config=None, ): out_dir = Path(to_absolute_path(config.train.out_dir)) best_dev_loss = torch.finfo(torch.float32).max last_dev_loss = torch.finfo(torch.float32).max in_lf0_idx = config.data.in_lf0_idx in_rest_idx = config.data.in_rest_idx if in_lf0_idx is None or in_rest_idx is None: raise ValueError("in_lf0_idx and in_rest_idx must be specified") pitch_reg_weight = config.train.pitch_reg_weight if "sample_rate" not in config.data: logger.warning( "sample_rate is not found in the data config. Fallback to 48000." ) sr = 48000 else: sr = config.data.sample_rate if "feats_criterion" not in config.train: logger.warning( "feats_criterion is not found in the train config. Fallback to MSE." ) feats_criterion = "mse" else: feats_criterion = config.train.feats_criterion if dist.is_initialized() and dist.get_rank() != 0: def tqdm(x, **kwargs): return x else: from tqdm import tqdm train_iter = 1 for epoch in tqdm(range(1, config.train.nepochs + 1)): for phase in data_loaders.keys(): train = phase.startswith("train") # https://pytorch.org/docs/stable/data.html#torch.utils.data.distributed.DistributedSampler if dist.is_initialized() and train and samplers[phase] is not None: samplers[phase].set_epoch(epoch) running_loss = 0 running_metrics = {} evaluated = False for in_feats, out_feats, lengths in tqdm( data_loaders[phase], desc=f"{phase} iter", leave=False ): # NOTE: This is needed for pytorch's PackedSequence lengths, indices = torch.sort(lengths, dim=0, descending=True) in_feats, out_feats = ( in_feats[indices].to(device), out_feats[indices].to(device), ) # Compute denormalized log-F0 in the musical scores with torch.no_grad(): lf0_score_denorm = ( in_feats[:, :, in_lf0_idx] - in_scaler.min_[in_lf0_idx] ) / in_scaler.scale_[in_lf0_idx] # Fill zeros for rest and padded frames lf0_score_denorm *= (in_feats[:, :, in_rest_idx] <= 0).float() lf0_score_denorm[make_pad_mask(lengths)] = 0 # Compute time-variant pitch regularization weight vector # NOTE: the current impl. is very slow if pitch_reg_weight > 0.0: pitch_reg_dyn_ws = compute_batch_pitch_regularization_weight( lf0_score_denorm, decay_size=config.train.pitch_reg_decay_size, ) else: pitch_reg_dyn_ws = 1.0 if (not train) and (not evaluated): eval_model( phase, epoch, model, in_feats, out_feats, lengths, config.model, out_scaler, writer, sr=sr, lf0_score_denorm=lf0_score_denorm, use_world_codec=config.data.use_world_codec, vocoder=vocoder, vocoder_in_scaler=vocoder_in_scaler, vocoder_config=vocoder_config, max_num_eval_utts=config.train.max_num_eval_utts, ) evaluated = True loss, log_metrics = train_step( logger=logger, model=model, model_config=config.model, optim_config=config.train.optim, optimizer=optimizer, grad_scaler=grad_scaler, train=train, in_feats=in_feats, out_feats=out_feats, lengths=lengths, out_scaler=out_scaler, feats_criterion=feats_criterion, stream_wise_loss=config.train.stream_wise_loss, stream_weights=config.model.stream_weights, pitch_reg_dyn_ws=pitch_reg_dyn_ws, pitch_reg_weight=pitch_reg_weight, ) if train: if writer is not None: for key, val in log_metrics.items(): writer.add_scalar(f"{key}_Step/{phase}", val, train_iter) train_iter += 1 running_loss += loss.item() for k, v in log_metrics.items(): try: running_metrics[k] += float(v) except KeyError: running_metrics[k] = float(v) ave_loss = running_loss / len(data_loaders[phase]) logger.info("[%s] [Epoch %s]: loss %s", phase, epoch, ave_loss) if writer is not None: writer.add_scalar(f"Loss_Epoch/{phase}", ave_loss, epoch) if use_mlflow: mlflow.log_metric(f"{phase}_loss", ave_loss, step=epoch) for k, v in running_metrics.items(): ave_v = v / len(data_loaders[phase]) if writer is not None: writer.add_scalar(f"{k}_Epoch/{phase}", ave_v, epoch) if use_mlflow: mlflow.log_metric(f"{phase}_{k}", ave_v, step=epoch) if not train: last_dev_loss = ave_loss if not train and ave_loss < best_dev_loss: best_dev_loss = ave_loss save_checkpoint( logger, out_dir, model, optimizer, lr_scheduler, epoch, is_best=True ) lr_scheduler.step() if epoch % config.train.checkpoint_epoch_interval == 0: save_checkpoint( logger, out_dir, model, optimizer, lr_scheduler, epoch, is_best=False ) save_checkpoint( logger, out_dir, model, optimizer, lr_scheduler, config.train.nepochs ) logger.info("The best loss was %s", best_dev_loss) if use_mlflow: mlflow.log_metric("best_dev_loss", best_dev_loss, step=epoch) mlflow.log_artifacts(out_dir) return last_dev_loss @hydra.main(config_path="conf/train_acoustic", config_name="config") def my_app(config: DictConfig) -> None: if "max_time_frames" in config.data and config.data.max_time_frames > 0: collate_fn = partial( collate_fn_random_segments, max_time_frames=config.data.max_time_frames ) else: if "reduction_factor" in config.model.netG: collate_fn = partial( collate_fn_default, reduction_factor=config.model.netG.reduction_factor, ) else: collate_fn = collate_fn_default if config.train.use_ddp: dist.init_process_group("nccl") rank = dist.get_rank() device_id = rank % torch.cuda.device_count() torch.cuda.set_device(device_id) device = torch.device("cuda" if torch.cuda.is_available() else "cpu") ( model, optimizer, lr_scheduler, grad_scaler, data_loaders, samplers, writer, logger, in_scaler, out_scaler, ) = setup(config, device, collate_fn) path = config.train.pretrained_vocoder_checkpoint if path is not None and len(path) > 0: logger.info(f"Loading pretrained vocoder checkpoint from {path}") vocoder, vocoder_in_scaler, vocoder_config = load_vocoder( path, device, config.model ) else: vocoder, vocoder_in_scaler, vocoder_config = None, None, None check_resf0_config(logger, model, config, in_scaler, out_scaler) out_scaler = PyTorchStandardScaler( torch.from_numpy(out_scaler.mean_), torch.from_numpy(out_scaler.scale_) ).to(device) use_mlflow = config.mlflow.enabled if use_mlflow: with mlflow.start_run() as run: # NOTE: modify out_dir when running with mlflow config.train.out_dir = f"{config.train.out_dir}/{run.info.run_id}" save_configs(config) log_params_from_omegaconf_dict(config) last_dev_loss = train_loop( config=config, logger=logger, device=device, model=model, optimizer=optimizer, lr_scheduler=lr_scheduler, grad_scaler=grad_scaler, data_loaders=data_loaders, samplers=samplers, writer=writer, in_scaler=in_scaler, out_scaler=out_scaler, use_mlflow=use_mlflow, vocoder=vocoder, vocoder_in_scaler=vocoder_in_scaler, vocoder_config=vocoder_config, ) else: save_configs(config) last_dev_loss = train_loop( config=config, logger=logger, device=device, model=model, optimizer=optimizer, lr_scheduler=lr_scheduler, grad_scaler=grad_scaler, data_loaders=data_loaders, samplers=samplers, writer=writer, in_scaler=in_scaler, out_scaler=out_scaler, use_mlflow=use_mlflow, vocoder=vocoder, vocoder_in_scaler=vocoder_in_scaler, vocoder_config=vocoder_config, ) return last_dev_loss def entry(): my_app() if __name__ == "__main__": my_app()
21,429
36.929204
103
py
nnsvs
nnsvs-master/nnsvs/bin/synthesis.py
import os from os.path import join import hydra import joblib import numpy as np import torch from hydra.utils import to_absolute_path from nnmnkwii.io import hts from nnsvs.gen import ( postprocess_acoustic, postprocess_waveform, predict_acoustic, predict_timing, predict_waveform, ) from nnsvs.logger import getLogger from nnsvs.util import extract_static_scaler, init_seed, load_utt_list, load_vocoder from omegaconf import DictConfig, OmegaConf from scipy.io import wavfile from tqdm.auto import tqdm @hydra.main(config_path="conf/synthesis", config_name="config") def my_app(config: DictConfig) -> None: global logger logger = getLogger(config.verbose) logger.info(OmegaConf.to_yaml(config)) if not torch.cuda.is_available(): device = torch.device("cpu") else: device = torch.device(config.device) # timelag timelag_config = OmegaConf.load(to_absolute_path(config.timelag.model_yaml)) timelag_model = hydra.utils.instantiate(timelag_config.netG).to(device) checkpoint = torch.load( to_absolute_path(config.timelag.checkpoint), map_location=lambda storage, loc: storage, ) timelag_model.load_state_dict(checkpoint["state_dict"]) timelag_in_scaler = joblib.load(to_absolute_path(config.timelag.in_scaler_path)) timelag_out_scaler = joblib.load(to_absolute_path(config.timelag.out_scaler_path)) timelag_model.eval() # duration duration_config = OmegaConf.load(to_absolute_path(config.duration.model_yaml)) duration_model = hydra.utils.instantiate(duration_config.netG).to(device) checkpoint = torch.load( to_absolute_path(config.duration.checkpoint), map_location=lambda storage, loc: storage, ) duration_model.load_state_dict(checkpoint["state_dict"]) duration_in_scaler = joblib.load(to_absolute_path(config.duration.in_scaler_path)) duration_out_scaler = joblib.load(to_absolute_path(config.duration.out_scaler_path)) duration_model.eval() # acoustic model acoustic_config = OmegaConf.load(to_absolute_path(config.acoustic.model_yaml)) acoustic_model = hydra.utils.instantiate(acoustic_config.netG).to(device) checkpoint = torch.load( to_absolute_path(config.acoustic.checkpoint), map_location=lambda storage, loc: storage, ) acoustic_model.load_state_dict(checkpoint["state_dict"]) acoustic_in_scaler = joblib.load(to_absolute_path(config.acoustic.in_scaler_path)) acoustic_out_scaler = joblib.load(to_absolute_path(config.acoustic.out_scaler_path)) acoustic_model.eval() # NOTE: this is used for GV post-filtering acoustic_out_static_scaler = extract_static_scaler( acoustic_out_scaler, acoustic_config ) # Vocoder if config.vocoder.checkpoint is not None and len(config.vocoder.checkpoint) > 0: vocoder, vocoder_in_scaler, vocoder_config = load_vocoder( to_absolute_path(config.vocoder.checkpoint), device, acoustic_config, ) else: vocoder, vocoder_in_scaler, vocoder_config = None, None, None if config.synthesis.vocoder_type != "world": logger.warning("Vocoder checkpoint is not specified") logger.info(f"Use world instead of {config.synthesis.vocoder_type}.") config.synthesis.vocoder_type = "world" # Run synthesis for each utt. binary_dict, numeric_dict = hts.load_question_set( to_absolute_path(config.synthesis.qst) ) in_dir = to_absolute_path(config.in_dir) out_dir = to_absolute_path(config.out_dir) os.makedirs(out_dir, exist_ok=True) utt_ids = load_utt_list(to_absolute_path(config.utt_list)) logger.info("Processes %s utterances...", len(utt_ids)) for utt_id in tqdm(utt_ids): labels = hts.load(join(in_dir, f"{utt_id}.lab")) hts_frame_shift = int(config.synthesis.frame_period * 1e4) labels.frame_shift = hts_frame_shift init_seed(1234) if config.synthesis.ground_truth_duration: duration_modified_labels = labels else: duration_modified_labels = predict_timing( device=device, labels=labels, binary_dict=binary_dict, numeric_dict=numeric_dict, timelag_model=timelag_model, timelag_config=timelag_config, timelag_in_scaler=timelag_in_scaler, timelag_out_scaler=timelag_out_scaler, duration_model=duration_model, duration_config=duration_config, duration_in_scaler=duration_in_scaler, duration_out_scaler=duration_out_scaler, log_f0_conditioning=config.synthesis.log_f0_conditioning, allowed_range=config.timelag.allowed_range, allowed_range_rest=config.timelag.allowed_range_rest, force_clip_input_features=config.timelag.force_clip_input_features, frame_period=config.synthesis.frame_period, ) # Predict acoustic features acoustic_features = predict_acoustic( device=device, labels=duration_modified_labels, acoustic_model=acoustic_model, acoustic_config=acoustic_config, acoustic_in_scaler=acoustic_in_scaler, acoustic_out_scaler=acoustic_out_scaler, binary_dict=binary_dict, numeric_dict=numeric_dict, subphone_features=config.synthesis.subphone_features, log_f0_conditioning=config.synthesis.log_f0_conditioning, force_clip_input_features=config.acoustic.force_clip_input_features, f0_shift_in_cent=config.synthesis.pre_f0_shift_in_cent, ) # NOTE: the output of this function is tuple of features # e.g., (mgc, lf0, vuv, bap) multistream_features = postprocess_acoustic( device=device, acoustic_features=acoustic_features, duration_modified_labels=duration_modified_labels, binary_dict=binary_dict, numeric_dict=numeric_dict, acoustic_config=acoustic_config, acoustic_out_static_scaler=acoustic_out_static_scaler, postfilter_model=None, # NOTE: learned post-filter is not supported postfilter_config=None, postfilter_out_scaler=None, sample_rate=config.synthesis.sample_rate, frame_period=config.synthesis.frame_period, relative_f0=config.synthesis.relative_f0, feature_type=config.synthesis.feature_type, post_filter_type=config.synthesis.post_filter_type, trajectory_smoothing=config.synthesis.trajectory_smoothing, trajectory_smoothing_cutoff=config.synthesis.trajectory_smoothing_cutoff, trajectory_smoothing_cutoff_f0=config.synthesis.trajectory_smoothing_cutoff_f0, vuv_threshold=config.synthesis.vuv_threshold, f0_shift_in_cent=config.synthesis.post_f0_shift_in_cent, vibrato_scale=1.0, force_fix_vuv=config.synthesis.force_fix_vuv, ) # Generate waveform by vocoder wav = predict_waveform( device=device, multistream_features=multistream_features, vocoder=vocoder, vocoder_config=vocoder_config, vocoder_in_scaler=vocoder_in_scaler, sample_rate=config.synthesis.sample_rate, frame_period=config.synthesis.frame_period, use_world_codec=config.synthesis.use_world_codec, feature_type=config.synthesis.feature_type, vocoder_type=config.synthesis.vocoder_type, vuv_threshold=config.synthesis.vuv_threshold, ) wav = postprocess_waveform( wav=wav, sample_rate=config.synthesis.sample_rate, dtype=np.int16, peak_norm=False, loudness_norm=False, ) out_wav_path = join(out_dir, f"{utt_id}.wav") wavfile.write( out_wav_path, rate=config.synthesis.sample_rate, data=wav.astype(np.int16) ) def entry(): my_app() # pylint: disable=no-value-for-parameter if __name__ == "__main__": my_app() # pylint: disable=no-value-for-parameter
8,358
39.381643
91
py
nnsvs
nnsvs-master/nnsvs/bin/gen_static_features.py
import os from os.path import exists, join import hydra import joblib import numpy as np import pyworld import torch from hydra.utils import to_absolute_path from nnsvs.acoustic_models.util import pad_inference from nnsvs.base import PredictionType from nnsvs.gen import get_windows from nnsvs.logger import getLogger from nnsvs.mdn import mdn_get_most_probable_sigma_and_mu from nnsvs.multistream import ( get_static_features, get_static_stream_sizes, multi_stream_mlpg, split_streams, ) from nnsvs.postfilters import variance_scaling from nnsvs.util import StandardScaler, load_utt_list from omegaconf import DictConfig, OmegaConf from tqdm import tqdm logger = None use_cuda = torch.cuda.is_available() @torch.no_grad() def _gen_static_features( model, model_config, in_feats, gt_feats, out_scaler, gta=False ): if model.prediction_type() == PredictionType.PROBABILISTIC: if gta: raise ValueError("GTA not supported for probabilistic models for now") else: max_mu, max_sigma = model.inference(in_feats, [in_feats.shape[1]]) if np.any(model_config.has_dynamic_features): # Apply denormalization # (B, T, D_out) -> (T, D_out) max_sigma_sq = ( max_sigma.squeeze(0).cpu().data.numpy() ** 2 * out_scaler.var_ ) max_mu = out_scaler.inverse_transform(max_mu.squeeze(0).cpu().data.numpy()) # Apply MLPG # (T, D_out) -> (T, static_dim) out_feats = multi_stream_mlpg( max_mu, max_sigma_sq, get_windows(model_config.num_windows), model_config.stream_sizes, model_config.has_dynamic_features, ) else: # (T, D_out) out_feats = max_mu.squeeze(0).cpu().data.numpy() out_feats = out_scaler.inverse_transform(out_feats) else: if gta: if hasattr(model, "reduction_factor"): out_feats = pad_inference( model, in_feats, [in_feats.shape[1]], model.reduction_factor, mode="replicate", y=gt_feats, mdn=model.prediction_type() in [PredictionType.PROBABILISTIC, 3], ) if model.prediction_type() == PredictionType.PROBABILISTIC: # (sigma, mu) -> (mu) out_feats = mdn_get_most_probable_sigma_and_mu(out_feats)[1] elif model.prediction_type() == 3: # (mu, sigma) -> (mu) out_feats = out_feats[0] else: out_feats = model(in_feats, [in_feats.shape[1]], gt_feats) # out + residual case out_feats = out_feats[0] if isinstance(out_feats, tuple) else out_feats # multiple outputs case out_feats = out_feats[-1] if isinstance(out_feats, list) else out_feats out_feats = out_feats.squeeze(0).cpu().data.numpy() else: out_feats = ( model.inference(in_feats, [in_feats.shape[1]]) .squeeze(0) .cpu() .data.numpy() ) out_feats = out_scaler.inverse_transform(out_feats) # Apply MLPG if necessary if np.any(model_config.has_dynamic_features): out_feats = multi_stream_mlpg( out_feats, out_scaler.var_, get_windows(model_config.num_windows), model_config.stream_sizes, model_config.has_dynamic_features, ) return out_feats.astype(np.float32) @hydra.main(config_path="conf/gen_static_features", config_name="config") def my_app(config: DictConfig) -> None: global logger logger = getLogger(config.verbose) logger.info(OmegaConf.to_yaml(config)) device = torch.device("cuda" if use_cuda else "cpu") utt_list = to_absolute_path(config.utt_list) in_dir = to_absolute_path(config.in_dir) gt_dir = to_absolute_path(config.gt_dir) out_dir = to_absolute_path(config.out_dir) utt_ids = load_utt_list(utt_list) os.makedirs(out_dir, exist_ok=True) model_config = OmegaConf.load(to_absolute_path(config.model.model_yaml)) model = hydra.utils.instantiate(model_config.netG).to(device) checkpoint = torch.load( to_absolute_path(config.model.checkpoint), map_location=lambda storage, loc: storage, ) model.load_state_dict(checkpoint["state_dict"]) model.eval() out_scaler = joblib.load(to_absolute_path(config.out_scaler_path)) mean_ = get_static_features( out_scaler.mean_.reshape(1, 1, out_scaler.mean_.shape[-1]), model_config.num_windows, model_config.stream_sizes, model_config.has_dynamic_features, ) mean_ = np.concatenate(mean_, -1).reshape(1, -1) var_ = get_static_features( out_scaler.var_.reshape(1, 1, out_scaler.var_.shape[-1]), model_config.num_windows, model_config.stream_sizes, model_config.has_dynamic_features, ) var_ = np.concatenate(var_, -1).reshape(1, -1) scale_ = get_static_features( out_scaler.scale_.reshape(1, 1, out_scaler.scale_.shape[-1]), model_config.num_windows, model_config.stream_sizes, model_config.has_dynamic_features, ) scale_ = np.concatenate(scale_, -1).reshape(1, -1) static_scaler = StandardScaler(mean_, var_, scale_) static_stream_sizes = get_static_stream_sizes( model_config.stream_sizes, model_config.has_dynamic_features, model_config.num_windows, ) for utt_id in tqdm(utt_ids): if not exists(join(gt_dir, utt_id + "-feats.npy")): print(f"Skip {utt_id}") continue in_feats = ( torch.from_numpy(np.load(join(in_dir, utt_id + "-feats.npy"))) .unsqueeze(0) .to(device) ) gt_feats = ( torch.from_numpy(np.load(join(gt_dir, utt_id + "-feats.npy"))) .unsqueeze(0) .to(device) ) static_feats = _gen_static_features( model, model_config, in_feats, gt_feats, out_scaler, config.gta ) outs = split_streams(static_feats, static_stream_sizes) if len(outs) == 4: mgc, lf0, vuv, bap = outs mgc_end_dim = static_stream_sizes[0] bap_start_dim = sum(static_stream_sizes[:3]) bap_end_dim = sum(static_stream_sizes[:4]) if config.gv_postfilter: # mgc mgc = variance_scaling( static_scaler.var_.reshape(-1)[:mgc_end_dim], mgc, offset=config.mgc_offset, ) # bap bap = variance_scaling( static_scaler.var_.reshape(-1)[bap_start_dim:bap_end_dim], bap, offset=config.bap_offset, ) if config.mgc2sp: # 24k: 1024 # 48: 2048 fft_size = pyworld.get_cheaptrick_fft_size(config.sample_rate) sp = np.log( pyworld.decode_spectral_envelope( mgc.astype(np.float64), config.sample_rate, fft_size ).astype(np.float32) ) static_feats = np.concatenate([sp, lf0, vuv, bap], axis=-1) else: static_feats = np.concatenate([mgc, lf0, vuv, bap], axis=-1) elif len(outs) == 3: mel, lf0, vuv = outs # NOTE: may add postfilter here static_feats = np.concatenate([mel, lf0, vuv], axis=-1) else: raise ValueError("Invalid number of streams") gt_feats = gt_feats.squeeze(0) if len(static_feats) != len(gt_feats): print(static_feats.shape, gt_feats.shape) raise RuntimeError(f"Length mismatch in {utt_id}") if config.normalize: assert not config.mgc2sp, "need to compute normalization stats" static_feats = static_scaler.transform(static_feats) out_path = join(out_dir, f"{utt_id}-feats.npy") np.save(out_path, static_feats.astype(np.float32), allow_pickle=False) def entry(): my_app() if __name__ == "__main__": my_app()
8,527
34.385892
87
py
nnsvs
nnsvs-master/nnsvs/bin/train.py
from functools import partial from pathlib import Path import hydra import mlflow import numpy as np import torch import torch.distributed as dist from hydra.utils import to_absolute_path from nnmnkwii import metrics from nnsvs.base import PredictionType from nnsvs.mdn import mdn_get_most_probable_sigma_and_mu, mdn_loss from nnsvs.multistream import split_streams from nnsvs.train_util import ( collate_fn_default, collate_fn_random_segments, get_stream_weight, log_params_from_omegaconf_dict, save_checkpoint, save_configs, setup, ) from nnsvs.util import PyTorchStandardScaler, make_non_pad_mask from omegaconf import DictConfig from torch import nn from torch.cuda.amp import autocast @torch.no_grad() def compute_distortions(pred_out_feats, out_feats, lengths, out_scaler): assert pred_out_feats.shape == out_feats.shape out_feats = out_scaler.inverse_transform(out_feats) pred_out_feats = out_scaler.inverse_transform(pred_out_feats) dist = {} try: dist["ObjEval_RMSE"] = np.sqrt( metrics.mean_squared_error(out_feats, pred_out_feats, lengths=lengths) ) except ZeroDivisionError: pass return dist def train_step( model, optimizer, grad_scaler, train, in_feats, out_feats, lengths, out_scaler, feats_criterion="mse", stream_wise_loss=False, stream_weights=None, stream_sizes=None, ): model.train() if train else model.eval() optimizer.zero_grad() if feats_criterion in ["l2", "mse"]: criterion = nn.MSELoss(reduction="none") elif feats_criterion in ["l1", "mae"]: criterion = nn.L1Loss(reduction="none") else: raise RuntimeError("not supported criterion") prediction_type = ( model.module.prediction_type() if isinstance(model, nn.DataParallel) else model.prediction_type() ) # Apply preprocess if required (e.g., FIR filter for shallow AR) # defaults to no-op if isinstance(model, nn.DataParallel): out_feats = model.module.preprocess_target(out_feats) else: out_feats = model.preprocess_target(out_feats) # Run forward with autocast(enabled=grad_scaler is not None): pred_out_feats = model(in_feats, lengths) # Mask (B, T, 1) mask = make_non_pad_mask(lengths).unsqueeze(-1).to(in_feats.device) # Compute loss if prediction_type == PredictionType.PROBABILISTIC: pi, sigma, mu = pred_out_feats # (B, max(T)) or (B, max(T), D_out) mask_ = mask if len(pi.shape) == 4 else mask.squeeze(-1) # Compute loss and apply mask with autocast(enabled=grad_scaler is not None): loss = mdn_loss(pi, sigma, mu, out_feats, reduce=False) loss = loss.masked_select(mask_).mean() else: if stream_wise_loss: w = get_stream_weight(stream_weights, stream_sizes).to(in_feats.device) streams = split_streams(out_feats, stream_sizes) pred_streams = split_streams(pred_out_feats, stream_sizes) loss = 0 for pred_stream, stream, sw in zip(pred_streams, streams, w): with autocast(enabled=grad_scaler is not None): loss += ( sw * criterion( pred_stream.masked_select(mask), stream.masked_select(mask) ).mean() ) else: with autocast(enabled=grad_scaler is not None): loss = criterion( pred_out_feats.masked_select(mask), out_feats.masked_select(mask) ).mean() if prediction_type == PredictionType.PROBABILISTIC: with torch.no_grad(): pred_out_feats_ = mdn_get_most_probable_sigma_and_mu(pi, sigma, mu)[1] else: pred_out_feats_ = pred_out_feats distortions = compute_distortions(pred_out_feats_, out_feats, lengths, out_scaler) if train: if grad_scaler is not None: grad_scaler.scale(loss).backward() grad_scaler.step(optimizer) grad_scaler.update() else: loss.backward() optimizer.step() return loss, distortions def train_loop( config, logger, device, model, optimizer, lr_scheduler, grad_scaler, data_loaders, samplers, writer, out_scaler, use_mlflow, ): out_dir = Path(to_absolute_path(config.train.out_dir)) best_dev_loss = torch.finfo(torch.float32).max last_dev_loss = torch.finfo(torch.float32).max if "feats_criterion" not in config.train: logger.warning( "feats_criterion is not found in the train config. Fallback to MSE." ) feats_criterion = "mse" else: feats_criterion = config.train.feats_criterion if dist.is_initialized() and dist.get_rank() != 0: def tqdm(x, **kwargs): return x else: from tqdm import tqdm train_iter = 1 for epoch in tqdm(range(1, config.train.nepochs + 1)): for phase in data_loaders.keys(): train = phase.startswith("train") # https://pytorch.org/docs/stable/data.html#torch.utils.data.distributed.DistributedSampler if dist.is_initialized() and train and samplers[phase] is not None: samplers[phase].set_epoch(epoch) running_loss = 0 running_metrics = {} for in_feats, out_feats, lengths in tqdm( data_loaders[phase], desc=f"{phase} iter", leave=False ): # NOTE: This is needed for pytorch's PackedSequence lengths, indices = torch.sort(lengths, dim=0, descending=True) in_feats, out_feats = ( in_feats[indices].to(device), out_feats[indices].to(device), ) loss, log_metrics = train_step( model=model, optimizer=optimizer, grad_scaler=grad_scaler, train=train, in_feats=in_feats, out_feats=out_feats, lengths=lengths, out_scaler=out_scaler, feats_criterion=feats_criterion, stream_wise_loss=config.train.stream_wise_loss, stream_weights=config.model.stream_weights, stream_sizes=config.model.stream_sizes, ) if train: if writer is not None: for key, val in log_metrics.items(): writer.add_scalar(f"{key}_Step/{phase}", val, train_iter) train_iter += 1 running_loss += loss.item() for k, v in log_metrics.items(): try: running_metrics[k] += float(v) except KeyError: running_metrics[k] = float(v) ave_loss = running_loss / len(data_loaders[phase]) if writer is not None: writer.add_scalar(f"Loss/{phase}", ave_loss, epoch) if use_mlflow: mlflow.log_metric(f"{phase}_loss", ave_loss, step=epoch) ave_loss = running_loss / len(data_loaders[phase]) logger.info("[%s] [Epoch %s]: loss %s", phase, epoch, ave_loss) if writer is not None: writer.add_scalar(f"Loss_Epoch/{phase}", ave_loss, epoch) if use_mlflow: mlflow.log_metric(f"{phase}_loss", ave_loss, step=epoch) for k, v in running_metrics.items(): ave_v = v / len(data_loaders[phase]) if writer is not None: writer.add_scalar(f"{k}_Epoch/{phase}", ave_v, epoch) if use_mlflow: mlflow.log_metric(f"{phase}_{k}", ave_v, step=epoch) if not train: last_dev_loss = ave_loss if not train and ave_loss < best_dev_loss: best_dev_loss = ave_loss save_checkpoint( logger, out_dir, model, optimizer, lr_scheduler, epoch, is_best=True ) lr_scheduler.step() if epoch % config.train.checkpoint_epoch_interval == 0: save_checkpoint( logger, out_dir, model, optimizer, lr_scheduler, epoch, is_best=False ) save_checkpoint( logger, out_dir, model, optimizer, lr_scheduler, config.train.nepochs ) logger.info("The best loss was %s", best_dev_loss) if use_mlflow: mlflow.log_metric("best_dev_loss", best_dev_loss, step=epoch) mlflow.log_artifacts(out_dir) return last_dev_loss @hydra.main(config_path="conf/train", config_name="config") def my_app(config: DictConfig) -> None: if "max_time_frames" in config.data and config.data.max_time_frames > 0: collate_fn = partial( collate_fn_random_segments, max_time_frames=config.data.max_time_frames ) else: collate_fn = collate_fn_default if config.train.use_ddp: dist.init_process_group("nccl") rank = dist.get_rank() device_id = rank % torch.cuda.device_count() torch.cuda.set_device(device_id) device = torch.device("cuda" if torch.cuda.is_available() else "cpu") ( model, optimizer, lr_scheduler, grad_scaler, data_loaders, samplers, writer, logger, _, out_scaler, ) = setup(config, device, collate_fn) out_scaler = PyTorchStandardScaler( torch.from_numpy(out_scaler.mean_), torch.from_numpy(out_scaler.scale_) ).to(device) use_mlflow = config.mlflow.enabled if use_mlflow: with mlflow.start_run() as run: # NOTE: modify out_dir when running with mlflow config.train.out_dir = f"{config.train.out_dir}/{run.info.run_id}" save_configs(config) log_params_from_omegaconf_dict(config) last_dev_loss = train_loop( config=config, logger=logger, device=device, model=model, optimizer=optimizer, lr_scheduler=lr_scheduler, grad_scaler=grad_scaler, data_loaders=data_loaders, samplers=samplers, writer=writer, out_scaler=out_scaler, use_mlflow=use_mlflow, ) else: save_configs(config) last_dev_loss = train_loop( config=config, logger=logger, device=device, model=model, optimizer=optimizer, lr_scheduler=lr_scheduler, grad_scaler=grad_scaler, data_loaders=data_loaders, samplers=samplers, writer=writer, out_scaler=out_scaler, use_mlflow=use_mlflow, ) return last_dev_loss def entry(): my_app() if __name__ == "__main__": my_app()
11,217
31.705539
103
py
nnsvs
nnsvs-master/nnsvs/bin/anasyn.py
import os from os.path import join import hydra import numpy as np import pysptk import pyworld import torch from hydra.utils import to_absolute_path from nnsvs.dsp import bandpass_filter from nnsvs.gen import gen_world_params from nnsvs.logger import getLogger from nnsvs.multistream import get_static_stream_sizes, split_streams from nnsvs.svs import load_vocoder from nnsvs.util import init_seed, load_utt_list from omegaconf import DictConfig, OmegaConf from scipy.io import wavfile from tqdm.auto import tqdm @torch.no_grad() def anasyn( device, acoustic_features, acoustic_config, vocoder=None, vocoder_config=None, vocoder_in_scaler=None, sample_rate=48000, frame_period=5, vuv_threshold=0.5, use_world_codec=True, feature_type="world", vocoder_type="world", ): static_stream_sizes = get_static_stream_sizes( acoustic_config.stream_sizes, acoustic_config.has_dynamic_features, acoustic_config.num_windows, ) # Split multi-stream features streams = split_streams(acoustic_features, static_stream_sizes) # Generate WORLD parameters if feature_type == "world": assert len(streams) == 4 mgc, lf0, vuv, bap = streams elif feature_type == "melf0": mel, lf0, vuv = split_streams(acoustic_features, [80, 1, 1]) else: raise ValueError(f"Unknown feature type: {feature_type}") # Waveform generation by (1) WORLD or (2) neural vocoder if vocoder_type == "world": f0, spectrogram, aperiodicity = gen_world_params( mgc, lf0, vuv, bap, sample_rate, vuv_threshold=vuv_threshold, use_world_codec=use_world_codec, ) wav = pyworld.synthesize( f0, spectrogram, aperiodicity, sample_rate, frame_period, ) elif vocoder_type == "pwg": # NOTE: So far vocoder models are trained on binary V/UV features vuv = (vuv > vuv_threshold).astype(np.float32) if feature_type == "world": voc_inp = ( torch.from_numpy( vocoder_in_scaler.transform( np.concatenate([mgc, lf0, vuv, bap], axis=-1) ) ) .float() .to(device) ) elif feature_type == "melf0": voc_inp = ( torch.from_numpy( vocoder_in_scaler.transform( np.concatenate([mel, lf0, vuv], axis=-1) ) ) .float() .to(device) ) wav = vocoder.inference(voc_inp).view(-1).to("cpu").numpy() elif vocoder_type == "usfgan": if feature_type == "world": fftlen = pyworld.get_cheaptrick_fft_size(sample_rate) use_mcep_aperiodicity = bap.shape[-1] > 5 if use_mcep_aperiodicity: mcep_aperiodicity_order = bap.shape[-1] - 1 alpha = pysptk.util.mcepalpha(sample_rate) aperiodicity = pysptk.mc2sp( np.ascontiguousarray(bap).astype(np.float64), fftlen=fftlen, alpha=alpha, ) else: aperiodicity = pyworld.decode_aperiodicity( np.ascontiguousarray(bap).astype(np.float64), sample_rate, fftlen ) # fill aperiodicity with ones for unvoiced regions aperiodicity[vuv.reshape(-1) < vuv_threshold, 0] = 1.0 # WORLD fails catastrophically for out of range aperiodicity aperiodicity = np.clip(aperiodicity, 0.0, 1.0) # back to bap if use_mcep_aperiodicity: bap = pysptk.sp2mc( aperiodicity, order=mcep_aperiodicity_order, alpha=alpha, ) else: bap = pyworld.code_aperiodicity(aperiodicity, sample_rate).astype( np.float32 ) aux_feats = ( torch.from_numpy( vocoder_in_scaler.transform(np.concatenate([mgc, bap], axis=-1)) ) .float() .to(device) ) elif feature_type == "melf0": # NOTE: So far vocoder models are trained on binary V/UV features vuv = (vuv > vuv_threshold).astype(np.float32) aux_feats = ( torch.from_numpy(vocoder_in_scaler.transform(mel)).float().to(device) ) contf0 = np.exp(lf0) if vocoder_config.data.sine_f0_type in ["contf0", "cf0"]: f0_inp = contf0 elif vocoder_config.data.sine_f0_type == "f0": f0_inp = contf0 f0_inp[vuv < vuv_threshold] = 0 wav = vocoder.inference(f0_inp, aux_feats).view(-1).to("cpu").numpy() return wav def post_process(wav, sample_rate): wav = bandpass_filter(wav, sample_rate) if np.max(wav) > 10: if np.abs(wav).max() > 32767: wav = wav / np.abs(wav).max() # data is likely already in [-32768, 32767] wav = wav.astype(np.int16) else: if np.abs(wav).max() > 1.0: wav = wav / np.abs(wav).max() wav = (wav * 32767.0).astype(np.int16) return wav @hydra.main(config_path="conf/synthesis", config_name="config") def my_app(config: DictConfig) -> None: global logger logger = getLogger(config.verbose) logger.info(OmegaConf.to_yaml(config)) if not torch.cuda.is_available(): device = torch.device("cpu") else: device = torch.device(config.device) acoustic_config = OmegaConf.load(to_absolute_path(config.acoustic.model_yaml)) # Vocoder if config.vocoder.checkpoint is not None and len(config.vocoder.checkpoint) > 0: vocoder, vocoder_in_scaler, vocoder_config = load_vocoder( to_absolute_path(config.vocoder.checkpoint), device, acoustic_config, ) else: vocoder, vocoder_in_scaler, vocoder_config = None, None, None if config.synthesis.vocoder_type != "world": logger.warning("Vocoder checkpoint is not specified") logger.info(f"Use world instead of {config.synthesis.vocoder_type}.") config.synthesis.vocoder_type = "world" # Run synthesis for each utt. in_dir = to_absolute_path(config.in_dir) out_dir = to_absolute_path(config.out_dir) os.makedirs(out_dir, exist_ok=True) utt_ids = load_utt_list(to_absolute_path(config.utt_list)) logger.info("Processes %s utterances...", len(utt_ids)) for utt_id in tqdm(utt_ids): acoustic_features = np.load(join(in_dir, f"{utt_id}-feats.npy")) init_seed(1234) wav = anasyn( device=device, acoustic_features=acoustic_features, acoustic_config=acoustic_config, vocoder=vocoder, vocoder_config=vocoder_config, vocoder_in_scaler=vocoder_in_scaler, sample_rate=config.synthesis.sample_rate, frame_period=config.synthesis.frame_period, use_world_codec=config.synthesis.use_world_codec, feature_type=config.synthesis.feature_type, vocoder_type=config.synthesis.vocoder_type, vuv_threshold=config.synthesis.vuv_threshold, ) wav = post_process(wav, config.synthesis.sample_rate) out_wav_path = join(out_dir, f"{utt_id}.wav") wavfile.write( out_wav_path, rate=config.synthesis.sample_rate, data=wav.astype(np.int16) ) def entry(): my_app() # pylint: disable=no-value-for-parameter if __name__ == "__main__": my_app() # pylint: disable=no-value-for-parameter
7,921
33.146552
86
py
nnsvs
nnsvs-master/nnsvs/layers/conv.py
from torch import nn from torch.nn.utils import weight_norm def WNConv1d(*args, **kwargs): return weight_norm(nn.Conv1d(*args, **kwargs)) class ResnetBlock(nn.Module): def __init__(self, dim, dilation=1): super().__init__() self.block = nn.Sequential( nn.LeakyReLU(0.2), nn.ReflectionPad1d(dilation), WNConv1d(dim, dim, kernel_size=3, dilation=dilation), nn.LeakyReLU(0.2), WNConv1d(dim, dim, kernel_size=1), ) self.shortcut = WNConv1d(dim, dim, kernel_size=1) def forward(self, x): return self.shortcut(x) + self.block(x)
640
26.869565
65
py
nnsvs
nnsvs-master/nnsvs/layers/layer_norm.py
# Adapted from https://github.com/espnet/espnet # Copyright 2019 Shigeki Karita # Apache 2.0 (http://www.apache.org/licenses/LICENSE-2.0) """Layer normalization module.""" import torch class LayerNorm(torch.nn.LayerNorm): """Layer normalization module. Args: nout (int): Output dim size. dim (int): Dimension to be normalized. """ def __init__(self, nout, dim=-1): """Construct an LayerNorm object.""" super(LayerNorm, self).__init__(nout, eps=1e-12) self.dim = dim def forward(self, x): """Apply layer normalization. Args: x (torch.Tensor): Input tensor. Returns: torch.Tensor: Normalized tensor. """ if self.dim == -1: return super(LayerNorm, self).forward(x) return ( super(LayerNorm, self) .forward(x.transpose(self.dim, -1)) .transpose(self.dim, -1) )
953
25.5
59
py
nnsvs
nnsvs-master/nnsvs/wavenet/modules.py
import torch from nnsvs.wavenet import conv from torch import nn def Conv1d(in_channels, out_channels, kernel_size, *args, **kwargs): """Weight-normalized Conv1d layer.""" m = conv.Conv1d(in_channels, out_channels, kernel_size, *args, **kwargs) return nn.utils.weight_norm(m) def Conv1d1x1(in_channels, out_channels, bias=True): """1x1 Weight-normalized Conv1d layer.""" return Conv1d(in_channels, out_channels, kernel_size=1, bias=bias) class ResSkipBlock(nn.Module): """Convolution block with residual and skip connections. Args: residual_channels (int): Residual connection channels. gate_channels (int): Gated activation channels. kernel_size (int): Kernel size of convolution layers. skip_out_channels (int): Skip connection channels. dilation (int): Dilation factor. cin_channels (int): Local conditioning channels. args (list): Additional arguments for Conv1d. kwargs (dict): Additional arguments for Conv1d. """ def __init__( self, residual_channels, # 残差結合のチャネル数 gate_channels, # ゲートのチャネル数 kernel_size, # カーネルサイズ skip_out_channels, # スキップ結合のチャネル数 dilation=1, # dilation factor cin_channels=80, # 条件付特徴量のチャネル数 *args, **kwargs, ): super().__init__() self.padding = (kernel_size - 1) * dilation # 1 次元膨張畳み込み (dilation == 1 のときは、通常の1 次元畳み込み) self.conv = Conv1d( residual_channels, gate_channels, kernel_size, *args, padding=self.padding, dilation=dilation, **kwargs, ) # local conditioning 用の 1x1 convolution self.conv1x1c = Conv1d1x1(cin_channels, gate_channels, bias=False) # ゲート付き活性化関数のために、1 次元畳み込みの出力は2 分割されることに注意 gate_out_channels = gate_channels // 2 self.conv1x1_out = Conv1d1x1(gate_out_channels, residual_channels) self.conv1x1_skip = Conv1d1x1(gate_out_channels, skip_out_channels) def forward(self, x, c): """Forward step Args: x (torch.Tensor): Input signal. c (torch.Tensor): Local conditioning signal. Returns: tuple: Tuple of output signal and skip connection signal """ return self._forward(x, c, False) def incremental_forward(self, x, c): """Incremental forward Args: x (torch.Tensor): Input signal. c (torch.Tensor): Local conditioning signal. Returns: tuple: Tuple of output signal and skip connection signal """ return self._forward(x, c, True) def _forward(self, x, c, is_incremental): # 残差接続用に入力を保持 residual = x # メインの dilated convolutionの計算 # 推論時と学習時で入力のテンソルのshapeが異なるのに注意 if is_incremental: splitdim = -1 # (B, T, C) x = self.conv.incremental_forward(x) else: splitdim = 1 # (B, C, T) x = self.conv(x) # 因果性を保証するために、出力をシフトする x = x[:, :, : -self.padding] # チャンネル方向で出力を分割 a, b = x.split(x.size(splitdim) // 2, dim=splitdim) # local conditioning c = self._conv1x1_forward(self.conv1x1c, c, is_incremental) ca, cb = c.split(c.size(splitdim) // 2, dim=splitdim) a, b = a + ca, b + cb # ゲート付き活性化関数 x = torch.tanh(a) * torch.sigmoid(b) # スキップ接続用の出力を計算 s = self._conv1x1_forward(self.conv1x1_skip, x, is_incremental) # 残差接続の要素和行う前に、次元数を合わせる x = self._conv1x1_forward(self.conv1x1_out, x, is_incremental) x = x + residual return x, s def _conv1x1_forward(self, conv, x, is_incremental): if is_incremental: x = conv.incremental_forward(x) else: x = conv(x) return x def clear_buffer(self): """Clear input buffer.""" for c in [ self.conv, self.conv1x1_out, self.conv1x1_skip, self.conv1x1c, ]: if c is not None: c.clear_buffer()
4,184
28.680851
76
py
nnsvs
nnsvs-master/nnsvs/wavenet/wavenet.py
import torch from nnsvs.wavenet.modules import Conv1d1x1, ResSkipBlock from torch import nn from torch.nn import functional as F class WaveNet(nn.Module): """WaveNet Args: in_dim (int): the dimension of the input out_dim (int): the dimension of the output layers (int): the number of layers stacks (int): the number of residual stacks residual_channels (int): the number of residual channels gate_channels (int): the number of channels for the gating function skip_out_channels (int): the number of channels in the skip output kernel_size (int): the size of the convolutional kernel """ def __init__( self, in_dim=334, out_dim=206, layers=10, stacks=1, residual_channels=64, gate_channels=128, skip_out_channels=64, kernel_size=3, ): super().__init__() self.in_dim = in_dim self.out_dim = out_dim self.first_conv = Conv1d1x1(out_dim, residual_channels) self.main_conv_layers = nn.ModuleList() layers_per_stack = layers // stacks for layer in range(layers): dilation = 2 ** (layer % layers_per_stack) conv = ResSkipBlock( residual_channels, gate_channels, kernel_size, skip_out_channels, dilation=dilation, cin_channels=in_dim, ) self.main_conv_layers.append(conv) self.last_conv_layers = nn.ModuleList( [ nn.ReLU(), Conv1d1x1(skip_out_channels, skip_out_channels), nn.ReLU(), Conv1d1x1(skip_out_channels, out_dim), ] ) def forward(self, c, x, lengths=None): """Forward step Args: c (torch.Tensor): the conditional features (B, T, C) x (torch.Tensor): the target features (B, T, C) Returns: torch.Tensor: the output waveform """ x = x.transpose(1, 2) c = c.transpose(1, 2) x = self.first_conv(x) skips = 0 for f in self.main_conv_layers: x, h = f(x, c) skips += h x = skips for f in self.last_conv_layers: x = f(x) # (B, C, T) -> (B, T, C) x = x.transpose(1, 2) return x def inference(self, c, num_time_steps=100, tqdm=lambda x: x): """Inference step Args: c (torch.Tensor): the local conditioning feature (B, T, C) num_time_steps (int): the number of time steps to generate tqdm (lambda): a tqdm function to track progress Returns: torch.Tensor: the output waveform """ self.clear_buffer() # Local conditioning B = c.shape[0] outputs = [] # 自己回帰生成における初期値 current_input = torch.zeros(B, 1, self.out_dim).to(c.device) if tqdm is None: ts = range(num_time_steps) else: ts = tqdm(range(num_time_steps)) # 逐次的に生成 for t in ts: # 時刻 t における入力は、時刻 t-1 における出力 if t > 0: current_input = outputs[-1] # 時刻 t における条件付け特徴量 ct = c[:, t, :].unsqueeze(1) x = current_input x = self.first_conv.incremental_forward(x) skips = 0 for f in self.main_conv_layers: x, h = f.incremental_forward(x, ct) skips += h x = skips for f in self.last_conv_layers: if hasattr(f, "incremental_forward"): x = f.incremental_forward(x) else: x = f(x) # Softmax によって、出力をカテゴリカル分布のパラメータに変換 x = F.softmax(x.view(B, -1), dim=1) # カテゴリカル分布からサンプリング x = torch.distributions.OneHotCategorical(x).sample() outputs += [x.data] # T x B x C # 各時刻における出力を結合 outputs = torch.stack(outputs) # B x T x C outputs = outputs.transpose(0, 1).contiguous() self.clear_buffer() return outputs def clear_buffer(self): """Clear the internal buffer.""" self.first_conv.clear_buffer() for f in self.main_conv_layers: f.clear_buffer() for f in self.last_conv_layers: try: f.clear_buffer() except AttributeError: pass def remove_weight_norm_(self): """Remove weight normalization of the model""" def _remove_weight_norm(m): try: torch.nn.utils.remove_weight_norm(m) except ValueError: return self.apply(_remove_weight_norm)
4,862
27.273256
75
py
nnsvs
nnsvs-master/nnsvs/wavenet/conv.py
import torch from packaging import version from torch import nn from torch.nn import functional as F torch_is_ge_180 = version.parse(torch.__version__) >= version.parse("1.8.0") class Conv1d(nn.Conv1d): """Extended nn.Conv1d for incremental dilated convolutions""" def __init__(self, *args, **kwargs): super().__init__(*args, **kwargs) self.clear_buffer() self._linearized_weight = None if torch_is_ge_180: self.register_full_backward_hook(self._clear_linearized_weight) else: self.register_backward_hook(self._clear_linearized_weight) def incremental_forward(self, input): # input: (B, T, C) if self.training: raise RuntimeError("incremental_forward only supports eval mode") # run forward pre hooks (e.g., weight norm) for hook in self._forward_pre_hooks.values(): hook(self, input) # reshape weight weight = self._get_linearized_weight() kw = self.kernel_size[0] dilation = self.dilation[0] bsz = input.size(0) # input: bsz x len x dim if kw > 1: input = input.data if self.input_buffer is None: self.input_buffer = input.new( bsz, kw + (kw - 1) * (dilation - 1), input.size(2) ) self.input_buffer.zero_() else: # shift buffer self.input_buffer[:, :-1, :] = self.input_buffer[:, 1:, :].clone() # append next input self.input_buffer[:, -1, :] = input[:, -1, :] input = self.input_buffer if dilation > 1: input = input[:, 0::dilation, :].contiguous() with torch.no_grad(): output = F.linear(input.view(bsz, -1), weight, self.bias) return output.view(bsz, 1, -1) def clear_buffer(self): self.input_buffer = None def _get_linearized_weight(self): if self._linearized_weight is None: kw = self.kernel_size[0] # nn.Conv1d assert self.weight.size() == (self.out_channels, self.in_channels, kw) weight = self.weight.transpose(1, 2).contiguous() assert weight.size() == (self.out_channels, kw, self.in_channels) self._linearized_weight = weight.view(self.out_channels, -1) return self._linearized_weight def _clear_linearized_weight(self, *args): self._linearized_weight = None
2,508
34.842857
82
py
nnsvs
nnsvs-master/nnsvs/tacotron/postnet.py
# Acknowledgement: some of the code was adapted from ESPnet # Copyright 2019 Nagoya University (Tomoki Hayashi) # Apache 2.0 (http://www.apache.org/licenses/LICENSE-2.0) from torch import nn class Postnet(nn.Module): """Post-Net of Tacotron 2 Args: in_dim (int): dimension of input layers (int): number of layers channels (int): number of channels kernel_size (int): kernel size dropout (float): dropout rate """ def __init__( self, in_dim, layers=5, channels=512, kernel_size=5, dropout=0.5, ): super().__init__() postnet = nn.ModuleList() for layer in range(layers): in_channels = in_dim if layer == 0 else channels out_channels = in_dim if layer == layers - 1 else channels postnet += [ nn.Conv1d( in_channels, out_channels, kernel_size, stride=1, padding=(kernel_size - 1) // 2, bias=False, ), nn.BatchNorm1d(out_channels), ] if layer != layers - 1: postnet += [nn.Tanh()] postnet += [nn.Dropout(dropout)] self.postnet = nn.Sequential(*postnet) def forward(self, xs): """Forward step Args: xs (torch.Tensor): input sequence Returns: torch.Tensor: output sequence """ return self.postnet(xs)
1,571
25.644068
70
py
nnsvs
nnsvs-master/nnsvs/tacotron/encoder.py
# The code was adapted from ttslearn https://github.com/r9y9/ttslearn # Acknowledgement: some of the code was adapted from ESPnet # Copyright 2019 Nagoya University (Tomoki Hayashi) # Apache 2.0 (http://www.apache.org/licenses/LICENSE-2.0) from torch import nn from torch.nn.utils.rnn import pack_padded_sequence, pad_packed_sequence def encoder_init(m): if isinstance(m, nn.Conv1d): nn.init.xavier_uniform_(m.weight, nn.init.calculate_gain("relu")) class Encoder(nn.Module): """Encoder of Tacotron 2 Args: in_dim (int): dimension of embeddings hidden_dim (int): dimension of hidden units conv_layers (int): number of convolutional layers conv_channels (int): number of convolutional channels conv_kernel_size (int): size of convolutional kernel dropout (float): dropout rate """ def __init__( self, in_dim=512, hidden_dim=512, conv_layers=3, conv_channels=512, conv_kernel_size=5, dropout=0.5, ): super(Encoder, self).__init__() convs = nn.ModuleList() for layer in range(conv_layers): in_channels = in_dim if layer == 0 else conv_channels convs += [ nn.Conv1d( in_channels, conv_channels, conv_kernel_size, padding=(conv_kernel_size - 1) // 2, bias=False, ), nn.BatchNorm1d(conv_channels), nn.ReLU(), nn.Dropout(dropout), ] self.convs = nn.Sequential(*convs) self.blstm = nn.LSTM( conv_channels, hidden_dim // 2, 1, batch_first=True, bidirectional=True ) self.apply(encoder_init) def forward(self, seqs, in_lens): """Forward step Args: seqs (torch.Tensor): input sequences (B, T, C) in_lens (torch.Tensor): input sequence lengths Returns: torch.Tensor: encoded sequences """ out = self.convs(seqs.transpose(1, 2)).transpose(1, 2) if not isinstance(in_lens, list): in_lens = in_lens.to("cpu") out = pack_padded_sequence(out, in_lens, batch_first=True) out, _ = self.blstm(out) out, _ = pad_packed_sequence(out, batch_first=True) return out
2,409
29.897436
83
py
nnsvs
nnsvs-master/nnsvs/tacotron/decoder.py
# The code was adapted from ttslearn https://github.com/r9y9/ttslearn # NonAttentiveDecoder is added to the original code. # Acknowledgement: some of the code was adapted from ESPnet # Copyright 2019 Nagoya University (Tomoki Hayashi) # Apache 2.0 (http://www.apache.org/licenses/LICENSE-2.0) import torch import torch.nn.functional as F from nnsvs.base import BaseModel, PredictionType from nnsvs.mdn import MDNLayer, mdn_get_most_probable_sigma_and_mu, mdn_get_sample from nnsvs.util import init_weights from torch import nn def decoder_init(m): if isinstance(m, nn.Conv1d): nn.init.xavier_uniform_(m.weight, nn.init.calculate_gain("tanh")) class ZoneOutCell(nn.Module): def __init__(self, cell, zoneout=0.1): super().__init__() self.cell = cell self.hidden_size = cell.hidden_size self.zoneout = zoneout def forward(self, inputs, hidden): next_hidden = self.cell(inputs, hidden) next_hidden = self._zoneout(hidden, next_hidden, self.zoneout) return next_hidden def _zoneout(self, h, next_h, prob): h_0, c_0 = h h_1, c_1 = next_h h_1 = self._apply_zoneout(h_0, h_1, prob) c_1 = self._apply_zoneout(c_0, c_1, prob) return h_1, c_1 def _apply_zoneout(self, h, next_h, prob): if self.training: if prob > 0.0: mask = h.new(*h.size()).bernoulli_(prob) else: mask = 0 return mask * h + (1 - mask) * next_h else: return prob * h + (1 - prob) * next_h class Prenet(nn.Module): """Pre-Net of Tacotron. Args: in_dim (int) : dimension of input layers (int) : number of pre-net layers hidden_dim (int) : dimension of hidden layer dropout (float) : dropout rate """ def __init__( self, in_dim, layers=2, hidden_dim=256, dropout=0.5, eval_dropout=True ): super().__init__() self.dropout = dropout self.eval_dropout = eval_dropout prenet = nn.ModuleList() for layer in range(layers): prenet += [ nn.Linear(in_dim if layer == 0 else hidden_dim, hidden_dim), nn.ReLU(), ] self.prenet = nn.Sequential(*prenet) def forward(self, x): """Forward step Args: x (torch.Tensor) : input tensor Returns: torch.Tensor : output tensor """ for layer in self.prenet: if self.eval_dropout: x = F.dropout(layer(x), self.dropout, training=True) else: x = F.dropout(layer(x), self.dropout, training=self.training) return x class NonAttentiveDecoder(BaseModel): """Decoder of Tacotron w/o attention mechanism Args: in_dim (int) : dimension of encoder hidden layer out_dim (int) : dimension of output layers (int) : number of LSTM layers hidden_dim (int) : dimension of hidden layer prenet_layers (int) : number of pre-net layers prenet_hidden_dim (int) : dimension of pre-net hidden layer prenet_dropout (float) : dropout rate of pre-net zoneout (float) : zoneout rate reduction_factor (int) : reduction factor attention_hidden_dim (int) : dimension of attention hidden layer attention_conv_channel (int) : number of attention convolution channels attention_conv_kernel_size (int) : kernel size of attention convolution downsample_by_conv (bool) : if True, downsample by convolution initial_value (float) : initial value for the autoregressive decoder. """ def __init__( self, in_dim=512, out_dim=80, layers=2, hidden_dim=1024, prenet_layers=2, prenet_hidden_dim=256, prenet_dropout=0.5, zoneout=0.1, reduction_factor=1, downsample_by_conv=False, init_type="none", eval_dropout=True, prenet_noise_std=0.0, initial_value=0.0, ): super().__init__() self.out_dim = out_dim self.reduction_factor = reduction_factor self.prenet_dropout = prenet_dropout self.prenet_noise_std = prenet_noise_std self.initial_value = initial_value if prenet_layers > 0: self.prenet = Prenet( out_dim, prenet_layers, prenet_hidden_dim, prenet_dropout, eval_dropout=eval_dropout, ) lstm_in_dim = in_dim + prenet_hidden_dim else: self.prenet = None prenet_hidden_dim = 0 lstm_in_dim = in_dim + out_dim self.lstm = nn.ModuleList() for layer in range(layers): lstm = nn.LSTMCell( lstm_in_dim if layer == 0 else hidden_dim, hidden_dim, ) self.lstm += [ZoneOutCell(lstm, zoneout)] proj_in_dim = in_dim + hidden_dim self.feat_out = nn.Linear(proj_in_dim, out_dim * reduction_factor, bias=False) if reduction_factor > 1 and downsample_by_conv: self.conv_downsample = nn.Conv1d( in_dim, in_dim, kernel_size=reduction_factor, stride=reduction_factor, groups=in_dim, ) else: self.conv_downsample = None init_weights(self, init_type) def _zero_state(self, hs): init_hs = hs.new_zeros(hs.size(0), self.lstm[0].hidden_size) return init_hs def is_autoregressive(self): return True def forward(self, encoder_outs, in_lens, decoder_targets=None): """Forward step Args: encoder_outs (torch.Tensor): encoder outputs (B, T, C) in_lens (torch.Tensor): input lengths decoder_targets (torch.Tensor): decoder targets for teacher-forcing. (B, T, C) Returns: torch.Tensor: the output (B, C, T) """ is_inference = decoder_targets is None if not is_inference: assert encoder_outs.shape[1] == decoder_targets.shape[1] # Adjust number of frames according to the reduction factor # (B, Lmax, out_dim) -> (B, Lmax/r, out_dim) if self.reduction_factor > 1 and not is_inference: decoder_targets = decoder_targets[ :, self.reduction_factor - 1 :: self.reduction_factor ] if self.reduction_factor > 1: if self.conv_downsample is not None: encoder_outs = self.conv_downsample( encoder_outs.transpose(1, 2) ).transpose(1, 2) else: encoder_outs = encoder_outs[ :, self.reduction_factor - 1 :: self.reduction_factor ] h_list, c_list = [], [] for _ in range(len(self.lstm)): h_list.append(self._zero_state(encoder_outs)) c_list.append(self._zero_state(encoder_outs)) go_frame = ( encoder_outs.new_zeros(encoder_outs.size(0), self.out_dim) + self.initial_value ) prev_out = go_frame if not is_inference and self.prenet is not None: prenet_outs = self.prenet(decoder_targets) outs = [] for t in range(encoder_outs.shape[1]): # Pre-Net if self.prenet is not None: if is_inference: prenet_out = self.prenet(prev_out) else: prenet_out = prenet_outs[:, t, :] elif self.prenet_noise_std > 0: prenet_out = ( prev_out + torch.randn_like(prev_out) * self.prenet_noise_std ) else: prenet_out = F.dropout(prev_out, self.prenet_dropout, training=True) # LSTM xs = torch.cat([encoder_outs[:, t], prenet_out], dim=1) h_list[0], c_list[0] = self.lstm[0](xs, (h_list[0], c_list[0])) for i in range(1, len(self.lstm)): h_list[i], c_list[i] = self.lstm[i]( h_list[i - 1], (h_list[i], c_list[i]) ) # Output hcs = torch.cat([h_list[-1], encoder_outs[:, t]], dim=1) outs.append(self.feat_out(hcs).view(encoder_outs.size(0), self.out_dim, -1)) # Update decoder input for the next time step if is_inference: prev_out = outs[-1][:, :, -1] # (1, out_dim) else: # Teacher forcing prev_out = decoder_targets[:, t, :] outs = torch.cat(outs, dim=2) # (B, out_dim, Lmax) if self.reduction_factor > 1: outs = outs.view(outs.size(0), self.out_dim, -1) # (B, out_dim, Lmax) # (B, C, T) -> (B, T, C) return outs.transpose(1, 2) class MDNNonAttentiveDecoder(BaseModel): """Non-atteive decoder with MDN Each decoder step outputs the parameters of MDN. Args: in_dim (int): input dimension out_dim (int): output dimension layers (int): number of LSTM layers hidden_dim (int): hidden dimension prenet_layers (int): number of prenet layers prenet_hidden_dim (int): prenet hidden dimension prenet_dropout (float): prenet dropout rate zoneout (float): zoneout rate reduction_factor (int): reduction factor downsample_by_conv (bool): if True, use conv1d to downsample the input num_gaussians (int): number of Gaussians sampling_mode (str): sampling mode init_type (str): initialization type eval_dropout (bool): if True, use dropout in evaluation initial_value (float) : initial value for the autoregressive decoder. """ def __init__( self, in_dim=512, out_dim=80, layers=2, hidden_dim=1024, prenet_layers=2, prenet_hidden_dim=256, prenet_dropout=0.5, zoneout=0.1, reduction_factor=1, downsample_by_conv=False, num_gaussians=8, sampling_mode="mean", init_type="none", eval_dropout=True, prenet_noise_std=0.0, initial_value=0.0, ): super().__init__() self.out_dim = out_dim self.reduction_factor = reduction_factor self.prenet_dropout = prenet_dropout self.prenet_noise_std = prenet_noise_std self.num_gaussians = num_gaussians self.sampling_mode = sampling_mode assert sampling_mode in ["mean", "random"] self.initial_value = initial_value if prenet_layers > 0: self.prenet = Prenet( out_dim, prenet_layers, prenet_hidden_dim, prenet_dropout, eval_dropout=eval_dropout, ) lstm_in_dim = in_dim + prenet_hidden_dim else: self.prenet = None prenet_hidden_dim = 0 lstm_in_dim = in_dim + out_dim self.lstm = nn.ModuleList() for layer in range(layers): lstm = nn.LSTMCell( lstm_in_dim if layer == 0 else hidden_dim, hidden_dim, ) self.lstm += [ZoneOutCell(lstm, zoneout)] proj_in_dim = in_dim + hidden_dim self.feat_out = MDNLayer( proj_in_dim, out_dim * reduction_factor, num_gaussians=num_gaussians, dim_wise=True, ) if reduction_factor > 1 and downsample_by_conv: self.conv_downsample = nn.Conv1d( in_dim, in_dim, kernel_size=reduction_factor, stride=reduction_factor, groups=in_dim, ) else: self.conv_downsample = None init_weights(self, init_type) def _zero_state(self, hs): init_hs = hs.new_zeros(hs.size(0), self.lstm[0].hidden_size) return init_hs def is_autoregressive(self): return True def prediction_type(self): return PredictionType.PROBABILISTIC def forward(self, encoder_outs, in_lens, decoder_targets=None): is_inference = decoder_targets is None if not is_inference: assert encoder_outs.shape[1] == decoder_targets.shape[1] # Adjust number of frames according to the reduction factor # (B, Lmax, out_dim) -> (B, Lmax/r, out_dim) if self.reduction_factor > 1 and not is_inference: decoder_targets = decoder_targets[ :, self.reduction_factor - 1 :: self.reduction_factor ] if self.reduction_factor > 1: if self.conv_downsample is not None: encoder_outs = self.conv_downsample( encoder_outs.transpose(1, 2) ).transpose(1, 2) else: encoder_outs = encoder_outs[ :, self.reduction_factor - 1 :: self.reduction_factor ] h_list, c_list = [], [] for _ in range(len(self.lstm)): h_list.append(self._zero_state(encoder_outs)) c_list.append(self._zero_state(encoder_outs)) go_frame = ( encoder_outs.new_zeros(encoder_outs.size(0), self.out_dim) + self.initial_value ) prev_out = go_frame if not is_inference and self.prenet is not None: prenet_outs = self.prenet(decoder_targets) mus = [] log_pis = [] log_sigmas = [] # NOTE: only used for inference mus_inf = [] for t in range(encoder_outs.shape[1]): # Pre-Net if self.prenet is not None: if is_inference: prenet_out = self.prenet(prev_out) else: prenet_out = prenet_outs[:, t, :] elif self.prenet_noise_std > 0: prenet_out = ( prev_out + torch.randn_like(prev_out) * self.prenet_noise_std ) else: prenet_out = F.dropout(prev_out, self.prenet_dropout, training=True) # LSTM xs = torch.cat([encoder_outs[:, t], prenet_out], dim=1) h_list[0], c_list[0] = self.lstm[0](xs, (h_list[0], c_list[0])) for i in range(1, len(self.lstm)): h_list[i], c_list[i] = self.lstm[i]( h_list[i - 1], (h_list[i], c_list[i]) ) # Output hcs = torch.cat([h_list[-1], encoder_outs[:, t]], dim=1) log_pi, log_sigma, mu = self.feat_out(hcs.unsqueeze(1)) # (B, 1, num_gaussians, out_dim*reduction_factor) # -> (B, reduction_factor, num_gaussians, out_dim) log_pi = ( log_pi.transpose(1, 2) .view(encoder_outs.size(0), self.num_gaussians, -1, self.out_dim) .transpose(1, 2) ) log_sigma = ( log_sigma.transpose(1, 2) .view(encoder_outs.size(0), self.num_gaussians, -1, self.out_dim) .transpose(1, 2) ) mu = ( mu.transpose(1, 2) .view(encoder_outs.size(0), self.num_gaussians, -1, self.out_dim) .transpose(1, 2) ) mus.append(mu) log_pis.append(log_pi) log_sigmas.append(log_sigma) # Update decoder input for the next time step if is_inference: # (B, reduction_factor, out_dim) if self.sampling_mode == "mean": _, mu = mdn_get_most_probable_sigma_and_mu(log_pi, log_sigma, mu) elif self.sampling_mode == "random": mu = mdn_get_sample(log_pi, log_sigma, mu) # Feed last sample for the feedback loop prev_out = mu[:, -1] mus_inf.append(mu) else: # Teacher forcing prev_out = decoder_targets[:, t, :] mus = torch.cat(mus, dim=1) # (B, out_dim, Lmax) log_pis = torch.cat(log_pis, dim=1) # (B, out_dim, Lmax) log_sigmas = torch.cat(log_sigmas, dim=1) # (B, out_dim, Lmax) if is_inference: mu = torch.cat(mus_inf, dim=1) # TODO: may need to track sigma. For now we only use mu return mu, mu else: return log_pis, log_sigmas, mus
16,660
33.281893
90
py
nnsvs
nnsvs-master/nnsvs/usfgan/__init__.py
import numpy as np import torch from nnsvs.usfgan.utils import SignalGenerator, dilated_factor from torch import nn class USFGANWrapper(nn.Module): def __init__(self, config, generator): super().__init__() self.generator = generator self.config = config def inference(self, f0, aux_feats): """Inference for USFGAN Args: f0 (numpy.ndarray): F0 (T, 1) aux_feats (Tensor): Auxiliary features (T, C) """ signal_generator = SignalGenerator( sample_rate=self.config.data.sample_rate, hop_size=self.config.data.hop_size, sine_amp=self.config.data.sine_amp, noise_amp=self.config.data.noise_amp, signal_types=self.config.data.signal_types, ) assert self.config.data.sine_f0_type in ["contf0", "cf0", "f0"] assert self.config.data.df_f0_type in ["contf0", "cf0", "f0"] device = aux_feats.device is_sifigan = "aux_context_window" not in self.config.generator if is_sifigan: dfs = [] for df, us in zip( self.config.data.dense_factors, np.cumprod(self.config.generator.upsample_scales), ): dfs += [ np.repeat( dilated_factor(f0.copy(), self.config.data.sample_rate, df), us ) ] df = [ torch.FloatTensor(np.array(df)).view(1, 1, -1).to(device) for df in dfs ] c = aux_feats.unsqueeze(0).transpose(2, 1).to(device) else: df = dilated_factor( np.squeeze(f0.copy()), self.config.data.sample_rate, self.config.data.dense_factor, ) df = df.repeat(self.config.data.hop_size, axis=0) pad_fn = nn.ReplicationPad1d(self.config.generator.aux_context_window) c = pad_fn(aux_feats.unsqueeze(0).transpose(2, 1)).to(device) df = torch.FloatTensor(df).view(1, 1, -1).to(device) f0 = torch.FloatTensor(f0).unsqueeze(0).transpose(2, 1).to(device) in_signal = signal_generator(f0) y = self.generator(in_signal, c, df)[0] return y
2,281
33.575758
87
py
nnsvs
nnsvs-master/nnsvs/usfgan/models/discriminator.py
# -*- coding: utf-8 -*- # Copyright 2022 Reo Yoneyama (Nagoya University) # MIT License (https://opensource.org/licenses/MIT) """Discriminator modules. References: - https://github.com/bigpon/QPPWG - https://github.com/jik876/hifi-gan """ import copy from logging import getLogger from tkinter import W # A logger for this file logger = getLogger(__name__) import torch import torch.nn as nn import torch.nn.functional as F from nnsvs.usfgan.layers import Conv1d, Conv2d from torchaudio.functional import spectrogram class PWGDiscriminator(nn.Module): """Parallel WaveGAN Discriminator module.""" def __init__( self, in_channels=1, out_channels=1, kernel_size=3, layers=10, conv_channels=64, dilation_factor=1, nonlinear_activation="LeakyReLU", nonlinear_activation_params={"negative_slope": 0.2}, bias=True, use_weight_norm=True, ): """Initialize Parallel WaveGAN Discriminator module. Args: in_channels (int): Number of input channels. out_channels (int): Number of output channels. kernel_size (int): Number of output channels. layers (int): Number of conv layers. conv_channels (int): Number of chnn layers. dilation_factor (int): Dilation factor. For example, if dilation_factor = 2, the dilation will be 2, 4, 8, ..., and so on. nonlinear_activation (str): Nonlinear function after each conv. nonlinear_activation_params (dict): Nonlinear function parameters bias (bool): Whether to use bias parameter in conv. use_weight_norm (bool) Whether to use weight norm. If set to true, it will be applied to all of the conv layers. """ super(PWGDiscriminator, self).__init__() assert (kernel_size - 1) % 2 == 0, "Not support even number kernel size." assert dilation_factor > 0, "Dilation factor must be > 0." self.conv_layers = nn.ModuleList() conv_in_channels = in_channels for i in range(layers - 1): if i == 0: dilation = 1 else: dilation = i if dilation_factor == 1 else dilation_factor ** i conv_in_channels = conv_channels padding = (kernel_size - 1) // 2 * dilation conv_layer = [ Conv1d( conv_in_channels, conv_channels, kernel_size=kernel_size, padding=padding, dilation=dilation, bias=bias, ), getattr(nn, nonlinear_activation)( inplace=True, **nonlinear_activation_params ), ] self.conv_layers += conv_layer padding = (kernel_size - 1) // 2 conv_last_layer = Conv1d( conv_in_channels, out_channels, kernel_size=kernel_size, padding=padding, bias=bias, ) self.conv_layers += [conv_last_layer] # apply weight norm if use_weight_norm: self.apply_weight_norm() def forward(self, x, return_fmaps=False): """Calculate forward propagation. Args: x (Tensor): Input noise signal (B, 1, T). return_fmaps (bool): Whether to return feature maps. Returns: Tensor: Output tensor (B, 1, T) """ fmaps = [] for f in self.conv_layers: x = f(x) if return_fmaps: fmaps.append(x) if return_fmaps: return [x], fmaps else: return [x] def apply_weight_norm(self): """Apply weight normalization module from all of the layers.""" def _apply_weight_norm(m): if isinstance(m, nn.Conv1d) or isinstance(m, nn.Conv2d): nn.utils.weight_norm(m) logger.debug(f"Weight norm is applied to {m}.") self.apply(_apply_weight_norm) def remove_weight_norm(self): """Remove weight normalization module from all of the layers.""" def _remove_weight_norm(m): try: logger.debug(f"Weight norm is removed from {m}.") nn.utils.remove_weight_norm(m) except ValueError: # this module didn't have weight norm return self.apply(_remove_weight_norm) class HiFiGANPeriodDiscriminator(nn.Module): """HiFiGAN period discriminator module.""" def __init__( self, in_channels=1, out_channels=1, period=3, kernel_sizes=[5, 3], channels=32, downsample_scales=[3, 3, 3, 3, 1], max_downsample_channels=1024, bias=True, nonlinear_activation="LeakyReLU", nonlinear_activation_params={"negative_slope": 0.1}, use_weight_norm=True, use_spectral_norm=False, ): """Initialize HiFiGANPeriodDiscriminator module. Args: in_channels (int): Number of input channels. out_channels (int): Number of output channels. period (int): Period. kernel_sizes (list): Kernel sizes of initial conv layers and the final conv layer. channels (int): Number of initial channels. downsample_scales (list): List of downsampling scales. max_downsample_channels (int): Number of maximum downsampling channels. use_additional_convs (bool): Whether to use additional conv layers in residual blocks. bias (bool): Whether to add bias parameter in convolution layers. nonlinear_activation (str): Activation function module name. nonlinear_activation_params (dict): Hyperparameters for activation function. use_weight_norm (bool): Whether to use weight norm. If set to true, it will be applied to all of the conv layers. use_spectral_norm (bool): Whether to use spectral norm. If set to true, it will be applied to all of the conv layers. """ super().__init__() assert len(kernel_sizes) == 2 assert kernel_sizes[0] % 2 == 1, "Kernel size must be odd number." assert kernel_sizes[1] % 2 == 1, "Kernel size must be odd number." self.period = period self.convs = nn.ModuleList() in_chs = in_channels out_chs = channels for downsample_scale in downsample_scales: self.convs += [ nn.Sequential( nn.Conv2d( in_chs, out_chs, (kernel_sizes[0], 1), (downsample_scale, 1), padding=((kernel_sizes[0] - 1) // 2, 0), ), getattr(nn, nonlinear_activation)(**nonlinear_activation_params), ) ] in_chs = out_chs # NOTE(kan-bayashi): Use downsample_scale + 1? out_chs = min(out_chs * 4, max_downsample_channels) self.output_conv = nn.Conv2d( out_chs, out_channels, (kernel_sizes[1] - 1, 1), 1, padding=((kernel_sizes[1] - 1) // 2, 0), ) if use_weight_norm and use_spectral_norm: raise ValueError("Either use use_weight_norm or use_spectral_norm.") # apply weight norm if use_weight_norm: self.apply_weight_norm() # apply spectral norm if use_spectral_norm: self.apply_spectral_norm() def forward(self, x, return_fmaps=False): """Calculate forward propagation. Args: c (Tensor): Input tensor (B, in_channels, T). return_fmaps (bool): Whether to return feature maps. Returns: list: List of each layer's tensors. """ # transform 1d to 2d -> (B, C, T/P, P) b, c, t = x.shape if t % self.period != 0: n_pad = self.period - (t % self.period) x = F.pad(x, (0, n_pad), "reflect") t += n_pad x = x.view(b, c, t // self.period, self.period) # forward conv fmap = [] for f in self.convs: x = f(x) if return_fmaps: fmap.append(x) x = self.output_conv(x) out = torch.flatten(x, 1, -1) if return_fmaps: return out, fmap else: return out def apply_weight_norm(self): """Apply weight normalization module from all of the layers.""" def _apply_weight_norm(m): if isinstance(m, nn.Conv2d): nn.utils.weight_norm(m) logger.debug(f"Weight norm is applied to {m}.") self.apply(_apply_weight_norm) def apply_spectral_norm(self): """Apply spectral normalization module from all of the layers.""" def _apply_spectral_norm(m): if isinstance(m, nn.Conv2d): nn.utils.spectral_norm(m) logger.debug(f"Spectral norm is applied to {m}.") self.apply(_apply_spectral_norm) class HiFiGANMultiPeriodDiscriminator(nn.Module): """HiFiGAN multi-period discriminator module.""" def __init__( self, periods=[2, 3, 5, 7, 11], discriminator_params={ "in_channels": 1, "out_channels": 1, "kernel_sizes": [5, 3], "channels": 32, "downsample_scales": [3, 3, 3, 3, 1], "max_downsample_channels": 1024, "bias": True, "nonlinear_activation": "LeakyReLU", "nonlinear_activation_params": {"negative_slope": 0.1}, "use_weight_norm": True, "use_spectral_norm": False, }, ): """Initialize HiFiGANMultiPeriodDiscriminator module. Args: periods (list): List of periods. discriminator_params (dict): Parameters for hifi-gan period discriminator module. The period parameter will be overwritten. """ super().__init__() self.discriminators = nn.ModuleList() for period in periods: params = copy.deepcopy(discriminator_params) params["period"] = period self.discriminators += [HiFiGANPeriodDiscriminator(**params)] def forward(self, x, return_fmaps=False): """Calculate forward propagation. Args: x (Tensor): Input noise signal (B, 1, T). return_fmaps (bool): Whether to return feature maps. Returns: List: List of list of each discriminator outputs, which consists of each layer output tensors. """ outs, fmaps = [], [] for f in self.discriminators: if return_fmaps: out, fmap = f(x, return_fmaps) fmaps.extend(fmap) else: out = f(x) outs.append(out) if return_fmaps: return outs, fmaps else: return outs class HiFiGANScaleDiscriminator(nn.Module): """HiFi-GAN scale discriminator module.""" def __init__( self, in_channels=1, out_channels=1, kernel_sizes=[15, 41, 5, 3], channels=128, max_downsample_channels=1024, max_groups=16, bias=True, downsample_scales=[2, 2, 4, 4, 1], nonlinear_activation="LeakyReLU", nonlinear_activation_params={"negative_slope": 0.1}, use_weight_norm=True, use_spectral_norm=False, ): """Initialize HiFiGAN scale discriminator module. Args: in_channels (int): Number of input channels. out_channels (int): Number of output channels. kernel_sizes (list): List of four kernel sizes. The first will be used for the first conv layer, and the second is for downsampling part, and the remaining two are for output layers. channels (int): Initial number of channels for conv layer. max_downsample_channels (int): Maximum number of channels for downsampling layers. bias (bool): Whether to add bias parameter in convolution layers. downsample_scales (list): List of downsampling scales. nonlinear_activation (str): Activation function module name. nonlinear_activation_params (dict): Hyperparameters for activation function. use_weight_norm (bool): Whether to use weight norm. If set to true, it will be applied to all of the conv layers. use_spectral_norm (bool): Whether to use spectral norm. If set to true, it will be applied to all of the conv layers. """ super().__init__() self.layers = nn.ModuleList() # check kernel size is valid assert len(kernel_sizes) == 4 for ks in kernel_sizes: assert ks % 2 == 1 # add first layer self.layers += [ nn.Sequential( nn.Conv1d( in_channels, channels, # NOTE(kan-bayashi): Use always the same kernel size kernel_sizes[0], bias=bias, padding=(kernel_sizes[0] - 1) // 2, ), getattr(nn, nonlinear_activation)(**nonlinear_activation_params), ) ] # add downsample layers in_chs = channels out_chs = channels # NOTE(kan-bayashi): Remove hard coding? groups = 4 for downsample_scale in downsample_scales: self.layers += [ nn.Sequential( nn.Conv1d( in_chs, out_chs, kernel_size=kernel_sizes[1], stride=downsample_scale, padding=(kernel_sizes[1] - 1) // 2, groups=groups, bias=bias, ), getattr(nn, nonlinear_activation)(**nonlinear_activation_params), ) ] in_chs = out_chs # NOTE(kan-bayashi): Remove hard coding? out_chs = min(in_chs * 2, max_downsample_channels) # NOTE(kan-bayashi): Remove hard coding? groups = min(groups * 4, max_groups) # add final layers out_chs = min(in_chs * 2, max_downsample_channels) self.layers += [ nn.Sequential( nn.Conv1d( in_chs, out_chs, kernel_size=kernel_sizes[2], stride=1, padding=(kernel_sizes[2] - 1) // 2, bias=bias, ), getattr(nn, nonlinear_activation)(**nonlinear_activation_params), ) ] self.last_layer = nn.Conv1d( out_chs, out_channels, kernel_size=kernel_sizes[3], stride=1, padding=(kernel_sizes[3] - 1) // 2, bias=bias, ) if use_weight_norm and use_spectral_norm: raise ValueError("Either use use_weight_norm or use_spectral_norm.") # apply weight norm if use_weight_norm: self.apply_weight_norm() # apply spectral norm if use_spectral_norm: self.apply_spectral_norm() def forward(self, x, return_fmaps=False): """Calculate forward propagation. Args: x (Tensor): Input noise signal (B, 1, T). return_fmaps (bool): Whether to return feature maps. Returns: List: List of output tensors of each layer. """ fmap = [] for f in self.layers: x = f(x) if return_fmaps: fmap.append(x) out = self.last_layer(x) if return_fmaps: return out, fmap else: return out def apply_weight_norm(self): """Apply weight normalization module from all of the layers.""" def _apply_weight_norm(m): if isinstance(m, nn.Conv2d): nn.utils.weight_norm(m) logger.debug(f"Weight norm is applied to {m}.") self.apply(_apply_weight_norm) def apply_spectral_norm(self): """Apply spectral normalization module from all of the layers.""" def _apply_spectral_norm(m): if isinstance(m, nn.Conv2d): nn.utils.spectral_norm(m) logger.debug(f"Spectral norm is applied to {m}.") self.apply(_apply_spectral_norm) class HiFiGANMultiScaleDiscriminator(nn.Module): """HiFi-GAN multi-scale discriminator module.""" def __init__( self, scales=3, downsample_pooling="AvgPool1d", # follow the official implementation setting downsample_pooling_params={ "kernel_size": 4, "stride": 2, "padding": 2, }, discriminator_params={ "in_channels": 1, "out_channels": 1, "kernel_sizes": [15, 41, 5, 3], "channels": 128, "max_downsample_channels": 1024, "max_groups": 16, "bias": True, "downsample_scales": [2, 2, 4, 4, 1], "nonlinear_activation": "LeakyReLU", "nonlinear_activation_params": {"negative_slope": 0.1}, }, follow_official_norm=False, ): """Initialize HiFiGAN multi-scale discriminator module. Args: scales (int): Number of multi-scales. downsample_pooling (str): Pooling module name for downsampling of the inputs. downsample_pooling_params (dict): Parameters for the above pooling module. discriminator_params (dict): Parameters for hifi-gan scale discriminator module. follow_official_norm (bool): Whether to follow the norm setting of the official implementation. The first discriminator uses spectral norm and the other discriminators use weight norm. """ super().__init__() self.discriminators = nn.ModuleList() # add discriminators for i in range(scales): params = copy.deepcopy(discriminator_params) if follow_official_norm: if i == 0: params["use_weight_norm"] = False params["use_spectral_norm"] = True else: params["use_weight_norm"] = True params["use_spectral_norm"] = False self.discriminators += [HiFiGANScaleDiscriminator(**params)] self.pooling = getattr(nn, downsample_pooling)(**downsample_pooling_params) def forward(self, x, return_fmaps=False): """Calculate forward propagation. Args: x (Tensor): Input noise signal (B, 1, T). return_fmaps (bool): Whether to return feature maps. Returns: List: List of list of each discriminator outputs, which consists of each layer output tensors. """ outs, fmaps = [], [] for f in self.discriminators: if return_fmaps: out, fmap = f(x, return_fmaps) fmaps.extend(fmap) else: out = f(x) outs.append(out) x = self.pooling(x) if return_fmaps: return outs, fmaps else: return outs class HiFiGANMultiScaleMultiPeriodDiscriminator(nn.Module): """HiFi-GAN multi-scale + multi-period discriminator module.""" def __init__( self, # Multi-scale discriminator related scales=3, scale_downsample_pooling="AvgPool1d", scale_downsample_pooling_params={ "kernel_size": 4, "stride": 2, "padding": 2, }, scale_discriminator_params={ "in_channels": 1, "out_channels": 1, "kernel_sizes": [15, 41, 5, 3], "channels": 128, "max_downsample_channels": 1024, "max_groups": 16, "bias": True, "downsample_scales": [2, 2, 4, 4, 1], "nonlinear_activation": "LeakyReLU", "nonlinear_activation_params": {"negative_slope": 0.1}, }, follow_official_norm=True, # Multi-period discriminator related periods=[2, 3, 5, 7, 11], period_discriminator_params={ "in_channels": 1, "out_channels": 1, "kernel_sizes": [5, 3], "channels": 32, "downsample_scales": [3, 3, 3, 3, 1], "max_downsample_channels": 1024, "bias": True, "nonlinear_activation": "LeakyReLU", "nonlinear_activation_params": {"negative_slope": 0.1}, "use_weight_norm": True, "use_spectral_norm": False, }, ): """Initialize HiFiGAN multi-scale + multi-period discriminator module. Args: scales (int): Number of multi-scales. scale_downsample_pooling (str): Pooling module name for downsampling of the inputs. scale_downsample_pooling_params (dict): Parameters for the above pooling module. scale_discriminator_params (dict): Parameters for hifi-gan scale discriminator module. follow_official_norm (bool): Whether to follow the norm setting of the official implementation. The first discriminator uses spectral norm and the other discriminators use weight norm. periods (list): List of periods. period_discriminator_params (dict): Parameters for hifi-gan period discriminator module. The period parameter will be overwritten. """ super().__init__() self.msd = HiFiGANMultiScaleDiscriminator( scales=scales, downsample_pooling=scale_downsample_pooling, downsample_pooling_params=scale_downsample_pooling_params, discriminator_params=scale_discriminator_params, follow_official_norm=follow_official_norm, ) self.mpd = HiFiGANMultiPeriodDiscriminator( periods=periods, discriminator_params=period_discriminator_params, ) def forward(self, x, return_fmaps=False): """Calculate forward propagation. Args: x (Tensor): Input noise signal (B, 1, T). return_fmaps (bool): Whether to return feature maps. Returns: List: List of list of each discriminator outputs, which consists of each layer output tensors. Multi scale and multi period ones are concatenated. """ if return_fmaps: msd_outs, msd_fmaps = self.msd(x, return_fmaps) mpd_outs, mpd_fmaps = self.mpd(x, return_fmaps) outs = msd_outs + mpd_outs fmaps = msd_fmaps + mpd_fmaps return outs, fmaps else: msd_outs = self.msd(x) mpd_outs = self.mpd(x) outs = msd_outs + mpd_outs return outs class UnivNetSpectralDiscriminator(nn.Module): """UnivNet spectral discriminator module.""" def __init__( self, fft_size, hop_size, win_length, window="hann_window", kernel_sizes=[(3, 9), (3, 9), (3, 9), (3, 9), (3, 3), (3, 3)], strides=[(1, 1), (1, 2), (1, 2), (1, 2), (1, 1), (1, 1)], channels=32, bias=True, nonlinear_activation="LeakyReLU", nonlinear_activation_params={"negative_slope": 0.2}, use_weight_norm=True, ): """Initialize HiFiGAN scale discriminator module. Args: kernel_sizes (list): List of four kernel sizes. The first will be used for the first conv layer, and the second is for downsampling part, and the remaining two are for output layers. strides (list): channels (int): Initial number of channels for conv layer. bias (bool): Whether to add bias parameter in convolution layers. nonlinear_activation (str): Activation function module name. nonlinear_activation_params (dict): Hyperparameters for activation function. use_weight_norm (bool): Whether to use weight norm. If set to true, it will be applied to all of the conv layers. """ super().__init__() self.fft_size = fft_size self.hop_size = hop_size self.win_length = win_length self.register_buffer("window", getattr(torch, window)(win_length)) self.layers = nn.ModuleList() # check kernel size is valid assert len(kernel_sizes) == len(strides) # add first layer self.layers += [ nn.Sequential( nn.Conv2d( 1, channels, kernel_sizes[0], stride=strides[0], bias=bias, ), getattr(nn, nonlinear_activation)(**nonlinear_activation_params), ) ] for i in range(1, len(kernel_sizes) - 2): self.layers += [ nn.Sequential( nn.Conv2d( channels, channels, kernel_size=kernel_sizes[i], stride=strides[i], bias=bias, ), getattr(nn, nonlinear_activation)(**nonlinear_activation_params), ) ] # add final layers self.layers += [ nn.Sequential( nn.Conv2d( channels, channels, kernel_size=kernel_sizes[-2], stride=strides[-2], bias=bias, ), getattr(nn, nonlinear_activation)(**nonlinear_activation_params), ) ] self.layers += [ nn.Conv2d( channels, 1, kernel_size=kernel_sizes[-1], stride=strides[-1], bias=bias, ) ] # apply weight norm if use_weight_norm: self.apply_weight_norm() def forward(self, x, return_fmaps=False): """Calculate forward propagation. Args: x (Tensor): Input noise signal (B, 1, T). return_fmaps (bool): Whether to return feature maps. Returns: List: List of output tensors of each layer. """ x = spectrogram( x, pad=self.win_length // 2, window=self.window, n_fft=self.fft_size, hop_length=self.hop_size, win_length=self.win_length, power=1.0, normalized=False, ).transpose(-1, -2) fmap = [] for f in self.layers: x = f(x) if return_fmaps: fmap.append(x) if return_fmaps: return x, fmap else: return x def apply_weight_norm(self): """Apply weight normalization module from all of the layers.""" def _apply_weight_norm(m): if isinstance(m, nn.Conv2d): nn.utils.weight_norm(m) logger.debug(f"Weight norm is applied to {m}.") self.apply(_apply_weight_norm) class UnivNetMultiResolutionSpectralDiscriminator(nn.Module): """UnivNet multi-resolution spectral discriminator module.""" def __init__( self, fft_sizes=[1024, 2048, 512], hop_sizes=[120, 240, 50], win_lengths=[600, 1200, 240], window="hann_window", discriminator_params={ "channels": 32, "kernel_sizes": [(3, 9), (3, 9), (3, 9), (3, 9), (3, 3), (3, 3)], "strides": [(1, 1), (1, 2), (1, 2), (1, 2), (1, 1), (1, 1)], "bias": True, "nonlinear_activation": "LeakyReLU", "nonlinear_activation_params": {"negative_slope": 0.2}, }, ): """Initialize UnivNetMultiResolutionSpectralDiscriminator module. Args: scales (int): Number of multi-scales. downsample_pooling (str): Pooling module name for downsampling of the inputs. downsample_pooling_params (dict): Parameters for the above pooling module. discriminator_params (dict): Parameters for hifi-gan scale discriminator module. follow_official_norm (bool): Whether to follow the norm setting of the official implementation. The first discriminator uses spectral norm and the other discriminators use weight norm. """ super().__init__() assert len(fft_sizes) == len(hop_sizes) == len(win_lengths) self.discriminators = nn.ModuleList() # add discriminators for i in range(len(fft_sizes)): params = copy.deepcopy(discriminator_params) self.discriminators += [ UnivNetSpectralDiscriminator( fft_size=fft_sizes[i], hop_size=hop_sizes[i], win_length=win_lengths[i], window=window, **params, ) ] def forward(self, x, return_fmaps=False): """Calculate forward propagation. Args: x (Tensor): Input noise signal (B, 1, T). return_fmaps (bool): Whether to return feature maps. Returns: List: List of list of each discriminator outputs, which consists of each layer output tensors. """ outs, fmaps = [], [] for f in self.discriminators: if return_fmaps: out, fmap = f(x, return_fmaps) fmaps.extend(fmap) else: out = f(x) outs.append(out) if return_fmaps: return outs, fmaps else: return outs class UnivNetMultiResolutionMultiPeriodDiscriminator(nn.Module): """UnivNet multi-resolution + multi-period discriminator module.""" def __init__( self, # Multi-resolution discriminator related fft_sizes=[1024, 2048, 512], hop_sizes=[120, 240, 50], win_lengths=[600, 1200, 240], window="hann_window", spectral_discriminator_params={ "channels": 32, "kernel_sizes": [(3, 9), (3, 9), (3, 9), (3, 9), (3, 3), (3, 3)], "strides": [(1, 1), (1, 2), (1, 2), (1, 2), (1, 1), (1, 1)], "bias": True, "nonlinear_activation": "LeakyReLU", "nonlinear_activation_params": {"negative_slope": 0.2}, }, # Multi-period discriminator related periods=[2, 3, 5, 7, 11], period_discriminator_params={ "in_channels": 1, "out_channels": 1, "kernel_sizes": [5, 3], "channels": 32, "downsample_scales": [3, 3, 3, 3, 1], "max_downsample_channels": 1024, "bias": True, "nonlinear_activation": "LeakyReLU", "nonlinear_activation_params": {"negative_slope": 0.1}, "use_weight_norm": True, "use_spectral_norm": False, }, ): """Initialize UnivNetMultiResolutionMultiPeriodDiscriminator module. Args: sperctral_discriminator_params (dict): Parameters for hifi-gan scale discriminator module. periods (list): List of periods. period_discriminator_params (dict): Parameters for hifi-gan period discriminator module. The period parameter will be overwritten. """ super().__init__() self.mrd = UnivNetMultiResolutionSpectralDiscriminator( fft_sizes=fft_sizes, hop_sizes=hop_sizes, win_lengths=win_lengths, window=window, discriminator_params=spectral_discriminator_params, ) self.mpd = HiFiGANMultiPeriodDiscriminator( periods=periods, discriminator_params=period_discriminator_params, ) def forward(self, x, return_fmaps=False): """Calculate forward propagation. Args: x (Tensor): Input noise signal (B, 1, T). return_fmaps (bool): Whether to return feature maps. Returns: List: List of list of each discriminator outputs, which consists of each layer output tensors. Multi scale and multi period ones are concatenated. """ if return_fmaps: mrd_outs, mrd_fmaps = self.mrd(x, return_fmaps) mpd_outs, mpd_fmaps = self.mpd(x, return_fmaps) outs = mrd_outs + mpd_outs fmaps = mrd_fmaps + mpd_fmaps return outs, fmaps else: mrd_outs = self.mrd(x) mpd_outs = self.mpd(x) outs = mrd_outs + mpd_outs return outs
33,567
33.218145
108
py
nnsvs
nnsvs-master/nnsvs/usfgan/models/generator.py
