Revert "Remove old triton upload"

README.md

@@ -82,8 +82,20 @@ dataset.upload_dataset(task="text_clf", split="train", module_dir="path_to_imdb_
 
 
 ## Model upload examples
-
-There is an exciting new model upload experience now in private preview. We'd love for you to try it out and give us feedback! If you're interested, please sign up for private preview [here](https://forms.gle/MSx7QNxmug2oFZYD6).
+### How to start
+Please refer to this [doc](https://github.com/Clarifai/clarifai-python/tree/master/clarifai/models/model_serving)
+### Examples
+| Model type  |  Example    |
+| ----------- | ----------- |
+| [multimodal-embedder](./model_upload/multimodal_embedder/) | [CLIP](./model_upload/multimodal_embedder/clip/) |
+| [text-classifier](./model_upload/text_classifier) | [xlm-roberta](./model_upload/text_classifier/xlm-roberta/) |
+| [text-embedder](./model_upload/text_embedder/) | [instructor-xl](./model_upload/text_embedder/instructor-xl/) |
+| [text-to-image](./model_upload/text_to_image/) | [sd-v1.5](./model_upload/text_to_image/sd-v1.5/) |
+| [text-to-text](./model_upload/text_to_text/) | [bart-summarize](./model_upload/text_to_text/bart-summarize/), [vllm model](./model_upload/vllm_text_to_text/example/) |
+| [visual_classsifier](./model_upload/visual_classsifier/) | [age_vit](./model_upload/visual_classsifier/age_vit/) |
+| [visual_detector](./model_upload/visual_detector/) | [yolof](./model_upload/visual_detector/yolof/), [faster-rcnn_torchserve](./model_upload/visual_detector/faster-rcnn_torchserve/) |
+| [visual_embedder](./model_upload/visual_embedder) | [vit-base](./model_upload/visual_embedder/vit-base/) |
+| [visual_segmenter](./model_upload/visual_segmenter) | [segformer-b2](./model_upload/visual_segmenter/segformer-b2/) |
 
 
 ## Note

model_upload/README.md

@@ -0,0 +1,19 @@
+## Clarifai Model Upload Examples
+
+A collection of pre-built models for different tasks. To run inference locally using any of the examples here, some models may require other files such as checkpoints be downloaded before testing and/or deployment to Clarifai as they are ommitted here due to github file size limits.
+
+See the Readme files under each model to see detail instruction of how to download the additional files and deploy the models.
+
+## Prerequisite
+
+Install latest `clarifai`
+
+```bash
+pip install --upgrade clarifai
+```
+
+Login to the platform for uploading model
+
+```
+clarifai login
+```

model_upload/multimodal_embedder/README.md

@@ -0,0 +1,39 @@
+## Multimodal Embedder Model Examples
+
+These can be used on the fly with minimal or no changes to test deploy image classification models to the Clarifai platform. See the required files section for each model below.
+
+* ### [CLIP](./clip/)
+
+	Required files to run tests locally:
+
+	Download the [model checkpoint from huggingface](https://huggingface.co/openai/clip-vit-base-patch32) and store it under `clip/checkpoint/`
+
+
+	```
+	$ pip install huggingface-hub
+	$ huggingface-cli download openai/clip-vit-base-patch32 --local-dir clip/checkpoint/ --local-dir-use-symlinks False --exclude *.msgpack *.h5
+	```
+
+	Install dependecies to test locally
+
+	```bash
+	$ pip install -r clip/requirements.txt
+	```
+
+	Deploy the model to Clarifai:
+
+	>Note: set `--no-test` flag for `build` and `upload` command to disable testing
+
+	1. Build
+
+	```bash
+	$ clarifai build model ./clip
+	```
+
+	upload `*.clarifai` file to storage to obtain direct download url
+
+	2. Upload
+
+	```bash
+	$ clarifai upload model ./clip --url <your_url> 
+	```

model_upload/multimodal_embedder/clip/clarifai_config.yaml

@@ -0,0 +1,31 @@
+# Sample config of inference_parameters and labels
+# For detail, please refer to docs
+# --------------------
+# inference_parameters:
+#   - path: boolean_var
+#     default_value: true
+#     field_type: 1
+#     description: a boolean variable
+#   - path: string_var
+#     default_value: "a string"
+#     field_type: 2
+#     description: a string variable
+#   - path: number_var
+#     default_value: 1
+#     field_type: 3
+#     description: a number variable
+#   - path: secret_string_var
+#     default_value: "YOUR_SECRET"
+#     field_type: 21
+#     description: a string variable contains secret like API key
+
+clarifai_model:
+  clarifai_model_id: ''
+  clarifai_user_app_id: ''
+  description: ''
+  inference_parameters: []
+  labels: []
+  type: multimodal-embedder
+serving_backend:
+  triton:
+    max_batch_size: 4

