Skip to content

P
Pytorch Widedeep

Greenplum / Pytorch Widedeep
大约 1 年 前同步成功

通知 9
Star 1
Fork 0
P
Pytorch Widedeep

体验新版 GitCode，发现更多精彩内容 >>

未验证 提交 263663e5 编写于 作者: J Javier 提交者： GitHub
浏览文件
操作

Merge pull request #8 from jrzaurin/codestyle 

Codestyle
上级 9a969dc0 bd4b19fb
展开全部 隐藏空白更改
内联 并排
Showing with 3533 addition and 2531 deletion
.gitignore
浏览文件 @ 263663e5
......@@ -9,15 +9,22 @@ __pycache__*
# jupyter / ipython
.ipynb_checkpoints
Untitled*.ipynb
# data related dirs
data/
model_weights/
weights/
# tests
.coverage
.pytest_cache
# sublime
*.sublime-workspace
sftp*-config.json
# misc
.DS_store
\ No newline at end of file
.DS_store
#mypy
.mypy_cache
\ No newline at end of file
.travis.yml
浏览文件 @ 263663e5
......@@ -3,6 +3,26 @@ language: python
python:
- "3.6"
- "3.7"
matrix:
fast_finish: true
- name: "Code Style (Black/Flake8)"
install:
- pip install --upgrade pip
- pip install black
- pip install flake8
script:
# Black code style
- black --check --diff pytorch_widedeep tests examples setup.py
# Stop the build if there are Python syntax errors or undefined names
- flake8 . --count --select=E901,E999,F821,F822,F823 --ignore=E266 --show-source --statistics
# exit-zero treats all errors as warnings. The GitHub editor is 127 chars wide
- flake8 . --count --exit-zero --max-complexity=10 --max-line-length=127 --ignore=E203,E266,E501,E722,F401,F403,F405,W503,C901 --statistics
# - name: "static type check"
# script:
# - mypy pytorch_widedeep --ignore-missing-imports --no-strict-optional
after_success: skip
install:
- pip install --upgrade pip
- pip install .
......
docs/01_Preprocessors_and_utils.md
浏览文件 @ 263663e5
......@@ -586,10 +586,10 @@ texts[:2]
```python
text_preprocessor = TextPreprocessor()
X_text = text_preprocessor.fit_transform(df, text_col='description')
text_preprocessor = TextPreprocessor(text_col='description')
X_text = text_preprocessor.fit_transform(df)
# From here on, any new observation can be prepared by simply running `.transform`
# new_X_text = text_preprocessor.transform(new_df, text_col='description')
# new_X_text = text_preprocessor.transform(new_df)
```
The vocabulary contains 6400 tokens
......@@ -1681,21 +1681,21 @@ padded_seq[0]
```python
image_preprocessor = wd.preprocessing.ImagePreprocessor()
X_images = image_preprocessor.fit_transform(df, img_col="id", img_path="data/airbnb/property_picture/")
image_preprocessor = wd.preprocessing.ImagePreprocessor(img_col='id', img_path="data/airbnb/property_picture/")
X_images = image_preprocessor.fit_transform(df)
# From here on, any new observation can be prepared by simply running `.transform`
# new_X_images = image_preprocessor.transform(new_df, img_col="id", img_path="data/airbnb/property_picture/")
# new_X_images = image_preprocessor.transform(new_df)
```
Reading Images from data/airbnb/property_picture/
8%|▊ | 83/1001 [00:00<00:02, 410.41it/s]
4%|▍ | 40/1001 [00:00<00:02, 391.71it/s]
Resizing
100%|██████████| 1001/1001 [00:02<00:00, 419.09it/s]
100%|██████████| 1001/1001 [00:02<00:00, 415.12it/s]
Computing normalisation metrics
......@@ -1731,7 +1731,7 @@ prop_imgnames = sample(prop_imgnames, 10)
print(prop_imgnames)
```
['510940.jpg', '469775.jpg', '499555.jpg', '519598.jpg', '429444.jpg', '391898.jpg', '550146.jpg', '294247.jpg', '469630.jpg', '369374.jpg']
['512853.jpg', '460396.jpg', '92352.jpg', '472203.jpg', '534665.jpg', '529070.jpg', '549281.jpg', '499163.jpg', '218915.jpg', '526627.jpg']
......@@ -1768,7 +1768,7 @@ for i,im in enumerate(prop_imgs):
print([im.shape for im in prop_imgs])
```
[(426, 639, 3), (426, 639, 3), (426, 639, 3), (426, 639, 3), (426, 639, 3), (426, 639, 3), (426, 639, 3), (426, 639, 3), (426, 639, 3), (426, 639, 3)]
[(426, 639, 3), (426, 639, 3), (426, 639, 3), (426, 639, 3), (426, 360, 3), (426, 639, 3), (426, 639, 3), (426, 639, 3), (426, 639, 3), (426, 639, 3)]
......
docs/02_Model_Components.md
浏览文件 @ 263663e5
......@@ -113,11 +113,11 @@ deepdense(X_deep)
tensor([[ 0.0000, -0.0000, -2.2326, 3.5109, -0.0000, -1.2939, -0.0000, -0.0000],
[-1.1958, -0.0000, -0.0000, -0.0000, 0.0000, 0.0000, -0.0000, -0.0000],
[-0.0000, -0.0000, 0.0000, 0.0000, -1.4503, -0.0000, -0.0000, 3.9983],
[-1.0092, 0.5101, -0.0000, -1.4664, -0.0000, -0.0000, -1.0064, -0.0000],
[-1.1821, 0.0000, 0.0000, -1.5362, 0.0000, 0.0000, 3.9992, -0.0000]],
tensor([[ 1.9317, -0.0000, 1.3663, -0.3984, -0.0000, -0.0000, -0.0000, -1.2662],
[ 0.0000, -1.5337, -0.0000, 0.0726, -0.4231, 3.9977, -0.0000, -0.0000],
[-0.0000, -1.5839, 3.2978, -1.7084, -1.0877, -0.9574, 0.0000, -0.0000],
[-0.0000, 1.6664, -1.6006, 0.0000, -0.0000, -0.9844, -0.0000, -0.0521],
[ 2.4249, 0.0000, -0.0000, -0.0000, 0.0000, -0.0000, 2.6460, 0.0000]],
grad_fn=<MulBackward0>)
......@@ -339,10 +339,10 @@ deepimage(X_img)
tensor([[-1.7624e-03, 8.7372e-03, -3.9061e-03, 2.2110e-01, -1.9655e-03,
7.9629e-02, 2.5455e-01, -1.6910e-03],
[-2.8030e-05, 1.6680e-01, -3.5123e-03, 1.5065e-01, 5.2558e-02,
5.4472e-02, 4.2029e-02, 2.3403e-02]], grad_fn=<LeakyReluBackward1>)
tensor([[ 8.4865e-02, -3.4401e-03, -9.1973e-04, 3.4269e-01, 3.2816e-02,
1.9682e-02, -8.0740e-04, 9.4898e-03],
[ 1.5473e-01, -6.2664e-03, -9.3413e-05, 3.8768e-01, -1.9963e-03,
1.1729e-01, -2.7111e-03, 1.8670e-01]], grad_fn=<LeakyReluBackward1>)
......
docs/03_Binary_Classification_with_Defaults.md
浏览文件 @ 263663e5
......@@ -307,8 +307,7 @@ Have a look to notebooks one and two if you want to get a good understanding of
```python
wide_cols = ['education', 'relationship','workclass','occupation','native_country','gender']
crossed_cols = [('education', 'occupation'), ('native_country', 'occupation')]
cat_embed_cols = [('education',16), ('relationship',8), ('workclass',16),
('occupation',16),('native_country',16)]
cat_embed_cols = [('education',16), ('relationship',8), ('workclass',16), ('occupation',16),('native_country',16)]
continuous_cols = ["age","hours_per_week"]
target_col = 'income_label'
```
......@@ -445,16 +444,16 @@ model.fit(X_wide=X_wide, X_deep=X_deep, target=target, n_epochs=5, batch_size=25
Training
epoch 1: 100%|██████████| 153/153 [00:02<00:00, 55.95it/s, loss=0.41, metrics={'acc': 0.811}]
valid: 100%|██████████| 39/39 [00:00<00:00, 119.88it/s, loss=0.362, metrics={'acc': 0.8152}]
epoch 2: 100%|██████████| 153/153 [00:02<00:00, 58.75it/s, loss=0.35, metrics={'acc': 0.8347}]
valid: 100%|██████████| 39/39 [00:00<00:00, 99.19it/s, loss=0.353, metrics={'acc': 0.8352}]
epoch 3: 100%|██████████| 153/153 [00:02<00:00, 59.79it/s, loss=0.344, metrics={'acc': 0.8372}]
valid: 100%|██████████| 39/39 [00:00<00:00, 113.76it/s, loss=0.349, metrics={'acc': 0.8376}]
epoch 4: 100%|██████████| 153/153 [00:02<00:00, 56.89it/s, loss=0.341, metrics={'acc': 0.8389}]
valid: 100%|██████████| 39/39 [00:00<00:00, 104.95it/s, loss=0.347, metrics={'acc': 0.8388}]
epoch 5: 100%|██████████| 153/153 [00:02<00:00, 56.49it/s, loss=0.338, metrics={'acc': 0.8404}]
valid: 100%|██████████| 39/39 [00:00<00:00, 105.79it/s, loss=0.346, metrics={'acc': 0.8405}]
epoch 1: 100%|██████████| 153/153 [00:02<00:00, 56.52it/s, loss=0.412, metrics={'acc': 0.7993}]
valid: 100%|██████████| 39/39 [00:00<00:00, 123.12it/s, loss=0.352, metrics={'acc': 0.8071}]
epoch 2: 100%|██████████| 153/153 [00:02<00:00, 59.55it/s, loss=0.351, metrics={'acc': 0.8351}]
valid: 100%|██████████| 39/39 [00:00<00:00, 121.98it/s, loss=0.346, metrics={'acc': 0.8359}]
epoch 3: 100%|██████████| 153/153 [00:02<00:00, 59.82it/s, loss=0.346, metrics={'acc': 0.8377}]
valid: 100%|██████████| 39/39 [00:00<00:00, 121.88it/s, loss=0.344, metrics={'acc': 0.8384}]
epoch 4: 100%|██████████| 153/153 [00:02<00:00, 58.97it/s, loss=0.342, metrics={'acc': 0.8392}]
valid: 100%|██████████| 39/39 [00:00<00:00, 122.20it/s, loss=0.342, metrics={'acc': 0.84}]
epoch 5: 100%|██████████| 153/153 [00:02<00:00, 58.28it/s, loss=0.34, metrics={'acc': 0.8406}]
valid: 100%|██████████| 39/39 [00:00<00:00, 116.57it/s, loss=0.341, metrics={'acc': 0.8413}]
As you can see, you can run a wide and deep model in just a few lines of code
docs/04_Binary_Classification_Varying_Parameters.md
浏览文件 @ 263663e5
......@@ -307,8 +307,7 @@ Have a look to notebooks one and two if you want to get a good understanding of
```python
wide_cols = ['education', 'relationship','workclass','occupation','native_country','gender']
crossed_cols = [('education', 'occupation'), ('native_country', 'occupation')]
cat_embed_cols = [('education',16), ('relationship',8), ('workclass',16),
('occupation',16),('native_country',16)]
cat_embed_cols = [('education',16), ('relationship',8), ('workclass',16), ('occupation',16),('native_country',16)]
continuous_cols = ["age","hours_per_week"]
target_col = 'income_label'
```
......@@ -477,26 +476,26 @@ model.fit(X_wide=X_wide, X_deep=X_deep, target=target, n_epochs=10, batch_size=2
Training
epoch 1: 100%|██████████| 153/153 [00:03<00:00, 44.53it/s, loss=0.783, metrics={'acc': 0.6151}]
valid: 100%|██████████| 39/39 [00:00<00:00, 116.60it/s, loss=0.469, metrics={'acc': 0.6493}]
epoch 2: 100%|██████████| 153/153 [00:03<00:00, 47.32it/s, loss=0.529, metrics={'acc': 0.7565}]
valid: 100%|██████████| 39/39 [00:00<00:00, 95.90it/s, loss=0.396, metrics={'acc': 0.7685}]
epoch 3: 100%|██████████| 153/153 [00:03<00:00, 46.55it/s, loss=0.457, metrics={'acc': 0.7907}]
valid: 100%|██████████| 39/39 [00:00<00:00, 116.63it/s, loss=0.372, metrics={'acc': 0.798}]
epoch 4: 100%|██████████| 153/153 [00:03<00:00, 49.69it/s, loss=0.421, metrics={'acc': 0.8038}]
valid: 100%|██████████| 39/39 [00:00<00:00, 128.04it/s, loss=0.366, metrics={'acc': 0.8091}]
epoch 5: 100%|██████████| 153/153 [00:03<00:00, 50.27it/s, loss=0.398, metrics={'acc': 0.815}]
valid: 100%|██████████| 39/39 [00:00<00:00, 131.81it/s, loss=0.36, metrics={'acc': 0.8188}]
epoch 6: 100%|██████████| 153/153 [00:03<00:00, 50.16it/s, loss=0.388, metrics={'acc': 0.817}]
valid: 100%|██████████| 39/39 [00:00<00:00, 130.33it/s, loss=0.36, metrics={'acc': 0.8204}]
epoch 7: 100%|██████████| 153/153 [00:03<00:00, 50.06it/s, loss=0.386, metrics={'acc': 0.8175}]
valid: 100%|██████████| 39/39 [00:00<00:00, 133.34it/s, loss=0.359, metrics={'acc': 0.8208}]
epoch 8: 100%|██████████| 153/153 [00:03<00:00, 50.43it/s, loss=0.387, metrics={'acc': 0.8189}]
valid: 100%|██████████| 39/39 [00:00<00:00, 124.92it/s, loss=0.359, metrics={'acc': 0.8221}]
epoch 9: 100%|██████████| 153/153 [00:03<00:00, 50.34it/s, loss=0.385, metrics={'acc': 0.8185}]
valid: 100%|██████████| 39/39 [00:00<00:00, 130.44it/s, loss=0.358, metrics={'acc': 0.8219}]
epoch 10: 100%|██████████| 153/153 [00:03<00:00, 50.29it/s, loss=0.384, metrics={'acc': 0.8191}]
valid: 100%|██████████| 39/39 [00:00<00:00, 129.58it/s, loss=0.358, metrics={'acc': 0.8225}]
epoch 1: 100%|██████████| 153/153 [00:03<00:00, 47.06it/s, loss=0.731, metrics={'acc': 0.6468}]
valid: 100%|██████████| 39/39 [00:00<00:00, 118.07it/s, loss=0.418, metrics={'acc': 0.6785}]
epoch 2: 100%|██████████| 153/153 [00:03<00:00, 49.72it/s, loss=0.51, metrics={'acc': 0.7637}]
valid: 100%|██████████| 39/39 [00:00<00:00, 114.38it/s, loss=0.376, metrics={'acc': 0.7765}]
epoch 3: 100%|██████████| 153/153 [00:03<00:00, 49.32it/s, loss=0.448, metrics={'acc': 0.7927}]
valid: 100%|██████████| 39/39 [00:00<00:00, 113.23it/s, loss=0.361, metrics={'acc': 0.8007}]
epoch 4: 100%|██████████| 153/153 [00:03<00:00, 48.23it/s, loss=0.413, metrics={'acc': 0.8079}]
valid: 100%|██████████| 39/39 [00:00<00:00, 118.17it/s, loss=0.355, metrics={'acc': 0.8132}]
epoch 5: 100%|██████████| 153/153 [00:03<00:00, 48.27it/s, loss=0.395, metrics={'acc': 0.8149}]
valid: 100%|██████████| 39/39 [00:00<00:00, 114.71it/s, loss=0.352, metrics={'acc': 0.8191}]
epoch 6: 100%|██████████| 153/153 [00:03<00:00, 48.97it/s, loss=0.387, metrics={'acc': 0.8157}]
valid: 100%|██████████| 39/39 [00:00<00:00, 116.88it/s, loss=0.352, metrics={'acc': 0.8199}]
epoch 7: 100%|██████████| 153/153 [00:03<00:00, 48.56it/s, loss=0.388, metrics={'acc': 0.8153}]
valid: 100%|██████████| 39/39 [00:00<00:00, 116.33it/s, loss=0.352, metrics={'acc': 0.8195}]
epoch 8: 100%|██████████| 153/153 [00:03<00:00, 48.54it/s, loss=0.383, metrics={'acc': 0.8184}]
valid: 100%|██████████| 39/39 [00:00<00:00, 93.05it/s, loss=0.351, metrics={'acc': 0.822}]
epoch 9: 100%|██████████| 153/153 [00:03<00:00, 45.71it/s, loss=0.385, metrics={'acc': 0.8196}]
valid: 100%|██████████| 39/39 [00:00<00:00, 110.48it/s, loss=0.351, metrics={'acc': 0.8229}]
epoch 10: 100%|██████████| 153/153 [00:03<00:00, 48.44it/s, loss=0.382, metrics={'acc': 0.8194}]
valid: 100%|██████████| 39/39 [00:00<00:00, 114.40it/s, loss=0.35, metrics={'acc': 0.8228}]
......@@ -646,7 +645,7 @@ model.history.epoch
print(model.history._history)
```
{'train_loss': [0.7826161832591287, 0.5294494130253012, 0.45743006565212424, 0.4206276263286865, 0.3982163554702709, 0.3881325295158461, 0.3862898593244989, 0.38681577603801404, 0.38500378529230755, 0.38388273743243], 'train_acc': [0.6151, 0.7565, 0.7907, 0.8038, 0.815, 0.817, 0.8175, 0.8189, 0.8185, 0.8191], 'val_loss': [0.4694176025879689, 0.3960292133001181, 0.37219820802028364, 0.3658289725963886, 0.3600605313594525, 0.35951805343994725, 0.35915129765486103, 0.3585702692851042, 0.3578468553530864, 0.3576407875770178], 'val_acc': [0.6493, 0.7685, 0.798, 0.8091, 0.8188, 0.8204, 0.8208, 0.8221, 0.8219, 0.8225]}
{'train_loss': [0.7313813343157176, 0.5101876866583731, 0.44813506724008545, 0.41332343941420513, 0.3945406624694276, 0.3871746306715448, 0.3884129401515512, 0.38312816230300206, 0.3847907395923839, 0.3817657043341718], 'train_acc': [0.6468, 0.7637, 0.7927, 0.8079, 0.8149, 0.8157, 0.8153, 0.8184, 0.8196, 0.8194], 'val_loss': [0.41844800649545133, 0.3759944920356457, 0.36132928041311413, 0.3554159953044011, 0.3523857922126085, 0.3518377657120044, 0.35156664175864977, 0.35120767278549, 0.35089820012068135, 0.35047405576094603], 'val_acc': [0.6785, 0.7765, 0.8007, 0.8132, 0.8191, 0.8199, 0.8195, 0.822, 0.8229, 0.8228]}
......@@ -669,71 +668,69 @@ model.get_embeddings(col_name='education', cat_encoding_dict=preprocess_deep.enc
{'11th': array([-1.08425401e-01, 5.09871461e-04, 1.25755548e-01, -1.20801523e-01,
-2.56043434e-01, -3.55644524e-02, -8.66190940e-02, -1.39202878e-01,
1.11087626e-04, 4.54997361e-01, -2.31609955e-01, -1.36443637e-02,
8.78131837e-02, -3.07353675e-01, -1.10240346e-02, 6.45920560e-02],
dtype=float32),
'HS-grad': array([ 0.19832617, 0.12040217, -0.5314197 , 0.35005897, -0.15391229,
-0.22196807, 0.09345723, 0.06745315, 0.25015768, 0.08744714,
0.24480642, -0.08957793, 0.27947524, -0.26326123, -0.19119193,
-0.10995993], dtype=float32),
'Assoc-acdm': array([ 0.06525454, -0.2618052 , -0.09840333, 0.10541438, 0.33471954,
-0.04292247, 0.10712572, 0.34287837, -0.18687049, -0.13836485,
-0.1715912 , 0.15273218, -0.03476759, -0.07450581, 0.56081617,
0.29201028], dtype=float32),
'Some-college': array([-0.45491776, -0.17205039, 0.21580465, -0.2539856 , 0.02358766,
-0.05496917, -0.01120283, 0.09221312, -0.12831998, 0.17159238,
0.196605 , -0.2090644 , -0.11193639, -0.18394227, -0.16056207,
0.02444198], dtype=float32),
'10th': array([-0.5581912 , -0.20644131, 0.1300292 , -0.10135209, 0.4538276 ,
-0.27146348, 0.12652951, 0.5233289 , 0.01145706, -0.05667543,
0.43509725, -0.74307233, 0.00139265, 0.07225899, 0.0781986 ,
-0.2610258 ], dtype=float32),
'Prof-school': array([-6.5744489e-02, 1.3956554e-01, 5.7986474e-01, 2.7874210e-01,
-2.4446699e-01, 7.9873689e-02, -3.8569799e-01, 2.2757685e-01,
-3.8109139e-02, 3.3144853e-01, -3.8229354e-02, 2.9802489e-01,
-1.5467829e-01, 5.4805580e-04, -2.1627106e-01, -2.6592135e-02],
dtype=float32),
'7th-8th': array([ 0.10858492, 0.42190084, 0.07536066, -0.11707054, 0.05351719,
0.32636967, 0.14053936, 0.45679298, -0.2558197 , -0.47910702,
0.4725715 , -0.0981419 , 0.3462793 , 0.07776859, -0.45930195,
0.12625834], dtype=float32),
'Bachelors': array([ 0.01805384, -0.10573057, 0.25564098, -0.27709666, -0.16297452,
-0.1851758 , -0.5702467 , -0.23569717, 0.067039 , -0.28916818,
-0.22313781, -0.23893505, 0.37708414, 0.17465928, -0.47459307,
0.04889947], dtype=float32),
'Masters': array([ 0.11953138, 0.11543513, -0.3954705 , 0.32583147, 0.23851769,
-0.6448425 , 0.00705628, 0.10673986, -0.08305098, -0.10872949,
-0.46080047, -0.05367367, -0.18693425, 0.14182107, -0.39178014,
-0.23969549], dtype=float32),
'Doctorate': array([ 0.04873321, -0.19027464, -0.10777274, -0.17476888, 0.47248197,
-0.2873778 , -0.29792303, -0.06811561, 0.16541322, -0.17425427,
-0.09404507, 0.06525683, 0.06408301, 0.38656166, 0.13369907,
0.10825544], dtype=float32),
'5th-6th': array([ 0.08566641, 0.03589746, 0.17174615, 0.08747724, 0.2698885 ,
0.08344392, -0.23652045, 0.31357667, 0.3546634 , -0.29814255,
0.10943606, 0.45218074, -0.0614133 , -0.31987205, 0.34947518,
0.07603104], dtype=float32),
'Assoc-voc': array([-0.07830544, 0.0278313 , 0.34295908, -0.27213913, -0.20097388,
0.10972344, 0.14000823, -0.24098383, -0.16614872, 0.19084413,
-0.02334382, 0.5209352 , 0.24089335, -0.1350642 , -0.23480216,
-0.32963687], dtype=float32),
'9th': array([ 0.12994888, 0.02475524, -0.12875263, 0.0097373 , 0.38253692,
-0.2718543 , 0.13766348, 0.27845392, -0.2036348 , -0.20567507,
-0.11305337, -0.47028974, 0.07009655, -0.29621345, -0.17303236,
0.15854478], dtype=float32),
'12th': array([-0.15079321, -0.26879913, -0.5159767 , 0.30044943, 0.0295292 ,
-0.32494095, 0.20975012, 0.35193697, -0.5034315 , -0.14420179,
0.06113023, 0.22398257, 0.0087006 , 0.09041765, -0.09754901,
-0.21647781], dtype=float32),
'1st-4th': array([-0.3199786 , 0.10094872, -0.10035568, 0.10014401, -0.09340642,
-0.00761677, 0.50759906, 0.288856 , -0.18745485, 0.05442255,
0.6481828 , 0.18515776, 0.21597311, -0.21534163, 0.01798662,
-0.22816893], dtype=float32),
'Preschool': array([ 0.10035816, -0.24015287, 0.00935481, 0.05356123, -0.18744251,
-0.39735606, 0.03849271, -0.2864288 , -0.10379744, 0.20251973,
0.14565234, -0.24607188, -0.14268415, 0.1209868 , 0.04662501,
0.41015574], dtype=float32)}
{'11th': array([-0.04807916, 0.21404432, 0.12517522, -0.154123 , 0.06864536,
0.00092955, -0.38516527, -0.18440197, 0.15861034, 0.12012056,
0.55413646, -0.16920644, 0.1356924 , -0.37921003, 0.53833497,
0.08743049], dtype=float32),
'HS-grad': array([ 0.37504154, 0.34191516, 0.27299362, 0.22921972, 0.07420117,
