6.Autoregression_Functions.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"},"colab":{"name":"6.Autoregression_Functions.ipynb","provenance":[],"collapsed_sections":[]}},"nbformat_minor":0,"nbformat":4,"cells":[{"cell_type":"code","source":["def inspect_prediction(X, pred):\n","    pred_len = len(pred)\n","    n_signals = X.shape[1]\n","    \n","    figs, axs = plt.subplots(n_signals, 1, sharex=True, figsize=(30,30))\n","    for i, col in enumerate(TARGET_LABELS):\n","        axs[i].plot(np.arange(pred_len), X[-pred_len*2:-pred_len,i])\n","        axs[i].plot(np.arange(pred_len, 2*pred_len), X[-pred_len:,i], color='green')\n","        axs[i].plot(np.arange(pred_len, 2*pred_len), pred[:,i], color='orange')\n","        axs[i].set_title(col)\n","    plt.show()"],"metadata":{"execution":{"iopub.status.busy":"2022-01-14T14:12:20.130465Z","iopub.execute_input":"2022-01-14T14:12:20.130745Z","iopub.status.idle":"2022-01-14T14:12:20.140240Z","shell.execute_reply.started":"2022-01-14T14:12:20.130709Z","shell.execute_reply":"2022-01-14T14:12:20.139515Z"},"trusted":true,"id":"SK0mD4vudPzX"},"execution_count":null,"outputs":[]},{"cell_type":"code","source":["def get_batch_ending_at(X, end_index, in_len, batch_size, DEBUG = False):\n","    start_index = end_index - in_len * batch_size\n","    batch = []\n","    for i in range(start_index, end_index, in_len):\n","        batch.append(X[i : i + in_len])\n","        \n","        if DEBUG:\n","            print('[@get_batch_ending_at] window indexes:', i, i+in_len-1)\n","        \n","    return np.array(batch)\n","    \n","\n","def autoreg_prediction(X,\n","                       model,\n","                       pred_len=864,\n","                       add_sin=True,\n","                       add_cos=True,\n","                       frequency = FREQUENCY,\n","                       batch_size = BATCH_SIZE,\n","                       DEBUG=False):\n","    \n","    X = np.array(X, dtype='float32')\n","    if DEBUG:\n","        f_name = 'autoreg_prediction'\n","        print(f'\\n[@{f_name}] Converting tensor to np array')\n","        print('\\tX starting shape:', X.shape)\n","        \n","    # Add sin and cosine columns\n","    if add_sin:\n","        sin_col = [np.sin(index * (2 * np.pi / frequency)) for index in range(len(X))]\n","        sin_col = np.expand_dims(sin_col, axis=-1)\n","        X = np.append(X, sin_col, axis=1)\n","        if DEBUG:\n","            print(f'\\n[@{f_name}] Adding sin column to X')\n","            print('\\tX shape after sin column addition:', X.shape)\n","    if add_cos:\n","        cos_col = [np.cos(index * (2 * np.pi / frequency)) for index in range(len(X))]\n","        cos_col = np.expand_dims(cos_col, axis=-1)\n","        X = np.append(X, cos_col, axis=1)\n","        if DEBUG:\n","            print(f'\\n[@{f_name}] Adding cos column to X')\n","            print('\\tX shape after cos column addition:', X.shape)\n","        \n","    \n","    model_in = model.layers[0].input_shape[-2]\n","    model_out = model.layers[-1].output_shape[-2]\n","    if DEBUG:\n","        print(f'\\n[@{f_name}] Reading model input and output shapes')\n","        print('\\tmodel_in:', model_in)\n","        print('\\tmodel_out:', model_out)\n","    \n","    i = 0\n","    prediction = []\n","    while len(prediction) == 0 or len(prediction) < pred_len:\n","        i += 1\n","        \n","        batch = get_batch_ending_at(X, len(X), model_in, batch_size, DEBUG=False)\n","        if DEBUG:\n","            print(f'\\n[@{f_name}] Predicting batch {str(i)}')\n","            print(f'\\tshape batch {str(i)}:', batch.shape)\n","\n","        temp_pred = model.predict(batch)[-1]\n","        if DEBUG:\n","            print(f'\\tshape prediction {str(i)}:', temp_pred.shape)\n","        \n","        if add_sin:\n","            sin_col = [np.sin(index * (2 * np.pi / frequency)) for index in range(len(X), len(X) + len(temp_pred))]\n","            sin_col = np.expand_dims(sin_col, axis=-1)\n","            temp_pred = np.append(temp_pred, sin_col, axis=1)\n","            if DEBUG:\n","                print(f'\\n[@{f_name}] Adding sin column to prediction {str(i)}')\n","                print('\\tPrediction shape after sin column addition:', temp_pred.shape)\n","            \n","        if add_cos:\n","            cos_col = [np.cos(index * (2 * np.pi / frequency)) for index in range(len(X), len(X) + len(temp_pred))]\n","            cos_col = np.expand_dims(cos_col, axis=-1)\n","            temp_pred = np.append(temp_pred, cos_col, axis=1)\n","            if DEBUG:\n","                