Restart your kernel for this notebook to use update package.
\n", "\n", "\n", "" ] }, { "cell_type": "markdown", "metadata": { "button": false, "new_sheet": false, "run_control": { "read_only": false } }, "source": [ "#### Setup SageMaker environment and variables" ] }, { "cell_type": "code", "execution_count": null, "metadata": { "button": false, "new_sheet": false, "run_control": { "read_only": false }, "tags": [] }, "outputs": [], "source": [ "import os, sys, json, sagemaker, pandas as pd, boto3, numpy as np, re\n", "from sagemaker import get_execution_role,image_uris, model_uris, script_uris, hyperparameters\n", "from sagemaker.tensorflow import TensorFlow\n", "\n", "sess = sagemaker.Session()\n", "\n", "role = get_execution_role()\n", "region = sess.boto_session.region_name\n", "account = sess.boto_session.client(\"sts\").get_caller_identity()[\"Account\"]\n", "smclient = boto3.Session().client('sagemaker')\n", "bucket = sess.default_bucket()\n", "stack_name= \"active-learning-stack\"\n", "key = \"sagemaker-byoal-pretrained\"\n", "\n" ] }, { "cell_type": "markdown", "metadata": {}, "source": [ "#### Select a pre-trained model\n", "***\n", "In this lab, we are using a pretrained BERT TensorFlow uncased model . For further details and other pretrained models that are available to use for transfer learning with the Text Classification - TensorFlow algorithm - please check [TensorFlow Hub Models](https://docs.aws.amazon.com/sagemaker/latest/dg/text-classification-tensorflow-Models.html) . This model is also available through Jumpstart .\n", "\n", "A complete list of JumpStart models can also be accessed at [JumpStart Models](https://sagemaker.readthedocs.io/en/stable/doc_utils/jumpstart.html#).\n", "***" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [], "source": [ "model_id = \"tensorflow-tc-bert-en-uncased-L-12-H-768-A-12-2\"" ] }, { "cell_type": "code", "execution_count": null, "metadata": { "tags": [] }, "outputs": [], "source": [ "model_id, model_version = model_id, \"*\"\n", "training_instance_type=\"ml.p3.2xlarge\"\n", "inference_instance_typ=\"ml.m5.xlarge\"\n", "\n", "# Retrieve the docker image\n", "train_image_uri = image_uris.retrieve(\n", " region=None,\n", " framework=None,\n", " model_id=model_id,\n", " model_version=model_version,\n", " image_scope=\"training\",\n", " instance_type=training_instance_type,\n", ").replace(\"/\", \"\\/\")\n", "\n", "# Retrieve the training script\n", "train_source_uri = script_uris.retrieve(\n", " model_id=model_id, model_version=model_version, script_scope=\"training\"\n", ").replace(\"/\", \"\\/\")\n", "\n", "\n", "# Retrieve the pre-trained model tarball to further fine-tune\n", "base_model_uri = model_uris.retrieve(\n", " model_id=model_id, model_version=model_version, model_scope=\"training\"\n", ").replace(\"/\", \"\\/\")\n", "\n", "inference_instance_type = \"ml.m5.xlarge\"\n", "\n", "# Retrieve the inference docker container uri.\n", "deploy_image_uri = image_uris.retrieve(\n", " region=None,\n", " framework=None,\n", " image_scope=\"inference\",\n", " model_id=model_id,\n", " model_version=model_version,\n", " instance_type=inference_instance_type,\n", ").replace(\"/\", \"\\/\")\n", "print(deploy_image_uri)\n", "\n" ] }, { "cell_type": "markdown", "metadata": {}, "source": [ "#### Build and deploy AWS CloudFormation template using AWS SAM" ] }, { "cell_type": "markdown", "metadata": {}, "source": [ "Update template.yaml . The template.yaml is leverage by AWS SAM to build out the backend infrastructure consisting of State machine, helper utilities and AWS lambda functions\n", "\n", "Update configurations variables for Active Learning process :\n", "\n", "- `byom` : (Bring your own model) . This parameter determines whether the Lambda function returns the Amazon Blazing Text Algorithm model or your own custom model docker