model_upload/multimodal_embedder/clip/inference.py

@@ -0,0 +1,62 @@
+# User model inference script.
+
+import os
+from pathlib import Path
+from typing import Dict, Union
+from clarifai.models.model_serving.model_config import *  # noqa
+import torch
+from transformers import CLIPModel, CLIPProcessor
+
+class InferenceModel(MultiModalEmbedder):
+  """User model inference class."""
+
+  def __init__(self) -> None:
+    """
+    Load inference time artifacts that are called frequently .e.g. models, tokenizers, etc.
+    in this method so they are loaded only once for faster inference.
+    """
+    # current directory
+    self.base_path: Path = os.path.dirname(__file__)
+    # local checkpoint for openai/clip-vit-base-patch32
+    self.model = CLIPModel.from_pretrained(os.path.join(self.base_path, "checkpoint"))
+    self.model.eval()
+    self.processor = CLIPProcessor.from_pretrained(os.path.join(self.base_path, "checkpoint"))
+
+  def predict(self, input_data: list,
+              inference_parameters: Dict[str, Union[str, float, int, bool]] = {}) -> list:
+    """ Custom prediction function for `multimodal-embedder` model.
+
+    Args:
+      input_data (List[_MultiModalInputTypeDict]): List of dict of key-value: `image`(np.ndarray) and `text` (str)
+      inference_parameters (Dict[str, Union[str, float, int, bool]]): your inference parameters
+
+    Returns:
+      list of EmbeddingOutput
+
+    """
+
+    outputs = []
+    for inp in input_data:
+      image, text = inp.get("image", None), inp.get("text", None)
+      with torch.no_grad():
+        inputs = self.processor(text=text, images=image, return_tensors="pt", padding=True)
+        if text is not None:
+          inputs = self.processor(text=text, return_tensors="pt", padding=True)
+          embeddings = self.model.get_text_features(**inputs)
+        else:
+          inputs = self.processor(images=image, return_tensors="pt", padding=True)
+          embeddings = self.model.get_image_features(**inputs)
+      embeddings = embeddings.squeeze().cpu().numpy()
+      outputs.append(EmbeddingOutput(embedding_vector=embeddings))
+
+    return outputs
+
+
+if __name__ == "__main__":
+
+  # Dummy test
+  model = InferenceModel()
+  input = dict(text="Hi")
+
+  output = model.predict([input])
+  print(output[0])

model_upload/multimodal_embedder/clip/requirements.txt

@@ -0,0 +1,4 @@
+clarifai
+tritonclient[all]
+transformers==4.38.0
+torch==2.1.1

model_upload/multimodal_embedder/clip/test.py

@@ -0,0 +1,40 @@
+import unittest
+
+from clarifai.models.model_serving.repo_build import BaseTest
+
+
+class CustomTest(unittest.TestCase):
+  """
+  BaseTest loads the InferenceModel from the inference.py file in the current working directory.
+  To execute the predict method of the InferenceModel, use the predict method in BaseTest.
+  It takes the exact same inputs and inference parameters, returning the same outputs as InferenceModel.predict.
+  The difference is that BaseTest.predict verifies your_infer_parameters against config.clarifai_models.inference_parameters and checks the output values.
+
+  For example, test input value of visual-classifier
+
+  def test_input(self):
+    import cv2
+    path = "path/to/image"
+    img = cv2.imread(path)
+    outputs = self.model.predict([img], infer_param1=..., infer_param2=...)
+    print(outputs)
+    assert outputs
+
+  """
+
+  def setUp(self) -> None:
+    your_infer_parameter = dict(
+    )  # for example dict(float_var=0.12, string_var="test", _secret_string_var="secret")
+    self.model = BaseTest(your_infer_parameter)
+
+  def test_default_cases(self):
+    """Test your model with dummy inputs.
+    In general, you only need to run this test to check your InferneceModel implementation.
+    In case the default inputs makes your model failed for some reason (not because of assert in `test_with_default_inputs`),
+    you can comment out this test.
+    """
+    self.model.test_with_default_inputs()
+
+  def test_specific_case1(self):
+    """ Implement your test case"""
+    pass

model_upload/text_classifier/README.md

@@ -0,0 +1,39 @@
+## Text Classification Model Examples
+
+These can be used on the fly with minimal or no changes to test deploy text classification models to the Clarifai platform. See the required files section for each model below.
+
+* ### [XLM-Roberta Tweet Sentiment Classifier](./xlm-roberta/)
+
+	Required files to run tests locally:
+
+	Download the [model checkpoint](https://huggingface.co/cardiffnlp/twitter-xlm-roberta-base-sentiment/tree/main) and store it under `xlm-roberta/checkpoint/`
+
+	```
+	$ pip install huggingface-hub
+	$ huggingface-cli download cardiffnlp/twitter-xlm-roberta-base-sentiment --local-dir xlm-roberta/checkpoint/ --local-dir-use-symlinks False --exclude *.msgpack *.h5
+	```
+
+	Install dependecies to test locally
+
+	```bash
+	$ pip install -r xlm-roberta/requirements.txt
+	```
+
+	Deploy the model to Clarifai:
+
+	>Note: set `--no-test` flag for `build` and `upload` command to disable testing
+
+	1. Build
+
+	```bash
+	$ clarifai build model ./xlm-roberta
+	```
+
+	upload `*.clarifai` file to storage to obtain direct download url
+
+	2. Upload
+
+	```bash
+	$ clarifai upload model ./xlm-roberta --url <your_url> 
+	```
+

