cleanup to rainforest competition

.gitignore

@@ -1,3 +1,6 @@
+nbs/data
+nbs/models
+nbs/preds
 *.bak
 .gitattributes
 .last_checked

Makefile

@@ -0,0 +1,35 @@
+.ONESHELL:
+SHELL := /bin/bash
+SRC = $(wildcard nbs/*.ipynb)
+
+all: dl_pipeline docs
+
+dl_pipeline: $(SRC)
+	nbdev_build_lib
+	touch dl_pipeline
+
+sync:
+	nbdev_update_lib
+
+docs_serve: docs
+	cd docs && bundle exec jekyll serve
+
+docs: $(SRC)
+	nbdev_build_docs
+	touch docs
+
+test:
+	nbdev_test_nbs
+
+release: pypi
+	nbdev_conda_package
+	nbdev_bump_version
+
+pypi: dist
+	twine upload --repository pypi dist/*
+
+dist: clean
+	python setup.py sdist bdist_wheel
+
+clean:
+	rm -rf dist

README.md

@@ -1,4 +1,24 @@
-# nbdev template
+# Project name here
+> Summary description here.
+
+
+This file will become your README and also the index of your documentation.
+
+## Install
+
+`pip install your_project_name`
+
+## How to use
+
+Fill me in please! Don't forget code examples:
+
+```
+1+1
+```
+
+
+
+
+    2
 
-Use this template to more easily create your nbdev project.
 

dl_pipeline/__init__.py

@@ -0,0 +1 @@
+__version__ = "0.0.1"

