This is the template of image-classification task for ABEJA Platform.
This template uses transfer-learning from VGG16 ImageNet. Making it easy to try building ML model, this template uses many hard-coding parameters. You can change the parameters by setting environmental variables or editing code directly.
If you want to use GPU, you need to edit "requirements-local.txt" like the below.
# tensorflow==1.13.1
tensorflow-gpu==1.13.1
- Python 3.6.x
- [For local] Install abeja-sdk
- abeja/all-cpu:19.04
- abeja/all-gpu:19.04
- Transfer learning from VGG16 ImageNet.
- Allow only 1 category
- Allow only 1 dataset
| env | type | description |
|---|---|---|
| BATCH_SIZE | int | Batch size. Default 32. |
| EPOCHS | int | Epoch number. This template applies "Early stopping". Default 50. |
| LEARNING_RATE | float | Learning rate. Need to be from 0.0 to 1.0. Default 0.0001. |
| ADAM_BETA_1 | float | Adam parameter "beta_1". Need to be from 0.0 to 1.0. Default 0.9. |
| ADAM_BETA_2 | float | Adam parameter "beta_2". Need to be from 0.0 to 1.0. Default 0.999. |
| ADAM_EPSILON | float | Adam parameter "epsilon". Need to be from 0.0. Default None = K.epsilon(). |
| ADAM_DECAY | float | Adam parameter "decay". Need to be from 0.0. Default 0.0. |
| DROPOUT | float | Dropout of the last layer (Transfer learning). Need to be from 0.0 to 1.0. Default 0.5. |
| DROPOUT_SEED | int | Random seed for Dropout. Default 42. |
| EARLY_STOPPING_TEST_SIZE | float | Test data size for "Early stopping". Need to be from 0.0 to 1.0. Default 0.2. |
| EARLY_STOPPING_PATIENCE | int | Number of patience for "Early stopping". Default 5. |
| IMG_ROWS | int | Image rows. Automatically resize to this size. Default 224. |
| IMG_COLS | int | Image cols. Automatically resize to this size. Default 224. |
| NB_CHANNELS | int | Image channels. If grayscale, then 1. If color, then 3. Default 3. |
| RANDOM_SEED | int | Random seed. Use it for a data shuffling. Default 42. |
| USE_ON_MEMORY | bool | Load data on memory. If you use a big dataset, set it to false. Default true |
| USE_CACHE | bool | Image cache. If you use a big dataset, set it to false. If USE_ON_MEMORY=true, then USE_CACHE=true automatically. Default true |
| NUM_DATA_LOAD_THREAD | int | Number of thread image loads. MUST NOT over BATCH_SIZE. Default 1 |
| ROTATION_RANGE | int | Degree range for random rotations. Default 20 |
| WIDTH_SHIFT_RANGE | float | Fraction of total width, if < 1, or pixels if >= 1. Default 0.05. |
| HEIGHT_SHIFT_RANGE | float | Fraction of total height, if < 1, or pixels if >= 1. Default 0.05. |
| BRIGHTNESS_RANGE | string | CSV format of two floats. Range for picking a brightness shift value from. Default None. |
| SHEAR_RANGE | float | Shear Intensity. Default 0.. |
| ZOOM_RANGE | float | Range for random zoom. [lower, upper] = [1-zoom_range, 1+zoom_range]. Default 0.. |
| CHANNEL_SHIFT_RANGE | float | Range for random channel shifts. Default 0.. |
| FILL_MODE | string | Points outside the boundaries of the input are filled according to the given mode. One of {"constant", "nearest", "reflect" or "wrap"}. Default nearest. |
| CVAL | float | Value used for points outside the boundaries when fill_mode = "constant". Default 0.. |
| HORIZONTAL_FLIP | bool | Randomly flip inputs horizontally. Default True. |
| VERTICAL_FLIP | bool | Randomly flip inputs vertically. Default False. |
| RESCALE | float | Rescaling factor. If 0, no rescaling is applied. Default 0.. |
| DATA_FORMAT | string | Image data format, either "channels_first" or "channels_last". Default channels_last. |
| DTYPE | string | Dtype to use for the generated arrays. Default float32. |
Use requirements-local.txt.
$ pip install -r requirements-local.txt
Set environment variables.
| env | type | description |
|---|---|---|
| ABEJA_ORGANIZATION_ID | str | Your organization ID. |
| ABEJA_PLATFORM_USER_ID | str | Your user ID. |
| ABEJA_PLATFORM_PERSONAL_ACCESS_TOKEN | str | Your Access Token. |
| DATASET_ID | str | Dataset ID. |
$ DATASET_ID='xxx' ABEJA_ORGANIZATION_ID='xxx' ABEJA_PLATFORM_USER_ID='user-xxx' ABEJA_PLATFORM_PERSONAL_ACCESS_TOKEN='xxx' python train.py