0.34922913, 0.19239122, -0.42343035, -0.845824 , -0.07287297,
0.27455565, 0.19505064, 0.07062761, -0.5201107 , 0.37823108,
0.46134958], dtype=float32),
'Assoc-acdm': array([ 0.22331461, 0.15005238, 0.13472553, -0.16886246, -0.12053325,
-0.04233408, -0.08905135, -0.54481906, 0.24300168, -0.21069968,
-0.00685616, -0.38423738, -0.00281451, 0.10599079, -0.05224385,
0.2891064 ], dtype=float32),
'Some-college': array([-0.09498356, -0.16801773, -0.09181987, 0.05381393, -0.03607363,
-0.05759075, 0.09382061, 0.33274302, -0.11906563, 0.14481838,
-0.1765725 , 0.20070277, 0.2960993 , -0.02055654, 0.26645136,
0.4075843 ], dtype=float32),
'10th': array([-0.12961714, -0.27546212, 0.24345328, -0.24318363, 0.31552687,
0.16653115, -0.05234893, 0.06825106, 0.2388588 , 0.10887478,
-0.12004007, -0.00373614, -0.0223387 , 0.133562 , 0.29672143,
0.03046475], dtype=float32),
'Prof-school': array([-0.1589678 , -0.07629952, 0.00763621, 0.13788143, 0.4114019 ,
0.07717889, -0.17072953, 0.29419565, -0.18929462, -0.09182461,
-0.08409152, 0.01395322, -0.20351669, 0.18333136, -0.03983613,
-0.31888708], dtype=float32),
'7th-8th': array([ 0.39654806, 0.26095334, -0.3147828 , -0.41267306, -0.23983437,
-0.08034727, 0.4807234 , 0.3054779 , -0.3085564 , -0.07860225,
-0.1279486 , -0.2846014 , 0.1358583 , 0.24006395, -0.18911272,
-0.2299538 ], dtype=float32),
'Bachelors': array([ 0.35242578, -0.03246311, 0.15835243, -0.06434399, 0.03403192,
0.0088449 , 0.00627425, -0.31485453, -0.30984947, -0.23008366,
-0.09467663, 0.17246258, -0.09432375, 0.07691337, 0.70925283,
0.18795769], dtype=float32),
'Masters': array([-0.14503758, 0.0048258 , 0.58242404, 0.28511924, -0.13773848,
0.35109136, 0.05824559, 0.3609631 , 0.4700086 , 0.4251728 ,
-0.2538366 , -0.00297809, 0.1424264 , -0.12481072, -0.09403807,
0.00634856], dtype=float32),
'Doctorate': array([-0.12487873, -0.1699961 , 0.2220065 , -0.04808738, 0.09443628,
-0.21019349, -0.23745097, 0.28523713, 0.05516997, -0.04004707,
0.3316393 , 0.18710822, 0.4153885 , -0.12905155, 0.03055826,
0.0664137 ], dtype=float32),
'5th-6th': array([ 0.21891987, -0.13600409, -0.03123563, 0.16288632, -0.03479639,
-0.4221951 , 0.4688111 , 0.08145971, -0.29254073, 0.18396533,
-0.20204993, -0.03327556, -0.2558647 , 0.56448 , -0.30299884,
0.07629355], dtype=float32),
'Assoc-voc': array([-0.01987046, -0.06434393, 0.00226 , 0.08150155, -0.33775425,
-0.13507745, 0.12741297, 0.0542295 , 0.09895965, 0.067229 ,
-0.1718493 , 0.01054914, 0.10441845, -0.18814586, -0.01663602,
0.03088147], dtype=float32),
'9th': array([-0.24095939, 0.2750888 , 0.01418325, -0.36754113, 0.5431856 ,
-0.19582956, 0.03485603, 0.22838333, -0.05723334, 0.10631263,
0.06331363, -0.09572615, 0.21977316, -0.02579625, -0.13822857,
0.28736743], dtype=float32),
'12th': array([-0.20278502, -0.19245535, -0.04846343, 0.14459866, 0.25858438,
0.15333128, 0.5074635 , -0.15141617, -0.19331448, -0.2630267 ,
-0.1378872 , -0.16868882, 0.4048257 , -0.34108582, -0.23098588,
0.2859633 ], dtype=float32),
'1st-4th': array([-0.53678703, 0.19669479, -0.18026853, 0.33791658, 0.14260627,
0.20269199, 0.00518189, 0.01120056, 0.01568659, 0.28752655,
0.3359768 , 0.01758064, 0.11630564, -0.35470524, -0.05704446,
0.41216984], dtype=float32),
'Preschool': array([ 0.10326536, -0.02895411, 0.11348445, 0.03685748, 0.55893034,
-0.2522173 , -0.07186767, -0.30955225, -0.17825711, 0.02907414,
-0.61121726, 0.40596214, 0.63471395, 0.3304132 , 0.05272925,
-0.4266447 ], dtype=float32)}
docs/05_Regression_with_Images_and_Text.md
浏览文件 @ 263663e5
......@@ -729,8 +729,8 @@ X_deep = deep_preprocessor.fit_transform(df)
```python
text_preprocessor = TextPreprocessor(word_vectors_path=word_vectors_path)
X_text = text_preprocessor.fit_transform(df, text_col)
text_preprocessor = TextPreprocessor(word_vectors_path=word_vectors_path, text_col=text_col)
X_text = text_preprocessor.fit_transform(df)
```
The vocabulary contains 6400 tokens
......@@ -742,19 +742,19 @@ X_text = text_preprocessor.fit_transform(df, text_col)
```python
image_processor = ImagePreprocessor()
X_images = image_processor.fit_transform(df, img_col, img_path)
image_processor = ImagePreprocessor(img_col = img_col, img_path = img_path)
X_images = image_processor.fit_transform(df)
```
Reading Images from data/airbnb/property_picture
4%|▍ | 43/1001 [00:00<00:02, 428.36it/s]
9%|▊ | 87/1001 [00:00<00:02, 428.89it/s]
Resizing
100%|██████████| 1001/1001 [00:02<00:00, 425.98it/s]
100%|██████████| 1001/1001 [00:02<00:00, 426.92it/s]
Computing normalisation metrics
......@@ -804,8 +804,8 @@ model.fit(X_wide=X_wide, X_deep=X_deep, X_text=X_text, X_img=X_images,
Training
epoch 1: 100%|██████████| 25/25 [02:04<00:00, 4.97s/it, loss=117]
valid: 100%|██████████| 7/7 [00:14<00:00, 2.00s/it, loss=122]
epoch 1: 100%|██████████| 25/25 [02:06<00:00, 5.05s/it, loss=118]
valid: 100%|██████████| 7/7 [00:14<00:00, 2.01s/it, loss=226]
### Regression with varying parameters and a FC-Head receiving the full deep side
......@@ -1050,8 +1050,8 @@ model.compile(method='regression', initializers=initializers, optimizers=optimiz
lr_schedulers=schedulers, callbacks=callbacks, transforms=transforms)
```
/Users/javier/pytorch-widedeep/pytorch_widedeep/initializers.py:32: UserWarning: No initializer found for deephead
if self.verbose: warnings.warn("No initializer found for {}".format(name))
/Users/javier/pytorch-widedeep/pytorch_widedeep/initializers.py:31: UserWarning: No initializer found for deephead
warnings.warn("No initializer found for {}".format(name))
......@@ -1221,8 +1221,8 @@ model.fit(X_wide=X_wide, X_deep=X_deep, X_text=X_text, X_img=X_images,
Training
epoch 1: 100%|██████████| 25/25 [02:08<00:00, 5.15s/it, loss=128]
valid: 100%|██████████| 7/7 [00:14<00:00, 2.05s/it, loss=95.5]
epoch 1: 100%|██████████| 25/25 [02:04<00:00, 4.97s/it, loss=127]
valid: 100%|██████████| 7/7 [00:14<00:00, 2.02s/it, loss=94]
we have only run one epoch, but let's check that the LRHistory callback records the lr values for each group
......
docs/06_WarmUp_Model_Components.md
浏览文件 @ 263663e5
......@@ -92,36 +92,36 @@ model.fit(X_wide=X_wide, X_deep=X_deep, target=target, n_epochs=5, batch_size=25
Warming up wide for 5 epochs
epoch 1: 100%|██████████| 153/153 [00:01<00:00, 133.55it/s, loss=0.469, metrics={'acc': 0.7963}]
epoch 2: 100%|██████████| 153/153 [00:00<00:00, 156.42it/s, loss=0.372, metrics={'acc': 0.8119}]
epoch 3: 100%|██████████| 153/153 [00:01<00:00, 152.65it/s, loss=0.363, metrics={'acc': 0.8183}]
epoch 4: 100%|██████████| 153/153 [00:01<00:00, 144.01it/s, loss=0.36, metrics={'acc': 0.8218}]
epoch 5: 100%|██████████| 153/153 [00:00<00:00, 155.56it/s, loss=0.359, metrics={'acc': 0.8241}]
epoch 1: 100%|██████████| 153/153 [00:01<00:00, 131.88it/s, loss=0.471, metrics={'acc': 0.7946}]
epoch 2: 100%|██████████| 153/153 [00:00<00:00, 154.81it/s, loss=0.373, metrics={'acc': 0.8115}]
epoch 3: 100%|██████████| 153/153 [00:01<00:00, 151.56it/s, loss=0.364, metrics={'acc': 0.8182}]
epoch 4: 100%|██████████| 153/153 [00:00<00:00, 154.22it/s, loss=0.362, metrics={'acc': 0.8216}]
epoch 5: 100%|██████████| 153/153 [00:01<00:00, 152.65it/s, loss=0.36, metrics={'acc': 0.8238}]
0%| | 0/153 [00:00<?, ?it/s]
Warming up deepdense for 5 epochs
epoch 1: 100%|██████████| 153/153 [00:02<00:00, 64.03it/s, loss=0.4, metrics={'acc': 0.8214}]
epoch 2: 100%|██████████| 153/153 [00:02<00:00, 64.30it/s, loss=0.347, metrics={'acc': 0.8237}]
epoch 3: 100%|██████████| 153/153 [00:02<00:00, 64.18it/s, loss=0.341, metrics={'acc': 0.8258}]
epoch 4: 100%|██████████| 153/153 [00:02<00:00, 64.98it/s, loss=0.337, metrics={'acc': 0.8277}]
epoch 5: 100%|██████████| 153/153 [00:02<00:00, 65.70it/s, loss=0.333, metrics={'acc': 0.8294}]
epoch 1: 100%|██████████| 153/153 [00:02<00:00, 64.04it/s, loss=0.395, metrics={'acc': 0.8222}]
epoch 2: 100%|██████████| 153/153 [00:02<00:00, 65.34it/s, loss=0.349, metrics={'acc': 0.8242}]
epoch 3: 100%|██████████| 153/153 [00:02<00:00, 65.05it/s, loss=0.343, metrics={'acc': 0.8262}]
epoch 4: 100%|██████████| 153/153 [00:02<00:00, 64.93it/s, loss=0.339, metrics={'acc': 0.8279}]
epoch 5: 100%|██████████| 153/153 [00:02<00:00, 65.15it/s, loss=0.335, metrics={'acc': 0.8295}]
0%| | 0/153 [00:00<?, ?it/s]
Training
epoch 1: 100%|██████████| 153/153 [00:02<00:00, 58.97it/s, loss=0.343, metrics={'acc': 0.8442}]
valid: 100%|██████████| 39/39 [00:00<00:00, 123.14it/s, loss=0.349, metrics={'acc': 0.8436}]
epoch 2: 100%|██████████| 153/153 [00:02<00:00, 59.48it/s, loss=0.333, metrics={'acc': 0.8457}]
valid: 100%|██████████| 39/39 [00:00<00:00, 114.05it/s, loss=0.346, metrics={'acc': 0.8447}]
epoch 3: 100%|██████████| 153/153 [00:02<00:00, 59.82it/s, loss=0.331, metrics={'acc': 0.8471}]
valid: 100%|██████████| 39/39 [00:00<00:00, 122.31it/s, loss=0.345, metrics={'acc': 0.8457}]
epoch 4: 100%|██████████| 153/153 [00:02<00:00, 59.20it/s, loss=0.329, metrics={'acc': 0.8474}]
valid: 100%|██████████| 39/39 [00:00<00:00, 113.66it/s, loss=0.345, metrics={'acc': 0.8459}]
epoch 5: 100%|██████████| 153/153 [00:02<00:00, 59.15it/s, loss=0.328, metrics={'acc': 0.8479}]
valid: 100%|██████████| 39/39 [00:00<00:00, 118.58it/s, loss=0.345, metrics={'acc': 0.8462}]
epoch 1: 100%|██████████| 153/153 [00:02<00:00, 58.31it/s, loss=0.345, metrics={'acc': 0.8415}]
valid: 100%|██████████| 39/39 [00:00<00:00, 120.77it/s, loss=0.346, metrics={'acc': 0.8416}]
epoch 2: 100%|██████████| 153/153 [00:02<00:00, 58.33it/s, loss=0.335, metrics={'acc': 0.8446}]
valid: 100%|██████████| 39/39 [00:00<00:00, 117.88it/s, loss=0.344, metrics={'acc': 0.8438}]
epoch 3: 100%|██████████| 153/153 [00:02<00:00, 58.43it/s, loss=0.331, metrics={'acc': 0.8457}]
valid: 100%|██████████| 39/39 [00:00<00:00, 109.26it/s, loss=0.343, metrics={'acc': 0.8449}]
epoch 4: 100%|██████████| 153/153 [00:02<00:00, 58.08it/s, loss=0.329, metrics={'acc': 0.8457}]
valid: 100%|██████████| 39/39 [00:00<00:00, 120.23it/s, loss=0.344, metrics={'acc': 0.8446}]
epoch 5: 100%|██████████| 153/153 [00:02<00:00, 58.75it/s, loss=0.327, metrics={'acc': 0.8464}]
valid: 100%|██████████| 39/39 [00:00<00:00, 119.22it/s, loss=0.344, metrics={'acc': 0.8453}]
### Warm up Gradually: The "felbo" and the "howard" routines
......@@ -187,11 +187,11 @@ X_wide = prepare_wide.fit_transform(df)
prepare_deep = DeepPreprocessor(embed_cols=cat_embed_cols, continuous_cols=continuous_cols)
X_deep = prepare_deep.fit_transform(df)
text_processor = TextPreprocessor(word_vectors_path=word_vectors_path)
X_text = text_processor.fit_transform(df, text_col)
text_processor = TextPreprocessor(word_vectors_path=word_vectors_path, text_col=text_col)
X_text = text_processor.fit_transform(df)
image_processor = ImagePreprocessor()
X_images = image_processor.fit_transform(df, img_col, img_path)
image_processor = ImagePreprocessor(img_col=img_col, img_path=img_path)
X_images = image_processor.fit_transform(df)
```
The vocabulary contains 6400 tokens
......@@ -202,12 +202,12 @@ X_images = image_processor.fit_transform(df, img_col, img_path)
Reading Images from data/airbnb/property_picture
8%|▊ | 84/1001 [00:00<00:02, 413.82it/s]
3%|▎ | 29/1001 [00:00<00:03, 282.66it/s]
Resizing
100%|██████████| 1001/1001 [00:02<00:00, 419.78it/s]
100%|██████████| 1001/1001 [00:03<00:00, 327.44it/s]
Computing normalisation metrics
......@@ -501,56 +501,56 @@ model.fit(X_wide=X_wide, X_deep=X_deep, X_text=X_text, X_img=X_images, target=ta
Warming up wide for 1 epochs
epoch 1: 100%|██████████| 25/25 [00:00<00:00, 42.89it/s, loss=127]
epoch 1: 100%|██████████| 25/25 [00:00<00:00, 57.98it/s, loss=127]
0%| | 0/25 [00:00<?, ?it/s]
Warming up deepdense for 1 epochs
epoch 1: 100%|██████████| 25/25 [00:00<00:00, 43.28it/s, loss=117]
epoch 1: 100%|██████████| 25/25 [00:00<00:00, 45.81it/s, loss=116]
0%| | 0/25 [00:00<?, ?it/s]
Warming up deeptext for 1 epochs
epoch 1: 100%|██████████| 25/25 [00:05<00:00, 4.78it/s, loss=132]
epoch 1: 100%|██████████| 25/25 [00:04<00:00, 5.37it/s, loss=132]
0%| | 0/25 [00:00<?, ?it/s]
Warming up deepimage, layer 1 of 5
epoch 1: 100%|██████████| 25/25 [01:11<00:00, 2.86s/it, loss=119]
epoch 1: 100%|██████████| 25/25 [01:10<00:00, 2.83s/it, loss=119]
0%| | 0/25 [00:00<?, ?it/s]
Warming up deepimage, layer 2 of 5
epoch 1: 100%|██████████| 25/25 [01:35<00:00, 3.81s/it, loss=108]
epoch 1: 100%|██████████| 25/25 [01:34<00:00, 3.76s/it, loss=108]
0%| | 0/25 [00:00<?, ?it/s]
Warming up deepimage, layer 3 of 5
epoch 1: 100%|██████████| 25/25 [01:58<00:00, 4.76s/it, loss=105]
epoch 1: 100%|██████████| 25/25 [01:57<00:00, 4.69s/it, loss=106]
0%| | 0/25 [00:00<?, ?it/s]
Warming up deepimage, layer 4 of 5
epoch 1: 100%|██████████| 25/25 [02:25<00:00, 5.80s/it, loss=105]
epoch 1: 100%|██████████| 25/25 [02:24<00:00, 5.79s/it, loss=105]
0%| | 0/25 [00:00<?, ?it/s]
Warming up deepimage, layer 5 of 5
epoch 1: 100%|██████████| 25/25 [03:04<00:00, 7.38s/it, loss=106]
epoch 1: 100%|██████████| 25/25 [03:01<00:00, 7.26s/it, loss=105]
0%| | 0/25 [00:00<?, ?it/s]
Training
epoch 1: 100%|██████████| 25/25 [02:05<00:00, 5.00s/it, loss=130]
valid: 100%|██████████| 7/7 [00:14<00:00, 2.04s/it, loss=125]
epoch 1: 100%|██████████| 25/25 [02:05<00:00, 5.03s/it, loss=129]
valid: 100%|██████████| 7/7 [00:14<00:00, 2.11s/it, loss=103]
And one would access to the `felbo` routine by changing the `param`, `warm_routine` to `'felbo'`
examples/02_Model_Components.ipynb
浏览文件 @ 263663e5
......@@ -46,7 +46,7 @@
},
{
"cell_type": "code",
"execution_count": 3,
"execution_count": 2,
"metadata": {},
"outputs": [
{
......@@ -57,7 +57,7 @@
")"
]
},
"execution_count": 3,
"execution_count": 2,
"metadata": {},
"output_type": "execute_result"
}
......@@ -78,7 +78,7 @@
},
{
"cell_type": "code",
"execution_count": 4,
"execution_count": 3,
"metadata": {},
"outputs": [],
"source": [
......@@ -89,7 +89,7 @@
},
{
"cell_type": "code",
"execution_count": 5,
"execution_count": 4,
"metadata": {},
"outputs": [],
"source": [
......@@ -112,7 +112,7 @@
},
{
"cell_type": "code",
"execution_count": 7,
"execution_count": 5,
"metadata": {},
"outputs": [],
"source": [
......@@ -122,7 +122,7 @@
},
{
"cell_type": "code",
"execution_count": 8,
"execution_count": 6,
"metadata": {},
"outputs": [
{
......@@ -153,7 +153,7 @@
")"
]
},
"execution_count": 8,
"execution_count": 6,
"metadata": {},
"output_type": "execute_result"
}
......@@ -164,21 +164,21 @@
},
{
"cell_type": "code",
"execution_count": 9,
"execution_count": 7,
"metadata": {},
"outputs": [
{
"data": {
"text/plain": [
"tensor([[ 0.0000, -0.0000, -2.2326, 3.5109, -0.0000, -1.2939, -0.0000, -0.0000],\n",
" [-1.1958, -0.0000, -0.0000, -0.0000, 0.0000, 0.0000, -0.0000, -0.0000],\n",
" [-0.0000, -0.0000, 0.0000, 0.0000, -1.4503, -0.0000, -0.0000, 3.9983],\n",
" [-1.0092, 0.5101, -0.0000, -1.4664, -0.0000, -0.0000, -1.0064, -0.0000],\n",
" [-1.1821, 0.0000, 0.0000, -1.5362, 0.0000, 0.0000, 3.9992, -0.0000]],\n",
"tensor([[ 1.9317, -0.0000, 1.3663, -0.3984, -0.0000, -0.0000, -0.0000, -1.2662],\n",
" [ 0.0000, -1.5337, -0.0000, 0.0726, -0.4231, 3.9977, -0.0000, -0.0000],\n",
" [-0.0000, -1.5839, 3.2978, -1.7084, -1.0877, -0.9574, 0.0000, -0.0000],\n",
" [-0.0000, 1.6664, -1.6006, 0.0000, -0.0000, -0.9844, -0.0000, -0.0521],\n",
" [ 2.4249, 0.0000, -0.0000, -0.0000, 0.0000, -0.0000, 2.6460, 0.0000]],\n",
" grad_fn=<MulBackward0>)"
]
},
"execution_count": 9,
"execution_count": 7,
"metadata": {},
"output_type": "execute_result"
}
......@@ -202,7 +202,7 @@
},
{
"cell_type": "code",
"execution_count": 10,
"execution_count": 8,
"metadata": {},
"outputs": [],
"source": [
......@@ -221,7 +221,7 @@
},
{
"cell_type": "code",
"execution_count": 12,
"execution_count": 9,
"metadata": {},
"outputs": [],
"source": [
......@@ -230,7 +230,7 @@
},
{
"cell_type": "code",
"execution_count": 13,
"execution_count": 10,
"metadata": {},
"outputs": [],
"source": [
......@@ -239,7 +239,7 @@
},
{
"cell_type": "code",
"execution_count": 14,
"execution_count": 11,
"metadata": {},
"outputs": [
{
......@@ -251,7 +251,7 @@
")"
]
},
"execution_count": 14,
"execution_count": 11,
"metadata": {},
"output_type": "execute_result"
}
......@@ -269,7 +269,7 @@
},
{
"cell_type": "code",
"execution_count": 15,
"execution_count": 12,
"metadata": {},
"outputs": [],
"source": [
......@@ -279,7 +279,7 @@
},
{
"cell_type": "code",
"execution_count": 16,
"execution_count": 13,
"metadata": {},
"outputs": [
{
......@@ -299,7 +299,7 @@
")"
]
},
"execution_count": 16,
"execution_count": 13,
"metadata": {},
"output_type": "execute_result"
}
......@@ -326,7 +326,7 @@
},
{
"cell_type": "code",
"execution_count": 18,
"execution_count": 14,
"metadata": {},
"outputs": [],
"source": [
......@@ -344,7 +344,7 @@
},
{
"cell_type": "code",
"execution_count": 20,
"execution_count": 15,
"metadata": {},
"outputs": [],
"source": [
......@@ -353,7 +353,7 @@
},
{
"cell_type": "code",
"execution_count": 21,
"execution_count": 16,
"metadata": {},
"outputs": [],
"source": [
......@@ -362,7 +362,7 @@
},
{
"cell_type": "code",
"execution_count": 22,
"execution_count": 17,
"metadata": {},
"outputs": [
{
......@@ -467,7 +467,7 @@
")"
]
},
"execution_count": 22,
"execution_count": 17,
"metadata": {},
"output_type": "execute_result"
}
......@@ -478,19 +478,19 @@
},
{
"cell_type": "code",
"execution_count": 23,
"execution_count": 18,
"metadata": {},
"outputs": [
{
"data": {
"text/plain": [
"tensor([[-1.7624e-03, 8.7372e-03, -3.9061e-03, 2.2110e-01, -1.9655e-03,\n",
" 7.9629e-02, 2.5455e-01, -1.6910e-03],\n",
" [-2.8030e-05, 1.6680e-01, -3.5123e-03, 1.5065e-01, 5.2558e-02,\n",
" 5.4472e-02, 4.2029e-02, 2.3403e-02]], grad_fn=<LeakyReluBackward1>)"
"tensor([[ 8.4865e-02, -3.4401e-03, -9.1973e-04, 3.4269e-01, 3.2816e-02,\n",
" 1.9682e-02, -8.0740e-04, 9.4898e-03],\n",
" [ 1.5473e-01, -6.2664e-03, -9.3413e-05, 3.8768e-01, -1.9963e-03,\n",
" 1.1729e-01, -2.7111e-03, 1.8670e-01]], grad_fn=<LeakyReluBackward1>)"
]
},
"execution_count": 23,
"execution_count": 18,
"metadata": {},
"output_type": "execute_result"
}
......@@ -508,7 +508,7 @@
},
{
"cell_type": "code",
"execution_count": 24,
"execution_count": 19,
"metadata": {},
"outputs": [],
"source": [
......@@ -587,7 +587,7 @@
},
{
"cell_type": "code",
"execution_count": 26,
"execution_count": 20,
"metadata": {},
"outputs": [],
"source": [
......