dl_pipeline/_nbdev.py

@@ -0,0 +1,122 @@
+# AUTOGENERATED BY NBDEV! DO NOT EDIT!
+
+__all__ = ["index", "modules", "custom_doc_links", "git_url"]
+
+index = {"seed_everything": "00_core.ipynb",
+         "after_loss": "00vision_losses.ipynb",
+         "cross_entropy": "00vision_losses.ipynb",
+         "cross_entropy_mixup": "00vision_losses.ipynb",
+         "binary_cross_entropy": "00vision_losses.ipynb",
+         "binary_cross_entropy_mixup": "00vision_losses.ipynb",
+         "binary_cross_entropy_scaled_mixup": "00vision_losses.ipynb",
+         "focal_loss": "00vision_losses.ipynb",
+         "get_loss": "00vision_losses.ipynb",
+         "gem": "00vision_models.ipynb",
+         "GeM": "00vision_models.ipynb",
+         "AdaptiveConcatPool2d_GeM": "00vision_models.ipynb",
+         "MobileNetV2": "00vision_models.ipynb",
+         "ResNet_": "00vision_models.ipynb",
+         "ResNet18_swsl": "00vision_models.ipynb",
+         "ResNet50_swsl": "00vision_models.ipynb",
+         "ResNet50_32x4d_swsl": "00vision_models.ipynb",
+         "xResNet_": "00vision_models.ipynb",
+         "xResNet50_ssa": "00vision_models.ipynb",
+         "ResNeSt_": "00vision_models.ipynb",
+         "ResNeSt50": "00vision_models.ipynb",
+         "ResNeSt101": "00vision_models.ipynb",
+         "ResNeSt200": "00vision_models.ipynb",
+         "ResNeSt269": "00vision_models.ipynb",
+         "ResNeSt50_fast_1s1x64d": "00vision_models.ipynb",
+         "ResNeSt50_fast_1s2x40d": "00vision_models.ipynb",
+         "ResNeSt50_fast_1s4x24d": "00vision_models.ipynb",
+         "ResNeSt50_fast_2s1x64d": "00vision_models.ipynb",
+         "ResNeSt50_fast_2s2x40d": "00vision_models.ipynb",
+         "ResNeSt50_fast_4s1x64d": "00vision_models.ipynb",
+         "ResNeSt50_fast_4s2x40d": "00vision_models.ipynb",
+         "DenseNet_": "00vision_models.ipynb",
+         "DenseNet121": "00vision_models.ipynb",
+         "DenseNet169": "00vision_models.ipynb",
+         "DenseNet201": "00vision_models.ipynb",
+         "DenseNet161": "00vision_models.ipynb",
+         "DenseNetBlur121": "00vision_models.ipynb",
+         "EfficientNet_": "00vision_models.ipynb",
+         "EfficientNetB0": "00vision_models.ipynb",
+         "EfficientNetB1": "00vision_models.ipynb",
+         "EfficientNetB2": "00vision_models.ipynb",
+         "EfficientNetB3": "00vision_models.ipynb",
+         "EfficientNetB4": "00vision_models.ipynb",
+         "EfficientNetB5": "00vision_models.ipynb",
+         "EfficientNetB6": "00vision_models.ipynb",
+         "EfficientNetB7": "00vision_models.ipynb",
+         "Head": "00vision_models.ipynb",
+         "EmbResNeSt_": "00vision_models.ipynb",
+         "EmbResNeSt50": "00vision_models.ipynb",
+         "get_model": "00vision_models.ipynb",
+         "SampleEpisode": "00vision_triplet.ipynb",
+         "compute_distance_matrix": "00vision_triplet.ipynb",
+         "EpisodeDataLoader": "00vision_triplet.ipynb",
+         "get_preds": "kaggle_rfcx-species-audio-detection.ipynb",
+         "distance": "00vision_triplet.ipynb",
+         "remove_duplicates": "00vision_triplet.ipynb",
+         "map5": "00vision_triplet.ipynb",
+         "accuracy": "01audio_util.ipynb",
+         "AddGaussianSNR": "01audio_augmentations.ipynb",
+         "ClippingDistortion": "01audio_augmentations.ipynb",
+         "FrequencyMask": "01audio_augmentations.ipynb",
+         "TimeMask": "01audio_augmentations.ipynb",
+         "Gain": "01audio_augmentations.ipynb",
+         "PitchShift": "01audio_augmentations.ipynb",
+         "Shift": "01audio_augmentations.ipynb",
+         "TimeStretch": "01audio_augmentations.ipynb",
+         "MelSpectrogram": "01audio_augmentations.ipynb",
+         "SAMPLE_FILE": "01audio_core.ipynb",
+         "TensorAudio": "01audio_core.ipynb",
+         "TensorAudioLabel": "01audio_core.ipynb",
+         "load_npy": "01audio_core.ipynb",
+         "sample_file": "01audio_core.ipynb",
+         "melspectrogram": "01audio_core.ipynb",
+         "show_sample": "01audio_core.ipynb",
+         "audio2npy": "01audio_core.ipynb",
+         "Datasets": "01audio_dataset.ipynb",
+         "DataLoader": "01audio_dataset.ipynb",
+         "DataLoaders": "01audio_dataset.ipynb",
+         "RenameColumns": "01audio_dataset.ipynb",
+         "load_dataframe": "01audio_dataset.ipynb",
+         "group_labels": "01audio_dataset.ipynb",
+         "time2pix_image": "01audio_dataset.ipynb",
+         "time2pix_wave": "01audio_dataset.ipynb",
+         "pix2time": "01audio_dataset.ipynb",
+         "pix2pix_image": "01audio_dataset.ipynb",
+         "time_labels": "01audio_dataset.ipynb",
+         "audio_crop": "01audio_dataset.ipynb",
+         "TilesTransform": "01audio_dataset.ipynb",
+         "reorganize_batch": "01audio_dataset.ipynb",
+         "create_dataset_item": "01audio_dataset.ipynb",
+         "apply_augmentations": "01audio_dataset.ipynb",
+         "audio_augment": "kaggle_rfcx-species-audio-detection.ipynb",
+         "show_augmentations": "01audio_dataset.ipynb",
+         "mask2category": "01audio_util.ipynb",
+         "lrap": "01audio_util.ipynb",
+         "kfold_dataframes": "01audio_util.ipynb",
+         "OneHot": "01audio_util.ipynb",
+         "MixUp": "01audio_util.ipynb",
+         "LabelSED": "01audio_util.ipynb",
+         "train": "kaggle_rfcx-species-audio-detection.ipynb",
+         "test": "kaggle_rfcx-species-audio-detection.ipynb",
+         "main": "kaggle_rfcx-species-audio-detection.ipynb"}
+
+modules = ["core.py",
+           "vision/losses.py",
+           "vision/models.py",
+           "vision/triplet.py",
+           "audio/augmentations.py",
+           "audio/core.py",
+           "audio/dataset.py",
+           "audio/util.py",
+           "kaggle/rfcx_species_audio_detection.py"]
+
+doc_url = "https://mnpinto.github.io/dl_pipeline/"
+
+git_url = "https://github.com/mnpinto/dl_pipeline/tree/master/"
+
+def custom_doc_links(name): return None

