model_training.ipynb

{"metadata":{"kernelspec":{"language":"python","display_name":"Python 3","name":"python3"},"language_info":{"name":"python","version":"3.7.12","mimetype":"text/x-python","codemirror_mode":{"name":"ipython","version":3},"pygments_lexer":"ipython3","nbconvert_exporter":"python","file_extension":".py"}},"nbformat_minor":4,"nbformat":4,"cells":[{"cell_type":"markdown","source":"# Handwritten digit classification on Raspberry Pi Pico\n\n![Project in action](https://github.com/code2k13/rpipico_digit_classification/blob/main/picoml.gif?raw=true)\n\n#### 🎥 Link to youtube video: https://youtu.be/beKvz8K6b_4\n\nThis is a companion notebook to my project '[Handwritten digit classification on Pi Pico](https://ashishware.com/2022/09/03/pipico_digit_classification/)'. Visit the site for more details\n\nThis notebook contains code to:\n- Train a ML model on MINST dataset to recognize hadwritten digits\n- Export the model to pure python code so that it can be executed with CircuitPython","metadata":{}},{"cell_type":"markdown","source":"The [m2cgen](https://github.com/BayesWitnesses/m2cgen) is wonderful tool that can convert many ML models to pure python code","metadata":{}},{"cell_type":"code","source":"!pip install m2cgen","metadata":{"execution":{"iopub.status.busy":"2022-09-04T06:05:50.335361Z","iopub.execute_input":"2022-09-04T06:05:50.335873Z","iopub.status.idle":"2022-09-04T06:06:01.335105Z","shell.execute_reply.started":"2022-09-04T06:05:50.335832Z","shell.execute_reply":"2022-09-04T06:06:01.333503Z"},"trusted":true},"execution_count":15,"outputs":[]},{"cell_type":"code","source":"%matplotlib inline","metadata":{"execution":{"iopub.status.busy":"2022-09-04T06:06:01.338041Z","iopub.execute_input":"2022-09-04T06:06:01.338483Z","iopub.status.idle":"2022-09-04T06:06:01.346463Z","shell.execute_reply.started":"2022-09-04T06:06:01.338441Z","shell.execute_reply":"2022-09-04T06:06:01.345192Z"},"trusted":true},"execution_count":16,"outputs":[]},{"cell_type":"markdown","source":"Importing required libraries.","metadata":{}},{"cell_type":"code","source":"import matplotlib.pyplot as plt\nfrom sklearn import datasets, svm, metrics, ensemble,tree,linear_model\nfrom sklearn.model_selection import train_test_split\nimport m2cgen as m2c\nimport numpy as np\nimport cv2","metadata":{"execution":{"iopub.status.busy":"2022-09-04T06:06:01.348022Z","iopub.execute_input":"2022-09-04T06:06:01.348395Z","iopub.status.idle":"2022-09-04T06:06:01.361893Z","shell.execute_reply.started":"2022-09-04T06:06:01.348363Z","shell.execute_reply":"2022-09-04T06:06:01.360541Z"},"trusted":true},"execution_count":17,"outputs":[]},{"cell_type":"markdown","source":"Load the MINST hadwritten digit dataset.","metadata":{"execution":{"iopub.status.busy":"2022-09-03T04:41:55.808831Z","iopub.execute_input":"2022-09-03T04:41:55.809469Z","iopub.status.idle":"2022-09-03T04:41:55.836584Z","shell.execute_reply.started":"2022-09-03T04:41:55.809166Z","shell.execute_reply":"2022-09-03T04:41:55.835113Z"}}},{"cell_type":"code","source":"data = datasets.load_digits()","metadata":{"execution":{"iopub.status.busy":"2022-09-04T06:06:01.363892Z","iopub.execute_input":"2022-09-04T06:06:01.364256Z","iopub.status.idle":"2022-09-04T06:06:01.465165Z","shell.execute_reply.started":"2022-09-04T06:06:01.364223Z","shell.execute_reply":"2022-09-04T06:06:01.463661Z"},"trusted":true},"execution_count":18,"outputs":[]},{"cell_type":"markdown","source":"The