examples/03_Binary_Classification_with_Defaults.ipynb
浏览文件 @ 263663e5
......@@ -387,8 +387,7 @@
"source": [
"wide_cols = ['education', 'relationship','workclass','occupation','native_country','gender']\n",
"crossed_cols = [('education', 'occupation'), ('native_country', 'occupation')]\n",
"cat_embed_cols = [('education',16), ('relationship',8), ('workclass',16),\n",
" ('occupation',16),('native_country',16)]\n",
"cat_embed_cols = [('education',16), ('relationship',8), ('workclass',16), ('occupation',16),('native_country',16)]\n",
"continuous_cols = [\"age\",\"hours_per_week\"]\n",
"target_col = 'income_label'"
]
......@@ -592,16 +591,16 @@
"name": "stderr",
"output_type": "stream",
"text": [
"epoch 1: 100%|██████████| 153/153 [00:02<00:00, 55.95it/s, loss=0.41, metrics={'acc': 0.811}] \n",
"valid: 100%|██████████| 39/39 [00:00<00:00, 119.88it/s, loss=0.362, metrics={'acc': 0.8152}]\n",
"epoch 2: 100%|██████████| 153/153 [00:02<00:00, 58.75it/s, loss=0.35, metrics={'acc': 0.8347}] \n",
"valid: 100%|██████████| 39/39 [00:00<00:00, 99.19it/s, loss=0.353, metrics={'acc': 0.8352}]\n",
"epoch 3: 100%|██████████| 153/153 [00:02<00:00, 59.79it/s, loss=0.344, metrics={'acc': 0.8372}]\n",
"valid: 100%|██████████| 39/39 [00:00<00:00, 113.76it/s, loss=0.349, metrics={'acc': 0.8376}]\n",
"epoch 4: 100%|██████████| 153/153 [00:02<00:00, 56.89it/s, loss=0.341, metrics={'acc': 0.8389}]\n",
"valid: 100%|██████████| 39/39 [00:00<00:00, 104.95it/s, loss=0.347, metrics={'acc': 0.8388}]\n",
"epoch 5: 100%|██████████| 153/153 [00:02<00:00, 56.49it/s, loss=0.338, metrics={'acc': 0.8404}]\n",
"valid: 100%|██████████| 39/39 [00:00<00:00, 105.79it/s, loss=0.346, metrics={'acc': 0.8405}]\n"
"epoch 1: 100%|██████████| 153/153 [00:02<00:00, 56.52it/s, loss=0.412, metrics={'acc': 0.7993}]\n",
"valid: 100%|██████████| 39/39 [00:00<00:00, 123.12it/s, loss=0.352, metrics={'acc': 0.8071}]\n",
"epoch 2: 100%|██████████| 153/153 [00:02<00:00, 59.55it/s, loss=0.351, metrics={'acc': 0.8351}]\n",
"valid: 100%|██████████| 39/39 [00:00<00:00, 121.98it/s, loss=0.346, metrics={'acc': 0.8359}]\n",
"epoch 3: 100%|██████████| 153/153 [00:02<00:00, 59.82it/s, loss=0.346, metrics={'acc': 0.8377}]\n",
"valid: 100%|██████████| 39/39 [00:00<00:00, 121.88it/s, loss=0.344, metrics={'acc': 0.8384}]\n",
"epoch 4: 100%|██████████| 153/153 [00:02<00:00, 58.97it/s, loss=0.342, metrics={'acc': 0.8392}]\n",
"valid: 100%|██████████| 39/39 [00:00<00:00, 122.20it/s, loss=0.342, metrics={'acc': 0.84}] \n",
"epoch 5: 100%|██████████| 153/153 [00:02<00:00, 58.28it/s, loss=0.34, metrics={'acc': 0.8406}] \n",
"valid: 100%|██████████| 39/39 [00:00<00:00, 116.57it/s, loss=0.341, metrics={'acc': 0.8413}]\n"
]
}
],
......
examples/04_Binary_Classification_Varying_Parameters.ipynb
浏览文件 @ 263663e5
......@@ -381,14 +381,13 @@
},
{
"cell_type": "code",
"execution_count": 8,
"execution_count": 4,
"metadata": {},
"outputs": [],
"source": [
"wide_cols = ['education', 'relationship','workclass','occupation','native_country','gender']\n",
"crossed_cols = [('education', 'occupation'), ('native_country', 'occupation')]\n",
"cat_embed_cols = [('education',16), ('relationship',8), ('workclass',16),\n",
" ('occupation',16),('native_country',16)]\n",
"cat_embed_cols = [('education',16), ('relationship',8), ('workclass',16), ('occupation',16),('native_country',16)]\n",
"continuous_cols = [\"age\",\"hours_per_week\"]\n",
"target_col = 'income_label'"
]
......@@ -485,7 +484,7 @@
},
{
"cell_type": "code",
"execution_count": 9,
"execution_count": 8,
"metadata": {},
"outputs": [],
"source": [
......@@ -500,7 +499,7 @@
},
{
"cell_type": "code",
"execution_count": 10,
"execution_count": 9,
"metadata": {},
"outputs": [
{
......@@ -540,7 +539,7 @@
")"
]
},
"execution_count": 10,
"execution_count": 9,
"metadata": {},
"output_type": "execute_result"
}
......@@ -565,7 +564,7 @@
},
{
"cell_type": "code",
"execution_count": 11,
"execution_count": 10,
"metadata": {},
"outputs": [],
"source": [
......@@ -576,7 +575,7 @@
},
{
"cell_type": "code",
"execution_count": 12,
"execution_count": 11,
"metadata": {},
"outputs": [],
"source": [
......@@ -597,7 +596,7 @@
},
{
"cell_type": "code",
"execution_count": 13,
"execution_count": 12,
"metadata": {},
"outputs": [],
"source": [
......@@ -612,7 +611,7 @@
},
{
"cell_type": "code",
"execution_count": 14,
"execution_count": 13,
"metadata": {},
"outputs": [],
"source": [
......@@ -624,7 +623,7 @@
},
{
"cell_type": "code",
"execution_count": 15,
"execution_count": 14,
"metadata": {},
"outputs": [
{
......@@ -646,26 +645,26 @@
"name": "stderr",
"output_type": "stream",
"text": [
"epoch 1: 100%|██████████| 153/153 [00:03<00:00, 44.53it/s, loss=0.783, metrics={'acc': 0.6151}]\n",
"valid: 100%|██████████| 39/39 [00:00<00:00, 116.60it/s, loss=0.469, metrics={'acc': 0.6493}]\n",
"epoch 2: 100%|██████████| 153/153 [00:03<00:00, 47.32it/s, loss=0.529, metrics={'acc': 0.7565}]\n",
"valid: 100%|██████████| 39/39 [00:00<00:00, 95.90it/s, loss=0.396, metrics={'acc': 0.7685}]\n",
"epoch 3: 100%|██████████| 153/153 [00:03<00:00, 46.55it/s, loss=0.457, metrics={'acc': 0.7907}]\n",
"valid: 100%|██████████| 39/39 [00:00<00:00, 116.63it/s, loss=0.372, metrics={'acc': 0.798}]\n",
"epoch 4: 100%|██████████| 153/153 [00:03<00:00, 49.69it/s, loss=0.421, metrics={'acc': 0.8038}]\n",
"valid: 100%|██████████| 39/39 [00:00<00:00, 128.04it/s, loss=0.366, metrics={'acc': 0.8091}]\n",
"epoch 5: 100%|██████████| 153/153 [00:03<00:00, 50.27it/s, loss=0.398, metrics={'acc': 0.815}] \n",
"valid: 100%|██████████| 39/39 [00:00<00:00, 131.81it/s, loss=0.36, metrics={'acc': 0.8188}]\n",
"epoch 6: 100%|██████████| 153/153 [00:03<00:00, 50.16it/s, loss=0.388, metrics={'acc': 0.817}] \n",
"valid: 100%|██████████| 39/39 [00:00<00:00, 130.33it/s, loss=0.36, metrics={'acc': 0.8204}]\n",
"epoch 7: 100%|██████████| 153/153 [00:03<00:00, 50.06it/s, loss=0.386, metrics={'acc': 0.8175}]\n",
"valid: 100%|██████████| 39/39 [00:00<00:00, 133.34it/s, loss=0.359, metrics={'acc': 0.8208}]\n",
"epoch 8: 100%|██████████| 153/153 [00:03<00:00, 50.43it/s, loss=0.387, metrics={'acc': 0.8189}]\n",
"valid: 100%|██████████| 39/39 [00:00<00:00, 124.92it/s, loss=0.359, metrics={'acc': 0.8221}]\n",
"epoch 9: 100%|██████████| 153/153 [00:03<00:00, 50.34it/s, loss=0.385, metrics={'acc': 0.8185}]\n",
"valid: 100%|██████████| 39/39 [00:00<00:00, 130.44it/s, loss=0.358, metrics={'acc': 0.8219}]\n",
"epoch 10: 100%|██████████| 153/153 [00:03<00:00, 50.29it/s, loss=0.384, metrics={'acc': 0.8191}]\n",
"valid: 100%|██████████| 39/39 [00:00<00:00, 129.58it/s, loss=0.358, metrics={'acc': 0.8225}]\n"
"epoch 1: 100%|██████████| 153/153 [00:03<00:00, 47.06it/s, loss=0.731, metrics={'acc': 0.6468}]\n",
"valid: 100%|██████████| 39/39 [00:00<00:00, 118.07it/s, loss=0.418, metrics={'acc': 0.6785}]\n",
"epoch 2: 100%|██████████| 153/153 [00:03<00:00, 49.72it/s, loss=0.51, metrics={'acc': 0.7637}] \n",
"valid: 100%|██████████| 39/39 [00:00<00:00, 114.38it/s, loss=0.376, metrics={'acc': 0.7765}]\n",
"epoch 3: 100%|██████████| 153/153 [00:03<00:00, 49.32it/s, loss=0.448, metrics={'acc': 0.7927}]\n",
"valid: 100%|██████████| 39/39 [00:00<00:00, 113.23it/s, loss=0.361, metrics={'acc': 0.8007}]\n",
"epoch 4: 100%|██████████| 153/153 [00:03<00:00, 48.23it/s, loss=0.413, metrics={'acc': 0.8079}]\n",
"valid: 100%|██████████| 39/39 [00:00<00:00, 118.17it/s, loss=0.355, metrics={'acc': 0.8132}]\n",
"epoch 5: 100%|██████████| 153/153 [00:03<00:00, 48.27it/s, loss=0.395, metrics={'acc': 0.8149}]\n",
"valid: 100%|██████████| 39/39 [00:00<00:00, 114.71it/s, loss=0.352, metrics={'acc': 0.8191}]\n",
"epoch 6: 100%|██████████| 153/153 [00:03<00:00, 48.97it/s, loss=0.387, metrics={'acc': 0.8157}]\n",
"valid: 100%|██████████| 39/39 [00:00<00:00, 116.88it/s, loss=0.352, metrics={'acc': 0.8199}]\n",
"epoch 7: 100%|██████████| 153/153 [00:03<00:00, 48.56it/s, loss=0.388, metrics={'acc': 0.8153}]\n",
"valid: 100%|██████████| 39/39 [00:00<00:00, 116.33it/s, loss=0.352, metrics={'acc': 0.8195}]\n",
"epoch 8: 100%|██████████| 153/153 [00:03<00:00, 48.54it/s, loss=0.383, metrics={'acc': 0.8184}]\n",
"valid: 100%|██████████| 39/39 [00:00<00:00, 93.05it/s, loss=0.351, metrics={'acc': 0.822}] \n",
"epoch 9: 100%|██████████| 153/153 [00:03<00:00, 45.71it/s, loss=0.385, metrics={'acc': 0.8196}]\n",
"valid: 100%|██████████| 39/39 [00:00<00:00, 110.48it/s, loss=0.351, metrics={'acc': 0.8229}]\n",
"epoch 10: 100%|██████████| 153/153 [00:03<00:00, 48.44it/s, loss=0.382, metrics={'acc': 0.8194}]\n",
"valid: 100%|██████████| 39/39 [00:00<00:00, 114.40it/s, loss=0.35, metrics={'acc': 0.8228}]\n"
]
}
],
......@@ -675,7 +674,7 @@
},
{
"cell_type": "code",
"execution_count": 16,
"execution_count": 15,
"metadata": {},
"outputs": [
{
......@@ -800,7 +799,7 @@
" 'zero_grad']"
]
},
"execution_count": 16,
"execution_count": 15,
"metadata": {},
"output_type": "execute_result"
}
......@@ -818,7 +817,7 @@
},
{
"cell_type": "code",
"execution_count": 17,
"execution_count": 16,
"metadata": {},
"outputs": [
{
......@@ -827,7 +826,7 @@
"[0, 1, 2, 3, 4, 5, 6, 7, 8, 9]"
]
},
"execution_count": 17,
"execution_count": 16,
"metadata": {},
"output_type": "execute_result"
}
......@@ -838,14 +837,14 @@
},
{
"cell_type": "code",
"execution_count": 18,
"execution_count": 17,
"metadata": {},
"outputs": [
{
"name": "stdout",
"output_type": "stream",
"text": [
"{'train_loss': [0.7826161832591287, 0.5294494130253012, 0.45743006565212424, 0.4206276263286865, 0.3982163554702709, 0.3881325295158461, 0.3862898593244989, 0.38681577603801404, 0.38500378529230755, 0.38388273743243], 'train_acc': [0.6151, 0.7565, 0.7907, 0.8038, 0.815, 0.817, 0.8175, 0.8189, 0.8185, 0.8191], 'val_loss': [0.4694176025879689, 0.3960292133001181, 0.37219820802028364, 0.3658289725963886, 0.3600605313594525, 0.35951805343994725, 0.35915129765486103, 0.3585702692851042, 0.3578468553530864, 0.3576407875770178], 'val_acc': [0.6493, 0.7685, 0.798, 0.8091, 0.8188, 0.8204, 0.8208, 0.8221, 0.8219, 0.8225]}\n"
"{'train_loss': [0.7313813343157176, 0.5101876866583731, 0.44813506724008545, 0.41332343941420513, 0.3945406624694276, 0.3871746306715448, 0.3884129401515512, 0.38312816230300206, 0.3847907395923839, 0.3817657043341718], 'train_acc': [0.6468, 0.7637, 0.7927, 0.8079, 0.8149, 0.8157, 0.8153, 0.8184, 0.8196, 0.8194], 'val_loss': [0.41844800649545133, 0.3759944920356457, 0.36132928041311413, 0.3554159953044011, 0.3523857922126085, 0.3518377657120044, 0.35156664175864977, 0.35120767278549, 0.35089820012068135, 0.35047405576094603], 'val_acc': [0.6785, 0.7765, 0.8007, 0.8132, 0.8191, 0.8199, 0.8195, 0.822, 0.8229, 0.8228]}\n"
]
}
],
......@@ -855,7 +854,7 @@
},
{
"cell_type": "code",
"execution_count": 19,
"execution_count": 18,
"metadata": {},
"outputs": [
{
......@@ -881,81 +880,79 @@
},
{
"cell_type": "code",
"execution_count": 20,
"execution_count": 19,
"metadata": {},
"outputs": [
{
"data": {
"text/plain": [
"{'11th': array([-1.08425401e-01, 5.09871461e-04, 1.25755548e-01, -1.20801523e-01,\n",
" -2.56043434e-01, -3.55644524e-02, -8.66190940e-02, -1.39202878e-01,\n",
" 1.11087626e-04, 4.54997361e-01, -2.31609955e-01, -1.36443637e-02,\n",
" 8.78131837e-02, -3.07353675e-01, -1.10240346e-02, 6.45920560e-02],\n",
" dtype=float32),\n",
" 'HS-grad': array([ 0.19832617, 0.12040217, -0.5314197 , 0.35005897, -0.15391229,\n",
" -0.22196807, 0.09345723, 0.06745315, 0.25015768, 0.08744714,\n",
" 0.24480642, -0.08957793, 0.27947524, -0.26326123, -0.19119193,\n",
" -0.10995993], dtype=float32),\n",
" 'Assoc-acdm': array([ 0.06525454, -0.2618052 , -0.09840333, 0.10541438, 0.33471954,\n",
" -0.04292247, 0.10712572, 0.34287837, -0.18687049, -0.13836485,\n",
" -0.1715912 , 0.15273218, -0.03476759, -0.07450581, 0.56081617,\n",
" 0.29201028], dtype=float32),\n",
" 'Some-college': array([-0.45491776, -0.17205039, 0.21580465, -0.2539856 , 0.02358766,\n",
" -0.05496917, -0.01120283, 0.09221312, -0.12831998, 0.17159238,\n",
" 0.196605 , -0.2090644 , -0.11193639, -0.18394227, -0.16056207,\n",
" 0.02444198], dtype=float32),\n",
" '10th': array([-0.5581912 , -0.20644131, 0.1300292 , -0.10135209, 0.4538276 ,\n",
" -0.27146348, 0.12652951, 0.5233289 , 0.01145706, -0.05667543,\n",
" 0.43509725, -0.74307233, 0.00139265, 0.07225899, 0.0781986 ,\n",
" -0.2610258 ], dtype=float32),\n",
" 'Prof-school': array([-6.5744489e-02, 1.3956554e-01, 5.7986474e-01, 2.7874210e-01,\n",
" -2.4446699e-01, 7.9873689e-02, -3.8569799e-01, 2.2757685e-01,\n",
" -3.8109139e-02, 3.3144853e-01, -3.8229354e-02, 2.9802489e-01,\n",
" -1.5467829e-01, 5.4805580e-04, -2.1627106e-01, -2.6592135e-02],\n",
" dtype=float32),\n",
" '7th-8th': array([ 0.10858492, 0.42190084, 0.07536066, -0.11707054, 0.05351719,\n",
" 0.32636967, 0.14053936, 0.45679298, -0.2558197 , -0.47910702,\n",
" 0.4725715 , -0.0981419 , 0.3462793 , 0.07776859, -0.45930195,\n",
" 0.12625834], dtype=float32),\n",
" 'Bachelors': array([ 0.01805384, -0.10573057, 0.25564098, -0.27709666, -0.16297452,\n",
" -0.1851758 , -0.5702467 , -0.23569717, 0.067039 , -0.28916818,\n",
" -0.22313781, -0.23893505, 0.37708414, 0.17465928, -0.47459307,\n",
" 0.04889947], dtype=float32),\n",
" 'Masters': array([ 0.11953138, 0.11543513, -0.3954705 , 0.32583147, 0.23851769,\n",
" -0.6448425 , 0.00705628, 0.10673986, -0.08305098, -0.10872949,\n",
" -0.46080047, -0.05367367, -0.18693425, 0.14182107, -0.39178014,\n",
" -0.23969549], dtype=float32),\n",
" 'Doctorate': array([ 0.04873321, -0.19027464, -0.10777274, -0.17476888, 0.47248197,\n",
" -0.2873778 , -0.29792303, -0.06811561, 0.16541322, -0.17425427,\n",
" -0.09404507, 0.06525683, 0.06408301, 0.38656166, 0.13369907,\n",
" 0.10825544], dtype=float32),\n",
" '5th-6th': array([ 0.08566641, 0.03589746, 0.17174615, 0.08747724, 0.2698885 ,\n",
" 0.08344392, -0.23652045, 0.31357667, 0.3546634 , -0.29814255,\n",
" 0.10943606, 0.45218074, -0.0614133 , -0.31987205, 0.34947518,\n",
" 0.07603104], dtype=float32),\n",
" 'Assoc-voc': array([-0.07830544, 0.0278313 , 0.34295908, -0.27213913, -0.20097388,\n",
" 0.10972344, 0.14000823, -0.24098383, -0.16614872, 0.19084413,\n",
" -0.02334382, 0.5209352 , 0.24089335, -0.1350642 , -0.23480216,\n",
" -0.32963687], dtype=float32),\n",
" '9th': array([ 0.12994888, 0.02475524, -0.12875263, 0.0097373 , 0.38253692,\n",
" -0.2718543 , 0.13766348, 0.27845392, -0.2036348 , -0.20567507,\n",
" -0.11305337, -0.47028974, 0.07009655, -0.29621345, -0.17303236,\n",
" 0.15854478], dtype=float32),\n",
" '12th': array([-0.15079321, -0.26879913, -0.5159767 , 0.30044943, 0.0295292 ,\n",
" -0.32494095, 0.20975012, 0.35193697, -0.5034315 , -0.14420179,\n",
" 0.06113023, 0.22398257, 0.0087006 , 0.09041765, -0.09754901,\n",
" -0.21647781], dtype=float32),\n",
" '1st-4th': array([-0.3199786 , 0.10094872, -0.10035568, 0.10014401, -0.09340642,\n",
" -0.00761677, 0.50759906, 0.288856 , -0.18745485, 0.05442255,\n",
" 0.6481828 , 0.18515776, 0.21597311, -0.21534163, 0.01798662,\n",
" -0.22816893], dtype=float32),\n",
" 'Preschool': array([ 0.10035816, -0.24015287, 0.00935481, 0.05356123, -0.18744251,\n",
" -0.39735606, 0.03849271, -0.2864288 , -0.10379744, 0.20251973,\n",
" 0.14565234, -0.24607188, -0.14268415, 0.1209868 , 0.04662501,\n",
" 0.41015574], dtype=float32)}"
"{'11th': array([-0.04807916, 0.21404432, 0.12517522, -0.154123 , 0.06864536,\n",
" 0.00092955, -0.38516527, -0.18440197, 0.15861034, 0.12012056,\n",
" 0.55413646, -0.16920644, 0.1356924 , -0.37921003, 0.53833497,\n",
" 0.08743049], dtype=float32),\n",
" 'HS-grad': array([ 0.37504154, 0.34191516, 0.27299362, 0.22921972, 0.07420117,\n",
" 0.34922913, 0.19239122, -0.42343035, -0.845824 , -0.07287297,\n",
" 0.27455565, 0.19505064, 0.07062761, -0.5201107 , 0.37823108,\n",
" 0.46134958], dtype=float32),\n",
" 'Assoc-acdm': array([ 0.22331461, 0.15005238, 0.13472553, -0.16886246, -0.12053325,\n",
" -0.04233408, -0.08905135, -0.54481906, 0.24300168, -0.21069968,\n",
" -0.00685616, -0.38423738, -0.00281451, 0.10599079, -0.05224385,\n",
" 0.2891064 ], dtype=float32),\n",
" 'Some-college': array([-0.09498356, -0.16801773, -0.09181987, 0.05381393, -0.03607363,\n",
" -0.05759075, 0.09382061, 0.33274302, -0.11906563, 0.14481838,\n",
" -0.1765725 , 0.20070277, 0.2960993 , -0.02055654, 0.26645136,\n",
" 0.4075843 ], dtype=float32),\n",
" '10th': array([-0.12961714, -0.27546212, 0.24345328, -0.24318363, 0.31552687,\n",
" 0.16653115, -0.05234893, 0.06825106, 0.2388588 , 0.10887478,\n",
" -0.12004007, -0.00373614, -0.0223387 , 0.133562 , 0.29672143,\n",
" 0.03046475], dtype=float32),\n",
" 'Prof-school': array([-0.1589678 , -0.07629952, 0.00763621, 0.13788143, 0.4114019 ,\n",
" 0.07717889, -0.17072953, 0.29419565, -0.18929462, -0.09182461,\n",
" -0.08409152, 0.01395322, -0.20351669, 0.18333136, -0.03983613,\n",
" -0.31888708], dtype=float32),\n",
" '7th-8th': array([ 0.39654806, 0.26095334, -0.3147828 , -0.41267306, -0.23983437,\n",
