![\"Sample](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F6426215%2F984375a0877013033e4cbcc82405ffb3%2Foutput.png?generation=1701996123906415&alt=media\")
My one approach is to convert feature-added table data to image data and use PyTorch Image Models (Timm).
\n\nPrivate Score
\n0.816
Public Score
\n0.620
Score is bad, but it's my personal favorite.
\nI shared other codes to kaggle-OPSCP repo on Github
\nMy initial step involved enriching the dataset with calculated statistics for each gene and per 'sm_name' and 'cell_type'. This included mean, standard deviation, minimum, maximum, median, skewness, kurtosis, and mean-to-standard deviation ratios. These new features significantly increased the data dimensions.
\n ():\n \n stat_df = df.groupby(group_col)[cols].apply(stat_func)\n stat_df.columns = [ col cols]\n stat_df.reset_index().astype({group_col: })\n\n\ncell_type_mean = calculate_statistic(all_de_train, , genes, x: x.mean(), )\ncell_type_std = calculate_statistic(all_de_train, , genes, x: x.std(), )\n\nI converted the data into an image format to handle this massive feature space. This conversion was necesally for applying image-based models.
\n ():\n square_side = (math.ceil(math.sqrt(data.shape[])))\n padding_size = square_side ** - data.shape[]\n\n data_padded = np.pad(data, ((, ), (, padding_size)), , constant_values=)\n\n \n data_reshaped = data_padded.reshape(-, square_side, square_side, )\n\n \n data_rgb = np.repeat(data_reshaped, , -)\n\n data_rgb\n
I utilized the PyTorch Image Models library, which offers many pre-trained models. I fine-tuned these models, such as efficientnet_v2_l and regnet_y_800mf.
", "messages": [ { "id": 2553029, "postDate": "2023-12-08T00:59:05.760Z", "content": "My one approach is to convert feature-added table data to image data and use PyTorch Image Models (Timm).
\n\nPrivate Score
\n0.816
Public Score
\n0.620
Score is bad, but it's my personal favorite.
\nI shared other codes to kaggle-OPSCP repo on Github
\nMy initial step involved enriching the dataset with calculated statistics for each gene and per 'sm_name' and 'cell_type'. This included mean, standard deviation, minimum, maximum, median, skewness, kurtosis, and mean-to-standard deviation ratios. These new features significantly increased the data dimensions.
\n ():\n \n stat_df = df.groupby(group_col)[cols].apply(stat_func)\n stat_df.columns = [ col cols]\n stat_df.reset_index().astype({group_col: })\n\n\ncell_type_mean = calculate_statistic(all_de_train, , genes, x: x.mean(), )\ncell_type_std = calculate_statistic(all_de_train, , genes, x: x.std(), )\n\nI converted the data into an image format to handle this massive feature space. This conversion was necesally for applying image-based models.
\n ():\n square_side = (math.ceil(math.sqrt(data.shape[])))\n padding_size = square_side ** - data.shape[]\n\n data_padded = np.pad(data, ((, ), (, padding_size)), , constant_values=)\n\n \n data_reshaped = data_padded.reshape(-, square_side, square_side, )\n\n \n data_rgb = np.repeat(data_reshaped, , -)\n\n data_rgb\n
I utilized the PyTorch Image Models library, which offers many pre-trained models. I fine-tuned these models, such as efficientnet_v2_l and regnet_y_800mf.