# -*- coding: utf-8 -*- # Copyright 2022 Reo Yoneyama (Nagoya University) # MIT License (https://opensource.org/licenses/MIT) """Unified Source-Filter GAN Generator modules.""" from logging import getLogger import torch import torch.nn as nn from nnsvs.usfgan.layers import Conv1d1x1, ResidualBlocks, upsample from nnsvs.usfgan.layers.residual_block import PeriodicityEstimator from nnsvs.usfgan.utils import index_initial # A logger for this file logger = getLogger(__name__) class USFGANGenerator(nn.Module): """Unified Source-Filter GAN Generator module.""" def __init__( self, source_network_params={ "blockA": 30, "cycleA": 3, "blockF": 0, "cycleF": 0, "cascade_mode": 0, }, filter_network_params={ "blockA": 0, "cycleA": 0, "blockF": 30, "cycleF": 3, "cascade_mode": 0, }, in_channels=1, out_channels=1, residual_channels=64, gate_channels=128, skip_channels=64, aux_channels=80, aux_context_window=2, use_weight_norm=True, upsample_params={"upsample_scales": [5, 4, 3, 2]}, ): """Initialize USFGANGenerator module. Args: source_network_params (dict): Source-network parameters. filter_network_params (dict): Filter-network parameters. in_channels (int): Number of input channels. out_channels (int): Number of output channels. residual_channels (int): Number of channels in residual conv. gate_channels (int): Number of channels in gated conv. skip_channels (int): Number of channels in skip conv. aux_channels (int): Number of channels for auxiliary feature conv. aux_context_window (int): Context window size for auxiliary feature. use_weight_norm (bool): Whether to use weight norm. If set to true, it will be applied to all of the conv layers. upsample_params (dict): Upsampling network parameters. """ super().__init__() self.in_channels = in_channels self.out_channels = out_channels self.aux_channels = aux_channels self.n_ch = residual_channels # define first convolution self.conv_first = Conv1d1x1(in_channels, residual_channels) # define upsampling network self.upsample_net = getattr(upsample, "ConvInUpsampleNetwork")( **upsample_params, aux_channels=aux_channels, aux_context_window=aux_context_window, ) # define source/filter networks for params in [ source_network_params, filter_network_params, ]: params.update( { "residual_channels": residual_channels, "gate_channels": gate_channels, "skip_channels": skip_channels, "aux_channels": aux_channels, } ) self.source_network = ResidualBlocks(**source_network_params) self.filter_network = ResidualBlocks(**filter_network_params) # convert source signal to hidden representation self.conv_mid = Conv1d1x1(out_channels, skip_channels) # convert hidden representation to output signal self.conv_last = nn.Sequential( nn.ReLU(), Conv1d1x1(skip_channels, skip_channels), nn.ReLU(), Conv1d1x1(skip_channels, out_channels), ) # apply weight norm if use_weight_norm: self.apply_weight_norm() def forward(self, x, c, d): """Calculate forward propagation. Args: x (Tensor): Input noise signal (B, 1, T). c (Tensor): Local conditioning auxiliary features (B, C ,T'). d (Tensor): Input pitch-dependent dilated factors (B, 1, T). Returns: Tensor: Output tensor (B, 1, T) """ # index initialization batch_index, ch_index = index_initial(x.size(0), self.n_ch) # perform upsampling c = self.upsample_net(c) assert c.size(-1) == x.size(-1) # encode to hidden representation x = self.conv_first(x) # source excitation generation x = self.source_network(x, c, d, batch_index, ch_index) s = self.conv_last(x) x = self.conv_mid(s) # resonance filtering x = self.filter_network(x, c, d, batch_index, ch_index) x = self.conv_last(x) return x, s def remove_weight_norm(self): """Remove weight normalization module from all of the layers.""" def _remove_weight_norm(m): try: logger.debug(f"Weight norm is removed from {m}.") nn.utils.remove_weight_norm(m) except ValueError: # this module didn't have weight norm return self.apply(_remove_weight_norm) def apply_weight_norm(self): """Apply weight normalization module from all of the layers.""" def _apply_weight_norm(m): if isinstance(m, nn.Conv1d) or isinstance(m, nn.Conv2d): nn.utils.weight_norm(m) logger.debug(f"Weight norm is applied to {m}.") self.apply(_apply_weight_norm) class CascadeHnUSFGANGenerator(nn.Module): """Cascade hn-uSFGAN Generator module.""" def __init__( self, harmonic_network_params={ "blockA": 20, "cycleA": 4, "blockF": 0, "cycleF": 0, "cascade_mode": 0, }, noise_network_params={ "blockA": 0, "cycleA": 0, "blockF": 5, "cycleF": 5, "cascade_mode": 0, }, filter_network_params={ "blockA": 0, "cycleA": 0, "blockF": 30, "cycleF": 3, "cascade_mode": 0, }, periodicity_estimator_params={ "conv_blocks": 3, "kernel_size": 5, "dilation": 1, "padding_mode": "replicate", }, in_channels=1, out_channels=1, residual_channels=64, gate_channels=128, skip_channels=64, aux_channels=80, aux_context_window=2, use_weight_norm=True, upsample_params={"upsample_scales": [5, 4, 3, 2]}, ): """Initialize CascadeHnUSFGANGenerator module. Args: harmonic_network_params (dict): Periodic source generation network parameters. noise_network_params (dict): Aperiodic source generation network parameters. filter_network_params (dict): Filter network parameters. periodicity_estimator_params (dict): Periodicity estimation network parameters. in_channels (int): Number of input channels. out_channels (int): Number of output channels. residual_channels (int): Number of channels in residual conv. gate_channels (int): Number of channels in gated conv. skip_channels (int): Number of channels in skip conv. aux_channels (int): Number of channels for auxiliary feature conv. aux_context_window (int): Context window size for auxiliary feature. use_weight_norm (bool): Whether to use weight norm. If set to true, it will be applied to all of the conv layers. upsample_params (dict): Upsampling network parameters. """ super(CascadeHnUSFGANGenerator, self).__init__() self.in_channels = in_channels self.out_channels = out_channels self.aux_channels = aux_channels self.n_ch = residual_channels # define first convolution self.conv_first_sine = Conv1d1x1(in_channels, residual_channels) self.conv_first_noise = Conv1d1x1(in_channels, residual_channels) self.conv_merge = Conv1d1x1(residual_channels * 2, residual_channels) # define upsampling network self.upsample_net = getattr(upsample, "ConvInUpsampleNetwork")( **upsample_params, aux_channels=aux_channels, aux_context_window=aux_context_window, ) # define harmonic/noise/filter networks for params in [ harmonic_network_params, noise_network_params, filter_network_params, ]: params.update( { "residual_channels": residual_channels, "gate_channels": gate_channels, "skip_channels": skip_channels, "aux_channels": aux_channels, } ) self.harmonic_network = ResidualBlocks(**harmonic_network_params) self.noise_network = ResidualBlocks(**noise_network_params) self.filter_network = ResidualBlocks(**filter_network_params) # define periodicity estimator self.periodicity_estimator = PeriodicityEstimator( **periodicity_estimator_params, in_channels=aux_channels ) # convert hidden representation to output signal self.conv_last = nn.Sequential( nn.ReLU(), Conv1d1x1(skip_channels, skip_channels), nn.ReLU(), Conv1d1x1(skip_channels, out_channels), ) # apply weight norm if use_weight_norm: self.apply_weight_norm() def forward(self, x, c, d): """Calculate forward propagation. Args: x (Tensor): Input noise signal (B, 1, T). c (Tensor): Local conditioning auxiliary features (B, C ,T'). d (Tensor): Input pitch-dependent dilated factors (B, 1, T). Returns: Tensor: Output tensor (B, 1, T) """ # index initialization batch_index, ch_index = index_initial(x.size(0), self.n_ch) # upsample auxiliary features c = self.upsample_net(c) assert c.size(-1) == x.size(-1) # estimate periodicity a = self.periodicity_estimator(c) # assume the first channel is sine and the other is noise sine, noise = torch.chunk(x, 2, 1) # encode to hidden representation h = self.conv_first_sine(sine) n = self.conv_first_noise(noise) # generate periodic and aperiodic source latent features h = self.harmonic_network(h, c, d, batch_index, ch_index) h = a * h n = self.conv_merge(torch.cat([h, n], dim=1)) n = self.noise_network(n, c, d, batch_index, ch_index) n = (1.0 - a) * n # merge periodic and aperiodic latent features s = h + n # resonance filtering x = self.filter_network(s, c, d, batch_index, ch_index) x = self.conv_last(x) # convert to 1d signal for regularization loss s = self.conv_last(s) # just for debug with torch.no_grad(): h = self.conv_last(h) n = self.conv_last(n) return x, s, h, n, a def remove_weight_norm(self): """Remove weight normalization module from all of the layers.""" def _remove_weight_norm(m): try: logger.debug(f"Weight norm is removed from {m}.") nn.utils.remove_weight_norm(m) except ValueError: # this module didn't have weight norm return self.apply(_remove_weight_norm) def apply_weight_norm(self): """Apply weight normalization module from all of the layers.""" def _apply_weight_norm(m): if isinstance(m, nn.Conv1d) or isinstance(m, nn.Conv2d): nn.utils.weight_norm(m) logger.debug(f"Weight norm is applied to {m}.") self.apply(_apply_weight_norm) class ParallelHnUSFGANGenerator(nn.Module): """Parallel hn-uSFGAN Generator module.""" def __init__( self, harmonic_network_params={ "blockA": 20, "cycleA": 4, "blockF": 0, "cycleF": 0, "cascade_mode": 0, }, noise_network_params={ "blockA": 0, "cycleA": 0, "blockF": 5, "cycleF": 5, "cascade_mode": 0, }, filter_network_params={ "blockA": 0, "cycleA": 0, "blockF": 30, "cycleF": 3, "cascade_mode": 0, }, periodicity_estimator_params={ "conv_blocks": 3, "kernel_size": 5, "dilation": 1, "padding_mode": "replicate", }, in_channels=1, out_channels=1, residual_channels=64, gate_channels=128, skip_channels=64, aux_channels=80, aux_context_window=2, use_weight_norm=True, upsample_params={"upsample_scales": [5, 4, 3, 2]}, ): """Initialize ParallelHnUSFGANGenerator module. Args: harmonic_network_params (dict): Periodic source generation network parameters. noise_network_params (dict): Aperiodic source generation network parameters. filter_network_params (dict): Filter network parameters. periodicity_estimator_params (dict): Periodicity estimation network parameters. in_channels (int): Number of input channels. out_channels (int): Number of output channels. residual_channels (int): Number of channels in residual conv. gate_channels (int): Number of channels in gated conv. skip_channels (int): Number of channels in skip conv. aux_channels (int): Number of channels for auxiliary feature conv. aux_context_window (int): Context window size for auxiliary feature. use_weight_norm (bool): Whether to use weight norm. If set to true, it will be applied to all of the conv layers. upsample_params (dict): Upsampling network parameters. """ super(ParallelHnUSFGANGenerator, self).__init__() self.in_channels = in_channels self.out_channels = out_channels self.aux_channels = aux_channels self.n_ch = residual_channels # define first convolution self.conv_first_sine = Conv1d1x1(in_channels, residual_channels) self.conv_first_noise = Conv1d1x1(in_channels, residual_channels) # define upsampling network self.upsample_net = getattr(upsample, "ConvInUpsampleNetwork")( **upsample_params, aux_channels=aux_channels, aux_context_window=aux_context_window, ) # define harmonic/noise/filter networks for params in [ harmonic_network_params, noise_network_params, filter_network_params, ]: params.update( { "residual_channels": residual_channels, "gate_channels": gate_channels, "skip_channels": skip_channels, "aux_channels": aux_channels, } ) self.harmonic_network = ResidualBlocks(**harmonic_network_params) self.noise_network = ResidualBlocks(**noise_network_params) self.filter_network = ResidualBlocks(**filter_network_params) # define periodicity estimator self.periodicity_estimator = PeriodicityEstimator( **periodicity_estimator_params, in_channels=aux_channels ) # convert hidden representation to output signal self.conv_last = nn.Sequential( nn.ReLU(), Conv1d1x1(skip_channels, skip_channels), nn.ReLU(), Conv1d1x1(skip_channels, out_channels), ) # apply weight norm if use_weight_norm: self.apply_weight_norm() def forward(self, x, c, d): """Calculate forward propagation. Args: x (Tensor): Input noise signal (B, 1, T). c (Tensor): Local conditioning auxiliary features (B, C ,T'). d (Tensor): Input pitch-dependent dilated factors (B, 1, T). Returns: Tensor: Output tensor (B, 1, T) """ # index initialization batch_index, ch_index = index_initial(x.size(0), self.n_ch) # upsample auxiliary features c = self.upsample_net(c) assert c.size(-1) == x.size(-1) # estimate periodicity a = self.periodicity_estimator(c) # assume the first channel is sine and the other is noise sine, noise = torch.chunk(x, 2, 1) # encode to hidden representation h = self.conv_first_sine(sine) n = self.conv_first_noise(noise) # generate periodic and aperiodic source latent features h = self.harmonic_network(h, c, d, batch_index, ch_index) n = self.noise_network(n, c, d, batch_index, ch_index) # merge periodic and aperiodic latent features h = a * h n = (1.0 - a) * n s = h + n # resonance filtering x = self.filter_network(s, c, d, batch_index, ch_index) x = self.conv_last(x) # convert to 1d signal for regularization loss s = self.conv_last(s) # just for debug with torch.no_grad(): h = self.conv_last(h) n = self.conv_last(n) return x, s, h, n, a def remove_weight_norm(self): """Remove weight normalization module from all of the layers.""" def _remove_weight_norm(m): try: logger.debug(f"Weight norm is removed from {m}.") nn.utils.remove_weight_norm(m) except ValueError: # this module didn't have weight norm return self.apply(_remove_weight_norm) def apply_weight_norm(self): """Apply weight normalization module from all of the layers.""" def _apply_weight_norm(m): if isinstance(m, nn.Conv1d) or isinstance(m, nn.Conv2d): nn.utils.weight_norm(m) logger.debug(f"Weight norm is applied to {m}.") self.apply(_apply_weight_norm)
18,402
32.766972
91
py
nnsvs
nnsvs-master/nnsvs/usfgan/layers/residual_block.py
# -*- coding: utf-8 -*- # Copyright 2022 Reo Yoneyama (Nagoya University) # MIT License (https://opensource.org/licenses/MIT) """Residual block modules. References: - https://github.com/bigpon/QPPWG - https://github.com/kan-bayashi/ParallelWaveGAN - https://github.com/r9y9/wavenet_vocoder """ import math import sys from logging import getLogger import torch import torch.nn as nn from nnsvs.usfgan.utils import pd_indexing # A logger for this file logger = getLogger(__name__) class Conv1d(nn.Conv1d): """Conv1d module with customized initialization.""" def __init__(self, *args, **kwargs): """Initialize Conv1d module.""" super(Conv1d, self).__init__(*args, **kwargs) def reset_parameters(self): """Reset parameters.""" nn.init.kaiming_normal_(self.weight, nonlinearity="relu") if self.bias is not None: nn.init.constant_(self.bias, 0.0) class Conv1d1x1(Conv1d): """1x1 Conv1d with customized initialization.""" def __init__(self, in_channels, out_channels, bias=True): """Initialize 1x1 Conv1d module.""" super(Conv1d1x1, self).__init__( in_channels, out_channels, kernel_size=1, padding=0, dilation=1, bias=bias ) class Conv2d(nn.Conv2d): """Conv2d module with customized initialization.""" def __init__(self, *args, **kwargs): """Initialize Conv2d module.""" super(Conv2d, self).__init__(*args, **kwargs) def reset_parameters(self): """Reset parameters.""" nn.init.kaiming_normal_(self.weight, mode="fan_out", nonlinearity="relu") if self.bias is not None: nn.init.constant_(self.bias, 0.0) class Conv2d1x1(Conv2d): """1x1 Conv2d with customized initialization.""" def __init__(self, in_channels, out_channels, bias=True): """Initialize 1x1 Conv2d module.""" super(Conv2d1x1, self).__init__( in_channels, out_channels, kernel_size=1, padding=0, dilation=1, bias=bias ) class FixedBlock(nn.Module): """Fixed block module in QPPWG.""" def __init__( self, residual_channels=64, gate_channels=128, skip_channels=64, aux_channels=80, kernel_size=3, dilation=1, bias=True, ): """Initialize Fixed ResidualBlock module. Args: residual_channels (int): Number of channels for residual connection. skip_channels (int): Number of channels for skip connection. aux_channels (int): Local conditioning channels i.e. auxiliary input dimension. dilation (int): Dilation size. bias (bool): Whether to add bias parameter in convolution layers. """ super(FixedBlock, self).__init__() padding = (kernel_size - 1) // 2 * dilation # dilation conv self.conv = Conv1d( residual_channels, gate_channels, kernel_size, padding=padding, padding_mode="reflect", dilation=dilation, bias=bias, ) # local conditioning if aux_channels > 0: self.conv1x1_aux = Conv1d1x1(aux_channels, gate_channels, bias=False) else: self.conv1x1_aux = None # conv output is split into two groups gate_out_channels = gate_channels // 2 self.conv1x1_out = Conv1d1x1(gate_out_channels, residual_channels, bias=bias) self.conv1x1_skip = Conv1d1x1(gate_out_channels, skip_channels, bias=bias) def forward(self, x, c): """Calculate forward propagation. Args: x (Tensor): Input tensor (B, residual_channels, T). c (Tensor): Local conditioning auxiliary tensor (B, aux_channels, T). Returns: Tensor: Output tensor for residual connection (B, residual_channels, T). Tensor: Output tensor for skip connection (B, skip_channels, T). """ residual = x x = self.conv(x) # split into two part for gated activation splitdim = 1 xa, xb = x.split(x.size(splitdim) // 2, dim=splitdim) # local conditioning if c is not None: assert self.conv1x1_aux is not None c = self.conv1x1_aux(c) ca, cb = c.split(c.size(splitdim) // 2, dim=splitdim) xa, xb = xa + ca, xb + cb x = torch.tanh(xa) * torch.sigmoid(xb) # for skip connection s = self.conv1x1_skip(x) # for residual connection x = (self.conv1x1_out(x) + residual) * math.sqrt(0.5) return x, s class AdaptiveBlock(nn.Module): """Adaptive block module in QPPWG.""" def __init__( self, residual_channels=64, gate_channels=128, skip_channels=64, aux_channels=80, bias=True, ): """Initialize Adaptive ResidualBlock module. Args: residual_channels (int): Number of channels for residual connection. skip_channels (int): Number of channels for skip connection. aux_channels (int): Local conditioning channels i.e. auxiliary input dimension. bias (bool): Whether to add bias parameter in convolution layers. """ super(AdaptiveBlock, self).__init__() # pitch-dependent dilation conv self.convP = Conv1d1x1(residual_channels, gate_channels, bias=bias) # past self.convC = Conv1d1x1(residual_channels, gate_channels, bias=bias) # current self.convF = Conv1d1x1(residual_channels, gate_channels, bias=bias) # future # local conditioning if aux_channels > 0: self.conv1x1_aux = Conv1d1x1(aux_channels, gate_channels, bias=False) else: self.conv1x1_aux = None # conv output is split into two groups gate_out_channels = gate_channels // 2 self.conv1x1_out = Conv1d1x1(gate_out_channels, residual_channels, bias=bias) self.conv1x1_skip = Conv1d1x1(gate_out_channels, skip_channels, bias=bias) def forward(self, xC, xP, xF, c): """Calculate forward propagation. Args: xC (Tensor): Current input tensor (B, residual_channels, T). xP (Tensor): Past input tensor (B, residual_channels, T). xF (Tensor): Future input tensor (B, residual_channels, T). c (Tensor): Local conditioning auxiliary tensor (B, aux_channels, T). Returns: Tensor: Output tensor for residual connection (B, residual_channels, T). Tensor: Output tensor for skip connection (B, skip_channels, T). """ residual = xC x = self.convC(xC) + self.convP(xP) + self.convF(xF) # split into two part for gated activation splitdim = 1 xa, xb = x.split(x.size(splitdim) // 2, dim=splitdim) # local conditioning if c is not None: assert self.conv1x1_aux is not None c = self.conv1x1_aux(c) ca, cb = c.split(c.size(splitdim) // 2, dim=splitdim) xa, xb = xa + ca, xb + cb x = torch.tanh(xa) * torch.sigmoid(xb) # for skip connection s = self.conv1x1_skip(x) # for residual connection x = (self.conv1x1_out(x) + residual) * math.sqrt(0.5) return x, s class ResidualBlocks(nn.Module): """Multiple residual blocks stacking module.""" def __init__( self, blockA, cycleA, blockF, cycleF, cascade_mode=0, residual_channels=64, gate_channels=128, skip_channels=64, aux_channels=80, ): """Initialize ResidualBlocks module. Args: blockA (int): Number of adaptive residual blocks. cycleA (int): Number of dilation cycles of adaptive residual blocks. blockF (int): Number of fixed residual blocks. cycleF (int): Number of dilation cycles of fixed residual blocks. cascade_mode (int): Cascaded mode (0: Adaptive->Fixed; 1: Fixed->Adaptive). residual_channels (int): Number of channels in residual conv. gate_channels (int): Number of channels in gated conv. skip_channels (int): Number of channels in skip conv. aux_channels (int): Number of channels for auxiliary feature conv. """ super(ResidualBlocks, self).__init__() # check the number of blocks and cycles cycleA = max(cycleA, 1) cycleF = max(cycleF, 1) assert blockA % cycleA == 0 self.blockA_per_cycle = blockA // cycleA assert blockF % cycleF == 0 blockF_per_cycle = blockF // cycleF # define adaptive residual blocks adaptive_blocks = nn.ModuleList() for block in range(blockA): conv = AdaptiveBlock( residual_channels=residual_channels, gate_channels=gate_channels, skip_channels=skip_channels, aux_channels=aux_channels, ) adaptive_blocks += [conv] # define fixed residual blocks fixed_blocks = nn.ModuleList() for block in range(blockF): dilation = 2 ** (block % blockF_per_cycle) conv = FixedBlock( residual_channels=residual_channels, gate_channels=gate_channels, skip_channels=skip_channels, aux_channels=aux_channels, dilation=dilation, ) fixed_blocks += [conv] # define cascaded structure if cascade_mode == 0: # adaptive->fixed self.conv_dilated = adaptive_blocks.extend(fixed_blocks) self.block_modes = [True] * blockA + [False] * blockF elif cascade_mode == 1: # fixed->adaptive self.conv_dilated = fixed_blocks.extend(adaptive_blocks) self.block_modes = [False] * blockF + [True] * blockA else: logger.error(f"Cascaded mode {cascade_mode} is not supported!") sys.exit(0) def forward(self, x, c, d, batch_index, ch_index): """Calculate forward propagation. Args: x (Tensor): Input noise signal (B, 1, T). c (Tensor): Local conditioning auxiliary features (B, C ,T). d (Tensor): Input pitch-dependent dilated factors (B, 1, T). Returns: Tensor: Output tensor (B, residual_channels, T). """ skips = 0 blockA_idx = 0 for f, mode in zip(self.conv_dilated, self.block_modes): if mode: # adaptive block dilation = 2 ** (blockA_idx % self.blockA_per_cycle) xP, xF = pd_indexing(x, d, dilation, batch_index, ch_index) x, h = f(x, xP, xF, c) blockA_idx += 1 else: # fixed block x, h = f(x, c) skips = h + skips skips *= math.sqrt(1.0 / len(self.conv_dilated)) return x class PeriodicityEstimator(nn.Module): """Periodicity estimator module.""" def __init__( self, in_channels, residual_channels=64, conv_layers=3, kernel_size=5, dilation=1, padding_mode="replicate", ): """Initialize USFGANGenerator module. Args: in_channels (int): Number of input channels. residual_channels (int): Number of channels in residual conv. conv_layers (int): # Number of convolution layers. kernel_size (int): Kernel size. dilation (int): Dilation size. padding_mode (str): Padding mode. """ super(PeriodicityEstimator, self).__init__() modules = [] for idx in range(conv_layers): conv1d = Conv1d( in_channels, residual_channels, kernel_size=kernel_size, dilation=dilation, padding=kernel_size // 2 * dilation, padding_mode=padding_mode, ) # initialize the initial outputs sigmoid(0)=0.5 to stabilize training if idx != conv_layers - 1: nonlinear = nn.ReLU(inplace=True) else: # NOTE: zero init induces nan or inf if weight normalization is used # nn.init.zeros_(conv1d.weight) nn.init.normal_(conv1d.weight, std=1e-4) nonlinear = nn.Sigmoid() modules += [conv1d, nonlinear] in_channels = residual_channels self.layers = nn.Sequential(*modules) def forward(self, x): """Calculate forward propagation. Args: x (Tensor): Input auxiliary features (B, C ,T). Returns: Tensor: Output tensor (B, residual_channels, T). """ return self.layers(x)
12,956
31.3925
91
py
nnsvs
nnsvs-master/nnsvs/usfgan/layers/cheaptrick.py
# -*- coding: utf-8 -*- # Copyright 2022 Reo Yoneyama (Nagoya University) # MIT License (https://opensource.org/licenses/MIT) """Spectral envelopes estimation module based on CheapTrick. References: - https://www.sciencedirect.com/science/article/pii/S0167639314000697 - https://github.com/mmorise/World """ import math import torch import torch.fft import torch.nn as nn class AdaptiveWindowing(nn.Module): """CheapTrick F0 adptive windowing module.""" def __init__( self, sampling_rate, hop_size, fft_size, f0_floor, f0_ceil, ): """Initialize AdaptiveWindowing module. Args: sampling_rate (int): Sampling rate. hop_size (int): Hop size. fft_size (int): FFT size. f0_floor (int): Minimum value of F0. f0_ceil (int): Maximum value of F0. """ super(AdaptiveWindowing, self).__init__() self.sampling_rate = sampling_rate self.hop_size = hop_size self.fft_size = fft_size self.register_buffer("window", torch.zeros((f0_ceil + 1, fft_size))) self.zero_padding = nn.ConstantPad2d((fft_size // 2, fft_size // 2, 0, 0), 0) # Pre-calculation of the window functions for f0 in range(f0_floor, f0_ceil + 1): half_win_len = round(1.5 * self.sampling_rate / f0) base_index = torch.arange( -half_win_len, half_win_len + 1, dtype=torch.int64 ) position = base_index / 1.5 / self.sampling_rate left = fft_size // 2 - half_win_len right = fft_size // 2 + half_win_len + 1 window = torch.zeros(fft_size) window[left:right] = 0.5 * torch.cos(math.pi * position * f0) + 0.5 average = torch.sum(window * window).pow(0.5) self.window[f0] = window / average def forward(self, x, f, power=False): """Calculate forward propagation. Args: x (Tensor): Waveform (B, fft_size // 2 + 1, T). f (Tensor): F0 sequence (B, T'). power (boot): Whether to use power or magnitude. Returns: Tensor: Power spectrogram (B, bin_size, T'). """ # Get the matrix of window functions corresponding to F0 x = self.zero_padding(x).unfold(1, self.fft_size, self.hop_size) windows = self.window[f] # Adaptive windowing and calculate power spectrogram. # In test, change x[:, : -1, :] to x. x = torch.abs(torch.fft.rfft(x[:, :-1, :] * windows)) x = x.pow(2) if power else x return x class AdaptiveLiftering(nn.Module): """CheapTrick F0 adptive windowing module.""" def __init__( self, sampling_rate, fft_size, f0_floor, f0_ceil, q1=-0.15, ): """Initialize AdaptiveLiftering module. Args: sampling_rate (int): Sampling rate. fft_size (int): FFT size. f0_floor (int): Minimum value of F0. f0_ceil (int): Maximum value of F0. q1 (float): Parameter to remove effect of adjacent harmonics. """ super(AdaptiveLiftering, self).__init__() self.sampling_rate = sampling_rate self.bin_size = fft_size // 2 + 1 self.q1 = q1 self.q0 = 1.0 - 2.0 * q1 self.register_buffer( "smoothing_lifter", torch.zeros((f0_ceil + 1, self.bin_size)) ) self.register_buffer( "compensation_lifter", torch.zeros((f0_ceil + 1, self.bin_size)) ) # Pre-calculation of the smoothing lifters and compensation lifters for f0 in range(f0_floor, f0_ceil + 1): smoothing_lifter = torch.zeros(self.bin_size) compensation_lifter = torch.zeros(self.bin_size) quefrency = torch.arange(1, self.bin_size) / sampling_rate smoothing_lifter[0] = 1.0 smoothing_lifter[1:] = torch.sin(math.pi * f0 * quefrency) / ( math.pi * f0 * quefrency ) compensation_lifter[0] = self.q0 + 2.0 * self.q1 compensation_lifter[1:] = self.q0 + 2.0 * self.q1 * torch.cos( 2.0 * math.pi * f0 * quefrency ) self.smoothing_lifter[f0] = smoothing_lifter self.compensation_lifter[f0] = compensation_lifter def forward(self, x, f, elim_0th=False): """Calculate forward propagation. Args: x (Tensor): Power spectrogram (B, bin_size, T'). f (Tensor): F0 sequence (B, T'). elim_0th (bool): Whether to eliminate cepstram 0th component. Returns: Tensor: Estimated spectral envelope (B, bin_size, T'). """ # Setting the smoothing lifter and compensation lifter smoothing_lifter = self.smoothing_lifter[f] compensation_lifter = self.compensation_lifter[f] # Calculating cepstrum tmp = torch.cat((x, torch.flip(x[:, :, 1:-1], [2])), dim=2) cepstrum = torch.fft.rfft(torch.log(torch.clamp(tmp, min=1e-7))).real # Set the 0th cepstrum to 0 if elim_0th: cepstrum[..., 0] = 0 # Liftering cepstrum with the lifters liftered_cepstrum = cepstrum * smoothing_lifter * compensation_lifter # Return the result to the spectral domain x = torch.fft.irfft(liftered_cepstrum)[:, :, : self.bin_size] return x class CheapTrick(nn.Module): """CheapTrick based spectral envelope estimation module.""" def __init__( self, sampling_rate, hop_size, fft_size, f0_floor=70, f0_ceil=340, uv_threshold=0, q1=-0.15, ): """Initialize AdaptiveLiftering module. Args: sampling_rate (int): Sampling rate. hop_size (int): Hop size. fft_size (int): FFT size. f0_floor (int): Minimum value of F0. f0_ceil (int): Maximum value of F0. uv_threshold (float): V/UV determining threshold. q1 (float): Parameter to remove effect of adjacent harmonics. """ super(CheapTrick, self).__init__() # fft_size must be larger than 3.0 * sampling_rate / f0_floor assert fft_size > 3.0 * sampling_rate / f0_floor self.f0_floor = f0_floor self.f0_ceil = f0_ceil self.uv_threshold = uv_threshold self.ada_wind = AdaptiveWindowing( sampling_rate, hop_size, fft_size, f0_floor, f0_ceil, ) self.ada_lift = AdaptiveLiftering( sampling_rate, fft_size, f0_floor, f0_ceil, q1, ) def forward(self, x, f, power=False, elim_0th=False): """Calculate forward propagation. Args: x (Tensor): Power spectrogram (B, T). f (Tensor): F0 sequence (B, T'). power (boot): Whether to use power or magnitude spectrogram. elim_0th (bool): Whether to eliminate cepstram 0th component. Returns: Tensor: Estimated spectral envelope (B, bin_size, T'). """ # Step0: Round F0 values to integers. voiced = (f > self.uv_threshold) * torch.ones_like(f) f = voiced * f + (1.0 - voiced) * self.f0_ceil f = torch.round(torch.clamp(f, min=self.f0_floor, max=self.f0_ceil)).to( torch.int64 ) # Step1: Adaptive windowing and calculate power or amplitude spectrogram. x = self.ada_wind(x, f, power) # Step3: Smoothing (log axis) and spectral recovery on the cepstrum domain. x = self.ada_lift(x, f, elim_0th) return x
7,834
30.849593
85
py
nnsvs
nnsvs-master/nnsvs/usfgan/layers/upsample.py
# -*- coding: utf-8 -*- """Upsampling module. This code is modified from https://github.com/r9y9/wavenet_vocoder. """ import numpy as np import torch import torch.nn.functional as F from nnsvs.usfgan.layers import Conv1d class Stretch2d(torch.nn.Module): """Stretch2d module.""" def __init__(self, x_scale, y_scale, mode="nearest"): """Initialize Stretch2d module. Args: x_scale (int): X scaling factor (Time axis in spectrogram). y_scale (int): Y scaling factor (Frequency axis in spectrogram). mode (str): Interpolation mode. """ super(Stretch2d, self).__init__() self.x_scale = x_scale self.y_scale = y_scale self.mode = mode def forward(self, x): """Calculate forward propagation. Args: x (Tensor): Input tensor (B, C, F, T). Returns: Tensor: Interpolated tensor (B, C, F * y_scale, T * x_scale) """ return F.interpolate( x, scale_factor=(self.y_scale, self.x_scale), mode=self.mode ) class Conv2d(torch.nn.Conv2d): """Conv2d module with customized initialization.""" def __init__(self, *args, **kwargs): """Initialize Conv2d module.""" super(Conv2d, self).__init__(*args, **kwargs) def reset_parameters(self): """Reset parameters.""" self.weight.data.fill_(1.0 / np.prod(self.kernel_size)) if self.bias is not None: torch.nn.init.constant_(self.bias, 0.0) class UpsampleNetwork(torch.nn.Module): """Upsampling network module.""" def __init__( self, upsample_scales, nonlinear_activation=None, nonlinear_activation_params={}, interpolate_mode="nearest", freq_axis_kernel_size=1, use_causal_conv=False, ): """Initialize upsampling network module. Args: upsample_scales (list): List of upsampling scales. nonlinear_activation (str): Activation function name. nonlinear_activation_params (dict): Arguments for specified activation function. interpolate_mode (str): Interpolation mode. freq_axis_kernel_size (int): Kernel size in the direction of frequency axis. """ super(UpsampleNetwork, self).__init__() self.use_causal_conv = use_causal_conv self.up_layers = torch.nn.ModuleList() for scale in upsample_scales: # interpolation layer stretch = Stretch2d(scale, 1, interpolate_mode) self.up_layers += [stretch] # conv layer assert ( freq_axis_kernel_size - 1 ) % 2 == 0, "Not support even number freq axis kernel size." freq_axis_padding = (freq_axis_kernel_size - 1) // 2 kernel_size = (freq_axis_kernel_size, scale * 2 + 1) if use_causal_conv: padding = (freq_axis_padding, scale * 2) else: padding = (freq_axis_padding, scale) conv = Conv2d(1, 1, kernel_size=kernel_size, padding=padding, bias=False) self.up_layers += [conv] # nonlinear if nonlinear_activation is not None: nonlinear = getattr(torch.nn, nonlinear_activation)( **nonlinear_activation_params ) self.up_layers += [nonlinear] def forward(self, c): """Calculate forward propagation. Args: c : Input tensor (B, C, T). Returns: Tensor: Upsampled tensor (B, C, T'), where T' = T * prod(upsample_scales). """ c = c.unsqueeze(1) # (B, 1, C, T) for f in self.up_layers: if self.use_causal_conv and isinstance(f, Conv2d): c = f(c)[..., : c.size(-1)] else: c = f(c) return c.squeeze(1) # (B, C, T') class ConvInUpsampleNetwork(torch.nn.Module): """Convolution + upsampling network module.""" def __init__( self, upsample_scales, nonlinear_activation=None, nonlinear_activation_params={}, interpolate_mode="nearest", freq_axis_kernel_size=1, aux_channels=80, aux_context_window=0, use_causal_conv=False, ): """Initialize convolution + upsampling network module. Args: upsample_scales (list): List of upsampling scales. nonlinear_activation (str): Activation function name. nonlinear_activation_params (dict): Arguments for specified activation function. mode (str): Interpolation mode. freq_axis_kernel_size (int): Kernel size in the direction of frequency axis. aux_channels (int): Number of channels of pre-convolutional layer. aux_context_window (int): Context window size of the pre-convolutional layer. use_causal_conv (bool): Whether to use causal structure. """ super(ConvInUpsampleNetwork, self).__init__() self.aux_context_window = aux_context_window self.use_causal_conv = use_causal_conv and aux_context_window > 0 # To capture wide-context information in conditional features kernel_size = ( aux_context_window + 1 if use_causal_conv else 2 * aux_context_window + 1 ) # NOTE(kan-bayashi): Here do not use padding because the input is already padded self.conv_in = Conv1d( aux_channels, aux_channels, kernel_size=kernel_size, bias=False ) self.upsample = UpsampleNetwork( upsample_scales=upsample_scales, nonlinear_activation=nonlinear_activation, nonlinear_activation_params=nonlinear_activation_params, interpolate_mode=interpolate_mode, freq_axis_kernel_size=freq_axis_kernel_size, use_causal_conv=use_causal_conv, ) def forward(self, c): """Calculate forward propagation. Args: c : Input tensor (B, C, T'). Returns: Tensor: Upsampled tensor (B, C, T), where T = (T' - aux_context_window * 2) * prod(upsample_scales). Note: The length of inputs considers the context window size. """ c_ = self.conv_in(c) c = c_[:, :, : -self.aux_context_window] if self.use_causal_conv else c_ return self.upsample(c)
6,484
32.25641
92
py
nnsvs
nnsvs-master/nnsvs/usfgan/utils/features.py
# -*- coding: utf-8 -*- # Copyright 2022 Reo Yoneyama (Nagoya University) # MIT License (https://opensource.org/licenses/MIT) """Feature-related functions. References: - https://github.com/bigpon/QPPWG """ import sys from logging import getLogger import numpy as np import torch from torch.nn.functional import interpolate # A logger for this file logger = getLogger(__name__) def validate_length(x, y, hop_size=None): """Validate length Args: x (ndarray): numpy array with x.shape[0] = len_x y (ndarray): numpy array with y.shape[0] = len_y hop_size (int): upsampling factor Returns: (ndarray): length adjusted x with same length y (ndarray): length adjusted y with same length x """ if hop_size is None: if x.shape[0] < y.shape[0]: y = y[: x.shape[0]] if x.shape[0] > y.shape[0]: x = x[: y.shape[0]] assert len(x) == len(y) else: if x.shape[0] > y.shape[0] * hop_size: x = x[: y.shape[0] * hop_size] if x.shape[0] < y.shape[0] * hop_size: mod_y = y.shape[0] * hop_size - x.shape[0] mod_y_frame = mod_y // hop_size + 1 y = y[:-mod_y_frame] x = x[: y.shape[0] * hop_size] assert len(x) == len(y) * hop_size return x, y def dilated_factor(batch_f0, fs, dense_factor): """Pitch-dependent dilated factor Args: batch_f0 (ndarray): the f0 sequence (T) fs (int): sampling rate dense_factor (int): the number of taps in one cycle Return: dilated_factors(np array): float array of the pitch-dependent dilated factors (T) """ batch_f0[batch_f0 == 0] = fs / dense_factor dilated_factors = np.ones(batch_f0.shape) * fs dilated_factors /= batch_f0 dilated_factors /= dense_factor assert np.all(dilated_factors > 0) return dilated_factors class SignalGenerator: """Input signal generator module.""" def __init__( self, sample_rate=24000, hop_size=120, sine_amp=0.1, noise_amp=0.003, signal_types=["sine", "noise"], ): """Initialize WaveNetResidualBlock module. Args: sample_rate (int): Sampling rate. hop_size (int): Hop size of input F0. sine_amp (float): Sine amplitude for NSF-based sine generation. noise_amp (float): Noise amplitude for NSF-based sine generation. signal_types (list): List of input signal types for generator. """ self.sample_rate = sample_rate self.hop_size = hop_size self.signal_types = signal_types self.sine_amp = sine_amp self.noise_amp = noise_amp # for signal_type in signal_types: # if not signal_type in ["noise", "sine", "uv"]: # logger.info(f"{signal_type} is not supported type for generator input.") # sys.exit(0) # logger.info(f"Use {signal_types} for generator input signals.") @torch.no_grad() def __call__(self, f0): signals = [] for typ in self.signal_types: if "noise" == typ: signals.append(self.random_noise(f0)) if "sine" == typ: signals.append(self.sinusoid(f0)) if "uv" == typ: signals.append(self.vuv_binary(f0)) input_batch = signals[0] for signal in signals[1:]: input_batch = torch.cat([input_batch, signal], axis=1) return input_batch @torch.no_grad() def random_noise(self, f0): """Calculate noise signals. Args: f0 (Tensor): F0 tensor (B, 1, T // hop_size). Returns: Tensor: Gaussian noise signals (B, 1, T). """ B, _, T = f0.size() noise = torch.randn((B, 1, T * self.hop_size), device=f0.device) return noise @torch.no_grad() def sinusoid(self, f0): """Calculate sine signals. Args: f0 (Tensor): F0 tensor (B, 1, T // hop_size). Returns: Tensor: Sines generated following NSF (B, 1, T). """ B, _, T = f0.size() vuv = interpolate((f0 > 0) * torch.ones_like(f0), T * self.hop_size) radious = (interpolate(f0, T * self.hop_size) / self.sample_rate) % 1 sine = vuv * torch.sin(torch.cumsum(radious, dim=2) * 2 * np.pi) * self.sine_amp if self.noise_amp > 0: noise_amp = vuv * self.noise_amp + (1.0 - vuv) * self.noise_amp / 3.0 noise = torch.randn((B, 1, T * self.hop_size), device=f0.device) * noise_amp sine = sine + noise return sine @torch.no_grad() def vuv_binary(self, f0): """Calculate V/UV binary sequences. Args: f0 (Tensor): F0 tensor (B, 1, T // hop_size). Returns: Tensor: V/UV binary sequences (B, 1, T). """ _, _, T = f0.size() uv = interpolate((f0 > 0) * torch.ones_like(f0), T * self.hop_size) return uv
5,103
27.198895
90
py
nnsvs
nnsvs-master/nnsvs/usfgan/utils/filters.py
# -*- coding: utf-8 -*- # Copyright 2020 Yi-Chiao Wu (Nagoya University) # based on a WaveNet script by Tomoki Hayashi (Nagoya University) # (https://github.com/kan-bayashi/PytorchWaveNetVocoder) # based on sprocket-vc script by Kazuhiro Kobayashi (Nagoya University) # (https://github.com/k2kobayashi/sprocket) # MIT License (https://opensource.org/licenses/MIT) """Filters.""" import numpy as np from scipy.signal import firwin, lfilter NUMTAPS = 255 def low_cut_filter(x, fs, cutoff=70): """Low-cut filter Args: x (ndarray): Waveform sequence fs (int): Sampling frequency cutoff (float): Cutoff frequency of low cut filter Return: (ndarray): Low cut filtered waveform sequence """ nyquist = fs // 2 norm_cutoff = cutoff / nyquist numtaps = NUMTAPS fil = firwin(numtaps, norm_cutoff, pass_zero=False) lcf_x = lfilter(fil, 1, x) return lcf_x def low_pass_filter(x, fs, cutoff=70): """Low-pass filter Args: x (ndarray): Waveform sequence fs (int): Sampling frequency cutoff (float): Cutoff frequency of low pass filter Return: (ndarray): Low pass filtered waveform sequence """ nyquist = fs // 2 norm_cutoff = cutoff / nyquist numtaps = NUMTAPS fil = firwin(numtaps, norm_cutoff, pass_zero=True) x_pad = np.pad(x, (numtaps, numtaps), "edge") lpf_x = lfilter(fil, 1, x_pad) lpf_x = lpf_x[numtaps + numtaps // 2 : -numtaps // 2] return lpf_x
1,509
24.166667
71
py
nnsvs
nnsvs-master/nnsvs/usfgan/utils/index.py
# -*- coding: utf-8 -*- # Copyright 2020 Yi-Chiao Wu (Nagoya University) # MIT License (https://opensource.org/licenses/MIT) """Indexing-related functions.""" import torch from torch.nn import ConstantPad1d as pad1d def pd_indexing(x, d, dilation, batch_index, ch_index): """Pitch-dependent indexing of past and future samples. Args: x (Tensor): Input feature map (B, C, T). d (Tensor): Input pitch-dependent dilated factors (B, 1, T). dilation (Int): Dilation size. batch_index (Tensor): Batch index ch_index (Tensor): Channel index Returns: Tensor: Past output tensor (B, out_channels, T) Tensor: Future output tensor (B, out_channels, T) """ (_, _, batch_length) = d.size() dilations = d * dilation # get past index idxP = torch.arange(-batch_length, 0).float() if torch.cuda.is_available(): idxP = idxP.cuda() idxP = torch.add(-dilations, idxP) idxP = idxP.round().long() maxP = -((torch.min(idxP) + batch_length)) assert maxP >= 0 idxP = (batch_index, ch_index, idxP) # padding past tensor xP = pad1d((maxP, 0), 0)(x) # get future index idxF = torch.arange(0, batch_length).float() if torch.cuda.is_available(): idxF = idxF.cuda() idxF = torch.add(dilations, idxF) idxF = idxF.round().long() maxF = torch.max(idxF) - (batch_length - 1) assert maxF >= 0 idxF = (batch_index, ch_index, idxF) # padding future tensor xF = pad1d((0, maxF), 0)(x) return xP[idxP], xF[idxF] def index_initial(n_batch, n_ch, tensor=True): """Tensor batch and channel index initialization. Args: n_batch (Int): Number of batch. n_ch (Int): Number of channel. tensor (bool): Return tensor or numpy array Returns: Tensor: Batch index Tensor: Channel index """ batch_index = [] for i in range(n_batch): batch_index.append([[i]] * n_ch) ch_index = [] for i in range(n_ch): ch_index += [[i]] ch_index = [ch_index] * n_batch if tensor: batch_index = torch.tensor(batch_index) ch_index = torch.tensor(ch_index) if torch.cuda.is_available(): batch_index = batch_index.cuda() ch_index = ch_index.cuda() return batch_index, ch_index
2,349
26.647059
68
py
nnsvs
nnsvs-master/nnsvs/transformer/encoder.py
# The code was adapted from https://github.com/jaywalnut310/vits import torch from torch import nn from torch.nn import functional as F from .attentions import MultiHeadAttention, convert_pad_shape class LayerNorm(nn.Module): def __init__(self, channels, eps=1e-5): super().__init__() self.channels = channels self.eps = eps self.gamma = nn.Parameter(torch.ones(channels)) self.beta = nn.Parameter(torch.zeros(channels)) def forward(self, x): x = x.transpose(1, -1) x = F.layer_norm(x, (self.channels,), self.gamma, self.beta, self.eps) return x.transpose(1, -1) class FFN(nn.Module): def __init__( self, in_channels, out_channels, filter_channels, kernel_size, p_dropout=0.0, activation=None, causal=False, ): super().__init__() self.in_channels = in_channels self.out_channels = out_channels self.filter_channels = filter_channels self.kernel_size = kernel_size self.p_dropout = p_dropout self.activation = activation self.causal = causal if causal: self.padding = self._causal_padding else: self.padding = self._same_padding self.conv_1 = nn.Conv1d(in_channels, filter_channels, kernel_size) self.conv_2 = nn.Conv1d(filter_channels, out_channels, kernel_size) self.drop = nn.Dropout(p_dropout) def forward(self, x, x_mask): x = self.conv_1(self.padding(x * x_mask)) if self.activation == "gelu": x = x * torch.sigmoid(1.702 * x) else: x = torch.relu(x) x = self.drop(x) x = self.conv_2(self.padding(x * x_mask)) return x * x_mask def _causal_padding(self, x): if self.kernel_size == 1: return x pad_l = self.kernel_size - 1 pad_r = 0 padding = [[0, 0], [0, 0], [pad_l, pad_r]] x = F.pad(x, convert_pad_shape(padding)) return x def _same_padding(self, x): if self.kernel_size == 1: return x pad_l = (self.kernel_size - 1) // 2 pad_r = self.kernel_size // 2 padding = [[0, 0], [0, 0], [pad_l, pad_r]] x = F.pad(x, convert_pad_shape(padding)) return x class Encoder(nn.Module): def __init__( self, hidden_channels, filter_channels, n_heads, n_layers, kernel_size=1, p_dropout=0.0, window_size=4, **kwargs, ): super().__init__() self.hidden_channels = hidden_channels self.filter_channels = filter_channels self.n_heads = n_heads self.n_layers = n_layers self.kernel_size = kernel_size self.p_dropout = p_dropout self.window_size = window_size self.drop = nn.Dropout(p_dropout) self.attn_layers = nn.ModuleList() self.norm_layers_1 = nn.ModuleList() self.ffn_layers = nn.ModuleList() self.norm_layers_2 = nn.ModuleList() for _ in range(self.n_layers): self.attn_layers.append( MultiHeadAttention( hidden_channels, hidden_channels, n_heads, p_dropout=p_dropout, window_size=window_size, ) ) self.norm_layers_1.append(LayerNorm(hidden_channels)) self.ffn_layers.append( FFN( hidden_channels, hidden_channels, filter_channels, kernel_size, p_dropout=p_dropout, ) ) self.norm_layers_2.append(LayerNorm(hidden_channels)) def forward(self, x, x_mask): attn_mask = x_mask.unsqueeze(2) * x_mask.unsqueeze(-1) x = x * x_mask for i in range(self.n_layers): y = self.attn_layers[i](x, x, attn_mask) y = self.drop(y) x = self.norm_layers_1[i](x + y) y = self.ffn_layers[i](x, x_mask) y = self.drop(y) x = self.norm_layers_2[i](x + y) x = x * x_mask return x
4,282
28.951049
78
py
nnsvs
nnsvs-master/nnsvs/transformer/attentions.py
# The code was adapted from https://github.com/jaywalnut310/vits import math import torch from torch import nn from torch.nn import functional as F def convert_pad_shape(pad_shape): ll = pad_shape[::-1] pad_shape = [item for sublist in ll for item in sublist] return pad_shape def sequence_mask(length, max_length=None): if max_length is None: max_length = length.max() x = torch.arange(max_length, dtype=length.dtype, device=length.device) return x.unsqueeze(0) < length.unsqueeze(1) class MultiHeadAttention(nn.Module): def __init__( self, channels, out_channels, n_heads, p_dropout=0.0, window_size=None, heads_share=True, block_length=None, proximal_bias=False, proximal_init=False, ): super().__init__() assert channels % n_heads == 0 self.channels = channels self.out_channels = out_channels self.n_heads = n_heads self.p_dropout = p_dropout self.window_size = window_size self.heads_share = heads_share self.block_length = block_length self.proximal_bias = proximal_bias self.proximal_init = proximal_init self.attn = None self.k_channels = channels // n_heads self.conv_q = nn.Conv1d(channels, channels, 1) self.conv_k = nn.Conv1d(channels, channels, 1) self.conv_v = nn.Conv1d(channels, channels, 1) self.conv_o = nn.Conv1d(channels, out_channels, 1) self.drop = nn.Dropout(p_dropout) if window_size is not None: n_heads_rel = 1 if heads_share else n_heads rel_stddev = self.k_channels ** -0.5 self.emb_rel_k = nn.Parameter( torch.randn(n_heads_rel, window_size * 2 + 1, self.k_channels) * rel_stddev ) self.emb_rel_v = nn.Parameter( torch.randn(n_heads_rel, window_size * 2 + 1, self.k_channels) * rel_stddev ) nn.init.xavier_uniform_(self.conv_q.weight) nn.init.xavier_uniform_(self.conv_k.weight) nn.init.xavier_uniform_(self.conv_v.weight) if proximal_init: with torch.no_grad(): self.conv_k.weight.copy_(self.conv_q.weight) self.conv_k.bias.copy_(self.conv_q.bias) def forward(self, x, c, attn_mask=None): q = self.conv_q(x) k = self.conv_k(c) v = self.conv_v(c) x, self.attn = self.attention(q, k, v, mask=attn_mask) x = self.conv_o(x) return x def attention(self, query, key, value, mask=None): # reshape [b, d, t] -> [b, n_h, t, d_k] b, d, t_s, t_t = (*key.size(), query.size(2)) query = query.view(b, self.n_heads, self.k_channels, t_t).transpose(2, 3) key = key.view(b, self.n_heads, self.k_channels, t_s).transpose(2, 3) value = value.view(b, self.n_heads, self.k_channels, t_s).transpose(2, 3) scores = torch.matmul(query / math.sqrt(self.k_channels), key.transpose(-2, -1)) if self.window_size is not None: assert ( t_s == t_t ), "Relative attention is only available for self-attention." key_relative_embeddings = self._get_relative_embeddings(self.emb_rel_k, t_s) rel_logits = self._matmul_with_relative_keys( query / math.sqrt(self.k_channels), key_relative_embeddings ) scores_local = self._relative_position_to_absolute_position(rel_logits) scores = scores + scores_local if self.proximal_bias: assert t_s == t_t, "Proximal bias is only available for self-attention." scores = scores + self._attention_bias_proximal(t_s).to( device=scores.device, dtype=scores.dtype ) if mask is not None: scores = scores.masked_fill(mask == 0, -1e4) if self.block_length is not None: assert ( t_s == t_t ), "Local attention is only available for self-attention." block_mask = ( torch.ones_like(scores) .triu(-self.block_length) .tril(self.block_length) ) scores = scores.masked_fill(block_mask == 0, -1e4) p_attn = F.softmax(scores, dim=-1) # [b, n_h, t_t, t_s] p_attn = self.drop(p_attn) output = torch.matmul(p_attn, value) if self.window_size is not None: relative_weights = self._absolute_position_to_relative_position(p_attn) value_relative_embeddings = self._get_relative_embeddings( self.emb_rel_v, t_s ) output = output + self._matmul_with_relative_values( relative_weights, value_relative_embeddings ) output = ( output.transpose(2, 3).contiguous().view(b, d, t_t) ) # [b, n_h, t_t, d_k] -> [b, d, t_t] return output, p_attn def _matmul_with_relative_values(self, x, y): """ x: [b, h, l, m] y: [h or 1, m, d] ret: [b, h, l, d] """ ret = torch.matmul(x, y.unsqueeze(0)) return ret def _matmul_with_relative_keys(self, x, y): """ x: [b, h, l, d] y: [h or 1, m, d] ret: [b, h, l, m] """ ret = torch.matmul(x, y.unsqueeze(0).transpose(-2, -1)) return ret def _get_relative_embeddings(self, relative_embeddings, length): # Pad first before slice to avoid using cond ops. pad_length = max(length - (self.window_size + 1), 0) slice_start_position = max((self.window_size + 1) - length, 0) slice_end_position = slice_start_position + 2 * length - 1 if pad_length > 0: padded_relative_embeddings = F.pad( relative_embeddings, convert_pad_shape([[0, 0], [pad_length, pad_length], [0, 0]]), ) else: padded_relative_embeddings = relative_embeddings used_relative_embeddings = padded_relative_embeddings[ :, slice_start_position:slice_end_position ] return used_relative_embeddings def _relative_position_to_absolute_position(self, x): """ x: [b, h, l, 2*l-1] ret: [b, h, l, l] """ batch, heads, length, _ = x.size() # Concat columns of pad to shift from relative to absolute indexing. x = F.pad(x, convert_pad_shape([[0, 0], [0, 0], [0, 0], [0, 1]])) # Concat extra elements so to add up to shape (len+1, 2*len-1). x_flat = x.view([batch, heads, length * 2 * length]) x_flat = F.pad(x_flat, convert_pad_shape([[0, 0], [0, 0], [0, length - 1]])) # Reshape and slice out the padded elements. x_final = x_flat.view([batch, heads, length + 1, 2 * length - 1])[ :, :, :length, length - 1 : ] return x_final def _absolute_position_to_relative_position(self, x): """ x: [b, h, l, l] ret: [b, h, l, 2*l-1] """ batch, heads, length, _ = x.size() # padd along column x = F.pad(x, convert_pad_shape([[0, 0], [0, 0], [0, 0], [0, length - 1]])) x_flat = x.view([batch, heads, length ** 2 + length * (length - 1)]) # add 0's in the beginning that will skew the elements after reshape x_flat = F.pad(x_flat, convert_pad_shape([[0, 0], [0, 0], [length, 0]])) x_final = x_flat.view([batch, heads, length, 2 * length])[:, :, :, 1:] return x_final def _attention_bias_proximal(self, length): """Bias for self-attention to encourage attention to close positions. Args: length: an integer scalar. Returns: a Tensor with shape [1, 1, length, length] """ r = torch.arange(length, dtype=torch.float32) diff = torch.unsqueeze(r, 0) - torch.unsqueeze(r, 1) return torch.unsqueeze(torch.unsqueeze(-torch.log1p(torch.abs(diff)), 0), 0)
8,150
36.911628
88
py
nnsvs
nnsvs-master/nnsvs/diffsinger/diffusion.py
from collections import deque from functools import partial import numpy as np import torch from nnsvs.base import BaseModel, PredictionType from tqdm import tqdm def extract(a, t, x_shape): b, *_ = t.shape out = a.gather(-1, t) return out.reshape(b, *((1,) * (len(x_shape) - 1))) def noise_like(shape, noise_fn, device, repeat=False): if repeat: resid = [1] * (len(shape) - 1) shape_one = (1, *shape[1:]) return noise_fn(*shape_one, device=device).repeat(shape[0], *resid) else: return noise_fn(*shape, device=device) def linear_beta_schedule(timesteps, min_beta=1e-4, max_beta=0.06): """ linear schedule """ betas = np.linspace(min_beta, max_beta, timesteps) return betas def cosine_beta_schedule(timesteps, s=0.008): """ cosine schedule as proposed in https://openreview.net/forum?id=-NEXDKk8gZ """ steps = timesteps + 1 x = np.linspace(0, steps, steps) alphas_cumprod = np.cos(((x / steps) + s) / (1 + s) * np.pi * 0.5) ** 2 alphas_cumprod = alphas_cumprod / alphas_cumprod[0] betas = 1 - (alphas_cumprod[1:] / alphas_cumprod[:-1]) return np.clip(betas, a_min=0, a_max=0.999) beta_schedule = { "cosine": cosine_beta_schedule, "linear": linear_beta_schedule, } class GaussianDiffusion(BaseModel): def __init__( self, in_dim, out_dim, denoise_fn, encoder=None, K_step=100, betas=None, schedule_type="linear", scheduler_params=None, norm_scale=10, pndm_speedup=None, ): super().__init__() self.in_dim = in_dim self.out_dim = out_dim self.denoise_fn = denoise_fn self.K_step = K_step self.pndm_speedup = pndm_speedup self.encoder = encoder self.norm_scale = norm_scale if scheduler_params is None: if schedule_type == "linear": scheduler_params = {"max_beta": 0.06} elif schedule_type == "cosine": scheduler_params = {"s": 0.008} if encoder is not None: assert encoder.in_dim == in_dim, "encoder input dim must match in_dim" assert out_dim == denoise_fn.in_dim, "denoise_fn input dim must match out_dim" if pndm_speedup: raise NotImplementedError("pndm_speedup is not implemented yet") if betas is not None: betas = ( betas.detach().cpu().numpy() if isinstance(betas, torch.Tensor) else betas ) else: betas = beta_schedule[schedule_type](K_step, **scheduler_params) alphas = 1.0 - betas alphas_cumprod = np.cumprod(alphas, axis=0) alphas_cumprod_prev = np.append(1.0, alphas_cumprod[:-1]) self.noise_list = deque(maxlen=4) to_torch = partial(torch.tensor, dtype=torch.float32) self.register_buffer("betas", to_torch(betas)) self.register_buffer("alphas_cumprod", to_torch(alphas_cumprod)) self.register_buffer("alphas_cumprod_prev", to_torch(alphas_cumprod_prev)) # calculations for diffusion q(x_t | x_{t-1}) and others self.register_buffer("sqrt_alphas_cumprod", to_torch(np.sqrt(alphas_cumprod))) self.register_buffer( "sqrt_one_minus_alphas_cumprod", to_torch(np.sqrt(1.0 - alphas_cumprod)) ) self.register_buffer( "log_one_minus_alphas_cumprod", to_torch(np.log(1.0 - alphas_cumprod)) ) self.register_buffer( "sqrt_recip_alphas_cumprod", to_torch(np.sqrt(1.0 / alphas_cumprod)) ) self.register_buffer( "sqrt_recipm1_alphas_cumprod", to_torch(np.sqrt(1.0 / alphas_cumprod - 1)) ) # calculations for posterior q(x_{t-1} | x_t, x_0) posterior_variance = ( betas * (1.0 - alphas_cumprod_prev) / (1.0 - alphas_cumprod) ) # above: equal to 1. / (1. / (1. - alpha_cumprod_tm1) + alpha_t / beta_t) self.register_buffer("posterior_variance", to_torch(posterior_variance)) # below: log calculation clipped because the posterior variance is 0 # at the beginning of the diffusion chain self.register_buffer( "posterior_log_variance_clipped", to_torch(np.log(np.maximum(posterior_variance, 1e-20))), ) self.register_buffer( "posterior_mean_coef1", to_torch(betas * np.sqrt(alphas_cumprod_prev) / (1.0 - alphas_cumprod)), ) self.register_buffer( "posterior_mean_coef2", to_torch( (1.0 - alphas_cumprod_prev) * np.sqrt(alphas) / (1.0 - alphas_cumprod) ), ) # NOTE: since the original impl. assume the data is distributed in [-1, 1] # let us (roughly) convert N(0,1) noramlized to data to [-1, 1] def _norm(self, x, a_max=10): return x / a_max def _denorm(self, x, a_max=10): return x * a_max def prediction_type(self): return PredictionType.DIFFUSION def q_mean_variance(self, x_start, t): mean = extract(self.sqrt_alphas_cumprod, t, x_start.shape) * x_start variance = extract(1.0 - self.alphas_cumprod, t, x_start.shape) log_variance = extract(self.log_one_minus_alphas_cumprod, t, x_start.shape) return mean, variance, log_variance def predict_start_from_noise(self, x_t, t, noise): return ( extract(self.sqrt_recip_alphas_cumprod, t, x_t.shape) * x_t - extract(self.sqrt_recipm1_alphas_cumprod, t, x_t.shape) * noise ) def q_posterior(self, x_start, x_t, t): posterior_mean = ( extract(self.posterior_mean_coef1, t, x_t.shape) * x_start + extract(self.posterior_mean_coef2, t, x_t.shape) * x_t ) posterior_variance = extract(self.posterior_variance, t, x_t.shape) posterior_log_variance_clipped = extract( self.posterior_log_variance_clipped, t, x_t.shape ) return posterior_mean, posterior_variance, posterior_log_variance_clipped def p_mean_variance(self, x, t, cond, clip_denoised: bool): noise_pred = self.denoise_fn(x, t, cond=cond) x_recon = self.predict_start_from_noise(x, t=t, noise=noise_pred) if clip_denoised: x_recon.clamp_(-1.0, 1.0) model_mean, posterior_variance, posterior_log_variance = self.q_posterior( x_start=x_recon, x_t=x, t=t ) return model_mean, posterior_variance, posterior_log_variance @torch.no_grad() def p_sample( self, x, t, cond, noise_fn=torch.randn, clip_denoised=True, repeat_noise=False ): b, *_, device = *x.shape, x.device model_mean, _, model_log_variance = self.p_mean_variance( x=x, t=t, cond=cond, clip_denoised=clip_denoised ) noise = noise_like(x.shape, noise_fn, device, repeat_noise) # no noise when t == 0 nonzero_mask = (1 - (t == 0).float()).reshape(b, *((1,) * (len(x.shape) - 1))) return model_mean + nonzero_mask * (0.5 * model_log_variance).exp() * noise @torch.no_grad() def p_sample_plms(self, x, t, interval, cond): """ Use the PLMS method from Pseudo Numerical Methods for Diffusion Models on Manifolds https://arxiv.org/abs/2202.09778. """ def get_x_pred(x, noise_t, t): a_t = extract(self.alphas_cumprod, t, x.shape) a_prev = extract( self.alphas_cumprod, torch.max(t - interval, torch.zeros_like(t)), x.shape, ) a_t_sq, a_prev_sq = a_t.sqrt(), a_prev.sqrt() x_delta = (a_prev - a_t) * ( (1 / (a_t_sq * (a_t_sq + a_prev_sq))) * x - 1 / (a_t_sq * (((1 - a_prev) * a_t).sqrt() + ((1 - a_t) * a_prev).sqrt())) * noise_t ) x_pred = x + x_delta return x_pred noise_list = self.noise_list noise_pred = self.denoise_fn(x, t, cond=cond) if len(noise_list) == 0: x_pred = get_x_pred(x, noise_pred, t) # noise_pred_prev = self.denoise_fn(x_pred, max(t - interval, 0), cond=cond) noise_pred_prev = self.denoise_fn( x_pred, torch.max(t - interval, torch.zeros_like(t)), cond=cond ) noise_pred_prime = (noise_pred + noise_pred_prev) / 2 elif len(noise_list) == 1: noise_pred_prime = (3 * noise_pred - noise_list[-1]) / 2 elif len(noise_list) == 2: noise_pred_prime = ( 23 * noise_pred - 16 * noise_list[-1] + 5 * noise_list[-2] ) / 12 elif len(noise_list) >= 3: noise_pred_prime = ( 55 * noise_pred - 59 * noise_list[-1] + 37 * noise_list[-2] - 9 * noise_list[-3] ) / 24 x_prev = get_x_pred(x, noise_pred_prime, t) noise_list.append(noise_pred) return x_prev def q_sample(self, x_start, t, noise=None): if noise is None: noise = torch.randn_like(x_start) return ( extract(self.sqrt_alphas_cumprod, t, x_start.shape) * x_start + extract(self.sqrt_one_minus_alphas_cumprod, t, x_start.shape) * noise ) def forward(self, cond, lengths=None, y=None): """Forward step Args: cond (torch.Tensor): conditioning features of shaep (B, T, encoder_hidden_dim) lengths (torch.Tensor): lengths of each sequence in the batch y (torch.Tensor): ground truth of shape (B, T, C) Returns: tuple of tensors (B, T, in_dim), (B, T, in_dim) """ B = cond.shape[0] device = cond.device if self.encoder is not None: cond = self.encoder(cond, lengths) # (B, M, T) cond = cond.transpose(1, 2) t = torch.randint(0, self.K_step, (B,), device=device).long() x = self._norm(y, self.norm_scale) x = x.transpose(1, 2)[:, None, :, :] # [B, 1, M, T] noise = torch.randn_like(x) x_noisy = self.q_sample(x_start=x, t=t, noise=noise) x_recon = self.denoise_fn(x_noisy, t, cond) noise = noise.squeeze(1).transpose(1, 2) x_recon = x_recon.squeeze(1).transpose(1, 2) return noise, x_recon def inference(self, cond, lengths=None): B = cond.shape[0] device = cond.device if self.encoder is not None: cond = self.encoder(cond, lengths) # (B, M, T) cond = cond.transpose(1, 2) t = self.K_step shape = (cond.shape[0], 1, self.out_dim, cond.shape[2]) x = torch.randn(shape, device=device) if self.pndm_speedup: self.noise_list = deque(maxlen=4) iteration_interval = int(self.pndm_speedup) for i in tqdm( reversed(range(0, t, iteration_interval)), desc="sample time step", total=t // iteration_interval, ): x = self.p_sample_plms( x, torch.full((B,), i, device=device, dtype=torch.long), iteration_interval, cond, ) else: for i in tqdm(reversed(range(0, t)), desc="sample time step", total=t): x = self.p_sample( x, torch.full((B,), i, device=device, dtype=torch.long), cond ) x = self._denorm(x[:, 0].transpose(1, 2), self.norm_scale) return x
11,770
33.928783
91
py
nnsvs
nnsvs-master/nnsvs/diffsinger/fs2.py
import math import torch from nnsvs.base import BaseModel from nnsvs.util import make_pad_mask from torch import nn from torch.nn import Parameter from torch.nn import functional as F def softmax(x, dim): return F.softmax(x, dim=dim, dtype=torch.float32) class PositionalEncoding(torch.nn.Module): """Positional encoding. Args: d_model (int): Embedding dimension. dropout_rate (float): Dropout rate. max_len (int): Maximum input length. reverse (bool): Whether to reverse the input position. """ def __init__(self, d_model, dropout_rate, max_len=5000, reverse=False): """Construct an PositionalEncoding object.""" super(PositionalEncoding, self).__init__() self.d_model = d_model self.reverse = reverse self.xscale = math.sqrt(self.d_model) self.dropout = torch.nn.Dropout(p=dropout_rate) self.pe = None self.extend_pe(torch.tensor(0.0).expand(1, max_len)) def extend_pe(self, x): """Reset the positional encodings.""" if self.pe is not None: if self.pe.size(1) >= x.size(1): if self.pe.dtype != x.dtype or self.pe.device != x.device: self.pe = self.pe.to(dtype=x.dtype, device=x.device) return pe = torch.zeros(x.size(1), self.d_model) if self.reverse: position = torch.arange( x.size(1) - 1, -1, -1.0, dtype=torch.float32 ).unsqueeze(1) else: position = torch.arange(0, x.size(1), dtype=torch.float32).unsqueeze(1) div_term = torch.exp( torch.arange(0, self.d_model, 2, dtype=torch.float32) * -(math.log(10000.0) / self.d_model) ) pe[:, 0::2] = torch.sin(position * div_term) pe[:, 1::2] = torch.cos(position * div_term) pe = pe.unsqueeze(0) self.pe = pe.to(device=x.device, dtype=x.dtype) def forward(self, x: torch.Tensor): """Add positional encoding. Args: x (torch.Tensor): Input tensor (batch, time, `*`). Returns: torch.Tensor: Encoded tensor (batch, time, `*`). """ self.extend_pe(x) x = x * self.xscale + self.pe[:, : x.size(1)] return self.dropout(x) class RelPositionalEncoding(PositionalEncoding): """Relative positional encoding module. See : Appendix B in https://arxiv.org/abs/1901.02860 Args: d_model (int): Embedding dimension. dropout_rate (float): Dropout rate. max_len (int): Maximum input length. """ def __init__(self, d_model, dropout_rate, max_len=5000): """Initialize class.""" super().__init__(d_model, dropout_rate, max_len, reverse=True) def forward(self, x): """Compute positional encoding. Args: x (torch.Tensor): Input tensor (batch, time, `*`). Returns: torch.Tensor: Encoded tensor (batch, time, `*`). torch.Tensor: Positional embedding tensor (1, time, `*`). """ self.extend_pe(x) x = x * self.xscale pos_emb = self.pe[:, : x.size(1)] return self.dropout(x) + self.dropout(pos_emb) class MultiheadAttention(nn.Module): def __init__( self, embed_dim, num_heads, kdim=None, vdim=None, dropout=0.0, bias=True, add_bias_kv=False, add_zero_attn=False, self_attention=False, encoder_decoder_attention=False, ): super().__init__() self.embed_dim = embed_dim self.kdim = kdim if kdim is not None else embed_dim self.vdim = vdim if vdim is not None else embed_dim self.qkv_same_dim = self.kdim == embed_dim and self.vdim == embed_dim self.num_heads = num_heads self.dropout = dropout self.head_dim = embed_dim // num_heads assert ( self.head_dim * num_heads == self.embed_dim ), "embed_dim must be divisible by num_heads" self.scaling = self.head_dim ** -0.5 self.self_attention = self_attention self.encoder_decoder_attention = encoder_decoder_attention assert not self.self_attention or self.qkv_same_dim, ( "Self-attention requires query, key and " "value to be of the same size" ) if self.qkv_same_dim: self.in_proj_weight = Parameter(torch.Tensor(3 * embed_dim, embed_dim)) else: self.k_proj_weight = Parameter(torch.Tensor(embed_dim, self.kdim)) self.v_proj_weight = Parameter(torch.Tensor(embed_dim, self.vdim)) self.q_proj_weight = Parameter(torch.Tensor(embed_dim, embed_dim)) if bias: self.in_proj_bias = Parameter(torch.Tensor(3 * embed_dim)) else: self.register_parameter("in_proj_bias", None) self.out_proj = nn.Linear(embed_dim, embed_dim, bias=bias) if add_bias_kv: self.bias_k = Parameter(torch.Tensor(1, 1, embed_dim)) self.bias_v = Parameter(torch.Tensor(1, 1, embed_dim)) else: self.bias_k = self.bias_v = None self.add_zero_attn = add_zero_attn self.reset_parameters() self.enable_torch_version = False if hasattr(F, "multi_head_attention_forward"): self.enable_torch_version = True else: self.enable_torch_version = False self.last_attn_probs = None def reset_parameters(self): if self.qkv_same_dim: nn.init.xavier_uniform_(self.in_proj_weight) else: nn.init.xavier_uniform_(self.k_proj_weight) nn.init.xavier_uniform_(self.v_proj_weight) nn.init.xavier_uniform_(self.q_proj_weight) nn.init.xavier_uniform_(self.out_proj.weight) if self.in_proj_bias is not None: nn.init.constant_(self.in_proj_bias, 0.0) nn.init.constant_(self.out_proj.bias, 0.0) if self.bias_k is not None: nn.init.xavier_normal_(self.bias_k) if self.bias_v is not None: nn.init.xavier_normal_(self.bias_v) def forward( self, query, key, value, key_padding_mask=None, incremental_state=None, need_weights=True, static_kv=False, attn_mask=None, before_softmax=False, need_head_weights=False, enc_dec_attn_constraint_mask=None, reset_attn_weight=None, ): """Input shape: Time x Batch x Channel Args: key_padding_mask (ByteTensor, optional): mask to exclude keys that are pads, of shape `(batch, src_len)`, where padding elements are indicated by 1s. need_weights (bool, optional): return the attention weights, averaged over heads (default: False). attn_mask (ByteTensor, optional): typically used to implement causal attention, where the mask prevents the attention from looking forward in time (default: None). before_softmax (bool, optional): return the raw attention weights and values before the attention softmax. need_head_weights (bool, optional): return the attention weights for each head. Implies *need_weights*. Default: return the average attention weights over all heads. """ if need_head_weights: need_weights = True tgt_len, bsz, embed_dim = query.size() assert embed_dim == self.embed_dim assert list(query.size()) == [tgt_len, bsz, embed_dim] if ( self.enable_torch_version and incremental_state is None and not static_kv and reset_attn_weight is None ): if self.qkv_same_dim: return F.multi_head_attention_forward( query, key, value, self.embed_dim, self.num_heads, self.in_proj_weight, self.in_proj_bias, self.bias_k, self.bias_v, self.add_zero_attn, self.dropout, self.out_proj.weight, self.out_proj.bias, self.training, key_padding_mask, need_weights, attn_mask, ) else: return F.multi_head_attention_forward( query, key, value, self.embed_dim, self.num_heads, torch.empty([0]), self.in_proj_bias, self.bias_k, self.bias_v, self.add_zero_attn, self.dropout, self.out_proj.weight, self.out_proj.bias, self.training, key_padding_mask, need_weights, attn_mask, use_separate_proj_weight=True, q_proj_weight=self.q_proj_weight, k_proj_weight=self.k_proj_weight, v_proj_weight=self.v_proj_weight, ) if incremental_state is not None: raise NotImplementedError() else: saved_state = None if self.self_attention: # self-attention q, k, v = self.in_proj_qkv(query) elif self.encoder_decoder_attention: # encoder-decoder attention q = self.in_proj_q(query) if key is None: assert value is None k = v = None else: k = self.in_proj_k(key) v = self.in_proj_v(key) else: q = self.in_proj_q(query) k = self.in_proj_k(key) v = self.in_proj_v(value) q *= self.scaling if self.bias_k is not None: assert self.bias_v is not None k = torch.cat([k, self.bias_k.repeat(1, bsz, 1)]) v = torch.cat([v, self.bias_v.repeat(1, bsz, 1)]) if attn_mask is not None: attn_mask = torch.cat( [attn_mask, attn_mask.new_zeros(attn_mask.size(0), 1)], dim=1 ) if key_padding_mask is not None: key_padding_mask = torch.cat( [ key_padding_mask, key_padding_mask.new_zeros(key_padding_mask.size(0), 1), ], dim=1, ) q = ( q.contiguous() .view(tgt_len, bsz * self.num_heads, self.head_dim) .transpose(0, 1) ) if k is not None: k = ( k.contiguous() .view(-1, bsz * self.num_heads, self.head_dim) .transpose(0, 1) ) if v is not None: v = ( v.contiguous() .view(-1, bsz * self.num_heads, self.head_dim) .transpose(0, 1) ) if saved_state is not None: raise NotImplementedError() src_len = k.size(1) # This is part of a workaround to get around fork/join parallelism # not supporting Optional types. if key_padding_mask is not None and key_padding_mask.shape == torch.Size([]): key_padding_mask = None if key_padding_mask is not None: assert key_padding_mask.size(0) == bsz assert key_padding_mask.size(1) == src_len if self.add_zero_attn: src_len += 1 k = torch.cat([k, k.new_zeros((k.size(0), 1) + k.size()[2:])], dim=1) v = torch.cat([v, v.new_zeros((v.size(0), 1) + v.size()[2:])], dim=1) if attn_mask is not None: attn_mask = torch.cat( [attn_mask, attn_mask.new_zeros(attn_mask.size(0), 1)], dim=1 ) if key_padding_mask is not None: key_padding_mask = torch.cat( [ key_padding_mask, torch.zeros(key_padding_mask.size(0), 1).type_as( key_padding_mask ), ], dim=1, ) attn_weights = torch.bmm(q, k.transpose(1, 2)) attn_weights = self.apply_sparse_mask(attn_weights, tgt_len, src_len, bsz) assert list(attn_weights.size()) == [bsz * self.num_heads, tgt_len, src_len] if attn_mask is not None: if len(attn_mask.shape) == 2: attn_mask = attn_mask.unsqueeze(0) elif len(attn_mask.shape) == 3: attn_mask = ( attn_mask[:, None] .repeat([1, self.num_heads, 1, 1]) .reshape(bsz * self.num_heads, tgt_len, src_len) ) attn_weights = attn_weights + attn_mask if enc_dec_attn_constraint_mask is not None: # bs x head x L_kv attn_weights = attn_weights.view(bsz, self.num_heads, tgt_len, src_len) attn_weights = attn_weights.masked_fill( enc_dec_attn_constraint_mask.unsqueeze(2).bool(), -1e9, ) attn_weights = attn_weights.view(bsz * self.num_heads, tgt_len, src_len) if key_padding_mask is not None: # don't attend to padding symbols attn_weights = attn_weights.view(bsz, self.num_heads, tgt_len, src_len) attn_weights = attn_weights.masked_fill( key_padding_mask.unsqueeze(1).unsqueeze(2), -1e9, ) attn_weights = attn_weights.view(bsz * self.num_heads, tgt_len, src_len) attn_logits = attn_weights.view(bsz, self.num_heads, tgt_len, src_len) if before_softmax: return attn_weights, v attn_weights_float = softmax(attn_weights, dim=-1) attn_weights = attn_weights_float.type_as(attn_weights) attn_probs = F.dropout( attn_weights_float.type_as(attn_weights), p=self.dropout, training=self.training, ) if reset_attn_weight is not None: if reset_attn_weight: self.last_attn_probs = attn_probs.detach() else: assert self.last_attn_probs is not None attn_probs = self.last_attn_probs attn = torch.bmm(attn_probs, v) assert list(attn.size()) == [bsz * self.num_heads, tgt_len, self.head_dim] attn = attn.transpose(0, 1).contiguous().view(tgt_len, bsz, embed_dim) attn = self.out_proj(attn) if need_weights: attn_weights = attn_weights_float.view( bsz, self.num_heads, tgt_len, src_len ).transpose(1, 0) if not need_head_weights: # average attention weights over heads attn_weights = attn_weights.mean(dim=0) else: attn_weights = None return attn, (attn_weights, attn_logits) def in_proj_qkv(self, query): return self._in_proj(query).chunk(3, dim=-1) def in_proj_q(self, query): if self.qkv_same_dim: return self._in_proj(query, end=self.embed_dim) else: bias = self.in_proj_bias if bias is not None: bias = bias[: self.embed_dim] return F.linear(query, self.q_proj_weight, bias) def in_proj_k(self, key): if self.qkv_same_dim: return self._in_proj(key, start=self.embed_dim, end=2 * self.embed_dim) else: weight = self.k_proj_weight bias = self.in_proj_bias if bias is not None: bias = bias[self.embed_dim : 2 * self.embed_dim] return F.linear(key, weight, bias) def in_proj_v(self, value): if self.qkv_same_dim: return self._in_proj(value, start=2 * self.embed_dim) else: weight = self.v_proj_weight bias = self.in_proj_bias if bias is not None: bias = bias[2 * self.embed_dim :] return F.linear(value, weight, bias) def _in_proj(self, input, start=0, end=None): weight = self.in_proj_weight bias = self.in_proj_bias weight = weight[start:end, :] if bias is not None: bias = bias[start:end] return F.linear(input, weight, bias) def apply_sparse_mask(self, attn_weights, tgt_len, src_len, bsz): return attn_weights def LayerNorm(normalized_shape, eps=1e-5, elementwise_affine=True): return torch.nn.LayerNorm(normalized_shape, eps, elementwise_affine) class BatchNorm1dTBC(nn.Module): def __init__(self, c): super(BatchNorm1dTBC, self).__init__() self.bn = nn.BatchNorm1d(c) def forward(self, x): """ :param x: [T, B, C] :return: [T, B, C] """ x = x.permute(1, 2, 0) # [B, C, T] x = self.bn(x) # [B, C, T] x = x.permute(2, 0, 1) # [T, B, C] return x def Linear(in_features, out_features, bias=True): m = nn.Linear(in_features, out_features, bias) nn.init.xavier_uniform_(m.weight) if bias: nn.init.constant_(m.bias, 0.0) return m class Swish(torch.autograd.Function): @staticmethod def forward(ctx, i): result = i * torch.sigmoid(i) ctx.save_for_backward(i) return result @staticmethod def backward(ctx, grad_output): i = ctx.saved_variables[0] sigmoid_i = torch.sigmoid(i) return grad_output * (sigmoid_i * (1 + i * (1 - sigmoid_i))) class CustomSwish(nn.Module): def forward(self, input_tensor): return Swish.apply(input_tensor) class TransformerFFNLayer(nn.Module): def __init__( self, hidden_size, filter_size, padding="SAME", kernel_size=1, dropout=0.0, act="gelu", ): super().__init__() self.kernel_size = kernel_size self.dropout = dropout self.act = act if padding == "SAME": self.ffn_1 = nn.Conv1d( hidden_size, filter_size, kernel_size, padding=kernel_size // 2 ) elif padding == "LEFT": self.ffn_1 = nn.Sequential( nn.ConstantPad1d((kernel_size - 1, 0), 0.0), nn.Conv1d(hidden_size, filter_size, kernel_size), ) self.ffn_2 = Linear(filter_size, hidden_size) if self.act == "swish": self.swish_fn = CustomSwish() def forward(self, x, incremental_state=None): # x: T x B x C if incremental_state is not None: assert incremental_state is None, "Nar-generation does not allow this." x = self.ffn_1(x.permute(1, 2, 0)).permute(2, 0, 1) x = x * self.kernel_size ** -0.5 if incremental_state is not None: x = x[-1:] if self.act == "gelu": x = F.gelu(x) if self.act == "relu": x = F.relu(x) if self.act == "swish": x = self.swish_fn(x) x = F.dropout(x, self.dropout, training=self.training) x = self.ffn_2(x) return x class EncSALayer(nn.Module): def __init__( self, c, num_heads, dropout, attention_dropout=0.1, relu_dropout=0.1, kernel_size=9, padding="SAME", norm="ln", act="gelu", ): super().__init__() self.c = c self.dropout = dropout self.num_heads = num_heads if num_heads > 0: if norm == "ln": self.layer_norm1 = LayerNorm(c) elif norm == "bn": self.layer_norm1 = BatchNorm1dTBC(c) self.self_attn = MultiheadAttention( self.c, num_heads, self_attention=True, dropout=attention_dropout, bias=False, ) if norm == "ln": self.layer_norm2 = LayerNorm(c) elif norm == "bn": self.layer_norm2 = BatchNorm1dTBC(c) self.ffn = TransformerFFNLayer( c, 4 * c, kernel_size=kernel_size, dropout=relu_dropout, padding=padding, act=act, ) def forward(self, x, encoder_padding_mask=None, **kwargs): layer_norm_training = kwargs.get("layer_norm_training", None) if layer_norm_training is not None: self.layer_norm1.training = layer_norm_training self.layer_norm2.training = layer_norm_training if self.num_heads > 0: residual = x x = self.layer_norm1(x) x, _, = self.self_attn( query=x, key=x, value=x, key_padding_mask=encoder_padding_mask ) x = F.dropout(x, self.dropout, training=self.training) x = residual + x x = x * (1 - encoder_padding_mask.float()).transpose(0, 1)[..., None] residual = x x = self.layer_norm2(x) x = self.ffn(x) x = F.dropout(x, self.dropout, training=self.training) x = residual + x x = x * (1 - encoder_padding_mask.float()).transpose(0, 1)[..., None] return x class TransformerEncoderLayer(nn.Module): def __init__( self, hidden_size, dropout, kernel_size=None, num_heads=2, norm="ln", padding="SAME", act="gelu", ): super().__init__() self.hidden_size = hidden_size self.dropout = dropout self.num_heads = num_heads self.op = EncSALayer( hidden_size, num_heads, dropout=dropout, attention_dropout=0.0, relu_dropout=dropout, kernel_size=kernel_size, padding=padding, norm=norm, act=act, ) def forward(self, x, **kwargs): return self.op(x, **kwargs) class FFTBlocks(nn.Module): def __init__( self, hidden_size, num_layers, ffn_kernel_size=9, dropout=0.1, num_heads=2, use_pos_embed=True, use_last_norm=True, norm="ln", use_pos_embed_alpha=True, ): super().__init__() self.num_layers = num_layers embed_dim = self.hidden_size = hidden_size self.dropout = dropout self.use_pos_embed = use_pos_embed self.use_last_norm = use_last_norm if use_pos_embed: self.padding_idx = 0 self.pos_embed_alpha = ( nn.Parameter(torch.Tensor([1])) if use_pos_embed_alpha else 1 ) self.embed_positions = RelPositionalEncoding(hidden_size, dropout_rate=0.0) self.layers = nn.ModuleList([]) self.layers.extend( [ TransformerEncoderLayer( self.hidden_size, self.dropout, kernel_size=ffn_kernel_size, num_heads=num_heads, ) for _ in range(self.num_layers) ] ) if self.use_last_norm: if norm == "ln": self.layer_norm = nn.LayerNorm(embed_dim) elif norm == "bn": self.layer_norm = BatchNorm1dTBC(embed_dim) else: self.layer_norm = None def forward(self, x, lengths, padding_mask=None): """ :param x: [B, T, C] :param padding_mask: [B, T] :return: [B, T, C] or [L, B, T, C] """ if padding_mask is None: padding_mask = make_pad_mask(lengths).to(x.device) # padding_mask = x.abs().sum(-1).eq(0).data if padding_mask is None else padding_mask nonpadding_mask_TB = ( 1 - padding_mask.transpose(0, 1).float()[:, :, None] ) # [T, B, 1] if self.use_pos_embed: # positions = self.pos_embed_alpha * self.embed_positions(x[..., 0]) positions = self.pos_embed_alpha * self.embed_positions(x) x = x + positions x = F.dropout(x, p=self.dropout, training=self.training) # B x T x C -> T x B x C x = x.transpose(0, 1) * nonpadding_mask_TB hiddens = [] for layer in self.layers: x = layer(x, encoder_padding_mask=padding_mask) * nonpadding_mask_TB hiddens.append(x) if self.use_last_norm: x = self.layer_norm(x) * nonpadding_mask_TB x = x.transpose(0, 1) # [B, T, C] return x class FFTBlocksEncoder(BaseModel): def __init__( self, in_dim, out_dim, hidden_dim, num_layers=2, ffn_kernel_size=9, dropout=0.1, num_heads=2, use_pos_embed=True, use_last_norm=True, norm="ln", use_pos_embed_alpha=True, reduction_factor=1, downsample_by_conv=True, in_ph_start_idx: int = 1, in_ph_end_idx: int = 50, embed_dim=None, ): super().__init__() self.in_dim = in_dim self.out_dim = out_dim self.in_ph_start_idx = in_ph_start_idx self.in_ph_end_idx = in_ph_end_idx self.num_vocab = in_ph_end_idx - in_ph_start_idx self.embed_dim = embed_dim self.reduction_factor = reduction_factor if self.embed_dim is not None: assert in_dim > self.num_vocab self.emb = nn.Embedding(self.num_vocab, embed_dim) self.fc_in = nn.Linear(in_dim - self.num_vocab, embed_dim) self.fc = nn.Linear(embed_dim, hidden_dim) else: self.emb = None self.fc_in = None self.fc = nn.Linear(in_dim, hidden_dim) if reduction_factor > 1 and downsample_by_conv: self.conv_downsample = nn.Conv1d( in_dim, in_dim, kernel_size=reduction_factor, stride=reduction_factor, groups=in_dim, ) else: self.conv_downsample = None self.encoder = FFTBlocks( hidden_size=hidden_dim, num_layers=num_layers, ffn_kernel_size=ffn_kernel_size, dropout=dropout, num_heads=num_heads, use_pos_embed=use_pos_embed, use_last_norm=use_last_norm, norm=norm, use_pos_embed_alpha=use_pos_embed_alpha, ) self.fc_out = nn.Linear(hidden_dim, out_dim * reduction_factor) def forward(self, x, lengths, y=None): if self.embed_dim is not None: x_first, x_ph_onehot, x_last = torch.split( x, [ self.in_ph_start_idx, self.num_vocab, self.in_dim - self.num_vocab - self.in_ph_start_idx, ], dim=-1, ) x_ph = torch.argmax(x_ph_onehot, dim=-1) # Make sure to have one-hot vector assert (x_ph_onehot.sum(-1) <= 1).all() x = self.emb(x_ph) + self.fc_in(torch.cat([x_first, x_last], dim=-1)) # Adjust lengths based on the reduction factor if self.reduction_factor > 1: lengths = (lengths / self.reduction_factor).long() if self.conv_downsample is not None: x = self.conv_downsample(x.transpose(1, 2)).transpose(1, 2) else: x = x[:, self.reduction_factor - 1 :: self.reduction_factor] x = self.fc(x) x = self.encoder(x, lengths) x = self.fc_out(x).view(x.shape[0], -1, self.out_dim) return x
28,175
32.703349
93
py
nnsvs
nnsvs-master/nnsvs/diffsinger/pe.py
import math import torch from torch import nn def denorm_f0( f0, uv, pitch_padding=None, min=None, max=None, pitch_norm="log", use_uv=True, f0_std=1.0, f0_mean=0.0, ): assert use_uv if pitch_norm == "standard": f0 = f0 * f0_std + f0_mean elif pitch_norm == "log": f0 = 2 ** f0 if min is not None: f0 = f0.clamp(min=min) if max is not None: f0 = f0.clamp(max=max) if uv is not None and use_uv: f0[uv > 0] = 0 if pitch_padding is not None: f0[pitch_padding] = 0 return f0 def LayerNorm(normalized_shape, eps=1e-5, elementwise_affine=True, export=False): if not export and torch.cuda.is_available(): try: from apex.normalization import FusedLayerNorm return FusedLayerNorm(normalized_shape, eps, elementwise_affine) except ImportError: pass return torch.nn.LayerNorm(normalized_shape, eps, elementwise_affine) class ConvNorm(torch.nn.Module): def __init__( self, in_channels, out_channels, kernel_size=1, stride=1, padding=None, dilation=1, bias=True, w_init_gain="linear", ): super(ConvNorm, self).__init__() if padding is None: assert kernel_size % 2 == 1 padding = int(dilation * (kernel_size - 1) / 2) self.conv = torch.nn.Conv1d( in_channels, out_channels, kernel_size=kernel_size, stride=stride, padding=padding, dilation=dilation, bias=bias, ) torch.nn.init.xavier_uniform_( self.conv.weight, gain=torch.nn.init.calculate_gain(w_init_gain) ) def forward(self, signal): conv_signal = self.conv(signal) return conv_signal def Linear(in_features, out_features, bias=True): m = nn.Linear(in_features, out_features, bias) nn.init.xavier_uniform_(m.weight) if bias: nn.init.constant_(m.bias, 0.0) return m def make_positions(tensor, padding_idx): """Replace non-padding symbols with their position numbers. Position numbers begin at padding_idx+1. Padding symbols are ignored. """ # The series of casts and type-conversions here are carefully # balanced to both work with ONNX export and XLA. In particular XLA # prefers ints, cumsum defaults to output longs, and ONNX doesn't know # how to handle the dtype kwarg in cumsum. mask = tensor.ne(padding_idx).int() return (torch.cumsum(mask, dim=1).type_as(mask) * mask).long() + padding_idx class SinusoidalPositionalEmbedding(nn.Module): """This module produces sinusoidal positional embeddings of any length. Padding symbols are ignored. """ def __init__(self, embedding_dim, padding_idx, init_size=1024): super().__init__() self.embedding_dim = embedding_dim self.padding_idx = padding_idx self.weights = SinusoidalPositionalEmbedding.get_embedding( init_size, embedding_dim, padding_idx, ) self.register_buffer("_float_tensor", torch.FloatTensor(1)) @staticmethod def get_embedding(num_embeddings, embedding_dim, padding_idx=None): """Build sinusoidal embeddings. This matches the implementation in tensor2tensor, but differs slightly from the description in Section 3.5 of "Attention Is All You Need". """ half_dim = embedding_dim // 2 emb = math.log(10000) / (half_dim - 1) emb = torch.exp(torch.arange(half_dim, dtype=torch.float) * -emb) emb = torch.arange(num_embeddings, dtype=torch.float).unsqueeze( 1 ) * emb.unsqueeze(0) emb = torch.cat([torch.sin(emb), torch.cos(emb)], dim=1).view( num_embeddings, -1 ) if embedding_dim % 2 == 1: # zero pad emb = torch.cat([emb, torch.zeros(num_embeddings, 1)], dim=1) if padding_idx is not None: emb[padding_idx, :] = 0 return emb def forward( self, input, incremental_state=None, timestep=None, positions=None, **kwargs ): """Input is expected to be of size [bsz x seqlen].""" bsz, seq_len = input.shape[:2] max_pos = self.padding_idx + 1 + seq_len if self.weights is None or max_pos > self.weights.size(0): # recompute/expand embeddings if needed self.weights = SinusoidalPositionalEmbedding.get_embedding( max_pos, self.embedding_dim, self.padding_idx, ) self.weights = self.weights.to(self._float_tensor) if incremental_state is not None: # positions is the same for every token when decoding a single step pos = timestep.view(-1)[0] + 1 if timestep is not None else seq_len return self.weights[self.padding_idx + pos, :].expand(bsz, 1, -1) positions = ( make_positions(input, self.padding_idx) if positions is None else positions ) return ( self.weights.index_select(0, positions.view(-1)) .view(bsz, seq_len, -1) .detach() ) def max_positions(self): """Maximum number of supported positions.""" return int(1e5) # an arbitrary large number class FSLayerNorm(torch.nn.LayerNorm): """Layer normalization module. :param int nout: output dim size :param int dim: dimension to be normalized """ def __init__(self, nout, dim=-1): """Construct an LayerNorm object.""" super(FSLayerNorm, self).__init__(nout, eps=1e-12) self.dim = dim def forward(self, x): """Apply layer normalization. :param torch.Tensor x: input tensor :return: layer normalized tensor :rtype torch.Tensor """ if self.dim == -1: return super(FSLayerNorm, self).forward(x) return super(FSLayerNorm, self).forward(x.transpose(1, -1)).transpose(1, -1) class PitchPredictor(torch.nn.Module): def __init__( self, idim, n_layers=5, n_chans=384, odim=2, kernel_size=5, dropout_rate=0.1, padding="SAME", ): """Initialize pitch predictor module. Args: idim (int): Input dimension. n_layers (int, optional): Number of convolutional layers. n_chans (int, optional): Number of channels of convolutional layers. kernel_size (int, optional): Kernel size of convolutional layers. dropout_rate (float, optional): Dropout rate. """ super(PitchPredictor, self).__init__() self.conv = torch.nn.ModuleList() self.kernel_size = kernel_size self.padding = padding for idx in range(n_layers): in_chans = idim if idx == 0 else n_chans self.conv += [ torch.nn.Sequential( torch.nn.ConstantPad1d( ((kernel_size - 1) // 2, (kernel_size - 1) // 2) if padding == "SAME" else (kernel_size - 1, 0), 0, ), torch.nn.Conv1d( in_chans, n_chans, kernel_size, stride=1, padding=0 ), torch.nn.ReLU(), FSLayerNorm(n_chans, dim=1), torch.nn.Dropout(dropout_rate), ) ] self.linear = torch.nn.Linear(n_chans, odim) self.embed_positions = SinusoidalPositionalEmbedding(idim, 0, init_size=4096) self.pos_embed_alpha = nn.Parameter(torch.Tensor([1])) def forward(self, xs): """ :param xs: [B, T, H] :return: [B, T, H] """ positions = self.pos_embed_alpha * self.embed_positions(xs[..., 0]) xs = xs + positions xs = xs.transpose(1, -1) # (B, idim, Tmax) for f in self.conv: xs = f(xs) # (B, C, Tmax) # NOTE: calculate in log domain xs = self.linear(xs.transpose(1, -1)) # (B, Tmax, H) return xs class Prenet(nn.Module): def __init__(self, in_dim=80, out_dim=256, kernel=5, n_layers=3, strides=None): super(Prenet, self).__init__() padding = kernel // 2 self.layers = [] self.strides = strides if strides is not None else [1] * n_layers for idx in range(n_layers): self.layers.append( nn.Sequential( nn.Conv1d( in_dim, out_dim, kernel_size=kernel, padding=padding, stride=self.strides[idx], ), nn.ReLU(), nn.BatchNorm1d(out_dim), ) ) in_dim = out_dim self.layers = nn.ModuleList(self.layers) self.out_proj = nn.Linear(out_dim, out_dim) def forward(self, x): """ :param x: [B, T, 80] :return: [L, B, T, H], [B, T, H] """ padding_mask = x.abs().sum(-1).eq(0).data # [B, T] nonpadding_mask_TB = 1 - padding_mask.float()[:, None, :] # [B, 1, T] x = x.transpose(1, 2) hiddens = [] for i, l in enumerate(self.layers): nonpadding_mask_TB = nonpadding_mask_TB[:, :, :: self.strides[i]] x = l(x) * nonpadding_mask_TB hiddens.append(x) hiddens = torch.stack(hiddens, 0) # [L, B, H, T] hiddens = hiddens.transpose(2, 3) # [L, B, T, H] x = self.out_proj(x.transpose(1, 2)) # [B, T, H] x = x * nonpadding_mask_TB.transpose(1, 2) return hiddens, x class ConvBlock(nn.Module): def __init__( self, idim=80, n_chans=256, kernel_size=3, stride=1, norm="gn", dropout=0 ): super().__init__() self.conv = ConvNorm(idim, n_chans, kernel_size, stride=stride) self.norm = norm if self.norm == "bn": self.norm = nn.BatchNorm1d(n_chans) elif self.norm == "in": self.norm = nn.InstanceNorm1d(n_chans, affine=True) elif self.norm == "gn": self.norm = nn.GroupNorm(n_chans // 16, n_chans) elif self.norm == "ln": self.norm = LayerNorm(n_chans // 16, n_chans) elif self.norm == "wn": self.conv = torch.nn.utils.weight_norm(self.conv.conv) self.dropout = nn.Dropout(dropout) self.relu = nn.ReLU() def forward(self, x): """ :param x: [B, C, T] :return: [B, C, T] """ x = self.conv(x) if not isinstance(self.norm, str): if self.norm == "none": pass elif self.norm == "ln": x = self.norm(x.transpose(1, 2)).transpose(1, 2) else: x = self.norm(x) x = self.relu(x) x = self.dropout(x) return x class ConvStacks(nn.Module): def __init__( self, idim=80, n_layers=5, n_chans=256, odim=32, kernel_size=5, norm="gn", dropout=0, strides=None, res=True, ): super().__init__() self.conv = torch.nn.ModuleList() self.kernel_size = kernel_size self.res = res self.in_proj = Linear(idim, n_chans) if strides is None: strides = [1] * n_layers else: assert len(strides) == n_layers for idx in range(n_layers): self.conv.append( ConvBlock( n_chans, n_chans, kernel_size, stride=strides[idx], norm=norm, dropout=dropout, ) ) self.out_proj = Linear(n_chans, odim) def forward(self, x, return_hiddens=False): """ :param x: [B, T, H] :return: [B, T, H] """ x = self.in_proj(x) x = x.transpose(1, -1) # (B, idim, Tmax) hiddens = [] for f in self.conv: x_ = f(x) x = x + x_ if self.res else x_ # (B, C, Tmax) hiddens.append(x) x = x.transpose(1, -1) x = self.out_proj(x) # (B, Tmax, H) if return_hiddens: hiddens = torch.stack(hiddens, 1) # [B, L, C, T] return x, hiddens return x class PitchExtractor(nn.Module): def __init__( self, n_mel_bins=80, conv_layers=2, hidden_size=256, predictor_hidden=-1, ffn_padding="SAME", predictor_kernel=5, pitch_type="frame", use_uv=True, ): super().__init__() self.hidden_size = hidden_size self.pitch_type = pitch_type assert pitch_type == "log" self.use_uv = use_uv self.predictor_hidden = ( predictor_hidden if predictor_hidden > 0 else self.hidden_size ) self.conv_layers = conv_layers self.mel_prenet = Prenet(n_mel_bins, self.hidden_size, strides=[1, 1, 1]) if self.conv_layers > 0: self.mel_encoder = ConvStacks( idim=self.hidden_size, n_chans=self.hidden_size, odim=self.hidden_size, n_layers=self.conv_layers, ) self.pitch_predictor = PitchPredictor( self.hidden_size, n_chans=self.predictor_hidden, n_layers=5, dropout_rate=0.1, odim=2, padding=ffn_padding, kernel_size=predictor_kernel, ) def forward(self, mel_input=None): mel_hidden = self.mel_prenet(mel_input)[1] if self.conv_layers > 0: mel_hidden = self.mel_encoder(mel_hidden) # log2(f0), uv pitch_pred = self.pitch_predictor(mel_hidden) lf0, uv = pitch_pred[:, :, 0], pitch_pred[:, :, 1] # f0 f0 = 2 ** lf0 # log(f0) lf0 = torch.log(f0) lf0[uv > 0] = 0 return lf0 class PitchExtractorWrapper(nn.Module): def __init__(self, **kwargs): super().__init__() self.model = PitchExtractor(**kwargs) def forward(self, x, lengths=None, y=None): return self.pitch_extractor(x)