" -0.08034727, 0.4807234 , 0.3054779 , -0.3085564 , -0.07860225,\n",
" -0.1279486 , -0.2846014 , 0.1358583 , 0.24006395, -0.18911272,\n",
" -0.2299538 ], dtype=float32),\n",
" 'Bachelors': array([ 0.35242578, -0.03246311, 0.15835243, -0.06434399, 0.03403192,\n",
" 0.0088449 , 0.00627425, -0.31485453, -0.30984947, -0.23008366,\n",
" -0.09467663, 0.17246258, -0.09432375, 0.07691337, 0.70925283,\n",
" 0.18795769], dtype=float32),\n",
" 'Masters': array([-0.14503758, 0.0048258 , 0.58242404, 0.28511924, -0.13773848,\n",
" 0.35109136, 0.05824559, 0.3609631 , 0.4700086 , 0.4251728 ,\n",
" -0.2538366 , -0.00297809, 0.1424264 , -0.12481072, -0.09403807,\n",
" 0.00634856], dtype=float32),\n",
" 'Doctorate': array([-0.12487873, -0.1699961 , 0.2220065 , -0.04808738, 0.09443628,\n",
" -0.21019349, -0.23745097, 0.28523713, 0.05516997, -0.04004707,\n",
" 0.3316393 , 0.18710822, 0.4153885 , -0.12905155, 0.03055826,\n",
" 0.0664137 ], dtype=float32),\n",
" '5th-6th': array([ 0.21891987, -0.13600409, -0.03123563, 0.16288632, -0.03479639,\n",
" -0.4221951 , 0.4688111 , 0.08145971, -0.29254073, 0.18396533,\n",
" -0.20204993, -0.03327556, -0.2558647 , 0.56448 , -0.30299884,\n",
" 0.07629355], dtype=float32),\n",
" 'Assoc-voc': array([-0.01987046, -0.06434393, 0.00226 , 0.08150155, -0.33775425,\n",
" -0.13507745, 0.12741297, 0.0542295 , 0.09895965, 0.067229 ,\n",
" -0.1718493 , 0.01054914, 0.10441845, -0.18814586, -0.01663602,\n",
" 0.03088147], dtype=float32),\n",
" '9th': array([-0.24095939, 0.2750888 , 0.01418325, -0.36754113, 0.5431856 ,\n",
" -0.19582956, 0.03485603, 0.22838333, -0.05723334, 0.10631263,\n",
" 0.06331363, -0.09572615, 0.21977316, -0.02579625, -0.13822857,\n",
" 0.28736743], dtype=float32),\n",
" '12th': array([-0.20278502, -0.19245535, -0.04846343, 0.14459866, 0.25858438,\n",
" 0.15333128, 0.5074635 , -0.15141617, -0.19331448, -0.2630267 ,\n",
" -0.1378872 , -0.16868882, 0.4048257 , -0.34108582, -0.23098588,\n",
" 0.2859633 ], dtype=float32),\n",
" '1st-4th': array([-0.53678703, 0.19669479, -0.18026853, 0.33791658, 0.14260627,\n",
" 0.20269199, 0.00518189, 0.01120056, 0.01568659, 0.28752655,\n",
" 0.3359768 , 0.01758064, 0.11630564, -0.35470524, -0.05704446,\n",
" 0.41216984], dtype=float32),\n",
" 'Preschool': array([ 0.10326536, -0.02895411, 0.11348445, 0.03685748, 0.55893034,\n",
" -0.2522173 , -0.07186767, -0.30955225, -0.17825711, 0.02907414,\n",
" -0.61121726, 0.40596214, 0.63471395, 0.3304132 , 0.05272925,\n",
" -0.4266447 ], dtype=float32)}"
]
},
"execution_count": 20,
"execution_count": 19,
"metadata": {},
"output_type": "execute_result"
}
......
examples/05_Regression_with_Images_and_Text.ipynb
浏览文件 @ 263663e5
......@@ -1022,7 +1022,7 @@
},
{
"cell_type": "code",
"execution_count": 7,
"execution_count": 10,
"metadata": {},
"outputs": [
{
......@@ -1038,13 +1038,13 @@
}
],
"source": [
"text_preprocessor = TextPreprocessor(word_vectors_path=word_vectors_path)\n",
"X_text = text_preprocessor.fit_transform(df, text_col)"
"text_preprocessor = TextPreprocessor(word_vectors_path=word_vectors_path, text_col=text_col)\n",
"X_text = text_preprocessor.fit_transform(df)"
]
},
{
"cell_type": "code",
"execution_count": 8,
"execution_count": 11,
"metadata": {},
"outputs": [
{
......@@ -1058,7 +1058,7 @@
"name": "stderr",
"output_type": "stream",
"text": [
" 4%|▍ | 43/1001 [00:00<00:02, 428.36it/s]"
" 9%|▊ | 87/1001 [00:00<00:02, 428.89it/s]"
]
},
{
......@@ -1072,7 +1072,7 @@
"name": "stderr",
"output_type": "stream",
"text": [
"100%|██████████| 1001/1001 [00:02<00:00, 425.98it/s]\n"
"100%|██████████| 1001/1001 [00:02<00:00, 426.92it/s]\n"
]
},
{
......@@ -1084,8 +1084,8 @@
}
],
"source": [
"image_processor = ImagePreprocessor()\n",
"X_images = image_processor.fit_transform(df, img_col, img_path)"
"image_processor = ImagePreprocessor(img_col = img_col, img_path = img_path)\n",
"X_images = image_processor.fit_transform(df)"
]
},
{
......@@ -1097,7 +1097,7 @@
},
{
"cell_type": "code",
"execution_count": 9,
"execution_count": 12,
"metadata": {},
"outputs": [],
"source": [
......@@ -1125,7 +1125,7 @@
},
{
"cell_type": "code",
"execution_count": 10,
"execution_count": 13,
"metadata": {},
"outputs": [],
"source": [
......@@ -1141,7 +1141,7 @@
},
{
"cell_type": "code",
"execution_count": 13,
"execution_count": 14,
"metadata": {},
"outputs": [],
"source": [
......@@ -1150,7 +1150,7 @@
},
{
"cell_type": "code",
"execution_count": 14,
"execution_count": 15,
"metadata": {},
"outputs": [
{
......@@ -1172,8 +1172,8 @@
"name": "stderr",
"output_type": "stream",
"text": [
"epoch 1: 100%|██████████| 25/25 [02:04<00:00, 4.97s/it, loss=117]\n",
"valid: 100%|██████████| 7/7 [00:14<00:00, 2.00s/it, loss=122]\n"
"epoch 1: 100%|██████████| 25/25 [02:06<00:00, 5.05s/it, loss=118]\n",
"valid: 100%|██████████| 7/7 [00:14<00:00, 2.01s/it, loss=226]\n"
]
}
],
......@@ -1193,7 +1193,7 @@
},
{
"cell_type": "code",
"execution_count": 15,
"execution_count": 16,
"metadata": {},
"outputs": [],
"source": [
......@@ -1217,7 +1217,7 @@
},
{
"cell_type": "code",
"execution_count": 16,
"execution_count": 17,
"metadata": {},
"outputs": [],
"source": [
......@@ -1233,7 +1233,7 @@
},
{
"cell_type": "code",
"execution_count": 17,
"execution_count": 18,
"metadata": {},
"outputs": [
{
......@@ -1386,7 +1386,7 @@
")"
]
},
"execution_count": 17,
"execution_count": 18,
"metadata": {},
"output_type": "execute_result"
}
......@@ -1406,7 +1406,7 @@
},
{
"cell_type": "code",
"execution_count": 18,
"execution_count": 19,
"metadata": {},
"outputs": [],
"source": [
......@@ -1420,7 +1420,7 @@
},
{
"cell_type": "code",
"execution_count": 19,
"execution_count": 20,
"metadata": {},
"outputs": [],
"source": [
......@@ -1433,7 +1433,7 @@
},
{
"cell_type": "code",
"execution_count": 20,
"execution_count": 21,
"metadata": {},
"outputs": [],
"source": [
......@@ -1446,7 +1446,7 @@
},
{
"cell_type": "code",
"execution_count": 21,
"execution_count": 22,
"metadata": {},
"outputs": [],
"source": [
......@@ -1472,15 +1472,15 @@
},
{
"cell_type": "code",
"execution_count": 22,
"execution_count": 23,
"metadata": {},
"outputs": [
{
"name": "stderr",
"output_type": "stream",
"text": [
"/Users/javier/pytorch-widedeep/pytorch_widedeep/initializers.py:32: UserWarning: No initializer found for deephead\n",
" if self.verbose: warnings.warn(\"No initializer found for {}\".format(name))\n"
"/Users/javier/pytorch-widedeep/pytorch_widedeep/initializers.py:31: UserWarning: No initializer found for deephead\n",
" warnings.warn(\"No initializer found for {}\".format(name))\n"
]
}
],
......@@ -1491,7 +1491,7 @@
},
{
"cell_type": "code",
"execution_count": 23,
"execution_count": 24,
"metadata": {},
"outputs": [
{
......@@ -1644,7 +1644,7 @@
")"
]
},
"execution_count": 23,
"execution_count": 24,
"metadata": {},
"output_type": "execute_result"
}
......@@ -1655,7 +1655,7 @@
},
{
"cell_type": "code",
"execution_count": 24,
"execution_count": 25,
"metadata": {},
"outputs": [
{
......@@ -1677,8 +1677,8 @@
"name": "stderr",
"output_type": "stream",
"text": [
"epoch 1: 100%|██████████| 25/25 [02:08<00:00, 5.15s/it, loss=128]\n",
"valid: 100%|██████████| 7/7 [00:14<00:00, 2.05s/it, loss=95.5]\n"
"epoch 1: 100%|██████████| 25/25 [02:04<00:00, 4.97s/it, loss=127]\n",
"valid: 100%|██████████| 7/7 [00:14<00:00, 2.02s/it, loss=94] \n"
]
}
],
......@@ -1696,7 +1696,7 @@
},
{
"cell_type": "code",
"execution_count": 25,
"execution_count": 26,
"metadata": {},
"outputs": [
{
......@@ -1721,7 +1721,7 @@
" 'lr_deephead_0': [0.001, 0.001]}"
]
},
"execution_count": 25,
"execution_count": 26,
"metadata": {},
"output_type": "execute_result"
}
......
examples/06_WarmUp_Model_Components.ipynb
浏览文件 @ 263663e5
......@@ -124,7 +124,7 @@
},
{
"cell_type": "code",
"execution_count": 8,
"execution_count": 7,
"metadata": {},
"outputs": [
{
......@@ -146,11 +146,11 @@
"name": "stderr",
"output_type": "stream",
"text": [
"epoch 1: 100%|██████████| 153/153 [00:01<00:00, 133.55it/s, loss=0.469, metrics={'acc': 0.7963}]\n",
"epoch 2: 100%|██████████| 153/153 [00:00<00:00, 156.42it/s, loss=0.372, metrics={'acc': 0.8119}]\n",
"epoch 3: 100%|██████████| 153/153 [00:01<00:00, 152.65it/s, loss=0.363, metrics={'acc': 0.8183}]\n",
"epoch 4: 100%|██████████| 153/153 [00:01<00:00, 144.01it/s, loss=0.36, metrics={'acc': 0.8218}] \n",
"epoch 5: 100%|██████████| 153/153 [00:00<00:00, 155.56it/s, loss=0.359, metrics={'acc': 0.8241}]\n",
"epoch 1: 100%|██████████| 153/153 [00:01<00:00, 131.88it/s, loss=0.471, metrics={'acc': 0.7946}]\n",
"epoch 2: 100%|██████████| 153/153 [00:00<00:00, 154.81it/s, loss=0.373, metrics={'acc': 0.8115}]\n",
"epoch 3: 100%|██████████| 153/153 [00:01<00:00, 151.56it/s, loss=0.364, metrics={'acc': 0.8182}]\n",
"epoch 4: 100%|██████████| 153/153 [00:00<00:00, 154.22it/s, loss=0.362, metrics={'acc': 0.8216}]\n",
"epoch 5: 100%|██████████| 153/153 [00:01<00:00, 152.65it/s, loss=0.36, metrics={'acc': 0.8238}] \n",
" 0%| | 0/153 [00:00<?, ?it/s]"
]
},
......@@ -165,11 +165,11 @@
"name": "stderr",
"output_type": "stream",
"text": [
"epoch 1: 100%|██████████| 153/153 [00:02<00:00, 64.03it/s, loss=0.4, metrics={'acc': 0.8214}] \n",
"epoch 2: 100%|██████████| 153/153 [00:02<00:00, 64.30it/s, loss=0.347, metrics={'acc': 0.8237}]\n",
"epoch 3: 100%|██████████| 153/153 [00:02<00:00, 64.18it/s, loss=0.341, metrics={'acc': 0.8258}]\n",
"epoch 4: 100%|██████████| 153/153 [00:02<00:00, 64.98it/s, loss=0.337, metrics={'acc': 0.8277}]\n",
"epoch 5: 100%|██████████| 153/153 [00:02<00:00, 65.70it/s, loss=0.333, metrics={'acc': 0.8294}]\n",
"epoch 1: 100%|██████████| 153/153 [00:02<00:00, 64.04it/s, loss=0.395, metrics={'acc': 0.8222}]\n",
"epoch 2: 100%|██████████| 153/153 [00:02<00:00, 65.34it/s, loss=0.349, metrics={'acc': 0.8242}]\n",
"epoch 3: 100%|██████████| 153/153 [00:02<00:00, 65.05it/s, loss=0.343, metrics={'acc': 0.8262}]\n",
"epoch 4: 100%|██████████| 153/153 [00:02<00:00, 64.93it/s, loss=0.339, metrics={'acc': 0.8279}]\n",
"epoch 5: 100%|██████████| 153/153 [00:02<00:00, 65.15it/s, loss=0.335, metrics={'acc': 0.8295}]\n",
" 0%| | 0/153 [00:00<?, ?it/s]"
]
},
......@@ -184,16 +184,16 @@
"name": "stderr",
"output_type": "stream",
"text": [
"epoch 1: 100%|██████████| 153/153 [00:02<00:00, 58.97it/s, loss=0.343, metrics={'acc': 0.8442}]\n",
"valid: 100%|██████████| 39/39 [00:00<00:00, 123.14it/s, loss=0.349, metrics={'acc': 0.8436}]\n",
"epoch 2: 100%|██████████| 153/153 [00:02<00:00, 59.48it/s, loss=0.333, metrics={'acc': 0.8457}]\n",
"valid: 100%|██████████| 39/39 [00:00<00:00, 114.05it/s, loss=0.346, metrics={'acc': 0.8447}]\n",
"epoch 3: 100%|██████████| 153/153 [00:02<00:00, 59.82it/s, loss=0.331, metrics={'acc': 0.8471}]\n",
"valid: 100%|██████████| 39/39 [00:00<00:00, 122.31it/s, loss=0.345, metrics={'acc': 0.8457}]\n",
"epoch 4: 100%|██████████| 153/153 [00:02<00:00, 59.20it/s, loss=0.329, metrics={'acc': 0.8474}]\n",
"valid: 100%|██████████| 39/39 [00:00<00:00, 113.66it/s, loss=0.345, metrics={'acc': 0.8459}]\n",
"epoch 5: 100%|██████████| 153/153 [00:02<00:00, 59.15it/s, loss=0.328, metrics={'acc': 0.8479}]\n",
"valid: 100%|██████████| 39/39 [00:00<00:00, 118.58it/s, loss=0.345, metrics={'acc': 0.8462}]\n"
"epoch 1: 100%|██████████| 153/153 [00:02<00:00, 58.31it/s, loss=0.345, metrics={'acc': 0.8415}]\n",
"valid: 100%|██████████| 39/39 [00:00<00:00, 120.77it/s, loss=0.346, metrics={'acc': 0.8416}]\n",
"epoch 2: 100%|██████████| 153/153 [00:02<00:00, 58.33it/s, loss=0.335, metrics={'acc': 0.8446}]\n",
"valid: 100%|██████████| 39/39 [00:00<00:00, 117.88it/s, loss=0.344, metrics={'acc': 0.8438}]\n",
"epoch 3: 100%|██████████| 153/153 [00:02<00:00, 58.43it/s, loss=0.331, metrics={'acc': 0.8457}]\n",
"valid: 100%|██████████| 39/39 [00:00<00:00, 109.26it/s, loss=0.343, metrics={'acc': 0.8449}]\n",
"epoch 4: 100%|██████████| 153/153 [00:02<00:00, 58.08it/s, loss=0.329, metrics={'acc': 0.8457}]\n",
"valid: 100%|██████████| 39/39 [00:00<00:00, 120.23it/s, loss=0.344, metrics={'acc': 0.8446}]\n",
"epoch 5: 100%|██████████| 153/153 [00:02<00:00, 58.75it/s, loss=0.327, metrics={'acc': 0.8464}]\n",
"valid: 100%|██████████| 39/39 [00:00<00:00, 119.22it/s, loss=0.344, metrics={'acc': 0.8453}]\n"
]
}
],
......@@ -233,7 +233,7 @@
},
{
"cell_type": "code",
"execution_count": 2,
"execution_count": 8,
"metadata": {},
"outputs": [],
"source": [
......@@ -243,7 +243,7 @@
},
{
"cell_type": "code",
"execution_count": 3,
"execution_count": 9,
"metadata": {},
"outputs": [],
"source": [
......@@ -270,7 +270,7 @@
},
{
"cell_type": "code",
"execution_count": 4,
"execution_count": 11,
"metadata": {},
"outputs": [
{
......@@ -289,7 +289,7 @@
"name": "stderr",
"output_type": "stream",
"text": [
" 8%|▊ | 84/1001 [00:00<00:02, 413.82it/s]"
" 3%|▎ | 29/1001 [00:00<00:03, 282.66it/s]"
]
},
{
......@@ -303,7 +303,7 @@
"name": "stderr",
"output_type": "stream",
"text": [
"100%|██████████| 1001/1001 [00:02<00:00, 419.78it/s]\n"
"100%|██████████| 1001/1001 [00:03<00:00, 327.44it/s]\n"
]
},
{
......@@ -323,16 +323,16 @@
"prepare_deep = DeepPreprocessor(embed_cols=cat_embed_cols, continuous_cols=continuous_cols)\n",
"X_deep = prepare_deep.fit_transform(df)\n",
"\n",
"text_processor = TextPreprocessor(word_vectors_path=word_vectors_path)\n",
"X_text = text_processor.fit_transform(df, text_col)\n",
"text_processor = TextPreprocessor(word_vectors_path=word_vectors_path, text_col=text_col)\n",
"X_text = text_processor.fit_transform(df)\n",
"\n",
"image_processor = ImagePreprocessor()\n",
"X_images = image_processor.fit_transform(df, img_col, img_path)"
"image_processor = ImagePreprocessor(img_col=img_col, img_path=img_path)\n",
"X_images = image_processor.fit_transform(df)"
]
},
{
"cell_type": "code",
"execution_count": 7,
"execution_count": 12,
"metadata": {},
"outputs": [],
"source": [
......@@ -551,7 +551,7 @@
},
{
"cell_type": "code",
"execution_count": 9,
"execution_count": 13,
"metadata": {},
"outputs": [
{
......@@ -632,7 +632,7 @@
" ), Linear(in_features=512, out_features=1, bias=True)]"
]
},
"execution_count": 9,
"execution_count": 13,
"metadata": {},
"output_type": "execute_result"
}
......@@ -652,7 +652,7 @@
},
{
"cell_type": "code",
"execution_count": 10,
"execution_count": 14,
"metadata": {},
"outputs": [],
"source": [
......@@ -661,7 +661,7 @@
},
{
"cell_type": "code",
"execution_count": 11,
"execution_count": 15,
"metadata": {},
"outputs": [
{
......@@ -683,7 +683,7 @@
"name": "stderr",
"output_type": "stream",
"text": [
"epoch 1: 100%|██████████| 25/25 [00:00<00:00, 42.89it/s, loss=127]\n",
"epoch 1: 100%|██████████| 25/25 [00:00<00:00, 57.98it/s, loss=127]\n",
" 0%| | 0/25 [00:00<?, ?it/s]"
]
},
......@@ -698,7 +698,7 @@
"name": "stderr",
"output_type": "stream",
"text": [
"epoch 1: 100%|██████████| 25/25 [00:00<00:00, 43.28it/s, loss=117]\n",
"epoch 1: 100%|██████████| 25/25 [00:00<00:00, 45.81it/s, loss=116]\n",
" 0%| | 0/25 [00:00<?, ?it/s]"
]
},
......@@ -713,7 +713,7 @@
"name": "stderr",
"output_type": "stream",
"text": [
"epoch 1: 100%|██████████| 25/25 [00:05<00:00, 4.78it/s, loss=132]\n",
"epoch 1: 100%|██████████| 25/25 [00:04<00:00, 5.37it/s, loss=132]\n",
" 0%| | 0/25 [00:00<?, ?it/s]"
]
},
......@@ -728,7 +728,7 @@
"name": "stderr",
"output_type": "stream",
"text": [
"epoch 1: 100%|██████████| 25/25 [01:11<00:00, 2.86s/it, loss=119]\n",
"epoch 1: 100%|██████████| 25/25 [01:10<00:00, 2.83s/it, loss=119]\n",
" 0%| | 0/25 [00:00<?, ?it/s]"
]
},
......@@ -743,7 +743,7 @@
"name": "stderr",
"output_type": "stream",
"text": [
"epoch 1: 100%|██████████| 25/25 [01:35<00:00, 3.81s/it, loss=108]\n",
"epoch 1: 100%|██████████| 25/25 [01:34<00:00, 3.76s/it, loss=108]\n",
" 0%| | 0/25 [00:00<?, ?it/s]"
]
},
......@@ -758,7 +758,7 @@
"name": "stderr",
"output_type": "stream",
"text": [
"epoch 1: 100%|██████████| 25/25 [01:58<00:00, 4.76s/it, loss=105]\n",
"epoch 1: 100%|██████████| 25/25 [01:57<00:00, 4.69s/it, loss=106]\n",
" 0%| | 0/25 [00:00<?, ?it/s]"
]
},
......@@ -773,7 +773,7 @@
"name": "stderr",
"output_type": "stream",
"text": [
"epoch 1: 100%|██████████| 25/25 [02:25<00:00, 5.80s/it, loss=105]\n",
"epoch 1: 100%|██████████| 25/25 [02:24<00:00, 5.79s/it, loss=105] \n",
" 0%| | 0/25 [00:00<?, ?it/s]"
]
},
......@@ -788,7 +788,7 @@
"name": "stderr",
"output_type": "stream",
"text": [
"epoch 1: 100%|██████████| 25/25 [03:04<00:00, 7.38s/it, loss=106] \n",
"epoch 1: 100%|██████████| 25/25 [03:01<00:00, 7.26s/it, loss=105] \n",
" 0%| | 0/25 [00:00<?, ?it/s]"
]
},
......@@ -803,8 +803,8 @@
"name": "stderr",
"output_type": "stream",
"text": [
"epoch 1: 100%|██████████| 25/25 [02:05<00:00, 5.00s/it, loss=130]\n",
"valid: 100%|██████████| 7/7 [00:14<00:00, 2.04s/it, loss=125] \n"
"epoch 1: 100%|██████████| 25/25 [02:05<00:00, 5.03s/it, loss=129]\n",
"valid: 100%|██████████| 7/7 [00:14<00:00, 2.11s/it, loss=103] \n"
]
}
],
......