image. . This is set to to `false` because we are using a pretrained TensorFlow BERT model\n", "- `byomimage` : Your own custom model image . Set this parameter to NotApplicable because we are using a pretrained BERT model\n", "- `sagemaker_program` : This parameter determines the training script to be used for training your model . In this case we are using a transfer learning script to fine tune the pretrained BERT model . Hence, setting this parameter to the script name transfer_learning.py\n", "- `sagemaker_submit_directory` : The directory which contains the training source code or other dependencies aside from the entry point file . Setting this parameter to the S3 bucket name where the training script is hosted\n", "- `pretrain_model` : This parameter determines if the active learning process is using pretrained models or not . This is set to to `true` since we are using a pretrained BERT model\n", "- `pretrain_algo_train_repo` : This parameter refers to the training image URI of our pretrained BERT algorithm . Because we are using a pretrained BERT model, setting this parameter value to the training image URI \n", "- `base_model_uri` : The URI of the pretrained BERT model which will be fine tuned with custom data. Because we are using a pretrained BERT model, setting this to parameter URI to the pretrained BERT model\n", "- `pretrain_algo_inference_repo` : This parameter refers to the inference image URI of our pretrained BERT algorithm . Since we are using a pretrained BERT model, setting this parameter value to the inference image URI \n", "\n", "\n", "Update Batch Strategy to `Single` in the `CreateTransformJob` step of the ActiveLearning-active-learning-stack state machine.\n", "This is because the TensorFlow algorithm used here supports `SingleRecord` \n", "- `BatchStrategy `: `SingleRecord`" ] }, { "cell_type": "code", "execution_count": null, "metadata": { "tags": [] }, "outputs": [], "source": [ "%%bash -s \"$train_image_uri\" \"$train_source_uri\" \"$base_model_uri\" \"$deploy_image_uri\"\n", "\n", "cd src\n", "\n", "# Update byom value to False \n", "sed -i \"0,/^\\([[:space:]]*byom: *\\).*/s//\\1\"false\"/\" template.yaml\n", "\n", "#Update byomimage value to NotApplicable\n", "sed -i \"0,/^\\([[:space:]]*byomimage: *\\).*/s//\\1\"NotApplicable\"/\" template.yaml\n", "\n", "#Update Pretrained Model flag to True\n", "sed -i \"0,/^\\([[:space:]]*pretrain_model: *\\).*/s//\\1\"true\"/\" template.yaml\n", "\n", "#Update Training script for PreTrained model\n", "sed -i \"0,/^\\([[:space:]]*sagemaker_program: *\\).*/s//\\1\"transfer_learning.py\"/\" template.yaml\n", "\n", "#Update Pretrained algorithm training repository\n", "sed -i \"0,/^\\([[:space:]]*pretrain_algo_train_repo: *\\).*/s//\\1\"${1}\"/\" template.yaml\n", "\n", "#Update SageMaker submit directory \n", "sed -i \"0,/^\\([[:space:]]*sagemaker_submit_directory: *\\).*/s//\\1\"${2}\"/\" template.yaml\n", "\n", "#Update based pretrained model URI\n", "sed -i \"0,/^\\([[:space:]]*base_model_uri: *\\).*/s//\\1\"${3}\"/\" template.yaml\n", "\n", "#Updated Pretrained algorithm inference repository \n", "sed -i \"0,/^\\([[:space:]]*pretrain_algo_inference_repo: *\\).*/s//\\1\"${4}\"/\" template.yaml\n", "\n", "#Update BatchStrategy to SingleRecord\n", "sed -i 's/MultiRecord/SingleRecord/g' template.yaml\n", "\n", "cat template.yaml\n", "\n" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [], "source": [ "%%bash -s \"$region\" \"$stack_name\" \"$bucket\"\n", "\n", "# Change directory to src folder\n", "cd src\n", "\n", "#Delete SAM build folder to clean up any exisiting SAM templates\n", "rm -rf .aws-sam \n", "\n", "#Build AWS SAM artifacts to deploy AWS Cloudformation template\n", "sam build\n", "\n", "#Deploy AWS SAM template\n", "sam deploy --region ${1} --stack-name ${2} --s3-bucket ${3} --capabilities CAPABILITY_IAM