14,548
30.087607
87
py
nnsvs
nnsvs-master/nnsvs/diffsinger/denoiser.py
import math from math import sqrt import torch import torch.nn as nn import torch.nn.functional as F class Mish(nn.Module): def forward(self, x): return x * torch.tanh(F.softplus(x)) class SinusoidalPosEmb(nn.Module): def __init__(self, dim): super().__init__() self.dim = dim def forward(self, x): device = x.device half_dim = self.dim // 2 emb = math.log(10000) / (half_dim - 1) emb = torch.exp(torch.arange(half_dim, device=device) * -emb) emb = x[:, None] * emb[None, :] emb = torch.cat((emb.sin(), emb.cos()), dim=-1) return emb def Conv1d(*args, **kwargs): layer = nn.Conv1d(*args, **kwargs) nn.init.kaiming_normal_(layer.weight) return layer @torch.jit.script def silu(x): return x * torch.sigmoid(x) class ResidualBlock(nn.Module): def __init__(self, encoder_hidden, residual_channels, dilation): super().__init__() self.dilated_conv = Conv1d( residual_channels, 2 * residual_channels, 3, padding=dilation, dilation=dilation, ) self.diffusion_projection = nn.Linear(residual_channels, residual_channels) self.conditioner_projection = Conv1d(encoder_hidden, 2 * residual_channels, 1) self.output_projection = Conv1d(residual_channels, 2 * residual_channels, 1) def forward(self, x, conditioner, diffusion_step): diffusion_step = self.diffusion_projection(diffusion_step).unsqueeze(-1) conditioner = self.conditioner_projection(conditioner) y = x + diffusion_step y = self.dilated_conv(y) + conditioner gate, filter = torch.chunk(y, 2, dim=1) y = torch.sigmoid(gate) * torch.tanh(filter) y = self.output_projection(y) residual, skip = torch.chunk(y, 2, dim=1) return (x + residual) / sqrt(2.0), skip class DiffNet(nn.Module): def __init__( self, in_dim=80, encoder_hidden_dim=256, residual_layers=20, residual_channels=256, dilation_cycle_length=4, ): super().__init__() self.in_dim = in_dim self.input_projection = Conv1d(in_dim, residual_channels, 1) self.diffusion_embedding = SinusoidalPosEmb(residual_channels) dim = residual_channels self.mlp = nn.Sequential( nn.Linear(dim, dim * 4), Mish(), nn.Linear(dim * 4, dim) ) self.residual_layers = nn.ModuleList( [ ResidualBlock( encoder_hidden_dim, residual_channels, 2 ** (i % dilation_cycle_length), ) for i in range(residual_layers) ] ) self.skip_projection = Conv1d(residual_channels, residual_channels, 1) self.output_projection = Conv1d(residual_channels, in_dim, 1) nn.init.zeros_(self.output_projection.weight) def forward(self, spec, diffusion_step, cond): """ :param spec: [B, 1, M, T] :param diffusion_step: [B, 1] :param cond: [B, M, T] :return: """ x = spec[:, 0] x = self.input_projection(x) # x [B, residual_channel, T] x = F.relu(x) diffusion_step = self.diffusion_embedding(diffusion_step) diffusion_step = self.mlp(diffusion_step) skip = [] for _, layer in enumerate(self.residual_layers): x, skip_connection = layer(x, cond, diffusion_step) skip.append(skip_connection) x = torch.sum(torch.stack(skip), dim=0) / sqrt(len(self.residual_layers)) x = self.skip_projection(x) x = F.relu(x) x = self.output_projection(x) # [B, 80, T] return x[:, None, :, :]
3,818
29.552
86
py
ShapeMOD
ShapeMOD-main/SA_lang/sa_utils.py
import torch import torch.nn as nn import numpy as np import tasks.ShapeAssembly as sa from copy import deepcopy MAX_CUBES = 10 PREC = 4 TRANS_NORM = 10. SQUARE_THRESH = .1 def loadObj(infile): tverts = [] ttris = [] with open(infile) as f: for line in f: ls = line.split() if len(ls) == 0: continue if ls[0] == 'v': tverts.append([ float(ls[1]), float(ls[2]), float(ls[3]) ]) elif ls[0] == 'f': ttris.append([ int(ls[1].split('//')[0])-1, int(ls[2].split('//')[0])-1, int(ls[3].split('//')[0])-1 ]) return tverts, ttris def make_function(name, args): args = [str(arg) for arg in args] return '{}({})'.format(name, ", ".join(args)) def assign(var_name, value): return '{} = {}'.format(var_name, value) def get_aligned_flags(nodes): caligned = [1] * len(nodes[0]['cubes']) for node in nodes: for i in range(1, len(caligned)): if not node['aligned'][i]: caligned[i] = 0 return caligned def NodeToDirTensor(props): for key in ['xdir', 'ydir', 'zdir']: if not isinstance(props[key], np.ndarray): props[key] = np.array(props[key]) return props def get_flex_aligned_flags(nodes, thresh): caligned = [1] * len(nodes[0]['cubes']) for node in nodes: for i in range(1, len(caligned)): if not isFlexAxisAligned( NodeToDirTensor(node['cubes'][i]), NodeToDirTensor(node['cubes'][0]), thresh ): caligned[i] = 0 return caligned def isAxisAligned(props, bbox, thresh): xdir = props['xdir'] ydir = props['ydir'] zdir = props['zdir'] xdir /= np.linalg.norm(xdir) ydir /= np.linalg.norm(ydir) zdir /= np.linalg.norm(zdir) if xdir.dot(bbox['xdir']) < thresh: return False if ydir.dot(bbox['ydir']) < thresh: return False if zdir.dot(bbox['zdir']) < thresh: return False return True def isSquare(a, b, c): a = np.array(a) b = np.array(b) c = np.array(c) v = np.abs(a - b)/max(a.item(), b.item(), c.item()) < SQUARE_THRESH return v def isSpecAxisAligned(cdir, axis, thresh): cdir /= np.linalg.norm(cdir) if cdir.dot(axis) >= thresh: return True else: return False def isFlexAxisAligned(props, bbox, thresh): if isAxisAligned(props, bbox, thresh): return True xsquare = isSquare(props['zd'], props['yd'], props['xd']) ysquare = isSquare(props['zd'], props['xd'], props['yd']) zsquare = isSquare(props['xd'], props['yd'], props['zd']) if xsquare and ysquare and zsquare: return True if xsquare: return isSpecAxisAligned(props['xdir'], bbox['xdir'], thresh) if ysquare: return isSpecAxisAligned(props['ydir'], bbox['ydir'], thresh) if zsquare: return isSpecAxisAligned(props['zdir'], bbox['zdir'], thresh) return False def vector_cos(norm1, norm2): norm1 = np.asarray(norm1) norm2 = np.asarray(norm2) dot = np.dot(norm1, norm2) magnitude = np.linalg.norm(norm1) * np.linalg.norm(norm2) if magnitude == 0.: return 0. return dot / float(magnitude) def orientProps(center, xd, yd, zd, xdir, ydir, zdir): rt = np.asarray([1., 0., 0.]) up = np.asarray([0., 1., 0.]) fwd = np.asarray([0., 0., 1.]) l = [ (xdir, xd, 0), (ydir, yd, 1), (zdir, zd, 2), (-1 * xdir, xd, 3), (-1 * ydir, yd, 4), (-1 * zdir, zd, 5) ] rtdir, rtd, rind = sorted( deepcopy(l), key=lambda x: vector_cos(rt, x[0]))[-1] if rind >= 3: l.pop(rind) l.pop((rind+3) % 6) else: l.pop((rind+3) % 6) l.pop(rind) for i in range(0, 4): p_ind = l[i][2] if p_ind > max(rind, (rind+3) % 6): l[i] = (l[i][0], l[i][1], l[i][2] - 2) elif p_ind > min(rind, (rind+3) % 6): l[i] = (l[i][0], l[i][1], l[i][2] - 1) updir, upd, upind = sorted( deepcopy(l), key=lambda x: vector_cos(up, x[0]))[-1] if upind >= 2: l.pop(upind) l.pop((upind+2) % 4) else: l.pop((upind+2) % 4) l.pop(upind) fwdir, fwd, _ = sorted(l, key=lambda x: vector_cos(fwd, x[0]))[-1] return { 'center': torch.tensor(center).float(), 'xd': torch.tensor(rtd).float(), 'yd': torch.tensor(upd).float(), 'zd': torch.tensor(fwd).float(), 'xdir': torch.tensor(rtdir).float(), 'ydir': torch.tensor(updir).float(), 'zdir': torch.tensor(fwdir).float() } def locallyClean(prog): cube_count = -1 switches = [] for line in prog: if 'Cuboid' in line: if 'Program_' in line: switches.append(( f'cube{cube_count}', line.split()[0] )) cube_count += 1 for a, b in switches: prog = [line.replace(b, a) for line in prog] clines = [] P = sa.Program() for line in prog: if "Cuboid(" in line: parse = P.parseCuboid(line) name = parse[0] x = float(max(parse[1].item(), 0.01)) y = float(max(parse[2].item(), 0.01)) z = float(max(parse[3].item(), 0.01)) aligned = str(parse[4]) clines.append("\t" + gen.assign( name, gen.make_function('Cuboid', [x, y, z, aligned]))) if "attach(" in line: parse = P.parseAttach(line) clines.append("\t" + gen.make_function( "attach", [parse[0], parse[1]] + [ torch.clamp(co, 0.0, 1.0).item() for co in parse[2:] ] )) if "squeeze(" in line: parse = P.parseSqueeze(line) clines.append("\t" + gen.make_function( "squeeze", [parse[0], parse[1], parse[2], parse[3]] + [max(min(co, 1.0), 0.0) for co in parse[4:]] )) if "reflect(" in line: parse = P.parseReflect(line) clines.append("\t" + gen.make_function( "reflect", [parse[0], parse[1]] )) if "translate(" in line: parse = P.parseTranslate(line) clines.append("\t" + make_function( "translate", [ parse[0], parse[1], max(int(parse[2]), 1), float(min(max(parse[3], 0.0), 1.0)) ] )) return clines def fillHP(name, progs, children, program_line_maps = None): hp = {'name': name} hp['prog'] = locallyClean(progs[name]) hp['children'] = [ fillHP(cn, progs, children, program_line_maps) if cn is not None else {} for cn in children[name] ] hp['prog_global_line_indices'] = program_line_maps[name] if program_line_maps else None if len(hp['children']) > MAX_CUBES: return {} return hp def loadFromText(text): progs = {} children = {} # This maps program line numbers to global line numbers. program_line_maps = {} prog_num = -1 cur_prog = [] cur_children = [] cur_program_line_map = [] for index, line in enumerate(text.split('\n')): line = line + '\n' ls = line.split() if len(ls) == 0: continue if ls[0] == 'Assembly': prog_num = int(ls[1].split('_')[1]) elif ls[0] == '}': progs[prog_num] = cur_prog program_line_maps[prog_num] = cur_program_line_map children[prog_num] = cur_children cur_prog = [] cur_children = [] cur_program_line_map = [] elif 'Cuboid' in line: if 'Program_' in line: cur_children.append(int(ls[0].split('_')[1])) else: cur_children.append(None) # For function invocations, append the line's arguments to the current program. appendable_commands = ['Cuboid', 'attach', 'reflect', 'translate', 'squeeze'] for appendable_command in appendable_commands: if appendable_command in line: cur_prog.append(line[1:-1]) cur_program_line_map.append(index) return fillHP(0, progs, children, program_line_maps) def loadHPFromFile(progfile): progs = {} children = {} prog_num = -1 cur_prog = [] cur_children = [] with open(progfile) as f: for line in f: ls = line.split() if ls[0] == 'Assembly': prog_num = int(ls[1].split('_')[1]) elif ls[0] == '}': progs[prog_num] = cur_prog children[prog_num] = cur_children cur_prog = [] cur_children = [] elif 'Cuboid' in line: if 'Program_' in line: cur_children.append(int(ls[0].split('_')[1])) else: cur_children.append(None) cur_prog.append(line[1:-1]) elif 'attach' in line: cur_prog.append(line[1:-1]) elif 'reflect' in line: cur_prog.append(line[1:-1]) elif 'translate' in line: cur_prog.append(line[1:-1]) elif 'squeeze' in line: cur_prog.append(line[1:-1]) return fillHP(0, progs, children) def locallyNormClean(prog, max_val): cube_count = -1 switches = [] for line in prog: if 'Cuboid' in line: if 'Program_' in line: switches.append(( f'cube{cube_count}', line.split()[0] )) cube_count += 1 for a, b in switches: prog = [line.replace(b, a) for line in prog] clines = [] P = sa.Program() for line in prog: if "Cuboid(" in line: parse = P.parseCuboid(line) name = parse[0] x = float(min(max(parse[1].item() / max_val, 0.01), 1.0)) y = float(min(max(parse[2].item() / max_val, 0.01), 1.0)) z = float(min(max(parse[3].item() / max_val, 0.01), 1.0)) clines.append("\t" + assign( name, make_function('Cuboid', [x, y, z]))) if "attach(" in line: parse = P.parseAttach(line) clines.append("\t" + make_function( "attach", [parse[0], parse[1]] + [ torch.clamp(co, 0.0, 1.0).item() for co in parse[2:] ] )) if "reflect(" in line: parse = P.parseReflect(line) clines.append("\t" + make_function( "reflect", [parse[0], parse[1]] )) if "translate(" in line: parse = P.parseTranslate(line) clines.append("\t" + make_function( "translate", [ parse[0], parse[1], max(int(parse[2]), 1), float(min(max(parse[3], 0.0), 1.0)) ] )) return clines def fillNormHP(name, progs, children, max_val=None): hp = {'name': name} if max_val is None: max_val = max([float(d) for d in progs[name][0][:-1].split('(')[1].split(',')]) hp['prog'] = locallyNormClean(progs[name], max_val) hp['children'] = [ fillNormHP(cn, progs, children, max_val) if cn is not None else {} for cn in children[name] ] if len(hp['children']) > MAX_CUBES: return {} return hp def loadNormHPFromFile(progfile): progs = {} children = {} prog_num = -1 cur_prog = [] cur_children = [] with open(progfile) as f: for line in f: ls = line.split() if ls[0] == 'Assembly': prog_num = int(ls[1].split('_')[1]) elif ls[0] == '}': progs[prog_num] = cur_prog children[prog_num] = cur_children cur_prog = [] cur_children = [] elif 'Cuboid' in line: if 'Program_' in line: cur_children.append(int(ls[0].split('_')[1])) else: cur_children.append(None) cur_prog.append(line[1:-1]) elif 'attach' in line: cur_prog.append(line[1:-1]) elif 'reflect' in line: cur_prog.append(line[1:-1]) elif 'translate' in line: cur_prog.append(line[1:-1]) return fillNormHP(0, progs, children) def getHierProgLines(root, prog_field): prog_count = 0 root["prog_num"] = prog_count lines = [] q = [root] while(len(q) > 0): node = q.pop(0) lines.append(f"Assembly Program_{node['prog_num']}" +" {") NAME_DICT = {} c = 0 for ret, line in node[prog_field]: params = line[1:] fn = line[0] for p_ret in ret: if len(node['children'][c]) > 0: prog_count += 1 name = f'Program_{prog_count}' node['children'][c]['prog_num'] = prog_count else: name = p_ret NAME_DICT[p_ret] = name c += 1 clean_ret = [NAME_DICT[r] for r in ret] clean_params = [NAME_DICT[p] if p in NAME_DICT else p for p in params] if len(ret) > 0: lines.append("\t" + assign(','.join(clean_ret), make_function(fn, clean_params))) else: lines.append("\t" + make_function(fn, clean_params)) lines.append("}") for c in node["children"]: if c is not None and len(c) > 0: if "prog_num" in c: q.append(c) return lines def sagetHierProgLines(root): prog_count = 0 root["prog_num"] = prog_count lines = [] q = [root] P = sa.Program() while(len(q) > 0): node = q.pop(0) lines.append(f"Assembly Program_{node['prog_num']}" +" {") NAME_DICT = {} c = 0 for line in node["prog"]: if "Cuboid(" in line: parse = P.parseCuboid(line) if len(node["children"][c]) > 0: prog_count += 1 name = f"Program_{prog_count}" node["children"][c]["prog_num"] = prog_count else: name = parse[0] NAME_DICT[parse[0]] = name x = round(float(parse[1]), PREC) y = round(float(parse[2]), PREC) z = round(float(parse[3]), PREC) aligned = str(parse[4]) lines.append("\t" + assign( name, make_function('Cuboid', [x,y,z,aligned])) ) c += 1 if "attach(" in line: parse = P.parseAttach(line) lines.append("\t" + make_function( "attach", [NAME_DICT[parse[0]], NAME_DICT[parse[1]]] + [round(co.item(), PREC) for co in parse[2:]] )) if "reflect(" in line: parse = P.parseReflect(line) lines.append("\t" + make_function( "reflect", [NAME_DICT[parse[0]], parse[1]] )) if "translate(" in line: parse = P.parseTranslate(line) lines.append("\t" + make_function( "translate", [NAME_DICT[parse[0]], parse[1], int(parse[2]), round(float(parse[3]), PREC)] )) if "rotate(" in line: parse = P.parseRotate(line) lines.append("\t" + make_function( "rotate", [NAME_DICT[parse[0]], int(parse[1]), round(float(parse[2]), PREC)] )) if "squeeze(" in line: parse = P.parseSqueeze(line) lines.append("\t" + make_function( "squeeze", [NAME_DICT[parse[0]], NAME_DICT[parse[1]], NAME_DICT[parse[2]], parse[3]] + [round(co, PREC) for co in parse[4:]] )) lines.append("}") for c in node["children"]: if c is not None and len(c) > 0: if "prog_num" in c: q.append(c) return lines def sawriteHierProg(hier_prog, outfile): lines = sagetHierProgLines(hier_prog) with open(outfile, 'w') as f: for line in lines: f.write(f'{line}\n') def writeHierProg(hier_prog, prog_field, outfile): lines = getHierProgLines(hier_prog, prog_field) with open(outfile, 'w') as f: for line in lines: f.write(f'{line}\n') def log_print(s, of): with open(of, 'a') as f: f.write(f"{s}\n") print(s) def generateObj(verts, faces): faces = faces.clone() faces += 1 obj = "" for a, b, c in verts.tolist(): obj += f'v {a} {b} {c}\n' for a, b, c in faces.tolist(): obj += f"f {a} {b} {c}\n" return obj def writeObj(verts, faces, outfile): faces = faces.clone() faces += 1 with open(outfile, 'w') as f: for a, b, c in verts.tolist(): f.write(f'v {a} {b} {c}\n') for a, b, c in faces.tolist(): f.write(f"f {a} {b} {c}\n") def writePC(pc, fn): with open(fn, 'w') as f: for a, b, c, _, _, _ in pc: f.write(f'v {a} {b} {c} \n') def writeSPC(pc, fn): with open(fn, 'w') as f: for a, b, c in pc: f.write(f'v {a} {b} {c} \n') def face_areas_normals(faces, vs): face_normals = torch.cross( vs[:, faces[:, 1], :] - vs[:, faces[:, 0], :], vs[:, faces[:, 2], :] - vs[:, faces[:, 1], :], dim=2, ) face_areas = torch.norm(face_normals, dim=2) + 1e-8 face_normals = face_normals / face_areas[:, :, None] face_areas = 0.5 * face_areas return face_areas, face_normals def sample_surface(faces, vs, count, return_normals=True): """ sample mesh surface sample method: http://mathworld.wolfram.com/TrianglePointPicking.html Args --------- vs: vertices (batch x nvs x 3d coordinate) faces: triangle faces (torch.long) (num_faces x 3) count: number of samples Return --------- samples: (count, 3) points in space on the surface of mesh face_index: (count,) indices of faces for each sampled point """ if torch.isnan(faces).any() or torch.isnan(vs).any(): assert False, 'saw nan in sample_surface' device = vs.device bsize, nvs, _ = vs.shape area, normal = face_areas_normals(faces, vs) area_sum = torch.sum(area, dim=1) assert not (area <= 0.0).any().item( ), "Saw negative probability while sampling" assert not (area_sum <= 0.0).any().item( ), "Saw negative probability while sampling" assert not (area > 1000000.0).any().item(), "Saw inf" assert not (area_sum > 1000000.0).any().item(), "Saw inf" dist = torch.distributions.categorical.Categorical( probs=area / (area_sum[:, None])) face_index = dist.sample((count,)) # pull triangles into the form of an origin + 2 vectors tri_origins = vs[:, faces[:, 0], :] tri_vectors = vs[:, faces[:, 1:], :].clone() tri_vectors -= tri_origins.repeat(1, 1, 2).reshape((bsize, len(faces), 2, 3)) # pull the vectors for the faces we are going to sample from face_index = face_index.transpose(0, 1) face_index = face_index[:, :, None].expand((bsize, count, 3)) tri_origins = torch.gather(tri_origins, dim=1, index=face_index) face_index2 = face_index[:, :, None, :].expand((bsize, count, 2, 3)) tri_vectors = torch.gather(tri_vectors, dim=1, index=face_index2) # randomly generate two 0-1 scalar components to multiply edge vectors by random_lengths = torch.rand( count, 2, 1, device=vs.device, dtype=tri_vectors.dtype) # points will be distributed on a quadrilateral if we use 2x [0-1] samples # if the two scalar components sum less than 1.0 the point will be # inside the triangle, so we find vectors longer than 1.0 and # transform them to be inside the triangle random_test = random_lengths.sum(dim=1).reshape(-1) > 1.0 random_lengths[random_test] -= 1.0 random_lengths = torch.abs(random_lengths) # multiply triangle edge vectors by the random lengths and sum sample_vector = (tri_vectors * random_lengths[None, :]).sum(dim=2) # finally, offset by the origin to generate # (n,3) points in space on the triangle samples = sample_vector + tri_origins if return_normals: samples = torch.cat((samples, torch.gather( normal, dim=1, index=face_index)), dim=2) return samples else: return samples
21,643
28.013405
133
py
ShapeMOD
ShapeMOD-main/SA_lang/parse_files/old_intersect.py
import torch import itertools import trimesh import scipy import numpy as np import faiss from copy import deepcopy DOING_PARSE = True DIM = 20 ATT_DIM = 50 device = torch.device("cuda") resource = faiss.StandardGpuResources() def robust_norm(var, dim=2): return ((var ** 2).sum(dim=dim) + 1e-8).sqrt() class collisionEngine(): def __init__(self, device): self.dimension = 3 self.k = 1 self.device = device def search_nn(self, index, query, k): _, I = index.search(query, k) I_var = torch.from_numpy(np.ascontiguousarray(I).astype(np.int64)) I_var = I_var.to(self.device) return I_var def findPCInter(self, predict_pc, gt_pc, index_predict, index_gt, thresh): predict_pc = predict_pc.to(self.device) gt_pc = gt_pc.to(self.device) predict_pc_size = predict_pc.size() gt_pc_size = gt_pc.size() predict_pc_np = np.ascontiguousarray( torch.transpose(predict_pc.data.clone(), 1, 2).cpu().numpy() ) # BxMx3 gt_pc_np = np.ascontiguousarray( torch.transpose(gt_pc.data.clone(), 1, 2).cpu().numpy() ) # BxNx3 # selected_gt: Bxkx3xM selected_gt_by_predict = torch.FloatTensor( predict_pc_size[0], self.k, predict_pc_size[1], predict_pc_size[2] ) # selected_predict: Bxkx3xN selected_predict_by_gt = torch.FloatTensor( gt_pc_size[0], self.k, gt_pc_size[1], gt_pc_size[2] ) selected_gt_by_predict = selected_gt_by_predict.to(self.device) selected_predict_by_gt = selected_predict_by_gt.to(self.device) # process each batch independently. for i in range(predict_pc_np.shape[0]): # database is gt_pc, predict_pc -> gt_pc ----------------------------------------------------------- I_var = self.search_nn(index_gt, predict_pc_np[i], self.k) # process nearest k neighbors for k in range(self.k): selected_gt_by_predict[i, k, ...] = gt_pc[i].index_select( 1, I_var[:, k] ) # database is predict_pc, gt_pc -> predict_pc ------------------------------------------------------- I_var = self.search_nn(index_predict, gt_pc_np[i], self.k) # process nearest k neighbors for k in range(self.k): selected_predict_by_gt[i, k, ...] = predict_pc[i].index_select( 1, I_var[:, k] ) # compute loss =================================================== # selected_gt(Bxkx3xM) vs predict_pc(Bx3xM) r_to_gt = predict_pc.unsqueeze(1).expand_as(selected_gt_by_predict) - selected_gt_by_predict gt_to_r = selected_predict_by_gt - gt_pc.unsqueeze(1).expand_as(selected_predict_by_gt) r_to_gt = robust_norm( selected_gt_by_predict - predict_pc.unsqueeze(1).expand_as(selected_gt_by_predict), dim=2 ) gt_to_r = robust_norm( selected_predict_by_gt - gt_pc.unsqueeze(1).expand_as(selected_predict_by_gt), dim=2 ) r_to_gt = r_to_gt.flatten() gt_to_r = gt_to_r.flatten() return (r_to_gt < thresh).nonzero().squeeze(), (gt_to_r < thresh).nonzero().squeeze() if DOING_PARSE: a = (torch.arange(DIM).float()/(DIM-1)) b = a.unsqueeze(0).unsqueeze(0).repeat(DIM, DIM, 1) c = a.unsqueeze(0).unsqueeze(2).repeat(DIM, 1, DIM) d = a.unsqueeze(1).unsqueeze(2).repeat(1, DIM, DIM) g = torch.stack((b,c,d), dim=3).view(-1, 3).to(device) fr = (g[:,0] == 1.).nonzero().squeeze().to(device) fl = (g[:,0] == 0.).nonzero().squeeze().to(device) ft = (g[:,1] == 1.).nonzero().squeeze().to(device) fbo = (g[:,1] == 0.).nonzero().squeeze().to(device) ff = (g[:,2] == 1.).nonzero().squeeze().to(device) fba = (g[:,2] == 0.).nonzero().squeeze().to(device) bb_mask = torch.ones((DIM**3), 3).float().to(device) * 100 bb_mask[torch.cat((ft, fbo)), :] *= 0. top_mask = torch.ones((DIM**3), 3).float().to(device) * 100 bot_mask = torch.ones((DIM**3), 3).float().to(device) * 100 top_mask[ft] *= 0. bot_mask[fbo] *= 0. s_xyz = g.unsqueeze(0).to(device) c_xyz = torch.tensor([ [0,0,0], [0,0,1], [0,1,0], [0,1,1], [1,0,0], [1,0,1], [1,1,0], [1,1,1], ]).unsqueeze(0).float() colEngine = collisionEngine(device) atta = (torch.arange(ATT_DIM).float()/(ATT_DIM-1)) attb = atta.unsqueeze(0).unsqueeze(0).repeat(ATT_DIM, ATT_DIM, 1) attc = atta.unsqueeze(0).unsqueeze(2).repeat(ATT_DIM, 1, ATT_DIM) attd = atta.unsqueeze(1).unsqueeze(2).repeat(1, ATT_DIM, ATT_DIM) attg = torch.stack((attb,attc,attd), dim=3).view(-1, 3) attxyz = attg.unsqueeze(0).to(device) def smp_pt(geom, pt): xdir = geom[6:9] / (geom[6:9].norm() + 1e-8) ydir = geom[9:12] / (geom[9:12].norm() + 1e-8) zdir = torch.cross(xdir, ydir) r = torch.stack(( xdir, ydir, zdir )).T return (r @ ((pt -.5) * geom[:3])) + geom[3:6] def findHiddenCubes(cubes): cube_geom = [] for c in cubes: cube_geom.append(torch.cat(( c['xd'].unsqueeze(0), c['yd'].unsqueeze(0), c['zd'].unsqueeze(0), c['center'], c['xdir'], c['ydir'] ))) scene_geom = torch.stack([c for c in cube_geom]).to(device) scene_corners = sampleCorners(cubes).view(-1,8,3) bad_inds = [] for ind in range(len(cubes)): points = scene_corners[ind,:,:] all_inside = False for nind in range(len(cubes)): if nind == ind: continue if all_inside: break O = smp_pt(scene_geom[nind], torch.zeros(3).to(device)) A = torch.stack([ scene_geom[nind][0] * scene_geom[nind][6:9], scene_geom[nind][1] * scene_geom[nind][9:12], scene_geom[nind][2] * torch.cross( scene_geom[nind][6:9], scene_geom[nind][9:12] ) ]).T B = (points.T - O.unsqueeze(1)).cpu() p = torch.tensor(np.linalg.solve(A.cpu(), B.cpu())).T if p.min() >= -0.01 and p.max() <= 1.01: all_inside = True if all_inside: bad_inds.append(ind) return bad_inds def findOverlapCubes(cubes, CTHRESH): cube_geom = [] for c in cubes: cube_geom.append(torch.cat(( c['xd'].unsqueeze(0), c['yd'].unsqueeze(0), c['zd'].unsqueeze(0), c['center'], c['xdir'], c['ydir'] ))) scene_geom = torch.stack([c for c in cube_geom]).to(device) scene_corners = sampleCube(cubes).view(-1,DIM**3,3) bad_inds = set() for ind in range(len(cubes)): points = scene_corners[ind,:,:] covered = False for nind in range(len(cubes)): if nind == ind or nind in bad_inds: continue if covered: break O = smp_pt(scene_geom[nind], torch.zeros(3).to(device)) A = torch.stack([ scene_geom[nind][0] * scene_geom[nind][6:9], scene_geom[nind][1] * scene_geom[nind][9:12], scene_geom[nind][2] * torch.cross( scene_geom[nind][6:9], scene_geom[nind][9:12] ) ]).T B = (points.T - O.unsqueeze(1)).cpu() p = torch.tensor(np.linalg.solve(A.cpu(), B.cpu())).T num_outside_s = (p - torch.clamp(p, 0.0, 1.0)).abs().sum(dim=1).nonzero().squeeze().shape if len(num_outside_s) == 0 or (((DIM**3 - num_outside_s[0])*1.) / DIM**3) > CTHRESH: covered = True if covered: bad_inds.add(ind) return bad_inds def samplePC(cubes, flip_bbox=False, split_bbox=False): cube_geom = [] for c in cubes: cube_geom.append(torch.cat(( c['xd'].unsqueeze(0), c['yd'].unsqueeze(0), c['zd'].unsqueeze(0), c['center'], c['xdir'], c['ydir'] ))) scene_geom = torch.stack([c for c in cube_geom]).to(device) ind_to_pc = {} for i in range(0, scene_geom.shape[0]): xyz = s_xyz s_inds = (torch.ones(1,xyz.shape[1]) * i).long().to(device) s_r = torch.cat( ( (scene_geom[s_inds][:, :, 6:9] / (scene_geom[s_inds][:, :, 6:9].norm(dim=2).unsqueeze(2) + 1e-8)).unsqueeze(3), (scene_geom[s_inds][:, :, 9:12] / (scene_geom[s_inds][:, :, 9:12].norm(dim=2).unsqueeze(2) + 1e-8)).unsqueeze(3), torch.cross( scene_geom[s_inds][:, :, 6:9] / (scene_geom[s_inds][:, :, 6:9].norm(dim=2).unsqueeze(2) + 1e-8), scene_geom[s_inds][:, :, 9:12] / (scene_geom[s_inds][:, :, 9:12].norm(dim=2).unsqueeze(2) + 1e-8) ).unsqueeze(3) ), dim = 3) s_out = ((s_r @ (((xyz - .5) * scene_geom[s_inds][:, :, :3]).unsqueeze(-1))).squeeze() + scene_geom[s_inds][:, :, 3:6]).squeeze() ind_to_pc[i] = s_out if flip_bbox: ind_to_pc[0] += bb_mask temp = ind_to_pc[0].clone() ind_to_pc[0][ft] = temp[fbo] ind_to_pc[0][fbo] = temp[ft] if split_bbox: bbox_pc = ind_to_pc.pop(0) ind_to_pc[-2] = bbox_pc.clone() + bot_mask ind_to_pc[-1] = bbox_pc.clone() + top_mask res = {} for key in ind_to_pc: index_cpu = faiss.IndexFlatL2(3) index = faiss.index_cpu_to_gpu( resource, torch.cuda.current_device(), index_cpu ) index.add( np.ascontiguousarray(ind_to_pc[key].cpu().numpy()) ) res[key] = (ind_to_pc[key], index) return res, scene_geom def sampleCorners(cubes): cube_geom = [] for c in cubes: cube_geom.append(torch.cat(( c['xd'].unsqueeze(0), c['yd'].unsqueeze(0), c['zd'].unsqueeze(0), c['center'], c['xdir'], c['ydir'] ))) scene_geom = torch.stack([c for c in cube_geom]).to(device) s_inds = torch.arange(scene_geom.shape[0]).unsqueeze(1).repeat(1,8).view(1, -1).to(device) xyz = c_xyz.repeat(1,scene_geom.shape[0],1).to(device) s_r = torch.cat( ( (scene_geom[s_inds][:, :, 6:9] / (scene_geom[s_inds][:, :, 6:9].norm(dim=2).unsqueeze(2) + 1e-8)).unsqueeze(3), (scene_geom[s_inds][:, :, 9:12] / (scene_geom[s_inds][:, :, 9:12].norm(dim=2).unsqueeze(2) + 1e-8)).unsqueeze(3), torch.cross( scene_geom[s_inds][:, :, 6:9] / (scene_geom[s_inds][:, :, 6:9].norm(dim=2).unsqueeze(2) + 1e-8), scene_geom[s_inds][:, :, 9:12] / (scene_geom[s_inds][:, :, 9:12].norm(dim=2).unsqueeze(2) + 1e-8) ).unsqueeze(3) ), dim = 3) s_out = ((s_r @ (((xyz - .5) * scene_geom[s_inds][:, :, :3]).unsqueeze(-1))).squeeze() + scene_geom[s_inds][:, :, 3:6]).squeeze() return s_out def sampleCube(cubes): cube_geom = [] for c in cubes: cube_geom.append(torch.cat(( c['xd'].unsqueeze(0), c['yd'].unsqueeze(0), c['zd'].unsqueeze(0), c['center'], c['xdir'], c['ydir'] ))) scene_geom = torch.stack([c for c in cube_geom]).to(device) s_inds = torch.arange(scene_geom.shape[0]).unsqueeze(1).repeat(1,DIM**3).view(1, -1).to(device) xyz = s_xyz.repeat(1,scene_geom.shape[0],1).to(device) s_r = torch.cat( ( (scene_geom[s_inds][:, :, 6:9] / (scene_geom[s_inds][:, :, 6:9].norm(dim=2).unsqueeze(2) + 1e-8)).unsqueeze(3), (scene_geom[s_inds][:, :, 9:12] / (scene_geom[s_inds][:, :, 9:12].norm(dim=2).unsqueeze(2) + 1e-8)).unsqueeze(3), torch.cross( scene_geom[s_inds][:, :, 6:9] / (scene_geom[s_inds][:, :, 6:9].norm(dim=2).unsqueeze(2) + 1e-8), scene_geom[s_inds][:, :, 9:12] / (scene_geom[s_inds][:, :, 9:12].norm(dim=2).unsqueeze(2) + 1e-8) ).unsqueeze(3) ), dim = 3) s_out = ((s_r @ (((xyz - .5) * scene_geom[s_inds][:, :, :3]).unsqueeze(-1))).squeeze() + scene_geom[s_inds][:, :, 3:6]).squeeze() return s_out def findInters(ind_to_pc, scene_geom, ind_pairs=None): inters = {} if ind_pairs is None: l = list(ind_to_pc.keys()) l.sort() ind_pairs = [(a,b) for a,b in list( itertools.product(l,l) ) if a < b ] for ind1, ind2 in ind_pairs: # This makes us catch more bbox intersections thresh = scene_geom[[0 if ind1 < 0 else ind1, ind2]][:,:3].max() / DIM c1_inds, c2_inds = colEngine.findPCInter( ind_to_pc[ind1][0].T.unsqueeze(0), ind_to_pc[ind2][0].T.unsqueeze(0), ind_to_pc[ind1][1], ind_to_pc[ind2][1], thresh, ) if len(c1_inds.shape) == 0 or len(c2_inds.shape) == 0 or c1_inds.shape[0] == 0 or c2_inds.shape[0] == 0: continue inters[f"{ind1}_{ind2}"] = (c1_inds, c2_inds) return inters def vector_cos(norm1, norm2): norm1 = np.asarray(norm1) norm2 = np.asarray(norm2) dot = np.dot(norm1, norm2) magnitude = np.linalg.norm(norm1) * np.linalg.norm(norm2) if magnitude == 0.: return 0. return dot / float(magnitude) def orientProps(center, xd, yd, zd, xdir, ydir, zdir): rt = np.asarray([1., 0., 0.]) up = np.asarray([0., 1., 0.]) fwd = np.asarray([0., 0., 1.]) l = [ (xdir, xd, 0), (ydir, yd, 1), (zdir, zd, 2), (-1 * xdir, xd, 3), (-1 * ydir, yd, 4), (-1 * zdir, zd, 5) ] rtdir, rtd, rind = sorted( deepcopy(l), key=lambda x: vector_cos(rt, x[0]))[-1] if rind >= 3: l.pop(rind) l.pop((rind+3) % 6) else: l.pop((rind+3) % 6) l.pop(rind) for i in range(0, 4): p_ind = l[i][2] if p_ind > max(rind, (rind+3) % 6): l[i] = (l[i][0], l[i][1], l[i][2] - 2) elif p_ind > min(rind, (rind+3) % 6): l[i] = (l[i][0], l[i][1], l[i][2] - 1) updir, upd, upind = sorted( deepcopy(l), key=lambda x: vector_cos(up, x[0]))[-1] if upind >= 2: l.pop(upind) l.pop((upind+2) % 4) else: l.pop((upind+2) % 4) l.pop(upind) fwdir, fwd, _ = sorted(l, key=lambda x: vector_cos(fwd, x[0]))[-1] return { 'center': torch.tensor(center).float(), 'xd': torch.tensor(rtd).float(), 'yd': torch.tensor(upd).float(), 'zd': torch.tensor(fwd).float(), 'xdir': torch.tensor(rtdir).float(), 'ydir': torch.tensor(updir).float(), 'zdir': torch.tensor(fwdir).float() } def points_obb(points, precision): try: to_origin, size = trimesh.bounds.oriented_bounds(points, angle_digits=precision) center = to_origin[:3, :3].transpose().dot(-to_origin[:3, 3]) xdir = to_origin[0, :3] ydir = to_origin[1, :3] except scipy.spatial.qhull.QhullError: print('WARNING: falling back to PCA OBB computation since the more accurate minimum OBB computation failed.') center = points.mean(axis=0, keepdims=True) points = points - center center = center[0, :] pca = PCA() pca.fit(points) pcomps = pca.components_ points_local = np.matmul(pcomps, points.transpose()).transpose() size = points_local.max(axis=0) - points_local.min(axis=0) xdir = pcomps[0, :] ydir = pcomps[1, :] box = torch.from_numpy(np.hstack([center, size, xdir, ydir]).reshape(1, -1)).to(torch.float32) box = box.cpu().numpy().squeeze() center = box[:3] size = box[3:6] xdir = box[6:9] xdir /= np.linalg.norm(xdir) ydir = box[9:] ydir /= np.linalg.norm(ydir) zdir = np.cross(xdir, ydir) zdir /= np.linalg.norm(zdir) return orientProps(center, size[0], size[1], size[2], xdir, ydir, zdir) # Checks if any cube1 face midpoints are inside pc2 def checkFaces(cube1, cube2, scene_geom): ind1 = 0 if cube1 < 0 else cube1 ind2 = 0 if cube2 < 0 else cube2 faces = { 'right': (torch.tensor([1.0,0.5,0.5],device=device), 0, 0), 'left': (torch.tensor([0.0,0.5,0.5],device=device), 0, 1), 'top': (torch.tensor([0.5,1.0,0.5],device=device), 1, 0), 'bot': (torch.tensor([0.5,0.0,0.5],device=device), 1, 1), 'front': (torch.tensor([0.5,0.5,1.0],device=device), 2, 0), 'back': (torch.tensor([0.5,0.5,0.0],device=device), 2, 1), } if cube2 == -1: faces.pop('bot') faces.pop('left') faces.pop('right') faces.pop('front') faces.pop('back') if cube2 == -2: faces.pop('top') faces.pop('left') faces.pop('right') faces.pop('front') faces.pop('back') best_score = 1e8 best_face= None best_pt = None O = smp_pt(scene_geom[ind2], torch.zeros(3).to(device)) A = torch.stack([ scene_geom[ind2][0] * scene_geom[ind2][6:9], scene_geom[ind2][1] * scene_geom[ind2][9:12], scene_geom[ind2][2] * torch.cross( scene_geom[ind2][6:9], scene_geom[ind2][9:12] ) ]).T for face in faces: smpt, fi, ft = faces[face] pt = smp_pt(scene_geom[ind1], smpt) B = pt - O p = torch.tensor(np.linalg.solve(A.cpu(), B.cpu())) if p.min() < -0.01 or p.max() > 1.01: continue if ind2 == 0 and (p[1] >= .05 and p[1] <= .95): continue score = (p[fi] - ft).abs() if score < best_score: best_face = face best_score = score best_pt = pt return best_pt def calcAttPoint(pair, pcs, scene_geom, ind_to_pc): cube1 = int(pair.split('_')[0]) cube2 = int(pair.split('_')[1]) ind1 = cube1 if cube1 > 0 else 0 ind2 = cube2 if cube2 > 0 else 0 if ind1 == ind2: return None pc1, pc2 = pcs dim1 = scene_geom[ind1][0] * scene_geom[ind1][1] * scene_geom[ind1][2] dim2 = scene_geom[ind2][0] * scene_geom[ind2][1] * scene_geom[ind2][2] att_point = None if cube1 > 0: att_point = checkFaces(cube1, cube2, scene_geom) if att_point is None and cube2 > 0: att_point = checkFaces(cube2, cube1, scene_geom) if att_point is None: if ind1 > ind2: att_point = getBestATTPoint(pc1, pc2, cube1, cube2, ind_to_pc, ind1, ind2, scene_geom) else: att_point = getBestATTPoint(pc2, pc1, cube2, cube1, ind_to_pc, ind2, ind1, scene_geom) if att_point is None: return # att_point is point in 3D space, transform into local coordinate frames and reutrn collision = [ind1, ind2, None, None] for i, ind in ((2, ind1), (3, ind2)): O = smp_pt(scene_geom[ind], torch.zeros(3).to(device)) A = torch.stack([ scene_geom[ind][0] * scene_geom[ind][6:9], scene_geom[ind][1] * scene_geom[ind][9:12], scene_geom[ind][2] * torch.cross( scene_geom[ind][6:9], scene_geom[ind][9:12] ) ]).T B = att_point - O p = np.linalg.solve(A.cpu(), B.cpu()) p = np.clip(p, 0.0, 1.0) collision[i] = p.tolist() return collision def calcAnyPoint(pair, pcs, scene_geom, ind_to_pc): cube1 = int(pair.split('_')[0]) cube2 = int(pair.split('_')[1]) ind1 = cube1 if cube1 > 0 else 0 ind2 = cube2 if cube2 > 0 else 0 if ind1 == ind2: return None pc1, pc2 = pcs dim1 = scene_geom[ind1][0] * scene_geom[ind1][1] * scene_geom[ind1][2] dim2 = scene_geom[ind2][0] * scene_geom[ind2][1] * scene_geom[ind2][2] att_point = None att_point = getanATTPoint(pc1, pc2, cube1, cube2, ind_to_pc, ind1, ind2, scene_geom) if att_point is None: att_point = getanATTPoint(pc2, pc1, cube2, cube1, ind_to_pc, ind2, ind1, scene_geom) if att_point is None: return # att_point is point in 3D space, transform into local coordinate frames and reutrn collision = [ind1, ind2, None, None] for i, ind in ((2, ind1), (3, ind2)): O = smp_pt(scene_geom[ind], torch.zeros(3).to(device)) A = torch.stack([ scene_geom[ind][0] * scene_geom[ind][6:9], scene_geom[ind][1] * scene_geom[ind][9:12], scene_geom[ind][2] * torch.cross( scene_geom[ind][6:9], scene_geom[ind][9:12] ) ]).T B = att_point - O p = np.linalg.solve(A.cpu(), B.cpu()) p = np.clip(p, 0.0, 1.0) collision[i] = p.tolist() return collision def calcAttachments(inters, scene_geom, ind_to_pc): attachments = [] for pair in inters: attachments.append(calcAttPoint(pair, inters[pair], scene_geom, ind_to_pc)) return [a for a in attachments if a is not None] # Takes in intersections and an ind # Calculates intersection of all covered points # Checks amount that each face can be shortened # Updates parts to be shortened # Returns None if nothing shortened # Returns tuple of intersections to recalculate def shorten_cube(inters, parts, ind, scene_geom): cov_inds = torch.zeros((DIM**3)).float().to(device) cov_inters = [] nparts = [] for pair in inters: ind1 = int(pair.split('_')[0]) ind2 = int(pair.split('_')[1]) if ind1 == ind: cov_inters.append(pair) cov_inds[inters[pair][0]] = 1. nparts.append(ind2) elif ind2 == ind: cov_inters.append(pair) cov_inds[inters[pair][1]] = 1. nparts.append(ind1) if cov_inds.sum().item() == 0: return None face_shorten = {} dirs = { 'right': (fr, torch.tensor([-.01, 0., 0.]).unsqueeze(0).repeat(400,1), 0), 'left': (fl, torch.tensor([.01, 0., 0.]).unsqueeze(0).repeat(400,1), 0), 'top': (ft, torch.tensor([0., -.01, 0.]).unsqueeze(0).repeat(400,1), 1), 'bot': (fbo, torch.tensor([0., .01, 0.]).unsqueeze(0).repeat(400,1), 1), 'front': (ff, torch.tensor([0., 0., -.01]).unsqueeze(0).repeat(400,1), 2), 'back': (fba, torch.tensor([0., 0., .01]).unsqueeze(0).repeat(400,1), 2) } cube_geom = scene_geom[ind] p_r = torch.cat( ( (cube_geom[6:9] / (cube_geom[6:9].norm(dim=0).unsqueeze(0) + 1e-8)).unsqueeze(1), (cube_geom[9:12] / (cube_geom[9:12].norm(dim=0).unsqueeze(0) + 1e-8)).unsqueeze(1), torch.cross( cube_geom[6:9] / (cube_geom[6:9].norm(dim=0).unsqueeze(0) + 1e-8), cube_geom[9:12] / (cube_geom[9:12].norm(dim=0).unsqueeze(0) + 1e-8) ).unsqueeze(1) ), dim = 1) for d in dirs: dinds, inc, ki = dirs[d] inc = inc.to(device) for nd in range(101): points = ((p_r @ ((((g[dinds] + (inc * nd)).unsqueeze(0) - .5) * cube_geom[:3]).unsqueeze(-1))).squeeze() + cube_geom[3:6]).squeeze() inside = False for nind in nparts: if inside: break O = smp_pt(scene_geom[nind], torch.zeros(3).to(device)) A = torch.stack([ scene_geom[nind][0] * scene_geom[nind][6:9], scene_geom[nind][1] * scene_geom[nind][9:12], scene_geom[nind][2] * torch.cross( scene_geom[nind][6:9], scene_geom[nind][9:12] ) ]).T B = (points.T - O.unsqueeze(1)).cpu() p = torch.tensor(np.linalg.solve(A.cpu(), B.cpu())).T if p.min() >= -0.01 and p.max() <= 1.01: inside = True if not inside: break face_shorten[d] = max((nd-1) / 100, 0) dims = [ ('right', 'left', 'xd', 'xdir'), ('top', 'bot', 'yd', 'ydir'), ('front', 'back', 'zd', 'zdir') ] for d1, d2, key, dn in dims: if face_shorten[d1] > 0 or face_shorten[d2] > 0: x = face_shorten[d1] * parts[ind][key] y = face_shorten[d2] * parts[ind][key] parts[ind][key] -= x + y parts[ind]['center'] += parts[ind][dn] * ((y-x) / 2) return cov_inters def getBestATTPoint(pc1, pc2, cube1, cube2, ind_to_pc, ind1, ind2, scene_geom): a = ind_to_pc[cube1][0][pc1] b = ind_to_pc[cube2][0][pc2] joint = torch.cat((a, b), dim=0) jmin = joint.min(dim=0).values jmax = joint.max(dim=0).values xd = (jmax[0] - jmin[0]).abs() yd = (jmax[1] - jmin[1]).abs() zd = (jmax[2] - jmin[2]).abs() center = (jmax + jmin) / 2 bbox_geom = torch.cat(( xd.unsqueeze(0), yd.unsqueeze(0), zd.unsqueeze(0), center, torch.tensor([1.0,0.0,0.0]).to(device), torch.tensor([0.0,1.0,0.0]).to(device) )) p_r = torch.cat( ( (bbox_geom[6:9] / (bbox_geom[6:9].norm(dim=0).unsqueeze(0) + 1e-8)).unsqueeze(1), (bbox_geom[9:12] / (bbox_geom[9:12].norm(dim=0).unsqueeze(0) + 1e-8)).unsqueeze(1), torch.cross( bbox_geom[6:9] / (bbox_geom[6:9].norm(dim=0).unsqueeze(0) + 1e-8), bbox_geom[9:12] / (bbox_geom[9:12].norm(dim=0).unsqueeze(0) + 1e-8) ).unsqueeze(1) ), dim = 1) points = ((p_r @ (((attxyz - .5) * bbox_geom[:3]).unsqueeze(-1))).squeeze() + bbox_geom[3:6]).squeeze() # for each point find atp1 and atp2 O1 = smp_pt(scene_geom[ind1], torch.zeros(3).to(device)) A1 = torch.stack([ scene_geom[ind1][0] * scene_geom[ind1][6:9], scene_geom[ind1][1] * scene_geom[ind1][9:12], scene_geom[ind1][2] * torch.cross( scene_geom[ind1][6:9], scene_geom[ind1][9:12] ) ]).T O2 = smp_pt(scene_geom[ind2], torch.zeros(3).to(device)) A2 = torch.stack([ scene_geom[ind2][0] * scene_geom[ind2][6:9], scene_geom[ind2][1] * scene_geom[ind2][9:12], scene_geom[ind2][2] * torch.cross( scene_geom[ind2][6:9], scene_geom[ind2][9:12] ) ]).T B1 = (points.T - O1.unsqueeze(1)).cpu() B2 = (points.T - O2.unsqueeze(1)).cpu() atp1 = torch.tensor(np.linalg.lstsq(A1.cpu(), B1, rcond=None)[0]).T atp2 = torch.tensor(np.linalg.lstsq(A2.cpu(), B2, rcond=None)[0]).T atps = torch.cat((atp1, atp2), dim = 1) ne_inds = (((atps >= -0.01).sum(dim=1) == 6).int() + ((atps <= 1.01).sum(dim=1) == 6)).int() if ind2 == 0: for i in range(1, 6): offset = 0.01 * i bb_ne_inds = (((atps[:, 4] <= offset).int() + (atps[:, 4] >= (1-offset)).int()) == 1).int() bb_ne_inds += ne_inds.int() bb_ne_inds = (bb_ne_inds == 3).nonzero().squeeze() if bb_ne_inds.sum() > 0: return points[bb_ne_inds].mean(dim=0) return None ne_inds = (ne_inds == 2).nonzero().squeeze() if ne_inds.sum() > 0: return points[ne_inds].mean(dim=0) return None def getanATTPoint(pc1, pc2, cube1, cube2, ind_to_pc, ind1, ind2, scene_geom): a = ind_to_pc[cube1][0][pc1] b = ind_to_pc[cube2][0][pc2] joint = torch.cat((a, b), dim=0) jmin = joint.min(dim=0).values jmax = joint.max(dim=0).values xd = (jmax[0] - jmin[0]).abs() yd = (jmax[1] - jmin[1]).abs() zd = (jmax[2] - jmin[2]).abs() center = (jmax + jmin) / 2 bbox_geom = torch.cat(( xd.unsqueeze(0), yd.unsqueeze(0), zd.unsqueeze(0), center, torch.tensor([1.0,0.0,0.0]).to(device), torch.tensor([0.0,1.0,0.0]).to(device) )) p_r = torch.cat( ( (bbox_geom[6:9] / (bbox_geom[6:9].norm(dim=0).unsqueeze(0) + 1e-8)).unsqueeze(1), (bbox_geom[9:12] / (bbox_geom[9:12].norm(dim=0).unsqueeze(0) + 1e-8)).unsqueeze(1), torch.cross( bbox_geom[6:9] / (bbox_geom[6:9].norm(dim=0).unsqueeze(0) + 1e-8), bbox_geom[9:12] / (bbox_geom[9:12].norm(dim=0).unsqueeze(0) + 1e-8) ).unsqueeze(1) ), dim = 1) points = ((p_r @ (((attxyz - .5) * bbox_geom[:3]).unsqueeze(-1))).squeeze() + bbox_geom[3:6]).squeeze() # for each point find atp1 and atp2 O1 = smp_pt(scene_geom[ind1], torch.zeros(3).to(device)) A1 = torch.stack([ scene_geom[ind1][0] * scene_geom[ind1][6:9], scene_geom[ind1][1] * scene_geom[ind1][9:12], scene_geom[ind1][2] * torch.cross( scene_geom[ind1][6:9], scene_geom[ind1][9:12] ) ]).T O2 = smp_pt(scene_geom[ind2], torch.zeros(3).to(device)) A2 = torch.stack([ scene_geom[ind2][0] * scene_geom[ind2][6:9], scene_geom[ind2][1] * scene_geom[ind2][9:12], scene_geom[ind2][2] * torch.cross( scene_geom[ind2][6:9], scene_geom[ind2][9:12] ) ]).T B1 = (points.T - O1.unsqueeze(1)).cpu() B2 = (points.T - O2.unsqueeze(1)).cpu() atp1 = torch.tensor(np.linalg.lstsq(A1.cpu(), B1, rcond=None)[0]).T atp2 = torch.tensor(np.linalg.lstsq(A2.cpu(), B2, rcond=None)[0]).T atps = torch.cat((atp1, atp2), dim = 1) ne_inds = (((atps >= -0.05).sum(dim=1) == 6).int() + ((atps <= 1.05).sum(dim=1) == 6)).int() if ind2 == 0: for i in range(1, 6): offset = 0.01 * i bb_ne_inds = (((atps[:, 4] <= offset).int() + (atps[:, 4] >= (1-offset)).int()) == 1).int() bb_ne_inds += ne_inds.int() bb_ne_inds = (bb_ne_inds == 3).nonzero().squeeze() if bb_ne_inds.sum() > 0: return points[bb_ne_inds].mean(dim=0) return None ne_inds = (ne_inds == 2).nonzero().squeeze() if ne_inds.sum() > 0: return points[ne_inds].mean(dim=0) return None
31,198
31.163918
149
py
ShapeMOD
ShapeMOD-main/SA_lang/parse_files/symmetry.py
import torch import numpy as np import json_parse as jp import old_intersect as inter from copy import deepcopy import math SDIM_THRESH = .15 SANG_THRESH = .1 SPOS_THRESH = .1 VATT_THRESH = .05 CATT_THRESH = .3 GROUP_THRESH = .1 CPT_THRESH = .1 def smp_pt(geom, pt): xdir = geom[6:9] / (geom[6:9].norm() + 1e-8) ydir = geom[9:12] / (geom[9:12].norm() + 1e-8) zdir = torch.cross(xdir, ydir) r = torch.stack(( xdir, ydir, zdir )).T return (r @ ((pt -.5) * geom[:3])) + geom[3:6] def smp_rel_pos(geom, gpt): O = smp_pt(geom, torch.zeros(3)) A = torch.stack([ geom[0] * geom[6:9], geom[1] * geom[9:12], geom[2] * torch.cross( geom[6:9], geom[9:12] ) ]).T B = gpt - O p = torch.tensor(np.linalg.solve(A.cpu(), B.cpu())) return p def approx_same_dims(a, b, m): if ((a['xd'] - b['xd']).abs() / m) > SDIM_THRESH or \ ((a['yd'] - b['yd']).abs() / m) > SDIM_THRESH or \ ((a['zd'] - b['zd']).abs() / m) > SDIM_THRESH: return False return True def approx_same_point(a, b, m, thresh=SPOS_THRESH): if (a-b).norm() / m > thresh: return False return True # Checks if pt is a valid place to be for cuboid c def isValidAtt(node, pt, cind): ct = cubeToTensor(node['cubes'][cind]) O = smp_pt(ct, torch.zeros(3)) A = torch.stack([ ct[0] * ct[6:9], ct[1] * ct[9:12], ct[2] * torch.cross( ct[6:9], ct[9:12] ) ]).T B = pt - O p = torch.tensor(np.linalg.solve(A.cpu(), B.cpu())) if cind == 0 and (p[1] - .5).abs() < .5 - VATT_THRESH: return False if (p - .5).abs().max() >= .5 + VATT_THRESH: return False return True # checks if pt location in c is close to att def isCloseAtt(node, pt, cind, att): ct = cubeToTensor(node['cubes'][cind]) O = smp_pt(ct, torch.zeros(3)) A = torch.stack([ ct[0] * ct[6:9], ct[1] * ct[9:12], ct[2] * torch.cross( ct[6:9], ct[9:12] ) ]).T B = pt - O p = torch.tensor(np.linalg.solve(A.cpu(), B.cpu())) if (p - torch.tensor(att)).norm() > CATT_THRESH: return False return True # checks if pt location in c is close to att def isClosePt(node, pt, cind, att): ct = cubeToTensor(node['cubes'][cind]) spt = smp_pt(ct, torch.tensor(att)) if (pt - spt).norm() > CPT_THRESH: return False return True def cubeToTensor(c): return torch.cat(( c['xd'].unsqueeze(0), c['yd'].unsqueeze(0), c['zd'].unsqueeze(0), c['center'], c['xdir'], c['ydir'] )) # Is there a translational symmetry from i -> o # if so return (axis, scale in terms of bbox dimension) def checkTranSym(node, i, o): c_i = node['cubes'][i] c_o = node['cubes'][o] if not approx_same_dims(c_i, c_o, max(node['cubes'][0]['xd'], node['cubes'][0]['yd'], node['cubes'][0]['zd'])): return None cdir = c_o['center'] - c_i['center'] scale = cdir.norm() cdir /= cdir.norm() if cdir.dot(node['cubes'][0]['xdir']) > 1 - SANG_THRESH: tn = 'tr_X' td = node['cubes'][0]['xd'] tdir = node['cubes'][0]['xdir'] elif cdir.dot(node['cubes'][0]['ydir']) > 1 - SANG_THRESH: tn = 'tr_Y' td = node['cubes'][0]['yd'] tdir = node['cubes'][0]['ydir'] elif cdir.dot(node['cubes'][0]['zdir']) > 1 - SANG_THRESH: tn = 'tr_Z' td = node['cubes'][0]['zd'] tdir = node['cubes'][0]['zdir'] else: return None for n, at1 in node['pc_atts'][i]: opt = smp_pt( cubeToTensor(c_i), torch.tensor(at1) ) tpt = opt + (tdir * scale) #if not isValidAtt(node, tpt, n) or not isCloseAtt(node, tpt, o, at1): if (isValidAtt(node, opt, n) and not isValidAtt(node, tpt, n)) or not isClosePt(node, tpt, o, at1): return None return tn, (scale/td).item() def getRefMatrixHomo(axis, center): m = center d = axis / axis.norm() refmat = torch.stack(( torch.stack((1 - 2 * d[0] * d[0], -2 * d[0] * d[1], -2 * d[0] * d[2], 2 * d[0] * d[0] * m[0] + 2 * d[0] * d[1] * m[1] + 2 * d[0] * d[2] * m[2])), torch.stack((-2 * d[1] * d[0], 1 - 2 * d[1] * d[1], -2 * d[1] * d[2], 2 * d[1] * d[0] * m[0] + 2 * d[1] * d[1] * m[1] + 2 * d[1] * d[2] * m[2])), torch.stack((-2 * d[2] * d[0], -2 * d[2] * d[1], 1 - 2 * d[2] * d[2], 2 * d[2] * d[0] * m[0] + 2 * d[2] * d[1] * m[1] + 2 * d[2] * d[2] * m[2])) )) return refmat def reflect_cube(c, center, ndir): ref_c = {} pad = torch.nn.ConstantPad1d((0, 1), 1.0) reflection = getRefMatrixHomo(ndir, center) nreflection = torch.cat((reflection[:,:3], torch.zeros(3,1)), dim=1) posHomo = pad(c['center']) ref_c['center'] = reflection @ posHomo xHomo = pad(c['xdir']) yHomo = pad(c['ydir']) zHomo = pad(c['zdir']) ref_c['xdir'] = nreflection @ xHomo ref_c['ydir'] = nreflection @ yHomo ref_c['zdir'] = nreflection @ zHomo return ref_c def reflect_point(p, center, ndir): pad = torch.nn.ConstantPad1d((0, 1), 1.0) reflection = getRefMatrixHomo(ndir, center) posHomo = pad(p) return reflection @ posHomo # Is there reflectional symmetry from i->o # If so return (plane) def checkRefSym(node, i, o): c_i = node['cubes'][i] c_o = node['cubes'][o] if not approx_same_dims(c_i, c_o, max(node['cubes'][0]['xd'], node['cubes'][0]['yd'], node['cubes'][0]['zd'])): return None cdir = c_o['center'] - c_i['center'] cdir /= cdir.norm() if cdir.dot(node['cubes'][0]['xdir']) > 1 - SANG_THRESH: rn = 'ref_X' rdir = node['cubes'][0]['xdir'] elif cdir.dot(node['cubes'][0]['ydir']) > 1 - SANG_THRESH: rn = 'ref_Y' rdir = node['cubes'][0]['ydir'] elif cdir.dot(node['cubes'][0]['zdir']) > 1 - SANG_THRESH: rn = 'ref_Z' rdir = node['cubes'][0]['zdir'] else: return None center = smp_pt(cubeToTensor(node['cubes'][0]), torch.tensor([.5, .5, .5])) c_ref = reflect_cube(c_i, center, rdir) if not approx_same_point( c_ref['center'], c_o['center'], max(node['cubes'][0]['xd'], node['cubes'][0]['yd'], node['cubes'][0]['zd']) ): return None for n, at1 in node['pc_atts'][i]: opt = smp_pt( cubeToTensor(c_i), torch.tensor(at1) ) rpt = reflect_point(opt, center, rdir) if rn == 'ref_X': rat1 = [1-at1[0], at1[1], at1[2]] elif rn == 'ref_Y': rat1 = [at1[0], 1-at1[1], at1[2]] elif rn == 'ref_Z': rat1 = [at1[0], at1[1], 1-at1[2]] #if not isValidAtt(node, rpt, n) or not isCloseAtt(node, rpt, o, rat1): if (isValidAtt(node, opt, n) and not isValidAtt(node, rpt, n)) or not isClosePt(node, rpt, o, rat1): return None return rn # takes in a dict of symmetries, if any of the translational symmetries have more than one element try to group them # returned group symmetries, and the longest symmetry seen def groupSyms(syms): gsyms = {} longest = 2 for st in syms: if 'tr' not in st: gsyms[st] = syms[st] elif len(syms[st]) == 1: gsyms[st] = (syms[st][0][0], [syms[st][0][1]]) else: best_l = 0 syms[st].sort() for l in range(0, len(syms[st])): md = syms[st][l][0] exps = (torch.tensor([(l+1)*s[0] / md for s in syms[st][:l+1]]) - torch.arange(1, l+2)).abs().mean() if exps > GROUP_THRESH: break best_l = l group = [s[1] for s in syms[st][:best_l+1]] longest = max(longest, len(group) + 1) gsyms[st] = (syms[st][best_l][0], group) return gsyms, longest # returns list of lists of indices, where each sub-list is in the same semantic group def getSemanticGroups(node): groups = {} #for i, l in enumerate(node['sem_labels'].argmax(dim=1).tolist()): for i, l in enumerate(node['children_names']): #spec_add(groups, l, i) spec_add(groups, 'all', i) return groups def spec_add(d, k, v): if k in d: d[k].append(v) else: d[k] = [v] def checkSameAtts(node, ind, oind): pc_atts = node['pc_atts'] i_atts = [i[0] for i in pc_atts[ind]] o_atts = [i[0] for i in pc_atts[oind]] i_atts.sort() o_atts.sort() return (i_atts == o_atts) # For the indices in group, returns all symmetries where index is canonical member def getSymsForGroup(node, group): if len(group) == 1: return {} if 'syms' in node: prev_syms = set([s[0] for s in node['syms']]) else: prev_syms = set([]) syms = {} for ind in group: nsyms = {} for oind in group: if ind == oind or ind in prev_syms or oind in prev_syms: continue if not checkSameAtts(node, ind, oind): continue trans_sym = checkTranSym(node, ind, oind) ref_sym = checkRefSym(node, ind, oind) if trans_sym is not None: spec_add(nsyms, trans_sym[0], (trans_sym[1], oind)) if ref_sym is not None: spec_add(nsyms, ref_sym, oind) ngsyms, longest = groupSyms(nsyms) if len(ngsyms) > 0: syms[ind] = (longest, ngsyms) return syms def getLongestTranSym(ind, sg, mv): for st in ['tr_X', 'tr_Y', 'tr_Z']: if st in sg[1] and len(sg[1][st][1]) == (mv-1) : ntr = sg[1][st][1] ns = [ind, st, len(sg[1][st][1]), sg[1][st][0], ntr] return ns, ntr # Any node that shows up in ntr, remove all symmetries that they are involved in in pot_syms def removePotSyms(pot_syms, ntr): for n in ntr: if n in pot_syms: pot_syms.pop(n) ntr = set(ntr) ps_inds = [] for i in pot_syms: sts_to_pop = [] for st in pot_syms[i][1]: if 'ref' in st: if pot_syms[i][1][st][0] in ntr: sts_to_pop.append(st) elif 'tr' in st: js_to_pop = [] for j in range(len(pot_syms[i][1][st][1])): if pot_syms[i][1][st][1][j] in ntr: js_to_pop = [j] + js_to_pop for j in js_to_pop: pot_syms[i][1][st][1].pop(j) if len(pot_syms[i][1][st][1]) == 0: sts_to_pop.append(st) elif len(js_to_pop) > 0: pot_syms[i] = (pot_syms[i][0] - len(js_to_pop), pot_syms[i][1]) for st in sts_to_pop: pot_syms[i][1].pop(st) ps_inds = [i] + ps_inds for i in ps_inds: if len(pot_syms[i][1]) == 0: pot_syms.pop(i) # Takes in all of the symmetries in a group, chooses 'best' one, removes all other ones from the other nodes def getMaxGroupSyms(pot_syms, syms = [], tr = []): if len(pot_syms) == 0: return syms, tr ml = 2 mi = None for ind in pot_syms: if pot_syms[ind][0] > ml: ml = pot_syms[ind][0] mi = ind if mi is not None: ns, ntr = getLongestTranSym(mi, pot_syms[mi], ml) removePotSyms(pot_syms, [mi] + ntr) return getMaxGroupSyms(pot_syms, syms + [ns], tr + ntr) for st in ['ref_X', 'ref_Y', 'ref_Z']: for ind in pot_syms: if st in pot_syms[ind][1]: ntr = pot_syms[ind][1][st] ns = [ind, st, ntr] removePotSyms(pot_syms, [ind] + ntr) return getMaxGroupSyms(pot_syms, syms + [ns], tr + ntr) for st in ['tr_X', 'tr_Y', 'tr_Z']: for ind in pot_syms: if st in pot_syms[ind][1]: ntr = [pot_syms[ind][1][st][1][0]] ns = [ind, st, 1, pot_syms[ind][1][st][0], ntr] removePotSyms(pot_syms, [ind] + ntr) return getMaxGroupSyms(pot_syms, syms + [ns], tr + ntr) # remove all nodes to remove from the list, from children, remove all attachments that mention these nodes def cleanNode(node, ntr, syms): if len(syms) == 0: if 'syms' not in node: node['syms'] = [] return sntr = [] num_cubes_per_sym = {} for sym in syms: sym_cubes = [] num_cubes_per_sym[sym[0]] = len(sym[-1]) for itr in sym[-1]: sntr.append(itr) sym_cubes.append(node['cubes'][itr]) sym[-1] = sym_cubes syms.sort() sym_cubes = [] for sym in syms: scubes = sym.pop(-1) sym_cubes += scubes sntr.sort(reverse=True) cube_map = {} count = 0 for i in range(len(node['cubes'])): if i not in ntr: cube_map[i] = count count += 1 for n in sntr: for key in ['children', 'children_names', 'cubes']: node[key].pop(n) atts = [] temp_atts = node.pop('attachments') for att in temp_atts: if att[0] not in ntr and att[1] not in ntr: atts.append((cube_map[att[0]], cube_map[att[1]], att[2], att[3])) if 'syms' in node: node['syms'] += syms else: node['syms'] = syms node['syms'] = [[cube_map[s[0]]] + s[1:] for s in node['syms']] node['attachments'] = atts def getMaxSyms(pot_group_syms): nodes_to_remove = set() syms = [] for pgs in pot_group_syms: if len(pgs) == 0: continue group_syms, gntr = getMaxGroupSyms(pgs) syms += group_syms nodes_to_remove = nodes_to_remove.union(set(gntr)) return syms, nodes_to_remove def getPerCubeAtts(node): pc_atts = {} for c1, c2, at1, at2 in node['attachments']: spec_add(pc_atts, c1, (c2, at1)) spec_add(pc_atts, c2, (c1, at2)) return pc_atts def addSimpSymmetries(node): for child in node['children']: if len(child) > 0: addSimpSymmetries(child) node['pc_atts'] = getPerCubeAtts(node) groups = getSemanticGroups(node) pot_group_syms = [getSymsForGroup(node, groups[g]) for g in groups] syms, nodes_to_remove = getMaxSyms(pot_group_syms) cleanNode(node, nodes_to_remove, syms) ########################## def checkXRefSym(node, i, o): c_i = node['cubes'][i] c_o = node['cubes'][o] if not approx_same_dims(c_i, c_o, max(node['cubes'][0]['xd'], node['cubes'][0]['yd'], node['cubes'][0]['zd'])): return False cdir = c_o['center'] - c_i['center'] cdir /= cdir.norm() if cdir.dot(node['cubes'][0]['xdir']) < 1 - SANG_THRESH: return False center = smp_pt(cubeToTensor(node['cubes'][0]), torch.tensor([.5, .5, .5])) rdir = node['cubes'][0]['xdir'] c_ref = reflect_cube(c_i, center, rdir) if not approx_same_point( c_ref['center'], c_o['center'], max(node['cubes'][0]['xd'], node['cubes'][0]['yd'], node['cubes'][0]['zd']) ): return False for n, at1 in node['pc_atts'][i]: opt = smp_pt( cubeToTensor(c_i), torch.tensor(at1) ) rpt = reflect_point(opt, center, rdir) rat1 = [1-at1[0], at1[1], at1[2]] #if not isCloseAtt(node, rpt, o, rat1): if not isClosePt(node, rpt, o, rat1): return False return True # Get all nodes that have a reflectional symmetry # For now just do this over x-axis # Check that all attachments are close in the opposite member def getRefSyms(node): nn = len(node['cubes']) ref_syms = {} for ind in range(1, nn): if len(node['children'][ind]) > 0: continue for oind in range(1, nn): if ind == oind: continue if len(node['children'][oind]) > 0: continue if checkXRefSym(node, ind, oind): spec_add(ref_syms, ind, oind) return ref_syms # Get all connected components that share the same symmetry, with more than one member # (for rotational just to take the ones with the "largest" # of members) def getRefConnComps(pot_ref_syms, node): groups = [] added = set() for i in pot_ref_syms: if i in added: continue group = [i] pc_atts = set(n[0] for n in node['pc_atts'][i]) added.add(i) again = True while(again): again = False for j in pot_ref_syms: if j not in added: if j in pc_atts: group.append(j) added.add(j) pc_atts = pc_atts.union(set(n[0] for n in node['pc_atts'][j])) again=True if len(group) > 1: groups.append([ group, [pot_ref_syms[g][0] for g in group] ]) return groups def allConnected(a, b, node): prev_atts = node['pc_atts'] for (_, ind), (_, oind) in zip(a, b): found = False for n, _ in prev_atts[ind]: if n == oind: found = True if found is False: return False return True # Make sure for each attachment to a cube 'outside' of the group, that the attachment is Valid def checkValidRefAtts(ref_ccs, node): valid_groups = [] center = smp_pt(cubeToTensor(node['cubes'][0]), torch.tensor([.5, .5, .5])) rdir = node['cubes'][0]['xdir'] for group in ref_ccs: isValid = True mems = set(group[0]) for i in group[0]: for n, at1 in node['pc_atts'][i]: if n in mems: continue opt = smp_pt( cubeToTensor(node['cubes'][i]), torch.tensor(at1) ) rpt = reflect_point(opt, center, rdir) if isValidAtt(node, opt, n) and not isValidAtt(node, rpt, n): isValid = False if isValid: valid_groups.append(group) return valid_groups # For any group of valid ref syms, turn it into a sub-program # This involves: # - re-ordering cubes/children # - finding new OBB # - refinding all attachments # Adding this information to node['syms'] def createRefSubPrograms(valid_ref_syms, node): node.pop('pc_atts') node.pop('attachments') new_progs = [] to_remove = [] to_head = [] count = 0 for group in valid_ref_syms: new_prog = { 'name': f'{node["name"]}_ssp_{count}', 'children': [], 'children_names': [], 'cubes': [], } count += 1 to_head.append(group[0][0]) for i in group[0]: to_remove.append(i) new_prog['children'].append(node['children'][i]) new_prog['children_names'].append(node['children_names'][i]) new_prog['cubes'].append(node['cubes'][i]) bbox = jp.getOBB(new_prog['cubes']) new_prog['children'] = [{}] + new_prog['children'] new_prog['children_names'] = ["bbox"] + new_prog['children_names'] new_prog['cubes'] = [bbox] + new_prog['cubes'] ind_to_pc, scene_geom = inter.samplePC(new_prog['cubes'], split_bbox=True) inters = inter.findInters(ind_to_pc, scene_geom) new_prog['attachments'] = inter.calcAttachments(inters, scene_geom, ind_to_pc) for o in group[1]: to_remove.append(o) new_progs.append(new_prog) for ind, new_prog in zip(to_head, new_progs): node['children'][ind] = new_prog node['cubes'][ind] = new_prog['cubes'][0] node['children_names'][ind] = new_prog['name'] cube_map = {} count = 0 for i in range(len(node['cubes'])): if i not in (set(to_remove) - set(to_head)): cube_map[i] = count count += 1 node['syms'] = [[cube_map[th], 'ref_X'] for th in to_head] to_remove.sort(reverse=True) to_head = set(to_head) for tr in to_remove: if tr in to_head: continue for key in ['children', 'cubes', 'children_names']: node[key].pop(tr) to_remove = set(to_remove) - to_head ind_to_pc, scene_geom = inter.samplePC(node['cubes'], split_bbox=True) inters = inter.findInters(ind_to_pc, scene_geom) node['attachments'] = inter.calcAttachments(inters, scene_geom, ind_to_pc) def addSymSubPrograms(node): addRefSymSubPrograms(node) def addRefSymSubPrograms(node): for child in node['children']: if len(child) > 0: addRefSymSubPrograms(child) node['pc_atts'] = getPerCubeAtts(node) pot_ref_syms = getRefSyms(node) if len(pot_ref_syms) == 0: return ref_ccs = getRefConnComps(pot_ref_syms, node) if len(ref_ccs) == 0: return valid_ref_syms = checkValidRefAtts(ref_ccs, node) if len(valid_ref_syms) == 0: return createRefSubPrograms(valid_ref_syms, node)
22,111
26.571072
153
py
ShapeMOD
ShapeMOD-main/SA_lang/parse_files/json_parse.py
import torch import sys import ast import numpy as np import random import os import pickle import old_intersect as inter import random from copy import deepcopy import networkx as nx import symmetry as sym VERBOSE = False DO_SHORTEN = True DO_SIMP_SYMMETRIES = True DO_SQUEEZE = True DO_VALID_CHECK = True DO_NORM_AA_CUBES = True DO_SEM_FLATTEN = True DO_SEM_REHIER = True SQUARE_THRESH = 0.1 SD_THRESH = 0.01 AA_THRESH = 0.995 #?? SCOL_MAP = { 'chair': set(['caster', 'mechanical_control']), 'table': set(['caster', 'cabinet_door', 'drawer', 'keyboard_tray']), 'storage': set(['drawer', 'cabinet_door', 'mirror', 'caster']) } SFLAT_MAP = { 'chair': set(['chair_back', 'chair_arm', 'chair_base', 'chair_seat', 'footrest', 'chair_head']), 'table': set(['tabletop', 'table_base']), 'storage': set(['cabinet_frame', 'cabinet_base']) } SRH_MAP = { 'storage':('cabinet_frame', set(['countertop', 'shelf', 'drawer', 'cabinet_door', 'mirror'])) } def isSquare(a, b, c): v = (a - b).abs()/max(a.item(), b.item(), c.item()) < SQUARE_THRESH return v def isAxisAligned(props, bbox, thresh=AA_THRESH): xdir = props['xdir'] ydir = props['ydir'] zdir = props['zdir'] xdir /= xdir.norm() ydir /= ydir.norm() zdir /= zdir.norm() if xdir.dot(bbox['xdir']) < thresh: return False if ydir.dot(bbox['ydir']) < thresh: return False if zdir.dot(bbox['zdir']) < thresh: return False return True def shouldChange(props, bbox): if isAxisAligned(props, bbox): return True xsquare = isSquare(props['zd'], props['yd'], props['xd']) ysquare = isSquare(props['zd'], props['xd'], props['yd']) zsquare = isSquare(props['xd'], props['yd'], props['zd']) if xsquare and ysquare and zsquare: return True if xsquare: return isSpecAxisAligned(props['xdir'], bbox['xdir']) if ysquare: return isSpecAxisAligned(props['ydir'], bbox['ydir']) if zsquare: return isSpecAxisAligned(props['zdir'], bbox['zdir']) return False def isSpecAxisAligned(cdir, axis): cdir /= cdir.norm() if cdir.dot(axis) >= AA_THRESH: return True else: return False def getDataPath(category): if 'storage' == category: category += 'furniture' return f"/home/{os.getenv('USER')}/data/{category}_hier/" def getSemOrder(category): if category == "chair": sem_order_path = "stats/part_semantics/PGP-Chair.txt" elif category == "storage": sem_order_path = "stats/part_semantics/PGP-Storage.txt" elif category == "table": sem_order_path = "stats/part_semantics/PGP-Table.txt" else: assert False, f'Invalid Category {category}' sem_order = {"bbox": "-1","other":"100"} with open(sem_order_path) as f: for line in f: sem_order[line.split()[1].split('/')[-1]] = line.split()[0] return sem_order def cubeOrder(cubes, names, sem_order): d = [] min_c = np.array([1e8,1e8,1e8]) max_c = np.array([-1e8,-1e8,-1e8]) for rw in cubes: min_c = np.min((min_c, rw['center'].numpy()), axis = 0) max_c = np.max((max_c, rw['center'].numpy()), axis = 0) mac = np.max(max_c) mic = np.min(min_c) for c_ind, (rw, name) in enumerate(zip(cubes, names)): sc = (rw['center'].numpy() - mic) / (mac - mic) x_r = round(sc[0]*2)/2 y_r = round(sc[1]*2)/2 z_r = round(sc[2]*2)/2 d.append(( int(sem_order[name]), x_r + y_r + z_r, x_r, y_r, z_r, sc[0], sc[1], sc[2], c_ind )) d.sort() return [c_ind for _,_,_,_,_,_,_,_,c_ind in d] Sbbox = { 'xdir': torch.tensor([1.0, 0.0, 0.0]), 'ydir': torch.tensor([0.0, 1.0, 0.0]), 'zdir': torch.tensor([0.0, 0.0, 1.0]), } def vector_cos(norm1, norm2): norm1 = np.asarray(norm1) norm2 = np.asarray(norm2) dot = np.dot(norm1, norm2) magnitude = np.linalg.norm(norm1) * np.linalg.norm(norm2) if magnitude == 0.: return 0. return dot / float(magnitude) def orientProps(center, xd, yd, zd, xdir, ydir, zdir): rt = np.asarray([1., 0., 0.]) up = np.asarray([0., 1., 0.]) fwd = np.asarray([0., 0., 1.]) l = [ (xdir, xd, 0), (ydir, yd, 1), (zdir, zd, 2), (-1 * xdir, xd, 3), (-1 * ydir, yd, 4), (-1 * zdir, zd, 5) ] rtdir, rtd, rind = sorted( deepcopy(l), key=lambda x: vector_cos(rt, x[0]))[-1] if rind >= 3: l.pop(rind) l.pop((rind+3) % 6) else: l.pop((rind+3) % 6) l.pop(rind) for i in range(0, 4): p_ind = l[i][2] if p_ind > max(rind, (rind+3) % 6): l[i] = (l[i][0], l[i][1], l[i][2] - 2) elif p_ind > min(rind, (rind+3) % 6): l[i] = (l[i][0], l[i][1], l[i][2] - 1) updir, upd, upind = sorted( deepcopy(l), key=lambda x: vector_cos(up, x[0]))[-1] if upind >= 2: l.pop(upind) l.pop((upind+2) % 4) else: l.pop((upind+2) % 4) l.pop(upind) fwdir, fwd, _ = sorted(l, key=lambda x: vector_cos(fwd, x[0]))[-1] return { 'center': torch.tensor(center).float(), 'xd': torch.tensor(rtd).float(), 'yd': torch.tensor(upd).float(), 'zd': torch.tensor(fwd).float(), 'xdir': torch.tensor(rtdir).float(), 'ydir': torch.tensor(updir).float(), 'zdir': torch.tensor(fwdir).float() } def jsonToProps(json): json = np.array(json) center = np.array(json[:3]) xd = json[3] yd = json[4] zd = json[5] xdir = json[6:9] xdir /= np.linalg.norm(xdir) ydir = json[9:] ydir /= np.linalg.norm(ydir) zdir = np.cross(xdir, ydir) zdir /= np.linalg.norm(zdir) if xd < SD_THRESH or yd < SD_THRESH or zd < SD_THRESH: return None props = orientProps(center, xd, yd, zd, xdir, ydir, zdir) if DO_NORM_AA_CUBES: if shouldChange(props, Sbbox): props['xdir'] = Sbbox['xdir'].clone().detach() props['ydir'] = Sbbox['ydir'].clone().detach() props['zdir'] = Sbbox['zdir'].clone().detach() return props def addAttachments(node, sem_order): co = cubeOrder(node['cubes'], node['children_names'], sem_order) for key in ['cubes', 'children', 'children_names']: node[key] = [node[key][i] for i in co] ind_to_pc, scene_geom = inter.samplePC(node['cubes'], split_bbox=True) inters = inter.findInters(ind_to_pc, scene_geom) node['attachments'] = inter.calcAttachments(inters, scene_geom, ind_to_pc) for child in node['children']: if len(child) > 0: addAttachments(child, sem_order) # From raw json, get graph structure and all leaf cuboids def getShapeHier(ind, category): data_path = getDataPath(category) with open(data_path + ind + ".json") as f: json = None for line in f: json = ast.literal_eval(line) hier = {} queue = [(json, hier)] while len(queue) > 0: json, node = queue.pop(0) if "children" not in json: continue name = json["label"] # Don't add sub-programs when we collapse collapse_names = SCOL_MAP[category] if name in collapse_names: continue while("children" in json and len(json["children"])) == 1: json = json["children"][0] if "children" not in json: continue node.update({ "children": [], "children_names": [], "cubes": [], "name": name }) for c in json["children"]: cprops = jsonToProps(c["box"]) if cprops is not None: new_c = {} queue.append((c, new_c)) node["children"].append(new_c) node["cubes"].append(cprops) node["children_names"].append(c["label"]) return hier def getOBB(parts): part_corners = inter.sampleCorners(parts) bbox = inter.points_obb(part_corners.cpu(), 1) return bbox def cleanHier(hier): for i in range(len(hier['children'])): if len(hier['children'][i]) > 0: cleanHier(hier['children'][i]) if len(hier['children']) == 0: for key in list(hier.keys()): hier.pop(key) def trimHier(hier): for i in range(len(hier['children'])): if len(hier['children'][i]) > 0: if len(hier['children'][i]['children']) == 1: hier['cubes'][i] = hier['children'][i]['cubes'][0] hier['children_names'][i] = hier['children'][i]['children_names'][0] hier['children'][i] = hier['children'][i]['children'][0] if len(hier['children'][i]) > 0: trimHier(hier['children'][i]) def fillHier(hier): for i in range(len(hier['children'])): if len(hier['children'][i]) > 0: hier['cubes'][i] = fillHier(hier['children'][i]) hier['bbox'] = getOBB(hier['cubes']) return deepcopy(hier['bbox']) # centers and orients root bounding box # propogates transformation to all cuboids # also instanties bounding box into the cube + children spots def normalizeHier(hier): hier.pop('bbox') rbbox = {} samps = inter.sampleCorners(hier['cubes']).cpu() dims = samps.max(dim=0).values - samps.min(dim=0).values rbbox['xd'] = dims[0] rbbox['yd'] = dims[1] rbbox['zd'] = dims[2] rbbox['center'] = (samps.max(dim=0).values + samps.min(dim=0).values) / 2 rbbox['xdir'] = torch.tensor([1.,0.,0.]) rbbox['ydir'] = torch.tensor([0.,1.,0.]) rbbox['zdir'] = torch.tensor([0.,0.,1.]) hier['bbox'] = rbbox offset = rbbox['center'] q = [hier] while len(q) > 0: n = q.pop(0) bbox = n.pop('bbox') n['children'] = [{}] + n['children'] n['children_names'] = ["bbox"] + n['children_names'] n['cubes'] = [bbox] + n['cubes'] for i in range(len(n['cubes'])): n['cubes'][i]['center'] = n['cubes'][i]['center'] - offset for c in n['children']: if len(c) > 0: q.append(c) def markLeafCubes(hier): parts = [] q = [hier] while(len(q) > 0): n = q.pop(0) n['leaf_inds'] = [] assert(len(n['cubes']) == len(n['children'])) for cu, ch in zip(n['cubes'], n['children']): if len(ch) > 0: q.append(ch) n['leaf_inds'].append(-1) else: n['leaf_inds'].append(len(parts)) parts.append(cu) return parts def replace_parts(hier, parts, key): q = [hier] while(len(q) > 0): n = q.pop(0) binds = [] for i in range(len(n[key])): if n[key][i] != -1: lpart = parts[n[key][i]] if lpart is None: binds.append(i) else: n['cubes'][i] = lpart binds.sort(reverse=True) for bi in binds: n['children'].pop(bi) n['cubes'].pop(bi) n['children_names'].pop(bi) for c in n['children']: if c is not None and len(c) > 0: q.append(c) n.pop(key) # Takes in a hierarchy of just leaf cuboids, # finds new parameters for leaf cuboids so that # part-to-part connections are as valid as possible def shortenLeaves(hier): if VERBOSE: print("Doing Shortening") parts = markLeafCubes(hier) bad_inds = inter.findHiddenCubes(parts) ind_to_pc, scene_geom = inter.samplePC(parts) inters = inter.findInters(ind_to_pc, scene_geom) dim_parts = [ (p['xd'] * p['yd'] * p['zd'], i) for i,p in enumerate(parts) ] dim_parts.sort() for _, ind in dim_parts: if ind in bad_inds: continue if VERBOSE: print(f"Shortening Leaf ind: {ind}") sres = inter.shorten_cube(inters, parts, ind, scene_geom) if sres is not None: t_ind_to_pc, t_scene_geom = inter.samplePC([parts[ind]]) ind_to_pc[ind] = t_ind_to_pc[0] scene_geom[ind] = t_scene_geom[0] sres = [(int(s.split('_')[0]), int(s.split('_')[1])) for s in sres] new_inters = inter.findInters(ind_to_pc, scene_geom, sres) inters.update(new_inters) if parts[ind]['xd'] < SD_THRESH or \ parts[ind]['yd'] < SD_THRESH or \ parts[ind]['zd'] < SD_THRESH: bad_inds.append(ind) for bi in bad_inds: parts[bi] = None replace_parts(hier, parts, 'leaf_inds') def make_conn_graph(num_nodes, attachments): edges = [] for (ind1, ind2, _, _) in attachments: edges.append((ind1, ind2)) G = nx.Graph() G.add_nodes_from(list(range(num_nodes))) G.add_edges_from(edges) return G def assertConnected(num_nodes, attachments): G = make_conn_graph(num_nodes, attachments) assert nx.number_connected_components(G) == 1, 'disconnected graph' def checkConnected(node): assertConnected(len(node['cubes']), node['attachments']) for c in node['children']: if len(c) > 0: checkConnected(c) def memoize(f): def helper(ind, category): cdir = "parse_cache" cached_res = os.listdir(cdir) if ind in cached_res: return pickle.load(open(cdir+"/"+ind, "rb")) else: hier = f(ind, category) pickle.dump(hier, open(cdir+"/"+ind, "wb")) return hier return helper def checkCubeNum(node): assert len(node['cubes']) <= 11, f"Saw program with {len(node['cubes'])} cubes" for c in node['children']: if len(c) > 0: checkCubeNum(c) def flattenNode(node): for c in node['children']: if len(c) > 0: flattenNode(c) # Now everything has one sub-program fcubes = [] fchildren = [] fchildren_names = [] for i in range(len(node['children'])): if len(node['children'][i]) == 0: fcubes.append(node['cubes'][i]) fchildren.append(node['children'][i]) fchildren_names.append(node['children_names'][i]) else: fcubes += node['children'][i]['cubes'] fchildren += node['children'][i]['children'] fchildren_names += node['children'][i]['children_names'] node['cubes'] = fcubes node['children'] = fchildren node['children_names'] = fchildren_names def semFlattenHier(hier, category): flat_names = SFLAT_MAP[category] q = [hier] while (len(q) > 0): node = q.pop(0) if node['name'] in flat_names: flattenNode(node) else: for c in node['children']: if len(c) > 0: q.append(c) def semReHier(hier, category): if category not in SRH_MAP: return rh_tar, rh_names = SRH_MAP[category] if rh_tar not in hier['children_names']: return rhinds = [] for i,name in enumerate(hier['children_names']): if name in rh_names: rhinds.append(i) rhinds.sort(reverse=True) ti = hier['children_names'].index(rh_tar) for i in rhinds: for key in ['children_names', 'cubes', 'children']: hier['children'][ti][key].append(hier[key][i]) for i in rhinds: for key in ['children_names', 'cubes', 'children']: hier[key].pop(i) if len(hier['children']) == 1: hier['children_names'] = hier['children'][0]['children_names'] hier['cubes'] = hier['children'][0]['cubes'] hier['children'] = hier['children'][0]['children'] right = torch.tensor([1.0,0.5,0.5]) left = torch.tensor([0.0,0.5,0.5]) top = torch.tensor([0.5,1.0,0.5]) bot = torch.tensor([0.5,0.0,0.5]) front = torch.tensor([0.5,0.5,1.0]) back = torch.tensor([0.5,0.5,0.0]) def isFaceAtt(face, oface): if not DO_SQUEEZE: return None face = torch.tensor(face) if (face-right).norm() < .1 and abs(oface[0] - 0.0) < .1: return 'right' if (face-left).norm() < .1 and abs(oface[0] - 1.0) < .1: return 'left' if (face-top).norm() < .1 and abs(oface[1] - 0.0) < .1: return 'top' if (face-bot).norm() < .1 and abs(oface[1] - 1.0) < .1: return 'bot' if (face-front).norm() < .1 and abs(oface[2] - 0.0) < .1: return 'front' if (face-back).norm() < .1 and abs(oface[2] - 1.0) < .1: return 'back' return None def preProc(node): node['aligned'] = [isAxisAligned(cube, node['cubes'][0]) for cube in node['cubes']] node['cubes'] = [{f : cube[f].tolist() for f in cube } for cube in node['cubes']] new_attachments = [] for i0, i1, a0, a1 in node['attachments']: if i0 == 0: if a0[1] < .5: i0 = -2 else: i0 = -1 # FLIP BBOX TOP AND BOT FOR CONSISTENCY a0[1] = 1-a0[1] f0 = isFaceAtt(a0, a1) f1 = isFaceAtt(a1, a0) assert i1 != 0, 'uh oh' new_attachments.append(( i0, i1, a0, a1, f0, f1 )) node['attachments'] = new_attachments def parseJsonToHier(ind, category, get_gt=False): sem_order = getSemOrder(category) hier = getShapeHier(ind, category) assert len(hier) > 0, 'saw empty hier' if DO_SEM_FLATTEN: semFlattenHier(hier, category) if DO_SEM_REHIER: semReHier(hier, category) if DO_SHORTEN: shortenLeaves(hier) cleanHier(hier) trimHier(hier) fillHier(hier) normalizeHier(hier) addAttachments(hier, sem_order) if get_gt: return hier if DO_VALID_CHECK: checkConnected(hier) if DO_SIMP_SYMMETRIES: sym.addSymSubPrograms(hier) sym.addSimpSymmetries(hier) if DO_VALID_CHECK: checkCubeNum(hier) return hier def memoize(f): def helper(ind, category, part): cdir = "parse_part_cache" cached_res = os.listdir(cdir) if f"{ind}_{part}" in cached_res: return pickle.load(open(f"{cdir}/{ind}_{part}", "rb")) else: try: hier = f(ind, category, part) except Exception as e: print(f"Failed {ind} with {e}") hier = None pickle.dump(hier, open(f"{cdir}/{ind}_{part}", "wb")) return hier return helper def genAllData(outdir, ind, category): hier = parseJsonToHier(ind, category) q = [hier] seen = set() while(len(q)>0): node = q.pop(0) for i, c in enumerate(node['children']): while node['children_names'][i] in seen and i > 0: node['children_names'][i] += '+' if len(c) > 0: c['name'] = node['children_names'][i] q.append(c) node.pop('children') preProc(node) part = node['name'] pickle.dump(node, open(f"{outdir}/{ind}_{part}", "wb")) def loadAllData(in_dir, max_files=int(1e8)): ninds = [] nodes = [] files = list(os.listdir(in_dir)) files.sort() for f in files[:max_files]: node = pickle.load(open(f"{in_dir}/{f}", "rb")) if node is not None: ninds.append(f) nodes.append(node) return ninds, nodes if __name__ == '__main__': category = sys.argv[1] outdir = sys.argv[2] PATH_TO_SA_DATA = sys.argv[3] inds = os.listdir(PATH_TO_SA_DATA) inds = [i.split('.')[0] for i in inds] #inds = ['173', '2307', '44366', '1282'] #inds += ['42231', '40507', '36402', '41162', '41830'] os.system(f'mkdir {outdir}') from tqdm import tqdm for ind in tqdm(inds): try: hier = genAllData(outdir, str(ind), category) except Exception as e: print(e)
20,758
25.95974
100
py
ShapeMOD
ShapeMOD-main/SA_lang/tasks/infer_recon_metrics.py
from ShapeAssembly import hier_execute import sa_utils as utils import torch import os import sys import math import faiss import numpy as np from valid import check_stability, check_rooted device = torch.device("cuda") class SimpChamferLoss(torch.nn.Module): def __init__(self, device): super(SimpChamferLoss, self).__init__() self.dimension = 3 self.gpu_id = torch.cuda.current_device() self.res = faiss.StandardGpuResources() def build_nn_index(self, database): """ :param database: numpy array of Nx3 :return: Faiss index, in CPU """ index_cpu = faiss.IndexFlatL2(self.dimension) index = faiss.index_cpu_to_gpu(self.res, self.gpu_id, index_cpu) index.add(database) return index def search_nn(self, index, query, k): D, I = index.search(query, 1) return np.sqrt(D) def getAvgDist(self, index, query): D, I = index.search(query, 2) m_d = math.sqrt(np.percentile(D[:,1],90)) return m_d def calc_metrics(self, predict_pc, gt_pc, threshes): """ :param predict_pc: Bx3xM Variable in GPU :param gt_pc: Bx3xN Variable in GPU :return: """ predict_pc_size = predict_pc.size() gt_pc_size = gt_pc.size() predict_pc_np = np.ascontiguousarray( torch.transpose(predict_pc.data.clone(), 1, 2).cpu().numpy() ) # BxMx3 gt_pc_np = np.ascontiguousarray( torch.transpose(gt_pc.data.clone(), 1, 2).cpu().numpy() ) # BxNx3 index_predict = self.build_nn_index(predict_pc_np[0]) index_gt = self.build_nn_index(gt_pc_np[0]) fwd_dist = self.search_nn(index_gt, predict_pc_np[0], 1) bwd_dist = self.search_nn(index_predict, gt_pc_np[0], 1) cd = (fwd_dist.mean() / 2) + (bwd_dist.mean() / 2) ones = np.ones(fwd_dist.shape) fscores = [] for thresh in threshes: if thresh == 'def': thresh = self.getAvgDist(index_gt, gt_pc_np[0]) precision = (100 / ones.shape[0]) * np.sum(ones[fwd_dist <= thresh]) recall = (100 / ones.shape[0]) * np.sum(ones[bwd_dist <= thresh]) fs = (2*precision*recall) / (precision + recall + 1e-8) fscores.append(fs) return [cd] + fscores class SimpCPUChamferLoss(torch.nn.Module): def __init__(self): super(SimpCPUChamferLoss, self).__init__() self.dimension = 3 def build_nn_index(self, database): """ :param database: numpy array of Nx3 :return: Faiss index, in CPU """ index = faiss.IndexFlatL2(self.dimension) index.add(database) return index def search_nn(self, index, query, k): D, I = index.search(query, 1) return np.sqrt(D) def getAvgDist(self, index, query): D, I = index.search(query, 2) m_d = math.sqrt(np.percentile(D[:,1],90)) return m_d def calc_metrics(self, predict_pc, gt_pc, threshes): """ :param predict_pc: Bx3xM Variable in GPU :param gt_pc: Bx3xN Variable in GPU :return: """ predict_pc_size = predict_pc.size() gt_pc_size = gt_pc.size() predict_pc_np = np.ascontiguousarray( torch.transpose(predict_pc.data.clone(), 1, 2).cpu().numpy() ) # BxMx3 gt_pc_np = np.ascontiguousarray( torch.transpose(gt_pc.data.clone(), 1, 2).cpu().numpy() ) # BxNx3 index_predict = self.build_nn_index(predict_pc_np[0]) index_gt = self.build_nn_index(gt_pc_np[0]) fwd_dist = self.search_nn(index_gt, predict_pc_np[0], 1) bwd_dist = self.search_nn(index_predict, gt_pc_np[0], 1) cd = (fwd_dist.mean() / 2) + (bwd_dist.mean() / 2) ones = np.ones(fwd_dist.shape) fscores = [] for thresh in threshes: if thresh == 'def': thresh = self.getAvgDist(index_gt, gt_pc_np[0]) precision = (100 / ones.shape[0]) * np.sum(ones[fwd_dist <= thresh]) recall = (100 / ones.shape[0]) * np.sum(ones[bwd_dist <= thresh]) fs = (2*precision*recall) / (precision + recall + 1e-8) fscores.append(fs) return [cd] + fscores chamfer = SimpCPUChamferLoss() def getSampMetrics(verts, faces, t_samps): p_samps = torch.clamp( utils.sample_surface(faces, verts.unsqueeze(0), t_samps.shape[0], False)[0], -1, 1) samp_metrics = chamfer.calc_metrics( p_samps.T.unsqueeze(0).float(), t_samps.T.unsqueeze(0).float(), [.05, .03, .01, 'def'] ) return { 'cd': samp_metrics[0], 'fscore-05': samp_metrics[1], 'fscore-03': samp_metrics[2], 'fscore-01': samp_metrics[3], 'fscore-def': samp_metrics[4], } def getShapeIoU(cubes, gt_cubes): pvoxels = shape_voxelize(cubes) tvoxels = shape_voxelize(gt_cubes) iou = 100 * ( (pvoxels & tvoxels).sum().item() / (pvoxels | tvoxels).sum().item() ) return iou def recon_metrics( recon_sets, outpath, exp_name, name, epoch, VERBOSE, num_gen ): misses = 0. results = { 'iou': [], 'cd': [], 'fscore-def': [], 'fscore-01': [], 'fscore-03': [], 'fscore-05': [], 'rooted': [], 'stable': [] } count = 0 for prog, gt_prog, prog_ind, gt_pts in recon_sets: gt_verts, gt_faces, gt_cubes = hier_execute(gt_prog, return_all = True) try: verts, faces, cubes = hier_execute(prog, return_all = True) assert not torch.isnan(verts).any(), 'saw nan vert' try: if check_rooted(verts, faces): results['rooted'].append(1.) else: results['rooted'].append(0.) if check_stability(verts, faces): results['stable'].append(1.) else: results['stable'].append(0.) except Exception as e: print(f"Failed rooted/stable with {e} ???") except Exception as e: misses += 1. if VERBOSE: print(f"failed recon metrics for {prog_ind} with {e}") continue gt_objs = os.listdir(f"{outpath}/{exp_name}/objs/gt/") try: sm = getSampMetrics(verts, faces, gt_pts) for k, v in sm.items(): if v is not None: results[k].append(v) except Exception as e: if VERBOSE: print(f"failed Samp Metrics for {prog_ind} with {e}") if count >= num_gen: continue if f"{prog_ind}.obj" not in gt_objs: utils.writeObj(gt_verts, gt_faces, f"{outpath}/{exp_name}/objs/gt/{prog_ind}.obj") if 'dsl_prog' in gt_prog: utils.writeHierProg(gt_prog, 'dsl_prog', f"{outpath}/{exp_name}/programs/gt/{prog_ind}.txt") else: utils.sawriteHierProg(gt_prog, f"{outpath}/{exp_name}/programs/gt/{prog_ind}.txt") try: utils.writeObj( verts, faces, f"{outpath}/{exp_name}/objs/{name}/{epoch}_{prog_ind}.obj" ) if 'dsl_prog' in prog: utils.writeHierProg( prog, 'dsl_prog', f"{outpath}/{exp_name}/programs/{name}/{epoch}_{prog_ind}.txt" ) else: utils.sawriteHierProg( prog, f"{outpath}/{exp_name}/programs/{name}/{epoch}_{prog_ind}.txt" ) count += 1 except Exception as e: if VERBOSE: print(f"Failed writing prog/obj for {prog_ind} with {e}") for key in results: if len(results[key]) > 0: res = torch.tensor(results[key]).mean().item() else: res = 0. results[key] = res return results, misses