examples/adult_script.py
浏览文件 @ 263663e5
......@@ -5,42 +5,57 @@ import torch
from pytorch_widedeep.preprocessing import WidePreprocessor, DeepPreprocessor
from pytorch_widedeep.models import Wide, DeepDense, WideDeep
from pytorch_widedeep.optim import RAdam
from pytorch_widedeep.initializers import *
from pytorch_widedeep.callbacks import *
from pytorch_widedeep.metrics import *
from pytorch_widedeep.initializers import KaimingNormal, XavierNormal
from pytorch_widedeep.callbacks import LRHistory, EarlyStopping, ModelCheckpoint
from pytorch_widedeep.metrics import BinaryAccuracy
use_cuda = torch.cuda.is_available()
if __name__ == '__main__':
if __name__ == "__main__":
df = pd.read_csv('data/adult/adult.csv.zip')
df = pd.read_csv("data/adult/adult.csv.zip")
df.columns = [c.replace("-", "_") for c in df.columns]
df['age_buckets'] = pd.cut(df.age, bins=[16, 25, 30, 35, 40, 45, 50, 55, 60, 91], labels=np.arange(9))
df['income_label'] = (df["income"].apply(lambda x: ">50K" in x)).astype(int)
df.drop('income', axis=1, inplace=True)
df["age_buckets"] = pd.cut(
df.age, bins=[16, 25, 30, 35, 40, 45, 50, 55, 60, 91], labels=np.arange(9)
)
df["income_label"] = (df["income"].apply(lambda x: ">50K" in x)).astype(int)
df.drop("income", axis=1, inplace=True)
df.head()
wide_cols = ['age_buckets', 'education', 'relationship','workclass','occupation',
'native_country','gender']
crossed_cols = [('education', 'occupation'), ('native_country', 'occupation')]
cat_embed_cols = [('education',10), ('relationship',8), ('workclass',10),
('occupation',10),('native_country',10)]
continuous_cols = ["age","hours_per_week"]
target = 'income_label'
wide_cols = [
"age_buckets",
"education",
"relationship",
"workclass",
"occupation",
"native_country",
"gender",
]
crossed_cols = [("education", "occupation"), ("native_country", "occupation")]
cat_embed_cols = [
("education", 10),
("relationship", 8),
("workclass", 10),
("occupation", 10),
("native_country", 10),
]
continuous_cols = ["age", "hours_per_week"]
target = "income_label"
target = df[target].values
prepare_wide = WidePreprocessor(wide_cols=wide_cols, crossed_cols=crossed_cols)
X_wide = prepare_wide.fit_transform(df)
prepare_deep = DeepPreprocessor(embed_cols=cat_embed_cols, continuous_cols=continuous_cols)
prepare_deep = DeepPreprocessor(
embed_cols=cat_embed_cols, continuous_cols=continuous_cols
)
X_deep = prepare_deep.fit_transform(df)
wide = Wide(
wide_dim=X_wide.shape[1],
output_dim=1)
wide = Wide(wide_dim=X_wide.shape[1], output_dim=1)
deepdense = DeepDense(
hidden_layers=[64,32],
dropout=[0.2,0.2],
hidden_layers=[64, 32],
dropout=[0.2, 0.2],
deep_column_idx=prepare_deep.deep_column_idx,
embed_input=prepare_deep.embeddings_input,
continuous_cols=continuous_cols)
continuous_cols=continuous_cols,
)
model = WideDeep(wide=wide, deepdense=deepdense)
wide_opt = torch.optim.Adam(model.wide.parameters())
......@@ -48,19 +63,24 @@ if __name__ == '__main__':
wide_sch = torch.optim.lr_scheduler.StepLR(wide_opt, step_size=3)
deep_sch = torch.optim.lr_scheduler.StepLR(deep_opt, step_size=5)
optimizers = {'wide': wide_opt, 'deepdense':deep_opt}
schedulers = {'wide': wide_sch, 'deepdense':deep_sch}
initializers = {'wide': KaimingNormal, 'deepdense':XavierNormal}
callbacks = [LRHistory(n_epochs=10), EarlyStopping, ModelCheckpoint(filepath='model_weights/wd_out')]
optimizers = {"wide": wide_opt, "deepdense": deep_opt}
schedulers = {"wide": wide_sch, "deepdense": deep_sch}
initializers = {"wide": KaimingNormal, "deepdense": XavierNormal}
callbacks = [
LRHistory(n_epochs=10),
EarlyStopping,
ModelCheckpoint(filepath="model_weights/wd_out"),
]
metrics = [BinaryAccuracy]
model.compile(
method='binary',
method="binary",
optimizers=optimizers,
lr_schedulers=schedulers,
initializers=initializers,
callbacks=callbacks,
metrics=metrics)
metrics=metrics,
)
model.fit(
X_wide=X_wide,
......@@ -68,4 +88,5 @@ if __name__ == '__main__':
target=target,
n_epochs=10,
batch_size=256,
val_split=0.2)
val_split=0.2,
)
examples/airbnb_data_preprocessing.py
浏览文件 @ 263663e5
......@@ -3,32 +3,59 @@ import numpy as np
import pandas as pd
import warnings
import os
import gender_guesser.detector as gender
from sklearn.preprocessing import MultiLabelBinarizer
from functools import reduce
from itertools import chain
from collections import Counter
from pathlib import Path
warnings.filterwarnings('ignore')
warnings.filterwarnings("ignore")
DATA_PATH = Path('data/airbnb')
fname = 'listings.csv.gz'
DATA_PATH = Path("data/airbnb")
fname = "listings.csv.gz"
if not os.path.exists(DATA_PATH):
os.makedirs(DATA_PATH)
df_original = pd.read_csv(DATA_PATH/fname)
df_original = pd.read_csv(DATA_PATH / fname)
print(df_original.shape)
df_original.head()
# this is just subjective. One can choose some other columns
keep_cols = ['id', 'host_id', 'description', 'house_rules', 'host_name',
'host_listings_count', 'host_identity_verified', 'neighbourhood_cleansed',
'latitude', 'longitude','is_location_exact','property_type', 'room_type',
'accommodates', 'bathrooms', 'bedrooms','beds', 'amenities', 'price',
'security_deposit', 'cleaning_fee','guests_included', 'extra_people',
'minimum_nights','instant_bookable','cancellation_policy',
'reviews_per_month']
keep_cols = [
"id",
"host_id",
"description",
"house_rules",
"host_name",
"host_listings_count",
"host_identity_verified",
"neighbourhood_cleansed",
"latitude",
"longitude",
"is_location_exact",
"property_type",
"room_type",
"accommodates",
"bathrooms",
"bedrooms",
"beds",
"amenities",
"price",
"security_deposit",
"cleaning_fee",
"guests_included",
"extra_people",
"minimum_nights",
"instant_bookable",
"cancellation_policy",
"reviews_per_month",
]
df = df_original[keep_cols]
df = df[~df.reviews_per_month.isna()]
df = df[~df.description.isna() ]
df = df[~df.description.isna()]
df = df[~df.host_listings_count.isna()]
print(df.shape)
......@@ -39,25 +66,23 @@ print(df.shape)
#
# I will simply include a binary column with 1/0 if the property has/has not
# house rules.
df['has_house_rules'] = df['house_rules']
df["has_house_rules"] = df["house_rules"]
df.has_house_rules.fillna(0, inplace=True)
df['has_house_rules'][df.has_house_rules!=0] = 1
df.drop('house_rules', axis=1, inplace=True)
df["has_house_rules"][df.has_house_rules != 0] = 1
df.drop("house_rules", axis=1, inplace=True)
# host_name
#
# I will use names to infer gender using `gender_guesser`
import gender_guesser.detector as gender
from collections import Counter
host_name = df.host_name.tolist()
d = gender.Detector()
host_gender = [d.get_gender(n) for n in host_name]
replace_dict = {'mostly_male': 'male', 'mostly_female': 'female', 'andy': 'unknown'}
host_gender = [replace_dict.get(item,item) for item in host_gender]
replace_dict = {"mostly_male": "male", "mostly_female": "female", "andy": "unknown"}
host_gender = [replace_dict.get(item, item) for item in host_gender]
Counter(host_gender)
df['host_gender'] = host_gender
df.drop('host_name', axis=1, inplace=True)
df["host_gender"] = host_gender
df.drop("host_name", axis=1, inplace=True)
df.head()
# property_type, room_type, accommodates, bathrooms, bedrooms, beds and
......@@ -65,45 +90,58 @@ df.head()
#
# Here some standard pre-processing
df.property_type.value_counts()
replace_prop_type = [val for val in df.property_type.unique().tolist() if val not in ['Apartment', 'House']]
replace_prop_type = {k:'other' for k in replace_prop_type}
replace_prop_type = [
val
for val in df.property_type.unique().tolist()
if val not in ["Apartment", "House"]
]
replace_prop_type = {k: "other" for k in replace_prop_type}
df.property_type.replace(replace_prop_type, inplace=True)
df['property_type'] = df.property_type.apply(lambda x: '_'.join(x.split(' ')).lower())
df["property_type"] = df.property_type.apply(lambda x: "_".join(x.split(" ")).lower())
df.room_type.value_counts()
df['room_type'] = df.room_type.apply(lambda x: '_'.join(x.split(' ')).lower())
df["room_type"] = df.room_type.apply(lambda x: "_".join(x.split(" ")).lower())
df['bathrooms'][(df.bathrooms.isna()) & (df.room_type == 'private_room')] = 0
df['bathrooms'][(df.bathrooms.isna()) & (df.room_type == 'entire_home/apt')] = 1
df["bathrooms"][(df.bathrooms.isna()) & (df.room_type == "private_room")] = 0
df["bathrooms"][(df.bathrooms.isna()) & (df.room_type == "entire_home/apt")] = 1
df.bedrooms.fillna(1, inplace=True)
df.beds.fillna(1, inplace=True)
# Encode some as categorical
categorical_cut = [('accommodates', 3),
('guests_included', 3),
('minimum_nights', 3),
('host_listings_count', 3),
('bathrooms', 1.5),
('bedrooms', 3),
('beds', 3)
]
categorical_cut = [
("accommodates", 3),
("guests_included", 3),
("minimum_nights", 3),
("host_listings_count", 3),
("bathrooms", 1.5),
("bedrooms", 3),
("beds", 3),
]
for col, cut in categorical_cut:
new_colname = col + '_catg'
df[new_colname] = df[col].apply(lambda x: cut if x >=cut else x)
new_colname = col + "_catg"
df[new_colname] = df[col].apply(lambda x: cut if x >= cut else x)
df[new_colname] = df[new_colname].round().astype(int)
# Amenities
#
# I will just add a number of dummy columns with 1/0 if the property has/has
# not that particular amenity
from functools import reduce
from itertools import chain
amenity_repls = (("\"", ""), ("{", ""), ("}", ""),(" / ", "_"), ("/", "_"), (" ", "_"), ("(s)", ""))
amenity_repls = (
('"', ""),
("{", ""),
("}", ""),
(" / ", "_"),
("/", "_"),
(" ", "_"),
("(s)", ""),
)
amenities_raw = df.amenities.str.lower().tolist()
amenities = [reduce(lambda a, kv: a.replace(*kv), amenity_repls, s).split(",") for s in amenities_raw]
amenities = [
reduce(lambda a, kv: a.replace(*kv), amenity_repls, s).split(",")
for s in amenities_raw
]
all_amenities = list(chain(*amenities))
all_amenities_count = Counter(all_amenities)
......@@ -112,71 +150,81 @@ all_amenities_count
# having a look to the list we see that one amenity is empty and two are
# "translation missing:..."
keep_amenities = []
for k,v in all_amenities_count.items():
if k and 'missing' not in k:
for k, v in all_amenities_count.items():
if k and "missing" not in k:
keep_amenities.append(k)
final_amenities = [[amenity for amenity in house_amenities if amenity in keep_amenities]
for house_amenities in amenities]
final_amenities = [
[amenity for amenity in house_amenities if amenity in keep_amenities]
for house_amenities in amenities
]
# some properties have no amenities aparently
final_amenities = [['no amenities'] if not amenity else amenity for amenity in final_amenities]
final_amenities = [['amenity_'+ amenity for amenity in amenities] for amenities in final_amenities]
final_amenities = [
["no amenities"] if not amenity else amenity for amenity in final_amenities
]
final_amenities = [
["amenity_" + amenity for amenity in amenities] for amenities in final_amenities
]
# let's build the dummy df
from sklearn.preprocessing import MultiLabelBinarizer
df_list_of_amenities = pd.DataFrame(
{'groups': final_amenities
}, columns=['groups'])
s = df_list_of_amenities['groups']
df_list_of_amenities = pd.DataFrame({"groups": final_amenities}, columns=["groups"])
s = df_list_of_amenities["groups"]
mlb = MultiLabelBinarizer()
df_amenities = pd.DataFrame(mlb.fit_transform(s), columns=mlb.classes_, index=df.index)
df.drop('amenities', axis=1, inplace=True)
df.drop("amenities", axis=1, inplace=True)
df = pd.concat([df, df_amenities], axis=1)
df.head()
# Price, security_deposit, cleaning_fee, extra_people
money_columns = ['price', 'security_deposit', 'cleaning_fee', 'extra_people']
tmp_money_df = df[money_columns].fillna('$0')
money_columns = ["price", "security_deposit", "cleaning_fee", "extra_people"]
tmp_money_df = df[money_columns].fillna("$0")
money_repls = (("$", ""), (",", ""))
for col in money_columns:
val_str = tmp_money_df[col].tolist()
val_num = [float(st) for st in [reduce(lambda a, kv: a.replace(*kv), money_repls, s) for s in val_str]]
val_num = [
float(st)
for st in [
reduce(lambda a, kv: a.replace(*kv), money_repls, s) for s in val_str
]
]
tmp_money_df[col] = val_num
high_price, high_deposit, high_cleaning_fee, high_extra_people = 1000,2000,200,100
high_price, high_deposit, high_cleaning_fee, high_extra_people = 1000, 2000, 200, 100
high_price_count = (tmp_money_df.price >= high_price).sum()
high_deposit_count = (tmp_money_df.security_deposit >= high_deposit).sum()
high_cleaning_fee_count = (tmp_money_df.cleaning_fee >= high_cleaning_fee).sum()
high_extra_people_count = (tmp_money_df.extra_people >= high_extra_people).sum()
print('properties with very high price: {}'.format(high_price_count))
print('properties with very high security deposit: {}'.format(high_deposit_count))
print('properties with very high cleaning fee: {}'.format(high_cleaning_fee_count))
print('properties with very high extra people cost: {}'.format(high_extra_people_count))
print("properties with very high price: {}".format(high_price_count))
print("properties with very high security deposit: {}".format(high_deposit_count))
print("properties with very high cleaning fee: {}".format(high_cleaning_fee_count))
print("properties with very high extra people cost: {}".format(high_extra_people_count))
# We will now just concat and we will drop high values later one
df.drop(money_columns, axis=1, inplace=True)
df = pd.concat([df, tmp_money_df], axis=1)
df = df[(df.price<high_price) &
(df.price!=0) &
(df.security_deposit<high_deposit) &
(df.cleaning_fee<high_cleaning_fee) &
(df.extra_people<high_extra_people)]
df = df[
(df.price < high_price)
& (df.price != 0)
& (df.security_deposit < high_deposit)
& (df.cleaning_fee < high_cleaning_fee)
& (df.extra_people < high_extra_people)
]
df.head()
print(df.shape)
# let's make sure there are no nan left
has_nan = df.isnull().any(axis=0)
has_nan = [df.columns[i] for i in np.where(has_nan)[0]]
if not has_nan: print('no NaN, all OK')
if not has_nan:
print("no NaN, all OK")
# Computing a proxi for yield
......@@ -192,11 +240,10 @@ if not has_nan: print('no NaN, all OK')
# Francisco model assumptions simply multiply my yield by 6 (3 * (1/0.5)) or
# by 72 (3 * 2 * 12) if you prefer per year.
df['yield'] = (df['price']+df['cleaning_fee']) * (df['reviews_per_month'])
df.drop(['price','cleaning_fee','reviews_per_month'], axis=1, inplace=True)
df["yield"] = (df["price"] + df["cleaning_fee"]) * (df["reviews_per_month"])
df.drop(["price", "cleaning_fee", "reviews_per_month"], axis=1, inplace=True)
