Enable 32-bit depth image output with EXR + Drop Python 2.7 support (#91

)

* Add EXR support with 32-bit depth

* Drop Python 2.7 support

* Fix circleci config

* Remove test

* Bump version

* Test fixes

.circleci/config.yml

@@ -7,10 +7,6 @@ workflows:
           filters:
             tags:
               only: /.*/
-      - build-py27:
-          filters:
-            tags:
-              only: /.*/
 jobs:
   build-py36:
     docker:
@@ -58,33 +54,3 @@ jobs:
           command: |
             . venv/bin/activate
             make docs
-  build-py27:
-    docker:
-      - image: continuumio/miniconda:latest
-    steps:
-      - checkout
-      - run:
-          name: install python dependencies
-          command: |
-            conda create -y -n dev python=2.7 pip --force
-            source activate dev
-      - run:
-          name: install make
-          command: |
-            apt-get update
-            apt-get install -y build-essential
-      - restore_cache:
-          key: v1-dependency-cache-{{ checksum "setup.py" }}-{{ checksum "Makefile" }}-{{ checksum "requirements.txt" }}-{{ checksum "requirements-dev.txt" }}
-      - run:
-          name: install python dependencies
-          command: |
-            make dev
-      - save_cache:
-          key: v1-dependency-cache-{{ checksum "setup.py" }}-{{ checksum "Makefile" }}-{{ checksum "requirements.txt" }}-{{ checksum "requirements-dev.txt"}}
-          paths:
-            - "/opt/conda/envs/dev/"
-      - run:
-          name: run tests
-          command: |
-            source activate dev
-            make test

CHANGELOG.md

@@ -6,6 +6,12 @@ The Runway Model SDK follows [semantic versioning](https://semver.org/). Be awar
 
 Until version 1.0.0, expect that minor version changes may introduce breaking changes. We will take care not to introduce new behavior, features, or breaking changes in patch releases. If you require stability and reproducible behavior you *may* pin to a version or version range of the model SDK like `runway-python>=0.2.0` or `runway-python>=0.2,<0.3`.
 
+## v.0.6.0
+
+- Drop Python 2 support.
+- Add ability to specify output format with the `X-Runway-Output-Format` header.
+- Add 32-bit depth support for the `image` data type with the EXR format.
+
 ## v.0.5.8
 
 - Fix issue when extracting tar archives that include non-ASCII filenames.

README.md

@@ -14,7 +14,7 @@ The Runway Model SDK allows you to port new and existing machine learning models
 
 ## Installing
 
-The SDK supports both Python 2.7 and Python 3, but we recommend using Python 3. You can install the module using either pip or pip3 like so:
+The SDK supports Python 3.6+. You can install the module using either pip or pip3 like so:
 
 ```
 pip3 install runway-python
@@ -32,7 +32,7 @@ See the *Importing Models into Runway* [tutorial](https://docs.runwayapp.ai/#/im
 
 ## Developing
 
-If you'd like to contribute to the development of the Runway Python SDK, you can clone and modify this repository by following the instructions below. At the time of this writing the SDK is compatible with Python 2.7 and Python 3, however our developer dependencies `requirements-dev.txt` require Python 3.
+If you'd like to contribute to the development of the Runway Python SDK, you can clone and modify this repository by following the instructions below.
 
 ```bash
 git clone https://github.com/runwayml/model-sdk runway-model-sdk
@@ -50,7 +50,7 @@ make dev
 
 ### Testing
 
-Automated tests for the Runway Model SDK are written using `pytest` and live in the `tests/` directory. We also provide support for code coverage via `pytest-cov`. Tests can be run in both Python 2.7 or Python 3 environments.
+Automated tests for the Runway Model SDK are written using `pytest` and live in the `tests/` directory. We also provide support for code coverage via `pytest-cov`. 
 
 ```bash
 ## Create and activate a python3 virtual environment if you need to.

docs/source/index.md

@@ -6,7 +6,7 @@ These documents serve as a reference for the [Runway](https://runwayml.com) Mode
 
 ## Installing
 
-This SDK supports both Python 2.7 and Python 3, but we recommend using Python 3. You can install the module using either `pip` or `pip3` like so:
+This SDK supports both Python 3.6+. You can install the module using either `pip` or `pip3` like so:
 
