データセットは，「そのままモデルの入力になるデータセット」として作り上げる。
「画像リストの CSVファイルの読み込み」, 「画像の整形 (変換)」で行ったことは，一回性のものであり，データセットとなって残るというものではない。
以下，データセットの構築：
1. all_image_paths.csv
画像ファイル
image0.jpg (5x3 画素)
image1.jpg (3x5 画素)
image2.jpg (4x3 画素)
が，つぎのように配置されている：
(パス)
├ category0
│ └ image0.jpg
└ category1
├ image1.jpg
└ image2.jpg |
これに対し，画像リストの CSVファイル all_image_paths.csv を作成する：
レコードの項目を，つぎの２つとする：
(1) 画像ファイルのパス
(2) ラベル (カテゴリー)
そこで all_image_paths.csv は，つぎの３行：
(パス)/category0/image0.jpg,0[改行コード(LF)]
(パス)/category1/image1.jpg,1[改行コード(LF)]
(パス)/category1/image2.jpg,1[改行コード(LF)]
python インタラクティブ・シェルで作業する。
<![CDATA[$ source [パス]/venv/bin/activate
(venv) $ python
>>> import tensorflow as tf
>>> tf.enable_eager_execution() ]]>
2. all_image_paths, all_image_labels
<![CDATA[>>> csv_name = '(パス)/all_image_paths.csv'
>>> csvfile = open(csv_name, 'r')
>>> import csv
>>> reader = csv.reader(csvfile, delimiter=',')
>>> all_image_paths = []
>>> all_image_labels = []
>>> for row in reader:
... all_image_paths.append(row[0])
... all_image_labels.append(row[1])
...
>>> csvfile.close()
>>> all_image_paths
['/home/pi/project/ai/ex/csv/category0/image0.jpg',
'/home/pi/project/ai/ex/csv/category1/image1.jpg',
'/home/pi/project/ai/ex/csv/category1/image2.jpg']
>>> all_image_labels
['0', '1', '1']
>>> ]]>
3. 関数 load_and_preprocess_image
- データの画像は，サイズがまちまちである。
モデルに合わせて，これらを定形にリサイズする。
── ここでは，5x3, 3x5, 4x3 の３つの画像を 3x3 にリサイズ。
- 画素値の範囲 [0, 255] を，モデルに合わせて，[m, n] に変換する。
── ここでは，範囲 [0, 1] に変換。
<![CDATA[>>> def raw_to_tensor(path):
... image_raw = tf.io.read_file(path)
... return tf.image.decode_image(image_raw, channels=3)
...
>>> def load_and_preprocess_image(path):
... image_tensor = raw_to_tensor(path)
... image_final = tf.image.resize(image_tensor, [3, 3])
... image_final /= 255.0
... return image_final
... ]]>
- 関数 load_and_preprocess_image による画像変換を確認：
<![CDATA[>>> import matplotlib.pyplot as plt
>>> for path in all_image_paths:
... image_tensor = raw_to_tensor(path)
... plt.imshow(image_tensor)
... plt.show()
... image_final = load_and_preprocess_image(path)
... plt.imshow(image_final)
... plt.show()
...
<matplotlib.image.AxesImage object at 0x526b1470>]]>
<![CDATA[<matplotlib.image.AxesImage object at 0x524b50d0>]]>
<![CDATA[<matplotlib.image.AxesImage object at 0x524eed70>]]>
<![CDATA[
<matplotlib.image.AxesImage object at 0x524a7f90>]]>
<![CDATA[
<matplotlib.image.AxesImage object at 0x5246eb30>]]>
<![CDATA[
<matplotlib.image.AxesImage object at 0x51f14cd0>]]>
4. path_ds, label_ds
<![CDATA[>>> path_ds = \
... tf.data.Dataset.from_tensor_slices(all_image_paths)
>>> path_ds
<DatasetV1Adapter shapes: (), types: tf.string>
>>> for i in path_ds:
... print(i)
...