8,389
30.660377
108
py
ShapeMOD
ShapeMOD-main/SA_lang/tasks/losses.py
import torch import torch.nn as nn import faiss import numpy as np import math import generate as gen import execute as ex from copy import deepcopy import json_parse as jp def robust_norm(var, dim=2): return ((var ** 2).sum(dim=dim) + 1e-8).sqrt() class FScore(): def __init__(self): self.dimension = 3 self.k = 1 self.forward_loss = torch.FloatTensor([0]) self.backward_loss = torch.FloatTensor([0]) def build_nn_index(self, database): """ :param database: numpy array of Nx3 :return: Faiss index, in CPU """ index_cpu = faiss.IndexFlatL2(self.dimension) index = index_cpu index.add(database) return index def search_nn(self, index, query, k): """ :param index: Faiss index :param query: numpy array of Nx3 :return: D: Variable of Nxk, type FloatTensor, in GPU I: Variable of Nxk, type LongTensor, in GPU """ D, I = index.search(query, k) D_var = torch.from_numpy(np.ascontiguousarray(D)) I_var = torch.from_numpy(np.ascontiguousarray(I).astype(np.int64)) return D_var, I_var def getAvgDist(self, index, query): D, I = index.search(query, 2) m_d = math.sqrt(np.percentile(D[:,1],90)) return m_d def getOpMatch(self, points): return (points.max(axis = 0) - points.min(axis = 0)).max() / 100 def score(self, predict_pc_6, gt_pc_6, use_normals=True): """ :param predict_pc: Bx3xM Variable in GPU :param gt_pc: Bx3xN Variable in GPU :return: """ predict_pc = predict_pc_6[:, :3, :].cpu() gt_pc = gt_pc_6[:, :3, :].cpu() # predict_pcn = predict_pc_6[:, 3:, :].cpu() gt_pcn = gt_pc_6[:, 3:, :].cpu() predict_pc_size = predict_pc.size() gt_pc_size = gt_pc.size() predict_pc_np = np.ascontiguousarray( torch.transpose(predict_pc.data.clone(), 1, 2).cpu().numpy() ) # BxMx3 gt_pc_np = np.ascontiguousarray( torch.transpose(gt_pc.data.clone(), 1, 2).cpu().numpy() ) # BxNx3 # selected_gt: Bxkx3xM selected_gt_by_predict = torch.FloatTensor( predict_pc_size[0], self.k, predict_pc_size[1], predict_pc_size[2] ) # selected_predict: Bxkx3xN selected_predict_by_gt = torch.FloatTensor( gt_pc_size[0], self.k, gt_pc_size[1], gt_pc_size[2] ) if use_normals: # normals selected_gt_by_predictn = torch.FloatTensor( predict_pc_size[0], self.k, predict_pc_size[1], predict_pc_size[2] ) selected_predict_by_gtn = torch.FloatTensor( gt_pc_size[0], self.k, gt_pc_size[1], gt_pc_size[2] ) # process each batch independently. for i in range(predict_pc_np.shape[0]): index_predict = self.build_nn_index(predict_pc_np[i]) index_gt = self.build_nn_index(gt_pc_np[i]) # database is gt_pc, predict_pc -> gt_pc ----------------------------------------------------------- _, I_var = self.search_nn(index_gt, predict_pc_np[i], self.k) # process nearest k neighbors for k in range(self.k): selected_gt_by_predict[i, k, ...] = gt_pc[i].index_select( 1, I_var[:, k] ) if use_normals: selected_gt_by_predictn[i, k, ...] = gt_pcn[i].index_select( 1, I_var[:, k] ) # database is predict_pc, gt_pc -> predict_pc ------------------------------------------------------- _, I_var = self.search_nn(index_predict, gt_pc_np[i], self.k) # process nearest k neighbors for k in range(self.k): selected_predict_by_gt[i, k, ...] = predict_pc[i].index_select( 1, I_var[:, k] ) if use_normals: selected_predict_by_gtn[i, k, ...] = predict_pcn[i].index_select( 1, I_var[:, k] ) index_gt = self.build_nn_index(gt_pc_np[i]) dist = self.getAvgDist(index_gt, gt_pc_np[0]) # compute loss =================================================== # selected_gt(Bxkx3xM) vs predict_pc(Bx3xM) r_to_gt = predict_pc.unsqueeze(1).expand_as(selected_gt_by_predict) - selected_gt_by_predict gt_to_r = selected_predict_by_gt - gt_pc.unsqueeze(1).expand_as(selected_predict_by_gt) r_to_gt = robust_norm( selected_gt_by_predict - predict_pc.unsqueeze(1).expand_as(selected_gt_by_predict), dim=2 ).cpu().detach().numpy() gt_to_r = robust_norm( selected_predict_by_gt - gt_pc.unsqueeze(1).expand_as(selected_predict_by_gt), dim=2 ).cpu().detach().numpy() r_to_gt = r_to_gt.flatten() gt_to_r = gt_to_r.flatten() ones = np.ones(r_to_gt.shape) dt = dist precision = (100 / ones.shape[0]) * np.sum(ones[r_to_gt < dt]) recall = (100 / ones.shape[0]) * np.sum(ones[gt_to_r < dt]) return (2*precision*recall) / (precision + recall + 1e-8) class ChamferLoss(nn.Module): def __init__(self, device): super(ChamferLoss, self).__init__() self.dimension = 3 self.k = 1 # self.device = device self.gpu_id = torch.cuda.current_device() self.faiss_gpu = hasattr(faiss, 'StandardGpuResources') if self.faiss_gpu: self.res = faiss.StandardGpuResources() self.flat_config = faiss.GpuIndexFlatConfig() self.flat_config.device = self.gpu_id def build_nn_index(self, database): """ :param database: numpy array of Nx3 :return: Faiss index, in CPU """ # index = faiss.GpuIndexFlatL2(self.res, self.dimension, self.flat_config) # dimension is 3 index_cpu = faiss.IndexFlatL2(self.dimension) if self.faiss_gpu: index = faiss.index_cpu_to_gpu(self.res, self.gpu_id, index_cpu) else: index = index_cpu index.add(database) return index def search_nn(self, index, query, k): """ :param index: Faiss index :param query: numpy array of Nx3 :return: D: Variable of Nxk, type FloatTensor, in GPU I: Variable of Nxk, type LongTensor, in GPU """ D, I = index.search(query, k) D_var = torch.from_numpy(np.ascontiguousarray(D)) I_var = torch.from_numpy(np.ascontiguousarray(I).astype(np.int64)) if self.gpu_id >= 0: D_var = D_var.to(self.device) I_var = I_var.to(self.device) return D_var, I_var def forward(self, predict_pc_6, gt_pc_6, thresh, keep_dim=False, use_normals=True): """ :param predict_pc: Bx3xM Variable in GPU :param gt_pc: Bx3xN Variable in GPU :return: """ predict_pc = predict_pc_6[:, :3, :] gt_pc = gt_pc_6[:, :3, :] # if use_normals: predict_pcn = predict_pc_6[:, 3:, :] gt_pcn = gt_pc_6[:, 3:, :] predict_pc_size = predict_pc.size() gt_pc_size = gt_pc.size() predict_pc_np = np.ascontiguousarray( torch.transpose(predict_pc.data.clone(), 1, 2).cpu().numpy() ) # BxMx3 gt_pc_np = np.ascontiguousarray( torch.transpose(gt_pc.data.clone(), 1, 2).cpu().numpy() ) # BxNx3 # selected_gt: Bxkx3xM selected_gt_by_predict = torch.FloatTensor( predict_pc_size[0], self.k, predict_pc_size[1], predict_pc_size[2] ) # selected_predict: Bxkx3xN selected_predict_by_gt = torch.FloatTensor( gt_pc_size[0], self.k, gt_pc_size[1], gt_pc_size[2] ) if use_normals: # normals selected_gt_by_predictn = torch.FloatTensor( predict_pc_size[0], self.k, predict_pc_size[1], predict_pc_size[2] ) selected_predict_by_gtn = torch.FloatTensor( gt_pc_size[0], self.k, gt_pc_size[1], gt_pc_size[2] ) if self.gpu_id >= 0: selected_gt_by_predict = selected_gt_by_predict.to(self.device) selected_predict_by_gt = selected_predict_by_gt.to(self.device) if use_normals: selected_gt_by_predictn = selected_gt_by_predictn.to(self.device) selected_predict_by_gtn = selected_predict_by_gtn.to(self.device) # process each batch independently. for i in range(predict_pc_np.shape[0]): index_predict = self.build_nn_index(predict_pc_np[i]) index_gt = self.build_nn_index(gt_pc_np[i]) # database is gt_pc, predict_pc -> gt_pc ----------------------------------------------------------- _, I_var = self.search_nn(index_gt, predict_pc_np[i], self.k) # process nearest k neighbors for k in range(self.k): selected_gt_by_predict[i, k, ...] = gt_pc[i].index_select( 1, I_var[:, k] ) if use_normals: selected_gt_by_predictn[i, k, ...] = gt_pcn[i].index_select( 1, I_var[:, k] ) # database is predict_pc, gt_pc -> predict_pc ------------------------------------------------------- _, I_var = self.search_nn(index_predict, gt_pc_np[i], self.k) # process nearest k neighbors for k in range(self.k): selected_predict_by_gt[i, k, ...] = predict_pc[i].index_select( 1, I_var[:, k] ) if use_normals: selected_predict_by_gtn[i, k, ...] = predict_pcn[i].index_select( 1, I_var[:, k] ) # compute loss =================================================== # selected_gt(Bxkx3xM) vs predict_pc(Bx3xM) forward_loss_element = robust_norm( selected_gt_by_predict - predict_pc.unsqueeze(1).expand_as(selected_gt_by_predict) ) self.forward_loss = forward_loss_element.mean() self.forward_loss_array = forward_loss_element.mean(dim=1).mean(dim=1) # selected_predict(Bxkx3xN) vs gt_pc(Bx3xN) backward_loss_element = robust_norm( selected_predict_by_gt - gt_pc.unsqueeze(1).expand_as(selected_predict_by_gt) ) # BxkxN self.backward_loss = backward_loss_element.mean() self.backward_loss_array = backward_loss_element.mean(dim=1).mean(dim=1) # if keep_dim: self.forward_loss = torch.relu(forward_loss_element - thresh).squeeze(1).mean(dim=1) self.backward_loss = torch.relu(backward_loss_element - thresh).squeeze(1).mean(dim=1) return self.forward_loss + self.backward_loss else: self.forward_loss = torch.relu(forward_loss_element.flatten() - thresh).mean() self.backward_loss = torch.relu(backward_loss_element.flatten() - thresh).mean() return self.forward_loss + self.backward_loss def __call__(self, predict_pc, gt_pc, ang_wt, keep_dim=False): # start_time = time.time() loss = self.forward(predict_pc, gt_pc, ang_wt, keep_dim) # print(time.time()w-start_time) return loss def weighted_mae_loss(input, target, weight): return torch.sum(weight * (input - target).abs()) closs = torch.nn.BCEWithLogitsLoss(reduction='sum', pos_weight = torch.tensor(0.2)) celoss = nn.CrossEntropyLoss(reduction='sum') # 0 -> start # 1 -> Cuboid # 2 -> attach # 3 -> reflect # 4 -> translate # 5 -> squeeze # 6 -> end # 0 - 7 -> commandIndex # 7 -> 18 -> cuboid1 # 18 -> 29 -> cuboid2 # 29 -> 40 -> cuboid3 # 40 -> 43 -> l,h,w # 43 -> 46 -> x1,y1,z1 # 46 -> 49 -> x2,y2,z2 # 49 - 52 -> symaxis # 52 -> num # 53 -> scale # 54 - 60 face # 60 - 62 - UV # 62 -> aligned # BBOX DIMS 40 42 class ProgLoss(nn.Module): def forward(self, pred, target, weight): if target.shape[1] > pred.shape[1]: target = target[:,:pred.shape[1],:] weight = weight[:,:pred.shape[1],:] commands = torch.argmax(target[:,:,:7], dim = 2).flatten() pcommands = torch.argmax(pred[:,:,:7], dim = 2).flatten() cmdc = (commands == pcommands).sum().item() * 1.0 cmd_loss = celoss( pred[:,:,:7].view(-1,7), commands ) cub_inds = (commands==1).nonzero().flatten() ap_inds = (commands==2).nonzero().flatten() ref_inds = (commands==3).nonzero().flatten() trans_inds = (commands==4).nonzero().flatten() sq_inds = (commands==5).nonzero().flatten() cub_prm_loss = 0. align_loss = 0. align_num_pos = 0. align_num_neg = 0. align_posc = 0. align_negc = 0. if cub_inds.sum() > 0: cub_prm_loss = weighted_mae_loss( pred[:,cub_inds,:], target[:,cub_inds,:], weight[:,cub_inds,:] ) if cub_inds.sum() > 1: align_pos_inds = cub_inds[1:][(target[:, cub_inds[1:], 62].flatten() == 1.).nonzero().flatten()] align_neg_inds = cub_inds[1:][(target[:, cub_inds[1:], 62].flatten() == 0.).nonzero().flatten()] align_num_pos += align_pos_inds.shape[0] align_num_neg += align_neg_inds.shape[0] align_loss = closs( pred[:, cub_inds[1:], 62].flatten(), target[:, cub_inds[1:], 62].flatten() ) align_posc = (pred[:, align_pos_inds, 62] > 0.).sum().float().item() align_negc = (pred[:, align_neg_inds, 62] <= 0.).sum().float().item() xyz_prm_loss = 0 cubc = 0 ap_cub_loss = 0 if ap_inds.sum() > 0: xyz_prm_loss = weighted_mae_loss( pred[:,ap_inds,:], target[:,ap_inds,:], weight[:,ap_inds,:] ) ap_cube_1s = torch.argmax(target[:,ap_inds,7:18], dim=2).flatten() ap_cube_2s = torch.argmax(target[:,ap_inds,18:29], dim=2).flatten() ap_pcube_1s = torch.argmax(pred[:,ap_inds,7:18], dim=2).flatten() ap_pcube_2s = torch.argmax(pred[:,ap_inds,18:29], dim=2).flatten() ap_cubc = ((ap_pcube_1s == ap_cube_1s) * (ap_pcube_2s == ap_cube_2s)).sum().item() * 1.0 cubc = ap_cubc ap_cub_loss = celoss( pred[:,ap_inds,7:18].view(-1, 11), ap_cube_1s ) ap_cub_loss += celoss( pred[:,ap_inds,18:29].view(-1, 11), ap_cube_2s ) cub_loss = ap_cub_loss axisc = 0. axis_loss = 0. sym_cubc = 0. sym_cub_loss = 0. sym_prm_loss = 0. if ref_inds.sum() > 0: ref_cube_1s = torch.argmax(target[:,ref_inds,7:18], dim=2).flatten() ref_pcube_1s = torch.argmax(pred[:,ref_inds,7:18], dim=2).flatten() ref_cubc = ((ref_pcube_1s == ref_cube_1s)).sum().item() * 1.0 ref_axis = torch.argmax(target[:, ref_inds, 49:52], dim=2).flatten() ref_paxis = torch.argmax(pred[:, ref_inds, 49:52], dim=2).flatten() ref_axisc = ((ref_paxis == ref_axis)).sum().item() * 1.0 ref_axis_loss = celoss( pred[:,ref_inds,49:52].view(-1, 3), ref_axis ) ref_cub_loss = celoss( pred[:,ref_inds,7:18].view(-1, 11), ref_cube_1s ) axisc += ref_axisc axis_loss += ref_axis_loss sym_cubc += ref_cubc sym_cub_loss += ref_cub_loss if trans_inds.sum() > 0: trans_cube_1s = torch.argmax(target[:,trans_inds,7:18], dim=2).flatten() trans_pcube_1s = torch.argmax(pred[:,trans_inds,7:18], dim=2).flatten() trans_cubc = ((trans_pcube_1s == trans_cube_1s)).sum().item() * 1.0 trans_axis = torch.argmax(target[:, trans_inds, 49:52], dim=2).flatten() trans_paxis = torch.argmax(pred[:, trans_inds, 49:52], dim=2).flatten() trans_axisc = ((trans_paxis == trans_axis)).sum().item() * 1.0 trans_axis_loss = celoss( pred[:,trans_inds,49:52].view(-1, 3), trans_axis ) trans_cub_loss = celoss( pred[:,trans_inds,7:18].view(-1, 11), trans_cube_1s ) trans_prm_loss = weighted_mae_loss( pred[:,trans_inds,:], target[:,trans_inds,:], weight[:,trans_inds,:] ) axisc += trans_axisc axis_loss += trans_axis_loss sym_cubc += trans_cubc sym_cub_loss += trans_cub_loss sym_prm_loss += trans_prm_loss uv_prm_loss = 0 sq_cubc = 0 sq_cub_loss = 0 face_loss = 0 facec = 0 if sq_inds.sum() > 0: uv_prm_loss = weighted_mae_loss( pred[:,sq_inds,:], target[:,sq_inds,:], weight[:,sq_inds,:] ) sq_cube_1s = torch.argmax(target[:,sq_inds,7:18], dim=2).flatten() sq_cube_2s = torch.argmax(target[:,sq_inds,18:29], dim=2).flatten() sq_cube_3s = torch.argmax(target[:,sq_inds,29:40], dim=2).flatten() faces = torch.argmax(target[:, sq_inds, 54:60], dim=2).flatten() sq_pcube_1s = torch.argmax(pred[:,sq_inds,7:18], dim=2).flatten() sq_pcube_2s = torch.argmax(pred[:,sq_inds,18:29], dim=2).flatten() sq_pcube_3s = torch.argmax(pred[:,sq_inds,29:40], dim=2).flatten() pfaces = torch.argmax(pred[:,sq_inds,54:60], dim=2).flatten() sq_cubc = ( (sq_pcube_1s == sq_cube_1s) * \ (sq_pcube_2s == sq_cube_2s) * \ (sq_pcube_3s == sq_cube_3s) ).sum().item() * 1.0 facec = (faces == pfaces).sum().item() * 1.0 sq_cub_loss = celoss( pred[:,sq_inds,7:18].view(-1, 11), sq_cube_1s ) sq_cub_loss += celoss( pred[:,sq_inds,18:29].view(-1, 11), sq_cube_2s ) sq_cub_loss += celoss( pred[:,sq_inds,29:40].view(-1, 11), sq_cube_3s ) face_loss = celoss( pred[:,sq_inds,54:60].view(-1, 6), faces ) losses = { 'cmd': cmd_loss, 'cub_prm': cub_prm_loss, 'xyz_prm': xyz_prm_loss, 'uv_prm': uv_prm_loss, 'sym_prm': sym_prm_loss,# 'cub': cub_loss, 'sq_cub': sq_cub_loss, 'sym_cub': sym_cub_loss,# 'axis': axis_loss,# 'face': face_loss, 'cmdc': cmdc, 'cubc': cubc, 'sq_cubc': sq_cubc, 'sym_cubc': sym_cubc,# 'axisc': axisc,# 'facec': facec, 'align': align_loss, 'palignc': align_posc, 'nalignc': align_negc, 'nan': align_num_neg, 'nap': align_num_pos, 'na': ap_inds.shape[0] * 1.0, 'nc': cub_inds.shape[0] * 1.0, 'ns': (ref_inds.shape[0] * 1.0) + (trans_inds.shape[0] * 1.0), 'nsq': sq_inds.shape[0] * 1.0 } return losses
20,155
33.811744
113
py
ShapeMOD
ShapeMOD-main/SA_lang/tasks/make_abs_data.py
import sys sys.path.append("./dsls") sys.path.append("../") sys.path.append("../../") from tqdm import tqdm import torch from ShapeMOD import DSL, Function, ProgNode, OrderedProg import pickle import re import sa_utils as utils import importlib def clamp(v, a, b): return min(max(v, a), b) def make_function(name, args): args = [str(arg) for arg in args] return '{}({})'.format(name, ", ".join(args)) def getCuboidDims(F, f_params, j_param_map, bbox_dims): params = [1.0] + bbox_dims.tolist() f_params = f_params.tolist() for var in j_param_map: if var[0] == 'd': params.append('dummy') else: params.append(f_params.pop(0)) cube_dims = [] for fn, logic in zip(F.structure, F.getLogic()): if fn != "Cuboid": continue cube_line = [] for var in logic[:3]: assert var[0] == 'f' sum = 0. for ind, scale in var[1]: sum += float(scale) * params[ind] cube_line.append(sum) cube_dims.append(cube_line) return cube_dims def tensorize(prog, func_map, max_cparams, max_dparams): if prog is None: return None nmap = { 'bbox': 0, 'X': 0, 'Y': 1, 'Z': 2, 'left': 0, 'right': 1, 'bot': 2, 'top': 3, 'back': 4, 'front': 5 } cn_map = {0:0} t_lines = [('START',)] count = 1 child_map = [[]] for ln, (ret, line) in enumerate(prog): child_inds = [] for n in ret: if 'cube' in n: nmap[f'temp{int(n[4:])}'] = count cn_map[int(n[4:])+1] = count child_inds.append(count) count += 1 elif 'bbox' in n: child_inds.append(0) t_lines.append([l.replace('cube','temp') if isinstance(l, str) else l for l in line]) child_map.append(child_inds) t_lines.append(('STOP',)) child_map.append([]) prog_funcs = [] prog_cvars = [] prog_dvars = [] for t_line in t_lines: prog_funcs.append(func_map[t_line[0]]) line_cvars = torch.zeros(max_cparams).float() line_dvars = torch.zeros(max_dparams).long() dv_count = 0 cv_count = 0 for par in t_line[1:]: if par in nmap: line_dvars[dv_count] = nmap[par] dv_count += 1 else: line_cvars[cv_count] = float(par) cv_count += 1 prog_cvars.append(line_cvars) prog_dvars.append(line_dvars) prog_funcs = torch.tensor(prog_funcs) prog_cvars = torch.stack(prog_cvars) prog_dvars = torch.stack(prog_dvars) return (prog_funcs, prog_cvars, prog_dvars, cn_map, child_map) def getCatTypeMap(dsl): cat_type_map = {} for n, f in dsl.library.items(): _cat_type_map = {} c_count = 0 seen_inds = set() for st, log in zip(f.getStructure(), f.getLogic()): for l in log: if l[0] == 'c' and l[1] == 'var' and l[2] not in seen_inds: seen_inds.add(l[2]) name = f'c_var_{c_count}' c_count += 1 if st == 'squeeze': _cat_type_map[name] = 'sq' elif st == 'translate' or st == 'reflect': _cat_type_map[name] = 'sym' cat_type_map[n] = _cat_type_map return cat_type_map def getBestProgram(dsl, node, order_thresh = 1): def scoreOrder(ord): sum = 0. place = 1. for o in ord: sum += o[1] * place place *= .1 return sum best_score = 1e8 best_program = None res = [] if len(node.orders) == 0: return None for o in node.orders[:dsl.abs_order_num]: canon_sig, line_attrs, ret_attrs = o.canon_info score, program, _ = dsl.getApproxBestOps(line_attrs, ret_attrs) res.append((score, canon_sig, program)) res.sort() best_score = res[0][0] order_res = [ (scoreOrder(r[1]), r[2]) for r in res if r[0] < (best_score + order_thresh) ] order_res.sort() best_program = order_res[0][1] return best_program def getBestPrograms(dsl, nodes): best_progs = [] for node in tqdm(nodes): bp = getBestProgram(dsl, node) best_progs.append(bp) return best_progs def form_training_data(dsl, nodes): max_cparams = 0 # start with attach max max_dparams = 0 # start with squeeze max func_map = {'START': 0, 'STOP': 1} cparam_map = {0: [], 1: []} dparam_map = {0: [], 1: []} jparam_map = {0: [], 1: []} num_cube_map = {0: 0, 1: 0} cat_type_map = getCatTypeMap(dsl) for n, f in dsl.library.items(): func_map[n] = len(func_map) interf = f.getInterface() structure = f.getStructure() cparam_map[func_map[n]] = [] dparam_map[func_map[n]] = [] jparam_map[func_map[n]] = [] for i in interf: _i = i.split('_')[0] if _i == 'i': jparam_map[func_map[n]].append(f'd_{len(dparam_map[func_map[n]])}') dparam_map[func_map[n]].append(_i) elif _i == 'c': jparam_map[func_map[n]].append(f'd_{len(dparam_map[func_map[n]])}') dparam_map[func_map[n]].append(cat_type_map[n][i]) else: jparam_map[func_map[n]].append(f'c_{len(cparam_map[func_map[n]])}') cparam_map[func_map[n]].append(_i) num_cube_map[func_map[n]] = structure.count("Cuboid") max_cparams = max(len(cparam_map[func_map[n]]), max_cparams) max_dparams = max(len(dparam_map[func_map[n]]), max_dparams) best_progs = getBestPrograms(dsl, nodes) best_progs = [tensorize(bp, func_map, max_cparams, max_dparams) for bp in best_progs] return best_progs, func_map, cparam_map, dparam_map, jparam_map, max_cparams, max_dparams, num_cube_map def writeData(dsl, out_name, input_data, category): full_nodes = pickle.load(open(input_data, 'rb')) simp_inds = list(set([n.ind.split('_')[0] for n in full_nodes])) inds = [] nodes = [] for node in full_nodes: ind = node.ind if ind.split('_')[0] in simp_inds: inds.append(ind) nodes.append(node) node_tensors, func_map, cparam_map, dparam_map, jparam_map, \ max_cparams, max_dparams, num_cube_map = form_training_data(dsl, nodes) training_data = [] metadata = { 'func_map': func_map, 'cparam_map': cparam_map, 'dparam_map': dparam_map, 'jparam_map': jparam_map, 'max_cparams': max_cparams, 'max_dparams': max_dparams, 'num_cube_map': num_cube_map, 'dsl': dsl, } metadata['rev_func_map'] = {v:k for k, v in metadata['func_map'].items()} for key in ('i', 'sq', 'sym'): metadata[f'max_d_{key}_params'] = max([ len([ __l for __l in _l if __l == key]) for _i, _l in metadata['dparam_map'].items() ]) for ind in simp_inds: good_ind = True if f'{ind}_{category}' not in inds: print("Failed a lookup on root") continue root = {'lookup': f'{ind}_{category}'} q = [root] while(len(q) > 0): node = q.pop(0) index = inds.index(node['lookup']) node['children_names'] = nodes[index].children_names if node_tensors[index] is None: good_ind = False q = [] break node['func_gt'] = node_tensors[index][0].numpy() node['cparam_gt'] = node_tensors[index][1].numpy() node['dparam_gt'] = node_tensors[index][2].numpy() node['child_gt'] = node_tensors[index][4] node['children'] = [] child_order = {v:k for k,v in node_tensors[index][3].items()} children_names = node.pop('children_names') for ci in range(len(children_names)): cn = children_names[child_order[ci]] c_lookup = f'{ind}_{cn}' if c_lookup in inds: c_node = {'lookup': c_lookup} q.append(c_node) node['children'].append(c_node) else: node['children'].append({}) if good_ind: training_data.append((ind, root)) else: print("Failed") pickle.dump(training_data, open(f'{out_name}_train.data', 'wb')) pickle.dump(metadata, open(f'{out_name}_train.meta', 'wb')) def indToFace(i): m = ['left', 'right', 'bot', 'top', 'back', 'front'] return m[i] def indToAxis(i): m = ['X', 'Y', 'Z'] return m[i] def indToCube(i): if i == 0: return 'bbox' else: return f'cube{i-1}' def getSALine(logic, d_line, d_ret): sa_line = [] for var in logic: if var[0] == 'f': sum = 0. for ind, scale in var[1]: sum += float(scale) * d_line[ind] sa_line.append(sum) elif var[0] == 'i': if var[1] == 'const': sa_line.append(var[2]) elif var[1] == 'ret': sa_line.append(d_ret[var[2]]) else: sa_line.append(d_line[var[2]]) elif var[0] == 'b': if var[1] == 'const': sa_line.append(var[2]) else: sa_line.append(d_line[var[2]]) elif var[0] == 'c': if var[1] == 'const': sa_line.append(var[2]) else: sa_line.append(d_line[var[2]]) else: assert False, 'bad sa line arg' return sa_line def makeSALines(F, d_line, d_ret, last_cube): ret_num = 0 sa_lines = [] for fn, logic in zip(F.structure, F.getLogic()): params = getSALine(logic, d_line, d_ret) if fn == 'Cuboid': params[0] = clamp(params[0], 0.01, d_line[1]) params[1] = clamp(params[1], 0.01, d_line[2]) params[2] = clamp(params[2], 0.01, d_line[3]) else: params.insert(0, last_cube) if fn == 'attach': for i in range(2, 8): params[i] = clamp(params[i], 0.0, 1.0) # FLIP BBOX ATTACHES if params[1] == 'bbox': params[6] = 1 - params[6] # Sem Validity that attaches to BBox need to be top or bot if params[6] <= .5 and params[6] > .1: params[6] = 0.1 elif params[6] >= .5 and params[6] <.9: params[6] = 0.9 elif fn == 'squeeze': for i in range(4, 6): params[i] = clamp(params[i], 0.0, 1.0) elif fn == 'translate': params[2] = max(round(params[2] * utils.TRANS_NORM), 1) params[3] = clamp(params[3], 0.0, 1.0) mf = make_function(fn, params) ret = "" if fn == "Cuboid": ret = f"{d_ret[ret_num]} = " last_cube = d_ret[ret_num] ret_num += 1 sa_lines.append(ret + mf) return sa_lines, last_cube def makeSAProg(dsl, dsl_prog): bbox_dims = dsl_prog[0][1][1:] sa_prog = [f"{dsl_prog[0][0][0]} = {utils.make_function(dsl_prog[0][1][0], bbox_dims)}"] last_cube = 'bbox' for d_ret, d_line in dsl_prog[1:]: new_lines, last_cube = makeSALines( dsl.library[d_line[0]], [1.0] + list(bbox_dims[:3]) + list(d_line[1:]), d_ret, last_cube ) sa_prog += new_lines return sa_prog def fillProgram(dsl, hier, meta, func_field, cparam_field, dparam_field): prog = [] cube_num = -1 f_gt = hier[func_field].tolist() cp_gt = hier[cparam_field].tolist() dp_gt = hier[dparam_field].tolist() inv_func_map = {v:k for k,v in meta['func_map'].items()} for f_num, cl_prm, dl_prm in zip(f_gt, cp_gt, dp_gt): if f_num <= 1: continue fn = inv_func_map[f_num] line = [fn] j_prm = meta['jparam_map'][f_num] for j in j_prm: info = j.split('_') index = int(info[1]) if info[0] == 'c': lookup = cl_prm ptype = meta['cparam_map'][f_num][index] else: lookup = dl_prm ptype = meta['dparam_map'][f_num][index] if ptype == 'f': line.append(round(float(lookup[index]), 2)) elif ptype == 'i': if lookup[index] == 0: line.append('bbox') else: line.append(f'cube{int(lookup[index])-1}') elif ptype == 'sq': line.append(indToFace(lookup[index])) elif ptype == 'sym': line.append(indToAxis(lookup[index])) elif ptype == 'b': line.append(True if lookup[index] > 0 else False) else: assert False, f'bad input {j_prm}' ret = [] struct = dsl.library[fn].getStructure() for s in struct: if s == "Cuboid": if cube_num < 0: ret.append(f'bbox') else: ret.append(f'cube{cube_num}') cube_num += 1 prog.append((tuple(ret), tuple(line))) hier['dsl_prog'] = prog hier['prog'] = makeSAProg(dsl, prog) for c in hier['children']: if len(c) > 0: fillProgram(dsl, c, meta, func_field, cparam_field, dparam_field) def testGetProgram(dsl, out_name): from ShapeAssembly import hier_execute data = pickle.load(open(f'{out_name}_train.data', "rb")) metadata = pickle.load(open(f'{out_name}_train.meta', "rb")) for d in data: print(d[0]) fillProgram(dsl, d[1], metadata, 'func_gt', 'cparam_gt', 'dparam_gt') prog_lines = utils.getHierProgLines(d[1], 'dsl_prog') for l in prog_lines: print(l) verts, faces = hier_execute(d[1]) utils.writeObj(verts, faces, f'{d[0]}_dsl_gt.obj') def main(disc, out_name, in_dir, cat, config_path): config_mod = importlib.import_module(config_path) config = config_mod.loadConfig() dsl = DSL(config) for name, func in disc.ADD_FUNCS: dsl.library[name] = Function(func, dsl) dsl.full_library[name] = Function(func, dsl) for name in disc.RM_FUNCS: dsl.library.pop(name) with torch.no_grad(): writeData(dsl, out_name, in_dir, cat) if __name__ == "__main__": out_name = sys.argv[1] input_data = sys.argv[2] cat = sys.argv[3] disc = importlib.import_module(f'dsls.{sys.argv[4]}') config_path = sys.argv[5] main(disc, out_name, input_data, cat, config_path)
15,809
29.057034
107
py
ShapeMOD
ShapeMOD-main/SA_lang/tasks/sem_valid.py
import torch, os, sys, random import sa_utils as utils import pickle import model_prog as mp from copy import deepcopy from torch.distributions import Categorical from make_abs_data import fillProgram, makeSALines, getCuboidDims from ShapeAssembly import hier_execute, Program import numpy as np MAX_TRIES = 10 REJECT_MODE = 'HARD' BE = 1.1 MAX_SEQ = 30 sq_map = { 0:'left', 1:'right', 2:'bot', 3:'top', 4:'back', 5:'front' } sym_map = { 0: 'X', 1: 'Y', 2: 'Z' } def sem_eval_forward(net, inp_seq, code, code_start, bb_dims, hier_ind, P, sample): bb_dims = bb_dims.unsqueeze(0).unsqueeze(0).repeat(1,inp_seq.shape[1],1) hier_oh = torch.zeros(1, inp_seq.shape[1], mp.MAX_DEPTH).to(mp.device) hier_oh[0, :, min(hier_ind, 2)] = 1.0 inp = net.inp_net( torch.cat( (inp_seq, bb_dims, hier_oh), dim=2) ) gru_out, h = net.gru(inp, code.view(1,1,-1)) out = torch.zeros(inp_seq.shape, device=mp.device).float() commands = None for _net in net.net_list: if _net.func is not None: assert commands is not None if _net.func != commands: continue if _net.line_cond is not None: line_cond = out[:,:,_net.line_cond[0]:_net.line_cond[1]] else: line_cond = torch.zeros(inp_seq.shape[0], inp_seq.shape[1], 0, device=mp.device) if _net.bb_cond is True: bb_cond = bb_dims else: bb_cond = torch.zeros(inp_seq.shape[0], inp_seq.shape[1], 0, device=mp.device) raw_out = _net(torch.cat(( gru_out, line_cond, bb_cond ), dim=2)) if _net._type == 'func': dist = torch.softmax(raw_out.squeeze(), dim=0) mask = torch.zeros(dist.shape).float() if len(P.cuboids) == 0: mask[2] = 1.0 else: if len(P.cuboids) >= 2: lcmd = P.last_command else: lcmd = 'start' mask[torch.tensor(P.cmd_to_func_masks[lcmd])] = 1.0 dist = dist * mask.to(mp.device) dist[0] = 0. if sample: cmd = Categorical( dist ).sample() else: cmd = dist.argmax().item() out[0,0, _net.start+cmd] = 1.0 assert commands == None commands = cmd elif _net._type == 'disc': dist = torch.softmax(raw_out.squeeze(), dim=0) if sample: m = Categorical( dist ).sample() else: m = dist.argmax().item() out[0,0, _net.start+m] = 1.0 elif _net._type == 'b': if sample: r = torch.distributions.Bernoulli( torch.sigmoid(raw_out.squeeze()) ).sample().float() else: r = (raw_out.squeeze() >= 0.).float() out[0,0,_net.start:_net.end] = r elif _net._type == 'f': r = raw_out.squeeze() out[0,0,_net.start:_net.end] = r double_enc = torch.cat(( gru_out, code_start.repeat(1, gru_out.shape[1], 1) ), dim = 2) child_pred = net.child_net( double_enc ) next_codes = net.next_code_net( double_enc ).view(inp_seq.shape[0], inp_seq.shape[1], net.max_children, -1) return out, next_codes, child_pred, h def decode_sa_line(net, P, out, bbox_dims, meta): f_num, cl_prm, dl_prm = net.decode_line(out) assert f_num != 0, 'this should never happen' if f_num == 1: return ['<END>'] inv_func_map = {v:k for k,v in meta['func_map'].items()} fn = inv_func_map[f_num] line = [fn] j_prm = meta['jparam_map'][f_num] for j in j_prm: info = j.split('_') index = int(info[1]) if info[0] == 'c': lookup = cl_prm ptype = meta['cparam_map'][f_num][index] else: lookup = dl_prm ptype = meta['dparam_map'][f_num][index] if ptype == 'f': line.append(round(float(lookup[index]), 2)) elif ptype == 'i': if lookup[index] == 0: line.append('bbox') else: line.append(f'cube{int(lookup[index])-1}') elif ptype == 'sq': line.append(sq_map[lookup[index].item()]) elif ptype == 'sym': line.append(sym_map[lookup[index].item()]) elif ptype == 'b': line.append(bool(lookup[index])) else: assert False, f'bad input {j_prm}' ret = [] struct = meta['dsl'].library[fn].getStructure() cube_num = len([c for c in P.cuboids.values() if c.parent is None]) for s in struct: if s == "Cuboid": if cube_num == 0: ret.append(f'bbox') else: ret.append(f'cube{cube_num-1}') cube_num += 1 d_ret = tuple(ret) d_line = tuple(line) sa_lines, _ = makeSALines( meta['dsl'].library[d_line[0]], [1.0] + bbox_dims + list(d_line[1:]) , d_ret, P.last_cuboid ) return sa_lines def insideBBox(P): bbox_corners = P.cuboids[f"bbox"].getCorners() maxb = bbox_corners.max(dim=0).values minb = bbox_corners.min(dim=0).values assert (maxb > 0).all() assert (minb < 0).all() maxb = (maxb + 0.1) * BE minb = (minb - 0.1) * BE for ci in P.cuboids: if ci == 'bbox': continue corners = P.cuboids[ci].getCorners() maxc = corners.max(dim=0).values minc = corners.min(dim=0).values if (maxc > maxb).any(): return False if (minc < minb).any(): return False return True def checkValidLine(net, P, out, bbox_dims, meta): sa_lines = decode_sa_line(net, P, out, bbox_dims, meta) attaches_to_add = [] cube_syms = [] for line in sa_lines: if '<END>' in line: # Make sure nothing is left unmoved for v in P.cube_attaches.values(): if len(v) == 0: return None, False if 'Cuboid(' in line: parse = P.parseCuboid(line) P.last_cuboid = parse[0] if 'bbox' not in line: P.last_command = 'Cuboid' P.cube_attaches[parse[0]] = [] if 'attach' in line: parse = P.parseAttach(line) if parse[0] != P.last_cuboid: assert False, 'how did this fail' return None, False P.last_command = 'attach' attaches_to_add.append((parse[0], parse[1])) if 'squeeze' in line: parse = P.parseSqueeze(line) if parse[0] != P.last_cuboid: assert False, 'how did this fail' return None, False P.last_command = 'squeeze' attaches_to_add.append((parse[0], parse[1])) attaches_to_add.append((parse[0], parse[2])) if 'translate' in line: parse = P.parseTranslate(line) if parse[0] != P.last_cuboid: assert False, 'how did this fail' return None, False P.last_command = 'translate' cube_syms.append(parse[0]) if 'reflect' in line: parse = P.parseReflect(line) if parse[0] != P.last_cuboid: assert False, 'how did this fail' return None, False P.last_command = 'reflect' cube_syms.append(parse[0]) try: P.execute(line) P.last_command_cuboid = 'Cuboid(' in line except Exception as e: if mp.VERBOSE: print(f"failed line {line} with {e}") # Return none return None, False for a, o in attaches_to_add: past_attaches = P.cube_attaches[a] # Attached to non bounding box more than once if o != 'bbox' and o in past_attaches: return None, False # Already has max attachments if len(past_attaches) == 2: return None, False P.cube_attaches[a].append(o) for c in cube_syms: if c in P.cube_syms: return None, False P.cube_syms.add(c) return P, insideBBox(P) def getSpecFuncs(meta, types): l = [] if 'end' in types: l.append(1) dsl = meta['dsl'] for k, v in meta['func_map'].items(): if k in dsl.library: if dsl.library[k].structure[0] in types: l.append(v) return l def sem_eval_prog(net, code, rejection_sample = True): try: return _sem_eval_prog(net, code, rejection_sample, None) except AssertionError as e: if e.args[0] != 'rejection sample': raise e return {} def _sem_eval_prog(net, code, rejection_sample, node): is_root = False if node is None: is_root = True bb_dims = net.bb_net(code) node = { 'depth': 0, 'bb_dims': bb_dims } if node['depth'] > mp.MAX_DEPTH: node.pop('depth') node.pop('bb_dims') return h = code.view(1,1, -1) h_start = h.clone() inp = net.getStartLine() out_lines = [] children = [] P = Program() P.reject = False P.cube_attaches = {} P.cube_syms = set() meta = net.metadata P.last_cuboid = None P.last_command = 'start' P.cmd_to_func_masks = { 'start': getSpecFuncs(meta, ('Cuboid')), 'Cuboid': getSpecFuncs(meta, ('attach', 'squeeze')), 'attach': getSpecFuncs(meta, ('attach', 'reflect', 'translate', 'Cuboid', 'end')), 'squeeze': getSpecFuncs(meta, ('reflect', 'translate', 'Cuboid', 'end')), 'reflect': getSpecFuncs(meta, ('Cuboid', 'end')), 'translate': getSpecFuncs(meta, ('Cuboid', 'end')) } P.cuboids.pop('bbox') for i in range(MAX_SEQ): should_add = True for j in range(MAX_TRIES): # Failed to make valid prog if j == (MAX_TRIES-1) or (P.reject and rejection_sample): P.reject = True should_add = False break new_inp, pnext, pchild, new_h = sem_eval_forward( net, inp, h, h_start, node['bb_dims'], node['depth'], P, j > 0 ) clean_out = new_inp.squeeze() new_P, valid = checkValidLine( net, deepcopy(P), clean_out, node['bb_dims'].tolist(), net.metadata ) if not valid: if REJECT_MODE == 'HARD': P.reject = True continue inp = new_inp h = new_h P = new_P break if not should_add: break fstart, fend = net.metadata['tl_map']['func'] func_ind = torch.argmax(clean_out[fstart:fend]).item() if func_ind == 1: break out_lines.append(clean_out) child_pred = pchild[0][0] next_codes = pnext[0][0] _, _c, _ = net.decode_line(clean_out) cube_dims = getCuboidDims( net.metadata['dsl'].library[ net.metadata['rev_func_map'][func_ind] ], _c, net.metadata['jparam_map'][func_ind], node['bb_dims'] ) cube_dims = torch.tensor(cube_dims,device=child_pred.device) for i in range(net.metadata['num_cube_map'][func_ind]): if child_pred[i].item() >= 0.0: child = { 'depth': node['depth']+1, 'bb_dims': cube_dims[i] } children.append(child) _sem_eval_prog(net, next_codes[i], rejection_sample, child) else: children.append({}) node['children'] = children out_funcs, out_cprms, out_dprms = net.split_lines(out_lines) node[mp.FUNC_PRED_FIELD] = torch.tensor(out_funcs) node[mp.CPARAM_PRED_FIELD] = torch.stack(out_cprms) if len(out_cprms) > 0 else torch.tensor([]) node[mp.DPARAM_PRED_FIELD] = torch.stack(out_dprms) if len(out_dprms) > 0 else torch.tensor([]) if P.reject and rejection_sample: assert False, 'rejection sample' return node
13,496
27.656051
99
py
ShapeMOD
ShapeMOD-main/SA_lang/tasks/pointnet_fd.py
from __future__ import print_function import argparse import os import random import torch import torch.nn as nn import torch.optim as optim import torch.utils.data as data from torch.autograd import Variable import json import torch.nn.functional as F from tqdm import tqdm import numpy as np import pickle from sa_utils import sample_surface from scipy import linalg def calculate_frechet_distance(mu1, sigma1, mu2, sigma2, eps=1e-6): """Numpy implementation of the Frechet Distance. The Frechet distance between two multivariate Gaussians X_1 ~ N(mu_1, C_1) and X_2 ~ N(mu_2, C_2) is d^2 = ||mu_1 - mu_2||^2 + Tr(C_1 + C_2 - 2*sqrt(C_1*C_2)). Stable version by Dougal J. Sutherland. Params: -- mu1 : Numpy array containing the activations of a layer of the inception net (like returned by the function 'get_predictions') for generated samples. -- mu2 : The sample mean over activations, precalculated on an representative data set. -- sigma1: The covariance matrix over activations for generated samples. -- sigma2: The covariance matrix over activations, precalculated on an representative data set. Returns: -- : The Frechet Distance. """ mu1 = np.atleast_1d(mu1) mu2 = np.atleast_1d(mu2) sigma1 = np.atleast_2d(sigma1) sigma2 = np.atleast_2d(sigma2) assert mu1.shape == mu2.shape, \ 'Training and test mean vectors have different lengths' assert sigma1.shape == sigma2.shape, \ 'Training and test covariances have different dimensions' diff = mu1 - mu2 # Product might be almost singular covmean, _ = linalg.sqrtm(sigma1.dot(sigma2), disp=False) if not np.isfinite(covmean).all(): msg = ('fid calculation produces singular product; ' 'adding %s to diagonal of cov estimates') % eps print(msg) offset = np.eye(sigma1.shape[0]) * eps covmean = linalg.sqrtm((sigma1 + offset).dot(sigma2 + offset)) # Numerical error might give slight imaginary component if np.iscomplexobj(covmean): if not np.allclose(np.diagonal(covmean).imag, 0, atol=1e-3): m = np.max(np.abs(covmean.imag)) raise ValueError('Imaginary component {}'.format(m)) covmean = covmean.real tr_covmean = np.trace(covmean) return (diff.dot(diff) + np.trace(sigma1) + np.trace(sigma2) - 2 * tr_covmean) class STN3d(nn.Module): def __init__(self): super(STN3d, self).__init__() self.conv1 = torch.nn.Conv1d(3, 64, 1) self.conv2 = torch.nn.Conv1d(64, 128, 1) self.conv3 = torch.nn.Conv1d(128, 1024, 1) self.fc1 = nn.Linear(1024, 512) self.fc2 = nn.Linear(512, 256) self.fc3 = nn.Linear(256, 9) self.relu = nn.ReLU() self.bn1 = nn.BatchNorm1d(64) self.bn2 = nn.BatchNorm1d(128) self.bn3 = nn.BatchNorm1d(1024) self.bn4 = nn.BatchNorm1d(512) self.bn5 = nn.BatchNorm1d(256) def forward(self, x): batchsize = x.size()[0] x = F.relu(self.bn1(self.conv1(x))) x = F.relu(self.bn2(self.conv2(x))) x = F.relu(self.bn3(self.conv3(x))) x = torch.max(x, 2, keepdim=True)[0] x = x.view(-1, 1024) x = F.relu(self.bn4(self.fc1(x))) x = F.relu(self.bn5(self.fc2(x))) x = self.fc3(x) iden = Variable(torch.from_numpy(np.array([1,0,0,0,1,0,0,0,1]).astype(np.float32))).view(1,9).repeat(batchsize,1).to(x.device) x = x + iden x = x.view(-1, 3, 3) return x class STNkd(nn.Module): def __init__(self, k=64): super(STNkd, self).__init__() self.conv1 = torch.nn.Conv1d(k, 64, 1) self.conv2 = torch.nn.Conv1d(64, 128, 1) self.conv3 = torch.nn.Conv1d(128, 1024, 1) self.fc1 = nn.Linear(1024, 512) self.fc2 = nn.Linear(512, 256) self.fc3 = nn.Linear(256, k*k) self.relu = nn.ReLU() self.bn1 = nn.BatchNorm1d(64) self.bn2 = nn.BatchNorm1d(128) self.bn3 = nn.BatchNorm1d(1024) self.bn4 = nn.BatchNorm1d(512) self.bn5 = nn.BatchNorm1d(256) self.k = k def forward(self, x): batchsize = x.size()[0] x = F.relu(self.bn1(self.conv1(x))) x = F.relu(self.bn2(self.conv2(x))) x = F.relu(self.bn3(self.conv3(x))) x = torch.max(x, 2, keepdim=True)[0] x = x.view(-1, 1024) x = F.relu(self.bn4(self.fc1(x))) x = F.relu(self.bn5(self.fc2(x))) x = self.fc3(x) iden = Variable(torch.from_numpy(np.eye(self.k).flatten().astype(np.float32))).view(1,self.k*self.k).repeat(batchsize,1) x = x + iden x = x.view(-1, self.k, self.k) return x class PointNetfeat(nn.Module): def __init__(self, global_feat = True, feature_transform = False): super(PointNetfeat, self).__init__() self.stn = STN3d() self.conv1 = torch.nn.Conv1d(3, 64, 1) self.conv2 = torch.nn.Conv1d(64, 128, 1) self.conv3 = torch.nn.Conv1d(128, 1024, 1) self.bn1 = nn.BatchNorm1d(64) self.bn2 = nn.BatchNorm1d(128) self.bn3 = nn.BatchNorm1d(1024) self.global_feat = global_feat self.feature_transform = feature_transform if self.feature_transform: self.fstn = STNkd(k=64) def forward(self, x): n_pts = x.size()[2] trans = self.stn(x) x = x.transpose(2, 1) x = torch.bmm(x, trans) x = x.transpose(2, 1) x = F.relu(self.bn1(self.conv1(x))) if self.feature_transform: trans_feat = self.fstn(x) x = x.transpose(2,1) x = torch.bmm(x, trans_feat) x = x.transpose(2,1) else: trans_feat = None pointfeat = x x = F.relu(self.bn2(self.conv2(x))) x = self.bn3(self.conv3(x)) x = torch.max(x, 2, keepdim=True)[0] x = x.view(-1, 1024) if self.global_feat: return x, trans, trans_feat else: x = x.view(-1, 1024, 1).repeat(1, 1, n_pts) return torch.cat([x, pointfeat], 1), trans, trans_feat class PointNetCls(nn.Module): def __init__(self, k=2, feature_transform=False): super(PointNetCls, self).__init__() self.feature_transform = feature_transform self.feat = PointNetfeat(global_feat=True, feature_transform=feature_transform) self.fc1 = nn.Linear(1024, 512) self.fc2 = nn.Linear(512, 256) self.fc3 = nn.Linear(256, k) self.dropout = nn.Dropout(p=0.3) self.bn1 = nn.BatchNorm1d(512) self.bn2 = nn.BatchNorm1d(256) self.relu = nn.ReLU() def get_emb(self, x): x1, trans, trans_feat = self.feat(x) x2 = F.relu(self.bn1(self.fc1(x1))) x3 = self.fc2(x2) return x3 def forward(self, x): x1, trans, trans_feat = self.feat(x) x2 = F.relu(self.bn1(self.fc1(x1))) x3 = F.relu(self.bn2(self.fc2(x2))) xy = self.fc3(x3) return F.log_softmax(xy, dim=1), trans, trans_feat, x1 class PointNetDenseCls(nn.Module): def __init__(self, k = 2, feature_transform=False): super(PointNetDenseCls, self).__init__() self.k = k self.feature_transform=feature_transform self.feat = PointNetfeat(global_feat=False, feature_transform=feature_transform) self.conv1 = torch.nn.Conv1d(1088, 512, 1) self.conv2 = torch.nn.Conv1d(512, 256, 1) self.conv3 = torch.nn.Conv1d(256, 128, 1) self.conv4 = torch.nn.Conv1d(128, self.k, 1) self.bn1 = nn.BatchNorm1d(512) self.bn2 = nn.BatchNorm1d(256) self.bn3 = nn.BatchNorm1d(128) def forward(self, x): batchsize = x.size()[0] n_pts = x.size()[2] x, trans, trans_feat = self.feat(x) x = F.relu(self.bn1(self.conv1(x))) x = F.relu(self.bn2(self.conv2(x))) x = F.relu(self.bn3(self.conv3(x))) x = self.conv4(x) x = x.transpose(2,1).contiguous() x = F.log_softmax(x.view(-1,self.k), dim=-1) x = x.view(batchsize, n_pts, self.k) return x, trans, trans_feat def feature_transform_regularizer(trans): d = trans.size()[1] batchsize = trans.size()[0] I = torch.eye(d)[None, :, :] loss = torch.mean(torch.norm(torch.bmm(trans, trans.transpose(2,1)) - I, dim=(1,2))) return loss def get_fd(s1, s2, device=None): class ClassificationDataset(): def __init__(self, samples): super(ClassificationDataset, self).__init__() self.datas = samples self.N = len(self.datas) def __len__(self): return self.N def __getitem__(self,index): return torch.clamp(self.datas[index], -1.0, 1.0) num_classes = 16 test_dataset1 = ClassificationDataset(s1) test_dataset2 = ClassificationDataset(s2) test_loader1 = torch.utils.data.DataLoader( test_dataset1, batch_size = 16, shuffle = True ) test_loader2 = torch.utils.data.DataLoader( test_dataset2, batch_size = 16, shuffle = True ) classifier = PointNetCls(k=num_classes, feature_transform=False) classifier.load_state_dict(torch.load("data/pgp_pointnet_cls_model_249.pth")) classifier = classifier.eval() if device is not None: classifier = classifier.to(device) def calculate_activation_statistics(loader): embs = [] for i, points in enumerate(loader, 0): points = points.transpose(2, 1) if device is not None: points = points.to(device) _, _, _, emb = classifier(points) embs.append(emb.detach().cpu()) act = torch.cat(embs, dim=0).numpy() mu = np.mean(act, axis=0) sigma = np.cov(act, rowvar=False) return mu, sigma m1, s1 = calculate_activation_statistics(test_loader1) m2, s2 = calculate_activation_statistics(test_loader2) return calculate_frechet_distance(m1, s1, m2, s2) if __name__ == '__main__': import os import utils import sys from ShapeAssembly import hier_execute def load_samps(fold, start, per): inds = os.listdir(fold) s = [] for ind in tqdm(inds[start:per+start]): hp = utils.loadHPFromFile(fold+ind) verts, faces = hier_execute(hp) s.append((verts, faces)) return s s1 = load_samps(sys.argv[1], int(sys.argv[4]), int(sys.argv[3])) s2 = load_samps(sys.argv[2], int(sys.argv[5]), int(sys.argv[3])) print(get_fd(s1, s2))
10,819
32.190184
134
py
ShapeMOD
ShapeMOD-main/SA_lang/tasks/infer_sem_valid.py
import torch, os, sys, random import sa_utils as utils import pickle import model_prog as mp from copy import deepcopy from torch.distributions import Categorical from make_abs_data import fillProgram, makeSALines, getCuboidDims from ShapeAssembly import hier_execute, Program import numpy as np MAX_TRIES = 50 BE = 1.5 MAX_SEQ = 30 sq_map = { 0:'left', 1:'right', 2:'bot', 3:'top', 4:'back', 5:'front' } sym_map = { 0: 'X', 1: 'Y', 2: 'Z' } def sem_eval_forward(net, inp_seq, code, code_start, bb_dims, hier_ind, P, samp_ind): bb_dims = bb_dims.unsqueeze(0).unsqueeze(0).repeat(1,inp_seq.shape[1],1) hier_oh = torch.zeros(1, inp_seq.shape[1], mp.MAX_DEPTH).to(mp.device) hier_oh[0, :, min(hier_ind, 2)] = 1.0 inp = net.inp_net( torch.cat( (inp_seq, bb_dims, hier_oh), dim=2) ) gru_out, h = net.gru(inp, code.view(1,1,-1)) out = torch.zeros(inp_seq.shape, device=mp.device).float() commands = None for _net in net.net_list: if _net.func is not None: assert commands is not None if _net.func != commands: continue if _net.line_cond is not None: line_cond = out[:,:,_net.line_cond[0]:_net.line_cond[1]] else: line_cond = torch.zeros(inp_seq.shape[0], inp_seq.shape[1], 0, device=mp.device) if _net.bb_cond is True: bb_cond = bb_dims else: bb_cond = torch.zeros(inp_seq.shape[0], inp_seq.shape[1], 0, device=mp.device) raw_out = _net(torch.cat(( gru_out, line_cond, bb_cond ), dim=2)) if _net._type == 'func': if samp_ind == 0: dist = torch.softmax(raw_out.squeeze(), dim=0) else: dist = torch.softmax( raw_out.squeeze() \ / (samp_ind * 1.0), dim=0 ) mask = torch.zeros(dist.shape).float() if len(P.cuboids) == 0: mask[2] = 1.0 else: if len(P.cuboids) >= 2: lcmd = P.last_command else: lcmd = 'start' mask[torch.tensor(P.cmd_to_func_masks[lcmd])] = 1.0 dist = dist * mask.to(mp.device) dist[0] = 0. if samp_ind > 0: cmd = Categorical( dist ).sample() else: cmd = dist.argmax().item() out[0,0, _net.start+cmd] = 1.0 assert commands == None commands = cmd elif _net._type == 'disc': dist = torch.softmax(raw_out.squeeze(), dim=0) if samp_ind > 0: m = Categorical( dist ).sample() else: m = dist.argmax().item() out[0,0, _net.start+m] = 1.0 elif _net._type == 'b': if samp_ind > 0: r = torch.distributions.Bernoulli( torch.sigmoid(raw_out.squeeze()/(1.0 * samp_ind)) ).sample().float() else: r = (raw_out.squeeze() >= 0.).float() out[0,0,_net.start:_net.end] = r elif _net._type == 'f': r = raw_out.squeeze() out[0,0,_net.start:_net.end] = r double_enc = torch.cat(( gru_out, code_start.repeat(1, gru_out.shape[1], 1) ), dim = 2) child_pred = net.child_net( double_enc ) next_codes = net.next_code_net( double_enc ).view(inp_seq.shape[0], inp_seq.shape[1], net.max_children, -1) return out, next_codes, child_pred, h def decode_sa_line(net, P, out, bbox_dims, meta): f_num, cl_prm, dl_prm = net.decode_line(out) assert f_num != 0, 'this should never happen' if f_num == 1: return ['<END>'] inv_func_map = {v:k for k,v in meta['func_map'].items()} fn = inv_func_map[f_num] line = [fn] j_prm = meta['jparam_map'][f_num] for j in j_prm: info = j.split('_') index = int(info[1]) if info[0] == 'c': lookup = cl_prm ptype = meta['cparam_map'][f_num][index] else: lookup = dl_prm ptype = meta['dparam_map'][f_num][index] if ptype == 'f': line.append(round(float(lookup[index]), 2)) elif ptype == 'i': if lookup[index] == 0: line.append('bbox') else: line.append(f'cube{int(lookup[index])-1}') elif ptype == 'sq': line.append(sq_map[lookup[index].item()]) elif ptype == 'sym': line.append(sym_map[lookup[index].item()]) elif ptype == 'b': line.append(bool(lookup[index])) else: assert False, f'bad input {j_prm}' ret = [] struct = meta['dsl'].library[fn].getStructure() cube_num = len([c for c in P.cuboids.values() if c.parent is None]) for s in struct: if s == "Cuboid": if cube_num == 0: ret.append(f'bbox') else: ret.append(f'cube{cube_num-1}') cube_num += 1 d_ret = tuple(ret) d_line = tuple(line) sa_lines, _ = makeSALines( meta['dsl'].library[d_line[0]], [1.0] + bbox_dims + list(d_line[1:]) , d_ret, P.last_cuboid ) return sa_lines def insideBBox(P): bbox_corners = P.cuboids[f"bbox"].getCorners() maxb = bbox_corners.max(dim=0).values minb = bbox_corners.min(dim=0).values assert (maxb > 0).all() assert (minb < 0).all() maxb = (maxb + 0.1) * BE minb = (minb - 0.1) * BE for ci in P.cuboids: if ci == 'bbox': continue corners = P.cuboids[ci].getCorners() maxc = corners.max(dim=0).values minc = corners.min(dim=0).values if (maxc > maxb).any(): return False if (minc < minb).any(): return False return True def checkValidLine(net, P, out, bbox_dims, meta): sa_lines = decode_sa_line(net, P, out, bbox_dims, meta) attaches_to_add = [] cube_syms = [] for line in sa_lines: if '<END>' in line: # Make sure nothing is left unmoved for v in P.cube_attaches.values(): if len(v) == 0: return None, False if 'Cuboid(' in line: parse = P.parseCuboid(line) P.last_cuboid = parse[0] if 'bbox' not in line: P.last_command = 'Cuboid' P.cube_attaches[parse[0]] = [] if 'attach' in line: parse = P.parseAttach(line) if parse[0] != P.last_cuboid: assert False, 'how did this fail' return None, False P.last_command = 'attach' attaches_to_add.append((parse[0], parse[1])) if 'squeeze' in line: parse = P.parseSqueeze(line) if parse[0] != P.last_cuboid: assert False, 'how did this fail' return None, False P.last_command = 'squeeze' attaches_to_add.append((parse[0], parse[1])) attaches_to_add.append((parse[0], parse[2])) if 'translate' in line: parse = P.parseTranslate(line) if parse[0] != P.last_cuboid: assert False, 'how did this fail' return None, False P.last_command = 'translate' cube_syms.append(parse[0]) if 'reflect' in line: parse = P.parseReflect(line) if parse[0] != P.last_cuboid: assert False, 'how did this fail' return None, False P.last_command = 'reflect' cube_syms.append(parse[0]) try: P.execute(line) P.last_command_cuboid = 'Cuboid(' in line except Exception as e: if mp.VERBOSE: print(f"failed line {line} with {e}") # Return none return None, False for a, o in attaches_to_add: past_attaches = P.cube_attaches[a] # Attached to non bounding box more than once if o != 'bbox' and o in past_attaches: return None, False # Already has max attachments if len(past_attaches) == 2: return None, False P.cube_attaches[a].append(o) for c in cube_syms: if c in P.cube_syms: return None, False P.cube_syms.add(c) return P, insideBBox(P) def getSpecFuncs(meta, types): l = [] if 'end' in types: l.append(1) dsl = meta['dsl'] for k, v in meta['func_map'].items(): if k in dsl.library: if dsl.library[k].structure[0] in types: l.append(v) return l def sem_eval_prog(net, code, node=None): is_root = False if node is None: is_root = True bb_dims = net.bb_net(code) node = { 'depth': 0, 'bb_dims': bb_dims } if node['depth'] > mp.MAX_DEPTH: node.pop('depth') node.pop('bb_dims') return h = code.view(1,1, -1) h_start = h.clone() inp = net.getStartLine() out_lines = [] children = [] P = Program() P.cube_attaches = {} P.cube_syms = set() meta = net.metadata P.last_cuboid = None P.last_command = 'start' P.cmd_to_func_masks = { 'start': getSpecFuncs(meta, ('Cuboid')), 'Cuboid': getSpecFuncs(meta, ('attach', 'squeeze')), 'attach': getSpecFuncs(meta, ('attach', 'reflect', 'translate', 'Cuboid', 'end')), 'squeeze': getSpecFuncs(meta, ('reflect', 'translate', 'Cuboid', 'end')), 'reflect': getSpecFuncs(meta, ('Cuboid', 'end')), 'translate': getSpecFuncs(meta, ('Cuboid', 'end')) } P.cuboids.pop('bbox') for i in range(MAX_SEQ): should_break = False for j in range(MAX_TRIES): # Failed to make valid prog if j == (MAX_TRIES-1): keys = list(node.keys()) for k in keys: if k in node: node.pop(k) should_break = True break new_inp, pnext, pchild, new_h = sem_eval_forward( net, inp, h, h_start, node['bb_dims'], node['depth'], P, j ) clean_out = new_inp.squeeze() new_P, valid = checkValidLine( net, deepcopy(P), clean_out, node['bb_dims'].tolist(), net.metadata ) if not valid: continue inp = new_inp h = new_h P = new_P break if should_break: break fstart, fend = net.metadata['tl_map']['func'] func_ind = torch.argmax(clean_out[fstart:fend]).item() if func_ind == 1: break out_lines.append(clean_out) child_pred = pchild[0][0] next_codes = pnext[0][0] _, _c, _ = net.decode_line(clean_out) cube_dims = getCuboidDims( net.metadata['dsl'].library[ net.metadata['rev_func_map'][func_ind] ], _c, net.metadata['jparam_map'][func_ind], node['bb_dims'] ) cube_dims = torch.tensor(cube_dims,device=child_pred.device) for i in range(net.metadata['num_cube_map'][func_ind]): if child_pred[i].item() >= 0.0: child = { 'depth': node['depth']+1, 'bb_dims': cube_dims[i] } children.append(child) sem_eval_prog(net, next_codes[i], child) else: children.append({}) if len(node) == 0: if not is_root: return node else: return dummy_prog(net) node['children'] = children out_funcs, out_cprms, out_dprms = net.split_lines(out_lines) node[mp.FUNC_PRED_FIELD] = torch.tensor(out_funcs) node[mp.CPARAM_PRED_FIELD] = torch.stack(out_cprms) if len(out_cprms) > 0 else torch.tensor([]) node[mp.DPARAM_PRED_FIELD] = torch.stack(out_dprms) if len(out_dprms) > 0 else torch.tensor([]) return node def dummy_prog(net): node = {} node['children'] = [{}, {}] node[mp.FUNC_PRED_FIELD] = torch.tensor([2,2]) node[mp.CPARAM_PRED_FIELD] = torch.ones(2, net.metadata['max_cparams']).float() node[mp.DPARAM_PRED_FIELD] = torch.zeros(2, net.metadata['max_dparams']).long() return node
13,859
27.401639
99
py
ShapeMOD
ShapeMOD-main/SA_lang/tasks/etw_pytorch_utils.py
# From https://github.com/erikwijmans from __future__ import ( division, absolute_import, with_statement, print_function, unicode_literals, ) import torch.nn as nn import os import torch import torch.nn as nn from torch.autograd.function import InplaceFunction from itertools import repeat import numpy as np import shutil import tqdm from scipy.stats import t as student_t import statistics as stats class SharedMLP(nn.Sequential): def __init__( self, args, bn=False, activation=nn.ReLU(inplace=True), preact=False, first=False, name="", ): # type: (SharedMLP, List[int], bool, Any, bool, bool, AnyStr) -> None super(SharedMLP, self).__init__() for i in range(len(args) - 1): self.add_module( name + "layer{}".format(i), Conv2d( args[i], args[i + 1], bn=(not first or not preact or (i != 0)) and bn, activation=activation if (not first or not preact or (i != 0)) else None, preact=preact, ), ) class _BNBase(nn.Sequential): def __init__(self, in_size, batch_norm=None, name=""): super(_BNBase, self).__init__() self.add_module(name + "bn", batch_norm(in_size)) nn.init.constant_(self[0].weight, 1.0) nn.init.constant_(self[0].bias, 0) class BatchNorm1d(_BNBase): def __init__(self, in_size, name=""): # type: (BatchNorm1d, int, AnyStr) -> None super(BatchNorm1d, self).__init__(in_size, batch_norm=nn.BatchNorm1d, name=name) class BatchNorm2d(_BNBase): def __init__(self, in_size, name=""): # type: (BatchNorm2d, int, AnyStr) -> None super(BatchNorm2d, self).__init__(in_size, batch_norm=nn.BatchNorm2d, name=name) class BatchNorm3d(_BNBase): def __init__(self, in_size, name=""): # type: (BatchNorm3d, int, AnyStr) -> None super(BatchNorm3d, self).__init__(in_size, batch_norm=nn.BatchNorm3d, name=name) class _ConvBase(nn.Sequential): def __init__( self, in_size, out_size, kernel_size, stride, padding, dilation, activation, bn, init, conv=None, norm_layer=None, bias=True, preact=False, name="", ): super(_ConvBase, self).__init__() bias = bias and (not bn) conv_unit = conv( in_size, out_size, kernel_size=kernel_size, stride=stride, padding=padding, dilation=dilation, bias=bias, ) init(conv_unit.weight) if bias: nn.init.constant_(conv_unit.bias, 0) if bn: if not preact: bn_unit = norm_layer(out_size) else: bn_unit = norm_layer(in_size) if preact: if bn: self.add_module(name + "normlayer", bn_unit) if activation is not None: self.add_module(name + "activation", activation) self.add_module(name + "conv", conv_unit) if not preact: if bn: self.add_module(name + "normlayer", bn_unit) if activation is not None: self.add_module(name + "activation", activation) class Conv1d(_ConvBase): def __init__( self, in_size, out_size, kernel_size=1, stride=1, padding=0, dilation=1, activation=nn.ReLU(inplace=True), bn=False, init=nn.init.kaiming_normal_, bias=True, preact=False, name="", norm_layer=BatchNorm1d, ): # type: (Conv1d, int, int, int, int, int, int, Any, bool, Any, bool, bool, AnyStr, _BNBase) -> None super(Conv1d, self).__init__( in_size, out_size, kernel_size, stride, padding, dilation, activation, bn, init, conv=nn.Conv1d, norm_layer=norm_layer, bias=bias, preact=preact, name=name, ) class Conv2d(_ConvBase): def __init__( self, in_size, out_size, kernel_size=(1, 1), stride=(1, 1), padding=(0, 0), dilation=(1, 1), activation=nn.ReLU(inplace=True), bn=False, init=nn.init.kaiming_normal_, bias=True, preact=False, name="", norm_layer=BatchNorm2d, ): # type: (Conv2d, int, int, Tuple[int, int], Tuple[int, int], Tuple[int, int], Tuple[int, int], Any, bool, Any, bool, bool, AnyStr, _BNBase) -> None super(Conv2d, self).__init__( in_size, out_size, kernel_size, stride, padding, dilation, activation, bn, init, conv=nn.Conv2d, norm_layer=norm_layer, bias=bias, preact=preact, name=name, ) class Conv3d(_ConvBase): def __init__( self, in_size, out_size, kernel_size=(1, 1, 1), stride=(1, 1, 1), padding=(0, 0, 0), dilation=(1, 1, 1), activation=nn.ReLU(inplace=True), bn=False, init=nn.init.kaiming_normal_, bias=True, preact=False, name="", norm_layer=BatchNorm3d, ): # type: (Conv3d, int, int, Tuple[int, int, int], Tuple[int, int, int], Tuple[int, int, int], Tuple[int, int, int], Any, bool, Any, bool, bool, AnyStr, _BNBase) -> None super(Conv3d, self).__init__( in_size, out_size, kernel_size, stride, padding, dilation, activation, bn, init, conv=nn.Conv3d, norm_layer=norm_layer, bias=bias, preact=preact, name=name, ) class FC(nn.Sequential): def __init__( self, in_size, out_size, activation=nn.ReLU(inplace=True), bn=False, init=None, preact=False, name="", ): # type: (FC, int, int, Any, bool, Any, bool, AnyStr) -> None super(FC, self).__init__() fc = nn.Linear(in_size, out_size, bias=not bn) if init is not None: init(fc.weight) if not bn: nn.init.constant_(fc.bias, 0) if preact: if bn: self.add_module(name + "bn", BatchNorm1d(in_size)) if activation is not None: self.add_module(name + "activation", activation) self.add_module(name + "fc", fc) if not preact: if bn: self.add_module(name + "bn", BatchNorm1d(out_size)) if activation is not None: self.add_module(name + "activation", activation) class _DropoutNoScaling(InplaceFunction): @staticmethod def _make_noise(input): return input.new().resize_as_(input) @staticmethod def symbolic(g, input, p=0.5, train=False, inplace=False): if inplace: return None n = g.appendNode( g.create("Dropout", [input]).f_("ratio", p).i_("is_test", not train) ) real = g.appendNode(g.createSelect(n, 0)) g.appendNode(g.createSelect(n, 1)) return real @classmethod def forward(cls, ctx, input, p=0.5, train=False, inplace=False): if p < 0 or p > 1: raise ValueError( "dropout probability has to be between 0 and 1, " "but got {}".format(p) ) ctx.p = p ctx.train = train ctx.inplace = inplace if ctx.inplace: ctx.mark_dirty(input) output = input else: output = input.clone() if ctx.p > 0 and ctx.train: ctx.noise = cls._make_noise(input) if ctx.p == 1: ctx.noise.fill_(0) else: ctx.noise.bernoulli_(1 - ctx.p) ctx.noise = ctx.noise.expand_as(input) output.mul_(ctx.noise) return output @staticmethod def backward(ctx, grad_output): if ctx.p > 0 and ctx.train: return grad_output.mul(ctx.noise), None, None, None else: return grad_output, None, None, None dropout_no_scaling = _DropoutNoScaling.apply class _FeatureDropoutNoScaling(_DropoutNoScaling): @staticmethod def symbolic(input, p=0.5, train=False, inplace=False): return None @staticmethod def _make_noise(input): return input.new().resize_( input.size(0), input.size(1), *repeat(1, input.dim() - 2) ) feature_dropout_no_scaling = _FeatureDropoutNoScaling.apply def group_model_params(model, **kwargs): # type: (nn.Module, ...) -> List[Dict] decay_group = [] no_decay_group = [] for name, param in model.named_parameters(): if name.find("normlayer") != -1 or name.find("bias") != -1: no_decay_group.append(param) else: decay_group.append(param) assert len(list(model.parameters())) == len(decay_group) + len(no_decay_group) return [ dict(params=decay_group, **kwargs), dict(params=no_decay_group, weight_decay=0.0, **kwargs), ] def checkpoint_state(model=None, optimizer=None, best_prec=None, epoch=None, it=None): optim_state = optimizer.state_dict() if optimizer is not None else None if model is not None: if isinstance(model, torch.nn.DataParallel): model_state = model.module.state_dict() else: model_state = model.state_dict() else: model_state = None return { "epoch": epoch, "it": it, "best_prec": best_prec, "model_state": model_state, "optimizer_state": optim_state, } def save_checkpoint(state, is_best, filename="checkpoint", bestname="model_best"): filename = "{}.pth.tar".format(filename) torch.save(state, filename) if is_best: shutil.copyfile(filename, "{}.pth.tar".format(bestname)) def load_checkpoint(model=None, optimizer=None, filename="checkpoint"): filename = "{}.pth.tar".format(filename) if os.path.isfile(filename): print("==> Loading from checkpoint '{}'".format(filename)) checkpoint = torch.load(filename) epoch = checkpoint["epoch"] it = checkpoint.get("it", 0.0) best_prec = checkpoint["best_prec"] if model is not None and checkpoint["model_state"] is not None: model.load_state_dict(checkpoint["model_state"]) if optimizer is not None and checkpoint["optimizer_state"] is not None: optimizer.load_state_dict(checkpoint["optimizer_state"]) print("==> Done") return it, epoch, best_prec else: print("==> Checkpoint '{}' not found".format(filename)) return None def variable_size_collate(pad_val=0, use_shared_memory=True): import collections _numpy_type_map = { "float64": torch.DoubleTensor, "float32": torch.FloatTensor, "float16": torch.HalfTensor, "int64": torch.LongTensor, "int32": torch.IntTensor, "int16": torch.ShortTensor, "int8": torch.CharTensor, "uint8": torch.ByteTensor, } def wrapped(batch): "Puts each data field into a tensor with outer dimension batch size" error_msg = "batch must contain tensors, numbers, dicts or lists; found {}" elem_type = type(batch[0]) if torch.is_tensor(batch[0]): max_len = 0 for b in batch: max_len = max(max_len, b.size(0)) numel = sum([int(b.numel() / b.size(0) * max_len) for b in batch]) if use_shared_memory: # If we're in a background process, concatenate directly into a # shared memory tensor to avoid an extra copy storage = batch[0].storage()._new_shared(numel) out = batch[0].new(storage) else: out = batch[0].new(numel) out = out.view( len(batch), max_len, *[batch[0].size(i) for i in range(1, batch[0].dim())] ) out.fill_(pad_val) for i in range(len(batch)): out[i, 0 : batch[i].size(0)] = batch[i] return out elif ( elem_type.__module__ == "numpy" and elem_type.__name__ != "str_" and elem_type.__name__ != "string_" ): elem = batch[0] if elem_type.__name__ == "ndarray": # array of string classes and object if re.search("[SaUO]", elem.dtype.str) is not None: raise TypeError(error_msg.format(elem.dtype)) return wrapped([torch.from_numpy(b) for b in batch]) if elem.shape == (): # scalars py_type = float if elem.dtype.name.startswith("float") else int return _numpy_type_map[elem.dtype.name](list(map(py_type, batch))) elif isinstance(batch[0], int): return torch.LongTensor(batch) elif isinstance(batch[0], float): return torch.DoubleTensor(batch) elif isinstance(batch[0], collections.Mapping): return {key: wrapped([d[key] for d in batch]) for key in batch[0]} elif isinstance(batch[0], collections.Sequence): transposed = zip(*batch) return [wrapped(samples) for samples in transposed] raise TypeError((error_msg.format(type(batch[0])))) return wrapped class TrainValSplitter: r""" Creates a training and validation split to be used as the sampler in a pytorch DataLoader Parameters --------- numel : int Number of elements in the entire training dataset percent_train : float Percentage of data in the training split shuffled : bool Whether or not shuffle which data goes to which split """ def __init__(self, numel, percent_train, shuffled=False): # type: (TrainValSplitter, int, float, bool) -> None indicies = np.array([i for i in range(numel)]) if shuffled: np.random.shuffle(indicies) self.train = torch.utils.data.sampler.SubsetRandomSampler( indicies[0 : int(percent_train * numel)] ) self.val = torch.utils.data.sampler.SubsetRandomSampler( indicies[int(percent_train * numel) : -1] ) class CrossValSplitter: r""" Class that creates cross validation splits. The train and val splits can be used in pytorch DataLoaders. The splits can be updated by calling next(self) or using a loop: for _ in self: .... Parameters --------- numel : int Number of elements in the training set k_folds : int Number of folds shuffled : bool Whether or not to shuffle which data goes in which fold """ def __init__(self, numel, k_folds, shuffled=False): # type: (CrossValSplitter, int, int, bool) -> None inidicies = np.array([i for i in range(numel)]) if shuffled: np.random.shuffle(inidicies) self.folds = np.array(np.array_split(inidicies, k_folds), dtype=object) self.current_v_ind = -1 self.val = torch.utils.data.sampler.SubsetRandomSampler(self.folds[0]) self.train = torch.utils.data.sampler.SubsetRandomSampler( np.concatenate(self.folds[1:], axis=0) ) self.metrics = {} def __iter__(self): self.current_v_ind = -1 return self def __len__(self): return len(self.folds) def __getitem__(self, idx): assert idx >= 0 and idx < len(self) self.val.inidicies = self.folds[idx] self.train.inidicies = np.concatenate( self.folds[np.arange(len(self)) != idx], axis=0 ) def __next__(self): self.current_v_ind += 1 if self.current_v_ind >= len(self): raise StopIteration self[self.current_v_ind] def update_metrics(self, to_post): # type: (CrossValSplitter, dict) -> None for k, v in to_post.items(): if k in self.metrics: self.metrics[k].append(v) else: self.metrics[k] = [v] def print_metrics(self): for name, samples in self.metrics.items(): xbar = stats.mean(samples) sx = stats.stdev(samples, xbar) tstar = student_t.ppf(1.0 - 0.025, len(samples) - 1) margin_of_error = tstar * sx / sqrt(len(samples)) print("{}: {} +/- {}".format(name, xbar, margin_of_error)) def set_bn_momentum_default(bn_momentum): def fn(m): if isinstance(m, (nn.BatchNorm1d, nn.BatchNorm2d, nn.BatchNorm3d)): m.momentum = bn_momentum return fn class BNMomentumScheduler(object): def __init__(self, model, bn_lambda, last_epoch=-1, setter=set_bn_momentum_default): if not isinstance(model, nn.Module): raise RuntimeError( "Class '{}' is not a PyTorch nn Module".format(type(model).__name__) ) self.model = model self.setter = setter self.lmbd = bn_lambda self.step(last_epoch + 1) self.last_epoch = last_epoch def step(self, epoch=None): if epoch is None: epoch = self.last_epoch + 1 self.last_epoch = epoch self.model.apply(self.setter(self.lmbd(epoch))) class Trainer(object): r""" Reasonably generic trainer for pytorch models Parameters ---------- model : pytorch model Model to be trained model_fn : function (model, inputs, labels) -> preds, loss, accuracy optimizer : torch.optim Optimizer for model checkpoint_name : str Name of file to save checkpoints to best_name : str Name of file to save best model to lr_scheduler : torch.optim.lr_scheduler Learning rate scheduler. .step() will be called at the start of every epoch bnm_scheduler : BNMomentumScheduler Batchnorm momentum scheduler. .step() will be called at the start of every epoch eval_frequency : int How often to run an eval log_name : str Name of file to output tensorboard_logger to """ def __init__( self, model, model_fn, optimizer, checkpoint_name="ckpt", best_name="best", lr_scheduler=None, bnm_scheduler=None, eval_frequency=-1, viz=None, ): self.model, self.model_fn, self.optimizer, self.lr_scheduler, self.bnm_scheduler = ( model, model_fn, optimizer, lr_scheduler, bnm_scheduler, ) self.checkpoint_name, self.best_name = checkpoint_name, best_name self.eval_frequency = eval_frequency self.training_best, self.eval_best = {}, {} self.viz = viz @staticmethod def _decode_value(v): if isinstance(v[0], float): return np.mean(v) elif isinstance(v[0], tuple): if len(v[0]) == 3: num = [l[0] for l in v] denom = [l[1] for l in v] w = v[0][2] else: num = [l[0] for l in v] denom = [l[1] for l in v] w = None return np.average( np.sum(num, axis=0) / (np.sum(denom, axis=0) + 1e-6), weights=w ) else: raise AssertionError("Unknown type: {}".format(type(v))) def _train_it(self, it, batch): self.model.train() if self.lr_scheduler is not None: self.lr_scheduler.step(it) if self.bnm_scheduler is not None: self.bnm_scheduler.step(it) self.optimizer.zero_grad() _, loss, eval_res = self.model_fn(self.model, batch) loss.backward() self.optimizer.step() return eval_res def eval_epoch(self, d_loader): self.model.eval() eval_dict = {} total_loss = 0.0 count = 1.0 for i, data in tqdm.tqdm( enumerate(d_loader, 0), total=len(d_loader), leave=False, desc="val" ): self.optimizer.zero_grad() _, loss, eval_res = self.model_fn(self.model, data, eval=True) total_loss += loss.item() count += 1 for k, v in eval_res.items(): if v is not None: eval_dict[k] = eval_dict.get(k, []) + [v] return total_loss / count, eval_dict def train( self, start_it, start_epoch, n_epochs, train_loader, test_loader=None, best_loss=0.0, ): r""" Call to begin training the model Parameters ---------- start_epoch : int Epoch to start at n_epochs : int Number of epochs to train for test_loader : torch.utils.data.DataLoader DataLoader of the test_data train_loader : torch.utils.data.DataLoader DataLoader of training data best_loss : float Testing loss of the best model """ eval_frequency = ( self.eval_frequency if self.eval_frequency > 0 else len(train_loader) ) it = start_it with tqdm.trange(start_epoch, n_epochs + 1, desc="epochs") as tbar, tqdm.tqdm( total=eval_frequency, leave=False, desc="train" ) as pbar: for epoch in tbar: for batch in train_loader: res = self._train_it(it, batch) it += 1 pbar.update() pbar.set_postfix(dict(total_it=it)) tbar.refresh() if self.viz is not None: self.viz.update("train", it, res) if (it % eval_frequency) == 0: pbar.close() if test_loader is not None: val_loss, res = self.eval_epoch(test_loader) if self.viz is not None: self.viz.update("val", it, res) is_best = val_loss < best_loss best_loss = min(best_loss, val_loss) save_checkpoint( checkpoint_state( self.model, self.optimizer, val_loss, epoch, it ), is_best, filename=self.checkpoint_name, bestname=self.best_name, ) pbar = tqdm.tqdm( total=eval_frequency, leave=False, desc="train" ) pbar.set_postfix(dict(total_it=it)) self.viz.flush() return best_loss if False: # Workaround for type hints without depending on the `typing` module from typing import * class Seq(nn.Sequential): def __init__(self, input_channels): super(Seq, self).__init__() self.count = 0 self.current_channels = input_channels def conv1d( self, out_size, kernel_size=1, stride=1, padding=0, dilation=1, activation=nn.ReLU(inplace=True), bn=False, init=nn.init.kaiming_normal_, bias=True, preact=False, name="", norm_layer=BatchNorm1d, ): # type: (Seq, int, int, int, int, int, Any, bool, Any, bool, bool, AnyStr) -> Seq self.add_module( str(self.count), Conv1d( self.current_channels, out_size, kernel_size=kernel_size, stride=stride, padding=padding, dilation=dilation, activation=activation, bn=bn, init=init, bias=bias, preact=preact, name=name, norm_layer=norm_layer, ), ) self.count += 1 self.current_channels = out_size return self def conv2d( self, out_size, kernel_size=(1, 1), stride=(1, 1), padding=(0, 0), dilation=(1, 1), activation=nn.ReLU(inplace=True), bn=False, init=nn.init.kaiming_normal_, bias=True, preact=False, name="", norm_layer=BatchNorm2d, ): # type: (Seq, int, Tuple[int, int], Tuple[int, int], Tuple[int, int], Tuple[int, int], Any, bool, Any, bool, bool, AnyStr) -> Seq self.add_module( str(self.count), Conv2d( self.current_channels, out_size, kernel_size=kernel_size, stride=stride, padding=padding, dilation=dilation, activation=activation, bn=bn, init=init, bias=bias, preact=preact, name=name, norm_layer=norm_layer, ), ) self.count += 1 self.current_channels = out_size return self def conv3d( self, out_size, kernel_size=(1, 1, 1), stride=(1, 1, 1), padding=(0, 0, 0), dilation=(1, 1, 1), activation=nn.ReLU(inplace=True), bn=False, init=nn.init.kaiming_normal_, bias=True, preact=False, name="", norm_layer=BatchNorm3d, ): # type: (Seq, int, Tuple[int, int], Tuple[int, int, int], Tuple[int, int, int], Tuple[int, int, int], Any, bool, Any, bool, bool, AnyStr) -> Seq self.add_module( str(self.count), Conv3d( self.current_channels, out_size, kernel_size=kernel_size, stride=stride, padding=padding, dilation=dilation, activation=activation, bn=bn, init=init, bias=bias, preact=preact, name=name, norm_layer=norm_layer, ), ) self.count += 1 self.current_channels = out_size return self def fc( self, out_size, activation=nn.ReLU(inplace=True), bn=False, init=None, preact=False, name="", ): # type: (Seq, int, Any, bool, Any, bool, AnyStr) -> None self.add_module( str(self.count), FC( self.current_channels, out_size, activation=activation, bn=bn, init=init, preact=preact, name=name, ), ) self.count += 1 self.current_channels = out_size return self def dropout(self, p=0.5): # type: (Seq, float) -> Seq self.add_module(str(self.count), nn.Dropout(p=0.5)) self.count += 1 return self def maxpool2d( self, kernel_size, stride=None, padding=0, dilation=1, return_indices=False, ceil_mode=False, ): self.add_module( str(self.count), nn.MaxPool2d( kernel_size=kernel_size, stride=stride, padding=padding, dilation=dilation, return_indices=return_indices, ceil_mode=ceil_mode, ), ) self.count += 1 return self
28,499
27.904665
175
py
ShapeMOD
ShapeMOD-main/SA_lang/tasks/ShapeAssembly.py