 ```bash
 pip3 install runway-python

docs/source/runway_yaml_file.md

@@ -12,7 +12,7 @@ A `runway.yml` file is required to be in the root file directory of each Runway
 ```yaml
 # Specify the version of the runway.yml spec.
 version: 0.1
-# Supported python versions are 2.7 and 3.6
+# Supported python versions are: 3.6
 python: 3.6
 # The command to run your model. This value is used as the CMD value in
 # the generated Docker image.
@@ -66,7 +66,7 @@ build_steps:
 ## Schema Reference
 
 - `version` (int, optional, default = `0.1`): This version specifies the schema of the configuration file not the version of the Runway Model SDK itself.
-- `python` (float, **required**): The Python version to use when running the model installing python dependencies. Currently supported values are `2.7` and `3.6`.
+- `python` (float, **required**): The Python version to use when running the model installing python dependencies. Only `3.6` is supported at this time.
 - `entrypoint` (string, **required**): The command to run your model. This value is used as the CMD value in the generated Docker image. A standard value for this field might be `entrypoint: python runway_model.py` where `runway_model.py` implements the `@runway.setup()`, `@runway.command()`, and most importantly the `runway.run()` functions.
 - `cuda` (float, **required if building for GPU**): The NVIDIA CUDA version to use in the production GPU runtime environment. The currently supported CUDA versions are `10`, `9.2`, and `9`.
 - `framework` (string, optional, default = `None`): The machine learning framework to pre-install during the build. Currently we support `"tensorflow"` and `"pytorch"` which will install the appropriate CPU or GPU packages of Tensorflow v1.12.0 and Pytorch v1.0 respectively depending on the build environment. If you require an ML framework other than Tensorflow or Pytorch, or a version of these libraries that is different than the versions provided by the ``frameworks`` object, you can omit this object and install these dependencies manually in the build steps.

requirements.txt

@@ -10,3 +10,4 @@ scipy>=1.2.1
 urllib3[secure]>=1.25.7
 Unidecode>=1.1.1
 flask-compress>=1.3.1
+imageio>=2.5.0

runway/__version__.py

@@ -1 +1 @@
-__version__ = '0.5.8'
+__version__ = '0.6.0'

runway/data_types.py

@@ -5,14 +5,11 @@
 import json
 import os
 import tarfile
-if sys.version_info[0] < 3:
-    from cStringIO import StringIO as IO
-else:
-    from io import BytesIO as IO
+from io import BytesIO as IO
 import numpy as np
 from scipy.spatial.distance import cdist
 from PIL import Image
-from .utils import is_url, extract_tarball, try_cast_np_scalar, download_file, get_color_palette
+from .utils import is_url, extract_tarball, try_cast_np_scalar, download_file, get_color_palette, encode_image
 from .exceptions import MissingArgumentError, InvalidArgumentError
 
 class BaseType(object):
@@ -35,7 +32,7 @@ def __init__(self, data_type, description=None):
         # functions to assign names to runway.data_types based on the dictionary keys.
         self.name = self.type
 
-    def serialize(self, value):
+    def serialize(self, value, output_format=None):
         raise NotImplementedError()
 
     def deserialize(self, value):
@@ -74,7 +71,7 @@ def setup(opts):
     def __init__(self, description=None):
         super(any, self).__init__('any', description=description)
 
-    def serialize(self, v):
+    def serialize(self, v, output_format=None):
         return v
 
     def deserialize(self, v):
@@ -121,7 +118,7 @@ def __init__(self, item_type=None, description=None, min_length=0, max_length=No
     def deserialize(self, items):
         return [self.item_type.deserialize(item) for item in items]
 
-    def serialize(self, items):
+    def serialize(self, items, output_format=None):
         return [self.item_type.serialize(item) for item in items]
 
     def to_dict(self):
@@ -193,11 +190,11 @@ def __init__(
         self.channels = channels
         if channels not in [1, 3, 4]:
             raise InvalidArgumentError(self.name or self.type, 'channels value needs to be 1, 3, or 4')
-        if default_output_format and default_output_format not in ['JPEG', 'PNG']:
-            msg = 'default_output_format needs to be JPEG or PNG'
+        if default_output_format and default_output_format.upper() not in ['JPEG', 'PNG']:
+            msg = 'default_output_format needs to be "JPEG" or "PNG'
             raise InvalidArgumentError(self.name, msg)
         if default_output_format:
-            self.default_output_format = default_output_format
+            self.default_output_format = default_output_format.upper()
         elif self.channels == 3:
             self.default_output_format = 'JPEG'
         else:
@@ -223,19 +220,23 @@ def deserialize(self, value):
             deserialized_image = deserialized_image.convert(self.get_pil_mode())
         return deserialized_image
 
-    def serialize(self, value):
+    def serialize(self, value, output_format=None):
+        if output_format is None:
+            output_format = self.default_output_format
+        should_output_32bit = output_format.upper() == 'EXR'
         if type(value) is np.ndarray:
-            im_pil = Image.fromarray(value.astype(np.uint8))
+            if not should_output_32bit:
+                value = value.astype(np.uint8)
+            im_pil = Image.fromarray(value)
         elif issubclass(type(value), Image.Image):
             im_pil = value
         else:
             raise InvalidArgumentError(self.name, 'value is not a PIL or numpy image')
-        buffer = IO()
-        if im_pil.mode != self.get_pil_mode():