", "rawMarkdown": "My one approach is to convert feature-added table data to image data and use PyTorch Image Models (Timm).\n\n[Code](https://www.kaggle.com/code/masaishi/opscp-m-nb019/notebook)\n\nPrivate Score\n0.816\n\nPublic Score\n0.620\n\nScore is bad, but it's my personal favorite.\n\nI shared other codes to [kaggle-OPSCP repo on Github](https://github.com/masaishi/kaggle-OPSCP)\n\n## Feature Engineering:\nMy initial step involved enriching the dataset with calculated statistics for each gene and per 'sm_name' and 'cell_type'. This included mean, standard deviation, minimum, maximum, median, skewness, kurtosis, and mean-to-standard deviation ratios. These new features significantly increased the data dimensions.\n```python\ndef calculate_statistic(df, group_col, cols, stat_func, stat_name):\n \"\"\"\n Function to calculate statistics for given columns grouped by a specific column.\n \"\"\"\n stat_df = df.groupby(group_col)[cols].apply(stat_func)\n stat_df.columns = [f'{group_col}_{stat_name}_{col}' for col in cols]\n return stat_df.reset_index().astype({group_col: str})\n\n# Example usage for calculating various statistics for 'cell_type'\ncell_type_mean = calculate_statistic(all_de_train, \"cell_type\", genes, lambda x: x.mean(), 'mean')\ncell_type_std = calculate_statistic(all_de_train, \"cell_type\", genes, lambda x: x.std(), 'std')\n# Additional statistics like min, max, median, skew, kurtosis, and ratio_mean_std are also calculated\n```\n\n## Converting Data to Image Format:\nI converted the data into an image format to handle this massive feature space. This conversion was necesally for applying image-based models.\n```python\ndef convert_to_image_format(data, output_size=(224, 224)):\n\t\tsquare_side = int(math.ceil(math.sqrt(data.shape[1])))\n\t\tpadding_size = square_side ** 2 - data.shape[1]\n\t\t\n\t\tdata_padded = np.pad(data, ((0, 0), (0, padding_size)), 'constant', constant_values=0)\n\t\t\n\t\t# We are reshaping to [N, H, W, C] because the final torch tensor needs to be [N, C, H, W]\n\t\tdata_reshaped = data_padded.reshape(-1, square_side, square_side, 1)\n\t\t\n\t\t# Expand the last dimension to three channels by repeating the data\n\t\tdata_rgb = np.repeat(data_reshaped, 3, -1)\n\t\t\n\t\treturn data_rgb\n```\n\n\n## Fine-Tuning PyTorch Image Models (timm):\nI utilized the PyTorch Image Models library, which offers many pre-trained models. I fine-tuned these models, such as efficientnet_v2_l and regnet_y_800mf.", "votes": 1 } ], "comments": [], "raw_markdown_by_id": { "2553029": "My one approach is to convert feature-added table data to image data and use PyTorch Image Models (Timm).\n\n[Code](https://www.kaggle.com/code/masaishi/opscp-m-nb019/notebook)\n\nPrivate Score\n0.816\n\nPublic Score\n0.620\n\nScore is bad, but it's my personal favorite.\n\nI shared other codes to [kaggle-OPSCP repo on Github](https://github.com/masaishi/kaggle-OPSCP)\n\n## Feature Engineering:\nMy initial step involved enriching the dataset with calculated statistics for each gene and per 'sm_name' and 'cell_type'. This included mean, standard deviation, minimum, maximum, median, skewness, kurtosis, and mean-to-standard deviation ratios. These new features significantly increased the data dimensions.\n```python\ndef calculate_statistic(df, group_col, cols, stat_func, stat_name):\n \"\"\"\n Function to calculate statistics for given columns grouped by a specific column.\n \"\"\"\n stat_df = df.groupby(group_col)[cols].apply(stat_func)\n stat_df.columns = [f'{group_col}_{stat_name}_{col}' for col in cols]\n return stat_df.reset_index().astype({group_col: str})\n\n# Example usage for calculating various statistics for 'cell_type'\ncell_type_mean = calculate_statistic(all_de_train, \"cell_type\", genes, lambda x: x.mean(), 'mean')\ncell_type_std = calculate_statistic(all_de_train, \"cell_type\", genes, lambda x: x.std(), 'std')\n# Additional statistics like min, max, median, skew, kurtosis, and ratio_mean_std are also calculated\n```\n\n## Converting Data to Image Format:\nI converted the data into an image format to handle this massive feature space. This conversion was necesally for applying image-based models.\n```python\ndef convert_to_image_format(data, output_size=(224, 224)):\n\t\tsquare_side = int(math.ceil(math.sqrt(data.shape[1])))\n\t\tpadding_size = square_side ** 2 - data.shape[1]\n\t\t\n\t\tdata_padded = np.pad(data, ((0, 0), (0, padding_size)), 'constant', constant_values=0)\n\t\t\n\t\t# We are reshaping to [N, H, W, C] because the final torch tensor needs to be [N, C, H, W]\n\t\tdata_reshaped = data_padded.reshape(-1, square_side, square_side, 1)\n\t\t\n\t\t# Expand the last dimension to three channels by repeating the data\n\t\tdata_rgb = np.repeat(data_reshaped, 3, -1)\n\t\t\n\t\treturn data_rgb\n```\n\n\n## Fine-Tuning PyTorch Image Models (timm):\nI utilized the PyTorch Image Models library, which offers many pre-trained models. I fine-tuned these models, such as efficientnet_v2_l and regnet_y_800mf." } }