tf.Tensor(b'(パス)/category0/image0.jpg', shape=(), dtype=string)
tf.Tensor(b'(パス)/category1/image1.jpg', shape=(), dtype=string)
tf.Tensor(b'(パス)/category1/image2.jpg', shape=(), dtype=string)
>>> label_ds = \
... tf.data.Dataset.from_tensor_slices(all_image_labels)
>>> label_ds
<DatasetV1Adapter shapes: (), types: tf.string>
>>> for i in label_ds:
... print(i)
...
tf.Tensor(b'0', shape=(), dtype=string)
tf.Tensor(b'1', shape=(), dtype=string)
tf.Tensor(b'1', shape=(), dtype=string)]]>
5. image_ds
<![CDATA[>>> def load_and_preprocess_image(path):
... image_raw = tf.io.read_file(path)
... image_tensor = tf.image.decode_image(image_raw, channels=3, expand_animations=False)
... image_final = tf.image.resize(image_tensor, [3, 3])
... image_final /= 255.0
... return image_final
...
>>> AUTOTUNE = tf.data.experimental.AUTOTUNE
>>> image_ds = \
... path_ds.map(load_and_preprocess_image, num_parallel_calls=AUTOTUNE)]]>
この後，つぎの WARNING メッセージが２回続く：
<![CDATA[WARNING:tensorflow:AutoGraph could not transform
<function load_and_preprocess_image at 0x620a6a98> and will run it as-is.
Please report this to the TensorFlow team.
When filing the bug, set the verbosity to 10 (on Linux,
`export AUTOGRAPH_VERBOSITY=10`) and attach the full output.
Cause: Unable to locate the source code of
<function load_and_preprocess_image at 0x620a6a98>.
Note that functions defined in certain environments,
like the interactive Python shell do not expose their source code.
If that is the case, you should to define them in a .py source file.
If you are certain the code is graph-compatible,
wrap the call using @tf.autograph.do_not_convert.
Original error: could not get source code]]>
また， def load_and_preprocess_image の定義で，
tf.image.decode_image(image_raw, channels=3, expand_animations=False)
の「, expand_animations=False」が無いと，つぎのエラーメッセｰが続く：
<![CDATA[Traceback (most recent call last):
File "<stdin>", line 2, in <module>
File "/home/pi/venv/lib/python3.7/site-packages/tensorflow_core/python/data/ops/dataset_ops.py",
line 1993, in map self, map_func, num_parallel_calls, preserve_cardinality=False))
File "/home/pi/venv/lib/python3.7/site-packages/tensorflow_core/python/data/ops/dataset_ops.py",
line 3582, in __init__ use_legacy_function=use_legacy_function)
File "/home/pi/venv/lib/python3.7/site-packages/tensorflow_core/python/data/ops/dataset_ops.py",
line 2819, in __init__ self._function = wrapper_fn._get_concrete_function_internal()
File "/home/pi/venv/lib/python3.7/site-packages/tensorflow_core/python/eager/function.py",
line 2326, in _get_concrete_function_internal *args, **kwargs)
File "/home/pi/venv/lib/python3.7/site-packages/tensorflow_core/python/eager/function.py",
ine 2320, in _get_concrete_function_internal_garbage_collected graph_function, _, _ =
self._maybe_define_function(args, kwargs)
File "/home/pi/venv/lib/python3.7/site-packages/tensorflow_core/python/eager/function.py",
line 2628, in _maybe_define_function graph_function = self._create_graph_function(args, kwargs)
File "/home/pi/venv/lib/python3.7/site-packages/tensorflow_core/python/eager/function.py",
line 2517, in _create_graph_function capture_by_value=self._capture_by_value),
File "/home/pi/venv/lib/python3.7/site-packages/tensorflow_core/python/framework/func_graph.py",
line 943, in func_graph_from_py_func func_outputs = python_func(*func_args, **func_kwargs)
File "/home/pi/venv/lib/python3.7/site-packages/tensorflow_core/python/data/ops/dataset_ops.py",
ine 2812, in wrapper_fn ret = _wrapper_helper(*args)
File "/home/pi/venv/lib/python3.7/site-packages/tensorflow_core/python/data/ops/dataset_ops.py",
line 2754, in _wrapper_helper ret = autograph.tf_convert(func, ag_ctx)(*nested_args)
File "/home/pi/venv/lib/python3.7/site-packages/tensorflow_core/python/autograph/impl/api.py",
line 237, in wrapperraise e.ag_error_metadata.to_exception(e)
ValueError: in converted code:
<stdin>:3 load_and_preprocess_image
/home/pi/venv/lib/python3.7/site-packages/tensorflow_core/python/ops/image_ops_impl.py:
1226 resize_images skip_resize_if_same=True)
/home/pi/venv/lib/python3.7/site-packages/tensorflow_core/python/ops/image_ops_impl.py:
1073 _resize_images_common raise ValueError('\'images\' contains no shape.')