images in dataset are 8x8 pixel. We need to train our model on larger images to get better results, so we upscale images to 12x12 pixel","metadata":{}},{"cell_type":"code","source":"image_ds  = []\nfor img in data.images:\n    res = cv2.resize(img, dsize=(12, 12))\n    res = res/16\n    image_ds.append(res)","metadata":{"execution":{"iopub.status.busy":"2022-09-04T06:06:01.469157Z","iopub.execute_input":"2022-09-04T06:06:01.469654Z","iopub.status.idle":"2022-09-04T06:06:01.490063Z","shell.execute_reply.started":"2022-09-04T06:06:01.469616Z","shell.execute_reply":"2022-09-04T06:06:01.488947Z"},"trusted":true},"execution_count":19,"outputs":[]},{"cell_type":"markdown","source":"Let's see a sample image","metadata":{}},{"cell_type":"code","source":"plt.imshow((image_ds[78]),cmap='gray')","metadata":{"execution":{"iopub.status.busy":"2022-09-04T06:06:01.491443Z","iopub.execute_input":"2022-09-04T06:06:01.491824Z","iopub.status.idle":"2022-09-04T06:06:01.689395Z","shell.execute_reply.started":"2022-09-04T06:06:01.491792Z","shell.execute_reply":"2022-09-04T06:06:01.688422Z"},"trusted":true},"execution_count":20,"outputs":[]},{"cell_type":"markdown","source":"Training a LinearSVC classifier on upscaled images","metadata":{}},{"cell_type":"code","source":"img_data = np.asarray(image_ds)\nflattened_data = img_data.reshape((len(image_ds), -1))\nclf = svm.LinearSVC()\nX_train, X_test, y_train, y_test = train_test_split(\n    flattened_data, data.target, test_size=0.4, shuffle=False\n)\nclf.fit(X_train, y_train)\npredicted = clf.predict(X_test)","metadata":{"_uuid":"8f2839f25d086af736a60e9eeb907d3b93b6e0e5","_cell_guid":"b1076dfc-b9ad-4769-8c92-a6c4dae69d19","execution":{"iopub.status.busy":"2022-09-04T06:06:01.691054Z","iopub.execute_input":"2022-09-04T06:06:01.692350Z","iopub.status.idle":"2022-09-04T06:06:01.950879Z","shell.execute_reply.started":"2022-09-04T06:06:01.692299Z","shell.execute_reply":"2022-09-04T06:06:01.949488Z"},"trusted":true},"execution_count":21,"outputs":[]},{"cell_type":"markdown","source":"Let's visualize a image from test set and output it's prediction","metadata":{}},{"cell_type":"code","source":"plt.imshow(X_test[12].reshape((12,12)))\nplt.show()\nprint(\"Predicted : \" + str(predicted[12]))","metadata":{"execution":{"iopub.status.busy":"2022-09-04T06:06:01.952616Z","iopub.execute_input":"2022-09-04T06:06:01.953538Z","iopub.status.idle":"2022-09-04T06:06:02.127097Z","shell.execute_reply.started":"2022-09-04T06:06:01.953490Z","shell.execute_reply":"2022-09-04T06:06:02.126058Z"},"trusted":true},"execution_count":22,"outputs":[]},{"cell_type":"code","source":" print(\n    f\"Classification report for classifier {clf}:\\n\"\n    f\"{metrics.classification_report(y_test, predicted)}\\n\"\n)","metadata":{"execution":{"iopub.status.busy":"2022-09-04T06:06:02.130459Z","iopub.execute_input":"2022-09-04T06:06:02.131055Z","iopub.status.idle":"2022-09-04T06:06:02.143116Z","shell.execute_reply.started":"2022-09-04T06:06:02.131021Z","shell.execute_reply":"2022-09-04T06:06:02.141781Z"},"trusted":true},"execution_count":23,"outputs":[]},{"cell_type":"markdown","source":"### Testing our model\n💡Tip :Use this block to debug. Paste image array dumped by the board to visualize it and see what's going on","metadata":{}},{"cell_type":"code","source":"arr = [0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0.648438, 0.820313, 0.828125, 0.75, 0.75, 