# Taken from https://github.com/rkjones4/ShapeAssembly import torch import re import numpy as np import math import ast import sys import faiss from copy import deepcopy """ This file contains all of the logic in the ShapeAssembly DSL. You can execute a ShapeAssembly program as follows: > from ShapeAssembly import ShapeAssembly > sa = ShapeAssembly() > lines = sa.load_lines({path_to_program}) > sa.run(lines, {out_file_name}) The classes in this file are: Cuboid -> Part Proxies represented as Cuboids in space AttPoint -> Points that live within the local coordinate frames of cuboids -> specify where cuboids should attach Program -> Parses lines, locally executes lines by creating Cuboids, AttPoints, and changing their attributes. ShapeAssembly -> Entrypoint to language logic """ # Params controlling execution behavior EPS = .01 SMALL_EPS = 1e-4 COS_DIST_THRESH = 0.9 # Helper function: write mesh to out file def writeObj(verts, faces, outfile): with open(outfile, 'w') as f: for a, b, c in verts.tolist(): f.write(f'v {a} {b} {c}\n') for a, b, c in faces.tolist(): f.write(f"f {a+1} {b+1} {c+1}\n") def samplePC(cubes, xyz1, xyz2): cube_geom = [] for c in cubes: cube_geom.append(torch.cat(( c['xd'].unsqueeze(0), c['yd'].unsqueeze(0), c['zd'].unsqueeze(0), c['center'], c['xdir'], c['ydir'] ))) scene_geom = torch.stack([c for c in cube_geom]) ind_to_pc = {} xyzs = [xyz1, xyz2] for i in range(0, scene_geom.shape[0]): xyz = xyzs[i] s_inds = (torch.ones(1,xyz.shape[1]) * i).long() s_r = torch.cat( ( (scene_geom[s_inds][:, :, 6:9] / (scene_geom[s_inds][:, :, 6:9].norm(dim=2).unsqueeze(2) + 1e-8)).unsqueeze(3), (scene_geom[s_inds][:, :, 9:12] / (scene_geom[s_inds][:, :, 9:12].norm(dim=2).unsqueeze(2) + 1e-8)).unsqueeze(3), torch.cross( scene_geom[s_inds][:, :, 6:9] / (scene_geom[s_inds][:, :, 6:9].norm(dim=2).unsqueeze(2) + 1e-8), scene_geom[s_inds][:, :, 9:12] / (scene_geom[s_inds][:, :, 9:12].norm(dim=2).unsqueeze(2) + 1e-8) ).unsqueeze(3) ), dim = 3) s_out = ((s_r @ (((xyz - .5) * scene_geom[s_inds][:, :, :3]).unsqueeze(-1))).squeeze() + scene_geom[s_inds][:, :, 3:6]).squeeze() ind_to_pc[i] = s_out res = {} for key in ind_to_pc: index = faiss.IndexFlatL2(3) index.add( np.ascontiguousarray(ind_to_pc[key].cpu().numpy()) ) res[key] = (ind_to_pc[key], index) return res # Helper function: given angle + normal compute a rotation matrix that will accomplish the operation def getRotMatrix(angle, normal): s = torch.sin(angle) c = torch.cos(angle) nx = normal[0] ny = normal[1] nz = normal[2] rotmat = torch.stack(( torch.stack((c + (1 - c) * nx * nx, (1 - c) * nx * ny - s * nz, (1 - c) * nx * nz + s * ny)), torch.stack(((1 - c) * nx * ny + s * nz, c + (1 - c) * ny * ny, (1 - c) * ny * nz - s * nx)), torch.stack(((1 - c) * nx * nz - s * ny, (1 - c) * ny * nz + s * nx, c + (1 - c) * nz * nz)) )) return rotmat # Helper function: Find a minimum rotation from the current direction to the target direction def findMinRotation(cur, target): assert(cur.norm() != 0) assert(target.norm() != 0) ncur = cur / cur.norm() ntarget = target / target.norm() normal = torch.cross(ncur, ntarget) # co-linear if normal.norm() == 0: r_x = torch.tensor([[1.0, 0.0, 0.0], [0.0, 0.0, -1.0], [0, 1.0, 0.0]]) r_y = torch.tensor([[0.0, 0, 1.0], [0.0, 1.0, 0.0], [ -1.0, 0.0, 0.0]]) r_z = torch.tensor([[0.0, -1.0, 0.0], [1.0, 0.0, 0.0], [0.0, 0.0, 1.0]]) if torch.dot(r_x @ ncur, ncur) != 0: cnormal = r_x @ ncur elif torch.dot(r_y @ ncur, ncur) != 0: cnormal = r_y @ cur elif torch.dot(r_z @ ncur, ncur) != 0: cnormal = r_z @ ncur assert(cnormal.norm() != 0) nnormal = cnormal / cnormal.norm() angle = torch.tensor(math.pi) else: nnormal = normal / normal.norm() angle = torch.acos(torch.dot(ncur, ntarget)) if angle == 0 or torch.isnan(angle).any(): return torch.eye(3) return getRotMatrix(angle, nnormal) class Cuboid(): """ Cuboids are the base (and only) objects of a ShapeAssembly program. Dims are their dimensions, pos is the center of the cuboid, rfnorm (right face), tfnorm (top face) and ffnorm (front face) specify the orientation of the cuboid. The bounding volume is just a non-visible cuboid. Cuboids marked with the aligned flag behavior differently under attachment operations. """ def __init__(self, name, aligned = False, vis = True): # The default cube is unit, axis-aligned, centered at the origin self.dims = torch.tensor([1.0,1.0,1.0]) self.pos = torch.tensor([0.0,0.0,0.0]) self.rfnorm = torch.tensor([1.0,0.0,0.0]) self.tfnorm = torch.tensor([0.0,1.0,0.0]) self.ffnorm = torch.tensor([0.0,0.0,1.0]) # Keep track of all attachment obligations this cube has self.attachments = [] self.move_atts = [] # The bbox is not visible, but is still a cuboid, otherwise this should be True self.is_visible = vis self.name = name self.parent = None self.parent_axis = None self.aligned = aligned # Rotate the cuboid by the rotation matrix def rotateCuboid(self, rotation): self.rfnorm = rotation @ self.rfnorm self.tfnorm = rotation @ self.tfnorm self.ffnorm = rotation @ self.ffnorm def flipCuboid(self, a_ind): transform = torch.ones(3) transform[a_ind] *= -1 self.pos = transform * self.pos self.rfnorm = -1 * (transform * self.rfnorm) self.tfnorm = -1 * (transform * self.tfnorm) self.ffnorm = -1 * (transform * self.ffnorm) # Get the corners of the cuboid def getCorners(self): xd = self.dims[0] / 2 yd = self.dims[1] / 2 zd = self.dims[2] / 2 corners = torch.stack(( (self.rfnorm * xd) + (self.tfnorm * yd) + (self.ffnorm * zd), (self.rfnorm * xd) + (self.tfnorm * yd) + (self.ffnorm * -1 * zd), (self.rfnorm * xd) + (self.tfnorm * -1 * yd) + (self.ffnorm * zd), (self.rfnorm * xd) + (self.tfnorm * -1 * yd) + (self.ffnorm * -1 * zd), (self.rfnorm * -1 * xd) + (self.tfnorm * yd) + (self.ffnorm * zd), (self.rfnorm * -1 * xd) + (self.tfnorm * yd) + (self.ffnorm * -1 * zd), (self.rfnorm * -1 * xd) + (self.tfnorm * -1 * yd) + (self.ffnorm * zd), (self.rfnorm * -1 * xd) + (self.tfnorm * -1 * yd) + (self.ffnorm * -1 * zd), )) return corners + self.pos # Get the global point specified by relative coordinates x,y,z def getPos(self, x, y, z): pt = torch.stack((x, y, z)) r = torch.stack(( self.rfnorm, self.tfnorm, self.ffnorm )).T t_dims = torch.stack((self.dims[0], self.dims[1], self.dims[2])) return (r @ ((pt - .5) * t_dims)) + self.pos # Get the relative position of global poiunt gpt def getRelPos(self, gpt, normalize = False): O = self.getPos( torch.tensor(0.), torch.tensor(0.), torch.tensor(0.) ) A = torch.stack([ self.dims[0].clone() * self.rfnorm.clone(), self.dims[1].clone() * self.tfnorm.clone(), self.dims[2].clone() * self.ffnorm.clone() ]).T B = gpt - O p = A.inverse() @ B if normalize: return torch.clamp(p, 0.0, 1.0) return p # Make the cuboid bigger by a multiplied factor of scale (either dim 3 or dim 1) def scaleCuboid(self, scale): self.dims *= scale # Make the cuboid bigger by an added factor of scale to a specific dimension def increaseDim(self, dim, inc): dim_to_scale = { "height": torch.tensor([0.0, 1.0, 0.0]), "width": torch.tensor([0.0, 0.0, 1.0]), "length": torch.tensor([1.0, 0.0, 0.0]) } s = dim_to_scale[dim] * inc self.dims += s # Move the center of the cuboid by the translation vector def translateCuboid(self, translation): self.pos += translation # Used to convert cuboid into triangles on its faces def getTriFaces(self): return [ [0, 2, 1], [1, 2, 3], [0, 4, 6], [0, 6, 2], [1, 3, 5], [3, 7, 5], [4, 5, 7], [4, 7, 6], [1, 5, 4], [0, 1, 4], [2, 6, 7], [2, 7, 3] ] # Get the triangulation of the cuboid corners, for visualization + sampling def getTris(self): if self.is_visible: verts = self.getCorners() faces = torch.tensor(self.getTriFaces(), dtype=torch.long) return verts, faces return None, None # Return any attachments that are on this cuboid def getAttachments(self): return self.attachments # Return the cuboid's parameterization def getParams(self): return torch.cat(( self.dims, self.pos, self.rfnorm, self.tfnorm, self.ffnorm )) class AttPoint(): """ Attachment Points live with the local coordinate frame [0, 1]^3 of a cuboid. They are used to connect cuboids together. """ def __init__(self, cuboid, x, y, z): self.cuboid = cuboid self.x = x self.y = y self.z = z # To get the global position, all we need is the cuboid+face info, and the relative uv pos def getPos(self): return self.cuboid.getPos(self.x, self.y, self.z) # If we scale the height of the cuboid, what is the rate of change of this AP def getChangeVectorHeight(self): norm = self.cuboid.tfnorm return (self.y - .5) * norm # If we scale the length of the cuboid, what is the rate of change of this AP def getChangeVectorLength(self): norm = self.cuboid.rfnorm return (self.x - .5) * norm # If we scale the width of the cuboid, what is the rate of change of this AP def getChangeVectorWidth(self): norm = self.cuboid.ffnorm return (self.z - .5) * norm # get rate of change of this AP when we change the specified dimension def getChangeVector(self, dim): dim_to_sf = { 'height': self.getChangeVectorHeight, 'length': self.getChangeVectorLength, 'width': self.getChangeVectorWidth, } return dim_to_sf[dim]() # If we scale the height of the cuboid, what direction does the AP move with def getChangeDirHeight(self): if self.y > .5: return 'top' elif self.y < .5: return 'bot' else: return 'none' # If we scale the length of the cuboid, what direction does the AP move with def getChangeDirLength(self): if self.x > .5: return 'right' elif self.x < .5: return 'left' else: return 'none' # If we scale the width of the cuboid, what direction does the AP move with def getChangeDirWidth(self): if self.z > .5: return 'front' elif self.z < .5: return 'back' else: return 'none' def getChangeDir(self, dim): dim_to_dir = { 'height': self.getChangeDirHeight, 'length': self.getChangeDirLength, 'width': self.getChangeDirWidth, } return dim_to_dir[dim]() class Program(): """ A program maintains a representation of entire shape, including all of the member cuboids and all of the attachment points. The execute function is the entrypoint of text programs. """ def __init__(self, cuboids = {}): self.cuboids = self.getBoundBox() self.cuboids.update(cuboids) self.commands = [] self.parameters = [] self.att_points = {} self.resource = None # CONSTANTS For PC Intersection DIM = 10 a = (torch.arange(DIM).float()/(DIM-1)) b = a.unsqueeze(0).unsqueeze(0).repeat(DIM, DIM, 1) c = a.unsqueeze(0).unsqueeze(2).repeat(DIM, 1, DIM) d = a.unsqueeze(1).unsqueeze(2).repeat(1, DIM, DIM) g = torch.stack((b,c,d), dim=3).view(-1, 3) self.s_xyz = g.unsqueeze(0) bb_top_inds = (g[:,1] == 1.0).nonzero().squeeze() bb_bot_inds = (g[:,1] == 0.0).nonzero().squeeze() self.bb_top_xyz = g[bb_top_inds].unsqueeze(0) self.bb_bot_xyz = g[bb_bot_inds].unsqueeze(0) def flip(self, flip_axis): if flip_axis == 'X': axis = 0 elif flip_axis == 'Y': axis = 1 elif flip_axis == 'Z': axis = 2 for name, c in self.cuboids.items(): if name == 'bbox': continue c.flipCuboid(axis) # Each program starts off with an invisible bounding box def getBoundBox(self): bbox = Cuboid("bbox", aligned = True, vis=False) return { "bbox": bbox } # Get the triangles in the current scene -> first index is bounding box so skipped def getShapeGeo(self): if len(self.cuboids) < 2: return None, None cuboids = list(self.cuboids.values()) verts = torch.tensor([],dtype=torch.float) faces = torch.tensor([],dtype=torch.long) for cube in cuboids[1:]: v, f = cube.getTris() if v is not None and f is not None: faces = torch.cat((faces, (f + verts.shape[0]))) verts = torch.cat((verts, v)) return verts, faces # Make an obj of the current scene def render(self, ofile = "output.obj"): verts, faces = self.getShapeGeo() writeObj(verts, faces, ofile) # Parses a cuboid text line def parseCuboid(self, line): s = re.split(r'[()]', line) name = s[0].split("=")[0].strip() dim0 = None dim1 = None dim2 = None aligned = False params = s[1].split(',') dim0 = torch.tensor(float(params[0])) dim1 = torch.tensor(float(params[1])) dim2 = torch.tensor(float(params[2])) if len(params) == 4: aligned = ast.literal_eval(params[3].strip()) assert isinstance(aligned, bool), 'aligned not a bool' return (name, dim0, dim1, dim2, aligned) # Construct a new cuboid, add it to state def executeCuboid(self, parse): name = parse[0] if name in self.cuboids: c = self.cuboids[name] c.dims = torch.stack((parse[1], parse[2], parse[3])) else: c = Cuboid( parse[0], aligned = parse[4], ) c.scaleCuboid(torch.stack((parse[1], parse[2], parse[3]))) self.cuboids.update({ parse[0]: c }) # Logic for cuboids with no previous attachment. Finds a translation to satisfy the attachment def first_attach(self, ap, gpos): cur_pos = ap.getPos() diff = gpos - cur_pos ap.cuboid.translateCuboid(diff) return True # Logic for unaligned cuboids with one previous attachment. Find a scale and rotation to satisfy the attachment def second_attach(self, ap, gpos, prev_att): p_ap = prev_att[0] p_gpos = prev_att[1] a = p_gpos b = ap.getPos() c = gpos if (b-c).norm() < SMALL_EPS: return True # Increase dimension to fix distance dist = (c-a).norm() min_dim = 'height' min_sf = 1e8 for dim in ('height', 'width', 'length'): nsf = ap.getChangeVector(dim) psf = p_ap.getChangeVector(dim) if nsf.abs().sum() + psf.abs().sum() < SMALL_EPS: continue cn = b - a dn = nsf - psf at = (dn**2).sum() bt = 2 * (cn*dn).sum() ct = (cn**2).sum() - (dist**2) # Take the positive solution of the quadratic equation sf = ((-1 * bt) + (bt**2 - (4*at*ct) ).sqrt()) / ((2 * at) + 1e-8) if abs(sf) < abs(min_sf) and (bt**2 - (4*at*ct)) > 0: min_sf = sf min_dim = dim if min_sf == 1e8: nsf = ap.getChangeVector('height') + \ ap.getChangeVector('length') + \ ap.getChangeVector('width') psf = p_ap.getChangeVector('height') + \ p_ap.getChangeVector('length') + \ p_ap.getChangeVector('width') cn = b - a dn = nsf - psf at = (dn**2).sum() bt = 2 * (cn*dn).sum() ct = (cn**2).sum() - (dist**2) # Take the positive solution of the quadratic equation sf = ((-1 * bt) + (bt**2 - (4*at*ct) ).sqrt()) / ((2 * at) + 1e-8) if not torch.isnan(sf) and (bt**2 - (4*at*ct)) > 0: ap.cuboid.increaseDim('height', sf) ap.cuboid.increaseDim('length', sf) ap.cuboid.increaseDim('width', sf) else: ap.cuboid.increaseDim(min_dim, min_sf) # Reset the position of the cuboid such that the previous attachment is satisfied diff = p_gpos - p_ap.getPos() ap.cuboid.translateCuboid(diff) # find rotation to match points nb = ap.getPos() - p_gpos nc = c - p_gpos # If we are already in the correct position, don't rotate if nb.norm() == 0 or nc.norm() == 0 or (nb-nc).norm() < SMALL_EPS: return True rot_mat = findMinRotation(nb, nc) ap.cuboid.rotateCuboid(rot_mat) # Reset the position of the cuboid such that the attachments are satisfied sdiff = p_gpos - p_ap.getPos() ap.cuboid.translateCuboid(sdiff) return True # Returns if all of the previous attachment points on cuboid are co-linear def checkColinear(self, prev_atts): if len(prev_atts) == 2: return True else: o = prev_atts[0][1] l = prev_atts[1][1] - o for i in range(2, len(prev_atts)): n = prev_atts[i][1] - o if torch.cross(n, l).norm() > EPS: return False return True # For cuboids with 2+ attachments, checks if changing orientation would improve the fit def doOrient(self, ap, gpos, prev_atts): p_ap0 = prev_atts[0][0] p_gpos0 = prev_atts[0][1] p_ap1 = prev_atts[1][0] p_gpos1 = prev_atts[1][1] a0 = p_gpos0 a1 = p_gpos1 b = ap.getPos() c = gpos axis = (a0 - a1) / ((a0 - a1).norm() + 1e-8) # project c onto the plane defined by b and the axis of rotation tc = (axis.dot(b) - axis.dot(c)) / (axis.dot(axis) + 1e-8) c_p = c + (tc * axis) # project a0 onto the plane defined by b and the axis of rotation, this is the origin of rotation ta = (axis.dot(b) - axis.dot(a0)) / (axis.dot(axis) + 1e-8) a_p = a0 + (ta * axis) # project center of the cube onto the plane defined by b and the axis of rotation center = ap.cuboid.pos to = (axis.dot(b) - axis.dot(center)) / (axis.dot(axis) + 1e-8) o_p = center + (to * axis) fn = (b - a_p) noc = c_p - a_p at = fn.dot(fn) bt = 2 * (b-a_p).dot(fn) ct = (b-a_p).dot((b-a_p)) - noc.dot(noc) sf = ((-1 * bt) + (bt**2 - (4*at*ct) ).sqrt()) / ((2 * at) + 1e-8) if torch.isnan(sf): return False b_p = b + (sf * fn) nob = b_p - a_p # If we are already in the correct position, don't rotate if nob.norm() == 0 or noc.norm() == 0 or (nob-noc).norm() < EPS: return True nb = nob / (nob.norm() + 1e-8) nc = noc / (noc.norm() + 1e-8) angle = torch.acos(torch.dot(nb, nc)) rot_mat = getRotMatrix(angle, axis) # make sure axis is oriented correctly if ((rot_mat @ nb) - nc).norm() > EPS: rot_mat = getRotMatrix(angle, -1*axis) ap.cuboid.rotateCuboid(rot_mat) diff = a0 - p_ap0.getPos() ap.cuboid.translateCuboid(diff) # For cuboids with 2+ attachments, checks if scaling any dimension would improve the fit def maybeScale(self, ap, gpos, prev_atts): b = ap.getPos() c = gpos t_dir = c - b if t_dir.norm() < EPS: return # Can only scale a dimension if all of the CV for that dimension are the same eq_cv_dims = [] for dim in ('height', 'width', 'length'): cv = prev_atts[0][0].getChangeVector(dim) add = True for p_ap in prev_atts: if (cv - p_ap[0].getChangeVector(dim)).norm() > EPS: add = False if add: eq_cv_dims.append(dim) dirs_to_avoid = set() best_dim = None cos_dist = COS_DIST_THRESH nt_dir = t_dir / (t_dir.norm() + 1e-8) # See if scaling any direction would improve the fit (i.e. has a valid solution to quadratic equation) for dim in eq_cv_dims: sf = ap.getChangeVector(dim) dr = ap.getChangeDir(dim) if dr in dirs_to_avoid: continue nsf = sf / (sf.norm() + 1e-8) if torch.isnan(nsf).any().tolist(): continue if nt_dir.dot(nsf) > cos_dist: best_dim = dim cost_dist = nt_dir.dot(nsf) if best_dim == None: return False axis = ap.getChangeVector(best_dim) paxis = prev_atts[0][0].getChangeVector(best_dim) d = 1.0 / ((axis-paxis).norm() + 1e-8) naxis = axis / (axis.norm() + 1e-8) t = (naxis.dot(b) - naxis.dot(c)) / (naxis.dot(naxis) + 1e-8) c_p = c + (t * naxis) offset = c_p - b c_g = c - offset sf = (c_g - b).norm() ap.cuboid.increaseDim(best_dim, sf*d) diff = prev_atts[0][1] - prev_atts[0][0].getPos() ap.cuboid.translateCuboid(diff) # Given cuboids a and b, find the closest pair of points in their local coordinate frames to one another def getClosestPoints(self, a, b, xyz1, xyz2, is_bbox): if not is_bbox: a_corners = a.getCorners() for i in range(a_corners.shape[0]): rp = b.getRelPos(a_corners[i]) if rp.min() >= 0.0 and rp.max() <= 1.0: return None, None b_corners = b.getCorners() for i in range(b_corners.shape[0]): rp = a.getRelPos(b_corners[i]) if rp.min() >= 0.0 and rp.max() <= 1.0: return None, None ind_to_pc = samplePC( [ { 'xd': a.dims[0].detach(), 'yd': a.dims[1].detach(), 'zd': a.dims[2].detach(), 'center': a.pos.detach(), 'xdir': a.rfnorm.detach(), 'ydir': a.tfnorm.detach(), }, { 'xd': b.dims[0].detach(), 'yd': b.dims[1].detach(), 'zd': b.dims[2].detach(), 'center': b.pos.detach(), 'xdir': b.rfnorm.detach(), 'ydir': b.tfnorm.detach() } ], xyz1, xyz2 ) a_pc = ind_to_pc[0][0].unsqueeze(0) b_pc = ind_to_pc[1][0].unsqueeze(0) b_index = ind_to_pc[1][1] a_query = np.ascontiguousarray( a_pc.data.clone().detach().cpu().numpy() ) D, I = b_index.search(a_query[0], 1) I_var = torch.from_numpy(np.ascontiguousarray(I).astype(np.int64)).squeeze() D_var = torch.from_numpy(np.ascontiguousarray(D).astype(np.float)).squeeze() a_ind = D_var.argmin() b_ind = I_var[a_ind] a_pt = a_pc[0, a_ind] b_pt = b_pc[0, b_ind] return a_pt.cpu(), b_pt.cpu() # For aligned cuboids with a previous attachments, # see if increasing any dimension would cause the fit to be improved def aligned_attach(self, ap, oap): if oap.cuboid.name != 'bbox': xyz2 = self.s_xyz else: if oap.y > 0.5: xyz2 = self.bb_top_xyz else: xyz2 = self.bb_bot_xyz b, c = self.getClosestPoints( ap.cuboid, oap.cuboid, self.s_xyz, xyz2, oap.cuboid.name == 'bbox' ) if b is None or c is None: return mv = max( ap.cuboid.dims.max().item(), oap.cuboid.dims.max().item() ) # Do a sampling of 20 in inter if (b-c).norm() < mv * .05: return rp = ap.cuboid.getRelPos(b, True) n_ap = AttPoint(ap.cuboid, rp[0], rp[1], rp[2]) o_ap = AttPoint(ap.cuboid, 1-rp[0], 1-rp[1], 1-rp[2]) ppos = o_ap.getPos() D_l = 0 D_h = 0 D_w = 0 l_cv = n_ap.getChangeVector('length')[0] if l_cv != 0: D_l = (c[0] - b[0]) / (2*l_cv) h_cv = n_ap.getChangeVector('height')[1] if h_cv != 0: D_h = (c[1] - b[1]) / (2*h_cv) w_cv = n_ap.getChangeVector('width')[2] if w_cv != 0: D_w = (c[2] - b[2]) / (2*w_cv) if D_l > 0: ap.cuboid.increaseDim("length", D_l) if D_h > 0: ap.cuboid.increaseDim("height", D_h) if D_w > 0: ap.cuboid.increaseDim("width", D_w) diff = ppos - o_ap.getPos() ap.cuboid.translateCuboid(diff) # Attachment for a non-aligned cuboid with 2+ previous attachments. Only try to orient the cuboid if its previous attachment are # all colinear. def gen_attach(self, ap, gpos, prev_atts): if self.checkColinear(prev_atts): self.doOrient(ap, gpos, prev_atts) self.maybeScale(ap, gpos, prev_atts) # Moves the attach point to the global position def attach(self, ap, gpos, oci, oap=None): assert ap.cuboid.name != "bbox", 'tried to move the bbox' if ap.cuboid.aligned: self.aligned_cube_attach(ap, gpos, oci, oap) else: self.free_cube_attach(ap, gpos, oci) # Aligned attachment def aligned_cube_attach(self, ap, gpos, oci, oap): prev_atts = ap.cuboid.getAttachments() if len(prev_atts) == 0: self.first_attach(ap, gpos) else: self.aligned_attach(ap, oap) prev_atts.append((ap, gpos, oci)) ap.cuboid.move_atts.append((ap, gpos, oci)) # Non-aligned attachment def free_cube_attach(self, ap, gpos, oci): prev_atts = ap.cuboid.getAttachments() if len(prev_atts) == 0: self.first_attach(ap, gpos) elif len(prev_atts) == 1: self.second_attach(ap, gpos, prev_atts[0]) else: self.gen_attach(ap, gpos, prev_atts) prev_atts.append((ap, gpos, oci)) ap.cuboid.move_atts.append((ap, gpos, oci)) # Parses an attach line def parseAttach(self, line): s = re.split(r'[()]', line) args = [a.strip() for a in s[1].split(',')] return ( args[0], args[1], torch.tensor(float(args[2])), torch.tensor(float(args[3])), torch.tensor(float(args[4])), torch.tensor(float(args[5])), torch.tensor(float(args[6])), torch.tensor(float(args[7])) ) # Execute an attach line, creates two attachment points, then figures out how to best satisfy new constraint def executeAttach(self, parse): ap1 = AttPoint( self.cuboids[parse[0]], parse[2], parse[3], parse[4], ) ap2 = AttPoint( self.cuboids[parse[1]], parse[5], parse[6], parse[7], ) ap_pt_name = f'{parse[0]}_to_{parse[1]}' # Attach points should have unique names while ap_pt_name in self.att_points: ap_pt_name += '_n' self.att_points[ap_pt_name] = ap2 ap2.cuboid.getAttachments().append((ap2, ap2.getPos(), ap1.cuboid.name)) self.attach(ap1, ap2.getPos(), ap2.cuboid.name, ap2) # Parses a reflect command def parseReflect(self, line): s = re.split(r'[()]', line) args = [a.strip() for a in s[1].split(',')] return ( args[0], args[1], ) # Parses a translate command def parseTranslate(self, line): s = re.split(r'[()]', line) args = [a.strip() for a in s[1].split(',')] return ( args[0], args[1], int(args[2]), float(args[3]) ) # Parses a queeze command def parseSqueeze(self, line): s = re.split(r'[()]', line) args = [a.strip() for a in s[1].split(',')] return ( args[0], args[1], args[2], args[3], float(args[4]), float(args[5]) ) # Help function for getting direction of reflect commands def getRefDir(self, d): bbox = self.cuboids['bbox'] if d == 'X': return bbox.rfnorm.clone() elif d == 'Y': return bbox.tfnorm.clone() elif d == 'Z': return bbox.ffnorm.clone() else: assert False, 'bad reflect argument' # Help function for getting direction + scale of translate commands def getTransDir(self, d): bbox = self.cuboids['bbox'] if d == 'X': return bbox.rfnorm.clone(), bbox.dims[0].clone() elif d == 'Y': return bbox.tfnorm.clone(), bbox.dims[1].clone() elif d == 'Z': return bbox.ffnorm.clone(), bbox.dims[2].clone() else: assert False, 'bad reflect argument' # Given an axis + a center, consructs a tranformation matrix to satisfy reflection def getRefMatrixHomo(self, axis, center): m = center d = axis / axis.norm() refmat = torch.stack(( torch.stack((1 - 2 * d[0] * d[0], -2 * d[0] * d[1], -2 * d[0] * d[2], 2 * d[0] * d[0] * m[0] + 2 * d[0] * d[1] * m[1] + 2 * d[0] * d[2] * m[2])), torch.stack((-2 * d[1] * d[0], 1 - 2 * d[1] * d[1], -2 * d[1] * d[2], 2 * d[1] * d[0] * m[0] + 2 * d[1] * d[1] * m[1] + 2 * d[1] * d[2] * m[2])), torch.stack((-2 * d[2] * d[0], -2 * d[2] * d[1], 1 - 2 * d[2] * d[2], 2 * d[2] * d[0] * m[0] + 2 * d[2] * d[1] * m[1] + 2 * d[2] * d[2] * m[2])) )) return refmat # Reflect a point p, about center and a direction ndir def reflect_point(self, p, center, ndir): pad = torch.nn.ConstantPad1d((0, 1), 1.0) reflection = self.getRefMatrixHomo(ndir, center) posHomo = pad(p) return reflection @ posHomo # Given angle, axis and center constructs a transformation matrix to satisfy rotation def getRotMatrixHomo(self, angle, axis, center): s = torch.sin(angle) c = torch.cos(angle) naxis = axis / axis.norm() nx = naxis[0] ny = naxis[1] nz = naxis[2] tmpmat = torch.stack(( torch.stack((c + (1 - c) * nx * nx, (1 - c) * nx * ny - s * nz, (1 - c) * nx * nz + s * ny)), torch.stack(((1 - c) * nx * ny + s * nz, c + (1 - c) * ny * ny, (1 - c) * ny * nz - s * nx)), torch.stack(((1 - c) * nx * nz - s * ny, (1 - c) * ny * nz + s * nx, c + (1 - c) * nz * nz)) )) rotmat = torch.cat((tmpmat, torch.zeros(3,1)), dim = 1) lc = (torch.matmul(torch.eye(3) - tmpmat, center)).unsqueeze(1) return torch.cat((rotmat[:,:3], lc), dim = 1) # Executes a reflect line by making + executing new Cuboid and attach lines def executeReflect(self, parse): c = self.cuboids[parse[0]] assert c.name != "bbox", 'tried to move the bbox' rdir = self.getRefDir(parse[1]) name = c.name + '_ref' self.executeCuboid([f'{name}', c.dims[0].clone(), c.dims[1].clone(), c.dims[2].clone(), c.aligned]) self.cuboids[f'{name}'].parent = c.name self.cuboids[f'{name}'].parent_axis = parse[1] atts = c.move_atts for att in atts: if parse[1] == 'X': x = 1 - att[0].x.clone() else: x = att[0].x.clone() if parse[1] == 'Y': y = 1 - att[0].y.clone() else: y = att[0].y.clone() if parse[1] == 'Z': z = 1 - att[0].z.clone() else: z = att[0].z.clone() n = att[2] cpt = att[0].getPos().clone() rpt = self.reflect_point(cpt, self.cuboids['bbox'].pos.clone(), rdir) rrpt = self.cuboids[n].getRelPos(rpt, True) self.executeAttach([f'{name}', f'{n}', x, y, z, rrpt[0], rrpt[1], rrpt[2]]) # Executes a translate line by making + executing new Cuboid and attach lines def executeTranslate(self, parse): c = self.cuboids[parse[0]] assert c.name != "bbox", 'tried to move the bbox' tdir, td = self.getTransDir(parse[1]) N = parse[2] scale = (td * parse[3]) / float(N) for i in range(1, N+1): name = c.name + f'_trans_{i}' self.executeCuboid([f'{name}', c.dims[0].clone(), c.dims[1].clone(), c.dims[2].clone(), c.aligned]) self.cuboids[f'{name}'].parent = c.name atts = c.move_atts for att in atts: x = att[0].x y = att[0].y z = att[0].z n = att[2] cpt = att[0].getPos() rpt = cpt + (tdir * scale * i) rrpt = self.cuboids[n].getRelPos(rpt, True) self.executeAttach([f'{name}', f'{n}', x, y, z, rrpt[0], rrpt[1], rrpt[2]]) # Helper function for finding opposite face def getOppFace(self, face): of = { 'right': 'left', 'left': 'right', 'top': 'bot', 'bot': 'top', 'front': 'back', 'back': 'front', } return of[face] # Local coordinate frame to center of face conversion def getFacePos(self, face): ft = { 'right': ([1.0, 0.5, 0.5], 0, 0.), 'left': ([0.0, 0.5, 0.5], 0, 1.), 'top': ([.5, 1.0, 0.5], 1, 0.), 'bot': ([.5, 0.0, 0.5], 1, 1.), 'front': ([.5, 0.5, 1.0], 2, 0.), 'back': ([.5, 0.5, 0.0], 2, 1.), } return ft[face] # Converts squeeze parameters into parameters needed for the two attachment operators. def getSqueezeAtt(self, face, u, v, is_bbox): at1, ind, val = self.getFacePos(face) # bbox is "flipped" if is_bbox: rval = 1-val else: rval = val at2 = torch.zeros(3).float() q = [u, v] for i in range(3): if i == ind: at2[i] = rval else: at2[i] = q.pop(0) return torch.tensor(at1).float(), at2 # Executes a squeeze line by making + executing new Cuboid and attach lines def executeSqueeze(self, parse): face = parse[3] oface = self.getOppFace(face) atc1, ato1 = self.getSqueezeAtt( face, parse[4], parse[5], parse[1] == 'bbox' ) atc2, ato2 = self.getSqueezeAtt( oface, parse[4], parse[5], parse[2] == 'bbox' ) self.executeAttach([parse[0], parse[1], atc1[0], atc1[1], atc1[2], ato1[0], ato1[1], ato1[2]]) self.executeAttach([parse[0], parse[2], atc2[0], atc2[1], atc2[2], ato2[0], ato2[1], ato2[2]]) # Clear cuboids + attachment points, but keep the commands that made them in memory def resetState(self): self.cuboids = self.getBoundBox() self.att_points = {} # Supported commands and their execution functions # Commands are first parsed to get their type + parameters. Then, the line is executed by calling to the appropriate execute function def execute(self, line): res = None if "Cuboid(" in line: parse = self.parseCuboid(line) self.executeCuboid(parse) elif "attach(" in line: parse = self.parseAttach(line) self.executeAttach(parse) elif "reflect(" in line: parse = self.parseReflect(line) res = self.executeReflect(parse) elif "translate(" in line: parse = self.parseTranslate(line) res = self.executeTranslate(parse) elif "squeeze(" in line: parse = self.parseSqueeze(line) res = self.executeSqueeze(parse) # return any new lines generated by macros return res # To re-run a program given a set of commands and parameters. Often used during fitting to unstructurd geometry. def runProgram(self, param_lines): self.resetState() command_to_func = { "Cuboid": self.executeCuboid, "attach": self.executeAttach, "squeeze": self.executeSqueeze, "translate": self.executeTranslate, "reflect": self.executeReflect } for command, parse in param_lines: func = command_to_func[command] func(parse) # ** Helper Functions FOR ShapeAssembly Class ** def lineToAttrs(line): P = Program() if "Cuboid(" in line: func = "Cuboid" parse = list(P.parseCuboid(line)) param_inds = [1,2,3] elif "attach(" in line: func = "attach" parse = list(P.parseAttach(line)) param_inds = [2,3,4,5,6,7] elif "reflect(" in line: func = "reflect" parse = list(P.parseReflect(line)) param_inds = [] elif "translate(" in line: func = "translate" parse = list(P.parseTranslate(line)) param_inds = [3] elif "squeeze(" in line: func = "squeeze" parse = list(P.parseSqueeze(line)) param_inds = [4,5] tensor = torch.nn.Parameter( torch.tensor([parse[i] for i in param_inds]) ) for i,j in enumerate(param_inds): parse[j] = tensor[i] return func, tensor, parse def make_hier_prog(lines): all_progs = {} root_name = None cur_name = None cur_prog = [] cur_children = [] for line in lines: if len(line) == 0: continue ls = line.strip().split() if ls[0] == 'Assembly': cur_name = ls[1] if root_name is None: root_name = cur_name elif ls[0] == '}': all_progs[cur_name] = (cur_prog, cur_children) cur_children = [] cur_prog = [] cur_name = None else: if 'Cuboid' in line: if 'Program_' in line: cur_children.append(ls[0]) else: cur_children.append(None) cur_prog.append(line[1:-1]) hp = {'name': root_name} q = [hp] while(len(q)) > 0: node = q.pop(0) prog, children = all_progs[node['name']] node['prog'] = prog node['children'] = [] for child in children: c = {} if child is not None: c = {'name': child} q.append(c) node['children'].append(c) return hp # Logic to create function text def make_function(name, args): args = [str(arg) for arg in args] return '{}({})'.format(name, ", ".join(args)) def assign(var_name, value): return '{} = {}'.format(var_name, value) def get_pos_delta(abox, rbox, pos): r = torch.stack(( abox.rfnorm, abox.tfnorm, abox.ffnorm )).T return (r @ pos) + (abox.pos - rbox.pos) # Given a cuboid cube, and its local program bounding volume rbox, and the actual placement of its bonding volume abox, find the correct transformation for cube def apply_delta(abox, rbox, cube): r = torch.stack(( abox.rfnorm, abox.tfnorm, abox.ffnorm )).T cube.dims *= (abox.dims / rbox.dims) cube.pos = (r @ cube.pos) + (abox.pos - rbox.pos) cube.rfnorm = r @ cube.rfnorm cube.tfnorm = r @ cube.tfnorm cube.ffnorm = r @ cube.ffnorm # Execute a hierarchical shapeassembly program, in a differentiable fashion def diff_hier_execute(root, param_dict, return_all = False): q = [(root, None, False)] scene = [] while len(q) > 0: node, bbox, flip_axis = q.pop(0) param_lines = param_dict[node['name']] if bbox is None: bbox = Cuboid("bbox", aligned = True, vis=False) bbox.dims = torch.stack(param_lines[0][1][1:4]) TP = Program() TP.runProgram(param_lines) if flip_axis: TP.flip(flip_axis) rbox = TP.cuboids.pop('bbox') add = [] for i, c_key in enumerate(TP.cuboids.keys()): flip = False cub = TP.cuboids[c_key] child = None if i+1 < len(node["children"]): child = node["children"][i+1] elif cub.parent is not None: pi = list(TP.cuboids.keys()).index(cub.parent) child = deepcopy(node["children"][pi+1]) if cub.parent_axis is not None and 'prog' in child: flip_axis = cub.parent_axis # cub is found through local execution, this brings it into global space apply_delta(bbox, rbox, cub) # if intermediate cuboid, add back into queue if child is not None and len(child) > 0: q.append((child, cub, flip_axis)) # if leave cuboid, save these cuboid to the add list else: add.append(cub) scene += add verts = torch.tensor([],dtype=torch.float) faces = torch.tensor([],dtype=torch.long) for cube in scene: v, f = cube.getTris() if v is not None and f is not None: faces = torch.cat((faces, (f + verts.shape[0]))) verts = torch.cat((verts, v)) if return_all: return verts, faces, scene return verts, faces # Execute a hierarchical shapeassembly program def hier_execute(root, return_all=False): bbox = Cuboid('bbox') bbox.dims = torch.tensor( [float(a) for a in re.split(r'[()]', root['prog'][0])[1].split(',')[:3]] ) q = [(root, bbox, None)] scene = [] count = 0 while len(q) > 0: node, bbox, flip_axis = q.pop(0) lines = node["prog"] TP = Program() for line in lines: TP.execute(line) if flip_axis: TP.flip(flip_axis) rbox = TP.cuboids.pop('bbox') add = [] i_map = {} i = 0 for c_key in TP.cuboids.keys(): cub = TP.cuboids[c_key] child = None flip_axis = False if '_ref' not in c_key and '_trans' not in c_key: child = node["children"][i+1] i_map[c_key] = i i += 1 else: pi = i_map[cub.parent] child = deepcopy(node["children"][pi+1]) if cub.parent_axis is not None and 'prog' in child: flip_axis = cub.parent_axis # cub is found through local execution, this brings it into global space apply_delta(bbox, rbox, cub) # if intermediate cuboid, add back into queue if child is not None and len(child) > 0: q.append((child, cub, flip_axis)) # if leave cuboid, save these cuboid to the add list else: add.append(cub) scene += add verts = torch.tensor([],dtype=torch.float) faces = torch.tensor([],dtype=torch.long) for cube in scene: v, f = cube.getTris() if v is not None and f is not None: faces = torch.cat((faces, (f + verts.shape[0]))) verts = torch.cat((verts, v)) if not return_all: return verts, faces else: return verts, faces, scene # Execute a hierarchical shapeassembly program def hier_execute_to_npz(root): bbox = Cuboid('bbox') bbox.dims = torch.tensor( [float(a) for a in re.split(r'[()]', root['prog'][0])[1].split(',')[:3]] ) q = [(root, bbox, None)] scene_cubes = [] scene_aps = [] count = 0 while len(q) > 0: node, bbox, flip_axis = q.pop(0) lines = node["prog"] TP = Program() for line in lines: TP.execute(line) if flip_axis: TP.flip(flip_axis) rbox = TP.cuboids.pop('bbox') add_cubes = [] add_aps = [ ( ap, get_pos_delta(bbox, rbox, TP.att_points[ap].getPos()) ) for ap in TP.att_points ] i_map = {} i = 0 remove_inds = set() for c_key in TP.cuboids.keys(): cub = TP.cuboids[c_key] child = None flip_axis = False if '_ref' not in c_key and '_trans' not in c_key: child = node["children"][i+1] i_map[c_key] = i i += 1 else: pi = i_map[cub.parent] child = deepcopy(node["children"][pi+1]) if cub.parent_axis is not None and 'prog' in child: flip_axis = cub.parent_axis # cub is found through local execution, this brings it into global space apply_delta(bbox, rbox, cub) # if intermediate cuboid, add back into queue if child is not None and len(child) > 0: for j, (ap_name, _) in enumerate(add_aps): remove_inds.add(j) q.append((child, cub, flip_axis)) # if leave cuboid, save these cuboid to the add list else: add_cubes.append(cub) scene_cubes += add_cubes scene_aps += [ pos for i, (_, pos) in enumerate(add_aps) if i not in remove_inds ] scene_cubes = [ torch.cat(( c.pos, c.dims, c.rfnorm, c.tfnorm )).numpy() for c in scene_cubes ] scene_cubes = np.array(scene_cubes) scene_aps = [ap.numpy() for ap in scene_aps] scene_aps = np.array(scene_aps) return scene_cubes, scene_aps class ShapeAssembly(): """ Contains entrypoint logic for parsing and executing entire programs""" # Execute a program without any hierarchy def run_local(self, lines, out_file): P = Program() for line in lines: P.execute(line) P.render(out_file) # Execute a program differentiable w.r.t. to the parameters in the param_dict def diff_run(self, hier, param_dict): return diff_hier_execute(hier, param_dict) # Execute a program def run(self, lines, out_file): hier_prog = make_hier_prog(lines) verts, faces = hier_execute(hier_prog) writeObj(verts, faces, out_file) # load lines from a program file def load_lines(self, prog_file): lines = [] with open(prog_file) as f: for line in f: lines.append(line) return lines # Convert a hierarchy + dictionary of parameters into a full ShapeAssembly program def fill_hier(self, hier, param_dict): q = [hier] while(len(q) > 0): node = q.pop(0) param_lines = param_dict.pop(node['name']) lines = [] for func, tparams in param_lines: params = [ round(p.item(), 2) if isinstance(p, torch.Tensor) else p \ for p in tparams ] if func == "Cuboid": lines.append( assign( params[0], make_function(func, params[1:]) ) ) else: lines.append( make_function(func, params) ) node['prog'] = lines for c in node['children']: if len(c) > 0: q.append(c) # Return a program hierarchy, a dictionary of # nodes to lines (as tensors), and a list # of all tensors in lines -> used during differentiable execution def make_hier_param_dict(self, lines): h = make_hier_prog(lines) q = [h] param_list = [] param_dict = {} while len(q) > 0: node = q.pop(0) prog = node.pop('prog') lines = [] for line in prog: func, tensor, parse = lineToAttrs(line) lines.append((func, parse)) if tensor.shape[0] > 0: param_list.append(tensor) param_dict[node['name']] = lines for c in node['children']: if len(c) > 0: q.append(c) return h, param_dict, param_list if __name__ == '__main__': mode, prog_file, out_file = sys.argv[1], sys.argv[2], sys.argv[3] sa = ShapeAssembly() lines = sa.load_lines(prog_file) if mode == 'run': sa.run(lines, out_file) elif mode == 'run_local': sa.run_local(lines, out_file) else: print(f'Mode {mode} not recognized')
52,690
30.875983
371
py
ShapeMOD
ShapeMOD-main/SA_lang/tasks/pc_encoder.py
import torch import torch.nn as nn from collections import namedtuple import etw_pytorch_utils as pt_utils from pointnet2.utils.pointnet2_modules import PointnetSAModule class PCEncoder(nn.Module): r""" PointNet2 with single-scale grouping Classification network Parameters ---------- num_classes: int Number of semantics classes to predict over -- size of softmax classifier input_channels: int = 3 Number of input channels in the feature descriptor for each point. If the point cloud is Nx9, this value should be 6 as in an Nx9 point cloud, 3 of the channels are xyz, and 6 are feature descriptors use_xyz: bool = True Whether or not to use the xyz position of a point as a feature """ def __init__(self, input_channels=3, use_xyz=True): super(PCEncoder, self).__init__() self.SA_modules = nn.ModuleList() self.SA_modules.append( PointnetSAModule( npoint=512, radius=0.2, nsample=64, mlp=[input_channels, 64, 64, 128], use_xyz=use_xyz, ) ) self.SA_modules.append( PointnetSAModule( npoint=128, radius=0.4, nsample=64, mlp=[128, 128, 128, 256], use_xyz=use_xyz, ) ) self.SA_modules.append( PointnetSAModule(mlp=[256, 256, 512, 1024], use_xyz=use_xyz) ) self.FC_layer = ( pt_utils.Seq(1024) .fc(256, bn=False, activation=None) ) def forward(self, pointcloud): # type: (Pointnet2SSG, torch.cuda.FloatTensor) -> pt_utils.Seq r""" Forward pass of the network Parameters ---------- pointcloud: Variable(torch.cuda.FloatTensor) (B, N, 3 + input_channels) tensor Point cloud to run predicts on Each point in the point-cloud MUST be formated as (x, y, z, features...) """ xyz = pointcloud.contiguous() features = pointcloud.transpose(1, 2).contiguous() for module in self.SA_modules: xyz, features = module(xyz, features) return self.FC_layer(features.squeeze(-1)) if __name__ == '__main__': device = torch.device('cuda') a = torch.randn(10,10000,3).to(device) enc = PCEncoder().to(device) print(enc(a).shape)
2,553
31.329114
112
py
ShapeMOD
ShapeMOD-main/SA_lang/tasks/infer_model_prog.py
import sys sys.path.append("../") sys.path.append("../../") import os import torch import torch.nn as nn import torch.nn.functional as F from ShapeAssembly import Program, hier_execute, make_hier_prog, ShapeAssembly import sa_utils as utils import infer_recon_metrics import argparse import ast import random from torch.utils.data import DataLoader import time import matplotlib.pyplot as plt from argparse import Namespace import pickle from make_abs_data import fillProgram from tqdm import tqdm import infer_sem_valid as sv import numpy import json from pc_encoder import PCEncoder MAX_LINES = 30 device = torch.device("cuda") sa = ShapeAssembly() outpath = "model_output" SEQ_LEN = 0 MAX_DEPTH = 4 I_LENGTH = 11 # cuboid index SQ_LENGTH = 6 # face index SYM_LENGTH = 3 # axis index VERBOSE = False closs = nn.BCEWithLogitsLoss(reduction='none') celoss = nn.CrossEntropyLoss(reduction='none') FUNC_PRED_FIELD = 'func_pred' CPARAM_PRED_FIELD = 'cparam_pred' DPARAM_PRED_FIELD = 'dparam_pred' TRAIN_LOG_INFO = [ ('Total loss', 'loss', 'batch_count'), ('Func loss', 'func', 'batch_count'), ('Float Param loss', 'f_prm', 'batch_count'), ('Disc Param loss', 'd_prm', 'batch_count'), ('Bool Param loss', 'b_prm', 'batch_count'), ('BBox Param Loss', 'bbox', 'batch_count'), ('Child loss', 'child', 'batch_count'), ('KL loss', 'kl', 'batch_count'), ('Func Correct %', 'func_corr', 'func_total'), ('Disc Correct %', 'd_corr', 'd_total'), ('Bool Correct %', 'b_corr', 'b_total'), ('Child Correct %', 'c_corr', 'c_total'), ('Float Mean Error', 'f_prm', 'f_norm'), ] EVAL_LOG_INFO = [ ('CD', 'cd', 'no_norm'), ('Fscore-Def', 'fscore-def', 'no_norm'), ('Fscore-01', 'fscore-01', 'no_norm'), ('Fscore-03', 'fscore-03', 'no_norm'), ('Fscore-05', 'fscore-05', 'no_norm'), ('Rooted', 'rooted', 'no_norm'), ('Stable', 'stable', 'no_norm'), ('Prog Creation %', 'prog_creation_perc', 'no_norm'), ('Missing Line Number %', 'miss_ln', 'num_progs'), ('Extra Line Number %', 'extra_ln', 'num_progs'), ('Corr Child Number %', 'cn_corr', 'num_progs'), ('Func Correct %', 'func_corr', 'func_total'), ('Disc Correct %', 'd_corr', 'd_total'), ('Bool Correct %', 'b_corr', 'b_total'), ('Child Correct %', 'child_corr', 'child_total'), ('Float Mean Error', 'f_prm', 'num_progs'), ('BBox Mean Error', 'bbox', 'num_progs') ] def weighted_mae_loss(input, target, weight): return torch.sum(weight * (input - target).abs()) def make_function(name, args): args = [str(arg) for arg in args] return '{}({})'.format(name, ", ".join(args)) def assign(var_name, value): return '{} = {}'.format(var_name, value) # Multi-layer perceptron helper function class MLP(nn.Module): def __init__(self, ind, hdim1, hdim2, odim): super(MLP, self).__init__() self.l1 = nn.Linear(ind, hdim1) self.l2 = nn.Linear(hdim1, hdim2) self.l3 = nn.Linear(hdim2, odim) def forward(self, x): x = F.leaky_relu(self.l1(x),.2) x = F.leaky_relu(self.l2(x),.2) return self.l3(x) # Unused, but can be used class simplePCEncoder(nn.Module): def __init__(self, feat_len): super(simplePCEncoder, self).__init__() self.conv1 = nn.Conv1d(3, 64, 1) self.conv2 = nn.Conv1d(64, 128, 1) self.conv3 = nn.Conv1d(128, 128, 1) self.conv4 = nn.Conv1d(128, feat_len, 1) self.bn1 = nn.BatchNorm1d(64) self.bn2 = nn.BatchNorm1d(128) self.bn3 = nn.BatchNorm1d(128) self.bn4 = nn.BatchNorm1d(feat_len) self.mlp2mu = nn.Linear(feat_len, feat_len) def forward(self, pc): net = pc.transpose(2, 1) net = torch.relu(self.bn1(self.conv1(net))) net = torch.relu(self.bn2(self.conv2(net))) net = torch.relu(self.bn3(self.conv3(net))) net = torch.relu(self.bn4(self.conv4(net))) net = net.max(dim=2)[0] return self.mlp2mu(net) def getBatchEncData(batch): pcs = torch.stack(batch,dim=0).to(device) return pcs def get_pc_encodings(batch, encoder): pcs = getBatchEncData(batch) codes = encoder(pcs) return codes def getShapeEvalResult(pred_node, gt_node): # Corr line %, Func Corr %, Child Corr %, Bool Corr %, Disc Corr %, Float Mean Distance, # child num corr, line numc corr results = {} ln = min(pred_node[FUNC_PRED_FIELD].shape[0], gt_node['e_func_gt'][1:-1].shape[0]) if pred_node[FUNC_PRED_FIELD].shape[0] < gt_node['e_func_gt'][1:-1].shape[0]: results['miss_ln'] = 1. else: results['miss_ln'] = 0. if pred_node[FUNC_PRED_FIELD].shape[0] > gt_node['e_func_gt'][1:-1].shape[0]: results['extra_ln'] = 1. else: results['extra_ln'] = 0. results['num_progs'] = 1. func_pred = pred_node[FUNC_PRED_FIELD][:ln].cpu() cparam_pred = pred_node[CPARAM_PRED_FIELD][:ln].cpu() dparam_pred = pred_node[DPARAM_PRED_FIELD][:ln].cpu() func_target = gt_node['e_func_gt'][1:1+ln].cpu() float_target = gt_node['e_float_target'][1:ln+1].cpu() float_mask = gt_node['e_float_mask'][1:ln+1].cpu() disc_target = gt_node['e_disc_target'][1:ln+1].cpu() disc_mask = gt_node['e_disc_mask'][1:ln+1].cpu() bool_target = gt_node['e_bool_target'][1:ln+1].cpu() bool_mask = gt_node['e_bool_mask'][1:ln+1].cpu() results['func_corr'] = ( func_pred == func_target ).float().sum().item() results['func_total'] = ln results['b_corr'] = ( (cparam_pred == bool_target).float() * bool_mask ).sum().item() results['b_total'] = (bool_mask.sum() + 1e-8).item() results['d_corr'] = ( (dparam_pred == disc_target).float() * disc_mask ).sum().item() results['d_total'] = (disc_mask.sum() + 1e-8).item() results['f_prm'] = (( ((cparam_pred - float_target).abs() * float_mask).sum() ) / (float_mask.sum() + 1e-8)).item() cn = min(len(pred_node['children']), len(gt_node['children'])) if len(pred_node['children']) == len(gt_node['children']): results['cn_corr'] = 1. else: results['cn_corr'] = 0. results['child_corr'] = 0. results['child_total'] = cn for pred_child, gt_child in zip(pred_node['children'][:cn], gt_node['children'][:cn]): if len(pred_child) == 0 and len(gt_child) == 0: results['child_corr'] += 1. elif len(pred_child) > 0 and len(gt_child) > 0: results['child_corr'] += 1. child_results = getShapeEvalResult(pred_child, gt_child) for key in child_results: results[key] += child_results[key] return results def getBatchDecData(progs): seq = torch.stack([p['seq'] for p in progs],dim=0) inp_seq = seq[:,:-1,:] tar_seq = seq[:,1:,:] seq_weight = torch.stack([p['seq_mask'] for p in progs],dim=0)[:,1:] fprm_weight = torch.stack([p['fprm_weight'] for p in progs], dim=0)[:, 1:] children = [p['children'] for p in progs] child_target = torch.stack([p['child_target'] for p in progs], dim=0)[:, 1:] child_weight = torch.stack([p['child_weight'] for p in progs], dim=0)[:, 1:] lnext_inds = ((child_target.bool()) & child_weight.bool()).nonzero().tolist() cnext_inds = [] for i in range(len(progs)): for j in progs[i]["exp_inds"]: cnext_inds.append([i,j]) return inp_seq, tar_seq, seq_weight, fprm_weight, children, \ child_target, child_weight, lnext_inds, cnext_inds # GRU recurrent Decoder class dslDecoder(nn.Module): def __init__(self, hidden_dim, metadata): super(dslDecoder, self).__init__() self.hidden_dim = hidden_dim self.metadata = metadata self.input_dim = metadata['tl_size'] self.bb_net = MLP(hidden_dim, hidden_dim, hidden_dim, 3) self.inp_net = MLP(self.input_dim + MAX_DEPTH + 3, hidden_dim, hidden_dim, hidden_dim) self.max_cparams = metadata['max_cparams'] self.num_funcs = len(metadata['cparam_map']) self.gru = nn.GRU(hidden_dim, hidden_dim, batch_first = True) self.max_children = metadata['max_children'] self.child_net = MLP(hidden_dim * 2, hidden_dim//2, hidden_dim//8, self.max_children) on = self.max_children * hidden_dim self.next_code_net = MLP(hidden_dim * 2, hidden_dim, hidden_dim, on) tl_map = metadata['tl_map'] func_net = MLP(hidden_dim, hidden_dim//2, hidden_dim//4, self.num_funcs) start, end = tl_map['func'] func_net.start = start func_net.end = end func_net.line_cond = None func_net.bb_cond = None func_net.name = 'func' func_net.func = None func_net._type = 'func' net_list = [func_net] for _,func in metadata['func_map'].items(): if len(metadata['dparam_map'][func]) > 0: low = 1e8 high = -1e8 for i, prm in enumerate(metadata['dparam_map'][func]): start, end = tl_map[f'{func}_{prm}_{i}'] mlp = MLP(hidden_dim, hidden_dim //2, hidden_dim//4, end-start) mlp.start = start mlp.end = end mlp.line_cond = None mlp.bb_cond = False mlp.func = func mlp.name = f'{func}_{prm}_{i}' mlp._type = 'disc' net_list.append(mlp) low = min(low, start) high = max(high, end) else: low = None high = None if f'{func}_f' in tl_map: start, end = tl_map[f'{func}_f'] if low is None or high is None: mlp = MLP(hidden_dim + 3, hidden_dim //2, hidden_dim//4, end-start) mlp.line_cond = None else: mlp = MLP(hidden_dim + 3 + high-low, hidden_dim //2, hidden_dim//4, end-start) mlp.line_cond = (low, high) mlp.start = start mlp.end = end mlp.bb_cond = True mlp.func = func mlp.name = f'{func}_f' mlp._type = 'f' net_list.append(mlp) if f'{func}_b' in tl_map: start, end = tl_map[f'{func}_b'] mlp = MLP(hidden_dim, hidden_dim //2, hidden_dim//4, end-start) mlp.start = start mlp.end = end mlp.line_cond = None mlp.bb_cond = False mlp.func = func mlp.name = f'{func}_b' mlp._type = 'b' net_list.append(mlp) self.net_list = nn.ModuleList(net_list) def train_forward(self, inp_seq, code, _bb_dims, _hier_ind, gt_seq): bb_dims = _bb_dims.unsqueeze(1).repeat(1,inp_seq.shape[1],1) hier_oh = torch.zeros( inp_seq.shape[0], inp_seq.shape[1], MAX_DEPTH, device=device ) hier_oh[ torch.arange(inp_seq.shape[0],device=device), :, _hier_ind ] = 1.0 inp = self.inp_net( torch.cat( (inp_seq, bb_dims, hier_oh), dim=2) ) gru_out, _ = self.gru(inp, code.unsqueeze(0).contiguous()) fstart, fend = self.metadata['tl_map']['func'] commands = torch.argmax(gt_seq[:,:,fstart:fend], dim = 2).flatten() flat_out = torch.zeros(commands.shape[0], inp_seq.shape[2], device=device).float() flat_gt_seq = gt_seq.reshape(commands.shape[0], -1) flat_bb_dims = bb_dims.reshape(commands.shape[0], -1) flat_gru_out = gru_out.reshape(commands.shape[0], -1) for net in self.net_list: if net.func is None: # Func net flat_out[:,net.start:net.end] = net(flat_gru_out) else: cmd_inds = (commands == net.func).nonzero().flatten() if net.line_cond is not None: line_cond = flat_gt_seq[cmd_inds, net.line_cond[0]:net.line_cond[1]] else: line_cond = torch.zeros(cmd_inds.shape[0], 0, device=device) if net.bb_cond is True: bb_cond = flat_bb_dims[cmd_inds,:] else: bb_cond = torch.zeros(cmd_inds.shape[0], 0, device=device) flat_out[cmd_inds, net.start:net.end] = net(torch.cat(( flat_gru_out[cmd_inds,:], line_cond, bb_cond ), dim=1)) out = flat_out.view(inp_seq.shape) double_enc = torch.cat(( gru_out, code.unsqueeze(1).repeat(1, gru_out.shape[1], 1) ), dim = 2) child_pred = self.child_net( double_enc ) next_codes = self.next_code_net( double_enc ).view(inp_seq.shape[0], inp_seq.shape[1], self.max_children, -1) return out, next_codes, child_pred def calc_loss(self, out, pchild, tar, child_tar, seq_weight, fprm_weight, child_weight): result_map = {} result_map['f_prm'] = weighted_mae_loss(out, tar, fprm_weight) tl_map = self.metadata['tl_map'] fstart, fend = tl_map['func'] with torch.no_grad(): commands = torch.argmax(tar[:,:,fstart:fend], dim = 2).flatten() pcommands = torch.argmax(out[:,:,fstart:fend], dim = 2).flatten() result_map['func_corr'] = ( (commands == pcommands).float() * seq_weight.flatten() ).sum().item() * 1.0 result_map['func_total'] = seq_weight.sum().item() result_map['func'] = (celoss( out[:,:,fstart:fend].view(-1,fend-fstart), commands ) * seq_weight.flatten()).sum() result_map['child'] = (closs(pchild, child_tar) * child_weight).sum() result_map['c_corr'] = (((pchild >= 0).float() == child_tar).float() * child_weight).sum().item() result_map['c_total'] = (child_weight.sum() + 1e-8).item() b_prm = torch.tensor(0,device=device).float() b_corr = 0 b_total = 0 d_prm = torch.tensor(0,device=device).float() d_corr = 0 d_total = 0 for key, (start, end) in tl_map.items(): if key == 'func': continue cmd = int(key.split('_')[0]) typ = key.split('_')[1] if typ == 'f': continue cmd_mask = (commands == cmd).float().flatten() if cmd_mask.sum() == 0: continue if typ == 'i' or typ == 'sq' or typ == 'sym': with torch.no_grad(): ktar = torch.argmax(tar[:,:,start:end], dim=2).flatten() kpout = torch.argmax(out[:,:,start:end], dim=2).flatten() d_corr += ( (kpout == ktar).float() * cmd_mask ).sum().item() * 1.0 d_total += cmd_mask.sum().item() d_prm += (celoss( out[:,:,start:end].view(-1, end-start), ktar ) * cmd_mask).sum() elif typ == 'b': with torch.no_grad(): ktar = tar[:, :, start:end].reshape(-1, end-start) kpout = (out[:,:, start:end].reshape(-1, end-start) >= 0).float() b_corr += ( (kpout == ktar).float() * cmd_mask.unsqueeze(-1) ).sum().item() * 1.0 b_total += cmd_mask.sum().item() * (end-start) b_prm += (closs( out[:,:,start:end].reshape(-1, end-start), ktar ) * cmd_mask.unsqueeze(-1)).sum() result_map['b_prm'] = b_prm result_map['b_corr'] = b_corr + 1e-8 result_map['b_total'] = b_total + 1e-8 result_map['d_prm'] = d_prm result_map['d_corr'] = d_corr + 1e-8 result_map['d_total'] = d_total + 1e-8 result_map['f_norm'] = (loss_config['f_prm'] * (fprm_weight.sum() + 1e-8)).item() *1.0 return result_map def getStartLine(self): l = torch.zeros(1,1,self.metadata['tl_size'],device=device).float() l[0,0,0] = 1.0 return l def decode_line(self, line): _cparam = torch.zeros(self.metadata['max_cparams'], device=device).float() _dparam = torch.zeros(self.metadata['max_dparams'], device=device).long() fstart, fend = self.metadata['tl_map']['func'] cmd = line[fstart:fend].argmax().item() float_preds = [] bool_preds = [] tl_map = self.metadata['tl_map'] if f'{cmd}_f' in tl_map: fstart, fend = tl_map[f'{cmd}_f'] float_preds = line[fstart:fend].tolist() if f'{cmd}_b' in tl_map: bstart, bend = tl_map[f'{cmd}_b'] bool_preds = line[bstart:bend].tolist() for i,prm in enumerate(self.metadata['cparam_map'][cmd]): if prm == 'f': v = float_preds.pop(0) elif prm == 'b': v = bool_preds.pop(0) _cparam[i] = v for i, prm in enumerate(self.metadata['dparam_map'][cmd]): istart, iend = tl_map[f'{cmd}_{prm}_{i}'] v = torch.argmax(line[istart:iend]).item() _dparam[i] = v return cmd, _cparam, _dparam def split_lines(self, lines): p_func = [] p_cparam = [] p_dparam = [] for line in lines: _f, _c, _d = self.decode_line(line) p_func.append(_f) p_cparam.append(_c) p_dparam.append(_d) return p_func, p_cparam, p_dparam def eval_forward(self, inp_seq, code, code_start, bb_dims, hier_ind): bb_dims = bb_dims.unsqueeze(0).unsqueeze(0).repeat(1,inp_seq.shape[1],1) hier_oh = torch.zeros(1, inp_seq.shape[1], MAX_DEPTH).to(device) hier_oh[0, :, min(hier_ind, 2)] = 1.0 inp = self.inp_net( torch.cat( (inp_seq, bb_dims, hier_oh), dim=2) ) gru_out, h = self.gru(inp, code.view(1,1,-1)) out = torch.zeros(inp_seq.shape, device=device).float() commands = None for net in self.net_list: if net.func is not None: assert commands is not None if net.func != commands: continue if net.line_cond is not None: line_cond = out[:,:,net.line_cond[0]:net.line_cond[1]] else: line_cond = torch.zeros(inp_seq.shape[0], inp_seq.shape[1], 0, device=device) if net.bb_cond is True: bb_cond = bb_dims else: bb_cond = torch.zeros(inp_seq.shape[0], inp_seq.shape[1], 0, device=device) raw_out = net(torch.cat(( gru_out, line_cond, bb_cond ), dim=2)) if net._type == 'func': cmd = torch.argmax(raw_out.squeeze()).item() out[0,0, net.start+cmd] = 1.0 assert commands == None commands = cmd elif net._type == 'disc': m = torch.argmax(raw_out.squeeze()).item() out[0,0, net.start+m] = 1.0 elif net._type == 'b': r = (raw_out.squeeze() >= 0.).float() out[0,0,net.start:net.end] = r elif net._type == 'f': bb_max = bb_cond.max().item() r = torch.clamp(raw_out.squeeze(), 0.0, 10.) out[0,0,net.start:net.end] = r double_enc = torch.cat(( gru_out, code_start.repeat(1, gru_out.shape[1], 1) ), dim = 2) child_pred = self.child_net( double_enc ) next_codes = self.next_code_net( double_enc ).view(inp_seq.shape[0], inp_seq.shape[1], self.max_children, -1) return out, next_codes, child_pred, h def train_progs(self, batch, codes, loss_config): result_map = {key:0. for key in loss_config} bbox_target = torch.stack([b['bbox_gt'] for b in batch], dim=0) bbox_pred = self.bb_net(codes) bbox_loss = (bbox_target - bbox_pred).abs().sum() result_map['bbox'] = bbox_loss qp = batch qe = codes qbb = bbox_target qhi = torch.zeros(len(batch), device=device).long() while len(qp) > 0: bs = min(len(batch), len(qp)) bprogs = qp[:bs] bencs = qe[:bs] bbb = qbb[:bs] bhi = qhi[:bs] qp = qp[bs:] qe = qe[bs:] qbb = qbb[bs:] qhi = qhi[bs:] inp_seq, tar_seq, seq_weights, fprm_weights, children, \ child_targets, child_weights, lnext_inds, cnext_inds = getBatchDecData(bprogs) pout, pnext, pchild = self.train_forward( inp_seq, bencs, bbb, bhi, tar_seq ) _result = self.calc_loss( pout, pchild, tar_seq, child_targets, seq_weights, fprm_weights, child_weights ) for key in _result: if key in result_map: result_map[key] += _result[key] else: result_map[key] = _result[key] _qp = [] _qe = [] _qbb = [] _qhi = [] for ((li, lj, lk), (ci, cj)) in zip(lnext_inds, cnext_inds): _qp.append(children[ci][cj]) _qe.append(pnext[li, lj, lk]) _qbb.append(children[ci][cj]['bbox_gt']) _qhi.append(bhi[li]+1) if len(_qp) > 0: qe = torch.cat((qe, torch.stack(_qe)), dim = 0) qbb = torch.cat((qbb, torch.stack(_qbb).to(device)), dim = 0) qp += _qp qhi = torch.cat((qhi, torch.stack(_qhi)), dim = 0) return result_map def writeConfigFile(args): os.system(f'mkdir {outpath} > /dev/null 2>&1') os.system(f'mkdir {outpath}/{args.exp_name} > /dev/null 2>&1') os.system(f'mkdir {outpath}/{args.exp_name}/plots > /dev/null 2>&1') os.system(f'mkdir {outpath}/{args.exp_name}/plots/train > /dev/null 2>&1') os.system(f'mkdir {outpath}/{args.exp_name}/plots/eval > /dev/null 2>&1') os.system(f'mkdir {outpath}/{args.exp_name}/programs > /dev/null 2>&1') os.system(f'mkdir {outpath}/{args.exp_name}/programs/train > /dev/null 2>&1') os.system(f'mkdir {outpath}/{args.exp_name}/programs/val > /dev/null 2>&1') os.system(f'mkdir {outpath}/{args.exp_name}/programs/test > /dev/null 2>&1') os.system(f'mkdir {outpath}/{args.exp_name}/programs/gt > /dev/null 2>&1') os.system(f'mkdir {outpath}/{args.exp_name}/objs > /dev/null 2>&1') os.system(f'mkdir {outpath}/{args.exp_name}/objs/train > /dev/null 2>&1') os.system(f'mkdir {outpath}/{args.exp_name}/objs/val > /dev/null 2>&1') os.system(f'mkdir {outpath}/{args.exp_name}/objs/test > /dev/null 2>&1') os.system(f'mkdir {outpath}/{args.exp_name}/objs/gt > /dev/null 2>&1') os.system(f'mkdir {outpath}/{args.exp_name}/models > /dev/null 2>&1') with open(f"{outpath}/{args.exp_name}/config.txt", "w") as f: f.write(f"{args}\n") def get_max_sq_len(nodes): m = 0 q = [n for _,n in nodes] while len(q) > 0: n = q.pop(0) l = n['func_gt'].shape[0] m = max(m, l) for c in n['children']: if len(c) > 0: q.append(c) return m def addTargets(node, metadata): gt_funcs = node['func_gt'] gt_cparams = node['cparam_gt'] gt_dparams = node['dparam_gt'] node['bbox_gt'] = torch.from_numpy(node['cparam_gt'][1][:3]).float().to(device) seq = torch.zeros(SEQ_LEN, metadata['tl_size'], device=device).float() fprm_weight = torch.zeros(SEQ_LEN, metadata['tl_size'], device=device).float() seq_mask = torch.zeros(SEQ_LEN, device=device).float() seq_end = 0 for i, fn in enumerate(gt_funcs.tolist()): seq_mask[i] = 1.0 seq_end = i line = torch.zeros(metadata['tl_size'], device=device).float() weight = torch.zeros(metadata['tl_size'], device=device).float() line[fn] = 1.0 float_vals = [] bool_vals = [] for j, tp in enumerate(metadata['cparam_map'][fn]): if tp == 'f': float_vals.append(gt_cparams[i][j].item()) elif tp == 'b': bool_vals.append(gt_cparams[i][j].item()) else: assert False, f'bad type {tp}' if len(float_vals) > 0: start, end = metadata['tl_map'][f'{fn}_f'] line[start:end] = torch.tensor(float_vals, device=device) weight[start:end] = 1.0 if len(bool_vals) > 0: start, end = metadata['tl_map'][f'{fn}_b'] line[start:end] = torch.tensor(bool_vals, device=device) for j, prm in enumerate(metadata['dparam_map'][fn]): tar = int(gt_dparams[i][j].item()) start, end = metadata['tl_map'][f'{fn}_{prm}_{j}'] line[start+tar] = 1.0 seq[i] = line fprm_weight[i] = weight node['seq'] = seq node['fprm_weight'] = fprm_weight node['seq_end'] = torch.tensor([seq_end],device=device).long() node['seq_mask'] = seq_mask child_target = torch.zeros(SEQ_LEN, metadata['max_children'], device=device).float() child_weight = torch.zeros(SEQ_LEN, metadata['max_children'], device=device).float() child_enc_mask = torch.zeros(SEQ_LEN, device = device).float() child_to_ln_map = {} for i, inds in enumerate(node['child_gt']): for j, cind in enumerate(inds): child_weight[i, j] = 1.0 child_to_ln_map[cind] = (i, j) if len(node['children'][cind]) > 0: child_target[i,j] = 1.0 child_enc_mask[i] = 1.0 node['child_target'] = child_target node['child_weight'] = child_weight node['child_to_ln_map'] = child_to_ln_map node['child_enc_mask'] = child_enc_mask node['exp_inds'] = [] for i, child in enumerate(node['children']): if len(child) > 0: node['exp_inds'].append(i) addTargets(child, metadata) gt_cparams = node['cparam_gt'] gt_dparams = node['dparam_gt'] bool_target = torch.zeros(gt_funcs.shape[0], metadata['max_cparams'], device=device) bool_mask = torch.zeros(gt_funcs.shape[0], metadata['max_cparams'], device=device) float_target = torch.zeros(gt_funcs.shape[0], metadata['max_cparams'], device=device) float_mask = torch.zeros(gt_funcs.shape[0], metadata['max_cparams'], device=device) for i, tf in enumerate(gt_funcs): for j, tp in enumerate(metadata['cparam_map'][tf]): if tp == 'f': float_target[i][j] = gt_cparams[i][j].item() float_mask[i][j] = 1.0 elif tp == 'b': bool_target[i][j] = gt_cparams[i][j].item() bool_mask[i][j] = 1.0 else: assert False, f'bad type {tp}' disc_target = torch.zeros(gt_funcs.shape[0], metadata['max_dparams'], device=device).long() disc_mask = torch.zeros(gt_funcs.shape[0], metadata['max_dparams'], device=device) for i, tf in enumerate(gt_funcs): for j, _ in enumerate(metadata['dparam_map'][tf]): disc_target[i][j] = gt_dparams[i][j].item() disc_mask[i][j] = 1.0 node['e_func_gt'] = torch.tensor(node['func_gt'], device=device).long() node['e_cparam_gt'] = torch.tensor(node['cparam_gt'], device=device).float() node['e_dparam_gt'] = torch.tensor(node['dparam_gt'], device=device).long() node['e_bool_target'] = bool_target.float() node['e_float_target'] = float_target.float() node['e_disc_target'] = disc_target.long() node['e_bool_mask'] = bool_mask.float() node['e_float_mask'] = float_mask.float() node['e_disc_mask'] = disc_mask.float() def _col(samples): return samples def _bcol(samples): return samples # Full encoder + decoder training logic for a single program (i.e. a batch) def model_train(batch, encoder, decoder, enc_opt, dec_opt, loss_config): codes = get_pc_encodings([b[2] for b in batch], encoder) shape_result = decoder.train_progs([b[1] for b in batch], codes, loss_config) loss = 0. for key in loss_config: loss += (loss_config[key] * shape_result[key]) / len(batch) shape_result[key] = (loss_config[key] * shape_result[key].item()) / len(batch) if torch.is_tensor(loss) and enc_opt is not None and dec_opt is not None: dec_opt.zero_grad() enc_opt.zero_grad() loss.backward() dec_opt.step() enc_opt.step() shape_result['loss'] = loss.item() return shape_result def model_train_results(dataset, encoder, decoder, enc_opt, dec_opt, loss_config): if enc_opt is not None and dec_opt is not None: decoder.train() encoder.train() else: decoder.eval() encoder.eval() ep_result = {} bc = 0. for batch in dataset: bc += 1. batch_result = model_train( batch, encoder, decoder, dec_opt, enc_opt, loss_config ) for key in batch_result: if key not in ep_result: ep_result[key] = batch_result[key] else: ep_result[key] += batch_result[key] ep_result['batch_count'] = bc return ep_result def model_eval( eval_train_dataset, eval_val_dataset, eval_test_dataset, encoder, decoder, exp_name, epoch, num_write, metadata ): eval_results = {} for name, dataset in [ ('train', eval_train_dataset), ('val', eval_val_dataset), ('test', eval_test_dataset) ]: if len(dataset) == 0: eval_results[name] = {} continue named_results = { 'count': 0., 'miss_hier_prog': 0., 'no_norm': 1.0 } recon_sets = [] for batch in dataset: assert len(batch) == 1, 'batch size 1' shape = batch[0] named_results['count'] += 1. code = get_pc_encodings([shape[2]], encoder) node = sv.sem_eval_prog(decoder, code.squeeze()) try: shape_result = getShapeEvalResult(node, shape[1]) shape_result['bbox'] = (node['bb_dims'] - shape[1]['bbox_gt']).abs().sum().item() except Exception as e: if VERBOSE: print(f"FAILED SHAPE EVAL RESULT WITH {e}") shape_result = {} for key in shape_result: if key not in named_results: named_results[key] = shape_result[key] else: named_results[key] += shape_result[key] try: fillProgram( metadata['dsl'], node, metadata, FUNC_PRED_FIELD, CPARAM_PRED_FIELD, DPARAM_PRED_FIELD, ) recon_sets.append((node, shape[1], shape[0], shape[2])) except Exception as e: if VERBOSE: print(f"Failed Recon Program with {e}") named_results[f'miss_hier_prog'] += 1. # For reconstruction, get metric performance recon_results, recon_misses = infer_recon_metrics.recon_metrics( recon_sets, outpath, exp_name, name, epoch, VERBOSE, num_write + 1 ) for key in recon_results: named_results[key] = recon_results[key] named_results[f'miss_hier_prog'] += recon_misses named_results[f'prog_creation_perc'] = ( named_results[f'count'] - named_results[f'miss_hier_prog'] ) / named_results[f'count'] eval_results[name] = named_results return eval_results def print_train_results(result, exp_name): res = "" for name, key, norm_key in TRAIN_LOG_INFO: if key in result: res += f" {name} : {round(result[key] / (result[norm_key]+1e-8), 2)}\n" utils.log_print(res, f"{outpath}/{exp_name}/log.txt") def print_eval_results(result, exp_name): res = "" for name, key, norm_key in EVAL_LOG_INFO: if key in result: res += f" {name} : {round(result[key] / (result[norm_key]+1e-8), 4)}\n" utils.log_print(res, f"{outpath}/{exp_name}/log.txt") def make_train_plots(train_result, val_result, train_plots, aepochs, exp_name): for name, key, norm_key in TRAIN_LOG_INFO: for rname, result in [('train', train_result), ('val', val_result)]: if key not in result: continue res = result[key] / (result[norm_key]+1e-8) if name not in train_plots[rname]: train_plots[rname][name] = [res] else: train_plots[rname][name].append(res) if name not in train_plots['train']: continue plt.clf() plt.plot(aepochs, train_plots['train'][name], label='train') if name in train_plots['val']: plt.plot(aepochs, train_plots['val'][name], label='val') plt.legend() plt.grid() plt.savefig(f"{outpath}/{exp_name}/plots/train/{name}.png") def make_eval_plots(eval_result, eval_plots, aepochs, exp_name): for name, key, norm_key in EVAL_LOG_INFO: for rname, result in [ ('train', eval_result['train']), ('val', eval_result['val']), ('test', eval_result['test']) ]: if key not in result: continue res = result[key] / (result[norm_key]+1e-8) if not name in eval_plots[rname]: eval_plots[rname][name] = [res] else: eval_plots[rname][name].append(res) if name not in eval_plots['train']: continue plt.clf() plt.plot(aepochs, eval_plots['train'][name], label='train') if name in eval_plots['val']: plt.plot(aepochs, eval_plots['val'][name], label='val') if name in eval_plots['test']: plt.plot(aepochs, eval_plots['test'][name], label='test') plt.legend() plt.grid() plt.savefig(f"{outpath}/{exp_name}/plots/eval/{name}.png") # Helper function for keeping consistent train/val splits def getInds(train_ind_file): inds = set() with open(train_ind_file) as f: for line in f: inds.add(line.strip()) return inds def get_tensor_layout(metadata): # Figure out size of tensor # map from (func, type) -> indices in tensor tl_map = {} size = len(metadata['func_map']) tl_map['func'] = (0, size) start = size size += (I_LENGTH * metadata['max_d_i_params']) \ + (SQ_LENGTH * metadata['max_d_sq_params']) \ + (SYM_LENGTH * metadata['max_d_sym_params']) for func, prms in metadata['dparam_map'].items(): i_start = start sq_start = i_start + I_LENGTH * metadata['max_d_i_params'] sym_start = sq_start + SQ_LENGTH * metadata['max_d_sq_params'] for i, _typ in enumerate(prms): if _typ == 'i': opt_len = I_LENGTH _start = i_start i_start += opt_len elif _typ == 'sq': opt_len = SQ_LENGTH _start = sq_start sq_start += opt_len elif _typ == 'sym': opt_len = SYM_LENGTH _start = sym_start sym_start += opt_len tl_map[f'{func}_{_typ}_{i}'] = ( _start, _start + opt_len, ) for func, prms in metadata['cparam_map'].items(): nf = 0 nb = 0 for prm in prms: if 'f' in prm: nf += 1 elif 'b' in prm: nb += 1 if nf > 0: size += nf tl_map[f'{func}_f'] = (size-nf, size) if nb > 0: size += nb tl_map[f'{func}_b'] = (size-nb, size) return size, tl_map # Main entry-point of modeling logic def run_train( dataset_path, exp_name, max_shapes, epochs, hidden_dim, eval_per, loss_config, rd_seed, print_per, num_write, dec_lr, enc_lr, save_per, category, batch_size ): random.seed(rd_seed) numpy.random.seed(rd_seed) torch.manual_seed(rd_seed) raw_data = pickle.load(open(f"{dataset_path}_train.data", "rb")) metadata = pickle.load(open(f"{dataset_path}_train.meta", "rb")) metadata['max_children'] = max([int(i) for i in metadata['num_cube_map'].values()]) metadata['rev_func_map'] = {v:k for k, v in metadata['func_map'].items()} for key in ('i', 'sq', 'sym'): metadata[f'max_d_{key}_params'] = max([len([ __l for __l in _l if __l == key]) for _l in metadata['dparam_map'].values()]) tl_size, tl_map = get_tensor_layout(metadata) metadata['tl_size'] = tl_size metadata['tl_map'] = tl_map all_inds = [] all_data = [] all_pc = [] good_inds = [] max_sq_len = get_max_sq_len(raw_data) print(f"Seq len: {max_sq_len}") global SEQ_LEN SEQ_LEN = max_sq_len for d in tqdm(raw_data): if len(all_inds) >= max_shapes: break if len(good_inds) > 0 and d[0] not in good_inds: continue addTargets(d[1], metadata) fillProgram( metadata['dsl'], d[1], metadata, 'func_gt', 'cparam_gt', 'dparam_gt' ) pc = numpy.load(f'pc_data/{category}/{d[0]}.pts.npy') tpc = torch.from_numpy(pc) all_data.append(d[1]) all_inds.append(d[0]) all_pc.append(tpc) samples = list(zip(all_inds, all_data, all_pc)) train_ind_file = f'pc_data_splits/{category}/train.txt' val_ind_file = f'pc_data_splits/{category}/val.txt' test_ind_file = f'pc_data_splits/{category}/test.txt' train_samples = [] val_samples = [] test_samples = [] train_inds = getInds(train_ind_file) val_inds = getInds(val_ind_file) test_inds = getInds(test_ind_file) misses = 0. num_parts = [] for (ind, prog, pc) in samples: if ind in train_inds or ind in good_inds: train_samples.append((ind, prog, pc)) elif ind in val_inds: val_samples.append((ind, prog, pc)) elif ind in test_inds: test_samples.append((ind, prog, pc)) else: misses += 1 if len(good_inds) > 0: val_samples = train_samples[:1] test_samples = train_samples[:1] print(f"Samples missed: {misses}") train_num = len(train_samples) val_num = len(val_samples) test_num = len(test_samples) train_dataset = DataLoader( train_samples, batch_size, shuffle=True, collate_fn = _bcol ) val_dataset = DataLoader( val_samples, batch_size, shuffle = False, collate_fn = _bcol ) num_eval = max(val_num, test_num, len(good_inds)) eval_train_dataset = DataLoader( train_samples[:num_eval], 1, shuffle=False, collate_fn = _col ) eval_val_dataset = DataLoader( val_samples[:num_eval], 1, shuffle = False, collate_fn = _col ) eval_test_dataset = DataLoader( test_samples[:num_eval], 1, shuffle = False, collate_fn = _col ) utils.log_print(f"Training size: {len(train_samples)}", f"{outpath}/{exp_name}/log.txt") utils.log_print(f"Validation size: {len(val_samples)}", f"{outpath}/{exp_name}/log.txt") utils.log_print(f"Test size: {len(test_samples)}", f"{outpath}/{exp_name}/log.txt") val_epochs = [] train_epochs = [] train_plots = {'train': {}, 'val': {}} eval_plots = {'train': {}, 'val': {}, 'test': {}} encoder = PCEncoder() decoder = dslDecoder( hidden_dim, metadata, ) encoder.to(device) decoder.to(device) dec_opt = torch.optim.Adam( decoder.parameters(), lr = dec_lr, eps = 1e-6 ) enc_opt = torch.optim.Adam( encoder.parameters(), lr = enc_lr, eps = 1e-6 ) print('training ...') for e in range(0, epochs): json.dump({ 'train': train_plots, 'eval': eval_plots, 'train_epochs': train_epochs, 'val_epochs': val_epochs, }, open(f"{outpath}/{exp_name}/res.json" ,'w')) decoder.epoch = e do_print = (e+1) % print_per == 0 t = time.time() if do_print: utils.log_print(f"\nEpoch {e}:", f"{outpath}/{exp_name}/log.txt") train_result = model_train_results( train_dataset, encoder, decoder, enc_opt, dec_opt, loss_config, ) if do_print: with torch.no_grad(): val_result = model_train_results( val_dataset, encoder, decoder, None, None, loss_config ) train_epochs.append(e) make_train_plots(train_result, val_result, train_plots, train_epochs, exp_name) utils.log_print( f"Train results: ", f"{outpath}/{exp_name}/log.txt" ) print_train_results(train_result, exp_name) utils.log_print( f"Val results: ", f"{outpath}/{exp_name}/log.txt" ) print_train_results(val_result, exp_name) utils.log_print( f" Time = {time.time() - t}", f"{outpath}/{exp_name}/log.txt" ) with torch.no_grad(): if (e+1) % eval_per == 0: decoder.eval() encoder.eval() t = time.time() eval_results = model_eval( eval_train_dataset, eval_val_dataset, eval_test_dataset, encoder, decoder, exp_name, e, num_write, metadata ) utils.log_print(f"Evaluation training set results:", f"{outpath}/{exp_name}/log.txt") print_eval_results(eval_results['train'], exp_name) utils.log_print(f"Evaluation validation set results:", f"{outpath}/{exp_name}/log.txt") print_eval_results(eval_results['val'], exp_name) utils.log_print(f"Evaluation test set results:", f"{outpath}/{exp_name}/log.txt") print_eval_results(eval_results['test'], exp_name) utils.log_print(f"Eval Time = {time.time() - t}", f"{outpath}/{exp_name}/log.txt") val_epochs.append(e) make_eval_plots(eval_results, eval_plots, val_epochs, exp_name) if (e+1) % save_per == 0: utils.log_print("Saving Models", f"{outpath}/{exp_name}/log.txt") torch.save(decoder.state_dict(), f"{outpath}/{exp_name}/models/decoder_{e}.pt") torch.save(encoder.state_dict(), f"{outpath}/{exp_name}/models/encoder_{e}.pt") def run_eval(args): raw_data = pickle.load(open(f"{args.dataset_path}_train.data", "rb")) metadata = pickle.load(open(f"{args.dataset_path}_train.meta", "rb")) metadata['max_children'] = max([int(i) for i in metadata['num_cube_map'].values()]) metadata['rev_func_map'] = {v:k for k, v in metadata['func_map'].items()} for key in ('i', 'sq', 'sym'): metadata[f'max_d_{key}_params'] = max([len([ __l for __l in _l if __l == key]) for _l in metadata['dparam_map'].values()]) tl_size, tl_map = get_tensor_layout(metadata) metadata['tl_size'] = tl_size metadata['tl_map'] = tl_map inds = [] pc_data = [] gt_progs = [] test_ind_file = f'pc_data_splits/{args.category}/test.txt' test_inds = getInds(test_ind_file) for d in tqdm(raw_data): if len(inds) > args.num_gen: break if d[0] not in test_inds: continue fillProgram( metadata['dsl'], d[1], metadata, 'func_gt', 'cparam_gt', 'dparam_gt' ) inds.append(d[0]) gt_progs.append(d[1]) pc = numpy.load(f'pc_data/{args.category}/{d[0]}.pts.npy') tpc = torch.from_numpy(pc).to(device) pc_data.append(tpc) encoder = PCEncoder() decoder = dslDecoder( args.hidden_dim, metadata, ) encoder.load_state_dict(torch.load( f'{args.exp_name}/models/encoder_{args.load_epoch}.pt' )) decoder.load_state_dict(torch.load( f'{args.exp_name}/models/decoder_{args.load_epoch}.pt' )) encoder.to(device) decoder.to(device) encoder.eval() decoder.eval() outname = f'{args.exp_name}/infer_output' os.system(f'mkdir {outname}') os.system(f'mkdir {args.exp_name}/infer_output/objs') os.system(f'mkdir {args.exp_name}/infer_output/progs') for ind, pc, gtprog in tqdm(list(zip(inds, pc_data, gt_progs))): code = get_pc_encodings([pc], encoder) node = sv.sem_eval_prog(decoder, code.squeeze()) fillProgram( metadata['dsl'], node, metadata, FUNC_PRED_FIELD, CPARAM_PRED_FIELD, DPARAM_PRED_FIELD, ) try: pverts, pfaces = hier_execute(node) gtverts, gtfaces = hier_execute(gtprog) except Exception: continue utils.writeObj( pverts, pfaces, f'{outname}/objs/{ind}_pred_prog.obj' ) utils.writeObj( gtverts, gtfaces, f'{outname}/objs/{ind}_gt_prog.obj' ) utils.writeHierProg(node, 'dsl_prog', f"{outname}/progs/{ind}_pred_prog.txt") utils.writeHierProg(gtprog, 'dsl_prog', f"{outname}/progs/{ind}_gt_prog.txt") utils.writeSPC( pc.cpu().numpy(), f'{outname}/objs/{ind}_gt_pc.obj' ) if __name__ == '__main__': parser = argparse.ArgumentParser(description="Run Visual Program Induction model") parser.add_argument('-ds', '--dataset_path', help='Path to program data, e.g. data/shapemod_chair', type = str) parser.add_argument('-en', '--exp_name', help='name of experiment', type = str) parser.add_argument('-c', '--category', type = str, help = 'category of PartNet') parser.add_argument('-ms', '--max_shapes', default = 100000, type = int, help = 'max number of shapes to train/evaluate on ') parser.add_argument('-e', '--epochs', default = 2000, type = int, help = 'number of epochs to run for') parser.add_argument('-hd', '--hidden_dim', default = 256, type = int, help = 'hidden dimension size') parser.add_argument('-prp', '--print_per', default = 10, type = int, help = 'how often to print out training set statistics') parser.add_argument('-evp', '--eval_per', default = 50, type = int, help = 'how often to run evaluation statistics') parser.add_argument('-sp', '--save_per', default = 10, type = int, help = 'how often to save the model') parser.add_argument('-enc_lr', '--enc_lr', default = 0.0002, type = float, help = 'encoder learning rate') parser.add_argument('-dec_lr', '--dec_lr', default = 0.0002, type = float, help = 'decoder learning rate') parser.add_argument('-rd', '--rd_seed', default = 42, type = int, help = 'random seed') parser.add_argument('-ng', '--num_gen', default = 1000, type = int, help = 'number of shapes to generate each generation period') parser.add_argument('-nw', '--num_write', default = 25, type = int, help = 'number of shapes to write to .obj each evaluation period') parser.add_argument('-f_lw', '--f_lw', default = 50., type = float, help = 'weight on loss of continuous parameters') parser.add_argument('-d_lw', '--d_lw', default = 1., type = float, help = 'weight on loss of discrete parameters') parser.add_argument('-b_lw', '--b_lw', default = 1., type = float, help = 'weight on loss of boolean parameters') parser.add_argument('-c_lw', '--c_lw', default = 1., type = float, help = 'weight on loss of child predictions') parser.add_argument('-fn_lw', '--fn_lw', default = 1., type = float, help = 'weight on loss of function predictions') parser.add_argument('-b', '--batch_size', default = 32, type=int, help = 'batch size') parser.add_argument('-le', '--load_epoch', default = None, type=int, help = 'model epoch to load from pre-trained model') parser.add_argument('-m', '--mode', default = 'train', type=str, help = 'whether to train new model or generate samples from pre-trained model') args = parser.parse_args() loss_config = { 'd_prm': args.d_lw, 'f_prm': args.f_lw, 'bbox': args.f_lw, 'b_prm': args.b_lw, 'child': args.c_lw, 'func': args.fn_lw, } if args.mode == 'train': writeConfigFile(args) run_train( dataset_path=args.dataset_path, exp_name=args.exp_name, max_shapes=args.max_shapes, epochs=args.epochs, hidden_dim=args.hidden_dim, eval_per=args.eval_per, loss_config=loss_config, rd_seed=args.rd_seed, print_per=args.print_per, num_write=args.num_write, enc_lr=args.enc_lr, dec_lr=args.dec_lr, save_per=args.save_per, category=args.category, batch_size=args.batch_size ) elif args.mode == 'infer': run_eval(args) else: print(f"Bad mode {args.mode}")
53,175
34.450667
148
py
ShapeMOD
ShapeMOD-main/SA_lang/tasks/model_prog.py
import sys sys.path.append("../") sys.path.append("../../") import os import torch import torch.nn as nn import torch.nn.functional as F from ShapeAssembly import Program, hier_execute, make_hier_prog, ShapeAssembly import sa_utils as utils import recon_metrics import gen_metrics import argparse import ast import random from torch.utils.data import DataLoader import time import matplotlib.pyplot as plt from argparse import Namespace import pickle from make_abs_data import fillProgram, getCuboidDims from tqdm import tqdm import sem_valid as sv import numpy import json MAX_LINES = 30 device = torch.device("cuda") sa = ShapeAssembly() outpath = "model_output" SEQ_LEN = 0 MAX_DEPTH = 4 I_LENGTH = 11 # cuboid index SQ_LENGTH = 6 # face index SYM_LENGTH = 3 # axis index VERBOSE = False closs = nn.BCEWithLogitsLoss(reduction='none') celoss = nn.CrossEntropyLoss(reduction='none') FUNC_PRED_FIELD = 'func_pred' CPARAM_PRED_FIELD = 'cparam_pred' DPARAM_PRED_FIELD = 'dparam_pred' USE_SEM_VALID = True TRAIN_LOG_INFO = [ ('Total loss', 'loss', 'batch_count'), ('Func loss', 'func', 'batch_count'), ('Float Param loss', 'f_prm', 'batch_count'), ('Disc Param loss', 'd_prm', 'batch_count'), ('Bool Param loss', 'b_prm', 'batch_count'), ('BBox Param Loss', 'bbox', 'batch_count'), ('Child loss', 'child', 'batch_count'), ('KL loss', 'kl', 'batch_count'), ('Func Correct %', 'func_corr', 'func_total'), ('Disc Correct %', 'd_corr', 'd_total'), ('Bool Correct %', 'b_corr', 'b_total'), ('Child Correct %', 'c_corr', 'c_total'), ('Float Mean Error', 'f_prm', 'f_norm'), ] EVAL_LOG_INFO = [ ('CD', 'cd', 'no_norm'), ('Fscore-Def', 'fscore-def', 'no_norm'), ('Fscore-03', 'fscore-03', 'no_norm'), ('Fscore-05', 'fscore-05', 'no_norm'), ('IoU', 'iou', 'no_norm'), ('Prog Creation %', 'prog_creation_perc', 'no_norm'), ('Missing Line Number %', 'miss_ln', 'num_progs'), ('Extra Line Number %', 'extra_ln', 'num_progs'), ('Corr Child Number %', 'cn_corr', 'num_progs'), ('Func Correct %', 'func_corr', 'func_total'), ('Disc Correct %', 'd_corr', 'd_total'), ('Bool Correct %', 'b_corr', 'b_total'), ('Child Correct %', 'child_corr', 'child_total'), ('Float Mean Error', 'f_prm', 'num_progs'), ('BBox Mean Error', 'bbox', 'num_progs') ] GEN_LOG_INFO = [ ('Prog Creation %', 'prog_creation_perc'), ('Number of parts', 'num_parts'), ('Rootedness', 'rootedness'), ('Stability', 'stability'), ('Val FD', 'val_fd'), ('Train FD', 'train_fd'), ('Generalization', 'gen'), ('Coverage', 'cov'), ('Variety', 'var'), ] def weighted_mae_loss(input, target, weight): return torch.sum(weight * (input - target).abs()) def make_function(name, args): args = [str(arg) for arg in args] return '{}({})'.format(name, ", ".join(args)) def assign(var_name, value): return '{} = {}'.format(var_name, value) # Multi-layer perceptron helper function class MLP(nn.Module): def __init__(self, ind, hdim1, hdim2, odim): super(MLP, self).__init__() self.l1 = nn.Linear(ind, hdim1) self.l2 = nn.Linear(hdim1, hdim2) self.l3 = nn.Linear(hdim2, odim) def forward(self, x): x = F.leaky_relu(self.l1(x),.2) x = F.leaky_relu(self.l2(x),.2) return self.l3(x) class Sampler(nn.Module): def __init__(self, feature_size, hidden_size): super(Sampler, self).__init__() self.mlp1 = nn.Linear(feature_size, hidden_size) self.mlp2mu = nn.Linear(hidden_size, feature_size) self.mlp2var = nn.Linear(hidden_size, feature_size) def forward(self, x, variational): encode = torch.relu(self.mlp1(x)) mu = self.mlp2mu(encode) if not variational: return mu else: logvar = self.mlp2var(encode) std = logvar.mul(0.5).exp_() eps = torch.randn_like(std) kld = mu.pow(2).add_(logvar.exp()).mul_(-1).add_(1).add_(logvar) return torch.cat([eps.mul(std).add_(mu), kld], 1) # Given a sample (a hierarchical program) get the encoding of it and the KL loss def get_encodings(batch, encoder, variational): hd = encoder.hidden_dim batch_out = encoder.get_latent_codes(batch, variational) enc = batch_out[:,:hd] kld = batch_out[:,hd:] if variational: return enc, -kld.sum() else: return enc, 0.0 def getBatchEncData(progs): seq = torch.stack([p[0]['seq'] for p in progs],dim=0) seq_ends = torch.cat([p[0]['seq_end'] for p in progs], dim=0) child_masks = torch.stack([p[0]['child_enc_mask'] for p in progs], dim=0).unsqueeze(-1) inds = torch.tensor([p[1] for p in progs]) par_inds = torch.tensor([p[2] for p in progs]) par_lns = torch.tensor([p[3] for p in progs]) par_l_cns = torch.tensor([p[4] for p in progs]) return seq, seq_ends, child_masks, inds, par_inds, par_lns, par_l_cns # Encodes an entire hierarchical program class dslEncoder(nn.Module): def __init__(self, hidden_dim, metadata): super(dslEncoder, self).__init__() self.hidden_dim = hidden_dim self.metadata = metadata self.input_dim = metadata['tl_size'] self.inp_net = MLP(self.input_dim, hidden_dim, hidden_dim, hidden_dim) self.h_start = torch.nn.Parameter( torch.randn(1, 1, hidden_dim, device=device, requires_grad=True) ) self.gru = nn.GRU(hidden_dim, hidden_dim, batch_first = True) self.sampler = Sampler(hidden_dim, hidden_dim) def encode_programs(self, batch): hier_levels = {} q = [(b, 0, 1e8, 1e8, 1e8) for b in batch] ind = 0 while len(q) > 0: n, hl, pind, pln, pl_cn = q.pop(0) if hl not in hier_levels: hier_levels[hl] = [(n, ind, pind, pln, pl_cn)] else: hier_levels[hl].append((n, ind, pind, pln, pl_cn)) c_ln_map = n['child_to_ln_map'] for cn, child in enumerate(n['children']): if len(child) > 0: q.append((child, hl+1, ind, c_ln_map[cn][0], c_ln_map[cn][1])) ind += 1 hl_keys = list(hier_levels.keys()) hl_keys.sort(reverse=True) child_codes = torch.zeros( ind, SEQ_LEN, self.metadata['max_children'], self.hidden_dim ).float().to(device) for hl in hl_keys[:-1]: qp = hier_levels[hl] while len(qp) > 0: bs = min(len(batch), len(qp)) bprogs = qp[:bs] qp = qp[bs:] seq, seq_ends, child_masks, inds, \ par_inds, par_lns, par_l_cns = getBatchEncData(bprogs) local_inp = self.inp_net(seq) child_encs = child_codes[inds].sum(dim=2) inp = (child_masks * child_encs) + ((1-child_masks) * local_inp) gru_out, _ = self.gru(inp, self.h_start.repeat(1,inp.shape[0],1)) h = gru_out[ torch.arange(seq_ends.shape[0], device=device), seq_ends ] child_codes[par_inds, par_lns, par_l_cns] = h # special case last one seq, seq_ends, child_masks, inds, _, _, _ = getBatchEncData(hier_levels[0]) local_inp = self.inp_net(seq) child_encs = child_codes[inds].sum(dim=2) inp = (child_masks * child_encs) + ((1-child_masks) * local_inp) gru_out, _ = self.gru(inp, self.h_start.repeat(1,inp.shape[0],1)) h = gru_out[ torch.arange(seq_ends.shape[0], device=device), seq_ends ] return h def get_latent_codes(self, batch, variational): zs = self.encode_programs(batch) return self.sampler(zs, variational) def getShapeEvalResult(pred_node, gt_node): # Corr line %, Func Corr %, Child Corr %, Bool Corr %, Disc Corr %, Float Mean Distance, # child num corr, line numc corr results = {} ln = min(pred_node[FUNC_PRED_FIELD].shape[0], gt_node['e_func_gt'][1:-1].shape[0]) if pred_node[FUNC_PRED_FIELD].shape[0] < gt_node['e_func_gt'][1:-1].shape[0]: results['miss_ln'] = 1. else: results['miss_ln'] = 0. if pred_node[FUNC_PRED_FIELD].shape[0] > gt_node['e_func_gt'][1:-1].shape[0]: results['extra_ln'] = 1. else: results['extra_ln'] = 0. results['num_progs'] = 1. func_pred = pred_node[FUNC_PRED_FIELD][:ln].cpu() cparam_pred = pred_node[CPARAM_PRED_FIELD][:ln].cpu() dparam_pred = pred_node[DPARAM_PRED_FIELD][:ln].cpu() func_target = gt_node['e_func_gt'][1:1+ln].cpu() float_target = gt_node['e_float_target'][1:ln+1].cpu() float_mask = gt_node['e_float_mask'][1:ln+1].cpu() disc_target = gt_node['e_disc_target'][1:ln+1].cpu() disc_mask = gt_node['e_disc_mask'][1:ln+1].cpu() bool_target = gt_node['e_bool_target'][1:ln+1].cpu() bool_mask = gt_node['e_bool_mask'][1:ln+1].cpu() results['func_corr'] = ( func_pred == func_target ).float().sum().item() results['func_total'] = ln results['b_corr'] = ( (cparam_pred == bool_target).float() * bool_mask ).sum().item() results['b_total'] = (bool_mask.sum() + 1e-8).item() results['d_corr'] = ( (dparam_pred == disc_target).float() * disc_mask ).sum().item() results['d_total'] = (disc_mask.sum() + 1e-8).item() results['f_prm'] = (( ((cparam_pred - float_target).abs() * float_mask).sum() ) / (float_mask.sum() + 1e-8)).item() cn = min(len(pred_node['children']), len(gt_node['children'])) if len(pred_node['children']) == len(gt_node['children']): results['cn_corr'] = 1. else: results['cn_corr'] = 0. results['child_corr'] = 0. results['child_total'] = cn for pred_child, gt_child in zip(pred_node['children'][:cn], gt_node['children'][:cn]): if len(pred_child) == 0 and len(gt_child) == 0: results['child_corr'] += 1. elif len(pred_child) > 0 and len(gt_child) > 0: results['child_corr'] += 1. child_results = getShapeEvalResult(pred_child, gt_child) for key in child_results: results[key] += child_results[key] return results def getBatchDecData(progs): seq = torch.stack([p['seq'] for p in progs],dim=0) inp_seq = seq[:,:-1,:] tar_seq = seq[:,1:,:] seq_weight = torch.stack([p['seq_mask'] for p in progs],dim=0)[:,1:] fprm_weight = torch.stack([p['fprm_weight'] for p in progs], dim=0)[:, 1:] children = [p['children'] for p in progs] child_target = torch.stack([p['child_target'] for p in progs], dim=0)[:, 1:] child_weight = torch.stack([p['child_weight'] for p in progs], dim=0)[:, 1:] lnext_inds = ((child_target.bool()) & child_weight.bool()).nonzero().tolist() cnext_inds = [] for i in range(len(progs)): for j in progs[i]["exp_inds"]: cnext_inds.append([i,j]) return inp_seq, tar_seq, seq_weight, fprm_weight, children, \ child_target, child_weight, lnext_inds, cnext_inds # GRU recurrent Decoder class dslDecoder(nn.Module): def __init__(self, hidden_dim, metadata): super(dslDecoder, self).