ValueError: 'images' contains no shape.]]>
- image_ds ができていることを確認：
<![CDATA[>>> for i in image_ds:
... print(i)
...
tf.Tensor(
[[[0.3098039 0.28235292 0.25098038]
[0.5424836 0.5555555 0.52941173]
[0.4339869 0.45359474 0.43006533]]
[[0.29019606 0.27843136 0.2117647 ]
[0.49281043 0.4575163 0.2980392 ]
[0.40653595 0.32418302 0.20261437]]
[[0.2588235 0.23137254 0.12156862]
[0.496732 0.38692808 0.20392156]
[0.42614383 0.33594772 0.17124183]]], shape=(3, 3, 3), dtype=float32)
tf.Tensor(
[[[0.7333333 0.772549 0.8705882 ]
[0.69019604 0.7098039 0.7843137 ]
[0.6823529 0.7215686 0.7294117 ]]
[[0.4954248 0.620915 0.77385616]
[0.4954248 0.5464052 0.669281 ]
[0.44183004 0.448366 0.47843134]]
[[0.5581699 0.66405225 0.8405228 ]
[0.5973856 0.67581695 0.7816993 ]
[0.5437909 0.53594774 0.54771245]]], shape=(3, 3, 3), dtype=float32)
tf.Tensor(
[[[0.4705882 0.36470586 0.18039215]
[0.496732 0.38039213 0.21437907]
[0.3215686 0.2732026 0.14640522]]
[[0.41568625 0.3490196 0.23137254]
[0.4980392 0.45620912 0.3215686 ]
[0.34640518 0.3307189 0.2457516 ]]
[[0.45490193 0.47450978 0.45098037]
[0.55032676 0.55163395 0.51241827]
[0.40130717 0.38169932 0.34248364]]], shape=(3, 3, 3), dtype=float32)]]>
- image_ds の画像を表示：
<![CDATA[>>> import matplotlib.pyplot as plt
>>> for image in image_ds:
... plt.imshow(image)
... plt.show()
...
<matplotlib.image.AxesImage object at 0x5b765210>]]>
<![CDATA[<matplotlib.image.AxesImage object at 0x50d748b0>]]>
<![CDATA[<matplotlib.image.AxesImage object at 0x693255b0>]]>
6. image_label_ds
<![CDATA[>>> image_label_ds \
... = tf.data.Dataset.zip((image_ds, label_ds))
>>> image_label_ds
<DatasetV1Adapter shapes: ((3, 3, 3), ()), types: (tf.float32, tf.string)>
>>> for i in image_label_ds:
... print(i)
...