CAPABILITY_AUTO_EXPAND\n", "\n" ] }, { "cell_type": "markdown", "metadata": { "button": false, "new_sheet": false, "run_control": { "read_only": false } }, "source": [ "#### Prepare labeling input manifest file" ] }, { "cell_type": "markdown", "metadata": { "button": false, "new_sheet": false, "run_control": { "read_only": false } }, "source": [ "We will create an input manifest file for our active learning workflow using the newsCorpora.csv file from the [UCI News Dataset](https://archive.ics.uci.edu/ml/datasets/News+Aggregator). This dataset contains a list of about 420,000 articles that fall into one of four categories: Business (b), Science & Technology (t), Entertainment (e) and Health & Medicine (m). We will randomly choose 10,000 articles from that file to create our dataset.\n", "\n", "For the active learning loop to start, 20% of the data must be labeled. To quickly test the active learning component, we will include 20% (`labeled_count`) of the original labels provided in the dataset in our input manifest. We use this partially-labeled dataset as the input to the active learning loop." ] }, { "cell_type": "code", "execution_count": null, "metadata": { "button": false, "new_sheet": false, "run_control": { "read_only": false }, "tags": [] }, "outputs": [], "source": [ "! wget -nc https://archive.ics.uci.edu/ml/machine-learning-databases/00359/NewsAggregatorDataset.zip --no-check-certificate && unzip -o NewsAggregatorDataset.zip" ] }, { "cell_type": "code", "execution_count": null, "metadata": { "button": false, "new_sheet": false, "run_control": { "read_only": false }, "tags": [] }, "outputs": [], "source": [ "column_names = [\"TITLE\", \"URL\", \"PUBLISHER\", \"CATEGORY\", \"STORY\", \"HOSTNAME\", \"TIMESTAMP\"]\n", "manifest_file = \"partially-labeled.manifest\"\n", "news_data_all = pd.read_csv(\"newsCorpora.csv\", names=column_names, header=None, delimiter=\"\\t\")\n", "news_data = news_data_all.sample(n=10000, random_state=42)\n", "news_data = news_data[[\"TITLE\", \"CATEGORY\"]]\n", "news_data " ] }, { "cell_type": "markdown", "metadata": { "button": false, "new_sheet": false, "run_control": { "read_only": false } }, "source": [ "We will clean our data set using *pandas*." ] }, { "cell_type": "code", "execution_count": null, "metadata": { "button": false, "new_sheet": false, "run_control": { "read_only": false }, "tags": [] }, "outputs": [], "source": [ "news_data[\"TITLE\"].replace('\"', \"\", inplace=True, regex=True)\n", "news_data[\"TITLE\"].replace(\"[^\\w\\s]\", \"\", inplace=True, regex=True)\n", "news_data[\"TITLE\"] = news_data[\"TITLE\"].str.split(\"\\n\").str[0]\n", "news_data[\"CATEGORY_LABELS\"]=news_data[\"CATEGORY\"]\n", "news_data[\"CATEGORY\"] = news_data[\"CATEGORY\"].astype(\"category\").cat.codes\n", "news_data" ] }, { "cell_type": "code", "execution_count": null, "metadata": { "button": false, "new_sheet": false, "run_control": { "read_only": false }, "tags": [] }, "outputs": [], "source": [ "fixed = news_data[\"TITLE\"].str.lower().replace('\"', \"\")" ] }, { "cell_type": "markdown", "metadata": { "button": false, "new_sheet": false, "run_control": { "read_only": false } }, "source": [ "The following cell will create our partially-labeled input manifest file, and push it to our S3 bucket. " ] }, { "cell_type": "code", "execution_count": null, "metadata": { "button": false, "new_sheet": false, "run_control": { "read_only": false }, "tags": [] }, "outputs": [], "source": [ "import json\n", "\n", "total = len(news_data)\n", "labeled_count = int(total / 5) # 20% of the dataset is labeled.