diff --git a/fastai_training_job/train_job_image/Dockerfile b/fastai_training_job/train_job_image/Dockerfile
new file mode 100644
index 0000000..c1417e8
--- /dev/null
+++ b/fastai_training_job/train_job_image/Dockerfile
@@ -0,0 +1,9 @@
+
+FROM fastdotai/fastai@sha256:c36b43104474006d8f8cd2a65f740bfd505693c670644c1d2dbedb5a6fb2de8a
+RUN pip install -U fire==0.4.0 pandas==1.3.5 google-cloud-pubsub==2.13.0 google-cloud-storage==1.35.0 gcsfs==2022.5.0  
+WORKDIR /app
+COPY train.py .
+COPY test_model.py .
+COPY gcp_utils.py .
+
+ENTRYPOINT ["python", "train.py"]
diff --git a/fastai_training_job/train_job_image/model_hyperparams_test.json b/fastai_training_job/train_job_image/model_hyperparams_test.json
new file mode 100644
index 0000000..6abca17
--- /dev/null
+++ b/fastai_training_job/train_job_image/model_hyperparams_test.json
@@ -0,0 +1,10 @@
+{
+    "TAG_COLUMN": "tag",
+    "RESIZE_VALUE": 64,
+    "METHOD": "squish",
+    "FREEZE_EPOCHS": 1,
+    "BASE_LR": 0.002,
+    "EPOCHS": 0,
+    "ARCH_RESNET": "resnet34",
+    "TEST_THR_DEFAULT": 0.5
+}
\ No newline at end of file
diff --git a/fastai_training_job/train_job_image/train.py b/fastai_training_job/train_job_image/train.py
new file mode 100644
index 0000000..19bdb99
--- /dev/null
+++ b/fastai_training_job/train_job_image/train.py
@@ -0,0 +1,107 @@
+import json
+import os
+import time
+
+import fire
+import numpy as np
+import pandas as pd
+from fastai.vision.all import *
+
+from gcp_utils import publish_metrics
+
+timestamp = time.strftime("%Y%m%d-%H%M%S")
+
+
+def train_evaluate(job_dir: str = None,
+                   training_dataset_path: str = None,
+                   training_images_path: str = None,
+                   model_version: str = None,
+                   model_hyperparms_path: str = None):
+
+    if torch.cuda.is_available():
+        print('GPU available')
+
+    # GCS gets mounted when the instance starts using gcsfuse. All prefixes exist under /gcs/
+    job_dir, training_images_path, model_hyperparms_path = [f.replace(
+        'gs://', '/gcs/') if f.startswith('gs') else f for f in [job_dir, training_images_path, model_hyperparms_path]]
+
+    # create subdir/prefix to export results
+    job_subdir_export = f'{job_dir}/{model_version}_{timestamp}'
+    os.makedirs(job_subdir_export)
+
+    # read hyperparams values
+    with open(model_hyperparms_path) as model_hparams_json:
+        model_hparams_dict = json.load(model_hparams_json)
+        print(f'hyperparameters: {model_hparams_dict}')
+
+    # read training data csv
+    print('Reading csv ...')
+    df_train = pd.read_csv(training_dataset_path)
+
+    print('Loading and transforming images...')
+    # define datablock
+    img_block_01 = DataBlock(blocks=(ImageBlock, MultiCategoryBlock),
+                             splitter=ColSplitter('is_valid'),
+                             get_x=ColReader(
+                                 'image_id', pref=f'{training_images_path}/', suff=''),
+                             get_y=ColReader(
+                                 model_hparams_dict['TAG_COLUMN'], label_delim=','),
+                             item_tfms=Resize(
+                                 model_hparams_dict['RESIZE_VALUE'], method=model_hparams_dict['METHOD']),
+                             batch_tfms=aug_transforms(do_flip=False))
+
+    # load images and transform
+    img_dls_01 = img_block_01.dataloaders(df_train)
+
+    # train
+    print('Starting training...')
+    fscore = F1ScoreMulti()
+    learn_01 = cnn_learner(dls=img_dls_01,
+                           arch=eval(model_hparams_dict['ARCH_RESNET']),
+                           metrics=[accuracy_multi, fscore])
+    learn_01.fine_tune(
+        model_hparams_dict['EPOCHS'],
+        base_lr=model_hparams_dict['BASE_LR'],
+        freeze_epochs=model_hparams_dict['FREEZE_EPOCHS'],
+        cbs=[SaveModelCallback(monitor='accuracy_multi',
+                               fname='model', with_opt=True)]
+    )
+    print('Training completed.')
+
+    # dict to add metrics
+    metrics_log = {}
+
+    # get final model metrics
+    # here accuracy_multi is maximized (can be changed to any other metric like valid_loss)
+    accuracy_multi_train = learn_01.recorder.metrics[0].value.item()
+    metrics_log['accuracy_multi_train'] = accuracy_multi_train
+
+    # save metrics
+    print("Saving metrics")
+    with open(f'{job_subdir_export}/metrics_log.json', "w") as f:
+        json.dump(metrics_log, f)
+    print(metrics_log)
+
+    # pickle model ------
+    model_name = f'model_{model_version}'
+
+    pickle_model_path = f'{job_subdir_export}/{model_name}.pkl'
+    print(f'Saving pickle model to {pickle_model_path}')
+    learn_01.export(f'{pickle_model_path}')
+
+    # jit model ----------
+    # save for native torch import later
+    dummy_inp = torch.randn(
+        [1, 3, model_hparams_dict['RESIZE_VALUE'], model_hparams_dict['RESIZE_VALUE']])  # dummy
+
+    jit_model_path = f'{job_subdir_export}/{model_name}.pt'
+    print(f'Saving jit model to {jit_model_path}')
+    torch.jit.save(torch.jit.trace(learn_01.model, dummy_inp),
+                   f'{job_subdir_export}/model_{model_version}.pt')
+
+    # save vocab with jit model
+    vocab = np.save('models/vocab.npy', learn_01.dls.vocab)
+
+
+if __name__ == "__main__":
+    fire.Fire(train_evaluate)