(<tf.Tensor: shape=(3, 3, 3), dtype=float32, numpy=
array([[[0.3098039 , 0.28235292, 0.25098038],
[0.5424836 , 0.5555555 , 0.52941173],
[0.4339869 , 0.45359474, 0.43006533]],
[[0.29019606, 0.27843136, 0.2117647 ],
[0.49281043, 0.4575163 , 0.2980392 ],
[0.40653595, 0.32418302, 0.20261437]],
[[0.2588235 , 0.23137254, 0.12156862],
[0.496732 , 0.38692808, 0.20392156],
[0.42614383, 0.33594772, 0.17124183]]], dtype=float32)>,
<tf.Tensor: shape=(), dtype=string, numpy=b'0'>)
(<tf.Tensor: shape=(3, 3, 3), dtype=float32, numpy=
array([[[0.7333333 , 0.772549 , 0.8705882 ],
[0.69019604, 0.7098039 , 0.7843137 ],
[0.6823529 , 0.7215686 , 0.7294117 ]],
[[0.4954248 , 0.620915 , 0.77385616],
[0.4954248 , 0.5464052 , 0.669281 ],
[0.44183004, 0.448366 , 0.47843134]],
[[0.5581699 , 0.66405225, 0.8405228 ],
[0.5973856 , 0.67581695, 0.7816993 ],
[0.5437909 , 0.53594774, 0.54771245]]], dtype=float32)>,
<tf.Tensor: shape=(), dtype=string, numpy=b'1'>)
(<tf.Tensor: shape=(3, 3, 3), dtype=float32, numpy=
array([[[0.4705882 , 0.36470586, 0.18039215],
[0.496732 , 0.38039213, 0.21437907],
[0.3215686 , 0.2732026 , 0.14640522]],
[[0.41568625, 0.3490196 , 0.23137254],
[0.4980392 , 0.45620912, 0.3215686 ],
[0.34640518, 0.3307189 , 0.2457516 ]],
[[0.45490193, 0.47450978, 0.45098037],
[0.55032676, 0.55163395, 0.51241827],
[0.40130717, 0.38169932, 0.34248364]]], dtype=float32)>,
<tf.Tensor: shape=(), dtype=string, numpy=b'1'>)]]>
7. shuffle
<![CDATA[>>> image_count = len(all_image_paths)
>>> ds = image_label_ds.shuffle(buffer_size=image_count)
>>> ds
<DatasetV1Adapter shapes: ((3, 3, ?), ()), types: (tf.float32, tf.string)>
>>> for i in ds:
... print(i)
...
(<tf.Tensor: shape=(3, 3, 3), dtype=float32, numpy=
array([[[0.4705882 , 0.36470586, 0.18039215],
[0.496732 , 0.38039213, 0.21437907],
[0.3215686 , 0.2732026 , 0.14640522]],
[[0.41568625, 0.3490196 , 0.23137254],
[0.4980392 , 0.45620912, 0.3215686 ],
[0.34640518, 0.3307189 , 0.2457516 ]],
[[0.45490193, 0.47450978, 0.45098037],
[0.55032676, 0.55163395, 0.51241827],
[0.40130717, 0.38169932, 0.34248364]]], dtype=float32)>,
<tf.Tensor: shape=(), dtype=string, numpy=b'1'>)
(<tf.Tensor: shape=(3, 3, 3), dtype=float32, numpy=
array([[[0.3098039 , 0.28235292, 0.25098038],
[0.5424836 , 0.5555555 , 0.52941173],
[0.4339869 , 0.45359474, 0.43006533]],
[[0.29019606, 0.27843136, 0.2117647 ],
[0.49281043, 0.4575163 , 0.2980392 ],
[0.40653595, 0.32418302, 0.20261437]],
[[0.2588235 , 0.23137254, 0.12156862],
[0.496732 , 0.38692808, 0.20392156],
[0.42614383, 0.33594772, 0.17124183]]], dtype=float32)>,
<tf.Tensor: shape=(), dtype=string, numpy=b'0'>)
(<tf.Tensor: shape=(3, 3, 3), dtype=float32, numpy=
array([[[0.7333333 , 0.772549 , 0.8705882 ],
[0.69019604, 0.7098039 , 0.7843137 ],
[0.6823529 , 0.7215686 , 0.7294117 ]],
[[0.4954248 , 0.620915 , 0.77385616],
[0.4954248 , 0.5464052 , 0.669281 ],
[0.44183004, 0.448366 , 0.47843134]],
[[0.5581699 , 0.66405225, 0.8405228 ],
[0.5973856 , 0.67581695, 0.7816993 ],
[0.5437909 , 0.53594774, 0.54771245]]], dtype=float32)>,
<tf.Tensor: shape=(), dtype=string, numpy=b'1'>)]]>
8. batch
モデル (NN) に学習用画像を喰わせるときは，BATCH_SIZE 枚ずつ喰わせる。
<![CDATA[>>> BATCH_SIZE = 2
>>> ds = ds.batch(BATCH_SIZE)
>>>> ds
<DatasetV1Adapter shapes: ((?, 3, 3, ?), (?,)), types: (tf.float32, tf.string)>
>>> for i in ds:
... print(i)
...