dl_pipeline/audio/all.py

@@ -0,0 +1,3 @@
+# AUTOGENERATED! DO NOT EDIT! File to edit: nbs/01audio_all.ipynb (unless otherwise specified).
+
+__all__ = []

dl_pipeline/audio/augmentations.py

@@ -0,0 +1,138 @@
+# AUTOGENERATED! DO NOT EDIT! File to edit: nbs/01audio_augmentations.ipynb (unless otherwise specified).
+
+__all__ = ['AddGaussianSNR', 'ClippingDistortion', 'FrequencyMask', 'TimeMask', 'Gain', 'PitchShift', 'Shift',
+           'TimeStretch', 'MelSpectrogram']
+
+# Cell
+import matplotlib.pyplot as plt
+import audiomentations as aug
+from nnAudio import Spectrogram
+from fastcore.all import *
+from fastai.vision.augment import RandTransform
+from fastai.vision.all import *
+from .core import *
+
+# Cell
+class AddGaussianSNR(Transform):
+    "Add Gaussian Signal-to-noise ratio (SNR) noise"
+    def __init__(self, sample_rate, min_SNR=0.001, max_SNR=1.0, p=0.5, **kwargs):
+        store_attr('min_SNR'), store_attr('max_SNR'), store_attr('p')
+        super().__init__(**kwargs)
+        self.tfm = partial(aug.AddGaussianSNR(min_SNR=min_SNR, max_SNR=max_SNR, p=p),
+                           sample_rate=sample_rate)
+    def encodes(self, wav:TensorAudio):
+        return TensorAudio(self.tfm(wav.float().numpy()))
+
+# Cell
+class ClippingDistortion(Transform):
+    "Apply clipping distortion"
+    def __init__(self, sample_rate, min_percentile_threshold=0,
+                 max_percentile_threshold=40, p=0.5, **kwargs):
+        store_attr('min_percentile_threshold'), store_attr('max_percentile_threshold')
+        store_attr('p')
+        super().__init__(**kwargs)
+        self.tfm = partial(aug.ClippingDistortion(
+            min_percentile_threshold=min_percentile_threshold,
+            max_percentile_threshold=max_percentile_threshold, p=p), sample_rate=sample_rate)
+    def encodes(self, wav:TensorAudio):
+        return TensorAudio(self.tfm(wav.float().numpy()))
+
+# Cell
+class FrequencyMask(Transform):
+    "Applies a frequency mask to a range of frequencies"
+    def __init__(self, sample_rate,min_frequency_band=0.0, max_frequency_band=0.5,
+                 p=0.5, **kwargs):
+        store_attr('min_frequency_band'), store_attr('max_frequency_band'), store_attr('p')
+        super().__init__(**kwargs)
+        self.tfm = partial(aug.FrequencyMask(min_frequency_band=min_frequency_band,
+            max_frequency_band=max_frequency_band, p=p), sample_rate=sample_rate)
+    def encodes(self, wav:TensorAudio):
+        return TensorAudio(self.tfm(wav.float().numpy()))
+
+
+# Cell
+class TimeMask(Transform):
+    "Applies a mask to a section of the audio clip"
+    def __init__(self, sample_rate, min_band_part=0.0, max_band_part=0.5, p=0.5, **kwargs):
+        store_attr('min_band_part'), store_attr('max_band_part'), store_attr('p')
+        super().__init__(**kwargs)
+        self.tfm = partial(aug.TimeMask(min_band_part=min_band_part,
+            max_band_part=max_band_part, p=p), sample_rate=sample_rate)
+    def encodes(self, wav:TensorAudio):
+        return TensorAudio(self.tfm(wav.float().numpy()))
+
+# Cell
+class Gain(Transform):
+    "Apply a random gain between 'min_gain_in_db' and 'max_gain_in_db'"
+    def __init__(self, sample_rate, min_gain_in_db=-12, max_gain_in_db=12, p=0.5, **kwargs):
+        store_attr('min_gain_in_db'), store_attr('max_gain_in_db'), store_attr('p')
+        super().__init__(**kwargs)