0.75, 0.742188, 0, 0, 0, 0, 0, 0.65625, 0.757813, 0.648438, 0.5625, 0, 0, 0, 0, 0, 0, 0, 0, 0.65625, 0.75, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0.734375, 0.742188, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0.75, 0.546875, 0.640625, 0.640625, 0, 0, 0, 0, 0, 0, 0, 0, 0.734375, 0.75, 0.742188, 0.742188, 0.75, 0.75, 0.546875, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0.734375, 0.648438, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0.75, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0.742188, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0.75, 0, 0, 0, 0, 0, 0, 0.539063, 0, 0, 0, 0.65625, 0.75, 0, 0, 0]\nb = np.asarray(arr).reshape((12,12))\nplt.imshow(b,cmap='gray')\nplt.show()\nprint('Prediction : ' + str(clf.predict([arr])[0]))","metadata":{"execution":{"iopub.status.busy":"2022-09-04T06:06:02.144874Z","iopub.execute_input":"2022-09-04T06:06:02.145551Z","iopub.status.idle":"2022-09-04T06:06:02.316407Z","shell.execute_reply.started":"2022-09-04T06:06:02.145500Z","shell.execute_reply":"2022-09-04T06:06:02.315162Z"},"trusted":true},"execution_count":24,"outputs":[]},{"cell_type":"markdown","source":"### Converting model to pure python code\n\nFirst convert code to a python file.","metadata":{}},{"cell_type":"code","source":"import numpy as np\nimport sys\nsys.setrecursionlimit(2147483647)\ncode = m2c.export_to_python(clf)\nwith open('svm.py', 'w') as f:\n    f.write(code)","metadata":{"execution":{"iopub.status.busy":"2022-09-04T06:06:02.317765Z","iopub.execute_input":"2022-09-04T06:06:02.318143Z","iopub.status.idle":"2022-09-04T06:06:02.585434Z","shell.execute_reply.started":"2022-09-04T06:06:02.318110Z","shell.execute_reply":"2022-09-04T06:06:02.584146Z"},"trusted":true},"execution_count":25,"outputs":[]},{"cell_type":"markdown","source":"Then, minimize the file so that it occupies less space on Pi Pico's flash memory","metadata":{}},{"cell_type":"code","source":"!pip install python-minimizer ","metadata":{"execution":{"iopub.status.busy":"2022-09-04T06:06:02.587700Z","iopub.execute_input":"2022-09-04T06:06:02.588181Z","iopub.status.idle":"2022-09-04T06:06:13.697456Z","shell.execute_reply.started":"2022-09-04T06:06:02.588136Z","shell.execute_reply":"2022-09-04T06:06:13.695908Z"},"trusted":true},"execution_count":26,"outputs":[]},{"cell_type":"code","source":"!python-minimizer svm.py -o svm_min.py","metadata":{"execution":{"iopub.status.busy":"2022-09-04T06:06:13.700317Z","iopub.execute_input":"2022-09-04T06:06:13.700891Z","iopub.status.idle":"2022-09-04T06:06:15.088046Z","shell.execute_reply.started":"2022-09-04T06:06:13.700834Z","shell.execute_reply":"2022-09-04T06:06:15.086512Z"},"trusted":true},"execution_count":27,"outputs":[]},{"cell_type":"code","source":"!ls -alh svm_min.py","metadata":{"execution":{"iopub.status.busy":"2022-09-04T06:06:15.091887Z","iopub.execute_input":"2022-09-04T06:06:15.092273Z","iopub.status.idle":"2022-09-04T06:06:16.176747Z","shell.execute_reply.started":"2022-09-04T06:06:15.092236Z","shell.execute_reply":"2022-09-04T06:06:16.175475Z"},"trusted":true},"execution_count":28,"outputs":[]},{"cell_type":"markdown","source":"We have a 43kb model that can classify hand written digits without any external dependency !\nFor next steps visit : https://ashishware.com2022/09/03/pipico_digit_classification/","metadata":{}},{"cell_type":"code","source":"","metadata":{},"execution_count":null,"outputs":[]}]}