model_upload/text_classifier/xlm-roberta/clarifai_config.yaml

@@ -0,0 +1,13 @@
+clarifai_model:
+  clarifai_model_id: ''
+  clarifai_user_app_id: ''
+  description: ''
+  inference_parameters: []
+  labels:
+  - Negative
+  - Neutral
+  - Positive
+  type: text-classifier
+serving_backend:
+  triton:
+    max_batch_size: 4

model_upload/text_classifier/xlm-roberta/inference.py

@@ -0,0 +1,56 @@
+# User model inference script.
+
+import os
+from pathlib import Path
+from typing import Dict, Union
+from clarifai.models.model_serving.model_config import *  # noqa
+
+import torch
+from scipy.special import softmax
+from transformers import AutoModelForSequenceClassification, AutoTokenizer
+
+class InferenceModel(TextClassifier):
+  """User model inference class."""
+
+  def __init__(self) -> None:
+    """
+    Load inference time artifacts that are called frequently .e.g. models, tokenizers, etc.
+    in this method so they are loaded only once for faster inference.
+    """
+    # current directory
+    self.base_path: Path = os.path.dirname(__file__)
+    self.checkpoint_path: Path = os.path.join(self.base_path, "checkpoint")
+    self.model = AutoModelForSequenceClassification.from_pretrained(self.checkpoint_path)
+    self.tokenizer = AutoTokenizer.from_pretrained(self.checkpoint_path)
+
+  def predict(self, input_data: list,
+              inference_parameters: Dict[str, Union[str, float, int]]) -> list:
+    """ Custom prediction function for `text-classifier` model.
+
+    Args:
+      input_data (List[str]): List of text
+      inference_parameters (Dict[str, Union[str, float, int]]): your inference parameters
+
+    Returns:
+      list of ClassifierOutput
+
+    """
+
+    outputs = []
+    for inp in input_data:
+      encoded_input = self.tokenizer(inp, return_tensors='pt')
+      output = self.model(**encoded_input)
+      scores = output[0][0].detach().numpy()
+      scores = softmax(scores)
+      outputs.append(ClassifierOutput(predicted_scores=scores))
+
+    return outputs
+
+if __name__ == "__main__":
+
+  # Dummy test
+  model = InferenceModel()
+  input = "How are you today?"
+
+  output = model.predict([input])
+  print(output[0])

model_upload/text_classifier/xlm-roberta/requirements.txt

@@ -0,0 +1,7 @@
+clarifai
+tritonclient[all]
+torch==1.13.1
+transformers==4.38.0
+scipy==1.10.1
+sentencepiece==0.1.99
+protobuf<4.21.3

model_upload/text_classifier/xlm-roberta/test.py

@@ -0,0 +1,40 @@
+import unittest
+
+from clarifai.models.model_serving.repo_build import BaseTest
+
+
+class CustomTest(unittest.TestCase):
+  """
+  BaseTest loads the InferenceModel from the inference.py file in the current working directory.
+  To execute the predict method of the InferenceModel, use the predict method in BaseTest.
+  It takes the exact same inputs and inference parameters, returning the same outputs as InferenceModel.predict.
+  The difference is that BaseTest.predict verifies your_infer_parameters against config.clarifai_models.inference_parameters and checks the output values.
+
+  For example, test input value of visual-classifier
+
+  def test_input(self):
+    import cv2
+    path = "path/to/image"
+    img = cv2.imread(path)
+    outputs = self.model.predict([img], infer_param1=..., infer_param2=...)
+    print(outputs)
+    assert outputs
+
+  """
+
+  def setUp(self) -> None:
+    your_infer_parameter = dict(
+    )  # for example dict(float_var=0.12, string_var="test", _secret_string_var="secret")
+    self.model = BaseTest(your_infer_parameter)
+
+  def test_default_cases(self):
+    """Test your model with dummy inputs.
+    In general, you only need to run this test to check your InferneceModel implementation.
+    In case the default inputs makes your model failed for some reason (not because of assert in `test_with_default_inputs`),
+    you can comment out this test.
+    """
+    self.model.test_with_default_inputs()
+
+  def test_specific_case1(self):
+    """ Implement your test case"""
+    pass