+        self.tfm = partial(aug.Gain(min_gain_in_db=min_gain_in_db,
+            max_gain_in_db=max_gain_in_db, p=p), sample_rate=sample_rate)
+    def encodes(self, wav:TensorAudio):
+        return TensorAudio(self.tfm(wav.float().numpy()))
+
+# Cell
+class PitchShift(Transform):
+    "Shift pitch by a random value of semitones between 'min_semitones' and 'max_semitones'"
+    def __init__(self, sample_rate, min_semitones=-4, max_semitones=4, p=0.5, **kwargs):
+        store_attr('min_semitones'), store_attr('max_semitones'), store_attr('p')
+        super().__init__(**kwargs)
+        self.tfm = partial(aug.PitchShift(min_semitones=min_semitones,
+            max_semitones=max_semitones, p=p), sample_rate=sample_rate)
+    def encodes(self, wav:TensorAudio):
+        return TensorAudio(self.tfm(wav.float().numpy()))
+
+# Cell
+class Shift(Transform):
+    "Shift pitch by a random value of semitones between 'min_semitones' and 'max_semitones'"
+    def __init__(self, sample_rate, min_fraction=-0.5, max_fraction=0.5,
+                 rollover=True, p=0.5, **kwargs):
+        store_attr('min_fraction'), store_attr('max_fraction')
+        store_attr('rollover'), store_attr('p')
+        super().__init__(**kwargs)
+        self.tfm = partial(aug.Shift(min_fraction=min_fraction, max_fraction=max_fraction,
+                                     rollover=rollover, p=p), sample_rate=sample_rate)
+    def encodes(self, wav:TensorAudio):
+        return TensorAudio(self.tfm(wav.float().numpy()))
+
+
+# Cell
+class TimeStretch(Transform):
+    "Shift pitch by a random value of semitones between 'min_semitones' and 'max_semitones'"
+    def __init__(self, sample_rate, min_rate=0.8, max_rate=1.25,
+                 leave_length_unchanged=True, p=0.5, **kwargs):
+        store_attr('min_rate'), store_attr('max_rate')
+        store_attr('leave_length_unchanged'), store_attr('p')
+        super().__init__(**kwargs)
+        self.tfm = partial(aug.TimeStretch(min_rate=min_rate, max_rate=max_rate,
+            leave_length_unchanged=leave_length_unchanged, p=p), sample_rate=sample_rate)
+    def encodes(self, wav:TensorAudio):
+        return TensorAudio(self.tfm(wav.float().numpy()))
+
+# Cell
+class MelSpectrogram(Transform):
+    "Shift pitch by a random value of semitones between 'min_semitones' and 'max_semitones'"
+    def __init__(self, sample_rate, n_mels=128, hop_length=512, eps=1e-6,
+                 normalize_spectro=True, device=torch.device("cuda:0"), **kwargs):
+        store_attr('sample_rate'), store_attr('n_mels'), store_attr('hop_length')
+        store_attr('eps')
+        super().__init__(**kwargs)
+        self.spectro = Spectrogram.MelSpectrogram(
+            sr=sample_rate, n_mels=n_mels, hop_length=hop_length,
+            verbose=False, **kwargs).to(device)
+        self.relu = nn.ReLU(inplace=True)
+        self.normalize_spectro = normalize_spectro
+        self.eps = eps
+        self.device = device
+
+    def encodes(self, x:TensorAudio):
+        with torch.no_grad():
+            d = x.device
+            x = x.to(self.device)
+            x = self.relu(self.spectro(x)).unsqueeze(1)
+            x = x.add(self.eps).log()
+            if self.normalize_spectro:
+                x = (x - x.mean((2,3))[...,None,None])/x.std((2,3))[...,None,None]
+            assert np.sum(np.isnan(x.detach().cpu().numpy())) == 0
+            return TensorImage(x.to(d))