\n", "label_map = {\n", " \"b\": \"Business\",\n", " \"e\": \"Entertainment\",\n", " \"m\": \"Health & Medicine\",\n", " \"t\": \"Science and Technology\",\n", "}\n", "labeled_series = pd.Series(\n", " data=news_data.iloc[:labeled_count].TITLE.values,\n", " index=news_data.iloc[:labeled_count].CATEGORY.values,\n", ")\n", "annotation_metadata = b\"\"\"{ \"category-metadata\" : { \"confidence\": 1.0, \"human-annotated\": \"yes\", \"type\": \"groundtruth/text-classification\"} }\"\"\"\n", "annotation_metadata_dict = json.loads(annotation_metadata)\n", "with open(manifest_file, \"w\") as outfile:\n", " for items in labeled_series.items():\n", " labeled_record = dict()\n", " labeled_record[\"source\"] = items[1]\n", " labeled_record[\"category\"] = int(items[0])\n", " labeled_record.update(annotation_metadata_dict)\n", " outfile.write(json.dumps(labeled_record) + \"\\n\")\n", "\n", "unlabeled_series = pd.Series(\n", " data=news_data.iloc[labeled_count:].TITLE.values,\n", " index=news_data.iloc[labeled_count:].CATEGORY.values,\n", ")\n", "with open(manifest_file, \"a\") as outfile:\n", " for items in unlabeled_series.items():\n", " outfile.write('{\"source\":\"' + items[1] + '\"}\\n')\n", "\n", "boto3.resource(\"s3\").Bucket(bucket).upload_file(manifest_file, key + \"/\" + manifest_file)\n", "manifest_file_uri = \"s3://{}/{}\".format(bucket, key + \"/\" + manifest_file)" ] }, { "cell_type": "code", "execution_count": null, "metadata": { "button": false, "new_sheet": false, "run_control": { "read_only": false }, "tags": [] }, "outputs": [], "source": [ "# Use s3 client to upload relevant json strings to s3.\n", "s3_client = boto3.client(\"s3\")" ] }, { "cell_type": "markdown", "metadata": { "button": false, "new_sheet": false, "run_control": { "read_only": false } }, "source": [ "This cell will specify the labels that workers will use to categorize the articles. To customize your labeling job, add your own labels here. To learn more, see [LabelCategoryConfigS3Uri](https://docs.aws.amazon.com/sagemaker/latest/APIReference/API_CreateLabelingJob.html#sagemaker-CreateLabelingJob-request-LabelCategoryConfigS3Uri)." ] }, { "cell_type": "code", "execution_count": null, "metadata": { "button": false, "new_sheet": false, "run_control": { "read_only": false }, "tags": [] }, "outputs": [], "source": [ "label_file_name = \"class_labels.json\"\n", "label_file = \"\"\"{\n", " \"document-version\": \"2018-11-28\",\n", " \"labels\": [\n", " {\n", " \"label\": \"Business\"\n", " },\n", " {\n", " \"label\": \"Entertainment\"\n", " },\n", " {\n", " \"label\": \"Health & Medicine\"\n", " },\n", " {\n", " \"label\": \"Science and Technology\"\n", " }\n", " ]\n", "}\"\"\"\n", "\n", "s3_client.put_object(Body=label_file, Bucket=bucket, Key=key + \"/\" + label_file_name)\n", "label_file_uri = \"s3://{}/{}\".format(bucket, key + \"/\" + label_file_name)" ] }, { "cell_type": "markdown", "metadata": {}, "source": [ "#### Prepare custom worker task template " ] }, { "cell_type": "markdown", "metadata": { "button": false, "new_sheet": false, "run_control": { "read_only": false } }, "source": [ "The following cell will specify our custom worker task template. This template will configure the UI that workers will see when they open our text classification labeling job tasks. To learn how to customize this cell, see [Creating your custom labeling task template](https://docs.aws.amazon.com/sagemaker/latest/dg/sms-custom-templates-step2.html)." ] }, { "cell_type": "code", "execution_count": null, "metadata": { "button": false, "new_sheet": false, "run_control": { "read_only": false }, "tags": [] }, "outputs": [], "source": [ "template_file_name = \"instructions.template\"\n", "template_file = r\"\"\"\n", "\n", "Example Business title:
US open: Stocks fall after Fed official hints at accelerated tapering.
\n",
"
Example Entertainment title:
CBS negotiates three more seasons for The Big Bang Theory
\n",
"
Example Health & Medicine title:
Blood Test Could Predict Alzheimer's. Good News?
\n",
"
Example Science and Technology (t) title:
Elephants tell human friend from foe by voice.
\n",
"