# we will focus in cases with yield below 600 (we lose ~3% of the data).
# No real reason for this, simply removing some "outliers"
df = df[df['yield'] <= 600]
df.to_csv(DATA_PATH/'listings_processed.csv', index=False)
df = df[df["yield"] <= 600]
df.to_csv(DATA_PATH / "listings_processed.csv", index=False)
print("data preprocessed finished. Final shape: {}".format(df.shape))
examples/airbnb_script.py
浏览文件 @ 263663e5
import numpy as np
import pandas as pd
import os
import torch
from pytorch_widedeep.preprocessing import (WidePreprocessor, DeepPreprocessor,
TextPreprocessor, ImagePreprocessor)
from pytorch_widedeep.models import (Wide, DeepDense, DeepText, DeepImage,
WideDeep)
from pytorch_widedeep.initializers import *
from pytorch_widedeep.callbacks import *
from torchvision.transforms import ToTensor, Normalize
from pytorch_widedeep.preprocessing import (
WidePreprocessor,
DeepPreprocessor,
TextPreprocessor,
ImagePreprocessor,
)
from pytorch_widedeep.models import Wide, DeepDense, DeepText, DeepImage, WideDeep
from pytorch_widedeep.initializers import KaimingNormal
from pytorch_widedeep.callbacks import EarlyStopping, ModelCheckpoint
from pytorch_widedeep.optim import RAdam
use_cuda = torch.cuda.is_available()
if __name__ == '__main__':
df = pd.read_csv('data/airbnb/airbnb_sample.csv')
use_cuda = torch.cuda.is_available()
crossed_cols = (['property_type', 'room_type'],)
already_dummies = [c for c in df.columns if 'amenity' in c] + ['has_house_rules']
wide_cols = ['is_location_exact', 'property_type', 'room_type', 'host_gender',
'instant_bookable'] + already_dummies
cat_embed_cols = [(c, 16) for c in df.columns if 'catg' in c] + \
[('neighbourhood_cleansed', 64), ('cancellation_policy', 16)]
continuous_cols = ['latitude', 'longitude', 'security_deposit', 'extra_people']
already_standard = ['latitude', 'longitude']
text_col = 'description'
word_vectors_path = 'data/glove.6B/glove.6B.100d.txt'
img_col = 'id'
img_path = 'data/airbnb/property_picture'
target = 'yield'
if __name__ == "__main__":
df = pd.read_csv("data/airbnb/airbnb_sample.csv")
crossed_cols = (["property_type", "room_type"],)
already_dummies = [c for c in df.columns if "amenity" in c] + ["has_house_rules"]
wide_cols = [
"is_location_exact",
"property_type",
"room_type",
"host_gender",
"instant_bookable",
] + already_dummies
cat_embed_cols = [(c, 16) for c in df.columns if "catg" in c] + [
("neighbourhood_cleansed", 64),
("cancellation_policy", 16),
]
continuous_cols = ["latitude", "longitude", "security_deposit", "extra_people"]
already_standard = ["latitude", "longitude"]
text_col = "description"
word_vectors_path = "data/glove.6B/glove.6B.100d.txt"
img_col = "id"
img_path = "data/airbnb/property_picture"
target = "yield"
target = df[target].values
prepare_wide = WidePreprocessor(wide_cols=wide_cols, crossed_cols=crossed_cols)
X_wide = prepare_wide.fit_transform(df)
prepare_deep = DeepPreprocessor(embed_cols=cat_embed_cols, continuous_cols=continuous_cols)
prepare_deep = DeepPreprocessor(
embed_cols=cat_embed_cols, continuous_cols=continuous_cols
)
X_deep = prepare_deep.fit_transform(df)
text_processor = TextPreprocessor(word_vectors_path=word_vectors_path)
X_text = text_processor.fit_transform(df, text_col)
text_processor = TextPreprocessor(
word_vectors_path=word_vectors_path, text_col=text_col
)
X_text = text_processor.fit_transform(df)
image_processor = ImagePreprocessor()
X_images = image_processor.fit_transform(df, img_col, img_path)
image_processor = ImagePreprocessor(img_col=img_col, img_path=img_path)
X_images = image_processor.fit_transform(df)
wide = Wide(
wide_dim=X_wide.shape[1],
output_dim=1)
wide = Wide(wide_dim=X_wide.shape[1], output_dim=1)
deepdense = DeepDense(
hidden_layers=[64,32],
dropout=[0.2,0.2],
hidden_layers=[64, 32],
dropout=[0.2, 0.2],
deep_column_idx=prepare_deep.deep_column_idx,
embed_input=prepare_deep.embeddings_input,
continuous_cols=continuous_cols)
continuous_cols=continuous_cols,
)
deeptext = DeepText(
vocab_size=len(text_processor.vocab.itos),
hidden_dim=64,
n_layers=3,
rnn_dropout=0.5,
padding_idx=1,
embedding_matrix=text_processor.embedding_matrix
)
embedding_matrix=text_processor.embedding_matrix,
)
deepimage = DeepImage(pretrained=True, head_layers=None)
model = WideDeep(wide=wide, deepdense=deepdense, deeptext=deeptext,
deepimage=deepimage)
model = WideDeep(
wide=wide, deepdense=deepdense, deeptext=deeptext, deepimage=deepimage
)
wide_opt = torch.optim.Adam(model.wide.parameters())
deep_opt = torch.optim.Adam(model.deepdense.parameters())
text_opt = RAdam(model.deeptext.parameters())
img_opt = RAdam(model.deepimage.parameters())
img_opt = RAdam(model.deepimage.parameters())
wide_sch = torch.optim.lr_scheduler.StepLR(wide_opt, step_size=5)
deep_sch = torch.optim.lr_scheduler.StepLR(deep_opt, step_size=3)
text_sch = torch.optim.lr_scheduler.StepLR(text_opt, step_size=5)
img_sch = torch.optim.lr_scheduler.StepLR(img_opt, step_size=3)
optimizers = {'wide': wide_opt, 'deepdense':deep_opt, 'deeptext':text_opt, 'deepimage': img_opt}
schedulers = {'wide': wide_sch, 'deepdense':deep_sch, 'deeptext':text_sch, 'deepimage': img_sch}
initializers = {'wide': KaimingNormal, 'deepdense':KaimingNormal, 'deeptext':KaimingNormal, 'deepimage':KaimingNormal}
mean = [0.406, 0.456, 0.485] #BGR
std = [0.225, 0.224, 0.229] #BGR
img_sch = torch.optim.lr_scheduler.StepLR(img_opt, step_size=3)
optimizers = {
"wide": wide_opt,
"deepdense": deep_opt,
"deeptext": text_opt,
"deepimage": img_opt,
}
schedulers = {
"wide": wide_sch,
"deepdense": deep_sch,
"deeptext": text_sch,
"deepimage": img_sch,
}
initializers = {
"wide": KaimingNormal,
"deepdense": KaimingNormal,
"deeptext": KaimingNormal,
"deepimage": KaimingNormal,
}
mean = [0.406, 0.456, 0.485] # BGR
std = [0.225, 0.224, 0.229] # BGR
transforms = [ToTensor, Normalize(mean=mean, std=std)]
callbacks = [EarlyStopping, ModelCheckpoint(filepath='model_weights/wd_out.pt')]
model.compile(method='regression', initializers=initializers, optimizers=optimizers,
lr_schedulers=schedulers, callbacks=callbacks, transforms=transforms)
model.fit(X_wide=X_wide, X_deep=X_deep, X_text=X_text, X_img=X_images,
target=target, n_epochs=1, batch_size=32, val_split=0.2)
# With warm_up
callbacks = [EarlyStopping, ModelCheckpoint(filepath="model_weights/wd_out.pt")]
model.compile(
method="regression",
initializers=initializers,
optimizers=optimizers,
lr_schedulers=schedulers,
callbacks=callbacks,
transforms=transforms,
)
model.fit(
X_wide=X_wide,
X_deep=X_deep,
X_text=X_text,
X_img=X_images,
target=target,
n_epochs=1,
batch_size=32,
val_split=0.2,
)
# # With warm_up
# child = list(model.deepimage.children())[0]
# img_layers = list(child.backbone.children())[4:8] + [list(model.deepimage.children())[1]]
# img_layers = img_layers[::-1]
......@@ -98,4 +142,4 @@ if __name__ == '__main__':
# model.fit(X_wide=X_wide, X_deep=X_deep, X_text=X_text, X_img=X_images,
# target=target, n_epochs=1, batch_size=32, val_split=0.2, warm_up=True,
# warm_epochs=1, warm_deepimage_gradual=True, warm_deepimage_layers=img_layers,
# warm_deepimage_max_lr=0.01, warm_routine='howard')
\ No newline at end of file
# warm_deepimage_max_lr=0.01, warm_routine='howard')
examples/airbnb_script_multiclass.py
浏览文件 @ 263663e5
......@@ -4,51 +4,58 @@ import torch
from pytorch_widedeep.preprocessing import WidePreprocessor, DeepPreprocessor
from pytorch_widedeep.models import Wide, DeepDense, WideDeep
from pytorch_widedeep.optim import RAdam
from pytorch_widedeep.initializers import *
from pytorch_widedeep.callbacks import *
from pytorch_widedeep.metrics import *
import pdb
from pytorch_widedeep.metrics import CategoricalAccuracy
use_cuda = torch.cuda.is_available()
if __name__ == '__main__':
if __name__ == "__main__":
df = pd.read_csv('data/airbnb/airbnb_sample.csv')
df = pd.read_csv("data/airbnb/airbnb_sample.csv")
crossed_cols = (['property_type', 'room_type'],)
already_dummies = [c for c in df.columns if 'amenity' in c] + ['has_house_rules']
wide_cols = ['is_location_exact', 'property_type', 'room_type', 'host_gender',
'instant_bookable'] + already_dummies
cat_embed_cols = [(c, 16) for c in df.columns if 'catg' in c] + \
[('neighbourhood_cleansed', 64), ('cancellation_policy', 16)]
continuous_cols = ['latitude', 'longitude', 'security_deposit', 'extra_people']
already_standard = ['latitude', 'longitude']
df['yield_cat'] = pd.cut(df['yield'], bins=[0.2, 65, 163, 600], labels=[0,1,2])
df.drop('yield', axis=1, inplace=True)
target = 'yield_cat'
crossed_cols = (["property_type", "room_type"],)
already_dummies = [c for c in df.columns if "amenity" in c] + ["has_house_rules"]
wide_cols = [
"is_location_exact",
"property_type",
"room_type",
"host_gender",
"instant_bookable",
] + already_dummies
cat_embed_cols = [(c, 16) for c in df.columns if "catg" in c] + [
("neighbourhood_cleansed", 64),
("cancellation_policy", 16),
]
continuous_cols = ["latitude", "longitude", "security_deposit", "extra_people"]
already_standard = ["latitude", "longitude"]
df["yield_cat"] = pd.cut(df["yield"], bins=[0.2, 65, 163, 600], labels=[0, 1, 2])
df.drop("yield", axis=1, inplace=True)
target = "yield_cat"
target = np.array(df[target].values)
prepare_wide = WidePreprocessor(wide_cols=wide_cols, crossed_cols=crossed_cols)
X_wide = prepare_wide.fit_transform(df)
prepare_deep = DeepPreprocessor(embed_cols=cat_embed_cols, continuous_cols=continuous_cols)
prepare_deep = DeepPreprocessor(
embed_cols=cat_embed_cols, continuous_cols=continuous_cols
)
X_deep = prepare_deep.fit_transform(df)
wide = Wide(
wide_dim=X_wide.shape[1],
output_dim=3)
wide = Wide(wide_dim=X_wide.shape[1], output_dim=3)
deepdense = DeepDense(
hidden_layers=[64,32],
dropout=[0.2,0.2],
hidden_layers=[64, 32],
dropout=[0.2, 0.2],
deep_column_idx=prepare_deep.deep_column_idx,
embed_input=prepare_deep.embeddings_input,
continuous_cols=continuous_cols)
continuous_cols=continuous_cols,
)
model = WideDeep(wide=wide, deepdense=deepdense, output_dim=3)
model.compile(method='multiclass', metrics=[CategoricalAccuracy])
model.compile(method="multiclass", metrics=[CategoricalAccuracy])
model.fit(
X_wide=X_wide,
X_deep=X_deep,
target=target,
n_epochs=1,
batch_size=32,
val_split=0.2)
\ No newline at end of file
val_split=0.2,
)
\ No newline at end of file
examples/download_images.py
浏览文件 @ 263663e5
......@@ -7,42 +7,44 @@ from urllib.request import urlretrieve
from pathlib import Path
def download_images(df, out_path, id_col, img_col):
download_error = []
counter = 0
for idx,row in tqdm(df.iterrows(), total=df.shape[0]):
if counter <1000:
img_path = str(out_path/'.'.join([str(row[id_col]), 'jpg']))
if os.path.isfile(img_path):
continue
else:
try:
urlretrieve(row[img_col], img_path)
counter+=1
except:
# print("Error downloading host image {}".format(row[id_col]))
download_error.append(row[id_col])
pass
pickle.dump(download_error, open(DATA_PATH/(id_col+'_download_error.p'), 'wb'))
if __name__ == '__main__':
DATA_PATH=Path('data/airbnb')
HOST_PATH = DATA_PATH/'host_picture'
PROP_PATH = DATA_PATH/'property_picture'
if not os.path.exists(DATA_PATH):
os.makedirs(DATA_PATH)
if not os.path.exists(HOST_PATH):
os.makedirs(HOST_PATH)
if not os.path.exists(PROP_PATH):
os.makedirs(PROP_PATH)
df_original = pd.read_csv(DATA_PATH/'listings.csv.gz')[['id', 'host_id', 'picture_url', 'host_picture_url']]
df_processed = pd.read_csv(DATA_PATH/'listings_processed.csv')[['id', 'host_id']]
df = df_processed.merge(df_original, on=['id', 'host_id'])
df_host = df.groupby('host_id').first().reset_index()
download_images(df_host, HOST_PATH, id_col='host_id', img_col='host_picture_url')
download_images(df, PROP_PATH, id_col='id', img_col='picture_url')
def download_images(df, out_path, id_col, img_col):
download_error = []
counter = 0
for idx, row in tqdm(df.iterrows(), total=df.shape[0]):
if counter < 1000:
img_path = str(out_path / ".".join([str(row[id_col]), "jpg"]))
if os.path.isfile(img_path):
continue
else:
try:
urlretrieve(row[img_col], img_path)
counter += 1
except:
# print("Error downloading host image {}".format(row[id_col]))
download_error.append(row[id_col])
pass
pickle.dump(download_error, open(DATA_PATH / (id_col + "_download_error.p"), "wb"))
if __name__ == "__main__":
DATA_PATH = Path("data/airbnb")
HOST_PATH = DATA_PATH / "host_picture"
PROP_PATH = DATA_PATH / "property_picture"
if not os.path.exists(DATA_PATH):
os.makedirs(DATA_PATH)
if not os.path.exists(HOST_PATH):
os.makedirs(HOST_PATH)
if not os.path.exists(PROP_PATH):
os.makedirs(PROP_PATH)
df_original = pd.read_csv(DATA_PATH / "listings.csv.gz")[
["id", "host_id", "picture_url", "host_picture_url"]
]
df_processed = pd.read_csv(DATA_PATH / "listings_processed.csv")[["id", "host_id"]]
df = df_processed.merge(df_original, on=["id", "host_id"])
df_host = df.groupby("host_id").first().reset_index()
download_images(df_host, HOST_PATH, id_col="host_id", img_col="host_picture_url")
download_images(df, PROP_PATH, id_col="id", img_col="picture_url")
pytorch_widedeep/__init__.py
浏览文件 @ 263663e5
......@@ -11,4 +11,4 @@ from pytorch_widedeep.version import __version__
from .utils import dense_utils
from .utils import text_utils
from .utils import fastai_transforms
from .utils import image_utils
\ No newline at end of file
from .utils import image_utils
pytorch_widedeep/callbacks.py
浏览文件 @ 263663e5
此差异已折叠。
pytorch_widedeep/initializers.py
浏览文件 @ 263663e5
import torch
import re
import warnings
......@@ -7,180 +6,181 @@ from .wdtypes import *
class Initializer(object):
def __call__(self, model:nn.Module):
raise NotImplementedError('Initializer must implement this method')
def __call__(self, model: nn.Module):
raise NotImplementedError("Initializer must implement this method")
class MultipleInitializer(object):
def __init__(self, initializers:Dict[str, Initializer], verbose=True):
self.verbose=verbose
instantiated_initializers = {}
for model_name, initializer in initializers.items():
if isinstance(initializer, type):
instantiated_initializers[model_name] = initializer()
else: instantiated_initializers[model_name] = initializer
self._initializers = instantiated_initializers
def apply(self, model:nn.Module):
for name, child in model.named_children():
try:
self._initializers[name](child)
except:
if self.verbose: warnings.warn("No initializer found for {}".format(name))
def __init__(self, initializers: Dict[str, Initializer], verbose=True):
self.verbose = verbose
instantiated_initializers = {}
for model_name, initializer in initializers.items():
if isinstance(initializer, type):
instantiated_initializers[model_name] = initializer()
else:
instantiated_initializers[model_name] = initializer
self._initializers = instantiated_initializers
def apply(self, model: nn.Module):
for name, child in model.named_children():
try:
self._initializers[name](child)
except:
if self.verbose:
warnings.warn("No initializer found for {}".format(name))
class Normal(Initializer):
def __init__(self, mean=0.0, std=1.0, bias=False, pattern='.'):
self.mean = mean
self.std = std
self.bias = bias
self.pattern = pattern
super(Normal, self).__init__()
def __call__(self, submodel:nn.Module):
for n,p in submodel.named_parameters():
if re.search(self.pattern, n):
if self.bias and ('bias' in n):
nn.init.normal_(p, mean=self.mean, std=self.std)
elif 'bias' in n:
pass
elif p.requires_grad:
nn.init.normal_(p, mean=self.mean, std=self.std)
def __init__(self, mean=0.0, std=1.0, bias=False, pattern="."):
self.mean = mean
self.std = std
self.bias = bias
self.pattern = pattern
super(Normal, self).__init__()
def __call__(self, submodel: nn.Module):
for n, p in submodel.named_parameters():
if re.search(self.pattern, n):
if self.bias and ("bias" in n):
nn.init.normal_(p, mean=self.mean, std=self.std)
elif "bias" in n:
pass
elif p.requires_grad:
nn.init.normal_(p, mean=self.mean, std=self.std)
class Uniform(Initializer):
def __init__(self, a=0, b=1, bias=False, pattern='.'):
self.a = a
self.b = b
self.bias = bias
self.pattern = pattern
super(Uniform, self).__init__()
def __call__(self, submodel:nn.Module):
for n,p in submodel.named_parameters():
if re.search(self.pattern, n):
if self.bias and ('bias' in n):
nn.init.uniform_(p, a=self.a, b=self.b)
elif 'bias' in n:
pass
elif p.requires_grad:
nn.init.uniform_(p, a=self.a, b=self.b)
def __init__(self, a=0, b=1, bias=False, pattern="."):
self.a = a
self.b = b
self.bias = bias
self.pattern = pattern
super(Uniform, self).__init__()
def __call__(self, submodel: nn.Module):
for n, p in submodel.named_parameters():
if re.search(self.pattern, n):
if self.bias and ("bias" in n):
nn.init.uniform_(p, a=self.a, b=self.b)
elif "bias" in n:
pass
elif p.requires_grad:
nn.init.uniform_(p, a=self.a, b=self.b)
class ConstantInitializer(Initializer):
def __init__(self, value, bias=False, pattern="."):
def __init__(self, value, bias=False, pattern='.'):
self.bias = bias
self.value = value
self.pattern = pattern
super(ConstantInitializer, self).__init__()
self.bias = bias
self.value = value
self.pattern = pattern
super(ConstantInitializer, self).__init__()
def __call__(self, submodel:nn.Module):
for n,p in submodel.named_parameters():
if re.search(self.pattern, n):
if self.bias and ('bias' in n):
nn.init.constant_(p, val=self.value)
elif ('bias' in n):
pass
elif p.requires_grad:
nn.init.constant_(p, val=self.value)
def __call__(self, submodel: nn.Module):
for n, p in submodel.named_parameters():
if re.search(self.pattern, n):
if self.bias and ("bias" in n):
nn.init.constant_(p, val=self.value)
elif "bias" in n:
pass
elif p.requires_grad:
nn.init.constant_(p, val=self.value)
class XavierUniform(Initializer):
def __init__(self, gain=1, pattern='.'):
self.gain = gain
self.pattern = pattern
super(XavierUniform, self).__init__()
def __call__(self, submodel:nn.Module):
for n,p in submodel.named_parameters():
if re.search(self.pattern, n):
if 'bias' in n:
nn.init.constant_(p, val=0)
elif p.requires_grad:
try:
nn.init.xavier_uniform_(p, gain=self.gain)
except: pass
def __init__(self, gain=1, pattern="."):
self.gain = gain
self.pattern = pattern
super(XavierUniform, self).__init__()
def __call__(self, submodel: nn.Module):
for n, p in submodel.named_parameters():
if re.search(self.pattern, n):
if "bias" in n:
nn.init.constant_(p, val=0)
elif p.requires_grad:
try:
nn.init.xavier_uniform_(p, gain=self.gain)
except:
pass
class XavierNormal(Initializer):
def __init__(self, gain=1, pattern='.'):
self.gain = gain
self.pattern = pattern
super(XavierNormal, self).__init__()
def __call__(self, submodel:nn.Module):
for n,p in submodel.named_parameters():
if re.search(self.pattern, n):
if 'bias' in n:
nn.init.constant_(p, val=0)
elif p.requires_grad:
try:
nn.init.xavier_normal_(p, gain=self.gain)
except: pass
def __init__(self, gain=1, pattern="."):
self.gain = gain
self.pattern = pattern
super(XavierNormal, self).__init__()
def __call__(self, submodel: nn.Module):
for n, p in submodel.named_parameters():
if re.search(self.pattern, n):
if "bias" in n:
nn.init.constant_(p, val=0)
elif p.requires_grad:
try:
nn.init.xavier_normal_(p, gain=self.gain)
except:
pass
class KaimingUniform(Initializer):
def __init__(self, a=0, mode='fan_in', nonlinearity='leaky_relu', pattern='.'):
self.a = a
self.mode = mode
self.nonlinearity = nonlinearity
self.pattern = pattern
super(KaimingUniform, self).__init__()
def __call__(self, submodel:nn.Module):
for n,p in submodel.named_parameters():
if re.search(self.pattern, n):
if 'bias' in n:
nn.init.constant_(p, val=0)
elif p.requires_grad:
try:
nn.init.kaiming_normal_(p, a=self.a, mode=self.mode, nonlinearity=self.nonlinearity)
except: pass
def __init__(self, a=0, mode="fan_in", nonlinearity="leaky_relu", pattern="."):
self.a = a
self.mode = mode
self.nonlinearity = nonlinearity
self.pattern = pattern
super(KaimingUniform, self).__init__()
def __call__(self, submodel: nn.Module):
for n, p in submodel.named_parameters():
if re.search(self.pattern, n):
if "bias" in n:
nn.init.constant_(p, val=0)
elif p.requires_grad:
try:
nn.init.kaiming_normal_(
p, a=self.a, mode=self.mode, nonlinearity=self.nonlinearity
)
except:
pass
class KaimingNormal(Initializer):
def __init__(self, a=0, mode='fan_in', nonlinearity='leaky_relu', pattern='.'):
self.a = a
self.mode = mode
self.nonlinearity = nonlinearity
self.pattern = pattern
super(KaimingNormal, self).__init__()
def __call__(self, submodel:nn.Module):
for n,p in submodel.named_parameters():
if re.search(self.pattern, n):
if 'bias' in n:
nn.init.constant_(p, val=0)
elif p.requires_grad:
try:
nn.init.kaiming_normal_(p, a=self.a, mode=self.mode, nonlinearity=self.nonlinearity)
except: pass
def __init__(self, a=0, mode="fan_in", nonlinearity="leaky_relu", pattern="."):
self.a = a
self.mode = mode
self.nonlinearity = nonlinearity
self.pattern = pattern
super(KaimingNormal, self).__init__()
def __call__(self, submodel: nn.Module):
for n, p in submodel.named_parameters():
if re.search(self.pattern, n):
if "bias" in n:
nn.init.constant_(p, val=0)
elif p.requires_grad:
try:
nn.init.kaiming_normal_(
p, a=self.a, mode=self.mode, nonlinearity=self.nonlinearity
)
except:
pass
class Orthogonal(Initializer):
def __init__(self, gain=1, pattern='.'):
self.gain = gain
self.pattern = pattern
super(Orthogonal, self).__init__()
def __call__(self, submodel:nn.Module):
for n,p in submodel.named_parameters():
if re.search(self.pattern, n):
if 'bias' in n:
nn.init.constant_(p, val=0)
elif p.requires_grad:
try:
nn.init.orthogonal_(p, gain=self.gain)
except: pass
def __init__(self, gain=1, pattern="."):
self.gain = gain
self.pattern = pattern
super(Orthogonal, self).__init__()
def __call__(self, submodel: nn.Module):
for n, p in submodel.named_parameters():
if re.search(self.pattern, n):
if "bias" in n:
nn.init.constant_(p, val=0)
elif p.requires_grad:
try:
nn.init.orthogonal_(p, gain=self.gain)
except:
pass
pytorch_widedeep/losses.py
浏览文件 @ 263663e5
......@@ -9,24 +9,28 @@ use_cuda = torch.cuda.is_available()
class FocalLoss(nn.Module):
def __init__(self, alpha:float=0.25, gamma:float=1.):
def __init__(self, alpha: float = 0.25, gamma: float = 1.0):
super().__init__()
self.alpha = alpha
self.gamma = gamma
def get_weight(self, x:Tensor, t:Tensor) -> Tensor:
def get_weight(self, x: Tensor, t: Tensor) -> Tensor:
p = x.sigmoid()
pt = p*t + (1-p)*(1-t)
w = self.alpha*t + (1-self.alpha)*(1-t)
return (w * (1-pt).pow(self.gamma)).detach()
pt = p * t + (1 - p) * (1 - t) # type: ignore
w = self.alpha * t + (1 - self.alpha) * (1 - t) # type: ignore
return (w * (1 - pt).pow(self.gamma)).detach() # type: ignore
def forward(self, input:Tensor, target:Tensor) -> Tensor:
def forward(self, input: Tensor, target: Tensor) -> Tensor: # type: ignore
if input.size(1) == 1:
input = torch.cat([1-input, input], axis=1)
input = torch.cat([1 - input, input], axis=1) # type: ignore
num_class = 2
else: num_class = input.size(1)
else:
num_class = input.size(1)
binary_target = torch.eye(num_class)[target.long()]
if use_cuda: binary_target = binary_target.cuda()
if use_cuda:
binary_target = binary_target.cuda()
binary_target = binary_target.contiguous()
weight = self.get_weight(input, binary_target)
return F.binary_cross_entropy_with_logits(input, binary_target, weight, reduction='mean')
\ No newline at end of file
return F.binary_cross_entropy_with_logits(
input, binary_target, weight, reduction="mean"
)
pytorch_widedeep/metrics.py
浏览文件 @ 263663e5
import numpy as np
import torch
from .callbacks import Callback
from .wdtypes import *
class Metric(object):
def __init__(self):
self._name = ""
def reset(self):
raise NotImplementedError('Custom Metrics must implement this function')
raise NotImplementedError("Custom Metrics must implement this function")
def __call__(self, y_pred:Tensor, y_true:Tensor):
raise NotImplementedError('Custom Metrics must implement this function')
def __call__(self, y_pred: Tensor, y_true: Tensor):
raise NotImplementedError("Custom Metrics must implement this function")
class MultipleMetrics(object):
def __init__(self, metrics:List[Metric], prefix:str=''):
def __init__(self, metrics: List[Metric], prefix: str = ""):
instantiated_metrics = []
for metric in metrics:
if isinstance(metric, type): instantiated_metrics.append(metric())
else: instantiated_metrics.append(metric)
if isinstance(metric, type):
instantiated_metrics.append(metric())
else:
instantiated_metrics.append(metric)
self._metrics = instantiated_metrics
self.prefix = prefix
......@@ -29,58 +31,56 @@ class MultipleMetrics(object):
for metric in self._metrics:
metric.reset()
def __call__(self, y_pred:Tensor, y_true:Tensor) -> Dict:
def __call__(self, y_pred: Tensor, y_true: Tensor) -> Dict:
logs = {}
for metric in self._metrics:
logs[self.prefix+metric._name] = metric(y_pred, y_true)
logs[self.prefix + metric._name] = metric(y_pred, y_true)
return logs
class MetricCallback(Callback):
def __init__(self, container:MultipleMetrics):
def __init__(self, container: MultipleMetrics):
self.container = container
def on_epoch_begin(self, epoch:int, logs:Optional[Dict]=None):
def on_epoch_begin(self, epoch: int, logs: Optional[Dict] = None):
self.container.reset()
class CategoricalAccuracy(Metric):
def __init__(self, top_k=1):
self.top_k = top_k
self.correct_count = 0
self.total_count = 0
self._name = 'acc'
self._name = "acc"
def reset(self):
self.correct_count = 0
self.total_count = 0
def __call__(self, y_pred:Tensor, y_true:Tensor) -> np.ndarray:
top_k = (y_pred.topk(self.top_k,1)[1])
true_k = y_true.view(len(y_true),1).expand_as(top_k)
def __call__(self, y_pred: Tensor, y_true: Tensor) -> np.ndarray:
top_k = y_pred.topk(self.top_k, 1)[1]
true_k = y_true.view(len(y_true), 1).expand_as(top_k) # type: ignore
self.correct_count += top_k.eq(true_k).float().sum().item()
self.total_count += len(y_pred)
self.total_count += len(y_pred) # type: ignore
accuracy = float(self.correct_count) / float(self.total_count)
return np.round(accuracy, 4)
class BinaryAccuracy(Metric):
def __init__(self):
self.correct_count = 0
self.total_count = 0
self._name = 'acc'
self._name = "acc"
def reset(self):
self.correct_count = 0
self.total_count = 0
def __call__(self, y_pred:Tensor, y_true:Tensor) -> np.ndarray:
def __call__(self, y_pred: Tensor, y_true: Tensor) -> np.ndarray:
y_pred_round = y_pred.round()
self.correct_count += y_pred_round.eq(y_true.view(-1,1)).float().sum().item()
self.total_count += len(y_pred)
self.correct_count += y_pred_round.eq(y_true.view(-1, 1)).float().sum().item()
self.total_count += len(y_pred) # type: ignore
accuracy = float(self.correct_count) / float(self.total_count)
return np.round(accuracy, 4)
\ No newline at end of file
return np.round(accuracy, 4)
pytorch_widedeep/models/__init__.py
浏览文件 @ 263663e5
......@@ -3,4 +3,4 @@ from .deep_dense import DeepDense
from .deep_text import DeepText
from .deep_image import DeepImage
from .wide_deep import WideDeep
\ No newline at end of file
from .wide_deep import WideDeep
pytorch_widedeep/models/_multiple_lr_scheduler.py
浏览文件 @ 263663e5
import torch
from ..wdtypes import *
class MultipleLRScheduler(object):
def __init__(self, scheds:Dict[str,LRScheduler]):
def __init__(self, scheds: Dict[str, LRScheduler]):
self._schedulers = scheds
def step(self):
......