__init__() self.hidden_dim = hidden_dim self.metadata = metadata self.input_dim = metadata['tl_size'] self.bb_net = MLP(hidden_dim, hidden_dim, hidden_dim, 3) self.inp_net = MLP(self.input_dim + MAX_DEPTH + 3, hidden_dim, hidden_dim, hidden_dim) self.max_cparams = metadata['max_cparams'] self.num_funcs = len(metadata['cparam_map']) self.gru = nn.GRU(hidden_dim, hidden_dim, batch_first = True) self.max_children = metadata['max_children'] self.child_net = MLP(hidden_dim * 2, hidden_dim//2, hidden_dim//8, self.max_children) on = self.max_children * hidden_dim self.next_code_net = MLP(hidden_dim * 2, hidden_dim, hidden_dim, on) tl_map = metadata['tl_map'] func_net = MLP(hidden_dim, hidden_dim//2, hidden_dim//4, self.num_funcs) start, end = tl_map['func'] func_net.start = start func_net.end = end func_net.line_cond = None func_net.bb_cond = None func_net.name = 'func' func_net.func = None func_net._type = 'func' net_list = [func_net] for _,func in metadata['func_map'].items(): if len(metadata['dparam_map'][func]) > 0: low = 1e8 high = -1e8 for i, prm in enumerate(metadata['dparam_map'][func]): start, end = tl_map[f'{func}_{prm}_{i}'] mlp = MLP(hidden_dim, hidden_dim //2, hidden_dim//4, end-start) mlp.start = start mlp.end = end mlp.line_cond = None mlp.bb_cond = False mlp.func = func mlp.name = f'{func}_{prm}_{i}' mlp._type = 'disc' net_list.append(mlp) low = min(low, start) high = max(high, end) else: low = None high = None if f'{func}_f' in tl_map: start, end = tl_map[f'{func}_f'] if low is None or high is None: mlp = MLP(hidden_dim + 3, hidden_dim //2, hidden_dim//4, end-start) mlp.line_cond = None else: mlp = MLP(hidden_dim + 3 + high-low, hidden_dim //2, hidden_dim//4, end-start) mlp.line_cond = (low, high) mlp.start = start mlp.end = end mlp.bb_cond = True mlp.func = func mlp.name = f'{func}_f' mlp._type = 'f' net_list.append(mlp) if f'{func}_b' in tl_map: start, end = tl_map[f'{func}_b'] mlp = MLP(hidden_dim, hidden_dim //2, hidden_dim//4, end-start) mlp.start = start mlp.end = end mlp.line_cond = None mlp.bb_cond = False mlp.func = func mlp.name = f'{func}_b' mlp._type = 'b' net_list.append(mlp) self.net_list = nn.ModuleList(net_list) def train_forward(self, inp_seq, code, _bb_dims, _hier_ind, gt_seq): bb_dims = _bb_dims.unsqueeze(1).repeat(1,inp_seq.shape[1],1) hier_oh = torch.zeros( inp_seq.shape[0], inp_seq.shape[1], MAX_DEPTH, device=device ) hier_oh[ torch.arange(inp_seq.shape[0],device=device), :, _hier_ind ] = 1.0 inp = self.inp_net( torch.cat( (inp_seq, bb_dims, hier_oh), dim=2) ) gru_out, _ = self.gru(inp, code.unsqueeze(0).contiguous()) fstart, fend = self.metadata['tl_map']['func'] commands = torch.argmax(gt_seq[:,:,fstart:fend], dim = 2).flatten() flat_out = torch.zeros(commands.shape[0], inp_seq.shape[2], device=device).float() flat_gt_seq = gt_seq.reshape(commands.shape[0], -1) flat_bb_dims = bb_dims.reshape(commands.shape[0], -1) flat_gru_out = gru_out.reshape(commands.shape[0], -1) for net in self.net_list: if net.func is None: # Func net flat_out[:,net.start:net.end] = net(flat_gru_out) else: cmd_inds = (commands == net.func).nonzero().flatten() if net.line_cond is not None: line_cond = flat_gt_seq[cmd_inds, net.line_cond[0]:net.line_cond[1]] else: line_cond = torch.zeros(cmd_inds.shape[0], 0, device=device) if net.bb_cond is True: bb_cond = flat_bb_dims[cmd_inds,:] else: bb_cond = torch.zeros(cmd_inds.shape[0], 0, device=device) flat_out[cmd_inds, net.start:net.end] = net(torch.cat(( flat_gru_out[cmd_inds,:], line_cond, bb_cond ), dim=1)) out = flat_out.view(inp_seq.shape) double_enc = torch.cat(( gru_out, code.unsqueeze(1).repeat(1, gru_out.shape[1], 1) ), dim = 2) child_pred = self.child_net( double_enc ) next_codes = self.next_code_net( double_enc ).view(inp_seq.shape[0], inp_seq.shape[1], self.max_children, -1) return out, next_codes, child_pred def calc_loss(self, out, pchild, tar, child_tar, seq_weight, fprm_weight, child_weight): result_map = {} result_map['f_prm'] = weighted_mae_loss(out, tar, fprm_weight) tl_map = self.metadata['tl_map'] fstart, fend = tl_map['func'] with torch.no_grad(): commands = torch.argmax(tar[:,:,fstart:fend], dim = 2).flatten() pcommands = torch.argmax(out[:,:,fstart:fend], dim = 2).flatten() result_map['func_corr'] = ( (commands == pcommands).float() * seq_weight.flatten() ).sum().item() * 1.0 result_map['func_total'] = seq_weight.sum().item() result_map['func'] = (celoss( out[:,:,fstart:fend].view(-1,fend-fstart), commands ) * seq_weight.flatten()).sum() result_map['child'] = (closs(pchild, child_tar) * child_weight).sum() result_map['c_corr'] = (((pchild >= 0).float() == child_tar).float() * child_weight).sum().item() result_map['c_total'] = (child_weight.sum() + 1e-8).item() b_prm = torch.tensor(0,device=device).float() b_corr = 0 b_total = 0 d_prm = torch.tensor(0,device=device).float() d_corr = 0 d_total = 0 for key, (start, end) in tl_map.items(): if key == 'func': continue cmd = int(key.split('_')[0]) typ = key.split('_')[1] if typ == 'f': continue cmd_mask = (commands == cmd).float().flatten() if cmd_mask.sum() == 0: continue if typ == 'i' or typ == 'sq' or typ == 'sym': with torch.no_grad(): ktar = torch.argmax(tar[:,:,start:end], dim=2).flatten() kpout = torch.argmax(out[:,:,start:end], dim=2).flatten() d_corr += ( (kpout == ktar).float() * cmd_mask ).sum().item() * 1.0 d_total += cmd_mask.sum().item() d_prm += (celoss( out[:,:,start:end].view(-1, end-start), ktar ) * cmd_mask).sum() elif typ == 'b': with torch.no_grad(): ktar = tar[:, :, start:end].reshape(-1, end-start) kpout = (out[:,:, start:end].reshape(-1, end-start) >= 0).float() b_corr += ( (kpout == ktar).float() * cmd_mask.unsqueeze(-1) ).sum().item() * 1.0 b_total += cmd_mask.sum().item() * (end-start) b_prm += (closs( out[:,:,start:end].reshape(-1, end-start), ktar ) * cmd_mask.unsqueeze(-1)).sum() result_map['b_prm'] = b_prm result_map['b_corr'] = b_corr + 1e-8 result_map['b_total'] = b_total + 1e-8 result_map['d_prm'] = d_prm result_map['d_corr'] = d_corr + 1e-8 result_map['d_total'] = d_total + 1e-8 result_map['f_norm'] = (loss_config['f_prm'] * (fprm_weight.sum() + 1e-8)).item() *1.0 return result_map def getStartLine(self): l = torch.zeros(1,1,self.metadata['tl_size'],device=device).float() l[0,0,0] = 1.0 return l def decode_line(self, line): _cparam = torch.zeros(self.metadata['max_cparams'], device=device).float() _dparam = torch.zeros(self.metadata['max_dparams'], device=device).long() fstart, fend = self.metadata['tl_map']['func'] cmd = line[fstart:fend].argmax().item() float_preds = [] bool_preds = [] tl_map = self.metadata['tl_map'] if f'{cmd}_f' in tl_map: fstart, fend = tl_map[f'{cmd}_f'] float_preds = line[fstart:fend].tolist() if f'{cmd}_b' in tl_map: bstart, bend = tl_map[f'{cmd}_b'] bool_preds = line[bstart:bend].tolist() for i,prm in enumerate(self.metadata['cparam_map'][cmd]): if prm == 'f': v = float_preds.pop(0) elif prm == 'b': v = bool_preds.pop(0) _cparam[i] = v for i, prm in enumerate(self.metadata['dparam_map'][cmd]): istart, iend = tl_map[f'{cmd}_{prm}_{i}'] v = torch.argmax(line[istart:iend]).item() _dparam[i] = v return cmd, _cparam, _dparam def split_lines(self, lines): p_func = [] p_cparam = [] p_dparam = [] for line in lines: _f, _c, _d = self.decode_line(line) p_func.append(_f) p_cparam.append(_c) p_dparam.append(_d) return p_func, p_cparam, p_dparam def eval_forward(self, inp_seq, code, code_start, bb_dims, hier_ind): bb_dims = bb_dims.unsqueeze(0).unsqueeze(0).repeat(1,inp_seq.shape[1],1) hier_oh = torch.zeros(1, inp_seq.shape[1], MAX_DEPTH).to(device) hier_oh[0, :, min(hier_ind, 2)] = 1.0 inp = self.inp_net( torch.cat( (inp_seq, bb_dims, hier_oh), dim=2) ) gru_out, h = self.gru(inp, code.view(1,1,-1)) out = torch.zeros(inp_seq.shape, device=device).float() commands = None for net in self.net_list: if net.func is not None: assert commands is not None if net.func != commands: continue if net.line_cond is not None: line_cond = out[:,:,net.line_cond[0]:net.line_cond[1]] else: line_cond = torch.zeros(inp_seq.shape[0], inp_seq.shape[1], 0, device=device) if net.bb_cond is True: bb_cond = bb_dims else: bb_cond = torch.zeros(inp_seq.shape[0], inp_seq.shape[1], 0, device=device) raw_out = net(torch.cat(( gru_out, line_cond, bb_cond ), dim=2)) if net._type == 'func': cmd = torch.argmax(raw_out.squeeze()).item() out[0,0, net.start+cmd] = 1.0 assert commands == None commands = cmd elif net._type == 'disc': m = torch.argmax(raw_out.squeeze()).item() out[0,0, net.start+m] = 1.0 elif net._type == 'b': r = (raw_out.squeeze() >= 0.).float() out[0,0,net.start:net.end] = r elif net._type == 'f': bb_max = bb_cond.max().item() r = torch.clamp(raw_out.squeeze(), 0.0, 10.) out[0,0,net.start:net.end] = r double_enc = torch.cat(( gru_out, code_start.repeat(1, gru_out.shape[1], 1) ), dim = 2) child_pred = self.child_net( double_enc ) next_codes = self.next_code_net( double_enc ).view(inp_seq.shape[0], inp_seq.shape[1], self.max_children, -1) return out, next_codes, child_pred, h def eval_prog(self, code, node = None): if node is None: bb_dims = self.bb_net(code) node = { 'depth': 0, 'bb_dims': bb_dims } if node['depth'] > MAX_DEPTH: node.pop('depth') return h = code.view(1,1, -1) h_start = h.clone() inp = self.getStartLine() out_lines = [] children = [] for i in range(MAX_LINES): inp, pnext, pchild, h = self.eval_forward( inp, h, h_start, node['bb_dims'], node['depth'] ) clean_out = inp.squeeze() fstart, fend = self.metadata['tl_map']['func'] func_ind = torch.argmax(clean_out[fstart:fend]).item() if func_ind == 1: break out_lines.append(clean_out) child_pred = pchild[0][0] next_codes = pnext[0][0] _, _c, _ = self.decode_line(clean_out) cube_dims = getCuboidDims( self.metadata['dsl'].library[ self.metadata['rev_func_map'][func_ind] ], _c, self.metadata['jparam_map'][func_ind], node['bb_dims'] ) cube_dims = torch.tensor(cube_dims,device=child_pred.device) for i in range(self.metadata['num_cube_map'][func_ind]): if child_pred[i].item() >= 0.0: child = { 'depth': node['depth']+1, 'bb_dims': cube_dims[i] } children.append(child) self.eval_prog(next_codes[i], child) else: children.append({}) node['children'] = children out_funcs, out_cprms, out_dprms = self.split_lines(out_lines) node[FUNC_PRED_FIELD] = torch.tensor(out_funcs) node[CPARAM_PRED_FIELD] = torch.stack(out_cprms) if len(out_cprms) > 0 else torch.tensor([]) node[DPARAM_PRED_FIELD] = torch.stack(out_dprms) if len(out_dprms) > 0 else torch.tensor([]) return node def train_progs(self, batch, codes, loss_config): result_map = {key:0. for key in loss_config} bbox_target = torch.stack([b['bbox_gt'] for b in batch], dim=0) bbox_pred = self.bb_net(codes) bbox_loss = (bbox_target - bbox_pred).abs().sum() result_map['bbox'] = bbox_loss qp = batch qe = codes qbb = bbox_target qhi = torch.zeros(len(batch), device=device).long() while len(qp) > 0: bs = min(len(batch), len(qp)) bprogs = qp[:bs] bencs = qe[:bs] bbb = qbb[:bs] bhi = qhi[:bs] qp = qp[bs:] qe = qe[bs:] qbb = qbb[bs:] qhi = qhi[bs:] inp_seq, tar_seq, seq_weights, fprm_weights, children, \ child_targets, child_weights, lnext_inds, cnext_inds = getBatchDecData(bprogs) pout, pnext, pchild = self.train_forward( inp_seq, bencs, bbb, bhi, tar_seq ) _result = self.calc_loss( pout, pchild, tar_seq, child_targets, seq_weights, fprm_weights, child_weights ) for key in _result: if key in result_map: result_map[key] += _result[key] else: result_map[key] = _result[key] _qp = [] _qe = [] _qbb = [] _qhi = [] for ((li, lj, lk), (ci, cj)) in zip(lnext_inds, cnext_inds): _qp.append(children[ci][cj]) _qe.append(pnext[li, lj, lk]) _qbb.append(children[ci][cj]['bbox_gt']) _qhi.append(bhi[li]+1) if len(_qp) > 0: qe = torch.cat((qe, torch.stack(_qe)), dim = 0) qbb = torch.cat((qbb, torch.stack(_qbb).to(device)), dim = 0) qp += _qp qhi = torch.cat((qhi, torch.stack(_qhi)), dim = 0) return result_map def writeConfigFile(args): os.system(f'mkdir {outpath} > /dev/null 2>&1') os.system(f'mkdir {outpath}/{args.exp_name} > /dev/null 2>&1') os.system(f'mkdir {outpath}/{args.exp_name}/plots > /dev/null 2>&1') os.system(f'mkdir {outpath}/{args.exp_name}/plots/train > /dev/null 2>&1') os.system(f'mkdir {outpath}/{args.exp_name}/plots/eval > /dev/null 2>&1') os.system(f'mkdir {outpath}/{args.exp_name}/plots/gen > /dev/null 2>&1') os.system(f'mkdir {outpath}/{args.exp_name}/programs > /dev/null 2>&1') os.system(f'mkdir {outpath}/{args.exp_name}/programs/train > /dev/null 2>&1') os.system(f'mkdir {outpath}/{args.exp_name}/programs/val > /dev/null 2>&1') os.system(f'mkdir {outpath}/{args.exp_name}/programs/gen > /dev/null 2>&1') os.system(f'mkdir {outpath}/{args.exp_name}/programs/gt > /dev/null 2>&1') os.system(f'mkdir {outpath}/{args.exp_name}/objs > /dev/null 2>&1') os.system(f'mkdir {outpath}/{args.exp_name}/objs/train > /dev/null 2>&1') os.system(f'mkdir {outpath}/{args.exp_name}/objs/val > /dev/null 2>&1') os.system(f'mkdir {outpath}/{args.exp_name}/objs/gen > /dev/null 2>&1') os.system(f'mkdir {outpath}/{args.exp_name}/objs/gt > /dev/null 2>&1') os.system(f'mkdir {outpath}/{args.exp_name}/models > /dev/null 2>&1') with open(f"{outpath}/{args.exp_name}/config.txt", "w") as f: f.write(f"{args}\n") def get_max_sq_len(nodes): m = 0 q = [n for _,n in nodes] while len(q) > 0: n = q.pop(0) l = n['func_gt'].shape[0] m = max(m, l) for c in n['children']: if len(c) > 0: q.append(c) return m def addTargets(node, metadata): gt_funcs = node['func_gt'] gt_cparams = node['cparam_gt'] gt_dparams = node['dparam_gt'] node['bbox_gt'] = torch.from_numpy(node['cparam_gt'][1][:3]).float().to(device) seq = torch.zeros(SEQ_LEN, metadata['tl_size'], device=device).float() fprm_weight = torch.zeros(SEQ_LEN, metadata['tl_size'], device=device).float() seq_mask = torch.zeros(SEQ_LEN, device=device).float() seq_end = 0 for i, fn in enumerate(gt_funcs.tolist()): seq_mask[i] = 1.0 seq_end = i line = torch.zeros(metadata['tl_size'], device=device).float() weight = torch.zeros(metadata['tl_size'], device=device).float() line[fn] = 1.0 float_vals = [] bool_vals = [] for j, tp in enumerate(metadata['cparam_map'][fn]): if tp == 'f': float_vals.append(gt_cparams[i][j].item()) elif tp == 'b': bool_vals.append(gt_cparams[i][j].item()) else: assert False, f'bad type {tp}' if len(float_vals) > 0: start, end = metadata['tl_map'][f'{fn}_f'] line[start:end] = torch.tensor(float_vals, device=device) weight[start:end] = 1.0 if len(bool_vals) > 0: start, end = metadata['tl_map'][f'{fn}_b'] line[start:end] = torch.tensor(bool_vals, device=device) for j, prm in enumerate(metadata['dparam_map'][fn]): tar = int(gt_dparams[i][j].item()) start, end = metadata['tl_map'][f'{fn}_{prm}_{j}'] line[start+tar] = 1.0 seq[i] = line fprm_weight[i] = weight node['seq'] = seq node['fprm_weight'] = fprm_weight node['seq_end'] = torch.tensor([seq_end],device=device).long() node['seq_mask'] = seq_mask child_target = torch.zeros(SEQ_LEN, metadata['max_children'], device=device).float() child_weight = torch.zeros(SEQ_LEN, metadata['max_children'], device=device).float() child_enc_mask = torch.zeros(SEQ_LEN, device = device).float() child_to_ln_map = {} for i, inds in enumerate(node['child_gt']): for j, cind in enumerate(inds): child_weight[i, j] = 1.0 child_to_ln_map[cind] = (i, j) if len(node['children'][cind]) > 0: child_target[i,j] = 1.0 child_enc_mask[i] = 1.0 node['child_target'] = child_target node['child_weight'] = child_weight node['child_to_ln_map'] = child_to_ln_map node['child_enc_mask'] = child_enc_mask node['exp_inds'] = [] for i, child in enumerate(node['children']): if len(child) > 0: node['exp_inds'].append(i) addTargets(child, metadata) gt_cparams = node['cparam_gt'] gt_dparams = node['dparam_gt'] bool_target = torch.zeros(gt_funcs.shape[0], metadata['max_cparams'], device=device) bool_mask = torch.zeros(gt_funcs.shape[0], metadata['max_cparams'], device=device) float_target = torch.zeros(gt_funcs.shape[0], metadata['max_cparams'], device=device) float_mask = torch.zeros(gt_funcs.shape[0], metadata['max_cparams'], device=device) for i, tf in enumerate(gt_funcs): for j, tp in enumerate(metadata['cparam_map'][tf]): if tp == 'f': float_target[i][j] = gt_cparams[i][j].item() float_mask[i][j] = 1.0 elif tp == 'b': bool_target[i][j] = gt_cparams[i][j].item() bool_mask[i][j] = 1.0 else: assert False, f'bad type {tp}' disc_target = torch.zeros(gt_funcs.shape[0], metadata['max_dparams'], device=device).long() disc_mask = torch.zeros(gt_funcs.shape[0], metadata['max_dparams'], device=device) for i, tf in enumerate(gt_funcs): for j, _ in enumerate(metadata['dparam_map'][tf]): disc_target[i][j] = gt_dparams[i][j].item() disc_mask[i][j] = 1.0 node['e_func_gt'] = torch.tensor(node['func_gt'], device=device).long() node['e_cparam_gt'] = torch.tensor(node['cparam_gt'], device=device).float() node['e_dparam_gt'] = torch.tensor(node['dparam_gt'], device=device).long() node['e_bool_target'] = bool_target.float() node['e_float_target'] = float_target.float() node['e_disc_target'] = disc_target.long() node['e_bool_mask'] = bool_mask.float() node['e_float_mask'] = float_mask.float() node['e_disc_mask'] = disc_mask.float() def _col(samples): return samples def _bcol(samples): progs = [s[1] for s in samples] return progs # Full encoder + decoder training logic for a single program (i.e. a batch) def model_train(batch, encoder, decoder, enc_opt, dec_opt, loss_config): codes, kl_loss = get_encodings(batch, encoder, 'kl' in loss_config) shape_result = decoder.train_progs(batch, codes, loss_config) if 'kl' in loss_config: shape_result['kl'] = kl_loss loss = 0. for key in loss_config: loss += (loss_config[key] * shape_result[key]) / len(batch) shape_result[key] = (loss_config[key] * shape_result[key].item()) / len(batch) if torch.is_tensor(loss) and enc_opt is not None and dec_opt is not None: dec_opt.zero_grad() enc_opt.zero_grad() loss.backward() dec_opt.step() enc_opt.step() shape_result['loss'] = loss.item() return shape_result def model_train_results(dataset, encoder, decoder, enc_opt, dec_opt, loss_config): if enc_opt is not None and dec_opt is not None: decoder.train() encoder.train() else: decoder.eval() encoder.eval() ep_result = {} bc = 0. for batch in dataset: bc += 1. batch_result = model_train( batch, encoder, decoder, dec_opt, enc_opt, loss_config ) for key in batch_result: if key not in ep_result: ep_result[key] = batch_result[key] else: ep_result[key] += batch_result[key] ep_result['batch_count'] = bc return ep_result def model_eval( eval_train_dataset, eval_val_dataset, encoder, decoder, exp_name, epoch, num_write, metadata ): eval_results = {} for name, dataset in [('train', eval_train_dataset), ('val', eval_val_dataset)]: if len(dataset) == 0: eval_results[name] = {} continue named_results = { 'count': 0., 'miss_hier_prog': 0., 'no_norm': 1.0 } recon_sets = [] for batch in dataset: assert len(batch) == 1, 'batch size 1' shape = batch[0] named_results['count'] += 1. # Always get maximum likelihood estimation (i.e. mean) of shape encoding at eval time code, _ = get_encodings([shape[1]], encoder, variational=False) if USE_SEM_VALID: node = sv.sem_eval_prog(decoder, code.squeeze()) else: node = decoder.eval_prog(code.squeeze()) try: shape_result = getShapeEvalResult(node, shape[1]) shape_result['bbox'] = (node['bb_dims'] - shape[1]['bbox_gt']).abs().sum().item() except Exception as e: if VERBOSE: print(f"FAILED SHAPE EVAL RESULT WITH {e}") shape_result = {} for key in shape_result: if key not in named_results: named_results[key] = shape_result[key] else: named_results[key] += shape_result[key] try: fillProgram( metadata['dsl'], node, metadata, FUNC_PRED_FIELD, CPARAM_PRED_FIELD, DPARAM_PRED_FIELD, ) recon_sets.append((node, shape[1], shape[0])) except Exception as e: if VERBOSE: print(f"Failed Recon Program with {e}") named_results[f'miss_hier_prog'] += 1. # For reconstruction, get metric performance recon_results, recon_misses = recon_metrics.recon_metrics( recon_sets, outpath, exp_name, name, epoch, VERBOSE, num_write + 1 ) for key in recon_results: named_results[key] = recon_results[key] named_results[f'miss_hier_prog'] += recon_misses named_results[f'prog_creation_perc'] = ( named_results[f'count'] - named_results[f'miss_hier_prog'] ) / named_results[f'count'] eval_results[name] = named_results return eval_results def model_gen(decoder, num_gen, metadata, exp_name, epoch, num_write, train_samps, val_samps): gen_progs = [] gen_prog_fails = 0. # Also generate a set of unconditional ShapeAssembly Programs for i in range(0, num_gen): code = torch.randn(args.hidden_dim).to(device) if USE_SEM_VALID: node = sv.sem_eval_prog(decoder, code.squeeze()) else: node = decoder.eval_prog(code.squeeze()) try: fillProgram( metadata['dsl'], node, metadata, FUNC_PRED_FIELD, CPARAM_PRED_FIELD, DPARAM_PRED_FIELD, ) gen_progs.append(node) except Exception as e: gen_prog_fails += 1. if VERBOSE: print(f"Failed generating new program with {e}") # Get metrics for unconditional generations gen_results, gen_misses = gen_metrics.gen_metrics( gen_progs, outpath, exp_name, epoch, VERBOSE, num_write, train_samps, val_samps ) gen_results['prog_creation_perc'] = (num_gen - gen_misses - gen_prog_fails) / (num_gen + 1e-8) return gen_results def print_train_results(result, exp_name): res = "" for name, key, norm_key in TRAIN_LOG_INFO: if key in result: res += f" {name} : {round(result[key] / (result[norm_key]+1e-8), 2)}\n" utils.log_print(res, f"{outpath}/{exp_name}/log.txt") def print_eval_results(result, exp_name): res = "" for name, key, norm_key in EVAL_LOG_INFO: if key in result: res += f" {name} : {round(result[key] / (result[norm_key]+1e-8), 2)}\n" utils.log_print(res, f"{outpath}/{exp_name}/log.txt") def print_gen_results(result, exp_name): res = "" for name, key in GEN_LOG_INFO: if key in result: res += f" {name} : {round(result[key], 2)}\n" utils.log_print(res, f"{outpath}/{exp_name}/log.txt") def make_train_plots(train_result, val_result, train_plots, aepochs, exp_name): for name, key, norm_key in TRAIN_LOG_INFO: for rname, result in [('train', train_result), ('val', val_result)]: if key not in result: continue res = result[key] / (result[norm_key]+1e-8) if name not in train_plots[rname]: train_plots[rname][name] = [res] else: train_plots[rname][name].append(res) if name not in train_plots['train']: continue plt.clf() plt.plot(aepochs, train_plots['train'][name], label='train') if name in train_plots['val']: plt.plot(aepochs, train_plots['val'][name], label='val') plt.legend() plt.grid() plt.savefig(f"{outpath}/{exp_name}/plots/train/{name}.png") def make_eval_plots(eval_result, eval_plots, aepochs, exp_name): for name, key, norm_key in EVAL_LOG_INFO: for rname, result in [('train', eval_result['train']), ('val', eval_result['val'])]: if key not in result: continue res = result[key] / (result[norm_key]+1e-8) if not name in eval_plots[rname]: eval_plots[rname][name] = [res] else: eval_plots[rname][name].append(res) if name not in eval_plots['train']: continue plt.clf() plt.plot(aepochs, eval_plots['train'][name], label='train') if name in eval_plots['val']: plt.plot(aepochs, eval_plots['val'][name], label='val') plt.legend() plt.grid() plt.savefig(f"{outpath}/{exp_name}/plots/eval/{name}.png") def make_gen_plots(gen_result, gen_plots, aepochs, exp_name): for name, key in GEN_LOG_INFO: if key not in gen_result: continue res = gen_result[key] if not name in gen_plots: gen_plots[name] = [res] else: gen_plots[name].append(res) plt.clf() plt.plot(aepochs, gen_plots[name]) plt.grid() plt.savefig(f"{outpath}/{exp_name}/plots/gen/{name}.png") # Helper function for keeping consistent train/val splits def getInds(train_ind_file): inds = set() with open(train_ind_file) as f: for line in f: inds.add(line.strip()) return inds def get_tensor_layout(metadata): # Figure out size of tensor # map from (func, type) -> indices in tensor tl_map = {} size = len(metadata['func_map']) tl_map['func'] = (0, size) start = size size += (I_LENGTH * metadata['max_d_i_params']) \ + (SQ_LENGTH * metadata['max_d_sq_params']) \ + (SYM_LENGTH * metadata['max_d_sym_params']) for func, prms in metadata['dparam_map'].items(): i_start = start sq_start = i_start + I_LENGTH * metadata['max_d_i_params'] sym_start = sq_start + SQ_LENGTH * metadata['max_d_sq_params'] for i, _typ in enumerate(prms): if _typ == 'i': opt_len = I_LENGTH _start = i_start i_start += opt_len elif _typ == 'sq': opt_len = SQ_LENGTH _start = sq_start sq_start += opt_len elif _typ == 'sym': opt_len = SYM_LENGTH _start = sym_start sym_start += opt_len tl_map[f'{func}_{_typ}_{i}'] = ( _start, _start + opt_len, ) for func, prms in metadata['cparam_map'].items(): nf = 0 nb = 0 for prm in prms: if 'f' in prm: nf += 1 elif 'b' in prm: nb += 1 if nf > 0: size += nf tl_map[f'{func}_f'] = (size-nf, size) if nb > 0: size += nb tl_map[f'{func}_b'] = (size-nb, size) return size, tl_map # Main entry-point of modeling logic def run_train( dataset_path, exp_name, max_shapes, epochs, hidden_dim, eval_per, variational, loss_config, rd_seed, print_per, num_gen, num_write, num_eval, dec_lr, enc_lr, gen_per, save_per, category, batch_size ): random.seed(rd_seed) numpy.random.seed(rd_seed) torch.manual_seed(rd_seed) raw_data = pickle.load(open(f"{dataset_path}_train.data", "rb")) metadata = pickle.load(open(f"{dataset_path}_train.meta", "rb")) metadata['max_children'] = max([int(i) for i in metadata['num_cube_map'].values()]) metadata['rev_func_map'] = {v:k for k, v in metadata['func_map'].items()} for key in ('i', 'sq', 'sym'): metadata[f'max_d_{key}_params'] = max([len([ __l for __l in _l if __l == key]) for _l in metadata['dparam_map'].values()]) tl_size, tl_map = get_tensor_layout(metadata) metadata['tl_size'] = tl_size metadata['tl_map'] = tl_map all_inds = [] all_data = [] good_inds = [] max_sq_len = get_max_sq_len(raw_data) print(f"Seq len: {max_sq_len}") global SEQ_LEN SEQ_LEN = max_sq_len for d in raw_data: if len(all_inds) >= max_shapes: break if len(good_inds) > 0 and d[0] not in good_inds: continue addTargets(d[1], metadata) fillProgram( metadata['dsl'], d[1], metadata, 'func_gt', 'cparam_gt', 'dparam_gt' ) all_data.append(d[1]) all_inds.append(d[0]) inds_and_progs = list(zip(all_inds, all_data)) samples = [(i, d) for i,d in inds_and_progs] train_ind_file = f'data_splits/{category}/train.txt' val_ind_file = f'data_splits/{category}/val.txt' train_samples = [] val_samples = [] train_inds = getInds(train_ind_file) val_inds = getInds(val_ind_file) misses = 0. num_parts = [] for (ind, prog) in samples: if ind in train_inds or ind in good_inds: train_samples.append((ind, prog)) elif ind in val_inds: val_samples.append((ind, prog)) else: misses += 1 if len(good_inds) > 0: val_samples = train_samples[:1] print(f"Samples missed: {misses}") train_num = len(train_samples) val_num = len(val_samples) train_surf_samples, val_surf_samples = gen_metrics.get_gt_surf_samples(category, num_gen) train_dataset = DataLoader( train_samples, batch_size, shuffle=True, collate_fn = _bcol ) eval_train_dataset = DataLoader( train_samples[:num_eval], 1, shuffle=False, collate_fn = _col ) val_dataset = DataLoader( val_samples, batch_size, shuffle = False, collate_fn = _bcol ) eval_val_dataset = DataLoader( val_samples[:num_eval//2], 1, shuffle = False, collate_fn = _col ) utils.log_print(f"Training size: {len(train_samples)}", f"{outpath}/{exp_name}/log.txt") utils.log_print(f"Validation size: {len(val_samples)}", f"{outpath}/{exp_name}/log.txt") val_epochs = [] train_epochs = [] gen_epochs = [] train_plots = {'train': {}, 'val': {}} eval_plots = {'train': {}, 'val': {}} gen_plots = {} encoder = dslEncoder(hidden_dim, metadata) decoder = dslDecoder( hidden_dim, metadata, ) encoder.to(device) decoder.to(device) dec_opt = torch.optim.Adam( decoder.parameters(), lr = dec_lr, eps = 1e-6 ) enc_opt = torch.optim.Adam( encoder.parameters(), lr = enc_lr, eps = 1e-6 ) print('training ...') for e in range(0, epochs): json.dump({ 'train': train_plots, 'eval': eval_plots, 'gen': gen_plots, 'train_epochs': train_epochs, 'val_epochs': val_epochs, 'gen_epochs': gen_epochs }, open(f"{outpath}/{exp_name}/res.json" ,'w')) decoder.epoch = e do_print = (e+1) % print_per == 0 t = time.time() if do_print: utils.log_print(f"\nEpoch {e}:", f"{outpath}/{exp_name}/log.txt") train_result = model_train_results( train_dataset, encoder, decoder, enc_opt, dec_opt, loss_config, ) if do_print: with torch.no_grad(): val_result = model_train_results( val_dataset, encoder, decoder, None, None, {k:v for k,v in loss_config.items() if k != 'kl'} ) train_epochs.append(e) make_train_plots(train_result, val_result, train_plots, train_epochs, exp_name) utils.log_print( f"Train results: ", f"{outpath}/{exp_name}/log.txt" ) print_train_results(train_result, exp_name) utils.log_print( f"Val results: ", f"{outpath}/{exp_name}/log.txt" ) print_train_results(val_result, exp_name) utils.log_print( f" Time = {time.time() - t}", f"{outpath}/{exp_name}/log.txt" ) with torch.no_grad(): if (e+1) % eval_per == 0: decoder.eval() encoder.eval() t = time.time() eval_results = model_eval( eval_train_dataset, eval_val_dataset, encoder, decoder, exp_name, e, num_write, metadata ) utils.log_print(f"Evaluation training set results:", f"{outpath}/{exp_name}/log.txt") print_eval_results(eval_results['train'], exp_name) utils.log_print(f"Evaluation validation set results:", f"{outpath}/{exp_name}/log.txt") print_eval_results(eval_results['val'], exp_name) utils.log_print(f"Eval Time = {time.time() - t}", f"{outpath}/{exp_name}/log.txt") val_epochs.append(e) make_eval_plots(eval_results, eval_plots, val_epochs, exp_name) if (e+1) % gen_per == 0: gen_epochs.append(e) t = time.time() gen_result = model_gen( decoder, num_gen, metadata, exp_name, e, num_write, train_surf_samples, val_surf_samples ) utils.log_print(f"Unconditional Generated Results:", f"{outpath}/{exp_name}/log.txt") print_gen_results(gen_result, exp_name) make_gen_plots(gen_result, gen_plots, gen_epochs, exp_name) utils.log_print(f"Gen Time = {time.time() - t}", f"{outpath}/{exp_name}/log.txt") if (e+1) % save_per == 0: utils.log_print("Saving Models", f"{outpath}/{exp_name}/log.txt") torch.save(decoder.state_dict(), f"{outpath}/{exp_name}/models/decoder_{e}.pt") torch.save(encoder.state_dict(), f"{outpath}/{exp_name}/models/encoder_{e}.pt") def make_gens(args): metadata = pickle.load(open(f"{args.dataset_path}_train.meta", "rb")) metadata['max_children'] = max([int(i) for i in metadata['num_cube_map'].values()]) metadata['rev_func_map'] = {v:k for k, v in metadata['func_map'].items()} for key in ('i', 'sq', 'sym'): metadata[f'max_d_{key}_params'] = max([len([ __l for __l in _l if __l == key]) for _l in metadata['dparam_map'].values()]) tl_size, tl_map = get_tensor_layout(metadata) metadata['tl_size'] = tl_size metadata['tl_map'] = tl_map decoder = dslDecoder( args.hidden_dim, metadata, ) decoder.load_state_dict(torch.load( f'{args.exp_name}/models/decoder_{args.load_epoch}.pt' )) decoder.to(device) decoder.eval() i = 0 outname = f'{args.exp_name}/gen_output' os.system(f'mkdir {outname}') while (i < args.num_gen): code = torch.randn(args.hidden_dim).to(device) node = sv.sem_eval_prog(decoder, code) try: fillProgram( metadata['dsl'], node, metadata, FUNC_PRED_FIELD, CPARAM_PRED_FIELD, DPARAM_PRED_FIELD ) sa_lines = utils.sagetHierProgLines(node) od_lines = utils.getHierProgLines(node, 'dsl_prog') verts, faces = hier_execute(node) utils.writeObj(verts, faces, f"{outname}/gen_{i}.obj") with open(f'{outname}/prog_sa_{i}.txt', 'w') as f: for line in sa_lines: f.write(f'{line}\n') with open(f'{outname}/prog_od_{i}.txt', 'w') as f: for line in od_lines: f.write(f'{line}\n') i += 1 except Exception as e: print(f"failed {i} with {e}") if __name__ == '__main__': parser = argparse.ArgumentParser(description="Run Novel Shape Generation model") parser.add_argument('-ds', '--dataset_path', help='Path to program data, e.g. data/shapemod_chair', type = str) parser.add_argument('-en', '--exp_name', help='name of experiment', type = str) parser.add_argument('-c', '--category', type = str, help = 'category of PartNet') parser.add_argument('-ms', '--max_shapes', default = 100000, type = int, help = 'max number of shapes to train/evaluate on ') parser.add_argument('-e', '--epochs', default = 5000, type = int, help = 'number of epochs to run for') parser.add_argument('-hd', '--hidden_dim', default = 256, type = int, help = 'hidden dimension size') parser.add_argument('-prp', '--print_per', default = 20, type = int, help = 'how often to print out training set statistics') parser.add_argument('-evp', '--eval_per', default = 100, type = int, help = 'how often to run evaluation statistics') parser.add_argument('-gp', '--gen_per', default = 100, type = int, help = 'how often to run novel generation statistics') parser.add_argument('-v', '--variational', default = "True", type = str, help = 'whether model should be a VAE or AE') parser.add_argument('-enc_lr', '--enc_lr', default = 0.0002, type = float, help = 'encoder learning rate') parser.add_argument('-dec_lr', '--dec_lr', default = 0.0002, type = float, help = 'decoder learning rate') parser.add_argument('-rd', '--rd_seed', default = 42, type = int, help = 'random seed') parser.add_argument('-ng', '--num_gen', default = 1000, type = int, help = 'number of shapes to generate each generation period') parser.add_argument('-nw', '--num_write', default = 25, type = int, help = 'number of shapes to write to .obj each evaluation period') parser.add_argument('-ne', '--num_eval', default = 200, type = int, help = 'number of shapes to run eval statistics on each evaluation period') parser.add_argument('-f_lw', '--f_lw', default = 50., type = float, help = 'weight on loss of continuous parameters') parser.add_argument('-d_lw', '--d_lw', default = 1., type = float, help = 'weight on loss of discrete parameters') parser.add_argument('-b_lw', '--b_lw', default = 1., type = float, help = 'weight on loss of boolean parameters') parser.add_argument('-c_lw', '--c_lw', default = 1., type = float, help = 'weight on loss of child predictions') parser.add_argument('-fn_lw', '--fn_lw', default = 1., type = float, help = 'weight on loss of function predictions') parser.add_argument('-kl_lw', '--kl_lw', default = 0.1, type = float, help = 'KL divergence loss weight') parser.add_argument('-b', '--batch_size', default = 64, type=int, help = 'batch size') parser.add_argument('-le', '--load_epoch', default = None, type=int, help = 'model epoch to load from pre-trained model') parser.add_argument('-m', '--mode', default = 'train', type=str, help = 'whether to train new model or generate samples from pre-trained model') args = parser.parse_args() args.variational = ast.literal_eval(args.variational) loss_config = { 'd_prm': args.d_lw, 'f_prm': args.f_lw, 'bbox': args.f_lw, 'b_prm': args.b_lw, 'child': args.c_lw, 'func': args.fn_lw, } if args.variational: loss_config['kl'] = args.kl_lw if args.mode == 'train': writeConfigFile(args) run_train( dataset_path=args.dataset_path, exp_name=args.exp_name, max_shapes=args.max_shapes, epochs=args.epochs, hidden_dim=args.hidden_dim, eval_per=args.eval_per, variational=args.variational, loss_config=loss_config, rd_seed=args.rd_seed, print_per=args.print_per, num_gen=args.num_gen, num_write=args.num_write, num_eval=args.num_eval, enc_lr=args.enc_lr, dec_lr=args.dec_lr, gen_per=args.gen_per, save_per=args.gen_per, category=args.category, batch_size=args.batch_size ) elif args.mode == 'gen': make_gens(args) else: print(f"Bad mode {args.mode}")
61,405
34.826138
167
py
ShapeMOD
ShapeMOD-main/SA_lang/tasks/recon_metrics.py
from ShapeAssembly import hier_execute import sa_utils as utils import torch import os import sys import math import faiss import numpy as np device = torch.device("cuda") NUM_SAMPS = 10000 V_DIM = 64 CD_MULT = 500. voxel_inds = ((np.indices((V_DIM, V_DIM, V_DIM)).T + .5) / (V_DIM//2)) -1. flat_voxel_inds = torch.from_numpy(voxel_inds.reshape(-1, 3)).float() class SimpChamferLoss(torch.nn.Module): def __init__(self, device): super(SimpChamferLoss, self).__init__() self.dimension = 3 self.gpu_id = torch.cuda.current_device() self.res = faiss.StandardGpuResources() def build_nn_index(self, database): """ :param database: numpy array of Nx3 :return: Faiss index, in CPU """ index_cpu = faiss.IndexFlatL2(self.dimension) index = faiss.index_cpu_to_gpu(self.res, self.gpu_id, index_cpu) index.add(database) return index def search_nn(self, index, query, k): D, I = index.search(query, 1) return np.sqrt(D) def getAvgDist(self, index, query): D, I = index.search(query, 2) m_d = math.sqrt(np.percentile(D[:,1],90)) return m_d def calc_metrics(self, predict_pc, gt_pc, threshes): """ :param predict_pc: Bx3xM Variable in GPU :param gt_pc: Bx3xN Variable in GPU :return: """ predict_pc_size = predict_pc.size() gt_pc_size = gt_pc.size() predict_pc_np = np.ascontiguousarray( torch.transpose(predict_pc.data.clone(), 1, 2).cpu().numpy() ) # BxMx3 gt_pc_np = np.ascontiguousarray( torch.transpose(gt_pc.data.clone(), 1, 2).cpu().numpy() ) # BxNx3 index_predict = self.build_nn_index(predict_pc_np[0]) index_gt = self.build_nn_index(gt_pc_np[0]) fwd_dist = self.search_nn(index_gt, predict_pc_np[0], 1) bwd_dist = self.search_nn(index_predict, gt_pc_np[0], 1) cd = (fwd_dist.mean() / 2) + (bwd_dist.mean() / 2) ones = np.ones(fwd_dist.shape) fscores = [] for thresh in threshes: if thresh == 'def': thresh = self.getAvgDist(index_gt, gt_pc_np[0]) precision = (100 / ones.shape[0]) * np.sum(ones[fwd_dist <= thresh]) recall = (100 / ones.shape[0]) * np.sum(ones[bwd_dist <= thresh]) fs = (2*precision*recall) / (precision + recall + 1e-8) fscores.append(fs) return [cd] + fscores class SimpCPUChamferLoss(torch.nn.Module): def __init__(self): super(SimpCPUChamferLoss, self).__init__() self.dimension = 3 def build_nn_index(self, database): """ :param database: numpy array of Nx3 :return: Faiss index, in CPU """ index = faiss.IndexFlatL2(self.dimension) index.add(database) return index def search_nn(self, index, query, k): D, I = index.search(query, 1) return np.sqrt(D) def getAvgDist(self, index, query): D, I = index.search(query, 2) m_d = math.sqrt(np.percentile(D[:,1],90)) return m_d def calc_metrics(self, predict_pc, gt_pc, threshes): """ :param predict_pc: Bx3xM Variable in GPU :param gt_pc: Bx3xN Variable in GPU :return: """ predict_pc_size = predict_pc.size() gt_pc_size = gt_pc.size() predict_pc_np = np.ascontiguousarray( torch.transpose(predict_pc.data.clone(), 1, 2).cpu().numpy() ) # BxMx3 gt_pc_np = np.ascontiguousarray( torch.transpose(gt_pc.data.clone(), 1, 2).cpu().numpy() ) # BxNx3 index_predict = self.build_nn_index(predict_pc_np[0]) index_gt = self.build_nn_index(gt_pc_np[0]) fwd_dist = self.search_nn(index_gt, predict_pc_np[0], 1) bwd_dist = self.search_nn(index_predict, gt_pc_np[0], 1) cd = (fwd_dist.mean() / 2) + (bwd_dist.mean() / 2) ones = np.ones(fwd_dist.shape) fscores = [] for thresh in threshes: if thresh == 'def': thresh = self.getAvgDist(index_gt, gt_pc_np[0]) precision = (100 / ones.shape[0]) * np.sum(ones[fwd_dist <= thresh]) recall = (100 / ones.shape[0]) * np.sum(ones[bwd_dist <= thresh]) fs = (2*precision*recall) / (precision + recall + 1e-8) fscores.append(fs) return [cd] + fscores chamfer = SimpCPUChamferLoss() def voxelize(cube_params): rotmat = torch.stack(( cube_params[6:9], cube_params[9:12], cube_params[12:15] )) cent_pts = flat_voxel_inds - cube_params[3:6].unsqueeze(0) rot_pts = (rotmat @ cent_pts.T).T cube_sdfs = ( rot_pts.abs() - (cube_params[:3].unsqueeze(0) * .5) ).max(dim=1).values return (cube_sdfs <= (1./V_DIM)).bool() def shape_voxelize(cubes): voxels = torch.zeros(64**3).bool().to(flat_voxel_inds.device) for cube in cubes: vox_occs = voxelize(cube.getParams().to(flat_voxel_inds.device)) voxels = voxels | vox_occs return voxels def getSampMetrics(verts, faces, gt_verts, gt_faces): p_samps = torch.clamp( utils.sample_surface(faces, verts.unsqueeze(0), NUM_SAMPS, False)[0], -1, 1) t_samps = torch.clamp( utils.sample_surface(gt_faces, gt_verts.unsqueeze(0), NUM_SAMPS, False)[0], -1, 1) samp_metrics = chamfer.calc_metrics( p_samps.T.unsqueeze(0).float(), t_samps.T.unsqueeze(0).float(), [.05, .03, 'def'] ) return { 'cd': samp_metrics[0] * CD_MULT, 'fscore-05': samp_metrics[1], 'fscore-03': samp_metrics[2], 'fscore-def': samp_metrics[3], } def getShapeIoU(cubes, gt_cubes): pvoxels = shape_voxelize(cubes) tvoxels = shape_voxelize(gt_cubes) iou = 100 * ( (pvoxels & tvoxels).sum().item() / (pvoxels | tvoxels).sum().item() ) return iou def recon_metrics( recon_sets, outpath, exp_name, name, epoch, VERBOSE, num_gen ): misses = 0. results = { 'iou': [], 'cd': [], 'fscore-def': [], 'fscore-03': [], 'fscore-05': [], } count = 0 for prog, gt_prog, prog_ind in recon_sets: gt_verts, gt_faces, gt_cubes = hier_execute(gt_prog, return_all = True) try: verts, faces, cubes = hier_execute(prog, return_all = True) assert not torch.isnan(verts).any(), 'saw nan vert' except Exception as e: misses += 1. if VERBOSE: print(f"failed recon metrics for {prog_ind} with {e}") continue gt_objs = os.listdir(f"{outpath}/{exp_name}/objs/gt/") try: sm = getSampMetrics(verts, faces, gt_verts, gt_faces) for k, v in sm.items(): if v is not None: results[k].append(v) except Exception as e: if VERBOSE: print(f"failed Samp Metrics for {prog_ind} with {e}") try: iou = getShapeIoU(cubes, gt_cubes) if iou is not None: results['iou'].append(iou) except Exception as e: if VERBOSE: print(f"failed Shape Iou for {prog_ind} with {e}") if count >= num_gen: continue if f"{prog_ind}.obj" not in gt_objs: utils.writeObj(gt_verts, gt_faces, f"{outpath}/{exp_name}/objs/gt/{prog_ind}.obj") if 'dsl_prog' in gt_prog: utils.writeHierProg(gt_prog, 'dsl_prog', f"{outpath}/{exp_name}/programs/gt/{prog_ind}.txt") else: utils.sawriteHierProg(gt_prog, f"{outpath}/{exp_name}/programs/gt/{prog_ind}.txt") try: utils.writeObj( verts, faces, f"{outpath}/{exp_name}/objs/{name}/{epoch}_{prog_ind}.obj" ) if 'dsl_prog' in prog: utils.writeHierProg( prog, 'dsl_prog', f"{outpath}/{exp_name}/programs/{name}/{epoch}_{prog_ind}.txt" ) else: utils.sawriteHierProg( prog, f"{outpath}/{exp_name}/programs/{name}/{epoch}_{prog_ind}.txt" ) count += 1 except Exception as e: if VERBOSE: print(f"Failed writing prog/obj for {prog_ind} with {e}") for key in results: if len(results[key]) > 0: res = torch.tensor(results[key]).mean().item() else: res = 0. results[key] = res return results, misses def simp_recon_metrics( recon_sets ): results = { 'iou': [], 'cd': [], 'fscore-def': [], 'fscore-03': [], 'fscore-05': [], } for prog, gt_prog in recon_sets: _results = {k:None for k in results} try: gt_verts, gt_faces, gt_cubes = hier_execute(gt_prog, return_all = True) verts, faces, cubes = hier_execute(prog, return_all = True) assert not torch.isnan(verts).any(), 'saw nan vert' sm = getSampMetrics(verts, faces, gt_verts, gt_faces) for k, v in sm.items(): _results[k] = v.item() iou = getShapeIoU(cubes, gt_cubes) _results['iou'] = iou except Exception as e: print(f"Faile with {e}") for k in _results: results[k].append(_results[k]) return results
9,981
30.488959
108
py
ShapeMOD
ShapeMOD-main/SA_lang/tasks/gen_metrics.py
from ShapeAssembly import hier_execute, ShapeAssembly, make_hier_prog import sa_utils as utils import torch import os import sys from valid import check_stability, check_rooted from pointnet_fd import get_fd from recon_metrics import chamfer, CD_MULT from tqdm import tqdm import faiss import numpy as np NUM_SAMPS = 2500 CD_NUM_SAMPS = 1024 device = torch.device("cuda") #device = torch.device("cpu") MAX_ROOTED_STABLE = 200 MAX_VARI = 200 sa = ShapeAssembly() class CDPairs(torch.nn.Module): def __init__(self, device, mem = 100 * 1024 * 1024): super(CDPairs, self).__init__() self.gpu_id = torch.cuda.current_device() self.res = faiss.StandardGpuResources() self.res.noTempMemory() self.res.setTempMemory(mem) self.dimension = 3 def build_nn_index(self, database): """ :param database: numpy array of Nx3 :return: Faiss index, in CPU """ index_cpu = faiss.IndexFlatL2(self.dimension) index = faiss.index_cpu_to_gpu(self.res, self.gpu_id, index_cpu) index.add(database) return index def search_nn(self, index, query, k): D, I = index.search(query, k) return np.sqrt(D) # B x D x 3 inputs def calc_cd(self, source, target, skip_same=False): source = np.ascontiguousarray(source.cpu().numpy()) target = np.ascontiguousarray(target.cpu().numpy()) d_source = [] d_target = [] for i in range(source.shape[0]): d_source.append(self.build_nn_index(source[i])) for i in range(target.shape[0]): d_target.append(self.build_nn_index(target[i])) min_dists = [] for i in tqdm(range(source.shape[0])): min_dist = 1e8 for j in range(target.shape[0]): if i == j and skip_same: continue fwd_dist = self.search_nn(d_target[j], source[i], 1) bwd_dist = self.search_nn(d_source[i], target[j], 1) dist = fwd_dist.mean() + bwd_dist.mean() if dist < min_dist: min_dist = dist min_dists.append(min_dist) return torch.tensor(min_dists).mean().item() cdpairs = CDPairs(device) def getGTSamples(progs, num): data = [] for prog in tqdm(progs[:num]): verts, faces = hier_execute(prog) verts = verts.to(device) faces = faces.to(device) try: samps = utils.sample_surface(faces, verts.unsqueeze(0), NUM_SAMPS, False).squeeze() data.append(samps) except Exception as e: print(f"Failed GT samples with {e}") return data def getSamples(meshes, VERBOSE, num_samps = NUM_SAMPS): data = [] for verts, faces in tqdm(meshes): verts = verts.to(device) faces = faces.to(device) try: samps = utils.sample_surface(faces, verts.unsqueeze(0), num_samps, False).squeeze() data.append(samps) except Exception as e: if VERBOSE: print(f"couldn't sample with {e}") return data # Helper function for keeping consistent train/val splits def getInds(train_ind_file): inds = set() with open(train_ind_file) as f: for line in f: inds.add(line.strip()) return inds def get_gt_surf_samples(category, num = 1000): train_ind_file = f'data_splits/{category}/train.txt' val_ind_file = f'data_splits/{category}/val.txt' os.system('mkdir gen_comps') os.system('mkdir gen_comps/gt_val') os.system('mkdir gen_comps/gt_train/') os.system(f'mkdir gen_comps/gt_val/{category}') os.system(f'mkdir gen_comps/gt_train/{category}') gt_train_files = set(os.listdir(f'gen_comps/gt_train/{category}/')) gt_val_files = set(os.listdir(f'gen_comps/gt_val/{category}/')) train_samples = [] val_samples = [] train_inds = list(getInds(train_ind_file)) val_inds = list(getInds(val_ind_file)) train_samples = [] val_samples = [] for ind in tqdm(train_inds[:num]): if ind+'.obj' in gt_train_files: verts, faces = utils.loadObj(f'gen_comps/gt_train/{category}/{ind}.obj') verts = torch.tensor(verts).float().to(device) faces = torch.tensor(faces).long().to(device) else: lines = sa.load_lines(f'data/{category}/{ind}.txt') hier_prog = make_hier_prog(lines) verts, faces = hier_execute(hier_prog) utils.writeObj(verts, faces, f'gen_comps/gt_train/{category}/{ind}.obj') train_samples.append((verts, faces)) for ind in tqdm(val_inds[:num]): if ind+'.obj' in gt_val_files: verts, faces = utils.loadObj(f'gen_comps/gt_val/{category}/{ind}.obj') verts = torch.tensor(verts).float().to(device) faces = torch.tensor(faces).long().to(device) else: lines = sa.load_lines(f'data/{category}/{ind}.txt') hier_prog = make_hier_prog(lines) verts, faces = hier_execute(hier_prog) utils.writeObj(verts, faces, f'gen_comps/gt_val/{category}/{ind}.obj') val_samples.append((verts, faces)) train_surf_samples = getSamples( train_samples[:num], True ) val_surf_samples = getSamples( val_samples[:num], True ) return train_surf_samples, val_surf_samples def getMinDist(p, samples): min_dist = 1e8 for s in samples: dist = chamfer.calc_metrics( p.squeeze().T.unsqueeze(0).cpu(), s.squeeze().T.unsqueeze(0).cpu(), [] )[0] min_dist = min(dist.item(), min_dist) return min_dist def gen_metrics( gen_progs, outpath, exp_name, epoch, VERBOSE, num_write, train_samps, val_samps ): misses = 0. results = { 'num_parts': [], 'rootedness': [], 'stability': [], 'gen': [], 'cov': [], 'var': [], } samples = [] print("DOING ROOTED AND STABLE") for i, prog in enumerate(gen_progs): try: verts, faces = hier_execute(prog) assert not torch.isnan(verts).any(), 'saw nan vert' if i < num_write: utils.writeObj(verts, faces, f"{outpath}/{exp_name}/objs/gen/{epoch}_{i}.obj") if 'dsl_prog' in prog: utils.writeHierProg(prog, 'dsl_prog', f"{outpath}/{exp_name}/programs/gen/{epoch}_{i}.txt") else: utils.sawriteHierProg(prog, f"{outpath}/{exp_name}/programs/gen/{epoch}_{i}.txt") results['num_parts'].append(verts.shape[0] / 8.0) samples.append((verts, faces)) except Exception as e: misses += 1. if VERBOSE: print(f"failed gen metrics for {i} with {e}") continue if i < MAX_ROOTED_STABLE: try: if check_rooted(verts, faces): results['rootedness'].append(1.) else: results['rootedness'].append(0.) if check_stability(verts, faces): results['stability'].append(1.) else: results['stability'].append(0.) except Exception as e: if VERBOSE: print(f"failed rooted/stable with {e}") for key in results: if len(results[key]) > 0: res = torch.tensor(results[key]).mean().item() else: res = 0. results[key] = res print("GETTING SAMPLES") gen_samps = getSamples(samples, VERBOSE) try: assert len(gen_samps) > 0, 'no gen samps' print("CALC GEN") gen = cdpairs.calc_cd( torch.stack([g[:CD_NUM_SAMPS] for g in gen_samps[:MAX_VARI]]), torch.stack([g[:CD_NUM_SAMPS] for g in train_samps[:MAX_VARI]]), False ) print("CALC COV") cov = cdpairs.calc_cd( torch.stack([g[:CD_NUM_SAMPS] for g in val_samps[:MAX_VARI]]), torch.stack([g[:CD_NUM_SAMPS] for g in gen_samps[:MAX_VARI]]), False ) print("CALC VAR") var = cdpairs.calc_cd( torch.stack([g[:CD_NUM_SAMPS] for g in gen_samps[:MAX_VARI]]), torch.stack([g[:CD_NUM_SAMPS] for g in gen_samps[:MAX_VARI]]), True ) results['gen'] = gen results['cov'] = cov results['var'] = var except Exception as e: results['gen'] = 1.0 results['cov'] = 1.0 results['var'] = 1.0 if VERBOSE: print(f"failed NN comparisons with {e}") try: results['val_fd'] = get_fd( [g.cpu() for g in gen_samps], [v.cpu() for v in val_samps], None) except Exception as e: results['val_fd'] = 100. if VERBOSE: print(f"failed getting val variance with {e}") try: results['train_fd'] = get_fd( [g.cpu() for g in gen_samps], [t.cpu() for t in train_samps], None) except Exception as e: results['train_fd'] = 100. if VERBOSE: print(f"failed getting train variance with {e}") return results, misses if __name__ == '__main__': from ShapeAssembly import ShapeAssembly, make_hier_prog, hier_execute import sys import time sa = ShapeAssembly() meshes = [] for ind in list(os.listdir('data/chair'))[:int(sys.argv[1])]: lines = sa.load_lines(f'data/chair/{ind}') hier_prog = make_hier_prog(lines) verts, faces = hier_execute(hier_prog) meshes.append((verts, faces)) gen_samps = getSamples(meshes, True) t = time.time() gen = cdpairs.calc_cd(torch.stack(gen_samps), torch.stack(gen_samps), True) print(gen) print(time.time() - t)
10,126
30.548287
111
py
ShapeMOD
ShapeMOD-main/SA_lang/tasks/valid.py
import argparse from functools import reduce import numpy as np import os import trimesh as tm from trimesh.collision import CollisionManager from trimesh.creation import box import pickle from tqdm import tqdm import pybullet as p import pybullet_data from trimesh.util import concatenate as meshconcat import xml.etree.ElementTree as xml import json import shutil import sa_utils as utils from sa_utils import sample_surface import torch import time import random import string def check_rooted(verts, faces): # Load up the mesh mesh = tm.Trimesh(vertices=verts, faces=faces) # Switch from y-up to z-up mesh.vertices = mesh.vertices[:, [0, 2, 1]] mesh.fix_normals() # Find the height of the ground plane z_ground = mesh.bounds[0][2] # Extract the individual cuboid parts comps = mesh.split().tolist() # Also create a thin box for the ground plane ground = box( extents = [10, 10, 0.01], transform = [ [1, 0, 0, 0], [0, 1, 0, 0], [0, 0, 1, z_ground - 0.01/2], [0, 0, 0, 1] ] ) comps.insert(0, ground) # Detect (approximate) intersections between parts # collision_dist = 0.005 * mesh.scale # collision_dist = 0.01 * mesh.scale collision_dist = 0.02 * mesh.scale adjacencies = {comp_index : [] for comp_index in range(len(comps))} manager = CollisionManager() for i in range(len(comps)-1): manager.add_object(str(i), comps[i]) for j in range(i+1, len(comps)): dist = manager.min_distance_single(comps[j]) if (dist < collision_dist): adjacencies[i].append(j) adjacencies[j].append(i) manager.remove_object(str(i)) # Run a DFS starting from the ground, check if everything is reachable visited = [False for comp in comps] stack = [0] # Index of 'ground' while len(stack) > 0: nindex = stack.pop() visited[nindex] = True for cindex in adjacencies[nindex]: if not visited[cindex]: stack.append(cindex) return all(visited) def obj2urdf(verts, faces, output_dir, density=1): if not os.path.exists(output_dir): os.makedirs(output_dir) # Load up the mesh mesh = tm.Trimesh(vertices=verts, faces=faces) # Switch from y-up to z-up mesh.vertices = mesh.vertices[:, [0, 2, 1]] mesh.fix_normals() # Extract the individual cuboid parts comps = mesh.split().tolist() # Detect (approximate) intersections between parts, to use for building joints # collision_dist = 0.005 * mesh.scale # collision_dist = 0.01 * mesh.scale collision_dist = 0.02 * mesh.scale adjacencies = {comp_index : [] for comp_index in range(len(comps))} manager = CollisionManager() for i in range(len(comps)-1): manager.add_object(str(i), comps[i]) for j in range(i+1, len(comps)): dist = manager.min_distance_single(comps[j]) if (dist < collision_dist): adjacencies[i].append(j) adjacencies[j].append(i) manager.remove_object(str(i)) # Compute connected components conn_comps = [] visited = [False for _ in comps] while not all(visited): conn_comp = set([]) start_idx = visited.index(False) stack = [start_idx] while len(stack) > 0: idx = stack.pop() visited[idx] = True conn_comp.add(idx) for nidx in adjacencies[idx]: if not visited[nidx]: stack.append(nidx) conn_comps.append(list(conn_comp)) # We export one URDF object file per connected component for i,conn_comp in enumerate(conn_comps): # Re-center this connected component ccmesh = meshconcat([comps[j] for j in conn_comp]) c = ccmesh.centroid transmat = [ [1, 0, 0, -c[0]], [0, 1, 0, -c[1]], [0, 0, 1, -c[2]], [0, 0, 0, 1] ] for j in conn_comp: comps[j].apply_transform(transmat) ccmesh.apply_transform(transmat) # Also, record where to start this mesh in the simulation # That's the x,y coords of the centroid, and -bbox bottom for the z (so it sits on the ground) # And the bbox diagonal (we use this for error thresholding) metadata = { 'start_pos': [c[0], c[1], -ccmesh.bounds[0][2]], 'volume': ccmesh.volume, 'height': abs(ccmesh.bounds[0][2] - ccmesh.bounds[1][2]), 'base': min(abs(ccmesh.bounds[0][0] - ccmesh.bounds[1][0]), abs(ccmesh.bounds[0][1] - ccmesh.bounds[1][1])) } with open(f'{output_dir}/assembly_{i}.json', 'w') as f: f.write(json.dumps(metadata)) # Build a directed tree by DFS root_idx = conn_comp[0] root = {'id': root_idx, 'children': []} fringe = [root] visited = set([root['id']]) while len(fringe) > 0: node = fringe.pop() for neighbor in adjacencies[node['id']]: if not (neighbor in visited): child_node = {'id': neighbor, 'children': []} node['children'].append(child_node) visited.add(child_node['id']) fringe.append(child_node) # Build up the URDF data structure urdf_root = xml.Element('robot') urdf_root.set('name', 'part_graph_shape') # Links for j in conn_comp: comp = comps[j] link = xml.SubElement(urdf_root, 'link') link.set('name', f'part_{j}') visual = xml.SubElement(link, 'visual') geometry = xml.SubElement(visual, 'geometry') mesh = xml.SubElement(geometry, 'mesh') mesh.set('filename', f'{output_dir}/part_{j}.stl') material = xml.SubElement(visual, 'material') material.set('name', 'gray') color = xml.SubElement(material, 'color') color.set('rgba', '0.5 0.5 0.5 1') collision = xml.SubElement(link, 'collision') geometry = xml.SubElement(collision, 'geometry') mesh = xml.SubElement(geometry, 'mesh') mesh.set('filename', f'{output_dir}/part_{j}.stl') inertial = xml.SubElement(link, 'inertial') mass = xml.SubElement(inertial, 'mass') mass.set('value', str(comp.volume * density)) inertia = xml.SubElement(inertial, 'inertia') inertia.set('ixx', '1.0') inertia.set('ixy', '0.0') inertia.set('ixz', '0.0') inertia.set('iyy', '1.0') inertia.set('iyz', '0.0') inertia.set('izz', '1.0') # Joints fringe = [root] while len(fringe) > 0: node = fringe.pop() for child_node in node['children']: joint = xml.SubElement(urdf_root, 'joint') joint.set('name', f'{node["id"]}_to_{child_node["id"]}') joint.set('type', 'fixed') parent = xml.SubElement(joint, 'parent') parent.set('link', f'part_{node["id"]}') child = xml.SubElement(joint, 'child') child.set('link', f'part_{child_node["id"]}') origin = xml.SubElement(joint, 'origin') origin.set('xyz', '0 0 0') fringe.append(child_node) # Save URDF file to disk # Have to make sure to split it into multiple lines, otherwise Bullet's URDF parser will # throw an error trying to load really large files as a single line... xmlstring = xml.tostring(urdf_root, encoding='unicode') xmlstring = '>\n'.join(xmlstring.split('>')) with open(f'{output_dir}/assembly_{i}.urdf', 'w') as f: f.write(xmlstring) # Write the parts to disk as STL files for the URDF to refer to for i,comp in enumerate(comps): comp.export(f'{output_dir}/part_{i}.stl') def check_stability(verts, faces, gui=False): # First, check if the file is even rooted. # If it's not rooted, it can't be stable if not check_rooted(verts, faces): return False # Start up the simulation mode = p.GUI if gui else p.DIRECT physicsClient = p.connect(mode) p.setGravity(0, 0, -9.8) # Load the ground plane p.setAdditionalSearchPath(pybullet_data.getDataPath()) # print(pybullet_data.getDataPath()) planeId = p.loadURDF("plane.urdf") # Convert the object to a URDF assembly, load it up # There may be more than one URDF, if the object had more than one connected component rand_str = ''.join(random.choices(string.ascii_lowercase, k=12)) obj2urdf(verts, faces, f'tmp/{rand_str}') objIds = [] startPositions = {} volumes = {} heights = {} bases = {} for urdf in [f for f in os.listdir(f'tmp/{rand_str}') if os.path.splitext(f)[1] == '.urdf']: with open(f'tmp/{rand_str}/{os.path.splitext(urdf)[0]}.json', 'r') as f: data = json.loads(f.read()) startPos = data['start_pos'] startOrientation = p.getQuaternionFromEuler([0,0,0]) objId = p.loadURDF(f"tmp/{rand_str}/{urdf}",startPos, startOrientation) objIds.append(objId) startPositions[objId] = startPos volumes[objId] = data['volume'] heights[objId] = data['height'] bases[objId] = data['base'] shutil.rmtree(f'tmp/{rand_str}') # Disable collisions between all objects (we only want collisions between objects and the ground) # That's because we want to check if the different components are *independently* stable, and # having them hit each other might muck up that judgment for i in range(0, len(objIds)-1): ni = p.getNumJoints(objIds[i]) for j in range(i+1, len(objIds)): nj = p.getNumJoints(objIds[j]) for k in range(-1, ni): for l in range(-1, nj): p.setCollisionFilterPair(objIds[i], objIds[j], k, l, False) import math # See if objects are stable from a small drop for objId in objIds: s = volumes[objId] b = bases[objId] v = s * b # 800, 4, 4, 4 p.applyExternalForce(objId, -1, (0, 0, 400*v), startPositions[objId], p.WORLD_FRAME) #p.applyExternalTorque(objId, -1, (0, 4*v, 0), p.WORLD_FRAME) #p.applyExternalTorque(objId, -1, (4*v, 0, 0), p.WORLD_FRAME) #p.applyExternalTorque(objId, -1, (0, 0, 80*v), p.WORLD_FRAME) # Run simulation if gui: for i in range(600): p.stepSimulation() time.sleep(1./600.) else: for i in range(1000): p.stepSimulation() for objId in objIds: endPos, _ = p.getBasePositionAndOrientation(objId) zend = endPos[2] zstart = startPositions[objId][2] zdiff = abs(zstart - zend) if zdiff > abs(0.025 * heights[objId]): p.disconnect() return False p.disconnect() return True
11,194
35.825658
119
py
ShapeMOD
ShapeMOD-main/SA_lang/tasks/pointnet2/setup.py
from __future__ import division, absolute_import, with_statement, print_function from setuptools import setup, find_packages from torch.utils.cpp_extension import BuildExtension, CUDAExtension import glob try: import builtins except: import __builtin__ as builtins builtins.__POINTNET2_SETUP__ = True import pointnet2 _ext_src_root = "pointnet2/_ext-src" _ext_sources = glob.glob("{}/src/*.cpp".format(_ext_src_root)) + glob.glob( "{}/src/*.cu".format(_ext_src_root) ) _ext_headers = glob.glob("{}/include/*".format(_ext_src_root)) requirements = ["etw_pytorch_utils==1.1.1", "h5py", "pprint", "enum34", "future"] setup( name="pointnet2", version=pointnet2.__version__, author="Erik Wijmans", packages=find_packages(), install_requires=requirements, ext_modules=[ CUDAExtension( name="pointnet2._ext", sources=_ext_sources, extra_compile_args={ "cxx": ["-O2", "-I{}".format("{}/include".format(_ext_src_root))], "nvcc": ["-O2", "-I{}".format("{}/include".format(_ext_src_root))], }, ) ], cmdclass={"build_ext": BuildExtension}, )
1,175
28.4
83
py
ShapeMOD
ShapeMOD-main/SA_lang/tasks/pointnet2/utils/pointnet2_utils.py
from __future__ import ( division, absolute_import, with_statement, print_function, unicode_literals, ) import torch from torch.autograd import Function import torch.nn as nn import etw_pytorch_utils as pt_utils import sys try: import builtins except: import __builtin__ as builtins try: import pointnet2._ext as _ext except ImportError: if not getattr(builtins, "__POINTNET2_SETUP__", False): raise ImportError( "Could not import _ext module.\n" "Please see the setup instructions in the README: " "https://github.com/erikwijmans/Pointnet2_PyTorch/blob/master/README.rst" ) if False: # Workaround for type hints without depending on the `typing` module from typing import * class RandomDropout(nn.Module): def __init__(self, p=0.5, inplace=False): super(RandomDropout, self).__init__() self.p = p self.inplace = inplace def forward(self, X): theta = torch.Tensor(1).uniform_(0, self.p)[0] return pt_utils.feature_dropout_no_scaling(X, theta, self.train, self.inplace) class FurthestPointSampling(Function): @staticmethod def forward(ctx, xyz, npoint): # type: (Any, torch.Tensor, int) -> torch.Tensor r""" Uses iterative furthest point sampling to select a set of npoint features that have the largest minimum distance Parameters ---------- xyz : torch.Tensor (B, N, 3) tensor where N > npoint npoint : int32 number of features in the sampled set Returns ------- torch.Tensor (B, npoint) tensor containing the set """ return _ext.furthest_point_sampling(xyz, npoint) @staticmethod def backward(xyz, a=None): return None, None furthest_point_sample = FurthestPointSampling.apply class GatherOperation(Function): @staticmethod def forward(ctx, features, idx): # type: (Any, torch.Tensor, torch.Tensor) -> torch.Tensor r""" Parameters ---------- features : torch.Tensor (B, C, N) tensor idx : torch.Tensor (B, npoint) tensor of the features to gather Returns ------- torch.Tensor (B, C, npoint) tensor """ _, C, N = features.size() ctx.for_backwards = (idx, C, N) return _ext.gather_points(features, idx) @staticmethod def backward(ctx, grad_out): idx, C, N = ctx.for_backwards grad_features = _ext.gather_points_grad(grad_out.contiguous(), idx, N) return grad_features, None gather_operation = GatherOperation.apply class ThreeNN(Function): @staticmethod def forward(ctx, unknown, known): # type: (Any, torch.Tensor, torch.Tensor) -> Tuple[torch.Tensor, torch.Tensor] r""" Find the three nearest neighbors of unknown in known Parameters ---------- unknown : torch.Tensor (B, n, 3) tensor of known features known : torch.Tensor (B, m, 3) tensor of unknown features Returns ------- dist : torch.Tensor (B, n, 3) l2 distance to the three nearest neighbors idx : torch.Tensor (B, n, 3) index of 3 nearest neighbors """ dist2, idx = _ext.three_nn(unknown, known) return torch.sqrt(dist2), idx @staticmethod def backward(ctx, a=None, b=None): return None, None three_nn = ThreeNN.apply class ThreeInterpolate(Function): @staticmethod def forward(ctx, features, idx, weight): # type(Any, torch.Tensor, torch.Tensor, torch.Tensor) -> Torch.Tensor r""" Performs weight linear interpolation on 3 features Parameters ---------- features : torch.Tensor (B, c, m) Features descriptors to be interpolated from idx : torch.Tensor (B, n, 3) three nearest neighbors of the target features in features weight : torch.Tensor (B, n, 3) weights Returns ------- torch.Tensor (B, c, n) tensor of the interpolated features """ B, c, m = features.size() n = idx.size(1) ctx.three_interpolate_for_backward = (idx, weight, m) return _ext.three_interpolate(features, idx, weight) @staticmethod def backward(ctx, grad_out): # type: (Any, torch.Tensor) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor] r""" Parameters ---------- grad_out : torch.Tensor (B, c, n) tensor with gradients of ouputs Returns ------- grad_features : torch.Tensor (B, c, m) tensor with gradients of features None None """ idx, weight, m = ctx.three_interpolate_for_backward grad_features = _ext.three_interpolate_grad( grad_out.contiguous(), idx, weight, m ) return grad_features, None, None three_interpolate = ThreeInterpolate.apply class GroupingOperation(Function): @staticmethod def forward(ctx, features, idx): # type: (Any, torch.Tensor, torch.Tensor) -> torch.Tensor r""" Parameters ---------- features : torch.Tensor (B, C, N) tensor of features to group idx : torch.Tensor (B, npoint, nsample) tensor containing the indicies of features to group with Returns ------- torch.Tensor (B, C, npoint, nsample) tensor """ B, nfeatures, nsample = idx.size() _, C, N = features.size() ctx.for_backwards = (idx, N) return _ext.group_points(features, idx) @staticmethod def backward(ctx, grad_out): # type: (Any, torch.tensor) -> Tuple[torch.Tensor, torch.Tensor] r""" Parameters ---------- grad_out : torch.Tensor (B, C, npoint, nsample) tensor of the gradients of the output from forward Returns ------- torch.Tensor (B, C, N) gradient of the features None """ idx, N = ctx.for_backwards grad_features = _ext.group_points_grad(grad_out.contiguous(), idx, N) return grad_features, None grouping_operation = GroupingOperation.apply class BallQuery(Function): @staticmethod def forward(ctx, radius, nsample, xyz, new_xyz): # type: (Any, float, int, torch.Tensor, torch.Tensor) -> torch.Tensor r""" Parameters ---------- radius : float radius of the balls nsample : int maximum number of features in the balls xyz : torch.Tensor (B, N, 3) xyz coordinates of the features new_xyz : torch.Tensor (B, npoint, 3) centers of the ball query Returns ------- torch.Tensor (B, npoint, nsample) tensor with the indicies of the features that form the query balls """ return _ext.ball_query(new_xyz, xyz, radius, nsample) @staticmethod def backward(ctx, a=None): return None, None, None, None ball_query = BallQuery.apply class QueryAndGroup(nn.Module): r""" Groups with a ball query of radius Parameters --------- radius : float32 Radius of ball nsample : int32 Maximum number of features to gather in the ball """ def __init__(self, radius, nsample, use_xyz=True): # type: (QueryAndGroup, float, int, bool) -> None super(QueryAndGroup, self).__init__() self.radius, self.nsample, self.use_xyz = radius, nsample, use_xyz def forward(self, xyz, new_xyz, features=None): # type: (QueryAndGroup, torch.Tensor. torch.Tensor, torch.Tensor) -> Tuple[Torch.Tensor] r""" Parameters ---------- xyz : torch.Tensor xyz coordinates of the features (B, N, 3) new_xyz : torch.Tensor centriods (B, npoint, 3) features : torch.Tensor Descriptors of the features (B, C, N) Returns ------- new_features : torch.Tensor (B, 3 + C, npoint, nsample) tensor """ idx = ball_query(self.radius, self.nsample, xyz, new_xyz) xyz_trans = xyz.transpose(1, 2).contiguous() grouped_xyz = grouping_operation(xyz_trans, idx) # (B, 3, npoint, nsample) grouped_xyz -= new_xyz.transpose(1, 2).unsqueeze(-1) if features is not None: grouped_features = grouping_operation(features, idx) if self.use_xyz: new_features = torch.cat( [grouped_xyz, grouped_features], dim=1 ) # (B, C + 3, npoint, nsample) else: new_features = grouped_features else: assert ( self.use_xyz ), "Cannot have not features and not use xyz as a feature!" new_features = grouped_xyz return new_features class GroupAll(nn.Module): r""" Groups all features Parameters --------- """ def __init__(self, use_xyz=True): # type: (GroupAll, bool) -> None super(GroupAll, self).__init__() self.use_xyz = use_xyz def forward(self, xyz, new_xyz, features=None): # type: (GroupAll, torch.Tensor, torch.Tensor, torch.Tensor) -> Tuple[torch.Tensor] r""" Parameters ---------- xyz : torch.Tensor xyz coordinates of the features (B, N, 3) new_xyz : torch.Tensor Ignored features : torch.Tensor Descriptors of the features (B, C, N) Returns ------- new_features : torch.Tensor (B, C + 3, 1, N) tensor """ grouped_xyz = xyz.transpose(1, 2).unsqueeze(2) if features is not None: grouped_features = features.unsqueeze(2) if self.use_xyz: new_features = torch.cat( [grouped_xyz, grouped_features], dim=1 ) # (B, 3 + C, 1, N) else: new_features = grouped_features else: new_features = grouped_xyz return new_features
10,413
26.049351
103
py
ShapeMOD
ShapeMOD-main/SA_lang/tasks/pointnet2/utils/linalg_utils.py
from __future__ import ( division, absolute_import, with_statement, print_function, unicode_literals, ) import torch from enum import Enum import numpy as np PDist2Order = Enum("PDist2Order", "d_first d_second") def pdist2(X, Z=None, order=PDist2Order.d_second): # type: (torch.Tensor, torch.Tensor, PDist2Order) -> torch.Tensor r""" Calculates the pairwise distance between X and Z D[b, i, j] = l2 distance X[b, i] and Z[b, j] Parameters --------- X : torch.Tensor X is a (B, N, d) tensor. There are B batches, and N vectors of dimension d Z: torch.Tensor Z is a (B, M, d) tensor. If Z is None, then Z = X Returns ------- torch.Tensor Distance matrix is size (B, N, M) """ if order == PDist2Order.d_second: if X.dim() == 2: X = X.unsqueeze(0) if Z is None: Z = X G = np.matmul(X, Z.transpose(-2, -1)) S = (X * X).sum(-1, keepdim=True) R = S.transpose(-2, -1) else: if Z.dim() == 2: Z = Z.unsqueeze(0) G = np.matmul(X, Z.transpose(-2, -1)) S = (X * X).sum(-1, keepdim=True) R = (Z * Z).sum(-1, keepdim=True).transpose(-2, -1) else: if X.dim() == 2: X = X.unsqueeze(0) if Z is None: Z = X G = np.matmul(X.transpose(-2, -1), Z) R = (X * X).sum(-2, keepdim=True) S = R.transpose(-2, -1) else: if Z.dim() == 2: Z = Z.unsqueeze(0) G = np.matmul(X.transpose(-2, -1), Z) S = (X * X).sum(-2, keepdim=True).transpose(-2, -1) R = (Z * Z).sum(-2, keepdim=True) return torch.abs(R + S - 2 * G).squeeze(0) def pdist2_slow(X, Z=None): if Z is None: Z = X D = torch.zeros(X.size(0), X.size(2), Z.size(2)) for b in range(D.size(0)): for i in range(D.size(1)): for j in range(D.size(2)): D[b, i, j] = torch.dist(X[b, :, i], Z[b, :, j]) return D if __name__ == "__main__": X = torch.randn(2, 3, 5) Z = torch.randn(2, 3, 3) print(pdist2(X, order=PDist2Order.d_first)) print(pdist2_slow(X)) print(torch.dist(pdist2(X, order=PDist2Order.d_first), pdist2_slow(X)))
2,335
26.482353
83
py
ShapeMOD
ShapeMOD-main/SA_lang/tasks/pointnet2/utils/pointnet2_modules.py
from __future__ import ( division, absolute_import, with_statement, print_function, unicode_literals, ) import torch import torch.nn as nn import torch.nn.functional as F import etw_pytorch_utils as pt_utils from pointnet2.utils import pointnet2_utils if False: # Workaround for type hints without depending on the `typing` module from typing import * class _PointnetSAModuleBase(nn.Module): def __init__(self): super(_PointnetSAModuleBase, self).__init__() self.npoint = None self.groupers = None self.mlps = None def forward(self, xyz, features=None): # type: (_PointnetSAModuleBase, torch.Tensor, torch.Tensor) -> Tuple[torch.Tensor, torch.Tensor] r""" Parameters ---------- xyz : torch.Tensor (B, N, 3) tensor of the xyz coordinates of the features features : torch.Tensor (B, N, C) tensor of the descriptors of the the features Returns ------- new_xyz : torch.Tensor (B, npoint, 3) tensor of the new features' xyz new_features : torch.Tensor (B, \sum_k(mlps[k][-1]), npoint) tensor of the new_features descriptors """ new_features_list = [] xyz_flipped = xyz.transpose(1, 2).contiguous() new_xyz = ( pointnet2_utils.gather_operation( xyz_flipped, pointnet2_utils.furthest_point_sample(xyz, self.npoint) ) .transpose(1, 2) .contiguous() if self.npoint is not None else None ) for i in range(len(self.groupers)): new_features = self.groupers[i]( xyz, new_xyz, features ) # (B, C, npoint, nsample) new_features = self.mlps[i](new_features) # (B, mlp[-1], npoint, nsample) new_features = F.max_pool2d( new_features, kernel_size=[1, new_features.size(3)] ) # (B, mlp[-1], npoint, 1) new_features = new_features.squeeze(-1) # (B, mlp[-1], npoint) new_features_list.append(new_features) return new_xyz, torch.cat(new_features_list, dim=1) class PointnetSAModuleMSG(_PointnetSAModuleBase): r"""Pointnet set abstrction layer with multiscale grouping Parameters ---------- npoint : int Number of features radii : list of float32 list of radii to group with nsamples : list of int32 Number of samples in each ball query mlps : list of list of int32 Spec of the pointnet before the global max_pool for each scale bn : bool Use batchnorm """ def __init__(self, npoint, radii, nsamples, mlps, bn=True, use_xyz=True): # type: (PointnetSAModuleMSG, int, List[float], List[int], List[List[int]], bool, bool) -> None super(PointnetSAModuleMSG, self).__init__() assert len(radii) == len(nsamples) == len(mlps) self.npoint = npoint self.groupers = nn.ModuleList() self.mlps = nn.ModuleList() for i in range(len(radii)): radius = radii[i] nsample = nsamples[i] self.groupers.append( pointnet2_utils.QueryAndGroup(radius, nsample, use_xyz=use_xyz) if npoint is not None else pointnet2_utils.GroupAll(use_xyz) ) mlp_spec = mlps[i] if use_xyz: mlp_spec[0] += 3 self.mlps.append(pt_utils.SharedMLP(mlp_spec, bn=bn)) class PointnetSAModule(PointnetSAModuleMSG): r"""Pointnet set abstrction layer Parameters ---------- npoint : int Number of features radius : float Radius of ball nsample : int Number of samples in the ball query mlp : list Spec of the pointnet before the global max_pool bn : bool Use batchnorm """ def __init__( self, mlp, npoint=None, radius=None, nsample=None, bn=True, use_xyz=True ): # type: (PointnetSAModule, List[int], int, float, int, bool, bool) -> None super(PointnetSAModule, self).__init__( mlps=[mlp], npoint=npoint, radii=[radius], nsamples=[nsample], bn=bn, use_xyz=use_xyz, ) class PointnetFPModule(nn.Module): r"""Propigates the features of one set to another Parameters ---------- mlp : list Pointnet module parameters bn : bool Use batchnorm """ def __init__(self, mlp, bn=True): # type: (PointnetFPModule, List[int], bool) -> None super(PointnetFPModule, self).__init__() self.mlp = pt_utils.SharedMLP(mlp, bn=bn) def forward(self, unknown, known, unknow_feats, known_feats): # type: (PointnetFPModule, torch.Tensor, torch.Tensor, torch.Tensor, torch.Tensor) -> torch.Tensor r""" Parameters ---------- unknown : torch.Tensor (B, n, 3) tensor of the xyz positions of the unknown features known : torch.Tensor (B, m, 3) tensor of the xyz positions of the known features unknow_feats : torch.Tensor (B, C1, n) tensor of the features to be propigated to known_feats : torch.Tensor (B, C2, m) tensor of features to be propigated Returns ------- new_features : torch.Tensor (B, mlp[-1], n) tensor of the features of the unknown features """ if known is not None: dist, idx = pointnet2_utils.three_nn(unknown, known) dist_recip = 1.0 / (dist + 1e-8) norm = torch.sum(dist_recip, dim=2, keepdim=True) weight = dist_recip / norm interpolated_feats = pointnet2_utils.three_interpolate( known_feats, idx, weight ) else: interpolated_feats = known_feats.expand( *(known_feats.size()[0:2] + [unknown.size(1)]) ) if unknow_feats is not None: new_features = torch.cat( [interpolated_feats, unknow_feats], dim=1 ) # (B, C2 + C1, n) else: new_features = interpolated_feats new_features = new_features.unsqueeze(-1) new_features = self.mlp(new_features) return new_features.squeeze(-1) if __name__ == "__main__": from torch.autograd import Variable torch.manual_seed(1) torch.cuda.manual_seed_all(1) xyz = Variable(torch.randn(2, 9, 3).cuda(), requires_grad=True) xyz_feats = Variable(torch.randn(2, 9, 6).cuda(), requires_grad=True) test_module = PointnetSAModuleMSG( npoint=2, radii=[5.0, 10.0], nsamples=[6, 3], mlps=[[9, 3], [9, 6]] ) test_module.cuda() print(test_module(xyz, xyz_feats)) # test_module = PointnetFPModule(mlp=[6, 6]) # test_module.cuda() # from torch.autograd import gradcheck # inputs = (xyz, xyz, None, xyz_feats) # test = gradcheck(test_module, inputs, eps=1e-6, atol=1e-4) # print(test) for _ in range(1): _, new_features = test_module(xyz, xyz_feats) new_features.backward(torch.cuda.FloatTensor(*new_features.size()).fill_(1)) print(new_features) print(xyz.grad)
7,338
30.229787
106
py
LINDA_DSS
LINDA_DSS-master/learning.py
# update for tensorflow from __future__ import absolute_import, division, print_function, unicode_literals import pandas as pd import numpy as np import seaborn as sns import random as rn import re import warnings import csv import tensorflow as tf # Force TensorFlow to single thread # Multiple threads are a potential source of non-reprocible research resulsts session_conf = tf.compat.v1.ConfigProto( intra_op_parallelism_threads=1, inter_op_parallelism_threads=1 ) # tf.set_random_seed() will make random number generation in the TensorFlow backend # have a well defined initial state # more details: https://www.tensorflow.org/api_docs/python/tf/set_random_seed tf.compat.v1.set_random_seed(515) # keras / deep learning libraries from tensorflow import keras from tensorflow.keras.models import Sequential from tensorflow.keras.models import model_from_json from tensorflow.keras.layers import Dense from tensorflow.keras.optimizers import Nadam from tensorflow.keras.callbacks import TensorBoard from tensorflow.keras.utils import plot_model # callbacks from tensorflow.keras import backend as K from tensorflow.keras.callbacks import EarlyStopping from tensorflow.keras.callbacks import ModelCheckpoint from tensorflow.keras.callbacks import ReduceLROnPlateau from matplotlib import pyplot as plt from matplotlib.pyplot import figure import matplotlib.image as mpimg import pylab as pl from pylab import savefig plt.style.use('seaborn-deep') from sklearn.model_selection import train_test_split from sklearn.preprocessing import OneHotEncoder, StandardScaler,MinMaxScaler # Bayesian networks from sklearn.preprocessing import KBinsDiscretizer from pylab import * import pyAgrum as gum import pyAgrum.lib.notebook as gnb # for classification purposes from pyAgrum.lib.bn2roc import showROC import warnings warnings.simplefilter(action='ignore', category=FutureWarning) from sklearn.metrics import roc_curve, auc from matplotlib import pyplot as plt from matplotlib.pyplot import figure import matplotlib.image as mpimg import pylab as pl from pylab import savefig plt.style.use('seaborn-deep') # RECALL ----------------------------------------------------------------------------- # def recall_m(y_true, y_pred): """Computes the recal measure of an evaluation setting Parameters ---------- y_true : list list of groundtruth labels y_pred : list list of predictions from blackbox Returns ------- recall : vector a vector with the recall values between the predictions and the groundtruths """ true_positives = K.sum(K.round(K.clip(y_true * y_pred, 0, 1))) possible_positives = K.sum(K.round(K.clip(y_true, 0, 1))) recall = true_positives / (possible_positives + K.epsilon()) return recall # PRECISION --------------------------------------------------------------------------- # def precision_m(y_true, y_pred): """Computes the precision measure of an evaluation setting Parameters ---------- y_true : list list of groundtruth labels y_pred : list list of predictions from blackbox Returns ------- precision : vector a vector with the precision values between the predictions and the groundtruths """ true_positives = K.sum(K.round(K.clip(y_true * y_pred, 0, 1))) predicted_positives = K.sum(K.round(K.clip(y_pred, 0, 1))) precision = true_positives / (predicted_positives + K.epsilon()) return precision # F1 ------------------------------------------------------------------------------------ # Computes the F1 measure of an evaluation setting # y_true: list of groundtruth labels # y_pred: list of predictions from blackbox def f1_m(y_true, y_pred): precision = precision_m(y_true, y_pred) recall = recall_m(y_true, y_pred) return 2*((precision*recall)/(precision+recall+K.epsilon())) # CREATE_MODEL -------------------------------------------------------------------------- # creates a neural network model with a certain number of hidden layers and a certain # number of neurons in each layer. # input_dim: an integer specifying the number of input neurons # output_dim: an integer specifying the number of output neurons (the number of labels) # hidden_layers: an integer specifying the number of hidden layers # loss_func: the loss function of the model. By default, it is applied the 'categorical_crossentropy' # optim: the optimisation algorithm used in the model. By default it is used the 'nadam' algorithm # metrics: a list of strings specifying the metrics to be evaluated ('accuracy', 'f1', 'recall','precision') def create_model(input_dim, output_dim, nodes, hidden_layers=1, loss_func='categorical_crossentropy', optim='nadam', metrics=['accuracy'], name='model'): model = Sequential(name=name) model.add( Dense(nodes, input_dim=input_dim, activation='relu')) # input layer for i in range(hidden_layers): # hidden layers model.add(Dense(nodes, activation='relu')) model.add(Dense(output_dim, activation='softmax')) # output layer if( optim == "nadam" ): # Compile model optim = keras.optimizers.Nadam(lr=0.0001, beta_1=0.9, beta_2=0.999) model.compile(loss=loss_func, optimizer=optim, metrics=metrics) return model # GRID_SEARCH ----------------------------------------------------------------------------- # Generates a set of models with different configurations, ranging from an # initial number of neurons to a maximum number of neurons # start_nodes: an integer specifying the initial number of neurons to generate a model from # max_nodes: an integer specifying the maximum number of neurons to generate a model from # max_hlayers: an integer specifying the maximum number of hidden layers to generate a model from # debug: boolean that acts as a flag. If True, it displays the characteristics of each model # metrics: a list of strings with the metrics to be evaluated def grid_search_model_generator(n_features, n_classes, start_nodes = 1, max_nodes = 12, max_hlayers = 5, debug = False, metrics = ['accuracy'] ): models = [] # generate different models with different neurons and different hidden layers for neurons in range(start_nodes, max_nodes+1): for hidden_layer in range(1, max_hlayers+1): model_name = "model_h" + str(hidden_layer) + "_N"+str(neurons) model = create_model(n_features, n_classes, neurons, hidden_layer, name=model_name, metrics = metrics) models.append( model ) # add the generated model to a list # plot general information for each model if( debug ): for model in models: model.summary() return models # PERFORM_GRID_SEARCH ------------------------------------------------------------------- # given a list of models with different configurations, fit the data to the models, # and evaluate the model. This function returns a list of training histories for each model # models: list of models # X_train: # Y_train: # X_validation: # Y_validation: # X_test: # Y_test: # batch_size: # epochs: def perform_grid_search( models, path, dataset_name, X_train, Y_train, X_validation, Y_validation, X_test, Y_test, batch_size, epochs ): HISTORY_DICT = {} # define the callebacks to take into consideration during training # stop training when convergence is achieved after 10 iterations early_stop = EarlyStopping(monitor='val_loss', patience=10, verbose=1, mode='min') # save the best model after every epoch model_checkpoint = ModelCheckpoint(path + "training/" + dataset_name + "/model_{epoch:02d}-{val_loss:.2f}.h5", monitor='val_loss', verbose=0, save_best_only=True, mode='min') callbacks_list = [early_stop, model_checkpoint] # grid search over each model for model in models: print('MODEL NAME:', model.name) history_callback = model.fit(X_train, Y_train, batch_size = batch_size, epochs = epochs, verbose=0, validation_data=(X_validation, Y_validation), callbacks=callbacks_list) score_test = model.evaluate( X_test, Y_test, verbose=0 ) score_train = model.evaluate( X_train, Y_train ) print('Test loss: ', format(score_test[0], '.4f'), '\tTrain loss: ', format(score_train[0], '.4f') ) print('Test accuracy: ', format(score_test[1], '.4f'), '\tTrain accu: ', format(score_train[1], '.4f') ) print('Abs accuracy: ', format( np.abs( score_test[1] - score_train[1] ), '.4f')) print('Abs loss: ', format( np.abs( score_test[0] - score_train[0] ), '.4f')) print('\n###########################################################\n') HISTORY_DICT[model.name] = [history_callback, model] return HISTORY_DICT # SAVE_MODEL ----------------------------------------------------------------------------- # saves a trained model into a json and hdf5 file # model: model to be saved # model_name: string with model name # path: string with path to save def save_model( model, model_name, path ): # serialize model to JSON model_json = model.to_json() with open(path + model_name+"_DUO.json", "w") as json_file: json_file.write(model_json) json_file.close() # serialize weights to HDF5 model.save_weights( path + model_name+"_DUO.h5") print("Saving files:") print(path + model_name+"_DUO.json") print(path + model_name+"_DUO.h5") print("Model saved to disk") # SAVE_MODEL_HISTORY ------------------------------------------------------------------- # saves a trained model into a csv file # model_hist: history of the model to be saved # model_name: string with model name # path: string with path to save def save_model_history( model_hist, model_name, path ): file = open(path + model_name + "_hist.csv", "w") w = csv.writer( file ) for key, val in model_hist.history.items(): w.writerow([key, val]) file.close() print(path + model_name+"_DUO.h5") print("Model history saved to disk") # LOAD_MODEL_HISTORY ------------------------------------------ # loads a saved model history into memory # model_name: the name of the model # path: path to model history def load_model_history( model_name, path): model_hist_loaded = {} values = [] # load dictionary r = open( path + model_name + "_hist.csv", "r").read() for line in r.split("\n"): if(len(line) == 0): continue metric = line.split(",\"[")[0] # extract metrics values_str = line.split(",\"[")[1].replace("]\"","").split(", ") # extract validation values values = [float(val_str) for val_str in values_str] model_hist_loaded.update( {metric : values} ) return model_hist_loaded # LOAD_MODEL ------------------------------------------ # loads a saved model into memory # model_name: the name of the model # path: path to model history def load_model( model_name, path ): json_file = open( path + model_name + "_DUO.json", 'r') loaded_model_json = json_file.read() json_file.close() # load weights into new model loaded_model = model_from_json(loaded_model_json) loaded_model.load_weights(path + model_name + "_DUO.h5") print("Loaded model from disk") return loaded_model def plot_model_history( model_history, metric ): plt.plot(model_history[ metric.lower() ], label='train') plt.plot(model_history["val_" + metric.lower()], label='validation') plt.ylabel(metric) plt.xlabel('Number of Epochs') plt.ylim([0, 1]) plt.legend() plt.show() def plot_ROC_Curve( model, X, Y, n_classes): Y_pred_proba = model.predict(X) # Compute ROC curve and ROC area for each class fpr = dict() tpr = dict() roc_auc = dict() for i in range(n_classes): fpr[i], tpr[i], _ = roc_curve(Y[:, i], Y_pred_proba[:, i]) roc_auc[i] = auc(fpr[i], tpr[i]) # Plot of a ROC curve for a specific class for i in range(n_classes): plt.figure() plt.plot(fpr[i], tpr[i], label='ROC curve (area = %0.2f)' % roc_auc[i]) plt.plot([0, 1], [0, 1], 'k--') plt.xlim([0.0, 1.0]) plt.ylim([0.0, 1.05]) plt.xlabel('False Positive Rate') plt.ylabel('True Positive Rate') plt.title('Receiver operating characteristic example') plt.legend(loc="lower right") plt.show() # ENCODE_DATA -------------------------------------------------------------------------- # Applies one hot encoder to data # data: a dataframe # class_var: string with class variable name def encode_data(data, class_var): feature_names = data.drop([class_var], axis=1).columns.tolist() X = data[ feature_names ].values y = data[class_var].values n_features = X.shape[1] n_classes = len(data[class_var].unique()) # create numerical encoding for attribute species enc = OneHotEncoder() Y = enc.fit_transform(y[:, np.newaxis]).toarray() # Scale data to have mean 0 and variance 1 # which is importance for convergence of the neural network scaler = MinMaxScaler() X_scaled = scaler.fit_transform(X) return X_scaled, Y, enc, scaler # LOAD_TRAINING_DATA --------------------------------------------------------------------- # loads into a multiarray format a training set previously saved in a .csv file # dataset_path: string containing the path where the files will be saved def load_training_data( dataset_path ): X_train = pd.read_csv(dataset_path.replace(".csv", "") + "_Xtrain.csv", index_col=False).values X_test = pd.read_csv(dataset_path.replace(".csv", "") + "_Xtest.csv", index_col=False).values X_validation =pd.read_csv(dataset_path.replace(".csv", "") + "_Xvalidation.csv",index_col=False).values Y_train = pd.read_csv(dataset_path.replace(".csv", "") + "_Ytrain.csv",index_col=False).values Y_test =pd.read_csv(dataset_path.replace(".csv", "") + "_Ytest.csv", index_col=False).values Y_validation = pd.read_csv(dataset_path.replace(".csv", "") + "_Yvalidation.csv", index_col=False).values return X_train, Y_train, X_test, Y_test, X_validation, Y_validation # GENERATE_SAVE_TRAINING_DATA ------------------------------------------------------------ # # dataset_path: string containing the path where the files will be saved # X: NxM matrix representing the training data # Y: NxC matrix representing the OneHotEconder of C classes def generate_save_training_data( dataset_path, X, Y): """Generates training, test and validation sets and stores this information into files Parameters ---------- dataset_path : str The file location of the spreadsheet samples : int, optional The number of permutations to generate from the original vector (default is 300) variance : int, optional Quantity to permute in each feature (default is 0.25) Returns ------- permutations : matrix a 2-D matrix with dimensions (samples, features) with all the permutations of the original vector """ # generate train, test and validation sets X_train, X_test, Y_train, Y_test = train_test_split(X, Y, test_size=0.3, random_state=515) X_validation, X_test, Y_validation, Y_test = train_test_split(X_test, Y_test, test_size=0.5, random_state=515) np.savetxt(dataset_path.replace(".csv", "") + "_Xtrain.csv", X_train, delimiter=",") np.savetxt(dataset_path.replace(".csv", "") + "_Xtest.csv", X_test, delimiter=",") np.savetxt(dataset_path.replace(".csv", "") + "_Xvalidation.csv", X_validation, delimiter=",") np.savetxt(dataset_path.replace(".csv", "") + "_Ytrain.csv", Y_train, delimiter=",") np.savetxt(dataset_path.replace(".csv", "") + "_Ytest.csv", Y_test, delimiter=",") np.savetxt(dataset_path.replace(".csv", "") + "_Yvalidation.csv", Y_validation, delimiter=",") ############################################################################## # BAYESIAN NETWORK EXPLANATIONS # ############################################################################## def compute_perm_range(feat, variance = 0.25): """ Parameters ---------- feat : float Value of a feature to be permuted samples : int, optional The number of permutations to generate from the original vector (default is 300) variance : int, optional Quantity to permute in each feature (default is 0.25) Returns ------- min_range : float minimum value that a feature can be permuted max_range : float maximum value that a feature can be permuted """ min_range = feat - variance max_range = feat + variance # features are scaled between 0 and 1 # if the permutation make the feature negative, this values is set to 0 if( min_range < 0 ): min_range = 0 # if the permutation make the feature bigger than 1, this values is set to 1 if( max_range > 1 ): max_range = 1 return min_range, max_range # PERMUTE_SINGLE_FEATURES_____________________________________________________________ # def permute_single_features( my_array, samples = 300, variance = 0.25 ): """Given a single array from which one pretends to generate local explanations from Draw samples from a uniform distribution within a range of feature_val +- variance Returns a matrix with a number of samples (by default 300) with permutations of each feature of the input vector Parameters ---------- my_array : np.array The datapoint to be locally explained samples : int, optional The number of permutations to generate from the original vector (default is 300) variance : int, optional Quantity to permute in each feature (default is 0.25) Returns ------- permutations : matrix a 2-D matrix with dimensions (samples, features) with all the permutations of the original vector """ # permutation result list permutations = [] # just keeping a controlled number of decimal places my_array = np.round(my_array,4) # keep a copy of the original array, since we will be changing the features my_array_backup = my_array # extract number of features num_features = my_array.shape[0] # add original vector to dataframe permutations.append( my_array_backup.tolist() ) # for each feature of the input feature vector, for feat in range(0, num_features): # get feature value my_array = my_array_backup feature_val = my_array[feat] # set permutation of feature between [ feat - variance ; feat + variance ] min_range, max_range = compute_perm_range( feature_val, variance ) # generate sample of random features within a range for perm in range(0, int(round(samples/num_features, 0))): # set the new vector my_array[feat] = np.abs(np.round(rn.uniform(min_range, max_range),4)) permutations.append( my_array.tolist() ) #rn.shuffle(permutations) return permutations def check_input( value ): if value < 0: return 0 if value > 1: return 1 return value def permute_single_features_circle( my_array, samples = 300, variance = 0.25 ): # permutation result list permutations = [] # just keeping a controlled number of decimal places my_array = np.round(my_array,4) # keep a copy of the original array, since we will be changing the features my_array_backup = my_array # extract number of features num_features = my_array.shape[0] # add original vector to dataframe permutations.append( my_array_backup.tolist() ) # for each feature of the input feature vector, for perm in range(0, int(round(samples/num_features, 0))): # generate sample of random features within a range temp1 = [] temp2 = [] for feat in range(0, num_features): theta = 2*math.pi*np.random.random() feature_val = my_array[feat] # set the new vector temp1.append( check_input( feature_val + np.round(np.random.uniform(0, variance),4)*math.cos(theta) )) temp2.append( check_input( feature_val + np.round(np.random.uniform(0, variance),4)*math.sin(theta) )) permutations.append( temp1 ) permutations.append( temp2 ) #rn.shuffle(permutations) return permutations[0:samples] # LEARNBN ------------------------------------------- # def learnBN( file_path, algorithm = "Hill Climbing" ): """Given a single array from which one pretends to generate local explanations from Draw samples from a uniform distribution within a range of feature_val +- variance Returns a matrix with a number of samples (by default 300) with permutations of each feature of the input vector Parameters ---------- my_array : np.array The datapoint to be locally explained samples : int, optional The number of permutations to generate from the original vector (default is 300) variance : int, optional Quantity to permute in each feature (default is 0.25) Returns ------- permutations : matrix a 2-D matrix with dimensions (samples, features) with all the permutations of the original vector """ learner = gum.BNLearner( file_path ) if( algorithm == "Hill Climbing"): print("Selecting Greedy Hill Climbing Algorithm") learner.useGreedyHillClimbing() if( algorithm == "Local Search" ): print("Selecting Local Search Algorithm") bn = learner.useLocalSearchWithTabuList() if( algorithm == "3off2"): print("Selecting 3Off2 Algorithm") learner.use3off2() if( algorithm == "miic" ): print("Selecting MIIC Algorithm") learner.useMIIC() learner.learnBN() bn = learner.learnBN() essencGraph = gum.EssentialGraph( bn ) infoBN = gnb.getInformation( bn ) return [ bn, infoBN, essencGraph ] # DISCRETIZE_DATAFRAME ------------------------------------------------------- # # def discretize_dataframe( df, class_var, num_bins=4 ): """Given a dataframe with continuous values, convert the continuous values into discrete ones by splitting the data into bins and by computing the respective quartiles Parameters ---------- df : pd.DataFrame The datapoint to be locally explained class_var : str The number of permutations to generate from the original vector (default is 300) num_bins : int, optional Quantity to permute in each feature (default is 0.25) Returns ------- permutations : matrix a 2-D matrix with dimensions (samples, features) with all the permutations of the original vector """ r=np.array(range(num_bins+1))/(1.0*num_bins) # quantiles are building using pandas.qcut # The "class" column is just copied. l=[] for col in df.columns.values: if col!=class_var: l.append( pd.DataFrame( pd.qcut( df[col],r, duplicates='drop',precision=2),columns=[col])) else: l.append( pd.DataFrame( df[col].values,columns=[col])) treated = pd.concat(l, join='outer', axis=1) return treated # SAVE_DISCRETIZED_DATAFRAME --------------------------------------------------- # def save_discretized_dataframe(indx, df_model, model_type, perm_type, bins, dataset_name, path, class_var): """Given a single array from which one pretends to generate local explanations from Draw samples from a uniform distribution within a range of feature_val +- variance Returns a matrix with a number of samples (by default 300) with permutations of each feature of the input vector Parameters ---------- my_array : np.array The datapoint to be locally explained samples : int, optional The number of permutations to generate from the original vector (default is 300) variance : int, optional Quantity to permute in each feature (default is 0.25) Returns ------- permutations : matrix a 2-D matrix with dimensions (samples, features) with all the permutations of the original vector """ file_path = path + dataset_name + "/" + str(indx) + "/" + re.sub( r"\.\w+", "", dataset_name ) + "_" + model_type +"_INDX_" + str(indx) + "_" + perm_type +".csv" df_discr = discretize_dataframe( df_model, bins, class_var ) print("Saving discretized dataset into: %s\n" %(file_path)) df_discr.to_csv( file_path, index=False) # WRAP_INFORMATION ------------------------------------------- # def wrap_information( local_data_dict ): true_positives = [] true_negatives = [] false_positives = [] false_negatives = [] for instance in local_data_dict: # wrap up true positives if( instance['prediction_type'] == 'TRUE POSITIVE'): true_positives.append(instance) # wrap up true negatives if( instance['prediction_type'] == 'TRUE NEGATIVE' ): true_negatives.append(instance) # wrap up false positives if( instance['prediction_type'] == 'FALSE POSITIVE' ): false_positives.append(instance) # wrap up false negatives if( instance['prediction_type'] == 'FALSE NEGATIVE' ): false_negatives.append(instance) return true_positives, true_negatives, false_positives, false_negatives # GENERATE_PERMUTATIONS ------------------------------------------- # def generate_permutations( instance, labels_lst, feature_names, class_var, encoder, scaler, model, samples = 300, variance = 0.25): # get datapoint in scaled feature space local_datapoint = np.array(instance['scaled_vector']) # get datapoint in original feature space local_datapoint_orig = np.array(instance['original_vector']) # permute features permutations = permute_single_features( local_datapoint, samples = samples, variance = variance ) #permutations = permute_single_features_circle( local_datapoint, samples = samples, variance = variance ) # convert permutations to original feature space permutations_orig = scaler.inverse_transform( permutations ) # compute predictions for each permuted instance predictions = encoder.inverse_transform( model.predict( permutations ) ) # convert prediction classes to labels labelled_predictions = [ labels_lst[ int(predictions[indx][0]) ] for indx in range(0, len(predictions))] # add all this information to a single dataframe df_local_permutations = pd.DataFrame( permutations_orig, columns = feature_names ) # add class variable to dataframe df_local_permutations[ class_var ] = labelled_predictions return df_local_permutations # GEBERATE_BN_EXPLANATIONS ------------------------------------------------------------ # def generate_BN_explanations(instance, label_lst, feature_names, class_var, encoder, scaler, model, path, dataset_name ): # necessary for starting Numpy generated random numbers in an initial state np.random.seed(515) # Necessary for starting core Python generated random numbers in a state rn.seed(515) indx = instance['index'] prediction_type = instance['prediction_type'].lower()+"s" prediction_type = prediction_type.replace(" ", "_") # generate permutations df = generate_permutations( instance, label_lst, feature_names, class_var, encoder, scaler, model) # discretize data df_discr = discretize_dataframe( df, class_var, num_bins=4 ) # save discretised dataframe (for debugging and reproduceability purposes) path_to_permutations = path + "feature_permutations/" + dataset_name.replace(".csv","") + "/" + prediction_type + "/" + str(indx) + ".csv" df_discr.to_csv( path_to_permutations, index=False) # normalise dataframe normalise_dataframe( path_to_permutations ) # learn BN bn, infoBN, essencGraph = learnBN( path_to_permutations.replace(".csv", "_norm.csv") ) # perform inference inference = gnb.getInference(bn, evs={},targets=df_discr.columns.to_list(), size='12') # show networks gnb.sideBySide(*[bn, inference, infoBN ], captions=[ "Bayesian Network", "Inference", "Information Network" ]) # save to file path_to_explanation = path + "explanations/" + dataset_name.replace(".csv", "") + "/BN/" + prediction_type + "/" gum.lib.bn2graph.dotize( bn , path_to_explanation + str(indx) + "_BN" ) gum.saveBN(bn,path_to_explanation + str(indx) + "_BN.net" ) return [bn, inference, infoBN] # GEBERATE_BN_EXPLANATIONSMB ------------------------------------------------------------ # def generate_BN_explanationsMB(instance, label_lst, feature_names, class_var, encoder, scaler, model, path, dataset_name, variance = 0.1, algorithm = "Hill Climbing" ): # necessary for starting Numpy generated random numbers in an initial state np.random.seed(515) # Necessary for starting core Python generated random numbers in a state rn.seed(515) indx = instance['index'] prediction_type = instance['prediction_type'].lower()+"s" prediction_type = prediction_type.replace(" ", "_") # generate permutations df = generate_permutations( instance, label_lst, feature_names, class_var, encoder, scaler, model, variance = variance) # discretize data df_discr = discretize_dataframe( df, class_var, num_bins=4 ) # save discretised dataframe (for debugging and reproduceability purposes) path_to_permutations = path + "feature_permutations/" + dataset_name.replace(".csv","") + "/" + prediction_type + "/" + str(indx) + ".csv" df_discr.to_csv( path_to_permutations, index=False) # normalise dataframe normalise_dataframe( path_to_permutations ) # learn BN bn, infoBN, essencGraph = learnBN( path_to_permutations.replace(".csv", "_norm.csv"), algorithm = algorithm) # perform inference inference = gnb.getInference(bn, evs={},targets=df_discr.columns.to_list(), size='12') # compute Markov Blanket markov_blanket = gum.MarkovBlanket(bn, class_var) # show networks # gnb.sideBySide(*[bn, inference, markov_blanket ], # captions=[ "Bayesian Network", "Inference", "Markov Blanket" ]) # save to file path_to_explanation = path + "explanations/" + dataset_name.replace(".csv", "") + "/BN/" + prediction_type + "/" gum.lib.bn2graph.dotize( bn , path_to_explanation + str(indx) + "_BN" ) gum.saveBN(bn,path_to_explanation + str(indx) + "_BN.net" ) return [bn, inference, infoBN, markov_blanket] # GENERATE_LOCAL_PREDICTIONS ------------------------------------------- # def generate_local_predictions( X, Y, model, scaler, encoder ): # get original vector orig_vec = np.round(scaler.inverse_transform(X),6) # generate all predictions for X predictions = model.predict( X ) # extrace the label of the prediction of X[indx] prediction_class = encoder.inverse_transform( predictions ) local_data_dict = [] for indx in range(0, orig_vec.shape[0]): ground_truth = np.expand_dims(Y[indx], axis=0) ground_truth_class = encoder.inverse_transform( ground_truth )[0][0] prediction = prediction_class[indx][0] # check if data point is a true positive if( ( int(prediction) == int(ground_truth_class) ) & (int(prediction)==1) & (int(ground_truth_class)==1) ): pred_type = "TRUE POSITIVE" # check if data point is a true negative if( ( int(prediction) == int(ground_truth_class) ) & (int(prediction)==0) & (int(ground_truth_class)==0) ): pred_type = "TRUE NEGATIVE" # check if data point is a false negative if( ( int(prediction) != int(ground_truth_class) ) & (int(prediction)==0) & (int(ground_truth_class)==1) ): pred_type = "FALSE NEGATIVE" # check if data point is a false positve if( ( int(prediction) != int(ground_truth_class) ) & (int(prediction)==1) & (int(ground_truth_class)==0) ): pred_type = "FALSE POSITIVE" local_data_dict.append( {'index' : indx, 'original_vector' : orig_vec[indx,:].tolist(), 'scaled_vector' : X[indx,:].tolist(), 'ground_truth' : ground_truth_class, 'predictions' : prediction, 'prediction_type' : pred_type}) return local_data_dict ################################################################################## # TEXT PROCESSING # # ############################################################################### # FIND ----------------------------------------------- # def find(s, ch): return [i for i, ltr in enumerate(s) if ltr == ch] # UNTOKENIZE ----------------------------------------------- # def untokenize( tokens, delim ): untokenized = tokens[0] for indx in range(1, len(tokens)): untokenized = untokenized + delim + tokens[indx] return untokenized # NORMALISE_LINE ------------------------------------------- # def normalise_line( my_str, class_label ): my_str = my_str.replace("\","+class_label, "") my_str = my_str.replace("-1e-05", "0.0000") tokens = my_str.split("\",\"") tokens_norm = [] for token in tokens: token = token.replace("]","") indxs = find(token, ".") indx_comma = find(token, ",")[0]+2 if( (len(token[indxs[1]+1 : -1 ]) >= 4) & (len( token[indxs[0]+1 : indx_comma-2 ]) >= 4) ): token_temp = token[0:indxs[0]] + "." + token[indxs[0] + 1 : indxs[0]+5] + ", " +token[indx_comma:indxs[1]] + token[indxs[1] : indxs[1]+5 ] + "]" if( (len(token[indxs[1]+1 : -1 ]) < 4) & (len( token[indxs[0]+1 : indx_comma-2 ]) >= 4) ): extra = "0"*(np.abs(len(token[indxs[1]+1 : -1 ]) - 4)) token_temp = token[0:indxs[0]] + "." + token[indxs[0] + 1 : indxs[0]+5] + ", " +token[indx_comma:indxs[1]] + token[indxs[1] : -1 ] + extra + "]" if( (len(token[indxs[1]+1 : -1 ]) >= 4) & (len( token[indxs[0]+1 : indx_comma-2 ]) < 4) ): extra = "0"*(np.abs(len( token[indxs[0]+1 : indx_comma-2 ]) - 4)) token_temp = token[0:indxs[0]] + "." + extra + ", " +token[indx_comma:indxs[1]] + token[indxs[1] : -1 ] + extra + "]" if( (len(token[indxs[1]+1 : -1 ]) < 4) & (len( token[indxs[0]+1 : indx_comma-2 ]) < 4) ): extra2 = "0"*(np.abs(len(token[indxs[1]+1 : -1 ]) - 4)) extra1 = "0"*(np.abs(len(token[indxs[0]+1 : -1 ]) - 4)) token_temp = token[0:indxs[0]] + "." + extra1 + ", " +token[indx_comma:indxs[1]] + token[indxs[1] : -1 ] + extra2 + "]" tokens_norm.append(token_temp) return untokenize( tokens_norm, "\",\"") + "\"," +class_label # NORMALISE_LINE ------------------------------------------- # def normalise_dataframe( path_to_permutations ): file = open(path_to_permutations,"r") f_write = open(path_to_permutations.replace(".csv", "_norm.csv"),"w") header = file.readline().replace("\n","") f_write.write( header + "\n") for line in file.readlines(): # get class class_label = line.split("\",")[-1].replace("\n","") # normalise dataframe input line_norm = normalise_line( line.replace("\n",""), class_label ) # write normalised input to file f_write.write(line_norm + "\n") file.close() f_write.close()
35,736