(<tf.Tensor: shape=(2, 3, 3, 3), dtype=float32, numpy=
array([[[[0.4705882 , 0.36470586, 0.18039215],
[0.496732 , 0.38039213, 0.21437907],
[0.3215686 , 0.2732026 , 0.14640522]],
[[0.41568625, 0.3490196 , 0.23137254],
[0.4980392 , 0.45620912, 0.3215686 ],
[0.34640518, 0.3307189 , 0.2457516 ]],
[[0.45490193, 0.47450978, 0.45098037],
[0.55032676, 0.55163395, 0.51241827],
[0.40130717, 0.38169932, 0.34248364]]],
[[[0.3098039 , 0.28235292, 0.25098038],
[0.5424836 , 0.5555555 , 0.52941173],
[0.4339869 , 0.45359474, 0.43006533]],
[[0.29019606, 0.27843136, 0.2117647 ],
[0.49281043, 0.4575163 , 0.2980392 ],
[0.40653595, 0.32418302, 0.20261437]],
[[0.2588235 , 0.23137254, 0.12156862],
[0.496732 , 0.38692808, 0.20392156],
[0.42614383, 0.33594772, 0.17124183]]]], dtype=float32)>,
<tf.Tensor: shape=(2,), dtype=string, numpy=array([b'1', b'0'], dtype=object)>)
(<tf.Tensor: shape=(1, 3, 3, 3), dtype=float32, numpy=
array([[[[0.7333333 , 0.772549 , 0.8705882 ],
[0.69019604, 0.7098039 , 0.7843137 ],
[0.6823529 , 0.7215686 , 0.7294117 ]],
[[0.4954248 , 0.620915 , 0.77385616],
[0.4954248 , 0.5464052 , 0.669281 ],
[0.44183004, 0.448366 , 0.47843134]],
[[0.5581699 , 0.66405225, 0.8405228 ],
[0.5973856 , 0.67581695, 0.7816993 ],
[0.5437909 , 0.53594774, 0.54771245]]]], dtype=float32)>,
<tf.Tensor: shape=(1,), dtype=string, numpy=array([b'1'], dtype=object)>)]]>
9. buffer
<![CDATA[>>>> ds = ds.prefetch(buffer_size=image_count)]]>
備考：つぎの shuffle は，エラーになる：
<![CDATA[>>> ds = image_label_ds.apply(
... tf.data.Dataset.shuffle(buffer_size= image_count) )
Traceback (most recent call last):
File "<stdin>", line 2, in <module>
TypeError: shuffle() missing 1 required positional argument: 'self'
>>> tf.data.Dataset.shuffle(buffer_size= image_count)
Traceback (most recent call last):
File "<stdin>", line 1, in <module>
TypeError: shuffle() missing 1 required positional argument: 'self']]>
|