+        if not should_output_32bit and im_pil.mode != self.get_pil_mode():
             im_pil = im_pil.convert(self.get_pil_mode())
-        im_pil.save(buffer, format=self.default_output_format)
-        body = base64.b64encode(buffer.getvalue()).decode('utf8')
-        return 'data:image/{format};base64,{body}'.format(format=self.default_output_format.lower(), body=body)
+        encoded = encode_image(im_pil, output_format)
+        body = base64.b64encode(encoded).decode('utf8')
+        return 'data:image/{format};base64,{body}'.format(format=output_format.lower(), body=body)
 
     def to_dict(self):
         ret = super(image, self).to_dict()
@@ -296,7 +297,7 @@ def __init__(self, length=None, description=None, default=None, sampling_mean=0,
     def deserialize(self, value):
         return np.array(value)
 
-    def serialize(self, value):
+    def serialize(self, value, output_format=None):
         return value.tolist()
 
     def to_dict(self):
@@ -353,7 +354,7 @@ def deserialize(self, value):
             raise InvalidArgumentError(self.name, msg)
         return value
 
-    def serialize(self, value):
+    def serialize(self, value, output_format=None):
         return value
 
     def to_dict(self):
@@ -400,7 +401,7 @@ def __init__(self, description=None, default=None, step=None, min=None, max=None
     def deserialize(self, value):
         return value
 
-    def serialize(self, value):
+    def serialize(self, value, output_format=None):
         return try_cast_np_scalar(value)
 
     def to_dict(self):
@@ -448,7 +449,7 @@ def __init__(self, description=None, default='', min_length=0, max_length=None):
     def deserialize(self, value):
         return value
 
-    def serialize(self, value):
+    def serialize(self, value, output_format=None):
         return str(value)
 
     def to_dict(self):
@@ -505,7 +506,7 @@ def deserialize(self, path_or_url):
                 raise InvalidArgumentError(self.name, 'file path does not have expected extension')
             return path_or_url
 
-    def serialize(self, value):
+    def serialize(self, value, output_format=None):
         return value
 
     def to_dict(self):
@@ -647,7 +648,7 @@ def deserialize(self, value):
             msg = 'unable to parse expected base64-encoded image'
             raise InvalidArgumentError(self.name, msg)
 
-    def serialize(self, value):
+    def serialize(self, value, output_format=None):
         if type(value) is np.ndarray:
             im_pil = Image.fromarray(value)
         elif issubclass(type(value), Image.Image):
@@ -710,7 +711,7 @@ def deserialize(self, value):
         self.validate(value)
         return value
 
-    def serialize(self, value):
+    def serialize(self, value, output_format=None):
         self.validate(value)
         return value
 
@@ -750,7 +751,7 @@ def deserialize(self, value):
         self.validate(value)
         return value
 
-    def serialize(self, value):
+    def serialize(self, value, output_format=None):
         value = [try_cast_np_scalar(item) for item in value]
         self.validate(value)
         return value
@@ -794,7 +795,7 @@ def deserialize(self, value):
         self.validate(value)
         return value
 
-    def serialize(self, value):
+    def serialize(self, value, output_format=None):
         value = [try_cast_np_scalar(item) for item in value]
         self.validate(value)
         return value
@@ -904,7 +905,7 @@ def deserialize(self, value):
         self.validate(value)
         return value
 
-    def serialize(self, value):
+    def serialize(self, value, output_format=None):
         self.validate(value)
         value = [[try_cast_np_scalar(pt[0]), try_cast_np_scalar(pt[1])] for pt in value]
         return value

runway/model.py

@@ -18,7 +18,7 @@
 from .exceptions import RunwayError, MissingInputError, MissingOptionError, \
     InferenceError, UnknownCommandError, SetupError
 from .data_types import *
-from .utils import gzipped, serialize_command, cast_to_obj, timestamp_millis, \
+from .utils import gzipped, parse_output_formats_from_header, serialize_command, cast_to_obj, timestamp_millis, \
         validate_post_request_body_is_json, get_json_or_none_if_invalid, argspec, \
         deserialize_data, serialize_data, generate_uuid
 from .__version__ import __version__ as model_sdk_version
@@ -126,6 +126,11 @@ def command_route(command_name):
                     raise UnknownCommandError(command_name)
                 inputs = self.commands[command_name]['inputs']
                 outputs = self.commands[command_name]['outputs']
+                output_formats_header = request.headers.get('X-Runway-Output-Format')
+                if output_formats_header:
+                    output_formats = parse_output_formats_from_header(output_formats_header)
+                else:
+                    output_formats = {}
                 input_dict = get_json_or_none_if_invalid(request)
                 deserialized_inputs = deserialize_data(input_dict, inputs)
                 self.millis_last_command = timestamp_millis()
@@ -144,7 +149,7 @@ def command_route(command_name):
                     raise reraise(InferenceError, InferenceError(repr(err)), sys.exc_info()[2])
                 if type(output_data) == tuple:
                     output_data, _ = output_data
-                return jsonify(serialize_data(output_data, outputs))
+                return jsonify(serialize_data(output_data, outputs, output_formats=output_formats))
             except RunwayError as err:
                 err.print_exception()
                 return jsonify(err.to_response()), err.code