37.099147
177
py
LINDA_DSS
LINDA_DSS-master/Experiments/impact_of_Learning_algo/learning.py
# update for tensorflow from __future__ import absolute_import, division, print_function, unicode_literals import pandas as pd import numpy as np import seaborn as sns import random as rn import re import warnings import csv import tensorflow as tf # Force TensorFlow to single thread # Multiple threads are a potential source of non-reprocible research resulsts session_conf = tf.compat.v1.ConfigProto( intra_op_parallelism_threads=1, inter_op_parallelism_threads=1 ) # tf.set_random_seed() will make random number generation in the TensorFlow backend # have a well defined initial state # more details: https://www.tensorflow.org/api_docs/python/tf/set_random_seed tf.compat.v1.set_random_seed(515) # keras / deep learning libraries from tensorflow import keras from tensorflow.keras.models import Sequential from tensorflow.keras.models import model_from_json from tensorflow.keras.layers import Dense from tensorflow.keras.optimizers import Nadam from tensorflow.keras.callbacks import TensorBoard from tensorflow.keras.utils import plot_model # callbacks from tensorflow.keras import backend as K from tensorflow.keras.callbacks import EarlyStopping from tensorflow.keras.callbacks import ModelCheckpoint from tensorflow.keras.callbacks import ReduceLROnPlateau from matplotlib import pyplot as plt from matplotlib.pyplot import figure import matplotlib.image as mpimg import pylab as pl from pylab import savefig plt.style.use('seaborn-deep') from sklearn.model_selection import train_test_split from sklearn.preprocessing import OneHotEncoder, StandardScaler,MinMaxScaler # Bayesian networks from sklearn.preprocessing import KBinsDiscretizer from pylab import * import pyAgrum as gum import pyAgrum.lib.notebook as gnb # for classification purposes from pyAgrum.lib.bn2roc import showROC import warnings warnings.simplefilter(action='ignore', category=FutureWarning) from sklearn.metrics import roc_curve, auc from matplotlib import pyplot as plt from matplotlib.pyplot import figure import matplotlib.image as mpimg import pylab as pl from pylab import savefig plt.style.use('seaborn-deep') # RECALL ----------------------------------------------------------------------------- # def recall_m(y_true, y_pred): """Computes the recal measure of an evaluation setting Parameters ---------- y_true : list list of groundtruth labels y_pred : list list of predictions from blackbox Returns ------- recall : vector a vector with the recall values between the predictions and the groundtruths """ true_positives = K.sum(K.round(K.clip(y_true * y_pred, 0, 1))) possible_positives = K.sum(K.round(K.clip(y_true, 0, 1))) recall = true_positives / (possible_positives + K.epsilon()) return recall # PRECISION --------------------------------------------------------------------------- # def precision_m(y_true, y_pred): """Computes the precision measure of an evaluation setting Parameters ---------- y_true : list list of groundtruth labels y_pred : list list of predictions from blackbox Returns ------- precision : vector a vector with the precision values between the predictions and the groundtruths """ true_positives = K.sum(K.round(K.clip(y_true * y_pred, 0, 1))) predicted_positives = K.sum(K.round(K.clip(y_pred, 0, 1))) precision = true_positives / (predicted_positives + K.epsilon()) return precision # F1 ------------------------------------------------------------------------------------ # Computes the F1 measure of an evaluation setting # y_true: list of groundtruth labels # y_pred: list of predictions from blackbox def f1_m(y_true, y_pred): precision = precision_m(y_true, y_pred) recall = recall_m(y_true, y_pred) return 2*((precision*recall)/(precision+recall+K.epsilon())) # CREATE_MODEL -------------------------------------------------------------------------- # creates a neural network model with a certain number of hidden layers and a certain # number of neurons in each layer. # input_dim: an integer specifying the number of input neurons # output_dim: an integer specifying the number of output neurons (the number of labels) # hidden_layers: an integer specifying the number of hidden layers # loss_func: the loss function of the model. By default, it is applied the 'categorical_crossentropy' # optim: the optimisation algorithm used in the model. By default it is used the 'nadam' algorithm # metrics: a list of strings specifying the metrics to be evaluated ('accuracy', 'f1', 'recall','precision') def create_model(input_dim, output_dim, nodes, hidden_layers=1, loss_func='categorical_crossentropy', optim='nadam', metrics=['accuracy'], name='model'): model = Sequential(name=name) model.add( Dense(nodes, input_dim=input_dim, activation='relu')) # input layer for i in range(hidden_layers): # hidden layers model.add(Dense(nodes, activation='relu')) model.add(Dense(output_dim, activation='softmax')) # output layer if( optim == "nadam" ): # Compile model optim = keras.optimizers.Nadam(lr=0.0001, beta_1=0.9, beta_2=0.999) model.compile(loss=loss_func, optimizer=optim, metrics=metrics) return model # GRID_SEARCH ----------------------------------------------------------------------------- # Generates a set of models with different configurations, ranging from an # initial number of neurons to a maximum number of neurons # start_nodes: an integer specifying the initial number of neurons to generate a model from # max_nodes: an integer specifying the maximum number of neurons to generate a model from # max_hlayers: an integer specifying the maximum number of hidden layers to generate a model from # debug: boolean that acts as a flag. If True, it displays the characteristics of each model # metrics: a list of strings with the metrics to be evaluated def grid_search_model_generator(n_features, n_classes, start_nodes = 1, max_nodes = 12, max_hlayers = 5, debug = False, metrics = ['accuracy'] ): models = [] # generate different models with different neurons and different hidden layers for neurons in range(start_nodes, max_nodes+1): for hidden_layer in range(1, max_hlayers+1): model_name = "model_h" + str(hidden_layer) + "_N"+str(neurons) model = create_model(n_features, n_classes, neurons, hidden_layer, name=model_name, metrics = metrics) models.append( model ) # add the generated model to a list # plot general information for each model if( debug ): for model in models: model.summary() return models # PERFORM_GRID_SEARCH ------------------------------------------------------------------- # given a list of models with different configurations, fit the data to the models, # and evaluate the model. This function returns a list of training histories for each model # models: list of models # X_train: # Y_train: # X_validation: # Y_validation: # X_test: # Y_test: # batch_size: # epochs: def perform_grid_search( models, path, dataset_name, X_train, Y_train, X_validation, Y_validation, X_test, Y_test, batch_size, epochs ): HISTORY_DICT = {} # define the callebacks to take into consideration during training # stop training when convergence is achieved after 10 iterations early_stop = EarlyStopping(monitor='val_loss', patience=10, verbose=1, mode='min') # save the best model after every epoch model_checkpoint = ModelCheckpoint(path + "training/" + dataset_name + "/model_{epoch:02d}-{val_loss:.2f}.h5", monitor='val_loss', verbose=0, save_best_only=True, mode='min') callbacks_list = [early_stop, model_checkpoint] # grid search over each model for model in models: print('MODEL NAME:', model.name) history_callback = model.fit(X_train, Y_train, batch_size = batch_size, epochs = epochs, verbose=0, validation_data=(X_validation, Y_validation), callbacks=callbacks_list) score_test = model.evaluate( X_test, Y_test, verbose=0 ) score_train = model.evaluate( X_train, Y_train ) print('Test loss: ', format(score_test[0], '.4f'), '\tTrain loss: ', format(score_train[0], '.4f') ) print('Test accuracy: ', format(score_test[1], '.4f'), '\tTrain accu: ', format(score_train[1], '.4f') ) print('Abs accuracy: ', format( np.abs( score_test[1] - score_train[1] ), '.4f')) print('Abs loss: ', format( np.abs( score_test[0] - score_train[0] ), '.4f')) print('\n###########################################################\n') HISTORY_DICT[model.name] = [history_callback, model] return HISTORY_DICT # SAVE_MODEL ----------------------------------------------------------------------------- # saves a trained model into a json and hdf5 file # model: model to be saved # model_name: string with model name # path: string with path to save def save_model( model, model_name, path ): # serialize model to JSON model_json = model.to_json() with open(path + model_name+"_DUO.json", "w") as json_file: json_file.write(model_json) json_file.close() # serialize weights to HDF5 model.save_weights( path + model_name+"_DUO.h5") print("Saving files:") print(path + model_name+"_DUO.json") print(path + model_name+"_DUO.h5") print("Model saved to disk") # SAVE_MODEL_HISTORY ------------------------------------------------------------------- # saves a trained model into a csv file # model_hist: history of the model to be saved # model_name: string with model name # path: string with path to save def save_model_history( model_hist, model_name, path ): file = open(path + model_name + "_hist.csv", "w") w = csv.writer( file ) for key, val in model_hist.history.items(): w.writerow([key, val]) file.close() print(path + model_name+"_DUO.h5") print("Model history saved to disk") # LOAD_MODEL_HISTORY ------------------------------------------ # loads a saved model history into memory # model_name: the name of the model # path: path to model history def load_model_history( model_name, path): model_hist_loaded = {} values = [] # load dictionary r = open( path + model_name + "_hist.csv", "r").read() for line in r.split("\n"): if(len(line) == 0): continue metric = line.split(",\"[")[0] # extract metrics values_str = line.split(",\"[")[1].replace("]\"","").split(", ") # extract validation values values = [float(val_str) for val_str in values_str] model_hist_loaded.update( {metric : values} ) return model_hist_loaded # LOAD_MODEL ------------------------------------------ # loads a saved model into memory # model_name: the name of the model # path: path to model history def load_model( model_name, path ): json_file = open( path + model_name + "_DUO.json", 'r') loaded_model_json = json_file.read() json_file.close() # load weights into new model loaded_model = model_from_json(loaded_model_json) loaded_model.load_weights(path + model_name + "_DUO.h5") print("Loaded model from disk") return loaded_model def plot_model_history( model_history, metric ): plt.plot(model_history[ metric.lower() ], label='train') plt.plot(model_history["val_" + metric.lower()], label='validation') plt.ylabel(metric) plt.xlabel('Number of Epochs') plt.ylim([0, 1]) plt.legend() plt.show() def plot_ROC_Curve( model, X, Y, n_classes): Y_pred_proba = model.predict(X) # Compute ROC curve and ROC area for each class fpr = dict() tpr = dict() roc_auc = dict() for i in range(n_classes): fpr[i], tpr[i], _ = roc_curve(Y[:, i], Y_pred_proba[:, i]) roc_auc[i] = auc(fpr[i], tpr[i]) # Plot of a ROC curve for a specific class for i in range(n_classes): plt.figure() plt.plot(fpr[i], tpr[i], label='ROC curve (area = %0.2f)' % roc_auc[i]) plt.plot([0, 1], [0, 1], 'k--') plt.xlim([0.0, 1.0]) plt.ylim([0.0, 1.05]) plt.xlabel('False Positive Rate') plt.ylabel('True Positive Rate') plt.title('Receiver operating characteristic example') plt.legend(loc="lower right") plt.show() # ENCODE_DATA -------------------------------------------------------------------------- # Applies one hot encoder to data # data: a dataframe # class_var: string with class variable name def encode_data(data, class_var): feature_names = data.drop([class_var], axis=1).columns.tolist() X = data[ feature_names ].values y = data[class_var].values n_features = X.shape[1] n_classes = len(data[class_var].unique()) # create numerical encoding for attribute species enc = OneHotEncoder() Y = enc.fit_transform(y[:, np.newaxis]).toarray() # Scale data to have mean 0 and variance 1 # which is importance for convergence of the neural network scaler = MinMaxScaler() X_scaled = scaler.fit_transform(X) return X_scaled, Y, enc, scaler # LOAD_TRAINING_DATA --------------------------------------------------------------------- # loads into a multiarray format a training set previously saved in a .csv file # dataset_path: string containing the path where the files will be saved def load_training_data( dataset_path ): X_train = pd.read_csv(dataset_path.replace(".csv", "") + "_Xtrain.csv", index_col=False).values X_test = pd.read_csv(dataset_path.replace(".csv", "") + "_Xtest.csv", index_col=False).values X_validation =pd.read_csv(dataset_path.replace(".csv", "") + "_Xvalidation.csv",index_col=False).values Y_train = pd.read_csv(dataset_path.replace(".csv", "") + "_Ytrain.csv",index_col=False).values Y_test =pd.read_csv(dataset_path.replace(".csv", "") + "_Ytest.csv", index_col=False).values Y_validation = pd.read_csv(dataset_path.replace(".csv", "") + "_Yvalidation.csv", index_col=False).values return X_train, Y_train, X_test, Y_test, X_validation, Y_validation # GENERATE_SAVE_TRAINING_DATA ------------------------------------------------------------ # # dataset_path: string containing the path where the files will be saved # X: NxM matrix representing the training data # Y: NxC matrix representing the OneHotEconder of C classes def generate_save_training_data( dataset_path, X, Y): """Generates training, test and validation sets and stores this information into files Parameters ---------- dataset_path : str The file location of the spreadsheet samples : int, optional The number of permutations to generate from the original vector (default is 300) variance : int, optional Quantity to permute in each feature (default is 0.25) Returns ------- permutations : matrix a 2-D matrix with dimensions (samples, features) with all the permutations of the original vector """ # generate train, test and validation sets X_train, X_test, Y_train, Y_test = train_test_split(X, Y, test_size=0.3, random_state=515) X_validation, X_test, Y_validation, Y_test = train_test_split(X_test, Y_test, test_size=0.5, random_state=515) np.savetxt(dataset_path.replace(".csv", "") + "_Xtrain.csv", X_train, delimiter=",") np.savetxt(dataset_path.replace(".csv", "") + "_Xtest.csv", X_test, delimiter=",") np.savetxt(dataset_path.replace(".csv", "") + "_Xvalidation.csv", X_validation, delimiter=",") np.savetxt(dataset_path.replace(".csv", "") + "_Ytrain.csv", Y_train, delimiter=",") np.savetxt(dataset_path.replace(".csv", "") + "_Ytest.csv", Y_test, delimiter=",") np.savetxt(dataset_path.replace(".csv", "") + "_Yvalidation.csv", Y_validation, delimiter=",") # GENERATE_CONFUSION_MATRIX_DATA --------------------------------- ############################################################################## # BAYESIAN NETWORK EXPLANATIONS # ############################################################################## def compute_perm_range(feat, variance = 0.25): """ Parameters ---------- feat : float Value of a feature to be permuted samples : int, optional The number of permutations to generate from the original vector (default is 300) variance : int, optional Quantity to permute in each feature (default is 0.25) Returns ------- min_range : float minimum value that a feature can be permuted max_range : float maximum value that a feature can be permuted """ min_range = feat - variance max_range = feat + variance # features are scaled between 0 and 1 # if the permutation make the feature negative, this values is set to 0 if( min_range < 0 ): min_range = 0 # if the permutation make the feature bigger than 1, this values is set to 1 if( max_range > 1 ): max_range = 1 return min_range, max_range # PERMUTE_SINGLE_FEATURES_____________________________________________________________ # def permute_single_features( my_array, samples = 300, variance = 0.25 ): """Given a single array from which one pretends to generate local explanations from Draw samples from a uniform distribution within a range of feature_val +- variance Returns a matrix with a number of samples (by default 300) with permutations of each feature of the input vector Parameters ---------- my_array : np.array The datapoint to be locally explained samples : int, optional The number of permutations to generate from the original vector (default is 300) variance : int, optional Quantity to permute in each feature (default is 0.25) Returns ------- permutations : matrix a 2-D matrix with dimensions (samples, features) with all the permutations of the original vector """ # permutation result list permutations = [] # just keeping a controlled number of decimal places my_array = np.round(my_array,4) # keep a copy of the original array, since we will be changing the features my_array_backup = my_array # extract number of features num_features = my_array.shape[0] # add original vector to dataframe permutations.append( my_array_backup.tolist() ) # for each feature of the input feature vector, for feat in range(0, num_features): # get feature value my_array = my_array_backup feature_val = my_array[feat] # set permutation of feature between [ feat - variance ; feat + variance ] min_range, max_range = compute_perm_range( feature_val, variance ) # generate sample of random features within a range for perm in range(0, int(round(samples/num_features, 0))): # set the new vector my_array[feat] = np.abs(np.round(rn.uniform(min_range, max_range),4)) permutations.append( my_array.tolist() ) #rn.shuffle(permutations) return permutations # LEARNBN ------------------------------------------- # def learnBN( file_path, algorithm = "Hill Climbing" ): """Given a single array from which one pretends to generate local explanations from Draw samples from a uniform distribution within a range of feature_val +- variance Returns a matrix with a number of samples (by default 300) with permutations of each feature of the input vector Parameters ---------- my_array : np.array The datapoint to be locally explained samples : int, optional The number of permutations to generate from the original vector (default is 300) variance : int, optional Quantity to permute in each feature (default is 0.25) Returns ------- permutations : matrix a 2-D matrix with dimensions (samples, features) with all the permutations of the original vector """ learner = gum.BNLearner( file_path ) if( algorithm == "Hill Climbing"): print("Selecting Greedy Hill Climbing Algorithm") learner.useGreedyHillClimbing() if( algorithm == "Local Search" ): print("Selecting Local Search Algorithm") bn = learner.useLocalSearchWithTabuList() if( algorithm == "3off2"): print("Selecting 3Off2 Algorithm") learner.use3off2() if( algorithm == "miic" ): print("Selecting MIIC Algorithm") learner.useMIIC() learner.learnBN() bn = learner.learnBN() essencGraph = gum.EssentialGraph( bn ) infoBN = gnb.getInformation( bn ) return [ bn, infoBN, essencGraph ] # DISCRETIZE_DATAFRAME ------------------------------------------------------- # # def discretize_dataframe( df, class_var, num_bins=4 ): """Given a dataframe with continuous values, convert the continuous values into discrete ones by splitting the data into bins and by computing the respective quartiles Parameters ---------- df : pd.DataFrame The datapoint to be locally explained class_var : str The number of permutations to generate from the original vector (default is 300) num_bins : int, optional Quantity to permute in each feature (default is 0.25) Returns ------- permutations : matrix a 2-D matrix with dimensions (samples, features) with all the permutations of the original vector """ r=np.array(range(num_bins+1))/(1.0*num_bins) # quantiles are building using pandas.qcut # The "class" column is just copied. l=[] for col in df.columns.values: if col!=class_var: l.append( pd.DataFrame( pd.qcut( df[col],r, duplicates='drop',precision=2),columns=[col])) else: l.append( pd.DataFrame( df[col].values,columns=[col])) treated = pd.concat(l, join='outer', axis=1) return treated # SAVE_DISCRETIZED_DATAFRAME --------------------------------------------------- # def save_discretized_dataframe(indx, df_model, model_type, perm_type, bins, dataset_name, path, class_var): """Given a single array from which one pretends to generate local explanations from Draw samples from a uniform distribution within a range of feature_val +- variance Returns a matrix with a number of samples (by default 300) with permutations of each feature of the input vector Parameters ---------- my_array : np.array The datapoint to be locally explained samples : int, optional The number of permutations to generate from the original vector (default is 300) variance : int, optional Quantity to permute in each feature (default is 0.25) Returns ------- permutations : matrix a 2-D matrix with dimensions (samples, features) with all the permutations of the original vector """ file_path = path + dataset_name + "/" + str(indx) + "/" + re.sub( r"\.\w+", "", dataset_name ) + "_" + model_type +"_INDX_" + str(indx) + "_" + perm_type +".csv" df_discr = discretize_dataframe( df_model, bins, class_var ) print("Saving discretized dataset into: %s\n" %(file_path)) df_discr.to_csv( file_path, index=False) # GENERATE_PERMUTATIONS ------------------------------------------- # def generate_permutations( instance, labels_lst, feature_names, class_var, encoder, scaler, model, samples = 300, variance = 0.25): # get datapoint in scaled feature space local_datapoint = np.array(instance['scaled_vector']) # get datapoint in original feature space local_datapoint_orig = np.array(instance['original_vector']) # permute features permutations = permute_single_features( local_datapoint, samples = samples, variance = variance ) # convert permutations to original feature space permutations_orig = scaler.inverse_transform( permutations ) # compute predictions for each permuted instance predictions = encoder.inverse_transform( model.predict( permutations ) ) # convert prediction classes to labels labelled_predictions = [ labels_lst[ int(predictions[indx][0]) ] for indx in range(0, len(predictions))] # add all this information to a single dataframe df_local_permutations = pd.DataFrame( permutations_orig, columns = feature_names ) # add class variable to dataframe df_local_permutations[ class_var ] = labelled_predictions return df_local_permutations # GEBERATE_BN_EXPLANATIONS ------------------------------------------------------------ # def generate_BN_explanations(instance, label_lst, feature_names, class_var, encoder, scaler, model, path, dataset_name ): # necessary for starting Numpy generated random numbers in an initial state np.random.seed(515) # Necessary for starting core Python generated random numbers in a state rn.seed(515) indx = instance['index'] prediction_type = instance['prediction_type'].lower()+"s" prediction_type = prediction_type.replace(" ", "_") # generate permutations df = generate_permutations( instance, label_lst, feature_names, class_var, encoder, scaler, model) # discretize data df_discr = discretize_dataframe( df, class_var, num_bins=4 ) # save discretised dataframe (for debugging and reproduceability purposes) path_to_permutations = path + "feature_permutations/" + dataset_name.replace(".csv","") + "/" + prediction_type + "/" + str(indx) + ".csv" df_discr.to_csv( path_to_permutations, index=False) # normalise dataframe normalise_dataframe( path_to_permutations ) # learn BN bn, infoBN, essencGraph = learnBN( path_to_permutations.replace(".csv", "_norm.csv") ) # perform inference inference = gnb.getInference(bn, evs={},targets=df_discr.columns.to_list(), size='12') # show networks gnb.sideBySide(*[bn, inference, infoBN ], captions=[ "Bayesian Network", "Inference", "Information Network" ]) # save to file path_to_explanation = path + "explanations/" + dataset_name.replace(".csv", "") + "/BN/" + prediction_type + "/" gum.lib.bn2graph.dotize( bn , path_to_explanation + str(indx) + "_BN" ) gum.saveBN(bn,path_to_explanation + str(indx) + "_BN.net" ) return [bn, inference, infoBN] # GEBERATE_BN_EXPLANATIONSMB ------------------------------------------------------------ # def generate_BN_explanationsMB(instance, label_lst, feature_names, class_var, encoder, scaler, model, path, dataset_name, variance = 0.25, algorithm = "Hill Climbing" ): # necessary for starting Numpy generated random numbers in an initial state np.random.seed(515) # Necessary for starting core Python generated random numbers in a state rn.seed(515) indx = instance['index'] prediction_type = instance['prediction_type'].lower()+"s" prediction_type = prediction_type.replace(" ", "_") # generate permutations df = generate_permutations( instance, label_lst, feature_names, class_var, encoder, scaler, model, variance = variance) # discretize data df_discr = discretize_dataframe( df, class_var, num_bins=4 ) # save discretised dataframe (for debugging and reproduceability purposes) path_to_permutations = path + "feature_permutations/" + dataset_name.replace(".csv","") + "/" + prediction_type + "/" + str(indx) + ".csv" df_discr.to_csv( path_to_permutations, index=False) # normalise dataframe normalise_dataframe( path_to_permutations ) # learn BN bn, infoBN, essencGraph = learnBN( path_to_permutations.replace(".csv", "_norm.csv"), algorithm = algorithm) # perform inference inference = gnb.getInference(bn, evs={},targets=df_discr.columns.to_list(), size='12') # compute Markov Blanket markov_blanket = gum.MarkovBlanket(bn, class_var) # show networks gnb.sideBySide(*[bn, inference, markov_blanket ], captions=[ "Bayesian Network", "Inference", "Markov Blanket" ]) # save to file path_to_explanation = path + "explanations/" + dataset_name.replace(".csv", "") + "/BN/" + prediction_type + "/" gum.lib.bn2graph.dotize( bn , path_to_explanation + str(indx) + "_BN" ) gum.saveBN(bn,path_to_explanation + str(indx) + "_BN.net" ) return [bn, inference, infoBN, markov_blanket] # GENERATE_LOCAL_PREDICTIONS ------------------------------------------- # def generate_local_predictions( X, Y, model, scaler, encoder ): # get original vector orig_vec = np.round(scaler.inverse_transform(X),6) # generate all predictions for X predictions = model.predict( X ) # extrace the label of the prediction of X[indx] prediction_class = encoder.inverse_transform( predictions ) local_data_dict = [] for indx in range(0, orig_vec.shape[0]): ground_truth = np.expand_dims(Y[indx], axis=0) ground_truth_class = encoder.inverse_transform( ground_truth )[0][0] prediction = prediction_class[indx][0] # check if data point is a true positive if( ( int(prediction) == int(ground_truth_class) ) & (int(prediction)==1) & (int(ground_truth_class)==1) ): pred_type = "TRUE POSITIVE" # check if data point is a true negative if( ( int(prediction) == int(ground_truth_class) ) & (int(prediction)==0) & (int(ground_truth_class)==0) ): pred_type = "TRUE NEGATIVE" # check if data point is a false negative if( ( int(prediction) != int(ground_truth_class) ) & (int(prediction)==0) & (int(ground_truth_class)==1) ): pred_type = "FALSE NEGATIVE" # check if data point is a false positve if( ( int(prediction) != int(ground_truth_class) ) & (int(prediction)==1) & (int(ground_truth_class)==0) ): pred_type = "FALSE POSITIVE" local_data_dict.append( {'index' : indx, 'original_vector' : orig_vec[indx,:].tolist(), 'scaled_vector' : X[indx,:].tolist(), 'ground_truth' : ground_truth_class, 'predictions' : prediction, 'prediction_type' : pred_type}) return local_data_dict ################################################################################## # TEXT PROCESSING # # ############################################################################### # FIND ----------------------------------------------- # def find(s, ch): return [i for i, ltr in enumerate(s) if ltr == ch] # UNTOKENIZE ----------------------------------------------- # def untokenize( tokens, delim ): untokenized = tokens[0] for indx in range(1, len(tokens)): untokenized = untokenized + delim + tokens[indx] return untokenized # NORMALISE_LINE ------------------------------------------- # def normalise_line( my_str, class_label ): my_str = my_str.replace("\","+class_label, "") my_str = my_str.replace("-1e-05", "0.0000") tokens = my_str.split("\",\"") tokens_norm = [] for token in tokens: token = token.replace("]","") indxs = find(token, ".") indx_comma = find(token, ",")[0]+2 if( (len(token[indxs[1]+1 : -1 ]) >= 4) & (len( token[indxs[0]+1 : indx_comma-2 ]) >= 4) ): token_temp = token[0:indxs[0]] + "." + token[indxs[0] + 1 : indxs[0]+5] + ", " +token[indx_comma:indxs[1]] + token[indxs[1] : indxs[1]+5 ] + "]" if( (len(token[indxs[1]+1 : -1 ]) < 4) & (len( token[indxs[0]+1 : indx_comma-2 ]) >= 4) ): extra = "0"*(np.abs(len(token[indxs[1]+1 : -1 ]) - 4)) token_temp = token[0:indxs[0]] + "." + token[indxs[0] + 1 : indxs[0]+5] + ", " +token[indx_comma:indxs[1]] + token[indxs[1] : -1 ] + extra + "]" if( (len(token[indxs[1]+1 : -1 ]) >= 4) & (len( token[indxs[0]+1 : indx_comma-2 ]) < 4) ): extra = "0"*(np.abs(len( token[indxs[0]+1 : indx_comma-2 ]) - 4)) token_temp = token[0:indxs[0]] + "." + extra + ", " +token[indx_comma:indxs[1]] + token[indxs[1] : -1 ] + extra + "]" if( (len(token[indxs[1]+1 : -1 ]) < 4) & (len( token[indxs[0]+1 : indx_comma-2 ]) < 4) ): extra2 = "0"*(np.abs(len(token[indxs[1]+1 : -1 ]) - 4)) extra1 = "0"*(np.abs(len(token[indxs[0]+1 : -1 ]) - 4)) token_temp = token[0:indxs[0]] + "." + extra1 + ", " +token[indx_comma:indxs[1]] + token[indxs[1] : -1 ] + extra2 + "]" tokens_norm.append(token_temp) return untokenize( tokens_norm, "\",\"") + "\"," +class_label # NORMALISE_LINE ------------------------------------------- # def normalise_dataframe( path_to_permutations ): file = open(path_to_permutations,"r") f_write = open(path_to_permutations.replace(".csv", "_norm.csv"),"w") header = file.readline().replace("\n","") f_write.write( header + "\n") for line in file.readlines(): # get class class_label = line.split("\",")[-1].replace("\n","") # normalise dataframe input line_norm = normalise_line( line.replace("\n",""), class_label ) # write normalised input to file f_write.write(line_norm + "\n") file.close() f_write.close()
33,430
37.783063
177
py
LM-LSTM-CRF
LM-LSTM-CRF-master/eval_wc.py
from __future__ import print_function import datetime import time import torch import torch.autograd as autograd import torch.nn as nn import torch.optim as optim import codecs from model.crf import * from model.lm_lstm_crf import * import model.utils as utils from model.evaluator import eval_wc import argparse import json import os import sys from tqdm import tqdm import itertools import functools if __name__ == "__main__": parser = argparse.ArgumentParser(description='Evaluating LM-BLSTM-CRF') parser.add_argument('--load_arg', default='./checkpoint/soa/check_wc_p_char_lstm_crf.json', help='path to arg json') parser.add_argument('--load_check_point', default='./checkpoint/soa/check_wc_p_char_lstm_crf.model', help='path to model checkpoint file') parser.add_argument('--gpu',type=int, default=0, help='gpu id') parser.add_argument('--eva_matrix', choices=['a', 'fa'], default='fa', help='use f1 and accuracy or f1 alone') parser.add_argument('--test_file', default='', help='path to test file, if set to none, would use test_file path in the checkpoint file') args = parser.parse_args() with open(args.load_arg, 'r') as f: jd = json.load(f) jd = jd['args'] checkpoint_file = torch.load(args.load_check_point, map_location=lambda storage, loc: storage) f_map = checkpoint_file['f_map'] l_map = checkpoint_file['l_map'] c_map = checkpoint_file['c_map'] in_doc_words = checkpoint_file['in_doc_words'] if args.gpu >= 0: torch.cuda.set_device(args.gpu) # load corpus if args.test_file: with codecs.open(args.test_file, 'r', 'utf-8') as f: test_lines = f.readlines() else: with codecs.open(jd['test_file'], 'r', 'utf-8') as f: test_lines = f.readlines() # converting format test_features, test_labels = utils.read_corpus(test_lines) # construct dataset test_dataset, forw_test, back_test = utils.construct_bucket_mean_vb_wc(test_features, test_labels, l_map, c_map, f_map, jd['caseless']) test_dataset_loader = [torch.utils.data.DataLoader(tup, 50, shuffle=False, drop_last=False) for tup in test_dataset] # build model ner_model = LM_LSTM_CRF(len(l_map), len(c_map), jd['char_dim'], jd['char_hidden'], jd['char_layers'], jd['word_dim'], jd['word_hidden'], jd['word_layers'], len(f_map), jd['drop_out'], large_CRF=jd['small_crf'], if_highway=jd['high_way'], in_doc_words=in_doc_words, highway_layers = jd['highway_layers']) ner_model.load_state_dict(checkpoint_file['state_dict']) if args.gpu >= 0: if_cuda = True torch.cuda.set_device(args.gpu) ner_model.cuda() packer = CRFRepack_WC(len(l_map), True) else: if_cuda = False packer = CRFRepack_WC(len(l_map), False) evaluator = eval_wc(packer, l_map, args.eva_matrix) print('start') if 'f' in args.eva_matrix: result = evaluator.calc_score(ner_model, test_dataset_loader) for label, (test_f1, test_pre, test_rec, test_acc, msg) in result.items(): print(jd['checkpoint'] +' : %s : test_f1: %.4f test_rec: %.4f test_pre: %.4f test_acc: %.4f | %s\n' % (label, test_f1, test_rec, test_pre, test_acc, msg)) else: test_acc = evaluator.calc_score(ner_model, test_dataset_loader) print(jd['checkpoint'] + ' test_acc: %.4f\n' % (test_acc)) print('end')
3,402
36.395604
307
py
LM-LSTM-CRF
LM-LSTM-CRF-master/train_wc.py
from __future__ import print_function import datetime import time import torch import torch.autograd as autograd import torch.nn as nn import torch.optim as optim import codecs from model.crf import * from model.lm_lstm_crf import * import model.utils as utils from model.evaluator import eval_wc import argparse import json import os import sys from tqdm import tqdm import itertools import functools def eprint(*args, **kwargs): print(*args, file=sys.stderr, **kwargs) if __name__ == "__main__": parser = argparse.ArgumentParser(description='Learning with LM-LSTM-CRF together with Language Model') parser.add_argument('--rand_embedding', action='store_true', help='random initialize word embedding') parser.add_argument('--emb_file', default='./embedding/glove.6B.100d.txt', help='path to pre-trained embedding') parser.add_argument('--train_file', default='./data/ner/eng.train.iobes', help='path to training file') parser.add_argument('--dev_file', default='./data/ner/eng.testa.iobes', help='path to development file') parser.add_argument('--test_file', default='./data/ner/eng.testb.iobes', help='path to test file') parser.add_argument('--gpu', type=int, default=0, help='gpu id') parser.add_argument('--batch_size', type=int, default=10, help='batch_size') parser.add_argument('--unk', default='unk', help='unknow-token in pre-trained embedding') parser.add_argument('--char_hidden', type=int, default=300, help='dimension of char-level layers') parser.add_argument('--word_hidden', type=int, default=300, help='dimension of word-level layers') parser.add_argument('--drop_out', type=float, default=0.55, help='dropout ratio') parser.add_argument('--epoch', type=int, default=200, help='maximum epoch number') parser.add_argument('--start_epoch', type=int, default=0, help='start point of epoch') parser.add_argument('--checkpoint', default='./checkpoint/', help='checkpoint path') parser.add_argument('--caseless', action='store_true', help='caseless or not') parser.add_argument('--char_dim', type=int, default=30, help='dimension of char embedding') parser.add_argument('--word_dim', type=int, default=100, help='dimension of word embedding') parser.add_argument('--char_layers', type=int, default=1, help='number of char level layers') parser.add_argument('--word_layers', type=int, default=1, help='number of word level layers') parser.add_argument('--lr', type=float, default=0.015, help='initial learning rate') parser.add_argument('--lr_decay', type=float, default=0.05, help='decay ratio of learning rate') parser.add_argument('--fine_tune', action='store_false', help='fine tune the diction of word embedding or not') parser.add_argument('--load_check_point', default='', help='path previous checkpoint that want to be loaded') parser.add_argument('--load_opt', action='store_true', help='also load optimizer from the checkpoint') parser.add_argument('--update', choices=['sgd', 'adam'], default='sgd', help='optimizer choice') parser.add_argument('--momentum', type=float, default=0.9, help='momentum for sgd') parser.add_argument('--clip_grad', type=float, default=5.0, help='clip grad at') parser.add_argument('--small_crf', action='store_false', help='use small crf instead of large crf, refer model.crf module for more details') parser.add_argument('--mini_count', type=float, default=5, help='thresholds to replace rare words with <unk>') parser.add_argument('--lambda0', type=float, default=1, help='lambda0') parser.add_argument('--co_train', action='store_true', help='cotrain language model') parser.add_argument('--patience', type=int, default=15, help='patience for early stop') parser.add_argument('--high_way', action='store_true', help='use highway layers') parser.add_argument('--highway_layers', type=int, default=1, help='number of highway layers') parser.add_argument('--eva_matrix', choices=['a', 'fa'], default='fa', help='use f1 and accuracy or accuracy alone') parser.add_argument('--least_iters', type=int, default=50, help='at least train how many epochs before stop') parser.add_argument('--shrink_embedding', action='store_true', help='shrink the embedding dictionary to corpus (open this if pre-trained embedding dictionary is too large, but disable this may yield better results on external corpus)') args = parser.parse_args() if args.gpu >= 0: torch.cuda.set_device(args.gpu) print('setting:') print(args) # load corpus print('loading corpus') with codecs.open(args.train_file, 'r', 'utf-8') as f: lines = f.readlines() with codecs.open(args.dev_file, 'r', 'utf-8') as f: dev_lines = f.readlines() with codecs.open(args.test_file, 'r', 'utf-8') as f: test_lines = f.readlines() dev_features, dev_labels = utils.read_corpus(dev_lines) test_features, test_labels = utils.read_corpus(test_lines) if args.load_check_point: if os.path.isfile(args.load_check_point): print("loading checkpoint: '{}'".format(args.load_check_point)) checkpoint_file = torch.load(args.load_check_point) args.start_epoch = checkpoint_file['epoch'] f_map = checkpoint_file['f_map'] l_map = checkpoint_file['l_map'] c_map = checkpoint_file['c_map'] in_doc_words = checkpoint_file['in_doc_words'] train_features, train_labels = utils.read_corpus(lines) else: print("no checkpoint found at: '{}'".format(args.load_check_point)) else: print('constructing coding table') # converting format train_features, train_labels, f_map, l_map, c_map = utils.generate_corpus_char(lines, if_shrink_c_feature=True, c_thresholds=args.mini_count, if_shrink_w_feature=False) f_set = {v for v in f_map} f_map = utils.shrink_features(f_map, train_features, args.mini_count) if args.rand_embedding: print("embedding size: '{}'".format(len(f_map))) in_doc_words = len(f_map) else: dt_f_set = functools.reduce(lambda x, y: x | y, map(lambda t: set(t), dev_features), f_set) dt_f_set = functools.reduce(lambda x, y: x | y, map(lambda t: set(t), test_features), dt_f_set) print("feature size: '{}'".format(len(f_map))) print('loading embedding') if args.fine_tune: # which means does not do fine-tune f_map = {'<eof>': 0} f_map, embedding_tensor, in_doc_words = utils.load_embedding_wlm(args.emb_file, ' ', f_map, dt_f_set, args.caseless, args.unk, args.word_dim, shrink_to_corpus=args.shrink_embedding) print("embedding size: '{}'".format(len(f_map))) l_set = functools.reduce(lambda x, y: x | y, map(lambda t: set(t), dev_labels)) l_set = functools.reduce(lambda x, y: x | y, map(lambda t: set(t), test_labels), l_set) for label in l_set: if label not in l_map: l_map[label] = len(l_map) print('constructing dataset') # construct dataset dataset, forw_corp, back_corp = utils.construct_bucket_mean_vb_wc(train_features, train_labels, l_map, c_map, f_map, args.caseless) dev_dataset, forw_dev, back_dev = utils.construct_bucket_mean_vb_wc(dev_features, dev_labels, l_map, c_map, f_map, args.caseless) test_dataset, forw_test, back_test = utils.construct_bucket_mean_vb_wc(test_features, test_labels, l_map, c_map, f_map, args.caseless) dataset_loader = [torch.utils.data.DataLoader(tup, args.batch_size, shuffle=True, drop_last=False) for tup in dataset] dev_dataset_loader = [torch.utils.data.DataLoader(tup, 50, shuffle=False, drop_last=False) for tup in dev_dataset] test_dataset_loader = [torch.utils.data.DataLoader(tup, 50, shuffle=False, drop_last=False) for tup in test_dataset] # build model print('building model') ner_model = LM_LSTM_CRF(len(l_map), len(c_map), args.char_dim, args.char_hidden, args.char_layers, args.word_dim, args.word_hidden, args.word_layers, len(f_map), args.drop_out, large_CRF=args.small_crf, if_highway=args.high_way, in_doc_words=in_doc_words, highway_layers = args.highway_layers) if args.load_check_point: ner_model.load_state_dict(checkpoint_file['state_dict']) else: if not args.rand_embedding: ner_model.load_pretrained_word_embedding(embedding_tensor) ner_model.rand_init(init_word_embedding=args.rand_embedding) if args.update == 'sgd': optimizer = optim.SGD(ner_model.parameters(), lr=args.lr, momentum=args.momentum) elif args.update == 'adam': optimizer = optim.Adam(ner_model.parameters(), lr=args.lr) if args.load_check_point and args.load_opt: optimizer.load_state_dict(checkpoint_file['optimizer']) crit_lm = nn.CrossEntropyLoss() crit_ner = CRFLoss_vb(len(l_map), l_map['<start>'], l_map['<pad>']) if args.gpu >= 0: if_cuda = True print('device: ' + str(args.gpu)) torch.cuda.set_device(args.gpu) crit_ner.cuda() crit_lm.cuda() ner_model.cuda() packer = CRFRepack_WC(len(l_map), True) else: if_cuda = False packer = CRFRepack_WC(len(l_map), False) tot_length = sum(map(lambda t: len(t), dataset_loader)) best_f1 = float('-inf') best_acc = float('-inf') track_list = list() start_time = time.time() epoch_list = range(args.start_epoch, args.start_epoch + args.epoch) patience_count = 0 evaluator = eval_wc(packer, l_map, args.eva_matrix) for epoch_idx, args.start_epoch in enumerate(epoch_list): epoch_loss = 0 ner_model.train() for f_f, f_p, b_f, b_p, w_f, tg_v, mask_v, len_v in tqdm( itertools.chain.from_iterable(dataset_loader), mininterval=2, desc=' - Tot it %d (epoch %d)' % (tot_length, args.start_epoch), leave=False, file=sys.stdout): f_f, f_p, b_f, b_p, w_f, tg_v, mask_v = packer.repack_vb(f_f, f_p, b_f, b_p, w_f, tg_v, mask_v, len_v) ner_model.zero_grad() scores = ner_model(f_f, f_p, b_f, b_p, w_f) loss = crit_ner(scores, tg_v, mask_v) epoch_loss += utils.to_scalar(loss) if args.co_train: cf_p = f_p[0:-1, :].contiguous() cb_p = b_p[1:, :].contiguous() cf_y = w_f[1:, :].contiguous() cb_y = w_f[0:-1, :].contiguous() cfs, _ = ner_model.word_pre_train_forward(f_f, cf_p) loss = loss + args.lambda0 * crit_lm(cfs, cf_y.view(-1)) cbs, _ = ner_model.word_pre_train_backward(b_f, cb_p) loss = loss + args.lambda0 * crit_lm(cbs, cb_y.view(-1)) loss.backward() nn.utils.clip_grad_norm_(ner_model.parameters(), args.clip_grad) optimizer.step() epoch_loss /= tot_length # update lr if args.update == 'sgd': utils.adjust_learning_rate(optimizer, args.lr / (1 + (args.start_epoch + 1) * args.lr_decay)) # eval & save check_point if 'f' in args.eva_matrix: dev_result = evaluator.calc_score(ner_model, dev_dataset_loader) for label, (dev_f1, dev_pre, dev_rec, dev_acc, msg) in dev_result.items(): print('DEV : %s : dev_f1: %.4f dev_rec: %.4f dev_pre: %.4f dev_acc: %.4f | %s\n' % (label, dev_f1, dev_rec, dev_pre, dev_acc, msg)) (dev_f1, dev_pre, dev_rec, dev_acc, msg) = dev_result['total'] if dev_f1 > best_f1: patience_count = 0 best_f1 = dev_f1 test_result = evaluator.calc_score(ner_model, test_dataset_loader) for label, (test_f1, test_pre, test_rec, test_acc, msg) in test_result.items(): print('TEST : %s : test_f1: %.4f test_rec: %.4f test_pre: %.4f test_acc: %.4f | %s\n' % (label, test_f1, test_rec, test_pre, test_acc, msg)) (test_f1, test_pre, test_rec, test_acc, msg) = test_result['total'] track_list.append( {'loss': epoch_loss, 'dev_f1': dev_f1, 'dev_acc': dev_acc, 'test_f1': test_f1, 'test_acc': test_acc}) print( '(loss: %.4f, epoch: %d, dev F1 = %.4f, dev acc = %.4f, F1 on test = %.4f, acc on test= %.4f), saving...' % (epoch_loss, args.start_epoch, dev_f1, dev_acc, test_f1, test_acc)) try: utils.save_checkpoint({ 'epoch': args.start_epoch, 'state_dict': ner_model.state_dict(), 'optimizer': optimizer.state_dict(), 'f_map': f_map, 'l_map': l_map, 'c_map': c_map, 'in_doc_words': in_doc_words }, {'track_list': track_list, 'args': vars(args) }, args.checkpoint + 'cwlm_lstm_crf') except Exception as inst: print(inst) else: patience_count += 1 print('(loss: %.4f, epoch: %d, dev F1 = %.4f, dev acc = %.4f)' % (epoch_loss, args.start_epoch, dev_f1, dev_acc)) track_list.append({'loss': epoch_loss, 'dev_f1': dev_f1, 'dev_acc': dev_acc}) else: dev_acc = evaluator.calc_score(ner_model, dev_dataset_loader) if dev_acc > best_acc: patience_count = 0 best_acc = dev_acc test_acc = evaluator.calc_score(ner_model, test_dataset_loader) track_list.append( {'loss': epoch_loss, 'dev_acc': dev_acc, 'test_acc': test_acc}) print( '(loss: %.4f, epoch: %d, dev acc = %.4f, acc on test= %.4f), saving...' % (epoch_loss, args.start_epoch, dev_acc, test_acc)) try: utils.save_checkpoint({ 'epoch': args.start_epoch, 'state_dict': ner_model.state_dict(), 'optimizer': optimizer.state_dict(), 'f_map': f_map, 'l_map': l_map, 'c_map': c_map, 'in_doc_words': in_doc_words }, {'track_list': track_list, 'args': vars(args) }, args.checkpoint + 'cwlm_lstm_crf') except Exception as inst: print(inst) else: patience_count += 1 print('(loss: %.4f, epoch: %d, dev acc = %.4f)' % (epoch_loss, args.start_epoch, dev_acc)) track_list.append({'loss': epoch_loss, 'dev_acc': dev_acc}) print('epoch: ' + str(args.start_epoch) + '\t in ' + str(args.epoch) + ' take: ' + str( time.time() - start_time) + ' s') if patience_count >= args.patience and args.start_epoch >= args.least_iters: break #print best if 'f' in args.eva_matrix: eprint(args.checkpoint + ' dev_f1: %.4f dev_rec: %.4f dev_pre: %.4f dev_acc: %.4f test_f1: %.4f test_rec: %.4f test_pre: %.4f test_acc: %.4f\n' % (dev_f1, dev_rec, dev_pre, dev_acc, test_f1, test_rec, test_pre, test_acc)) else: eprint(args.checkpoint + ' dev_acc: %.4f test_acc: %.4f\n' % (dev_acc, test_acc)) # printing summary print('setting:') print(args)
15,992
48.82243
297
py
LM-LSTM-CRF
LM-LSTM-CRF-master/seq_w.py
from __future__ import print_function import datetime import time import torch import torch.autograd as autograd import torch.nn as nn import torch.optim as optim import codecs from model.crf import * from model.lstm_crf import * import model.utils as utils from model.predictor import predict_w import argparse import json import os import sys from tqdm import tqdm import itertools import functools if __name__ == "__main__": parser = argparse.ArgumentParser(description='Evaluating LM-BLSTM-CRF') parser.add_argument('--load_arg', default='./checkpoint/ner/ner_4_cwlm_lstm_crf.json', help='path to arg json') parser.add_argument('--load_check_point', default='./checkpoint/ner/ner_4_cwlm_lstm_crf.model', help='path to model checkpoint file') parser.add_argument('--gpu',type=int, default=0, help='gpu id') parser.add_argument('--decode_type', choices=['label', 'string'], default='string', help='type of decode function, set `label` to couple label with text, or set `string` to insert label into test') parser.add_argument('--batch_size', type=int, default=50, help='size of batch') parser.add_argument('--input_file', default='data/ner2003/test.txt', help='path to input un-annotated corpus') parser.add_argument('--output_file', default='output.txt', help='path to output file') args = parser.parse_args() print('loading dictionary') with open(args.load_arg, 'r') as f: jd = json.load(f) jd = jd['args'] checkpoint_file = torch.load(args.load_check_point, map_location=lambda storage, loc: storage) f_map = checkpoint_file['f_map'] l_map = checkpoint_file['l_map'] if args.gpu >= 0: torch.cuda.set_device(args.gpu) # loading corpus print('loading corpus') with codecs.open(args.input_file, 'r', 'utf-8') as f: lines = f.readlines() # converting format features = utils.read_features(lines) # build model print('loading model') ner_model = LSTM_CRF(len(f_map), len(l_map), jd['embedding_dim'], jd['hidden'], jd['layers'], jd['drop_out'], large_CRF=jd['small_crf']) ner_model.load_state_dict(checkpoint_file['state_dict']) if args.gpu >= 0: if_cuda = True torch.cuda.set_device(args.gpu) ner_model.cuda() else: if_cuda = False decode_label = (args.decode_type == 'label') predictor = predict_w(if_cuda, f_map, l_map, f_map['<eof>'], l_map['<pad>'], l_map['<start>'], decode_label, args.batch_size, jd['caseless']) print('annotating') with open(args.output_file, 'w') as fout: predictor.output_batch(ner_model, features, fout)
2,630
36.056338
201
py
LM-LSTM-CRF
LM-LSTM-CRF-master/train_w.py
from __future__ import print_function import datetime import time import torch import torch.autograd as autograd import torch.nn as nn import torch.optim as optim import codecs from model.crf import * from model.lstm_crf import * import model.utils as utils from model.evaluator import eval_w import argparse import json import os import sys from tqdm import tqdm import itertools import functools def eprint(*args, **kwargs): print(*args, file=sys.stderr, **kwargs) if __name__ == "__main__": parser = argparse.ArgumentParser(description='Learning with BLSTM-CRF') parser.add_argument('--rand_embedding', action='store_true', help='random initialize word embedding') parser.add_argument('--emb_file', default='./embedding/glove.6B.100d.txt', help='path to pre-trained embedding') parser.add_argument('--train_file', default='./data/ner2003/eng.train.iobes', help='path to training file') parser.add_argument('--dev_file', default='./data/ner2003/eng.testa.iobes', help='path to development file') parser.add_argument('--test_file', default='./data/ner2003/eng.testb.iobes', help='path to test file') parser.add_argument('--gpu', type=int, default=0, help='gpu id, set to -1 if use cpu mode') parser.add_argument('--batch_size', type=int, default=10, help='batch size (10)') parser.add_argument('--unk', default='unk', help='unknow-token in pre-trained embedding') parser.add_argument('--checkpoint', default='./checkpoint/', help='path to checkpoint prefix') parser.add_argument('--hidden', type=int, default=100, help='hidden dimension') parser.add_argument('--drop_out', type=float, default=0.55, help='dropout ratio') parser.add_argument('--epoch', type=int, default=200, help='maximum epoch number') parser.add_argument('--start_epoch', type=int, default=0, help='start epoch idx') parser.add_argument('--caseless', action='store_true', help='caseless or not') parser.add_argument('--embedding_dim', type=int, default=100, help='dimension for word embedding') parser.add_argument('--layers', type=int, default=1, help='number of lstm layers') parser.add_argument('--lr', type=float, default=0.015, help='initial learning rate') parser.add_argument('--lr_decay', type=float, default=0.05, help='decay ratio of learning rate') parser.add_argument('--fine_tune', action='store_false', help='fine tune pre-trained embedding dictionary') parser.add_argument('--load_check_point', default='', help='path of checkpoint') parser.add_argument('--load_opt', action='store_true', help='load optimizer from ') parser.add_argument('--update', choices=['sgd', 'adam'], default='sgd', help='optimizer method') parser.add_argument('--momentum', type=float, default=0.9, help='momentum for sgd') parser.add_argument('--clip_grad', type=float, default=5.0, help='grad clip at') parser.add_argument('--small_crf', action='store_false', help='use small crf instead of large crf, refer model.crf module for more details') parser.add_argument('--mini_count', type=float, default=5, help='thresholds to replace rare words with <unk>') parser.add_argument('--eva_matrix', choices=['a', 'fa'], default='fa', help='use f1 and accuracy or accuracy alone') parser.add_argument('--patience', type=int, default=15, help='patience for early stop') parser.add_argument('--least_iters', type=int, default=50, help='at least train how many epochs before stop') parser.add_argument('--shrink_embedding', action='store_true', help='shrink the embedding dictionary to corpus (open this if pre-trained embedding dictionary is too large, but disable this may yield better results on external corpus)') args = parser.parse_args() if args.gpu >= 0: torch.cuda.set_device(args.gpu) print('setting:') print(args) # load corpus print('loading corpus') with codecs.open(args.train_file, 'r', 'utf-8') as f: lines = f.readlines() with codecs.open(args.dev_file, 'r', 'utf-8') as f: dev_lines = f.readlines() with codecs.open(args.test_file, 'r', 'utf-8') as f: test_lines = f.readlines() # converting format dev_features, dev_labels = utils.read_corpus(dev_lines) test_features, test_labels = utils.read_corpus(test_lines) if args.load_check_point: if os.path.isfile(args.load_check_point): print("loading checkpoint: '{}'".format(args.load_check_point)) checkpoint_file = torch.load(args.load_check_point) args.start_epoch = checkpoint_file['epoch'] f_map = checkpoint_file['f_map'] l_map = checkpoint_file['l_map'] train_features, train_labels = utils.read_corpus(lines) else: print("no checkpoint found at: '{}'".format(args.load_check_point)) else: print('constructing coding table') # converting format train_features, train_labels, f_map, l_map = utils.generate_corpus(lines, if_shrink_feature=True, thresholds=0) f_set = {v for v in f_map} f_map = utils.shrink_features(f_map, train_features, args.mini_count) dt_f_set = functools.reduce(lambda x, y: x | y, map(lambda t: set(t), dev_features), f_set) dt_f_set = functools.reduce(lambda x, y: x | y, map(lambda t: set(t), test_features), dt_f_set) if not args.rand_embedding: print("feature size: '{}'".format(len(f_map))) print('loading embedding') if args.fine_tune: # which means does not do fine-tune f_map = {'<eof>': 0} f_map, embedding_tensor, in_doc_words = utils.load_embedding_wlm(args.emb_file, ' ', f_map, dt_f_set,args.caseless,args.unk, args.embedding_dim, shrink_to_corpus=args.shrink_embedding) print("embedding size: '{}'".format(len(f_map))) l_set = functools.reduce(lambda x, y: x | y, map(lambda t: set(t), dev_labels)) l_set = functools.reduce(lambda x, y: x | y, map(lambda t: set(t), test_labels), l_set) for label in l_set: if label not in l_map: l_map[label] = len(l_map) # construct dataset dataset = utils.construct_bucket_mean_vb(train_features, train_labels, f_map, l_map, args.caseless) dev_dataset = utils.construct_bucket_mean_vb(dev_features, dev_labels, f_map, l_map, args.caseless) test_dataset = utils.construct_bucket_mean_vb(test_features, test_labels, f_map, l_map, args.caseless) dataset_loader = [torch.utils.data.DataLoader(tup, args.batch_size, shuffle=True, drop_last=False) for tup in dataset] dev_dataset_loader = [torch.utils.data.DataLoader(tup, 50, shuffle=False, drop_last=False) for tup in dev_dataset] test_dataset_loader = [torch.utils.data.DataLoader(tup, 50, shuffle=False, drop_last=False) for tup in test_dataset] # build model print('building model') ner_model = LSTM_CRF(len(f_map), len(l_map), args.embedding_dim, args.hidden, args.layers, args.drop_out, large_CRF=args.small_crf) if args.load_check_point: ner_model.load_state_dict(checkpoint_file['state_dict']) else: if not args.rand_embedding: ner_model.load_pretrained_embedding(embedding_tensor) print('random initialization') ner_model.rand_init(init_embedding=args.rand_embedding) if args.update == 'sgd': optimizer = optim.SGD(ner_model.parameters(), lr=args.lr, momentum=args.momentum) elif args.update == 'adam': optimizer = optim.Adam(ner_model.parameters(), lr=args.lr) if args.load_check_point and args.load_opt: optimizer.load_state_dict(checkpoint_file['optimizer']) crit = CRFLoss_vb(len(l_map), l_map['<start>'], l_map['<pad>']) if args.gpu >= 0: if_cuda = True print('device: ' + str(args.gpu)) torch.cuda.set_device(args.gpu) crit.cuda() ner_model.cuda() packer = CRFRepack(len(l_map), True) else: if_cuda = False packer = CRFRepack(len(l_map), False) if args.load_check_point: dev_f1, dev_acc = eval_batch(ner_model, dev_dataset_loader, pack, l_map) test_f1, test_acc = eval_batch(ner_model, test_dataset_loader, pack, l_map) print('(checkpoint: dev F1 = %.4f, dev acc = %.4f, F1 on test = %.4f, acc on test= %.4f)' % (dev_f1, dev_acc, test_f1, test_acc)) tot_length = sum(map(lambda t: len(t), dataset_loader)) best_f1 = float('-inf') best_acc = float('-inf') track_list = list() start_time = time.time() epoch_list = range(args.start_epoch, args.start_epoch + args.epoch) patience_count = 0 evaluator = eval_w(packer, l_map, args.eva_matrix) for epoch_idx, args.start_epoch in enumerate(epoch_list): epoch_loss = 0 ner_model.train() for feature, tg, mask in tqdm( itertools.chain.from_iterable(dataset_loader), mininterval=2, desc=' - Tot it %d (epoch %d)' % (tot_length, args.start_epoch), leave=False, file=sys.stdout): fea_v, tg_v, mask_v = packer.repack_vb(feature, tg, mask) ner_model.zero_grad() scores, hidden = ner_model.forward(fea_v) loss = crit.forward(scores, tg_v, mask_v) loss.backward() nn.utils.clip_grad_norm_(ner_model.parameters(), args.clip_grad) optimizer.step() epoch_loss += utils.to_scalar(loss) # update lr utils.adjust_learning_rate(optimizer, args.lr / (1 + (args.start_epoch + 1) * args.lr_decay)) # average epoch_loss /= tot_length # eval & save check_point if 'f' in args.eva_matrix: dev_result = evaluator.calc_score(ner_model, dev_dataset_loader) for label, (dev_f1, dev_pre, dev_rec, dev_acc, msg) in dev_result.items(): print('DEV : %s : dev_f1: %.4f dev_rec: %.4f dev_pre: %.4f dev_acc: %.4f | %s\n' % (label, dev_f1, dev_pre, dev_rec, dev_acc, msg)) (dev_f1, dev_pre, dev_rec, dev_acc, msg) = dev_result['total'] if dev_f1 > best_f1: patience_count = 0 best_f1 = dev_f1 test_result = evaluator.calc_score(ner_model, test_dataset_loader) for label, (test_f1, test_pre, test_rec, test_acc, msg) in test_result.items(): print('TEST : %s : test_f1: %.4f test_rec: %.4f test_pre: %.4f test_acc: %.4f | %s\n' % (label, test_f1, test_rec, test_pre, test_acc, msg)) (test_f1, test_rec, test_pre, test_acc, msg) = test_result['total'] track_list.append( {'loss': epoch_loss, 'dev_f1': dev_f1, 'dev_acc': dev_acc, 'test_f1': test_f1, 'test_acc': test_acc}) print( '(loss: %.4f, epoch: %d, dev F1 = %.4f, dev acc = %.4f, F1 on test = %.4f, acc on test= %.4f), saving...' % (epoch_loss, args.start_epoch, dev_f1, dev_acc, test_f1, test_acc)) try: utils.save_checkpoint({ 'epoch': args.start_epoch, 'state_dict': ner_model.state_dict(), 'optimizer': optimizer.state_dict(), 'f_map': f_map, 'l_map': l_map, }, {'track_list': track_list, 'args': vars(args) }, args.checkpoint + 'lstm_crf') except Exception as inst: print(inst) else: patience_count += 1 print('(loss: %.4f, epoch: %d, dev F1 = %.4f, dev acc = %.4f)' % (epoch_loss, args.start_epoch, dev_f1, dev_acc)) track_list.append({'loss': epoch_loss, 'dev_f1': dev_f1, 'dev_acc': dev_acc}) else: dev_acc = evaluator.calc_score(ner_model, dev_dataset_loader) if dev_acc > best_acc: patience_count = 0 best_acc = dev_acc test_acc = evaluator.calc_score(ner_model, test_dataset_loader) track_list.append( {'loss': epoch_loss, 'dev_acc': dev_acc, 'test_acc': test_acc}) print( '(loss: %.4f, epoch: %d, dev acc = %.4f, acc on test= %.4f), saving...' % (epoch_loss, args.start_epoch, dev_acc, test_acc)) try: utils.save_checkpoint({ 'epoch': args.start_epoch, 'state_dict': ner_model.state_dict(), 'optimizer': optimizer.state_dict(), 'f_map': f_map, 'l_map': l_map, }, {'track_list': track_list, 'args': vars(args) }, args.checkpoint + 'lstm_crf') except Exception as inst: print(inst) else: patience_count += 1 print('(loss: %.4f, epoch: %d, dev acc = %.4f)' % (epoch_loss, args.start_epoch, dev_acc)) track_list.append({'loss': epoch_loss, 'dev_acc': dev_acc}) print('epoch: ' + str(args.start_epoch) + '\t in ' + str(args.epoch) + ' take: ' + str( time.time() - start_time) + ' s') if patience_count >= args.patience and args.start_epoch >= args.least_iters: break #print best if 'f' in args.eva_matrix: eprint(args.checkpoint + ' dev_f1: %.4f dev_rec: %.4f dev_pre: %.4f dev_acc: %.4f test_f1: %.4f test_rec: %.4f test_pre: %.4f test_acc: %.4f\n' % (dev_f1, dev_rec, dev_pre, dev_acc, test_f1, test_rec, test_pre, test_acc)) else: eprint(args.checkpoint + ' dev_acc: %.4f test_acc: %.4f\n' % (dev_acc, test_acc)) # printing summary print('setting:') print(args)
14,299
45.278317
239
py
LM-LSTM-CRF
LM-LSTM-CRF-master/seq_wc.py
from __future__ import print_function import datetime import time import torch import torch.autograd as autograd import torch.nn as nn import torch.optim as optim import codecs from model.crf import * from model.lm_lstm_crf import * import model.utils as utils from model.predictor import predict_wc import argparse import json import os import sys from tqdm import tqdm import itertools import functools if __name__ == "__main__": parser = argparse.ArgumentParser(description='Evaluating LM-BLSTM-CRF') parser.add_argument('--load_arg', default='./checkpoint/ner/ner_4_cwlm_lstm_crf.json', help='path to arg json') parser.add_argument('--load_check_point', default='./checkpoint/ner/ner_4_cwlm_lstm_crf.model', help='path to model checkpoint file') parser.add_argument('--gpu',type=int, default=0, help='gpu id') parser.add_argument('--decode_type', choices=['label', 'string'], default='string', help='type of decode function, set `label` to couple label with text, or set `string` to insert label into test') parser.add_argument('--batch_size', type=int, default=50, help='size of batch') parser.add_argument('--input_file', default='data/ner2003/test.txt', help='path to input un-annotated corpus') parser.add_argument('--output_file', default='output.txt', help='path to output file') args = parser.parse_args() print('loading dictionary') with open(args.load_arg, 'r') as f: jd = json.load(f) jd = jd['args'] checkpoint_file = torch.load(args.load_check_point, map_location=lambda storage, loc: storage) f_map = checkpoint_file['f_map'] l_map = checkpoint_file['l_map'] c_map = checkpoint_file['c_map'] in_doc_words = checkpoint_file['in_doc_words'] if args.gpu >= 0: torch.cuda.set_device(args.gpu) # loading corpus print('loading corpus') with codecs.open(args.input_file, 'r', 'utf-8') as f: lines = f.readlines() # converting format features = utils.read_features(lines) # build model print('loading model') ner_model = LM_LSTM_CRF(len(l_map), len(c_map), jd['char_dim'], jd['char_hidden'], jd['char_layers'], jd['word_dim'], jd['word_hidden'], jd['word_layers'], len(f_map), jd['drop_out'], large_CRF=jd['small_crf'], if_highway=jd['high_way'], in_doc_words=in_doc_words, highway_layers = jd['highway_layers']) ner_model.load_state_dict(checkpoint_file['state_dict']) if args.gpu >= 0: if_cuda = True torch.cuda.set_device(args.gpu) ner_model.cuda() else: if_cuda = False decode_label = (args.decode_type == 'label') predictor = predict_wc(if_cuda, f_map, c_map, l_map, f_map['<eof>'], c_map['\n'], l_map['<pad>'], l_map['<start>'], decode_label, args.batch_size, jd['caseless']) print('annotating') with open(args.output_file, 'w') as fout: predictor.output_batch(ner_model, features, fout)
2,909
39.416667
307
py
LM-LSTM-CRF
LM-LSTM-CRF-master/eval_w.py
from __future__ import print_function import datetime import time import torch import torch.autograd as autograd import torch.nn as nn import torch.optim as optim import codecs from model.crf import * from model.lstm_crf import * import model.utils as utils from model.evaluator import eval_w import argparse import json import os import sys from tqdm import tqdm import itertools import functools if __name__ == "__main__": parser = argparse.ArgumentParser(description='Evaluating BLSTM-CRF') parser.add_argument('--load_arg', default='./checkpoint/soa/check_wc_p_char_lstm_crf.json', help='arg json file path') parser.add_argument('--load_check_point', default='./checkpoint/soa/check_wc_p_char_lstm_crf.model', help='checkpoint path') parser.add_argument('--gpu',type=int, default=0, help='gpu id') parser.add_argument('--eva_matrix', choices=['a', 'fa'], default='fa', help='use f1 and accuracy or accuracy alone') parser.add_argument('--test_file', default='', help='path to test file, if set to none, would use test_file path in the checkpoint file') args = parser.parse_args() with open(args.load_arg, 'r') as f: jd = json.load(f) jd = jd['args'] checkpoint_file = torch.load(args.load_check_point, map_location=lambda storage, loc: storage) f_map = checkpoint_file['f_map'] l_map = checkpoint_file['l_map'] if args.gpu >= 0: torch.cuda.set_device(args.gpu) # load corpus if args.test_file: with codecs.open(args.test_file, 'r', 'utf-8') as f: test_lines = f.readlines() else: with codecs.open(jd['test_file'], 'r', 'utf-8') as f: test_lines = f.readlines() # converting format test_features, test_labels = utils.read_corpus(test_lines) # construct dataset test_dataset = utils.construct_bucket_mean_vb(test_features, test_labels, f_map, l_map, jd['caseless']) test_dataset_loader = [torch.utils.data.DataLoader(tup, 50, shuffle=False, drop_last=False) for tup in test_dataset] # build model ner_model = LSTM_CRF(len(f_map), len(l_map), jd['embedding_dim'], jd['hidden'], jd['layers'], jd['drop_out'], large_CRF=jd['small_crf']) ner_model.load_state_dict(checkpoint_file['state_dict']) if args.gpu >= 0: if_cuda = True torch.cuda.set_device(args.gpu) ner_model.cuda() packer = CRFRepack(len(l_map), True) else: if_cuda = False packer = CRFRepack(len(l_map), False) evaluator = eval_w(packer, l_map, args.eva_matrix) if 'f' in args.eva_matrix: test_f1, test_pre, test_rec, test_acc = evaluator.calc_score(ner_model, test_dataset_loader) print(jd['checkpoint'] + ' test_f1: %.4f test_rec: %.4f test_pre: %.4f test_acc: %.4f\n' % (test_f1, test_rec, test_pre, test_acc)) else: test_acc = evaluator.calc_score(ner_model, test_dataset_loader) print(jd['checkpoint'] + ' test_acc: %.4f\n' % (test_acc))
2,982
32.897727
141
py
LM-LSTM-CRF
LM-LSTM-CRF-master/docs/source/conf.py
#!/usr/bin/env python3 # -*- coding: utf-8 -*- # # LM-LSTM-CRF documentation build configuration file, created by # sphinx-quickstart on Thu Sep 14 03:49:01 2017. # # This file is execfile()d with the current directory set to its # containing dir. # # Note that not all possible configuration values are present in this # autogenerated file. # # All configuration values have a default; values that are commented out # serve to show the default. # If extensions (or modules to document with autodoc) are in another directory, # add these directories to sys.path here. If the directory is relative to the # documentation root, use os.path.abspath to make it absolute, like shown here. import os import sys sys.path.insert(0, os.path.abspath('../..')) # -- General configuration ------------------------------------------------ # If your documentation needs a minimal Sphinx version, state it here. # # needs_sphinx = '1.0' # Add any Sphinx extension module names here, as strings. They can be # extensions coming with Sphinx (named 'sphinx.ext.*') or your custom # ones. extensions = [ 'sphinx.ext.autodoc', 'sphinx.ext.autosummary', 'sphinx.ext.doctest', 'sphinx.ext.intersphinx', 'sphinx.ext.todo', 'sphinx.ext.coverage', 'sphinx.ext.mathjax', 'sphinx.ext.napoleon', 'sphinx.ext.viewcode', ] napoleon_use_ivar = True # Add any paths that contain templates here, relative to this directory. templates_path = ['_templates'] # The suffix(es) of source filenames. # You can specify multiple suffix as a list of string: # # source_suffix = ['.rst', '.md'] source_suffix = '.rst' # The master toctree document. master_doc = 'index' # General information about the project. project = 'LM-LSTM-CRF' copyright = '2017, Liyuan Liu, Frank Xu, Jingbo Shang' author = 'Liyuan Liu, Frank Xu, Jingbo Shang' # The version info for the project you're documenting, acts as replacement for # |version| and |release|, also used in various other places throughout the # built documents. # # The short X.Y version. version = '' # The full version, including alpha/beta/rc tags. release = '' # The language for content autogenerated by Sphinx. Refer to documentation # for a list of supported languages. # # This is also used if you do content translation via gettext catalogs. # Usually you set "language" from the command line for these cases. language = None # List of patterns, relative to source directory, that match files and # directories to ignore when looking for source files. # This patterns also effect to html_static_path and html_extra_path exclude_patterns = [] # The name of the Pygments (syntax highlighting) style to use. pygments_style = 'sphinx' # If true, `todo` and `todoList` produce output, else they produce nothing. todo_include_todos = False # -- Options for HTML output ---------------------------------------------- # The theme to use for HTML and HTML Help pages. See the documentation for # a list of builtin themes. # html_theme = 'sphinx_rtd_theme' # Theme options are theme-specific and customize the look and feel of a theme # further. For a list of options available for each theme, see the # documentation. # # html_theme_options = {} html_theme_options = { 'collapse_navigation': False, 'display_version': True, } # Add any paths that contain custom static files (such as style sheets) here, # relative to this directory. They are copied after the builtin static files, # so a file named "default.css" will overwrite the builtin "default.css". html_static_path = ['_static'] # Custom sidebar templates, must be a dictionary that maps document names # to template names. # # This is required for the alabaster theme # refs: http://alabaster.readthedocs.io/en/latest/installation.html#sidebars html_sidebars = { '**': [ 'about.html', 'navigation.html', 'relations.html', # needs 'show_related': True theme option to display 'searchbox.html', 'donate.html', ] } # -- Options for HTMLHelp output ------------------------------------------ # Output file base name for HTML help builder. htmlhelp_basename = 'LM-LSTM-CRFdoc' # -- Options for LaTeX output --------------------------------------------- latex_elements = { # The paper size ('letterpaper' or 'a4paper'). # # 'papersize': 'letterpaper', # The font size ('10pt', '11pt' or '12pt'). # # 'pointsize': '10pt', # Additional stuff for the LaTeX preamble. # # 'preamble': '', # Latex figure (float) alignment # # 'figure_align': 'htbp', } # Grouping the document tree into LaTeX files. List of tuples # (source start file, target name, title, # author, documentclass [howto, manual, or own class]). latex_documents = [ (master_doc, 'LM-LSTM-CRF.tex', 'LM-LSTM-CRF Documentation', 'Liyuan Liu, Frank Xu, Jingbo Shang', 'manual'), ] # -- Options for manual page output --------------------------------------- # One entry per manual page. List of tuples # (source start file, name, description, authors, manual section). man_pages = [ (master_doc, 'lm-lstm-crf', 'LM-LSTM-CRF Documentation', [author], 1) ] # -- Options for Texinfo output ------------------------------------------- # Grouping the document tree into Texinfo files. List of tuples # (source start file, target name, title, author, # dir menu entry, description, category) texinfo_documents = [ (master_doc, 'LM-LSTM-CRF', 'LM-LSTM-CRF Documentation', author, 'LM-LSTM-CRF', 'One line description of project.', 'Miscellaneous'), ]
5,569
29.773481
79
py
LM-LSTM-CRF
LM-LSTM-CRF-master/model/highway.py
""" .. module:: highway :synopsis: highway network .. moduleauthor:: Liyuan Liu """ import torch import torch.nn as nn import model.utils as utils class hw(nn.Module): """Highway layers args: size: input and output dimension dropout_ratio: dropout ratio """ def __init__(self, size, num_layers = 1, dropout_ratio = 0.5): super(hw, self).__init__() self.size = size self.num_layers = num_layers self.trans = nn.ModuleList() self.gate = nn.ModuleList() self.dropout = nn.Dropout(p=dropout_ratio) for i in range(num_layers): tmptrans = nn.Linear(size, size) tmpgate = nn.Linear(size, size) self.trans.append(tmptrans) self.gate.append(tmpgate) def rand_init(self): """ random initialization """ for i in range(self.num_layers): utils.init_linear(self.trans[i]) utils.init_linear(self.gate[i]) def forward(self, x): """ update statics for f1 score args: x (ins_num, hidden_dim): input tensor return: output tensor (ins_num, hidden_dim) """ g = nn.functional.sigmoid(self.gate[0](x)) h = nn.functional.relu(self.trans[0](x)) x = g * h + (1 - g) * x for i in range(1, self.num_layers): x = self.dropout(x) g = nn.functional.sigmoid(self.gate[i](x)) h = nn.functional.relu(self.trans[i](x)) x = g * h + (1 - g) * x return x
1,607
24.52381
66
py
LM-LSTM-CRF
LM-LSTM-CRF-master/model/lm_lstm_crf.py