pytorch_widedeep/models/_multiple_optimizer.py
浏览文件 @ 263663e5
import torch
from ..wdtypes import *
class MultipleOptimizer(object):
def __init__(self, opts:Dict[str,Optimizer]):
def __init__(self, opts: Dict[str, Optimizer]):
self._optimizers = opts
def zero_grad(self):
......@@ -13,4 +11,4 @@ class MultipleOptimizer(object):
def step(self):
for _, op in self._optimizers.items():
op.step()
\ No newline at end of file
op.step()
pytorch_widedeep/models/_multiple_transforms.py
浏览文件 @ 263663e5
......@@ -2,15 +2,17 @@ from torchvision.transforms import Compose
from ..wdtypes import *
class MultipleTransforms(object):
def __init__(self, transforms:List):
class MultipleTransforms(object):
def __init__(self, transforms: List[Transforms]):
instantiated_transforms = []
for transform in transforms:
if isinstance(transform, type): instantiated_transforms.append(transform())
else: instantiated_transforms.append(transform)
self._transforms = instantiated_transforms
instantiated_transforms = []
for transform in transforms:
if isinstance(transform, type):
instantiated_transforms.append(transform())
else:
instantiated_transforms.append(transform)
self._transforms = instantiated_transforms
def __call__(self):
return Compose(self._transforms)
\ No newline at end of file
def __call__(self):
return Compose(self._transforms)
pytorch_widedeep/models/_warmup.py
浏览文件 @ 263663e5
此差异已折叠。
pytorch_widedeep/models/_wd_dataset.py
浏览文件 @ 263663e5
......@@ -28,26 +28,37 @@ class WideDeepDataset(Dataset):
transforms: MultipleTransforms() object (which is in itself a torchvision
Compose). See in models/_multiple_transforms.py
"""
def __init__(self, X_wide:Union[np.ndarray, sparse_matrix], X_deep:np.ndarray,
target:Optional[np.ndarray]=None, X_text:Optional[np.ndarray]=None,
X_img:Optional[np.ndarray]=None, transforms:Optional=None):
def __init__(
self,
X_wide: Union[np.ndarray, sparse_matrix],
X_deep: np.ndarray,
target: Optional[np.ndarray] = None,
X_text: Optional[np.ndarray] = None,
X_img: Optional[np.ndarray] = None,
transforms: Optional[Any] = None,
):
self.X_wide = X_wide
self.X_deep = X_deep
self.X_text = X_text
self.X_img = X_img
self.X_img = X_img
self.transforms = transforms
if self.transforms:
self.transforms_names = [tr.__class__.__name__ for tr in self.transforms.transforms]
else: self.transforms_names = []
self.transforms_names = [
tr.__class__.__name__ for tr in self.transforms.transforms
]
else:
self.transforms_names = []
self.Y = target
def __getitem__(self, idx:int):
def __getitem__(self, idx: int):
# X_wide and X_deep are assumed to be *always* present
if isinstance(self.X_wide, sparse_matrix):
X = Bunch(wide=np.array(self.X_wide[idx].todense()).squeeze())
else: X = Bunch(wide=self.X_wide[idx])
X.deepdense= self.X_deep[idx]
else:
X = Bunch(wide=self.X_wide[idx])
X.deepdense = self.X_deep[idx]
if self.X_text is not None:
X.deeptext = self.X_text[idx]
if self.X_img is not None:
......@@ -55,24 +66,29 @@ class WideDeepDataset(Dataset):
# be ingested by the conv layers
xdi = self.X_img[idx]
# if int must be uint8
if 'int' in str(xdi.dtype) and 'uint8' != str(xdi.dtype): xdi = xdi.astype('uint8')
if "int" in str(xdi.dtype) and "uint8" != str(xdi.dtype):
xdi = xdi.astype("uint8")
# if int float must be float32
if 'float' in str(xdi.dtype) and 'float32' != str(xdi.dtype): xdi = xdi.astype('float32')
if "float" in str(xdi.dtype) and "float32" != str(xdi.dtype):
xdi = xdi.astype("float32")
# if there are no transforms, or these do not include ToTensor(),
# then we need to replicate what Tensor() does -> transpose axis
# and normalize if necessary
if not self.transforms or 'ToTensor' not in self.transforms_names:
xdi = xdi.transpose(2,0,1)
if 'int' in str(xdi.dtype): xdi = (xdi/xdi.max()).astype('float32')
if not self.transforms or "ToTensor" not in self.transforms_names:
xdi = xdi.transpose(2, 0, 1)
if "int" in str(xdi.dtype):
xdi = (xdi / xdi.max()).astype("float32")
# if ToTensor() is included, simply apply transforms
if 'ToTensor' in self.transforms_names: xdi = self.transforms(xdi)
# else apply transforms on the result of calling torch.Tensor on
if "ToTensor" in self.transforms_names:
xdi = self.transforms(xdi)
# else apply transforms on the result of calling torch.tensor on
# xdi after all the previous manipulation
elif self.transforms: xdi = self.transforms(torch.Tensor(xdi))
elif self.transforms:
xdi = self.transforms(torch.tensor(xdi))
# fill the Bunch
X.deepimage = xdi
if self.Y is not None:
y = self.Y[idx]
y = self.Y[idx]
return X, y
else:
return X
......
pytorch_widedeep/models/deep_dense.py
浏览文件 @ 263663e5
......@@ -5,9 +5,10 @@ from torch import nn
from ..wdtypes import *
def dense_layer(inp:int, out:int, p:float=0., bn=False):
def dense_layer(inp: int, out: int, p: float = 0.0, bn=False):
layers = [nn.Linear(inp, out), nn.LeakyReLU(inplace=True)]
if bn: layers.append(nn.BatchNorm1d(out))
if bn:
layers.append(nn.BatchNorm1d(out))
layers.append(nn.Dropout(p))
return nn.Sequential(*layers)
......@@ -70,14 +71,17 @@ class DeepDense(nn.Module):
[ 6.7187e-02, -1.2821e-03, -3.0960e-04, 3.6123e-01]],
grad_fn=<LeakyReluBackward1>)
"""
def __init__(self,
deep_column_idx:Dict[str,int],
hidden_layers:List[int],
batchnorm:bool=False,
dropout:Optional[List[float]]=None,
embed_input:Optional[List[Tuple[str,int,int]]]=None,
embed_p:float=0.,
continuous_cols:Optional[List[str]]=None):
def __init__(
self,
deep_column_idx: Dict[str, int],
hidden_layers: List[int],
batchnorm: bool = False,
dropout: Optional[List[float]] = None,
embed_input: Optional[List[Tuple[str, int, int]]] = None,
embed_p: float = 0.0,
continuous_cols: Optional[List[str]] = None,
):
super(DeepDense, self).__init__()
self.embed_input = embed_input
......@@ -86,38 +90,52 @@ class DeepDense(nn.Module):
# Embeddings
if self.embed_input is not None:
self.embed_layers = nn.ModuleDict({'emb_layer_'+col: nn.Embedding(val, dim)
for col, val, dim in self.embed_input})
self.embed_layers = nn.ModuleDict(
{
"emb_layer_" + col: nn.Embedding(val, dim)
for col, val, dim in self.embed_input
}
)
self.embed_dropout = nn.Dropout(embed_p)
emb_inp_dim = np.sum([embed[2] for embed in self.embed_input])
else:
emb_inp_dim = 0
# Continuous
if self.continuous_cols is not None: cont_inp_dim = len(self.continuous_cols)
else: cont_inp_dim = 0
if self.continuous_cols is not None:
cont_inp_dim = len(self.continuous_cols)
else:
cont_inp_dim = 0
# Dense Layers
input_dim = emb_inp_dim + cont_inp_dim
hidden_layers = [input_dim] + hidden_layers
if not dropout: dropout = [0.]*len(hidden_layers)
if not dropout:
dropout = [0.0] * len(hidden_layers)
self.dense = nn.Sequential()
for i in range(1, len(hidden_layers)):
self.dense.add_module(
'dense_layer_{}'.format(i-1),
dense_layer( hidden_layers[i-1], hidden_layers[i], dropout[i-1], batchnorm))
"dense_layer_{}".format(i - 1),
dense_layer(
hidden_layers[i - 1], hidden_layers[i], dropout[i - 1], batchnorm
),
)
# the output_dim attribute will be used as input_dim when "merging" the models
self.output_dim = hidden_layers[-1]
def forward(self, X:Tensor)->Tensor:
def forward(self, X: Tensor) -> Tensor: # type: ignore
if self.embed_input is not None:
x = [self.embed_layers['emb_layer_'+col](X[:,self.deep_column_idx[col]].long())
for col,_,_ in self.embed_input]
x = torch.cat(x, 1)
x = self.embed_dropout(x)
x = [
self.embed_layers["emb_layer_" + col](
X[:, self.deep_column_idx[col]].long()
)
for col, _, _ in self.embed_input
]
x = torch.cat(x, 1) # type: ignore
x = self.embed_dropout(x) # type: ignore
if self.continuous_cols is not None:
cont_idx = [self.deep_column_idx[col] for col in self.continuous_cols]
x_cont = X[:, cont_idx].float()
x = torch.cat([x, x_cont], 1) if self.embed_input is not None else x_cont
return self.dense(x)
\ No newline at end of file
x = torch.cat([x, x_cont], 1) if self.embed_input is not None else x_cont # type: ignore
return self.dense(x) # type: ignore
pytorch_widedeep/models/deep_image.py
浏览文件 @ 263663e5
import torch
from ..wdtypes import *
from .deep_dense import dense_layer
......@@ -8,14 +6,23 @@ from torch import nn
from torchvision import models
def conv_layer(ni:int, nf:int, ks:int=3, stride:int=1, maxpool:bool=True,
adaptiveavgpool:bool=False):
def conv_layer(
ni: int,
nf: int,
ks: int = 3,
stride: int = 1,
maxpool: bool = True,
adaptiveavgpool: bool = False,
):
layer = nn.Sequential(
nn.Conv2d(ni, nf, kernel_size=ks, bias=True, stride=stride, padding=ks//2),
nn.Conv2d(ni, nf, kernel_size=ks, bias=True, stride=stride, padding=ks // 2),
nn.BatchNorm2d(nf, momentum=0.01),
nn.LeakyReLU(negative_slope=0.1, inplace=True))
if maxpool: layer.add_module('maxpool', nn.MaxPool2d(2, 2))
if adaptiveavgpool: layer.add_module('adaptiveavgpool', nn.AdaptiveAvgPool2d(output_size=(1, 1)))
nn.LeakyReLU(negative_slope=0.1, inplace=True),
)
if maxpool:
layer.add_module("maxpool", nn.MaxPool2d(2, 2))
if adaptiveavgpool:
layer.add_module("adaptiveavgpool", nn.AdaptiveAvgPool2d(output_size=(1, 1)))
return layer
......@@ -75,23 +82,25 @@ class DeepImage(nn.Module):
1.5416e-01, 3.9227e-01, 5.5048e-01]], grad_fn=<LeakyReluBackward1>)
"""
def __init__(self,
pretrained:bool=True,
resnet:int=18,
freeze:Union[str,int]=6,
head_layers:Optional[List[int]] = None,
head_dropout:Optional[List[float]]=None,
head_batchnorm:Optional[bool] = False):
def __init__(
self,
pretrained: bool = True,
resnet: int = 18,
freeze: Union[str, int] = 6,
head_layers: Optional[List[int]] = None,
head_dropout: Optional[List[float]] = None,
head_batchnorm: Optional[bool] = False,
):
super(DeepImage, self).__init__()
self.head_layers = head_layers
if pretrained:
if resnet==18:
if resnet == 18:
vision_model = models.resnet18(pretrained=True)
elif resnet==34:
elif resnet == 34:
vision_model = models.resnet34(pretrained=True)
elif resnet==50:
elif resnet == 50:
vision_model = models.resnet50(pretrained=True)
backbone_layers = list(vision_model.children())[:-1]
......@@ -104,7 +113,9 @@ class DeepImage(nn.Module):
frozen_layers.append(layer)
self.backbone = nn.Sequential(*frozen_layers)
if isinstance(freeze, int):
assert freeze < 8, "freeze' must be less than 8 when using resnet architectures"
assert (
freeze < 8
), "freeze' must be less than 8 when using resnet architectures"
frozen_layers = []
trainable_layers = backbone_layers[freeze:]
for layer in backbone_layers[:freeze]:
......@@ -120,30 +131,38 @@ class DeepImage(nn.Module):
conv_layer(64, 128, 1, maxpool=False),
conv_layer(128, 256, 1, maxpool=False),
conv_layer(256, 512, 1, maxpool=False, adaptiveavgpool=True),
)
)
# the output_dim attribute will be used as input_dim when "merging" the models
self.output_dim = 512
if self.head_layers is not None:
assert self.head_layers[0]==self.output_dim, (
assert self.head_layers[0] == self.output_dim, (
"The output dimension from the backbone ({}) is not consistent with "
"the expected input dimension ({}) of the fc-head".format(
self.output_dim, self.head_layers[0]))
if not head_dropout: head_dropout = [0.]*len(head_layers)
self.output_dim, self.head_layers[0]
)
)
if not head_dropout:
head_dropout = [0.0] * len(head_layers)
self.imagehead = nn.Sequential()
for i in range(1, len(head_layers)):
self.imagehead.add_module(
'dense_layer_{}'.format(i-1),
dense_layer(head_layers[i-1], head_layers[i], head_dropout[i-1], head_batchnorm)
)
"dense_layer_{}".format(i - 1),
dense_layer(
head_layers[i - 1],
head_layers[i],
head_dropout[i - 1],
head_batchnorm,
),
)
self.output_dim = head_layers[-1]
def forward(self, x:Tensor)->Tensor:
def forward(self, x: Tensor) -> Tensor: # type: ignore
x = self.backbone(x)
x = x.view(x.size(0), -1)
if self.head_layers is not None:
out = self.imagehead(x)
return out
else:
return x
\ No newline at end of file
return x
pytorch_widedeep/models/deep_text.py
浏览文件 @ 263663e5
......@@ -67,73 +67,104 @@ class DeepText(nn.Module):
[-0.0844, 0.0681, -0.1016, -0.0464],
[-0.0268, 0.0294, -0.0988, -0.0666]], grad_fn=<SelectBackward>)
"""
def __init__(self,
vocab_size:int,
hidden_dim:int=64,
n_layers:int=3,
rnn_dropout:float=0.,
bidirectional:bool=False,
padding_idx:int=1,
embed_dim:Optional[int]=None,
embedding_matrix:Optional[np.ndarray]=None,
head_layers:Optional[List[int]] = None,
head_dropout:Optional[List[float]]=None,
head_batchnorm:Optional[bool] = False):
def __init__(
self,
vocab_size: int,
hidden_dim: int = 64,
n_layers: int = 3,
rnn_dropout: float = 0.0,
bidirectional: bool = False,
padding_idx: int = 1,
embed_dim: Optional[int] = None,
embedding_matrix: Optional[np.ndarray] = None,
head_layers: Optional[List[int]] = None,
head_dropout: Optional[List[float]] = None,
head_batchnorm: Optional[bool] = False,
):
super(DeepText, self).__init__()
if embed_dim is not None and embedding_matrix is not None and not embed_dim==embedding_matrix.shape[1]:
if (
embed_dim is not None
and embedding_matrix is not None
and not embed_dim == embedding_matrix.shape[1]
):
warnings.warn(
'the input embedding dimension {} and the dimension of the '
'pretrained embeddings {} do not match. The pretrained embeddings '
'dimension ({}) will be used'.format(embed_dim, embedding_matrix.shape[1],
embedding_matrix.shape[1]), UserWarning)
"the input embedding dimension {} and the dimension of the "
"pretrained embeddings {} do not match. The pretrained embeddings "
"dimension ({}) will be used".format(
embed_dim, embedding_matrix.shape[1], embedding_matrix.shape[1]
),
UserWarning,
)
self.bidirectional = bidirectional
self.head_layers = head_layers
# Pre-trained Embeddings
if isinstance(embedding_matrix, np.ndarray):
self.word_embed = nn.Embedding(vocab_size, embedding_matrix.shape[1], padding_idx = padding_idx)
self.word_embed.weight = nn.Parameter(torch.Tensor(embedding_matrix))
assert (
embedding_matrix.dtype == "float32"
), "'embedding_matrix' must be of dtype 'float32', got dtype '{}'".format(
str(embedding_matrix.dtype)
)
self.word_embed = nn.Embedding(
vocab_size, embedding_matrix.shape[1], padding_idx=padding_idx
)
self.word_embed.weight = nn.Parameter(
torch.tensor(embedding_matrix), requires_grad=True
)
embed_dim = embedding_matrix.shape[1]
else:
self.word_embed = nn.Embedding(vocab_size, embed_dim, padding_idx = padding_idx)
self.word_embed = nn.Embedding(
vocab_size, embed_dim, padding_idx=padding_idx
)
# stack of RNNs (LSTMs)
self.rnn = nn.LSTM(embed_dim,
self.rnn = nn.LSTM(
embed_dim,
hidden_dim,
num_layers=n_layers,
bidirectional=bidirectional,
dropout=rnn_dropout,
batch_first=True)
batch_first=True,
)
# the output_dim attribute will be used as input_dim when "merging" the models
self.output_dim = hidden_dim*2 if bidirectional else hidden_dim
self.output_dim = hidden_dim * 2 if bidirectional else hidden_dim
if self.head_layers is not None:
assert self.head_layers[0]==self.output_dim, (
assert self.head_layers[0] == self.output_dim, (
"The hidden dimension from the stack or RNNs ({}) is not consistent with "
"the expected input dimension ({}) of the fc-head".format(
self.output_dim, self.head_layers[0]))
if not head_dropout: head_dropout = [0.]*len(head_layers)
self.output_dim, self.head_layers[0]
)
)
if not head_dropout:
head_dropout = [0.0] * len(head_layers)
self.texthead = nn.Sequential()
for i in range(1, len(head_layers)):
self.texthead.add_module(
'dense_layer_{}'.format(i-1),
dense_layer(head_layers[i-1], head_layers[i], head_dropout[i-1], head_batchnorm)
)
"dense_layer_{}".format(i - 1),
dense_layer(
head_layers[i - 1],
head_layers[i],
head_dropout[i - 1],
head_batchnorm,
),
)
self.output_dim = head_layers[-1]
def forward(self, X:Tensor)->Tensor:
def forward(self, X: Tensor) -> Tensor: # type: ignore
embed = self.word_embed(X.long())
o, (h, c) = self.rnn(embed)
if self.bidirectional:
last_h = torch.cat((h[-2], h[-1]), dim = 1)
last_h = torch.cat((h[-2], h[-1]), dim=1)
else:
last_h = h[-1]
if self.head_layers is not None:
out = self.head(last_h)
out = self.texthead(last_h)
return out
else:
return last_h
\ No newline at end of file
return last_h
pytorch_widedeep/models/wide.py
浏览文件 @ 263663e5
import torch
from torch import nn
from ..wdtypes import *
......@@ -10,32 +8,33 @@ class Wide(nn.Module):
Parameters
----------
wide_dim: Int
wide_dim: Int
size of the input tensor
output_dim: Int
output_dim: Int
size of the ouput tensor
Attributes
----------
wide_linear: nn.Module
wide_linear: nn.Module
the linear layer that comprises the wide branch of the model
Example
--------
>>> import torch
>>> from pytorch_widedeep.models import Wide
>>> X = torch.empty(4, 4).random_(2)
>>> wide = Wide(wide_dim=X.size(0), output_dim=1)
>>> wide(X)
tensor([[-0.8841],
[-0.8633],
[-1.2713],
[-0.4762]], grad_fn=<AddmmBackward>)
>>> import torch
>>> from pytorch_widedeep.models import Wide
>>> X = torch.empty(4, 4).random_(2)
>>> wide = Wide(wide_dim=X.size(0), output_dim=1)
>>> wide(X)
tensor([[-0.8841],
[-0.8633],
[-1.2713],
[-0.4762]], grad_fn=<AddmmBackward>)
"""
def __init__(self,wide_dim:int, output_dim:int=1):
def __init__(self, wide_dim: int, output_dim: int = 1):