print(f'\\n[@{f_name}] Adding cos column to prediction {str(i)}')\n","                print('\\tPrediction shape after sin column addition:', temp_pred.shape)\n","        \n","        prediction = np.append(prediction, temp_pred, axis=0) if len(prediction) != 0 else temp_pred\n","        if DEBUG:\n","            print(f'\\n[@{f_name}] Adding prediction {str(i)} to total prediction')\n","            print(f'\\tPrediction shape after concatenating prediction {str(i)}:', prediction.shape)\n","        \n","        X = np.append(X, temp_pred, axis=0)\n","        if DEBUG:\n","            print(f'\\n[@{f_name}] Adding prediction {str(i)} to dataframe')\n","            print('\\tX shape after concatenating prediction:', X.shape)\n","            print('##################################################################')\n","            \n","    return prediction[0 : pred_len]"],"metadata":{"execution":{"iopub.status.busy":"2022-01-14T14:12:20.104815Z","iopub.execute_input":"2022-01-14T14:12:20.105217Z","iopub.status.idle":"2022-01-14T14:12:20.127591Z","shell.execute_reply.started":"2022-01-14T14:12:20.105182Z","shell.execute_reply":"2022-01-14T14:12:20.126868Z"},"trusted":true,"id":"5zZeLITbdPzW"},"execution_count":null,"outputs":[]},{"cell_type":"code","source":["model = tfk.models.load_model('***')\n","pred = autoreg_prediction(X, model, add_cos=False, add_sin=False)\n","tempX = X*std + mean\n","tempPred = pred*std + mean\n","inspect_prediction(tempX, tempPred)\n","error = np.sqrt((tempX[-864:] - tempPred)**2)\n","print('mse:',error.mean())"],"metadata":{"execution":{"iopub.status.busy":"2022-01-14T14:12:31.041533Z","iopub.execute_input":"2022-01-14T14:12:31.041798Z","iopub.status.idle":"2022-01-14T14:12:43.796257Z","shell.execute_reply.started":"2022-01-14T14:12:31.041760Z","shell.execute_reply":"2022-01-14T14:12:43.795538Z"},"trusted":true,"id":"-Xgk6bF8dPzZ"},"execution_count":null,"outputs":[]},{"cell_type":"code","source":["# NB different from previous one\n","def get_batch_ending_at(X, end_index, in_len, batch_size, DEBUG=False):\n","    start_index = end_index - in_len * batch_size\n","    batch = []\n","    for i in range(start_index-1, end_index-1, in_len):\n","        batch.append(X[i : i + in_len])\n","        \n","        if DEBUG:\n","            print('[@get_batch_ending_at] window indexes:', i, i+in_len-1)\n","        \n","    return np.array(batch)\n","\n","def get_median_value(pred_vals, index, model_out, DEBUG=False):\n","    if DEBUG:\n","        print('[@get_median_value] index:', index)\n","        print(f'\\tPredictions {str(index)} shape: {np.array(pred_vals[index]).shape}')\n","    \n","    if not index in pred_vals.keys():\n","        print(f'index {str(index)} not valid')\n","        assert index in pred_vals.keys()\n","    \n","    if not len(pred_vals[index]) == model_out:\n","        print(f'index {str(index)}: wrong number of samples: {str(len(pred_vals))}, should be: {str(model_out)}')\n","        len(pred_vals[index]) == model_out\n","    pred = np.array(pred_vals[index])\n","    pred = np.median(pred, axis=0)\n","    return pred\n","\n","def overlapping_autoreg_prediction(X,\n","                                   model,\n","                                   pred_len=1152,\n","                                   add_sin=True,\n","                                   add_cos=True,\n","                                   frequency = FREQUENCY,\n","                                   batch_size = BATCH_SIZE,\n","                                   DEBUG=False):\n","    \n","    f_name = 'overlapping_autoreg_prediction'\n","    \n","    X = np.array(X, dtype='float32')\n","    if DEBUG:\n","        print(f'\\n[@{f_name}] Converting tensor to np array')\n","        print('\\tX starting shape:', X.shape)\n","        \n","    # Add sin and cosine columns\n","    if add_sin:\n","        sin_col = [np.sin(index * (2 * np.pi / frequency)) for index in range(len(X))]\n","        sin_col = np.expand_dims(sin_col, axis=-1)\n","        X = np.append(X, sin_col, axis=1)\n","        if DEBUG:\n","            print(f'\\n[@{f_name}] Adding sin column to X')\n","            print('\\tX shape after sin column addition:', X.shape)\n","    if add_cos:\n","        cos_col = [np.cos(index * (2 * np.pi / frequency)) for index in range(len(X))]\n","        cos_col = np.expand_dims(cos_col, axis=-1)\n","        X = np.append(X, cos_col, axis=1)\n","        if DEBUG:\n","            print(f'\\n[@{f_name}] Adding cos column to X')\n","            print('\\tX shape after cos column addition:', X.shape)\n","        \n","    \n","    model_in = model.layers[0].input_shape[-2]\n","    model_out = model.layers[-1].output_shape[-2]\n","    if DEBUG:\n","        print(f'\\n[@{f_name}] Reading model input and output shapes')\n","        print('\\tmodel_in:', model_in)\n","        print('\\tmodel_out:', model_out)\n","        \n","    pred_start_index = len(X)\n","    pred_end_index = pred_start_index + pred_len\n","    pred_vals = {i:[] for i in range(pred_start_index, pred_end_index)}\n","    if DEBUG:\n","        print(f'\\n[@{f_name}] Calculating prediction indexes')\n","        print('\\tpred_start_index:', pred_start_index)\n","        print('\\tpred_end_index:', pred_end_index)\n","        \n","    # First part (No Dataset Modifications)\n","    for i, batch_end in enumerate(range(len(X) - (model_out-1), len(X)+1)):\n","        \n","        batch = get_batch_ending_at(X, batch_end, model_in, batch_size, DEBUG=False)\n","        if DEBUG:\n","            print(f'\\n[@{f_name}] Predicting batch {str(i)}')\n","            print(f'\\tshape batch {str(i)}:', batch.shape)\n","            \n","        temp_pred = model.predict(batch)[-1]\n","        if DEBUG:\n","            print(f'\\tshape prediction {str(i)}:', temp_pred.shape)\n","            print('---------------------------------------------------------------')\n","            \n","        for j in range(model_out):\n","            if (batch_end + j) in pred_vals.keys():\n","                pred_vals[batch_end + j].append(temp_pred[j])\n","    \n","    # Second part (lenghten dataset with predictions)\n","    new_val_index = pred_start_index\n","    for batch_end in range(pred_start_index, pred_end_index+1):\n","        if DEBUG:\n","            print(f'\\n[@{f_name}] Concatenating item {str(new_val_index)} to dataset')\n","            \n","        if new_val_index != pred_end_index:\n","            new_val = get_median_value(pred_vals, new_val_index, model_out, DEBUG=False)\n","        \n","            if add_sin:\n","                new_val = np.append(new_val, np.sin(new_val_index * (2 * np.pi / frequency)))\n","            if add_cos:\n","                new_val = np.append(new_val, np.cos(new_val_index * (2 * np.pi / frequency)))\n","\n","            new_val = np.expand_dims(new_val, axis=0)\n","            if DEBUG:\n","                print('\\tItem shape:', new_val.shape)\n","\n","            X = np.append(X, new_val, axis=0)\n","            if DEBUG:\n","                print('\\tX shape:', X.shape)\n","        \n","        batch = get_batch_ending_at(X, batch_end, model_in, batch_size, DEBUG=False)\n","        if DEBUG:\n","            print(f'\\n[@{f_name}] Predicting batch ending at {str(batch_end)}')\n","            \n","        temp_pred = model.predict(batch)[-1]\n","        if DEBUG:\n","            print('\\tPrediction shape:', temp_pred.shape)\n","            \n","        for j in range(model_out):\n","            if (batch_end + 1 + j) in pred_vals.keys():\n","                pred_vals[batch_end + 1 + j].append(temp_pred[j])\n","        \n","        new_val_index += 1\n","        \n","    \n","    \n","    # Compute prediction as median\n","    final_pred = [[] for i in range(pred_len)]\n","    for i in pred_vals.keys():\n","        pred_vals[i] = np.array(pred_vals[i])\n","        final_pred[i - pred_start_index] =  get_median_value(pred_vals, i, model_out, DEBUG=False)\n","    \n","    final_pred = np.array(final_pred)\n","    if DEBUG:\n","        print(f'\\n[@{f_name}] Final Prediction shape: {final_pred.shape}')\n","    return final_pred"],"metadata":{"execution":{"iopub.status.busy":"2022-01-14T14:12:43.797599Z","iopub.execute_input":"2022-01-14T14:12:43.799166Z","iopub.status.idle":"2022-01-14T14:12:43.835190Z","shell.execute_reply.started":"2022-01-14T14:12:43.799124Z","shell.execute_reply":"2022-01-14T14:12:43.834438Z"},"trusted":true,"id":"kMp6JylYdPzZ"},"execution_count":null,"outputs":[]},{"cell_type":"code","source":["model = tfk.models.load_model('***')\n","pred = overlapping_autoreg_prediction(X[:-864], model, add_cos=False, add_sin=False)\n","inspect_prediction(X*std+mean, pred*std+mean)\n","tempX = X*std + mean\n","tempPred = pred*std + mean\n","error = np.sqrt((tempX[-864:] - tempPred)**2)\n","print('mse:',error.mean())"],"metadata":{"execution":{"iopub.status.busy":"2022-01-14T14:12:43.836531Z","iopub.execute_input":"2022-01-14T14:12:43.836790Z","iopub.status.idle":"2022-01-14T14:14:49.011585Z","shell.execute_reply.started":"2022-01-14T14:12:43.836755Z","shell.execute_reply":"2022-01-14T14:14:49.010721Z"},"trusted":true,"id":"mSmYLSDhdPzd"},"execution_count":null,"outputs":[]}]}