""" .. module:: lm_lstm_crf :synopsis: lm_lstm_crf .. moduleauthor:: Liyuan Liu """ import torch import torch.autograd as autograd import torch.nn as nn import torch.optim as optim import numpy as np import model.crf as crf import model.utils as utils import model.highway as highway class LM_LSTM_CRF(nn.Module): """LM_LSTM_CRF model args: tagset_size: size of label set char_size: size of char dictionary char_dim: size of char embedding char_hidden_dim: size of char-level lstm hidden dim char_rnn_layers: number of char-level lstm layers embedding_dim: size of word embedding word_hidden_dim: size of word-level blstm hidden dim word_rnn_layers: number of word-level lstm layers vocab_size: size of word dictionary dropout_ratio: dropout ratio large_CRF: use CRF_L or not, refer model.crf.CRF_L and model.crf.CRF_S for more details if_highway: use highway layers or not in_doc_words: number of words that occurred in the corpus (used for language model prediction) highway_layers: number of highway layers """ def __init__(self, tagset_size, char_size, char_dim, char_hidden_dim, char_rnn_layers, embedding_dim, word_hidden_dim, word_rnn_layers, vocab_size, dropout_ratio, large_CRF=True, if_highway = False, in_doc_words = 2, highway_layers = 1): super(LM_LSTM_CRF, self).__init__() self.char_dim = char_dim self.char_hidden_dim = char_hidden_dim self.char_size = char_size self.word_dim = embedding_dim self.word_hidden_dim = word_hidden_dim self.word_size = vocab_size self.if_highway = if_highway self.char_embeds = nn.Embedding(char_size, char_dim) self.forw_char_lstm = nn.LSTM(char_dim, char_hidden_dim, num_layers=char_rnn_layers, bidirectional=False, dropout=dropout_ratio) self.back_char_lstm = nn.LSTM(char_dim, char_hidden_dim, num_layers=char_rnn_layers, bidirectional=False, dropout=dropout_ratio) self.char_rnn_layers = char_rnn_layers self.word_embeds = nn.Embedding(vocab_size, embedding_dim) self.word_lstm = nn.LSTM(embedding_dim + char_hidden_dim * 2, word_hidden_dim // 2, num_layers=word_rnn_layers, bidirectional=True, dropout=dropout_ratio) self.word_rnn_layers = word_rnn_layers self.dropout = nn.Dropout(p=dropout_ratio) self.tagset_size = tagset_size if large_CRF: self.crf = crf.CRF_L(word_hidden_dim, tagset_size) else: self.crf = crf.CRF_S(word_hidden_dim, tagset_size) if if_highway: self.forw2char = highway.hw(char_hidden_dim, num_layers=highway_layers, dropout_ratio=dropout_ratio) self.back2char = highway.hw(char_hidden_dim, num_layers=highway_layers, dropout_ratio=dropout_ratio) self.forw2word = highway.hw(char_hidden_dim, num_layers=highway_layers, dropout_ratio=dropout_ratio) self.back2word = highway.hw(char_hidden_dim, num_layers=highway_layers, dropout_ratio=dropout_ratio) self.fb2char = highway.hw(2 * char_hidden_dim, num_layers=highway_layers, dropout_ratio=dropout_ratio) self.char_pre_train_out = nn.Linear(char_hidden_dim, char_size) self.word_pre_train_out = nn.Linear(char_hidden_dim, in_doc_words) self.batch_size = 1 self.word_seq_length = 1 def set_batch_size(self, bsize): """ set batch size """ self.batch_size = bsize def set_batch_seq_size(self, sentence): """ set batch size and sequence length """ tmp = sentence.size() self.word_seq_length = tmp[0] self.batch_size = tmp[1] def rand_init_embedding(self): """ random initialize char-level embedding """ utils.init_embedding(self.char_embeds.weight) def load_pretrained_word_embedding(self, pre_word_embeddings): """ load pre-trained word embedding args: pre_word_embeddings (self.word_size, self.word_dim) : pre-trained embedding """ assert (pre_word_embeddings.size()[1] == self.word_dim) self.word_embeds.weight = nn.Parameter(pre_word_embeddings) def rand_init(self, init_char_embedding=True, init_word_embedding=False): """ random initialization args: init_char_embedding: random initialize char embedding or not init_word_embedding: random initialize word embedding or not """ if init_char_embedding: utils.init_embedding(self.char_embeds.weight) if init_word_embedding: utils.init_embedding(self.word_embeds.weight) if self.if_highway: self.forw2char.rand_init() self.back2char.rand_init() self.forw2word.rand_init() self.back2word.rand_init() self.fb2char.rand_init() utils.init_lstm(self.forw_char_lstm) utils.init_lstm(self.back_char_lstm) utils.init_lstm(self.word_lstm) utils.init_linear(self.char_pre_train_out) utils.init_linear(self.word_pre_train_out) self.crf.rand_init() def word_pre_train_forward(self, sentence, position, hidden=None): """ output of forward language model args: sentence (char_seq_len, batch_size): char-level representation of sentence position (word_seq_len, batch_size): position of blank space in char-level representation of sentence hidden: initial hidden state return: language model output (word_seq_len, in_doc_word), hidden """ embeds = self.char_embeds(sentence) d_embeds = self.dropout(embeds) lstm_out, hidden = self.forw_char_lstm(d_embeds) tmpsize = position.size() position = position.unsqueeze(2).expand(tmpsize[0], tmpsize[1], self.char_hidden_dim) select_lstm_out = torch.gather(lstm_out, 0, position) d_lstm_out = self.dropout(select_lstm_out).view(-1, self.char_hidden_dim) if self.if_highway: char_out = self.forw2word(d_lstm_out) d_char_out = self.dropout(char_out) else: d_char_out = d_lstm_out pre_score = self.word_pre_train_out(d_char_out) return pre_score, hidden def word_pre_train_backward(self, sentence, position, hidden=None): """ output of backward language model args: sentence (char_seq_len, batch_size): char-level representation of sentence (inverse order) position (word_seq_len, batch_size): position of blank space in inversed char-level representation of sentence hidden: initial hidden state return: language model output (word_seq_len, in_doc_word), hidden """ embeds = self.char_embeds(sentence) d_embeds = self.dropout(embeds) lstm_out, hidden = self.back_char_lstm(d_embeds) tmpsize = position.size() position = position.unsqueeze(2).expand(tmpsize[0], tmpsize[1], self.char_hidden_dim) select_lstm_out = torch.gather(lstm_out, 0, position) d_lstm_out = self.dropout(select_lstm_out).view(-1, self.char_hidden_dim) if self.if_highway: char_out = self.back2word(d_lstm_out) d_char_out = self.dropout(char_out) else: d_char_out = d_lstm_out pre_score = self.word_pre_train_out(d_char_out) return pre_score, hidden def forward(self, forw_sentence, forw_position, back_sentence, back_position, word_seq, hidden=None): ''' args: forw_sentence (char_seq_len, batch_size) : char-level representation of sentence forw_position (word_seq_len, batch_size) : position of blank space in char-level representation of sentence back_sentence (char_seq_len, batch_size) : char-level representation of sentence (inverse order) back_position (word_seq_len, batch_size) : position of blank space in inversed char-level representation of sentence word_seq (word_seq_len, batch_size) : word-level representation of sentence hidden: initial hidden state return: crf output (word_seq_len, batch_size, tag_size, tag_size), hidden ''' self.set_batch_seq_size(forw_position) #embedding layer forw_emb = self.char_embeds(forw_sentence) back_emb = self.char_embeds(back_sentence) #dropout d_f_emb = self.dropout(forw_emb) d_b_emb = self.dropout(back_emb) #forward the whole sequence forw_lstm_out, _ = self.forw_char_lstm(d_f_emb)#seq_len_char * batch * char_hidden_dim back_lstm_out, _ = self.back_char_lstm(d_b_emb)#seq_len_char * batch * char_hidden_dim #select predict point forw_position = forw_position.unsqueeze(2).expand(self.word_seq_length, self.batch_size, self.char_hidden_dim) select_forw_lstm_out = torch.gather(forw_lstm_out, 0, forw_position) back_position = back_position.unsqueeze(2).expand(self.word_seq_length, self.batch_size, self.char_hidden_dim) select_back_lstm_out = torch.gather(back_lstm_out, 0, back_position) fb_lstm_out = self.dropout(torch.cat((select_forw_lstm_out, select_back_lstm_out), dim=2)) if self.if_highway: char_out = self.fb2char(fb_lstm_out) d_char_out = self.dropout(char_out) else: d_char_out = fb_lstm_out #word word_emb = self.word_embeds(word_seq) d_word_emb = self.dropout(word_emb) #combine word_input = torch.cat((d_word_emb, d_char_out), dim = 2) #word level lstm lstm_out, _ = self.word_lstm(word_input) d_lstm_out = self.dropout(lstm_out) #convert to crf crf_out = self.crf(d_lstm_out) crf_out = crf_out.view(self.word_seq_length, self.batch_size, self.tagset_size, self.tagset_size) return crf_out
10,117
38.678431
241
py
LM-LSTM-CRF
LM-LSTM-CRF-master/model/predictor.py
""" .. module:: predictor :synopsis: prediction method (for un-annotated text) .. moduleauthor:: Liyuan Liu """ import torch import torch.autograd as autograd import numpy as np import itertools import sys from tqdm import tqdm from model.crf import CRFDecode_vb from model.utils import * class predict: """Base class for prediction, provide method to calculate f1 score and accuracy args: if_cuda: if use cuda to speed up l_map: dictionary for labels label_seq: type of decode function, set `True` to couple label with text, or set 'False' to insert label into test batch_size: size of batch in decoding """ def __init__(self, if_cuda, l_map, label_seq = True, batch_size = 50): self.if_cuda = if_cuda self.l_map = l_map self.r_l_map = revlut(l_map) self.batch_size = batch_size if label_seq: self.decode_str = self.decode_l else: self.decode_str = self.decode_s def decode_l(self, feature, label): """ decode a sentence coupled with label args: feature (list): words list label (list): label list """ return '\n'.join(map(lambda t: t[0] + ' '+ self.r_l_map[t[1].item()], zip(feature, label))) def decode_s(self, feature, label): """ decode a sentence in the format of <> args: feature (list): words list label (list): label list """ chunks = "" current = None for f, y in zip(feature, label): label = self.r_l_map[y.item()] if label.startswith('B-'): if current is not None: chunks += "</"+current+"> " current = label[2:] chunks += "<"+current+"> " + f + " " elif label.startswith('S-'): if current is not None: chunks += " </"+current+"> " current = label[2:] chunks += "<"+current+"> " + f + " </"+current+"> " current = None elif label.startswith('I-'): if current is not None: base = label[2:] if base == current: chunks += f+" " else: chunks += "</"+current+"> <"+base+"> " + f + " " current = base else: current = label[2:] chunks += "<"+current+"> " + f + " " elif label.startswith('E-'): if current is not None: base = label[2:] if base == current: chunks += f + " </"+base+"> " current = None else: chunks += "</"+current+"> <"+base+"> " + f + " </"+base+"> " current = None else: current = label[2:] chunks += "<"+current+"> " + f + " </"+current+"> " current = None else: if current is not None: chunks += "</"+current+"> " chunks += f+" " current = None if current is not None: chunks += "</"+current+"> " return chunks def output_batch(self, ner_model, documents, fout): """ decode the whole corpus in the specific format by calling apply_model to fit specific models args: ner_model: sequence labeling model feature (list): list of words list fout: output file """ ner_model.eval() d_len = len(documents) for d_ind in tqdm( range(0, d_len), mininterval=1, desc=' - Process', leave=False, file=sys.stdout): fout.write('-DOCSTART- -DOCSTART- -DOCSTART-\n\n') features = documents[d_ind] f_len = len(features) for ind in range(0, f_len, self.batch_size): eind = min(f_len, ind + self.batch_size) labels = self.apply_model(ner_model, features[ind: eind]) labels = torch.unbind(labels, 1) for ind2 in range(ind, eind): f = features[ind2] l = labels[ind2 - ind][0: len(f) ] fout.write(self.decode_str(features[ind2], l) + '\n\n') def apply_model(self, ner_model, features): """ template function for apply_model args: ner_model: sequence labeling model feature (list): list of words list """ return None class predict_w(predict): """prediction class for word level model (LSTM-CRF) args: if_cuda: if use cuda to speed up f_map: dictionary for words l_map: dictionary for labels pad_word: word padding pad_label: label padding start_label: start label label_seq: type of decode function, set `True` to couple label with text, or set 'False' to insert label into test batch_size: size of batch in decoding caseless: caseless or not """ def __init__(self, if_cuda, f_map, l_map, pad_word, pad_label, start_label, label_seq = True, batch_size = 50, caseless=True): predict.__init__(self, if_cuda, l_map, label_seq, batch_size) self.decoder = CRFDecode_vb(len(l_map), start_label, pad_label) self.pad_word = pad_word self.f_map = f_map self.l_map = l_map self.caseless = caseless def apply_model(self, ner_model, features): """ apply_model function for LSTM-CRF args: ner_model: sequence labeling model feature (list): list of words list """ if self.caseless: features = list(map(lambda t: list(map(lambda x: x.lower(), t)), features)) features = encode_safe(features, self.f_map, self.f_map['<unk>']) f_len = max(map(lambda t: len(t) + 1, features)) masks = torch.ByteTensor(list(map(lambda t: [1] * (len(t) + 1) + [0] * (f_len - len(t) - 1), features))) word_features = torch.LongTensor(list(map(lambda t: t + [self.pad_word] * (f_len - len(t)), features))) if self.if_cuda: fea_v = autograd.Variable(word_features.transpose(0, 1)).cuda() mask_v = masks.transpose(0, 1).cuda() else: fea_v = autograd.Variable(word_features.transpose(0, 1)) mask_v = masks.transpose(0, 1).contiguous() scores, _ = ner_model(fea_v) decoded = self.decoder.decode(scores.data, mask_v) return decoded class predict_wc(predict): """prediction class for LM-LSTM-CRF args: if_cuda: if use cuda to speed up f_map: dictionary for words c_map: dictionary for chars l_map: dictionary for labels pad_word: word padding pad_char: word padding pad_label: label padding start_label: start label label_seq: type of decode function, set `True` to couple label with text, or set 'False' to insert label into test batch_size: size of batch in decoding caseless: caseless or not """ def __init__(self, if_cuda, f_map, c_map, l_map, pad_word, pad_char, pad_label, start_label, label_seq = True, batch_size = 50, caseless=True): predict.__init__(self, if_cuda, l_map, label_seq, batch_size) self.decoder = CRFDecode_vb(len(l_map), start_label, pad_label) self.pad_word = pad_word self.pad_char = pad_char self.f_map = f_map self.c_map = c_map self.l_map = l_map self.caseless = caseless def apply_model(self, ner_model, features): """ apply_model function for LM-LSTM-CRF args: ner_model: sequence labeling model feature (list): list of words list """ char_features = encode2char_safe(features, self.c_map) if self.caseless: word_features = encode_safe(list(map(lambda t: list(map(lambda x: x.lower(), t)), features)), self.f_map, self.f_map['<unk>']) else: word_features = encode_safe(features, self.f_map, self.f_map['<unk>']) fea_len = [list( map( lambda t: len(t) + 1, f) ) for f in char_features] forw_features = concatChar(char_features, self.c_map) word_len = max(map(lambda t: len(t) + 1, word_features)) char_len = max(map(lambda t: len(t[0]) + word_len - len(t[1]), zip(forw_features, word_features))) forw_t = list( map( lambda t: t + [self.pad_char] * ( char_len - len(t) ), forw_features ) ) back_t = torch.LongTensor( list( map( lambda t: t[::-1], forw_t ) ) ) forw_t = torch.LongTensor( forw_t ) forw_p = torch.LongTensor( list( map( lambda t: list(itertools.accumulate( t + [1] * (word_len - len(t) ) ) ), fea_len) ) ) back_p = torch.LongTensor( list( map( lambda t: [char_len - 1] + [ char_len - 1 - tup for tup in t[:-1] ], forw_p) ) ) masks = torch.ByteTensor(list(map(lambda t: [1] * (len(t) + 1) + [0] * (word_len - len(t) - 1), word_features))) word_t = torch.LongTensor(list(map(lambda t: t + [self.pad_word] * (word_len - len(t)), word_features))) if self.if_cuda: f_f = autograd.Variable(forw_t.transpose(0, 1)).cuda() f_p = autograd.Variable(forw_p.transpose(0, 1)).cuda() b_f = autograd.Variable(back_t.transpose(0, 1)).cuda() b_p = autograd.Variable(back_p.transpose(0, 1)).cuda() w_f = autograd.Variable(word_t.transpose(0, 1)).cuda() mask_v = masks.transpose(0, 1).cuda() else: f_f = autograd.Variable(forw_t.transpose(0, 1)) f_p = autograd.Variable(forw_p.transpose(0, 1)) b_f = autograd.Variable(back_t.transpose(0, 1)) b_p = autograd.Variable(back_p.transpose(0, 1)) w_f = autograd.Variable(word_t.transpose(0, 1)) mask_v = masks.transpose(0, 1) scores = ner_model(f_f, f_p, b_f, b_p, w_f) decoded = self.decoder.decode(scores.data, mask_v) return decoded
10,324
35.875
147
py
LM-LSTM-CRF
LM-LSTM-CRF-master/model/ner_dataset.py
""" .. module:: datasets :synopsis: datasets .. moduleauthor:: Liyuan Liu """ from torch.utils.data import Dataset class CRFDataset(Dataset): """Dataset Class for word-level model args: data_tensor (ins_num, seq_length): words label_tensor (ins_num, seq_length): labels mask_tensor (ins_num, seq_length): padding masks """ def __init__(self, data_tensor, label_tensor, mask_tensor): assert data_tensor.size(0) == label_tensor.size(0) assert data_tensor.size(0) == mask_tensor.size(0) self.data_tensor = data_tensor self.label_tensor = label_tensor self.mask_tensor = mask_tensor def __getitem__(self, index): return self.data_tensor[index], self.label_tensor[index], self.mask_tensor[index] def __len__(self): return self.data_tensor.size(0) class CRFDataset_WC(Dataset): """Dataset Class for char-aware model args: forw_tensor (ins_num, seq_length): forward chars forw_index (ins_num, seq_length): index of forward chars back_tensor (ins_num, seq_length): backward chars back_index (ins_num, seq_length): index of backward chars word_tensor (ins_num, seq_length): words label_tensor (ins_num, seq_length): labels: mask_tensor (ins_num, seq_length): padding masks len_tensor (ins_num, 2): length of chars (dim0) and words (dim1) """ def __init__(self, forw_tensor, forw_index, back_tensor, back_index, word_tensor, label_tensor, mask_tensor, len_tensor): assert forw_tensor.size(0) == label_tensor.size(0) assert forw_tensor.size(0) == mask_tensor.size(0) assert forw_tensor.size(0) == forw_index.size(0) assert forw_tensor.size(0) == back_tensor.size(0) assert forw_tensor.size(0) == back_index.size(0) assert forw_tensor.size(0) == word_tensor.size(0) assert forw_tensor.size(0) == len_tensor.size(0) self.forw_tensor = forw_tensor self.forw_index = forw_index self.back_tensor = back_tensor self.back_index = back_index self.word_tensor = word_tensor self.label_tensor = label_tensor self.mask_tensor = mask_tensor self.len_tensor = len_tensor def __getitem__(self, index): return self.forw_tensor[index], self.forw_index[index], self.back_tensor[index], self.back_index[index], self.word_tensor[index], self.label_tensor[index], self.mask_tensor[index], self.len_tensor[index] def __len__(self): return self.forw_tensor.size(0)
2,581
37.537313
211
py
LM-LSTM-CRF
LM-LSTM-CRF-master/model/utils.py
""" .. module:: utils :synopsis: utility tools .. moduleauthor:: Liyuan Liu, Frank Xu """ import codecs import csv import itertools from functools import reduce import numpy as np import shutil import torch import json import torch.nn as nn import torch.nn.init from model.ner_dataset import * zip = getattr(itertools, 'izip', zip) def to_scalar(var): """change the first element of a tensor to scalar """ return var.view(-1).data.tolist()[0] def argmax(vec): """helper function to calculate argmax of input vector at dimension 1 """ _, idx = torch.max(vec, 1) return to_scalar(idx) def log_sum_exp(vec, m_size): """ calculate log of exp sum args: vec (batch_size, vanishing_dim, hidden_dim) : input tensor m_size : hidden_dim return: batch_size, hidden_dim """ _, idx = torch.max(vec, 1) # B * 1 * M max_score = torch.gather(vec, 1, idx.view(-1, 1, m_size)).view(-1, 1, m_size) # B * M return max_score.view(-1, m_size) + torch.log(torch.sum(torch.exp(vec - max_score.expand_as(vec)), 1)).view(-1, m_size) # B * M def switch(vec1, vec2, mask): """ switch function for pytorch args: vec1 (any size) : input tensor corresponding to 0 vec2 (same to vec1) : input tensor corresponding to 1 mask (same to vec1) : input tensor, each element equals to 0/1 return: vec (*) """ catvec = torch.cat([vec1.view(-1, 1), vec2.view(-1, 1)], dim=1) switched_vec = torch.gather(catvec, 1, mask.long().view(-1, 1)) return switched_vec.view(-1) def encode2char_safe(input_lines, char_dict): """ get char representation of lines args: input_lines (list of strings) : input corpus char_dict (dictionary) : char-level dictionary return: forw_lines """ unk = char_dict['<u>'] forw_lines = [list(map(lambda m: list(map(lambda t: char_dict.get(t, unk), m)), line)) for line in input_lines] return forw_lines def concatChar(input_lines, char_dict): """ concat char into string args: input_lines (list of list of char) : input corpus char_dict (dictionary) : char-level dictionary return: forw_lines """ features = [[char_dict[' ']] + list(reduce(lambda x, y: x + [char_dict[' ']] + y, sentence)) + [char_dict['\n']] for sentence in input_lines] return features def encode_safe(input_lines, word_dict, unk): """ encode list of strings into word-level representation with unk """ lines = list(map(lambda t: list(map(lambda m: word_dict.get(m, unk), t)), input_lines)) return lines def encode(input_lines, word_dict): """ encode list of strings into word-level representation """ lines = list(map(lambda t: list(map(lambda m: word_dict[m], t)), input_lines)) return lines def encode2Tensor(input_lines, word_dict, unk): """ encode list of strings into word-level representation (tensor) with unk """ lines = list(map(lambda t: torch.LongTensor(list(map(lambda m: word_dict.get(m, unk), t))), input_lines)) return lines def generate_corpus_char(lines, if_shrink_c_feature=False, c_thresholds=1, if_shrink_w_feature=False, w_thresholds=1): """ generate label, feature, word dictionary, char dictionary and label dictionary args: lines : corpus if_shrink_c_feature: whether shrink char-dictionary c_threshold: threshold for shrinking char-dictionary if_shrink_w_feature: whether shrink word-dictionary w_threshold: threshold for shrinking word-dictionary """ features, labels, feature_map, label_map = generate_corpus(lines, if_shrink_feature=if_shrink_w_feature, thresholds=w_thresholds) char_count = dict() for feature in features: for word in feature: for tup in word: if tup not in char_count: char_count[tup] = 0 else: char_count[tup] += 1 if if_shrink_c_feature: shrink_char_count = [k for (k, v) in iter(char_count.items()) if v >= c_thresholds] char_map = {shrink_char_count[ind]: ind for ind in range(0, len(shrink_char_count))} else: char_map = {k: v for (v, k) in enumerate(char_count.keys())} char_map['<u>'] = len(char_map) # unk for char char_map[' '] = len(char_map) # concat for char char_map['\n'] = len(char_map) # eof for char return features, labels, feature_map, label_map, char_map def shrink_features(feature_map, features, thresholds): """ filter un-common features by threshold """ feature_count = {k: 0 for (k, v) in iter(feature_map.items())} for feature_list in features: for feature in feature_list: feature_count[feature] += 1 shrinked_feature_count = [k for (k, v) in iter(feature_count.items()) if v >= thresholds] feature_map = {shrinked_feature_count[ind]: (ind + 1) for ind in range(0, len(shrinked_feature_count))} #inserting unk to be 0 encoded feature_map['<unk>'] = 0 #inserting eof feature_map['<eof>'] = len(feature_map) return feature_map def generate_corpus(lines, if_shrink_feature=False, thresholds=1): """ generate label, feature, word dictionary and label dictionary args: lines : corpus if_shrink_feature: whether shrink word-dictionary threshold: threshold for shrinking word-dictionary """ features = list() labels = list() tmp_fl = list() tmp_ll = list() feature_map = dict() label_map = dict() for line in lines: if not (line.isspace() or (len(line) > 10 and line[0:10] == '-DOCSTART-')): line = line.rstrip('\n').split() tmp_fl.append(line[0]) if line[0] not in feature_map: feature_map[line[0]] = len(feature_map) + 1 #0 is for unk tmp_ll.append(line[-1]) if line[-1] not in label_map: label_map[line[-1]] = len(label_map) elif len(tmp_fl) > 0: features.append(tmp_fl) labels.append(tmp_ll) tmp_fl = list() tmp_ll = list() if len(tmp_fl) > 0: features.append(tmp_fl) labels.append(tmp_ll) label_map['<start>'] = len(label_map) label_map['<pad>'] = len(label_map) if if_shrink_feature: feature_map = shrink_features(feature_map, features, thresholds) else: #inserting unk to be 0 encoded feature_map['<unk>'] = 0 #inserting eof feature_map['<eof>'] = len(feature_map) return features, labels, feature_map, label_map def read_corpus(lines): """ convert corpus into features and labels """ features = list() labels = list() tmp_fl = list() tmp_ll = list() for line in lines: if not (line.isspace() or (len(line) > 10 and line[0:10] == '-DOCSTART-')): line = line.rstrip('\n').split() tmp_fl.append(line[0]) tmp_ll.append(line[-1]) elif len(tmp_fl) > 0: features.append(tmp_fl) labels.append(tmp_ll) tmp_fl = list() tmp_ll = list() if len(tmp_fl) > 0: features.append(tmp_fl) labels.append(tmp_ll) return features, labels def read_features(lines, multi_docs = True): """ convert un-annotated corpus into features """ if multi_docs: documents = list() features = list() tmp_fl = list() for line in lines: if_doc_end = (len(line) > 10 and line[0:10] == '-DOCSTART-') if not (line.isspace() or if_doc_end): line = line.split()[0] tmp_fl.append(line) else: if len(tmp_fl) > 0: features.append(tmp_fl) tmp_fl = list() if if_doc_end and len(features) > 0: documents.append(features) features = list() if len(tmp_fl) > 0: features.append(tmp_fl) if len(features) >0: documents.append(features) return documents else: features = list() tmp_fl = list() for line in lines: if not (line.isspace() or (len(line) > 10 and line[0:10] == '-DOCSTART-')): line = line.split()[0] tmp_fl.append(line) elif len(tmp_fl) > 0: features.append(tmp_fl) tmp_fl = list() if len(tmp_fl) > 0: features.append(tmp_fl) return features def shrink_embedding(feature_map, word_dict, word_embedding, caseless): """ shrink embedding dictionary to in-doc words only """ if caseless: feature_map = set([k.lower() for k in feature_map.keys()]) new_word_list = [k for k in word_dict.keys() if (k in feature_map)] new_word_dict = {k:v for (v, k) in enumerate(new_word_list)} new_word_list_ind = torch.LongTensor([word_dict[k] for k in new_word_list]) new_embedding = word_embedding[new_word_list_ind] return new_word_dict, new_embedding def encode_corpus(lines, f_map, l_map, if_lower = False): """ encode corpus into features and labels """ tmp_fl = [] tmp_ll = [] features = [] labels = [] for line in lines: if not (line.isspace() or (len(line) > 10 and line[0:10] == '-DOCSTART-')): line = line.rstrip('\n').split() tmp_fl.append(line[0]) tmp_ll.append(line[-1]) elif len(tmp_fl) > 0: features.append(tmp_fl) labels.append(tmp_ll) tmp_fl = list() tmp_ll = list() if len(tmp_fl) > 0: features.append(tmp_fl) labels.append(tmp_ll) if if_lower: features = list(map(lambda t: list(map(lambda x: x.lower(), t)), features)) feature_e = encode_safe(features, f_map, f_map['<unk>']) label_e = encode(labels, l_map) return feature_e, label_e def encode_corpus_c(lines, f_map, l_map, c_map): """ encode corpus into features (both word-level and char-level) and labels """ tmp_fl = [] tmp_ll = [] features = [] labels = [] for line in lines: if not (line.isspace() or (len(line) > 10 and line[0:10] == '-DOCSTART-')): line = line.rstrip('\n').split() tmp_fl.append(line[0]) tmp_ll.append(line[-1]) elif len(tmp_fl) > 0: features.append(tmp_fl) labels.append(tmp_ll) tmp_fl = list() tmp_ll = list() if len(tmp_fl) > 0: features.append(tmp_fl) labels.append(tmp_ll) feature_c = encode2char_safe(features, c_map) feature_e = encode_safe(features, f_map, f_map['<unk>']) label_e = encode(labels, l_map) return feature_c, feature_e, label_e def load_embedding(emb_file, delimiter, feature_map, caseless, unk, shrink_to_train=False): """ load embedding """ if caseless: feature_set = set([key.lower() for key in feature_map]) else: feature_set = set([key for key in feature_map]) word_dict = dict() embedding_array = list() for line in open(emb_file, 'r'): line = line.split(delimiter) vector = list(map(lambda t: float(t), filter(lambda n: n and not n.isspace(), line[1:]))) if shrink_to_train and line[0] not in feature_set: continue if line[0] == unk: word_dict['<unk>'] = len(word_dict) else: word_dict[line[0]] = len(word_dict) embedding_array.append(vector) embedding_tensor_1 = torch.FloatTensor(np.asarray(embedding_array)) emb_len = embedding_tensor_1.size(1) rand_embedding_count = 0 for key in feature_map: if caseless: key = key.lower() if key not in word_dict: word_dict[key] = len(word_dict) rand_embedding_count += 1 rand_embedding_tensor = torch.FloatTensor(rand_embedding_count, emb_len) init_embedding(rand_embedding_tensor) embedding_tensor = torch.cat((embedding_tensor_1, rand_embedding_tensor), 0) return word_dict, embedding_tensor def load_embedding_wlm(emb_file, delimiter, feature_map, full_feature_set, caseless, unk, emb_len, shrink_to_train=False, shrink_to_corpus=False): """ load embedding, indoc words would be listed before outdoc words args: emb_file: path to embedding file delimiter: delimiter of lines feature_map: word dictionary full_feature_set: all words in the corpus caseless: convert into casesless style unk: string for unknown token emb_len: dimension of embedding vectors shrink_to_train: whether to shrink out-of-training set or not shrink_to_corpus: whether to shrink out-of-corpus or not """ if caseless: feature_set = set([key.lower() for key in feature_map]) full_feature_set = set([key.lower() for key in full_feature_set]) else: feature_set = set([key for key in feature_map]) full_feature_set = set([key for key in full_feature_set]) #ensure <unk> is 0 word_dict = {v:(k+1) for (k,v) in enumerate(feature_set - set(['<unk>']))} word_dict['<unk>'] = 0 in_doc_freq_num = len(word_dict) rand_embedding_tensor = torch.FloatTensor(in_doc_freq_num, emb_len) init_embedding(rand_embedding_tensor) indoc_embedding_array = list() indoc_word_array = list() outdoc_embedding_array = list() outdoc_word_array = list() for line in open(emb_file, 'r'): line = line.split(delimiter) vector = list(map(lambda t: float(t), filter(lambda n: n and not n.isspace(), line[1:]))) if shrink_to_train and line[0] not in feature_set: continue if line[0] == unk: rand_embedding_tensor[0] = torch.FloatTensor(vector) #unk is 0 elif line[0] in word_dict: rand_embedding_tensor[word_dict[line[0]]] = torch.FloatTensor(vector) elif line[0] in full_feature_set: indoc_embedding_array.append(vector) indoc_word_array.append(line[0]) elif not shrink_to_corpus: outdoc_word_array.append(line[0]) outdoc_embedding_array.append(vector) embedding_tensor_0 = torch.FloatTensor(np.asarray(indoc_embedding_array)) if not shrink_to_corpus: embedding_tensor_1 = torch.FloatTensor(np.asarray(outdoc_embedding_array)) word_emb_len = embedding_tensor_0.size(1) assert(word_emb_len == emb_len) if shrink_to_corpus: embedding_tensor = torch.cat([rand_embedding_tensor, embedding_tensor_0], 0) else: embedding_tensor = torch.cat([rand_embedding_tensor, embedding_tensor_0, embedding_tensor_1], 0) for word in indoc_word_array: word_dict[word] = len(word_dict) in_doc_num = len(word_dict) if not shrink_to_corpus: for word in outdoc_word_array: word_dict[word] = len(word_dict) return word_dict, embedding_tensor, in_doc_num def calc_threshold_mean(features): """ calculate the threshold for bucket by mean """ lines_len = list(map(lambda t: len(t) + 1, features)) average = int(sum(lines_len) / len(lines_len)) lower_line = list(filter(lambda t: t < average, lines_len)) upper_line = list(filter(lambda t: t >= average, lines_len)) lower_average = int(sum(lower_line) / len(lower_line)) upper_average = int(sum(upper_line) / len(upper_line)) max_len = max(lines_len) return [lower_average, average, upper_average, max_len] def construct_bucket_mean_gd(input_features, input_label, word_dict, label_dict): """ Construct bucket by mean for greedy decode, word-level only """ # encode and padding features = encode_safe(input_features, word_dict, word_dict['<unk>']) labels = encode(input_label, label_dict) labels = list(map(lambda t: [label_dict['<start>']] + list(t), labels)) thresholds = calc_threshold_mean(features) return construct_bucket_gd(features, labels, thresholds, word_dict['<eof>'], label_dict['<pad>']) def construct_bucket_mean_vb(input_features, input_label, word_dict, label_dict, caseless): """ Construct bucket by mean for viterbi decode, word-level only """ # encode and padding if caseless: input_features = list(map(lambda t: list(map(lambda x: x.lower(), t)), input_features)) features = encode_safe(input_features, word_dict, word_dict['<unk>']) labels = encode(input_label, label_dict) labels = list(map(lambda t: [label_dict['<start>']] + list(t), labels)) thresholds = calc_threshold_mean(features) return construct_bucket_vb(features, labels, thresholds, word_dict['<eof>'], label_dict['<pad>'], len(label_dict)) def construct_bucket_mean_vb_wc(word_features, input_label, label_dict, char_dict, word_dict, caseless): """ Construct bucket by mean for viterbi decode, word-level and char-level """ # encode and padding char_features = encode2char_safe(word_features, char_dict) fea_len = [list(map(lambda t: len(t) + 1, f)) for f in char_features] forw_features = concatChar(char_features, char_dict) labels = encode(input_label, label_dict) labels = list(map(lambda t: [label_dict['<start>']] + list(t), labels)) thresholds = calc_threshold_mean(fea_len) if caseless: word_features = list(map(lambda t: list(map(lambda x: x.lower(), t)), word_features)) word_features = encode_safe(word_features, word_dict, word_dict['<unk>']) return construct_bucket_vb_wc(word_features, forw_features, fea_len, labels, thresholds, word_dict['<eof>'], char_dict['\n'], label_dict['<pad>'], len(label_dict)) def construct_bucket_vb_wc(word_features, forw_features, fea_len, input_labels, thresholds, pad_word_feature, pad_char_feature, pad_label, label_size): """ Construct bucket by thresholds for viterbi decode, word-level and char-level """ # construct corpus for language model pre-training forw_corpus = [pad_char_feature] for forw_feature in forw_features: forw_corpus.extend(forw_feature + [pad_char_feature]) back_corpus = forw_corpus[::-1] # two way construct, first build the bucket, then calculate padding length, then do the padding buckets = [[[], [], [], [], [], [], [], []] for ind in range(len(thresholds))] # forw, forw_ind, back, back_in, label, mask buckets_len = [0 for ind in range(len(thresholds))] # thresholds is the padded length for fea # buckets_len is the padded length for char for f_f, f_l in zip(forw_features, fea_len): cur_len_1 = len(f_l) + 1 idx = 0 while thresholds[idx] < cur_len_1: idx += 1 tmp_concat_len = len(f_f) + thresholds[idx] - len(f_l) if buckets_len[idx] < tmp_concat_len: buckets_len[idx] = tmp_concat_len # calc padding for f_f, f_l, w_f, i_l in zip(forw_features, fea_len, word_features, input_labels): cur_len = len(f_l) idx = 0 cur_len_1 = cur_len + 1 while thresholds[idx] < cur_len_1: idx += 1 padded_feature = f_f + [pad_char_feature] * (buckets_len[idx] - len(f_f)) # pad feature with <'\n'>, at least one padded_feature_len = f_l + [1] * (thresholds[idx] - len(f_l)) # pad feature length with <'\n'>, at least one padded_feature_len_cum = list(itertools.accumulate(padded_feature_len)) # start from 0, but the first is ' ', so the position need not to be -1 buckets[idx][0].append(padded_feature) # char buckets[idx][1].append(padded_feature_len_cum) buckets[idx][2].append(padded_feature[::-1]) buckets[idx][3].append([buckets_len[idx] - 1] + [buckets_len[idx] - 1 - tup for tup in padded_feature_len_cum[:-1]]) buckets[idx][4].append(w_f + [pad_word_feature] * (thresholds[idx] - cur_len)) #word buckets[idx][5].append([i_l[ind] * label_size + i_l[ind + 1] for ind in range(0, cur_len)] + [i_l[cur_len] * label_size + pad_label] + [pad_label * label_size + pad_label] * (thresholds[idx] - cur_len_1)) # has additional start, label buckets[idx][6].append([1] * cur_len_1 + [0] * (thresholds[idx] - cur_len_1)) # has additional start, mask buckets[idx][7].append([len(f_f) + thresholds[idx] - len(f_l), cur_len_1]) bucket_dataset = [CRFDataset_WC(torch.LongTensor(bucket[0]), torch.LongTensor(bucket[1]), torch.LongTensor(bucket[2]), torch.LongTensor(bucket[3]), torch.LongTensor(bucket[4]), torch.LongTensor(bucket[5]), torch.ByteTensor(bucket[6]), torch.LongTensor(bucket[7])) for bucket in buckets] return bucket_dataset, forw_corpus, back_corpus def construct_bucket_vb(input_features, input_labels, thresholds, pad_feature, pad_label, label_size): """ Construct bucket by thresholds for viterbi decode, word-level only """ buckets = [[[], [], []] for _ in range(len(thresholds))] for feature, label in zip(input_features, input_labels): cur_len = len(feature) idx = 0 cur_len_1 = cur_len + 1 while thresholds[idx] < cur_len_1: idx += 1 buckets[idx][0].append(feature + [pad_feature] * (thresholds[idx] - cur_len)) buckets[idx][1].append([label[ind] * label_size + label[ind + 1] for ind in range(0, cur_len)] + [ label[cur_len] * label_size + pad_label] + [pad_label * label_size + pad_label] * ( thresholds[idx] - cur_len_1)) buckets[idx][2].append([1] * cur_len_1 + [0] * (thresholds[idx] - cur_len_1)) bucket_dataset = [CRFDataset(torch.LongTensor(bucket[0]), torch.LongTensor(bucket[1]), torch.ByteTensor(bucket[2])) for bucket in buckets] return bucket_dataset def construct_bucket_gd(input_features, input_labels, thresholds, pad_feature, pad_label): """ Construct bucket by thresholds for greedy decode, word-level only """ buckets = [[[], [], []] for ind in range(len(thresholds))] for feature, label in zip(input_features, input_labels): cur_len = len(feature) cur_len_1 = cur_len + 1 idx = 0 while thresholds[idx] < cur_len_1: idx += 1 buckets[idx][0].append(feature + [pad_feature] * (thresholds[idx] - cur_len)) buckets[idx][1].append(label[1:] + [pad_label] * (thresholds[idx] - cur_len)) buckets[idx][2].append(label + [pad_label] * (thresholds[idx] - cur_len_1)) bucket_dataset = [CRFDataset(torch.LongTensor(bucket[0]), torch.LongTensor(bucket[1]), torch.LongTensor(bucket[2])) for bucket in buckets] return bucket_dataset def find_length_from_feats(feats, feat_to_ix): """ find length of unpadded features based on feature """ end_position = len(feats) - 1 for position, feat in enumerate(feats): if feat.data[0] == feat_to_ix['<eof>']: end_position = position break return end_position + 1 def find_length_from_labels(labels, label_to_ix): """ find length of unpadded features based on labels """ end_position = len(labels) - 1 for position, label in enumerate(labels): if label == label_to_ix['<pad>']: end_position = position break return end_position def revlut(lut): return {v: k for k, v in lut.items()} # Turn a sequence of IOB chunks into single tokens def iob_to_spans(sequence, lut, strict_iob2=False): """ convert to iob to span """ iobtype = 2 if strict_iob2 else 1 chunks = [] current = None for i, y in enumerate(sequence): label = lut[y] if label.startswith('B-'): if current is not None: chunks.append('@'.join(current)) current = [label.replace('B-', ''), '%d' % i] elif label.startswith('I-'): if current is not None: base = label.replace('I-', '') if base == current[0]: current.append('%d' % i) else: chunks.append('@'.join(current)) if iobtype == 2: print('Warning, type=IOB2, unexpected format ([%s] follows other tag type [%s] @ %d)' % ( label, current[0], i)) current = [base, '%d' % i] else: current = [label.replace('I-', ''), '%d' % i] if iobtype == 2: print('Warning, unexpected format (I before B @ %d) %s' % (i, label)) else: if current is not None: chunks.append('@'.join(current)) current = None if current is not None: chunks.append('@'.join(current)) return set(chunks) # Turn a sequence of IOBES chunks into single tokens def iobes_to_spans(sequence, lut, strict_iob2=False): """ convert to iobes to span """ iobtype = 2 if strict_iob2 else 1 chunks = [] current = None for i, y in enumerate(sequence): label = lut[y] if label.startswith('B-'): if current is not None: chunks.append('@'.join(current)) current = [label.replace('B-', ''), '%d' % i] elif label.startswith('S-'): if current is not None: chunks.append('@'.join(current)) current = None base = label.replace('S-', '') chunks.append('@'.join([base, '%d' % i])) elif label.startswith('I-'): if current is not None: base = label.replace('I-', '') if base == current[0]: current.append('%d' % i) else: chunks.append('@'.join(current)) if iobtype == 2: print('Warning') current = [base, '%d' % i] else: current = [label.replace('I-', ''), '%d' % i] if iobtype == 2: print('Warning') elif label.startswith('E-'): if current is not None: base = label.replace('E-', '') if base == current[0]: current.append('%d' % i) chunks.append('@'.join(current)) current = None else: chunks.append('@'.join(current)) if iobtype == 2: print('Warning') current = [base, '%d' % i] chunks.append('@'.join(current)) current = None else: current = [label.replace('E-', ''), '%d' % i] if iobtype == 2: print('Warning') chunks.append('@'.join(current)) current = None else: if current is not None: chunks.append('@'.join(current)) current = None if current is not None: chunks.append('@'.join(current)) return set(chunks) def fill_y(nc, yidx): """ fill y to dense matrix """ batchsz = yidx.shape[0] siglen = yidx.shape[1] dense = np.zeros((batchsz, siglen, nc), dtype=np.int) for i in range(batchsz): for j in range(siglen): idx = int(yidx[i, j]) if idx > 0: dense[i, j, idx] = 1 return dense def save_checkpoint(state, track_list, filename): """ save checkpoint """ with open(filename+'.json', 'w') as f: json.dump(track_list, f) torch.save(state, filename+'.model') def adjust_learning_rate(optimizer, lr): """ shrink learning rate for pytorch """ for param_group in optimizer.param_groups: param_group['lr'] = lr def init_embedding(input_embedding): """ Initialize embedding """ bias = np.sqrt(3.0 / input_embedding.size(1)) nn.init.uniform_(input_embedding, -bias, bias) def init_linear(input_linear): """ Initialize linear transformation """ bias = np.sqrt(6.0 / (input_linear.weight.size(0) + input_linear.weight.size(1))) nn.init.uniform_(input_linear.weight, -bias, bias) if input_linear.bias is not None: input_linear.bias.data.zero_() def init_lstm(input_lstm): """ Initialize lstm """ for ind in range(0, input_lstm.num_layers): weight = eval('input_lstm.weight_ih_l'+str(ind)) bias = np.sqrt(6.0 / (weight.size(0)/4 + weight.size(1))) nn.init.uniform_(weight, -bias, bias) weight = eval('input_lstm.weight_hh_l'+str(ind)) bias = np.sqrt(6.0 / (weight.size(0)/4 + weight.size(1))) nn.init.uniform_(weight, -bias, bias) if input_lstm.bias: for ind in range(0, input_lstm.num_layers): weight = eval('input_lstm.bias_ih_l'+str(ind)) weight.data.zero_() weight.data[input_lstm.hidden_size: 2 * input_lstm.hidden_size] = 1 weight = eval('input_lstm.bias_hh_l'+str(ind)) weight.data.zero_() weight.data[input_lstm.hidden_size: 2 * input_lstm.hidden_size] = 1
29,437
34.424789
243
py
LM-LSTM-CRF
LM-LSTM-CRF-master/model/lstm_crf.py
""" .. module:: lstm_crf :synopsis: lstm_crf .. moduleauthor:: Liyuan Liu """ import torch import torch.autograd as autograd import torch.nn as nn import model.crf as crf import model.utils as utils class LSTM_CRF(nn.Module): """LSTM_CRF model args: vocab_size: size of word dictionary tagset_size: size of label set embedding_dim: size of word embedding hidden_dim: size of word-level blstm hidden dim rnn_layers: number of word-level lstm layers dropout_ratio: dropout ratio large_CRF: use CRF_L or not, refer model.crf.CRF_L and model.crf.CRF_S for more details """ def __init__(self, vocab_size, tagset_size, embedding_dim, hidden_dim, rnn_layers, dropout_ratio, large_CRF=True): super(LSTM_CRF, self).__init__() self.embedding_dim = embedding_dim self.hidden_dim = hidden_dim self.vocab_size = vocab_size self.word_embeds = nn.Embedding(vocab_size, embedding_dim) self.lstm = nn.LSTM(embedding_dim, hidden_dim // 2, num_layers=rnn_layers, bidirectional=True, dropout=dropout_ratio) self.rnn_layers = rnn_layers self.dropout1 = nn.Dropout(p=dropout_ratio) self.dropout2 = nn.Dropout(p=dropout_ratio) self.tagset_size = tagset_size if large_CRF: self.crf = crf.CRF_L(hidden_dim, tagset_size) else: self.crf = crf.CRF_S(hidden_dim, tagset_size) self.batch_size = 1 self.seq_length = 1 def rand_init_hidden(self): """ random initialize hidden variable """ return autograd.Variable( torch.randn(2 * self.rnn_layers, self.batch_size, self.hidden_dim // 2)), autograd.Variable( torch.randn(2 * self.rnn_layers, self.batch_size, self.hidden_dim // 2)) def set_batch_size(self, bsize): """ set batch size """ self.batch_size = bsize def set_batch_seq_size(self, sentence): """ set batch size and sequence length """ tmp = sentence.size() self.seq_length = tmp[0] self.batch_size = tmp[1] def load_pretrained_embedding(self, pre_embeddings): """ load pre-trained word embedding args: pre_word_embeddings (self.word_size, self.word_dim) : pre-trained embedding """ assert (pre_embeddings.size()[1] == self.embedding_dim) self.word_embeds.weight = nn.Parameter(pre_embeddings) def rand_init_embedding(self): utils.init_embedding(self.word_embeds.weight) def rand_init(self, init_embedding=False): """ random initialization args: init_embedding: random initialize embedding or not """ if init_embedding: utils.init_embedding(self.word_embeds.weight) utils.init_lstm(self.lstm) self.crf.rand_init() def forward(self, sentence, hidden=None): ''' args: sentence (word_seq_len, batch_size) : word-level representation of sentence hidden: initial hidden state return: crf output (word_seq_len, batch_size, tag_size, tag_size), hidden ''' self.set_batch_seq_size(sentence) embeds = self.word_embeds(sentence) d_embeds = self.dropout1(embeds) lstm_out, hidden = self.lstm(d_embeds, hidden) lstm_out = lstm_out.view(-1, self.hidden_dim) d_lstm_out = self.dropout2(lstm_out) crf_out = self.crf(d_lstm_out) crf_out = crf_out.view(self.seq_length, self.batch_size, self.tagset_size, self.tagset_size) return crf_out, hidden
3,738
30.420168
118
py
LM-LSTM-CRF
LM-LSTM-CRF-master/model/evaluator.py
""" .. module:: evaluator :synopsis: evaluation method (f1 score and accuracy) .. moduleauthor:: Liyuan Liu, Frank Xu """ import torch import numpy as np import itertools import model.utils as utils from torch.autograd import Variable from model.crf import CRFDecode_vb class eval_batch: """Base class for evaluation, provide method to calculate f1 score and accuracy args: packer: provide method to convert target into original space [TODO: need to improve] l_map: dictionary for labels """ def __init__(self, packer, l_map): self.packer = packer self.l_map = l_map self.r_l_map = utils.revlut(l_map) self.totalp_counts={} self.truep_counts={} self.fn_counts={} self.fp_counts={} self.f1={} def reset(self): """ re-set all states """ self.correct_labels = 0 self.total_labels = 0 self.gold_count = 0 self.guess_count = 0 self.overlap_count = 0 self.totalp_counts={} self.truep_counts={} self.fn_counts={} self.fp_counts={} self.f1={} def calc_f1_batch(self, decoded_data, target_data): """ update statics for f1 score args: decoded_data (batch_size, seq_len): prediction sequence target_data (batch_size, seq_len): ground-truth """ batch_decoded = torch.unbind(decoded_data, 1) batch_targets = torch.unbind(target_data, 0) for decoded, target in zip(batch_decoded, batch_targets): gold = self.packer.convert_for_eval(target) # remove padding length = utils.find_length_from_labels(gold, self.l_map) gold = gold[:length] best_path = decoded[:length] correct_labels_i, total_labels_i, gold_count_i, guess_count_i, overlap_count_i = self.eval_instance(best_path.numpy(), gold.numpy()) self.correct_labels += correct_labels_i self.total_labels += total_labels_i self.gold_count += gold_count_i self.guess_count += guess_count_i self.overlap_count += overlap_count_i def calc_acc_batch(self, decoded_data, target_data): """ update statics for accuracy args: decoded_data (batch_size, seq_len): prediction sequence target_data (batch_size, seq_len): ground-truth """ batch_decoded = torch.unbind(decoded_data, 1) batch_targets = torch.unbind(target_data, 0) for decoded, target in zip(batch_decoded, batch_targets): gold = self.packer.convert_for_eval(target) # remove padding length = utils.find_length_from_labels(gold, self.l_map) gold = gold[:length].numpy() best_path = decoded[:length].numpy() self.total_labels += length self.correct_labels += np.sum(np.equal(best_path, gold)) def f1_score(self): """ calculate f1 score based on statics """ if self.guess_count == 0: return {'total': (0.0, 0.0, 0.0, 0.0, '')} precision = self.overlap_count / float(self.guess_count) recall = self.overlap_count / float(self.gold_count) if precision == 0.0 or recall == 0.0: return {'total': (0.0, 0.0, 0.0, 0.0, '')} f = 2 * (precision * recall) / (precision + recall) accuracy = float(self.correct_labels) / self.total_labels message="" self.f1['total'] = (f, precision, recall, accuracy, message) for label in self.totalp_counts: tp = self.truep_counts.get(label,1) fn = sum(self.fn_counts.get(label,{}).values()) fp = sum(self.fp_counts.get(label,{}).values()) # print(label, str(tp), str(fp), str(fn), str(self.totalp_counts.get(label,0))) precision = tp / float(tp+fp+1e-9) recall = tp / float(tp+fn+1e-9) f = 2 * (precision * recall) / (precision + recall+1e-9) message = str(self.fn_counts.get(label, {})) self.f1[label] = (f, precision, recall, 0, message) return self.f1 def acc_score(self): """ calculate accuracy score based on statics """ if 0 == self.total_labels: return 0.0 accuracy = float(self.correct_labels) / self.total_labels return accuracy def eval_instance(self, best_path, gold): """ update statics for one instance args: best_path (seq_len): predicted gold (seq_len): ground-truth """ total_labels = len(best_path) correct_labels = np.sum(np.equal(best_path, gold)) for i in range(total_labels): gold_label = self.r_l_map[gold[i]] guessed_label = self.r_l_map[best_path[i]] self.totalp_counts[gold_label] = 1 + self.totalp_counts.get(gold_label,0) if gold_label == guessed_label: self.truep_counts[gold_label] = 1 + self.truep_counts.get(gold_label,0) else: val = self.fn_counts.get(gold_label,{}) val[guessed_label] = 1+ val.get(guessed_label,0) self.fn_counts[gold_label]=val val2 = self.fp_counts.get(guessed_label,{}) val2[gold_label] = 1+ val2.get(gold_label,0) self.fp_counts[guessed_label] = val2 gold_chunks = utils.iobes_to_spans(gold, self.r_l_map) gold_count = len(gold_chunks) guess_chunks = utils.iobes_to_spans(best_path, self.r_l_map) guess_count = len(guess_chunks) overlap_chunks = gold_chunks & guess_chunks overlap_count = len(overlap_chunks) return correct_labels, total_labels, gold_count, guess_count, overlap_count class eval_w(eval_batch): """evaluation class for word level model (LSTM-CRF) args: packer: provide method to convert target into original space [TODO: need to improve] l_map: dictionary for labels score_type: use f1score with using 'f' """ def __init__(self, packer, l_map, score_type): eval_batch.__init__(self, packer, l_map) self.decoder = CRFDecode_vb(len(l_map), l_map['<start>'], l_map['<pad>']) if 'f' in score_type: self.eval_b = self.calc_f1_batch self.calc_s = self.f1_score else: self.eval_b = self.calc_acc_batch self.calc_s = self.acc_score def calc_score(self, ner_model, dataset_loader): """ calculate score for pre-selected metrics args: ner_model: LSTM-CRF model dataset_loader: loader class for test set """ ner_model.eval() self.reset() for feature, tg, mask in itertools.chain.from_iterable(dataset_loader): fea_v, _, mask_v = self.packer.repack_vb(feature, tg, mask) scores, _ = ner_model(fea_v) decoded = self.decoder.decode(scores.data, mask_v.data) self.eval_b(decoded, tg) return self.calc_s() class eval_wc(eval_batch): """evaluation class for LM-LSTM-CRF args: packer: provide method to convert target into original space [TODO: need to improve] l_map: dictionary for labels score_type: use f1score with using 'f' """ def __init__(self, packer, l_map, score_type): eval_batch.__init__(self, packer, l_map) self.decoder = CRFDecode_vb(len(l_map), l_map['<start>'], l_map['<pad>']) if 'f' in score_type: self.eval_b = self.calc_f1_batch self.calc_s = self.f1_score else: self.eval_b = self.calc_acc_batch self.calc_s = self.acc_score def calc_score(self, ner_model, dataset_loader): """ calculate score for pre-selected metrics args: ner_model: LM-LSTM-CRF model dataset_loader: loader class for test set """ ner_model.eval() self.reset() for f_f, f_p, b_f, b_p, w_f, tg, mask_v, len_v in itertools.chain.from_iterable(dataset_loader): f_f, f_p, b_f, b_p, w_f, _, mask_v = self.packer.repack_vb(f_f, f_p, b_f, b_p, w_f, tg, mask_v, len_v) scores = ner_model(f_f, f_p, b_f, b_p, w_f) decoded = self.decoder.decode(scores.data, mask_v.data) self.eval_b(decoded, tg) return self.calc_s()
8,538
33.01992
144
py
LM-LSTM-CRF
LM-LSTM-CRF-master/model/crf.py
""" .. module:: crf :synopsis: conditional random field .. moduleauthor:: Liyuan Liu """ import torch import torch.autograd as autograd import torch.nn as nn import torch.optim as optim import torch.sparse as sparse import model.utils as utils class CRF_L(nn.Module): """Conditional Random Field (CRF) layer. This version is used in Ma et al. 2016, has more parameters than CRF_S args: hidden_dim : input dim size tagset_size: target_set_size if_biase: whether allow bias in linear trans """ def __init__(self, hidden_dim, tagset_size, if_bias=True): super(CRF_L, self).__init__() self.tagset_size = tagset_size self.hidden2tag = nn.Linear(hidden_dim, self.tagset_size * self.tagset_size, bias=if_bias) def rand_init(self): """random initialization """ utils.init_linear(self.hidden2tag) def forward(self, feats): """ args: feats (batch_size, seq_len, hidden_dim) : input score from previous layers return: output from crf layer (batch_size, seq_len, tag_size, tag_size) """ return self.hidden2tag(feats).view(-1, self.tagset_size, self.tagset_size) class CRF_S(nn.Module): """Conditional Random Field (CRF) layer. This version is used in Lample et al. 2016, has less parameters than CRF_L. args: hidden_dim: input dim size tagset_size: target_set_size if_biase: whether allow bias in linear trans """ def __init__(self, hidden_dim, tagset_size, if_bias=True): super(CRF_S, self).__init__() self.tagset_size = tagset_size self.hidden2tag = nn.Linear(hidden_dim, self.tagset_size, bias=if_bias) self.transitions = nn.Parameter(torch.Tensor(self.tagset_size, self.tagset_size)) def rand_init(self): """random initialization """ utils.init_linear(self.hidden2tag) self.transitions.data.zero_() def forward(self, feats): """ args: feats (batch_size, seq_len, hidden_dim) : input score from previous layers return: output from crf layer ( (batch_size * seq_len), tag_size, tag_size) """ scores = self.hidden2tag(feats).view(-1, self.tagset_size, 1) ins_num = scores.size(0) crf_scores = scores.expand(ins_num, self.tagset_size, self.tagset_size) + self.transitions.view(1, self.tagset_size, self.tagset_size).expand(ins_num, self.tagset_size, self.tagset_size) return crf_scores class CRFRepack: """Packer for word level model args: tagset_size: target_set_size if_cuda: whether use GPU """ def __init__(self, tagset_size, if_cuda): self.tagset_size = tagset_size self.if_cuda = if_cuda def repack_vb(self, feature, target, mask): """packer for viterbi loss args: feature (Seq_len, Batch_size): input feature target (Seq_len, Batch_size): output target mask (Seq_len, Batch_size): padding mask return: feature (Seq_len, Batch_size), target (Seq_len, Batch_size), mask (Seq_len, Batch_size) """ if self.if_cuda: fea_v = feature.transpose(0, 1).cuda() tg_v = target.transpose(0, 1).unsqueeze(2).cuda() mask_v = mask.transpose(0, 1).cuda() else: fea_v = feature.transpose(0, 1) tg_v = target.transpose(0, 1).contiguous().unsqueeze(2) mask_v = mask.transpose(0, 1).contiguous() return fea_v, tg_v, mask_v def repack_gd(self, feature, target, current): """packer for greedy loss args: feature (Seq_len, Batch_size): input feature target (Seq_len, Batch_size): output target current (Seq_len, Batch_size): current state return: feature (Seq_len, Batch_size), target (Seq_len * Batch_size), current (Seq_len * Batch_size, 1, 1) """ if self.if_cuda: fea_v = feature.transpose(0, 1).cuda() ts_v = target.transpose(0, 1).cuda().view(-1) cs_v = current.transpose(0, 1).cuda().view(-1, 1, 1) else: fea_v = feature.transpose(0, 1) ts_v = target.transpose(0, 1).contiguous().view(-1) cs_v = current.transpose(0, 1).contiguous().view(-1, 1, 1) return fea_v, ts_v, cs_v def convert_for_eval(self, target): """convert target to original decoding args: target: input labels used in training return: output labels used in test """ return target % self.tagset_size class CRFRepack_WC: """Packer for model with char-level and word-level args: tagset_size: target_set_size if_cuda: whether use GPU """ def __init__(self, tagset_size, if_cuda): self.tagset_size = tagset_size self.if_cuda = if_cuda def repack_vb(self, fc_feature, fc_position, bc_feature, bc_position, word_feature, target, mask, batch_len): """packer for viterbi loss args: fc_feature (Char_Seq_len, Batch_size) : forward_char input feature fc_position (Word_Seq_len, Batch_size) : forward_char input position bc_feature (Char_Seq_len, Batch_size) : backward_char input feature bc_position (Word_Seq_len, Batch_size) : backward_char input position word_feature (Word_Seq_len, Batch_size) : input word feature target (Seq_len, Batch_size) : output target mask (Word_Seq_len, Batch_size) : padding mask batch_len (Batch_size, 2) : length of instances in one batch return: f_f (Char_Reduced_Seq_len, Batch_size), f_p (Word_Reduced_Seq_len, Batch_size), b_f (Char_Reduced_Seq_len, Batch_size), b_p (Word_Reduced_Seq_len, Batch_size), w_f (size Word_Seq_Len, Batch_size), target (Reduced_Seq_len, Batch_size), mask (Word_Reduced_Seq_len, Batch_size) """ mlen, _ = batch_len.max(0) mlen = mlen.squeeze() ocl = bc_feature.size(1) if self.if_cuda: fc_feature = fc_feature[:, 0:mlen[0]].transpose(0, 1).cuda() fc_position = fc_position[:, 0:mlen[1]].transpose(0, 1).cuda() bc_feature = bc_feature[:, -mlen[0]:].transpose(0, 1).cuda() bc_position = (bc_position[:, 0:mlen[1]] - ocl + mlen[0]).transpose(0, 1).cuda() word_feature = word_feature[:, 0:mlen[1]].transpose(0, 1).cuda() tg_v = target[:, 0:mlen[1]].transpose(0, 1).unsqueeze(2).cuda() mask_v = mask[:, 0:mlen[1]].transpose(0, 1).cuda() else: fc_feature = fc_feature[:, 0:mlen[0]].transpose(0, 1) fc_position = fc_position[:, 0:mlen[1]].transpose(0, 1) bc_feature = bc_feature[:, -mlen[0]:].transpose(0, 1) bc_position = (bc_position[:, 0:mlen[1]] - ocl + mlen[0]).transpose(0, 1) word_feature = word_feature[:, 0:mlen[1]].transpose(0, 1) tg_v = target[:, 0:mlen[1]].transpose(0, 1).unsqueeze(2) mask_v = mask[:, 0:mlen[1]].transpose(0, 1).contiguous() return fc_feature, fc_position, bc_feature, bc_position, word_feature, tg_v, mask_v def convert_for_eval(self, target): """convert for eval args: target: input labels used in training return: output labels used in test """ return target % self.tagset_size class CRFLoss_gd(nn.Module): """loss for greedy decode loss, i.e., although its for CRF Layer, we calculate the loss as .. math:: \sum_{j=1}^n \log (p(\hat{y}_{j+1}|z_{j+1}, \hat{y}_{j})) instead of .. math:: \sum_{j=1}^n \log (\phi(\hat{y}_{j-1}, \hat{y}_j, \mathbf{z}_j)) - \log (\sum_{\mathbf{y}' \in \mathbf{Y}(\mathbf{Z})} \prod_{j=1}^n \phi(y'_{j-1}, y'_j, \mathbf{z}_j) ) args: tagset_size: target_set_size start_tag: ind for <start> end_tag: ind for <pad> average_batch: whether average the loss among batch """ def __init__(self, tagset_size, start_tag, end_tag, average_batch=True): super(CRFLoss_gd, self).__init__() self.tagset_size = tagset_size self.average_batch = average_batch self.crit = nn.CrossEntropyLoss(size_average=self.average_batch) def forward(self, scores, target, current): """ args: scores (Word_Seq_len, Batch_size, target_size_from, target_size_to): crf scores target (Word_Seq_len, Batch_size): golden list current (Word_Seq_len, Batch_size): current state return: crf greedy loss """ ins_num = current.size(0) current = current.expand(ins_num, 1, self.tagset_size) scores = scores.view(ins_num, self.tagset_size, self.tagset_size) current_score = torch.gather(scores, 1, current).squeeze() return self.crit(current_score, target) class CRFLoss_vb(nn.Module): """loss for viterbi decode .. math:: \sum_{j=1}^n \log (\phi(\hat{y}_{j-1}, \hat{y}_j, \mathbf{z}_j)) - \log (\sum_{\mathbf{y}' \in \mathbf{Y}(\mathbf{Z})} \prod_{j=1}^n \phi(y'_{j-1}, y'_j, \mathbf{z}_j) ) args: tagset_size: target_set_size start_tag: ind for <start> end_tag: ind for <pad> average_batch: whether average the loss among batch """ def __init__(self, tagset_size, start_tag, end_tag, average_batch=True): super(CRFLoss_vb, self).__init__() self.tagset_size = tagset_size self.start_tag = start_tag self.end_tag = end_tag self.average_batch = average_batch def forward(self, scores, target, mask): """ args: scores (seq_len, bat_size, target_size_from, target_size_to) : crf scores target (seq_len, bat_size, 1) : golden state mask (size seq_len, bat_size) : mask for padding return: loss """ # calculate batch size and seq len seq_len = scores.size(0) bat_size = scores.size(1) # calculate sentence score tg_energy = torch.gather(scores.view(seq_len, bat_size, -1), 2, target).view(seq_len, bat_size) # seq_len * bat_size tg_energy = tg_energy.masked_select(mask).sum() # calculate forward partition score # build iter seq_iter = enumerate(scores) # the first score should start with <start> _, inivalues = seq_iter.__next__() # bat_size * from_target_size * to_target_size # only need start from start_tag partition = inivalues[:, self.start_tag, :].clone() # bat_size * to_target_size # iter over last scores for idx, cur_values in seq_iter: # previous to_target is current from_target # partition: previous results log(exp(from_target)), #(batch_size * from_target) # cur_values: bat_size * from_target * to_target cur_values = cur_values + partition.contiguous().view(bat_size, self.tagset_size, 1).expand(bat_size, self.tagset_size, self.tagset_size) cur_partition = utils.log_sum_exp(cur_values, self.tagset_size) # (bat_size * from_target * to_target) -> (bat_size * to_target) # partition = utils.switch(partition, cur_partition, mask[idx].view(bat_size, 1).expand(bat_size, self.tagset_size)).view(bat_size, -1) mask_idx = mask[idx, :].view(bat_size, 1).expand(bat_size, self.tagset_size) partition.masked_scatter_(mask_idx, cur_partition.masked_select(mask_idx)) #0 for partition, 1 for cur_partition #only need end at end_tag partition = partition[:, self.end_tag].sum() # average = mask.sum() # average_batch if self.average_batch: loss = (partition - tg_energy) / bat_size else: loss = (partition - tg_energy) return loss class CRFDecode_vb(): """Batch-mode viterbi decode args: tagset_size: target_set_size start_tag: ind for <start> end_tag: ind for <pad> average_batch: whether average the loss among batch """ def __init__(self, tagset_size, start_tag, end_tag, average_batch=True): self.tagset_size = tagset_size self.start_tag = start_tag self.end_tag = end_tag self.average_batch = average_batch def decode(self, scores, mask): """Find the optimal path with viterbe decode args: scores (size seq_len, bat_size, target_size_from, target_size_to) : crf scores mask (seq_len, bat_size) : mask for padding return: decoded sequence (size seq_len, bat_size) """ # calculate batch size and seq len seq_len = scores.size(0) bat_size = scores.size(1) mask = 1 - mask decode_idx = torch.LongTensor(seq_len-1, bat_size) # calculate forward score and checkpoint # build iter seq_iter = enumerate(scores) # the first score should start with <start> _, inivalues = seq_iter.__next__() # bat_size * from_target_size * to_target_size # only need start from start_tag forscores = inivalues[:, self.start_tag, :] # bat_size * to_target_size back_points = list() # iter over last scores for idx, cur_values in seq_iter: # previous to_target is current from_target # partition: previous results log(exp(from_target)), #(batch_size * from_target) # cur_values: bat_size * from_target * to_target cur_values = cur_values + forscores.contiguous().view(bat_size, self.tagset_size, 1).expand(bat_size, self.tagset_size, self.tagset_size) forscores, cur_bp = torch.max(cur_values, 1) cur_bp.masked_fill_(mask[idx].view(bat_size, 1).expand(bat_size, self.tagset_size), self.end_tag) back_points.append(cur_bp) pointer = back_points[-1][:, self.end_tag] decode_idx[-1] = pointer for idx in range(len(back_points)-2, -1, -1): back_point = back_points[idx] index = pointer.contiguous().view(-1,1) pointer = torch.gather(back_point, 1, index).view(-1) decode_idx[idx] = pointer return decode_idx
14,450
36.73107
287
py
MICRO
MICRO-main/codes/main.py
from datetime import datetime import math import os import random import sys from time import time from tqdm import tqdm import numpy as np import torch import torch.nn as nn import torch.nn.functional as F import torch.optim as optim import torch.sparse as sparse from utility.parser import parse_args from Models import MICRO from utility.batch_test import * from utility.logging import Logger args = parse_args() class Trainer(object): def __init__(self, data_config): # argument settings self.n_users = data_config['n_users'] self.n_items = data_config['n_items'] self.task_name = "%s_%s_%s" % (datetime.now().strftime('%Y-%m-%d %H:%M:%S'), args.dataset, args.cf_model,) self.logger = Logger(filename=self.task_name, is_debug=args.debug) self.logger.logging("PID: %d" % os.getpid()) self.logger.logging(str(args)) self.mess_dropout = eval(args.mess_dropout) self.lr = args.lr self.emb_dim = args.embed_size self.batch_size = args.batch_size self.weight_size = eval(args.weight_size) self.n_layers = len(self.weight_size) self.regs = eval(args.regs) self.decay = self.regs[0] self.norm_adj = data_config['norm_adj'] self.norm_adj = self.sparse_mx_to_torch_sparse_tensor(self.norm_adj).float().cuda() image_feats = np.load('../data/{}/image_feat.npy'.format(args.dataset)) text_feats = np.load('../data/{}/text_feat.npy'.format(args.dataset)) self.model = MICRO(self.n_users, self.n_items, self.emb_dim, self.weight_size, self.mess_dropout, image_feats, text_feats) self.model = self.model.cuda() self.optimizer = optim.Adam(self.model.parameters(), lr=self.lr) self.lr_scheduler = self.set_lr_scheduler() def set_lr_scheduler(self): fac = lambda epoch: 0.96 ** (epoch / 50) scheduler = optim.lr_scheduler.LambdaLR(self.optimizer, lr_lambda=fac) return scheduler def test(self, users_to_test, is_val): self.model.eval() with torch.no_grad(): ua_embeddings, ia_embeddings, *rest = self.model(self.norm_adj, build_item_graph=True) result = test_torch(ua_embeddings, ia_embeddings, users_to_test, is_val) return result def train(self): training_time_list = [] loss_loger, pre_loger, rec_loger, ndcg_loger, hit_loger = [], [], [], [], [] stopping_step = 0 should_stop = False cur_best_pre_0 = 0. n_batch = data_generator.n_train // args.batch_size + 1 best_recall = 0 for epoch in (range(args.epoch)): t1 = time() loss, mf_loss, emb_loss, reg_loss = 0., 0., 0., 0. contrastive_loss = 0. n_batch = data_generator.n_train // args.batch_size + 1 f_time, b_time, loss_time, opt_time, clip_time, emb_time = 0., 0., 0., 0., 0., 0. sample_time = 0. build_item_graph = True for idx in (range(n_batch)): self.model.train() self.optimizer.zero_grad() sample_t1 = time() users, pos_items, neg_items = data_generator.sample() sample_time += time() - sample_t1 ua_embeddings, ia_embeddings, image_item_embeds, text_item_embeds, fusion_embed = self.model(self.norm_adj, build_item_graph=build_item_graph) build_item_graph = False u_g_embeddings = ua_embeddings[users] pos_i_g_embeddings = ia_embeddings[pos_items] neg_i_g_embeddings = ia_embeddings[neg_items] batch_mf_loss, batch_emb_loss, batch_reg_loss = self.bpr_loss(u_g_embeddings, pos_i_g_embeddings, neg_i_g_embeddings) batch_contrastive_loss = 0 batch_contrastive_loss += self.model.batched_contrastive_loss(image_item_embeds,fusion_embed) batch_contrastive_loss += self.model.batched_contrastive_loss(text_item_embeds,fusion_embed) batch_contrastive_loss *= args.loss_ratio batch_loss = batch_mf_loss + batch_emb_loss + batch_reg_loss + batch_contrastive_loss batch_loss.backward(retain_graph=False) self.optimizer.step() loss += float(batch_loss) mf_loss += float(batch_mf_loss) emb_loss += float(batch_emb_loss) reg_loss += float(batch_reg_loss) contrastive_loss += float(batch_contrastive_loss) self.lr_scheduler.step() del ua_embeddings, ia_embeddings, u_g_embeddings, neg_i_g_embeddings, pos_i_g_embeddings if math.isnan(loss) == True: self.logger.logging('ERROR: loss is nan.') sys.exit() if (epoch + 1) % args.verbose != 0: perf_str = 'Epoch %d [%.1fs]: train==[%.5f=%.5f + %.5f + %.5f]' % ( epoch, time() - t1, loss, mf_loss, emb_loss, reg_loss) training_time_list.append(time() - t1) self.logger.logging(perf_str) continue t2 = time() users_to_test = list(data_generator.test_set.keys()) users_to_val = list(data_generator.val_set.keys()) ret = self.test(users_to_val, is_val=True) training_time_list.append(t2 - t1) t3 = time() loss_loger.append(loss) rec_loger.append(ret['recall']) pre_loger.append(ret['precision']) ndcg_loger.append(ret['ndcg']) hit_loger.append(ret['hit_ratio']) if args.verbose > 0: perf_str = 'Epoch %d [%.1fs + %.1fs]: train==[%.5f=%.5f + %.5f + %.5f], recall=[%.5f, %.5f], ' \ 'precision=[%.5f, %.5f], hit=[%.5f, %.5f], ndcg=[%.5f, %.5f]' % \ (epoch, t2 - t1, t3 - t2, loss, mf_loss, emb_loss, reg_loss, ret['recall'][0], ret['recall'][-1], ret['precision'][0], ret['precision'][-1], ret['hit_ratio'][0], ret['hit_ratio'][-1], ret['ndcg'][0], ret['ndcg'][-1]) self.logger.logging(perf_str) if ret['recall'][1] > best_recall: best_recall = ret['recall'][1] test_ret = self.test(users_to_test, is_val=False) self.logger.logging("Test_Recall@%d: %.5f" % (eval(args.Ks)[1], test_ret['recall'][1])) stopping_step = 0 elif stopping_step < args.early_stopping_patience: stopping_step += 1 self.logger.logging('#####Early stopping steps: %d #####' % stopping_step) else: self.logger.logging('#####Early stop! #####') break self.logger.logging(str(test_ret)) def bpr_loss(self, users, pos_items, neg_items): pos_scores = torch.sum(torch.mul(users, pos_items), dim=1) neg_scores = torch.sum(torch.mul(users, neg_items), dim=1) regularizer = 1./2*(users**2).sum() + 1./2*(pos_items**2).sum() + 1./2*(neg_items**2).sum() regularizer = regularizer / self.batch_size maxi = F.logsigmoid(pos_scores - neg_scores) mf_loss = -torch.mean(maxi) emb_loss = self.decay * regularizer reg_loss = 0.0 return mf_loss, emb_loss, reg_loss def sparse_mx_to_torch_sparse_tensor(self, sparse_mx): """Convert a scipy sparse matrix to a torch sparse tensor.""" sparse_mx = sparse_mx.tocoo().astype(np.float32) indices = torch.from_numpy( np.vstack((sparse_mx.row, sparse_mx.col)).astype(np.int64)) values = torch.from_numpy(sparse_mx.data) shape = torch.Size(sparse_mx.shape) return torch.sparse.FloatTensor(indices, values, shape) def set_seed(seed): np.random.seed(seed) random.seed(seed) torch.manual_seed(seed) # cpu torch.cuda.manual_seed_all(seed) # gpu if __name__ == '__main__': os.environ["CUDA_VISIBLE_DEVICES"] = str(args.gpu_id) set_seed(args.seed) config = dict() config['n_users'] = data_generator.n_users config['n_items'] = data_generator.n_items plain_adj, norm_adj, mean_adj = data_generator.get_adj_mat() config['norm_adj'] = norm_adj trainer = Trainer(data_config=config) trainer.train()
8,593
39.92381
158
py
MICRO
MICRO-main/codes/Models.py
import os import numpy as np from time import time import torch import torch.nn as nn import torch.nn.functional as F from utility.parser import parse_args from utility.norm import build_sim, build_knn_normalized_graph args = parse_args() class MICRO(nn.Module): def __init__(self, n_users, n_items, embedding_dim, weight_size, dropout_list, image_feats, text_feats): super().__init__() self.n_users = n_users self.n_items = n_items self.embedding_dim = embedding_dim self.weight_size = weight_size self.n_ui_layers = len(self.weight_size) self.weight_size = [self.embedding_dim] + self.weight_size self.user_embedding = nn.Embedding(n_users, self.embedding_dim) self.item_id_embedding = nn.Embedding(n_items, self.embedding_dim) nn.init.xavier_uniform_(self.user_embedding.weight) nn.init.xavier_uniform_(self.item_id_embedding.weight) if args.cf_model == 'ngcf': self.GC_Linear_list = nn.ModuleList() self.Bi_Linear_list = nn.ModuleList() self.dropout_list = nn.ModuleList() for i in range(self.n_ui_layers): self.GC_Linear_list.append(nn.Linear(self.weight_size[i], self.weight_size[i+1])) self.Bi_Linear_list.append(nn.Linear(self.weight_size[i], self.weight_size[i+1])) self.dropout_list.append(nn.Dropout(dropout_list[i])) self.image_embedding = nn.Embedding.from_pretrained(torch.Tensor(image_feats), freeze=False) self.text_embedding = nn.Embedding.from_pretrained(torch.Tensor(text_feats), freeze=False) image_adj = build_sim(self.image_embedding.weight.detach()) image_adj = build_knn_normalized_graph(image_adj, topk=args.topk, is_sparse=args.sparse, norm_type=args.norm_type) text_adj = build_sim(self.text_embedding.weight.detach()) text_adj = build_knn_normalized_graph(text_adj, topk=args.topk, is_sparse=args.sparse, norm_type=args.norm_type) self.text_original_adj = text_adj.cuda() self.image_original_adj = image_adj.cuda() self.image_trs = nn.Linear(image_feats.shape[1], args.embed_size) self.text_trs = nn.Linear(text_feats.shape[1], args.embed_size) self.softmax = nn.Softmax(dim=-1) self.query = nn.Sequential( nn.Linear(self.embedding_dim, self.embedding_dim), nn.Tanh(), nn.Linear(self.embedding_dim, 1, bias=False) ) self.tau = 0.5 def mm(self, x, y): if args.sparse: return torch.sparse.mm(x, y) else: return torch.mm(x, y) def sim(self, z1, z2): z1 = F.normalize(z1) z2 = F.normalize(z2) return torch.mm(z1, z2.t()) def batched_contrastive_loss(self, z1, z2, batch_size=4096): device = z1.device num_nodes = z1.size(0) num_batches = (num_nodes - 1) // batch_size + 1 f = lambda x: torch.exp(x / self.tau) indices = torch.arange(0, num_nodes).to(device) losses = [] for i in range(num_batches): mask = indices[i * batch_size:(i + 1) * batch_size] refl_sim = f(self.sim(z1[mask], z1)) # [B, N] between_sim = f(self.sim(z1[mask], z2)) # [B, N] losses.append(-torch.log( between_sim[:, i * batch_size:(i + 1) * batch_size].diag() / (refl_sim.sum(1) + between_sim.sum(1) - refl_sim[:, i * batch_size:(i + 1) * batch_size].diag()))) loss_vec = torch.cat(losses) return loss_vec.mean() def forward(self, adj, build_item_graph=False): image_feats = self.image_trs(self.image_embedding.weight) text_feats = self.text_trs(self.text_embedding.weight) if build_item_graph: self.image_adj = build_sim(image_feats) self.image_adj = build_knn_normalized_graph(self.image_adj, topk=args.topk, is_sparse=args.sparse, norm_type=args.norm_type) self.image_adj = (1 - args.lambda_coeff) * self.image_adj + args.lambda_coeff * self.image_original_adj self.text_adj = build_sim(text_feats) self.text_adj = build_knn_normalized_graph(self.text_adj, topk=args.topk, is_sparse=args.sparse, norm_type=args.norm_type) self.text_adj = (1 - args.lambda_coeff) * self.text_adj + args.lambda_coeff * self.text_original_adj else: self.image_adj = self.image_adj.detach() self.text_adj = self.text_adj.detach() image_item_embeds = self.item_id_embedding.weight text_item_embeds = self.item_id_embedding.weight for i in range(args.layers): image_item_embeds = self.mm(self.image_adj, image_item_embeds) for i in range(args.layers): text_item_embeds = self.mm(self.text_adj, text_item_embeds) att = torch.cat([self.query(image_item_embeds), self.query(text_item_embeds)], dim=-1) weight = self.softmax(att) h = weight[:, 0].unsqueeze(dim=1) * image_item_embeds + weight[:, 1].unsqueeze(dim=1) * text_item_embeds if args.cf_model == 'ngcf': ego_embeddings = torch.cat((self.user_embedding.weight, self.item_id_embedding.weight), dim=0) all_embeddings = [ego_embeddings] for i in range(self.n_ui_layers): side_embeddings = torch.sparse.mm(adj, ego_embeddings) sum_embeddings = F.leaky_relu(self.GC_Linear_list[i](side_embeddings)) bi_embeddings = torch.mul(ego_embeddings, side_embeddings) bi_embeddings = F.leaky_relu(self.Bi_Linear_list[i](bi_embeddings)) ego_embeddings = sum_embeddings + bi_embeddings ego_embeddings = self.dropout_list[i](ego_embeddings) norm_embeddings = F.normalize(ego_embeddings, p=2, dim=1) all_embeddings += [norm_embeddings] all_embeddings = torch.stack(all_embeddings, dim=1) all_embeddings = all_embeddings.mean(dim=1, keepdim=False) u_g_embeddings, i_g_embeddings = torch.split(all_embeddings, [self.n_users, self.n_items], dim=0) i_g_embeddings = i_g_embeddings + F.normalize(h, p=2, dim=1) return u_g_embeddings, i_g_embeddings, image_item_embeds, text_item_embeds, h elif args.cf_model == 'lightgcn': ego_embeddings = torch.cat((self.user_embedding.weight, self.item_id_embedding.weight), dim=0) all_embeddings = [ego_embeddings] for i in range(self.n_ui_layers): side_embeddings = torch.sparse.mm(adj, ego_embeddings) ego_embeddings = side_embeddings all_embeddings += [ego_embeddings] all_embeddings = torch.stack(all_embeddings, dim=1) all_embeddings = all_embeddings.mean(dim=1, keepdim=False) u_g_embeddings, i_g_embeddings = torch.split(all_embeddings, [self.n_users, self.n_items], dim=0) i_g_embeddings = i_g_embeddings + F.normalize(h, p=2, dim=1) return u_g_embeddings, i_g_embeddings, image_item_embeds, text_item_embeds, h elif args.cf_model == 'mf': return self.user_embedding.weight, self.item_id_embedding.weight + F.normalize(h, p=2, dim=1), image_item_embeds, text_item_embeds, h class MF(nn.Module): def __init__(self, n_users, n_items, embedding_dim, weight_size, dropout_list, image_feats=None, text_feats=None): super().__init__() self.n_users = n_users self.n_items = n_items self.embedding_dim = embedding_dim self.user_embedding = nn.Embedding(n_users, embedding_dim) self.item_embedding = nn.Embedding(n_items, embedding_dim) nn.init.xavier_uniform_(self.user_embedding.weight) nn.init.xavier_uniform_(self.item_embedding.weight) def forward(self, adj, build_item_graph=False): return self.user_embedding.weight, self.item_embedding.weight class NGCF(nn.Module): def __init__(self, n_users, n_items, embedding_dim, weight_size, dropout_list, image_feats=None, text_feats=None): super().__init__() self.n_users = n_users self.n_items = n_items self.embedding_dim = embedding_dim self.weight_size = weight_size self.n_ui_layers = len(self.weight_size) self.dropout_list = nn.ModuleList() self.GC_Linear_list = nn.ModuleList() self.Bi_Linear_list = nn.ModuleList() self.weight_size = [self.embedding_dim] + self.weight_size for i in range(self.n_ui_layers): self.GC_Linear_list.append(nn.Linear(self.weight_size[i], self.weight_size[i+1])) self.Bi_Linear_list.append(nn.Linear(self.weight_size[i], self.weight_size[i+1])) self.dropout_list.append(nn.Dropout(dropout_list[i])) self.user_embedding = nn.Embedding(n_users, embedding_dim) self.item_id_embedding = nn.Embedding(n_items, embedding_dim) nn.init.xavier_uniform_(self.user_embedding.weight) nn.init.xavier_uniform_(self.item_id_embedding.weight) def forward(self, adj, build_item_graph): ego_embeddings = torch.cat((self.user_embedding.weight, self.item_id_embedding.weight), dim=0) all_embeddings = [ego_embeddings] for i in range(self.n_ui_layers): side_embeddings = torch.sparse.mm(adj, ego_embeddings) sum_embeddings = F.leaky_relu(self.GC_Linear_list[i](side_embeddings)) bi_embeddings = torch.mul(ego_embeddings, side_embeddings) bi_embeddings = F.leaky_relu(self.Bi_Linear_list[i](bi_embeddings)) ego_embeddings = sum_embeddings + bi_embeddings ego_embeddings = self.dropout_list[i](ego_embeddings) norm_embeddings = F.normalize(ego_embeddings, p=2, dim=1) all_embeddings += [norm_embeddings] all_embeddings = torch.cat(all_embeddings, dim=1) u_g_embeddings, i_g_embeddings = torch.split(all_embeddings, [self.n_users, self.n_items], dim=0) return u_g_embeddings, i_g_embeddings class LightGCN(nn.Module): def __init__(self, n_users, n_items, embedding_dim, weight_size, dropout_list, image_feats=None, text_feats=None): super().__init__() self.n_users = n_users self.n_items = n_items self.embedding_dim = embedding_dim self.n_ui_layers = len(weight_size) self.user_embedding = nn.Embedding(n_users, embedding_dim) self.item_id_embedding = nn.Embedding(n_items, embedding_dim) nn.init.xavier_uniform_(self.user_embedding.weight) nn.init.xavier_uniform_(self.item_id_embedding.weight) def forward(self, adj, build_item_graph): ego_embeddings = torch.cat((self.user_embedding.weight, self.item_id_embedding.weight), dim=0) all_embeddings = [ego_embeddings] for i in range(self.n_ui_layers): side_embeddings = torch.sparse.mm(adj, ego_embeddings) ego_embeddings = side_embeddings all_embeddings += [ego_embeddings] all_embeddings = torch.stack(all_embeddings, dim=1) all_embeddings = all_embeddings.mean(dim=1, keepdim=False) u_g_embeddings, i_g_embeddings = torch.split(all_embeddings, [self.n_users, self.n_items], dim=0) return u_g_embeddings, i_g_embeddings
11,458
46.156379
149
py
MICRO
MICRO-main/codes/utility/norm.py
import torch def build_sim(context): context_norm = context.div(torch.norm(context, p=2, dim=-1, keepdim=True)) sim = torch.mm(context_norm, context_norm.transpose(1, 0)) return sim def build_knn_normalized_graph(adj, topk, is_sparse, norm_type): device = adj.device knn_val, knn_ind = torch.topk(adj, topk, dim=-1) if is_sparse: tuple_list = [[row, int(col)] for row in range(len(knn_ind)) for col in knn_ind[row]] row = [i[0] for i in tuple_list] col = [i[1] for i in tuple_list] i = torch.LongTensor([row, col]).to(device) v = knn_val.flatten() edge_index, edge_weight = get_sparse_laplacian(i, v, normalization=norm_type, num_nodes=adj.shape[0]) return torch.sparse_coo_tensor(edge_index, edge_weight, adj.shape) else: weighted_adjacency_matrix = (torch.zeros_like(adj)).scatter_(-1, knn_ind, knn_val) return get_dense_laplacian(weighted_adjacency_matrix, normalization=norm_type) def get_sparse_laplacian(edge_index, edge_weight, num_nodes, normalization='none'): from torch_scatter import scatter_add row, col = edge_index[0], edge_index[1] deg = scatter_add(edge_weight, row, dim=0, dim_size=num_nodes) if normalization == 'sym': deg_inv_sqrt = deg.pow_(-0.5) deg_inv_sqrt.masked_fill_(deg_inv_sqrt == float('inf'), 0) edge_weight = deg_inv_sqrt[row] * edge_weight * deg_inv_sqrt[col] elif normalization == 'rw': deg_inv = 1.0 / deg deg_inv.masked_fill_(deg_inv == float('inf'), 0) edge_weight = deg_inv[row] * edge_weight return edge_index, edge_weight def get_dense_laplacian(adj, normalization='none'): if normalization == 'sym': rowsum = torch.sum(adj, -1) d_inv_sqrt = torch.pow(rowsum, -0.5) d_inv_sqrt[torch.isinf(d_inv_sqrt)] = 0. d_mat_inv_sqrt = torch.diagflat(d_inv_sqrt) L_norm = torch.mm(torch.mm(d_mat_inv_sqrt, adj), d_mat_inv_sqrt) elif normalization == 'rw': rowsum = torch.sum(adj, -1) d_inv = torch.pow(rowsum, -1) d_inv[torch.isinf(d_inv)] = 0. d_mat_inv = torch.diagflat(d_inv) L_norm = torch.mm(d_mat_inv, adj) elif normalization == 'none': L_norm = adj return L_norm
2,279
40.454545
109
py
MICRO
MICRO-main/codes/utility/batch_test.py
import utility.metrics as metrics from utility.parser import parse_args from utility.load_data import Data import multiprocessing import heapq import torch import pickle import numpy as np from time import time cores = multiprocessing.cpu_count() // 5 args = parse_args() Ks = eval(args.Ks) data_generator = Data(path=args.data_path + args.dataset, batch_size=args.batch_size) USR_NUM, ITEM_NUM = data_generator.n_users, data_generator.n_items N_TRAIN, N_TEST = data_generator.n_train, data_generator.n_test BATCH_SIZE = args.batch_size def ranklist_by_heapq(user_pos_test, test_items, rating, Ks): item_score = {} for i in test_items: item_score[i] = rating[i] K_max = max(Ks) K_max_item_score = heapq.nlargest(K_max, item_score, key=item_score.get) r = [] for i in K_max_item_score: if i in user_pos_test: r.append(1) else: r.append(0) auc = 0. return r, auc def get_auc(item_score, user_pos_test): item_score = sorted(item_score.items(), key=lambda kv: kv[1]) item_score.reverse() item_sort = [x[0] for x in item_score] posterior = [x[1] for x in item_score] r = [] for i in item_sort: if i in user_pos_test: r.append(1) else: r.append(0) auc = metrics.auc(ground_truth=r, prediction=posterior) return auc def ranklist_by_sorted(user_pos_test, test_items, rating, Ks): item_score = {} for i in test_items: item_score[i] = rating[i] K_max = max(Ks) K_max_item_score = heapq.nlargest(K_max, item_score, key=item_score.get) r = [] for i in K_max_item_score: if i in user_pos_test: r.append(1) else: r.append(0) auc = get_auc(item_score, user_pos_test) return r, auc def get_performance(user_pos_test, r, auc, Ks): precision, recall, ndcg, hit_ratio = [], [], [], [] for K in Ks: precision.append(metrics.precision_at_k(r, K)) recall.append(metrics.recall_at_k(r, K, len(user_pos_test))) ndcg.append(metrics.ndcg_at_k(r, K)) hit_ratio.append(metrics.hit_at_k(r, K)) return {'recall': np.array(recall), 'precision': np.array(precision), 'ndcg': np.array(ndcg), 'hit_ratio': np.array(hit_ratio), 'auc': auc} def test_one_user(x): # user u's ratings for user u is_val = x[-1] rating = x[0] #uid u = x[1] #user u's items in the training set try: training_items = data_generator.train_items[u] except Exception: training_items = [] #user u's items in the test set if is_val: user_pos_test = data_generator.val_set[u] else: user_pos_test = data_generator.test_set[u] all_items = set(range(ITEM_NUM)) test_items = list(all_items - set(training_items)) if args.test_flag == 'part': r, auc = ranklist_by_heapq(user_pos_test, test_items, rating, Ks) else: r, auc = ranklist_by_sorted(user_pos_test, test_items, rating, Ks) return get_performance(user_pos_test, r, auc, Ks) def test_torch(ua_embeddings, ia_embeddings, users_to_test, is_val, drop_flag=False, batch_test_flag=False): result = {'precision': np.zeros(len(Ks)), 'recall': np.zeros(len(Ks)), 'ndcg': np.zeros(len(Ks)), 'hit_ratio': np.zeros(len(Ks)), 'auc': 0.} pool = multiprocessing.Pool(cores) u_batch_size = BATCH_SIZE * 2 i_batch_size = BATCH_SIZE test_users = users_to_test n_test_users = len(test_users) n_user_batchs = n_test_users // u_batch_size + 1 count = 0 for u_batch_id in range(n_user_batchs): start = u_batch_id * u_batch_size end = (u_batch_id + 1) * u_batch_size user_batch = test_users[start: end] if batch_test_flag: n_item_batchs = ITEM_NUM // i_batch_size + 1 rate_batch = np.zeros(shape=(len(user_batch), ITEM_NUM)) i_count = 0 for i_batch_id in range(n_item_batchs): i_start = i_batch_id * i_batch_size i_end = min((i_batch_id + 1) * i_batch_size, ITEM_NUM) item_batch = range(i_start, i_end) u_g_embeddings = ua_embeddings[user_batch] i_g_embeddings = ia_embeddings[item_batch] i_rate_batch = torch.matmul(u_g_embeddings, torch.transpose(i_g_embeddings, 0, 1)) rate_batch[:, i_start: i_end] = i_rate_batch i_count += i_rate_batch.shape[1] assert i_count == ITEM_NUM else: item_batch = range(ITEM_NUM) u_g_embeddings = ua_embeddings[user_batch] i_g_embeddings = ia_embeddings[item_batch] rate_batch = torch.matmul(u_g_embeddings, torch.transpose(i_g_embeddings, 0, 1)) rate_batch = rate_batch.detach().cpu().numpy() user_batch_rating_uid = zip(rate_batch, user_batch, [is_val] * len(user_batch)) batch_result = pool.map(test_one_user, user_batch_rating_uid) count += len(batch_result) for re in batch_result: result['precision'] += re['precision'] / n_test_users result['recall'] += re['recall'] / n_test_users result['ndcg'] += re['ndcg'] / n_test_users result['hit_ratio'] += re['hit_ratio'] / n_test_users result['auc'] += re['auc'] / n_test_users assert count == n_test_users pool.close() return result
5,454
31.088235
108
py
reformer-pytorch
reformer-pytorch-master/setup.py
from setuptools import setup, find_packages setup( name = 'reformer_pytorch', packages = find_packages(exclude=['examples', 'pretraining']), version = '1.4.4', license='MIT', description = 'Reformer, the Efficient Transformer, Pytorch', author = 'Phil Wang', author_email = 'lucidrains@gmail.com', url = 'https://github.com/lucidrains/reformer-pytorch', keywords = ['transformers', 'attention', 'artificial intelligence'], install_requires=[ 'axial-positional-embedding>=0.1.0', 'einops', 'local-attention', 'product-key-memory', 'torch' ], classifiers=[ 'Development Status :: 4 - Beta', 'Intended Audience :: Developers', 'Topic :: Scientific/Engineering :: Artificial Intelligence', 'License :: OSI Approved :: MIT License', 'Programming Language :: Python :: 3.6', ], )
841
29.071429
70
py