super(Wide, self).__init__()
self.wide_linear = nn.Linear(wide_dim, output_dim)
def forward(self, X:Tensor)->Tensor:
def forward(self, X: Tensor) -> Tensor: # type: ignore
out = self.wide_linear(X.float())
return out
pytorch_widedeep/optim/radam.py
浏览文件 @ 263663e5
import math
import torch
from torch.optim.optimizer import Optimizer, required
from torch.optim.optimizer import Optimizer
class RAdam(Optimizer):
class RAdam(Optimizer):
def __init__(self, params, lr=1e-3, betas=(0.9, 0.999), eps=1e-8, weight_decay=0):
if not 0.0 <= lr:
raise ValueError("Invalid learning rate: {}".format(lr))
......@@ -29,65 +29,73 @@ class RAdam(Optimizer):
for group in self.param_groups:
for p in group['params']:
for p in group["params"]:
if p.grad is None:
continue
grad = p.grad.data.float()
if grad.is_sparse:
raise RuntimeError('RAdam does not support sparse gradients')
raise RuntimeError("RAdam does not support sparse gradients")
p_data_fp32 = p.data.float()
state = self.state[p]
if len(state) == 0:
state['step'] = 0
state['exp_avg'] = torch.zeros_like(p_data_fp32)
state['exp_avg_sq'] = torch.zeros_like(p_data_fp32)
state["step"] = 0
state["exp_avg"] = torch.zeros_like(p_data_fp32)
state["exp_avg_sq"] = torch.zeros_like(p_data_fp32)
else:
state['exp_avg'] = state['exp_avg'].type_as(p_data_fp32)
state['exp_avg_sq'] = state['exp_avg_sq'].type_as(p_data_fp32)
state["exp_avg"] = state["exp_avg"].type_as(p_data_fp32)
state["exp_avg_sq"] = state["exp_avg_sq"].type_as(p_data_fp32)
exp_avg, exp_avg_sq = state['exp_avg'], state['exp_avg_sq']
beta1, beta2 = group['betas']
exp_avg, exp_avg_sq = state["exp_avg"], state["exp_avg_sq"]
beta1, beta2 = group["betas"]
exp_avg_sq.mul_(beta2).addcmul_(1 - beta2, grad, grad)
exp_avg.mul_(beta1).add_(1 - beta1, grad)
state['step'] += 1
buffered = self.buffer[int(state['step'] % 10)]
if state['step'] == buffered[0]:
state["step"] += 1
buffered = self.buffer[int(state["step"] % 10)]
if state["step"] == buffered[0]:
N_sma, step_size = buffered[1], buffered[2]
else:
buffered[0] = state['step']
beta2_t = beta2 ** state['step']
buffered[0] = state["step"]
beta2_t = beta2 ** state["step"]
N_sma_max = 2 / (1 - beta2) - 1
N_sma = N_sma_max - 2 * state['step'] * beta2_t / (1 - beta2_t)
N_sma = N_sma_max - 2 * state["step"] * beta2_t / (1 - beta2_t)
buffered[1] = N_sma
# more conservative since it's an approximated value
if N_sma >= 5:
step_size = math.sqrt((1 - beta2_t) * (N_sma - 4) / (N_sma_max - 4) * (N_sma - 2) / N_sma * N_sma_max / (N_sma_max - 2)) / (1 - beta1 ** state['step'])
step_size = math.sqrt(
(1 - beta2_t)
* (N_sma - 4)
/ (N_sma_max - 4)
* (N_sma - 2)
/ N_sma
* N_sma_max
/ (N_sma_max - 2)
) / (1 - beta1 ** state["step"])
else:
step_size = 1.0 / (1 - beta1 ** state['step'])
step_size = 1.0 / (1 - beta1 ** state["step"])
buffered[2] = step_size
if group['weight_decay'] != 0:
p_data_fp32.add_(-group['weight_decay'] * group['lr'], p_data_fp32)
if group["weight_decay"] != 0:
p_data_fp32.add_(-group["weight_decay"] * group["lr"], p_data_fp32)
# more conservative since it's an approximated value
if N_sma >= 5:
denom = exp_avg_sq.sqrt().add_(group['eps'])
p_data_fp32.addcdiv_(-step_size * group['lr'], exp_avg, denom)
denom = exp_avg_sq.sqrt().add_(group["eps"])
p_data_fp32.addcdiv_(-step_size * group["lr"], exp_avg, denom)
else:
p_data_fp32.add_(-step_size * group['lr'], exp_avg)
p_data_fp32.add_(-step_size * group["lr"], exp_avg)
p.data.copy_(p_data_fp32)
return loss
class PlainRAdam(Optimizer):
class PlainRAdam(Optimizer):
def __init__(self, params, lr=1e-3, betas=(0.9, 0.999), eps=1e-8, weight_decay=0):
if not 0.0 <= lr:
raise ValueError("Invalid learning rate: {}".format(lr))
......@@ -113,46 +121,58 @@ class PlainRAdam(Optimizer):
for group in self.param_groups:
for p in group['params']:
for p in group["params"]:
if p.grad is None:
continue
grad = p.grad.data.float()
if grad.is_sparse:
raise RuntimeError('RAdam does not support sparse gradients')
raise RuntimeError("RAdam does not support sparse gradients")
p_data_fp32 = p.data.float()
state = self.state[p]
if len(state) == 0:
state['step'] = 0
state['exp_avg'] = torch.zeros_like(p_data_fp32)
state['exp_avg_sq'] = torch.zeros_like(p_data_fp32)
state["step"] = 0
state["exp_avg"] = torch.zeros_like(p_data_fp32)
state["exp_avg_sq"] = torch.zeros_like(p_data_fp32)
else:
state['exp_avg'] = state['exp_avg'].type_as(p_data_fp32)
state['exp_avg_sq'] = state['exp_avg_sq'].type_as(p_data_fp32)
state["exp_avg"] = state["exp_avg"].type_as(p_data_fp32)
state["exp_avg_sq"] = state["exp_avg_sq"].type_as(p_data_fp32)
exp_avg, exp_avg_sq = state['exp_avg'], state['exp_avg_sq']
beta1, beta2 = group['betas']
exp_avg, exp_avg_sq = state["exp_avg"], state["exp_avg_sq"]
beta1, beta2 = group["betas"]
exp_avg_sq.mul_(beta2).addcmul_(1 - beta2, grad, grad)
exp_avg.mul_(beta1).add_(1 - beta1, grad)
state['step'] += 1
beta2_t = beta2 ** state['step']
state["step"] += 1
beta2_t = beta2 ** state["step"]
N_sma_max = 2 / (1 - beta2) - 1
N_sma = N_sma_max - 2 * state['step'] * beta2_t / (1 - beta2_t)
N_sma = N_sma_max - 2 * state["step"] * beta2_t / (1 - beta2_t)
if group['weight_decay'] != 0:
p_data_fp32.add_(-group['weight_decay'] * group['lr'], p_data_fp32)
if group["weight_decay"] != 0:
p_data_fp32.add_(-group["weight_decay"] * group["lr"], p_data_fp32)
# more conservative since it's an approximated value
if N_sma >= 5:
step_size = group['lr'] * math.sqrt((1 - beta2_t) * (N_sma - 4) / (N_sma_max - 4) * (N_sma - 2) / N_sma * N_sma_max / (N_sma_max - 2)) / (1 - beta1 ** state['step'])
denom = exp_avg_sq.sqrt().add_(group['eps'])
step_size = (
group["lr"]
* math.sqrt(
(1 - beta2_t)
* (N_sma - 4)
/ (N_sma_max - 4)
* (N_sma - 2)
/ N_sma
* N_sma_max
/ (N_sma_max - 2)
)
/ (1 - beta1 ** state["step"])
)
denom = exp_avg_sq.sqrt().add_(group["eps"])
p_data_fp32.addcdiv_(-step_size, exp_avg, denom)
else:
step_size = group['lr'] / (1 - beta1 ** state['step'])
step_size = group["lr"] / (1 - beta1 ** state["step"])
p_data_fp32.add_(-step_size, exp_avg)
p.data.copy_(p_data_fp32)
......@@ -161,8 +181,9 @@ class PlainRAdam(Optimizer):
class AdamW(Optimizer):
def __init__(self, params, lr=1e-3, betas=(0.9, 0.999), eps=1e-8, weight_decay=0, warmup = 0):
def __init__(
self, params, lr=1e-3, betas=(0.9, 0.999), eps=1e-8, weight_decay=0, warmup=0
):
if not 0.0 <= lr:
raise ValueError("Invalid learning rate: {}".format(lr))
if not 0.0 <= eps:
......@@ -172,8 +193,9 @@ class AdamW(Optimizer):
if not 0.0 <= betas[1] < 1.0:
raise ValueError("Invalid beta parameter at index 1: {}".format(betas[1]))
defaults = dict(lr=lr, betas=betas, eps=eps,
weight_decay=weight_decay, warmup = warmup)
defaults = dict(
lr=lr, betas=betas, eps=eps, weight_decay=weight_decay, warmup=warmup
)
super(AdamW, self).__init__(params, defaults)
def __setstate__(self, state):
......@@ -186,49 +208,53 @@ class AdamW(Optimizer):
for group in self.param_groups:
for p in group['params']:
for p in group["params"]:
if p.grad is None:
continue
grad = p.grad.data.float()
if grad.is_sparse:
raise RuntimeError('Adam does not support sparse gradients, please consider SparseAdam instead')
raise RuntimeError(
"Adam does not support sparse gradients, please consider SparseAdam instead"
)
p_data_fp32 = p.data.float()
state = self.state[p]
if len(state) == 0:
state['step'] = 0
state['exp_avg'] = torch.zeros_like(p_data_fp32)
state['exp_avg_sq'] = torch.zeros_like(p_data_fp32)
state["step"] = 0
state["exp_avg"] = torch.zeros_like(p_data_fp32)
state["exp_avg_sq"] = torch.zeros_like(p_data_fp32)
else:
state['exp_avg'] = state['exp_avg'].type_as(p_data_fp32)
state['exp_avg_sq'] = state['exp_avg_sq'].type_as(p_data_fp32)
state["exp_avg"] = state["exp_avg"].type_as(p_data_fp32)
state["exp_avg_sq"] = state["exp_avg_sq"].type_as(p_data_fp32)
exp_avg, exp_avg_sq = state['exp_avg'], state['exp_avg_sq']
beta1, beta2 = group['betas']
exp_avg, exp_avg_sq = state["exp_avg"], state["exp_avg_sq"]
beta1, beta2 = group["betas"]
state['step'] += 1
state["step"] += 1
exp_avg_sq.mul_(beta2).addcmul_(1 - beta2, grad, grad)
exp_avg.mul_(beta1).add_(1 - beta1, grad)
denom = exp_avg_sq.sqrt().add_(group['eps'])
bias_correction1 = 1 - beta1 ** state['step']
bias_correction2 = 1 - beta2 ** state['step']
denom = exp_avg_sq.sqrt().add_(group["eps"])
bias_correction1 = 1 - beta1 ** state["step"]
bias_correction2 = 1 - beta2 ** state["step"]
if group['warmup'] > state['step']:
scheduled_lr = 1e-8 + state['step'] * group['lr'] / group['warmup']
if group["warmup"] > state["step"]:
scheduled_lr = 1e-8 + state["step"] * group["lr"] / group["warmup"]
else:
scheduled_lr = group['lr']
scheduled_lr = group["lr"]
step_size = scheduled_lr * math.sqrt(bias_correction2) / bias_correction1
step_size = (
scheduled_lr * math.sqrt(bias_correction2) / bias_correction1
)
if group['weight_decay'] != 0:
p_data_fp32.add_(-group['weight_decay'] * scheduled_lr, p_data_fp32)
if group["weight_decay"] != 0:
p_data_fp32.add_(-group["weight_decay"] * scheduled_lr, p_data_fp32)
p_data_fp32.addcdiv_(-step_size, exp_avg, denom)
p.data.copy_(p_data_fp32)
return loss
\ No newline at end of file
return loss
pytorch_widedeep/preprocessing/__init__.py
浏览文件 @ 263663e5
from ._preprocessors import WidePreprocessor
from ._preprocessors import DeepPreprocessor
from ._preprocessors import TextPreprocessor
from ._preprocessors import ImagePreprocessor
\ No newline at end of file
from ._preprocessors import ImagePreprocessor
pytorch_widedeep/utils/__init__.py
浏览文件 @ 263663e5
from .dense_utils import *
from .text_utils import *
from .fastai_transforms import *
from .image_utils import *
\ No newline at end of file
from .image_utils import *
pytorch_widedeep/utils/dense_utils.py
浏览文件 @ 263663e5
import numpy as np
import pandas as pd
from ..wdtypes import *
......@@ -7,11 +6,14 @@ from ..wdtypes import *
pd.options.mode.chained_assignment = None
__all__ = ['label_encoder']
__all__ = ["label_encoder"]
def label_encoder(df_inp:pd.DataFrame, cols:Optional[List[str]]=None,
val_to_idx:Optional[Dict[str,Dict[str,int]]]=None):
def label_encoder(
df_inp: pd.DataFrame,
cols: Optional[List[str]] = None,
val_to_idx: Optional[Dict[str, Dict[str, int]]] = None,
):
r"""
Label-encode some features of a given dataset.
......@@ -33,12 +35,12 @@ def label_encoder(df_inp:pd.DataFrame, cols:Optional[List[str]]=None,
"""
df = df_inp.copy()
if cols == None:
cols = list(df.select_dtypes(include=['object']).columns)
if cols is None:
cols = list(df.select_dtypes(include=["object"]).columns)
if not val_to_idx:
val_types = dict()
for c in cols:
for c in cols: # type: ignore
val_types[c] = df[c].unique()
val_to_idx = dict()
for k, v in val_types.items():
......@@ -47,4 +49,4 @@ def label_encoder(df_inp:pd.DataFrame, cols:Optional[List[str]]=None,
for k, v in val_to_idx.items():
df[k] = df[k].apply(lambda x: v[x])
return df, val_to_idx
\ No newline at end of file
return df, val_to_idx
pytorch_widedeep/utils/image_utils.py
浏览文件 @ 263663e5
'''
"""
AspectAwarePreprocessor and SimplePreprocessor are directly taked from the
great series of Books "Deep Learning for Computer Vision" by Adrian
(https://www.pyimagesearch.com/author/adrian/). Check here
https://www.pyimagesearch.com/
Credit for the code here to ADRIAN ROSEBROCK
'''
"""
import numpy as np
import imutils
import cv2
from ..wdtypes import *
__all__ = ['AspectAwarePreprocessor', 'SimplePreprocessor']
__all__ = ["AspectAwarePreprocessor", "SimplePreprocessor"]
class AspectAwarePreprocessor:
def __init__(self, width:int, height:int, inter=cv2.INTER_AREA):
def __init__(self, width: int, height: int, inter=cv2.INTER_AREA):
self.width = width
self.height = height
self.inter = inter
def preprocess(self, image:np.ndarray)->np.ndarray:
def preprocess(self, image: np.ndarray) -> np.ndarray:
(h, w) = image.shape[:2]
dW = 0
dH = 0
if w < h:
image = imutils.resize(image, width=self.width,
inter=self.inter)
image = imutils.resize(image, width=self.width, inter=self.inter)
dH = int((image.shape[0] - self.height) / 2.0)
else:
image = imutils.resize(image, height=self.height,
inter=self.inter)
image = imutils.resize(image, height=self.height, inter=self.inter)
dW = int((image.shape[1] - self.width) / 2.0)
(h, w) = image.shape[:2]
image = image[dH:h - dH, dW:w - dW]
image = image[dH : h - dH, dW : w - dW]
return cv2.resize(image, (self.width, self.height),
interpolation=self.inter)
return cv2.resize(image, (self.width, self.height), interpolation=self.inter)
class SimplePreprocessor:
def __init__(self, width:int, height:int, inter=cv2.INTER_AREA):
def __init__(self, width: int, height: int, inter=cv2.INTER_AREA):
self.width = width
self.height = height
self.inter = inter
def preprocess(self, image:np.ndarray)->np.ndarray:
return cv2.resize(image, (self.width, self.height),
interpolation=self.inter)
def preprocess(self, image: np.ndarray) -> np.ndarray:
return cv2.resize(image, (self.width, self.height), interpolation=self.inter)
pytorch_widedeep/version.py
浏览文件 @ 263663e5
__version__ = '0.3.8'
\ No newline at end of file
__version__ = "0.3.8"
pytorch_widedeep/wdtypes.py
浏览文件 @ 263663e5
......@@ -2,23 +2,83 @@ import sys
from torch.nn import Module
from torch import Tensor
from torchvision.transforms import *
from torchvision.transforms import (
CenterCrop,
ColorJitter,
Compose,
FiveCrop,
Grayscale,
Lambda,
LinearTransformation,
Normalize,
Pad,
RandomAffine,
RandomApply,
RandomChoice,
RandomCrop,
RandomGrayscale,
RandomHorizontalFlip,
RandomOrder,
RandomResizedCrop,
RandomRotation,
RandomSizedCrop,
RandomVerticalFlip,
Resize,
Scale,
TenCrop,
ToPILImage,
ToTensor,
)
from torch.optim.optimizer import Optimizer
from torch.utils.data.dataloader import DataLoader
from torch.optim.lr_scheduler import _LRScheduler
from pathlib import PosixPath
from typing import (List, Any, Union, Dict, Callable, Optional, Tuple,
Generator, Collection, Iterable)
from typing import (
List,
Any,
Union,
Dict,
Callable,
Optional,
Tuple,
Generator,
Collection,
Iterable,
Match,
Iterator,
)
from scipy.sparse.csr import csr_matrix as sparse_matrix
from types import SimpleNamespace
SimpleNamespace = type(sys.implementation)
ListRules = Collection[Callable[[str],str]]
ListRules = Collection[Callable[[str], str]]
Tokens = Collection[Collection[str]]
Transforms= Union[CenterCrop, ColorJitter, Compose, FiveCrop, Grayscale,
Lambda, LinearTransformation, Normalize, Pad, RandomAffine,
RandomApply, RandomChoice, RandomCrop, RandomGrayscale,
RandomHorizontalFlip, RandomOrder, RandomResizedCrop, RandomRotation,
RandomSizedCrop, RandomVerticalFlip, Resize, Scale, TenCrop,
ToPILImage, ToTensor]
Transforms = Union[
CenterCrop,
ColorJitter,
Compose,
FiveCrop,
Grayscale,
Lambda,
LinearTransformation,
Normalize,
Pad,
RandomAffine,
RandomApply,
RandomChoice,
RandomCrop,
RandomGrayscale,
RandomHorizontalFlip,
RandomOrder,
RandomResizedCrop,
RandomRotation,
RandomSizedCrop,
RandomVerticalFlip,
Resize,
Scale,
TenCrop,
ToPILImage,
ToTensor,
]
LRScheduler = _LRScheduler
ModelParams = Generator[Tensor,Tensor,Tensor]
ModelParams = Generator[Tensor, Tensor, Tensor]
setup.py
浏览文件 @ 263663e5
此差异已折叠。
tests/test_data_utils/test_du_deep_image.py
浏览文件 @ 263663e5
import numpy as np
import pandas as pd
import os
import pytest
from pytorch_widedeep.preprocessing import ImagePreprocessor
full_path = os.path.realpath(__file__)
path = os.path.split(full_path)[0]
df = pd.DataFrame({'galaxies': ['galaxy1.png', 'galaxy2.png']})
img_col = 'galaxies'
imd_dir = os.path.join(path,'images')
processor = ImagePreprocessor()
X_imgs = processor.fit_transform(df, img_col, img_path=imd_dir)
df = pd.DataFrame({"galaxies": ["galaxy1.png", "galaxy2.png"]})
img_col = "galaxies"
imd_dir = os.path.join(path, "images")
processor = ImagePreprocessor(img_col=img_col, img_path=imd_dir)
X_imgs = processor.fit_transform(df)
###############################################################################
# There is not much to test here, since I only resize.
###############################################################################
def test_sizes():
img_width = X_imgs.shape[1]
img_height = X_imgs.shape[2]
assert np.all((img_width==processor.width, img_height==processor.height))
\ No newline at end of file
img_width = X_imgs.shape[1]
img_height = X_imgs.shape[2]
assert np.all((img_width == processor.width, img_height == processor.height))
tests/test_model_components/test_mc_wide.py
浏览文件 @ 263663e5
import torch
import pytest
from pytorch_widedeep.models import Wide
inp = torch.rand(10, 10)
model = Wide(10, 1)
###############################################################################
# Simply testing that it runs
###############################################################################
def test_wide():
out = model(inp)
assert out.size(0) == 10 and out.size(1) == 1
\ No newline at end of file
out = model(inp)
assert out.size(0) == 10 and out.size(1) == 1
Markdown is supported
0% 或 .
You are about to add 0 people to the discussion. Proceed with caution.
先完成此消息的编辑！
想要评论请 注册 或