{"cells":[{"metadata":{"_uuid":"8f2839f25d086af736a60e9eeb907d3b93b6e0e5","_cell_guid":"b1076dfc-b9ad-4769-8c92-a6c4dae69d19","trusted":true},"cell_type":"code","source":"package_paths = [\n    '../input/pytorch-image-models/pytorch-image-models-master', #'../input/efficientnet-pytorch-07/efficientnet_pytorch-0.7.0'\n    '../input/image-fmix/FMix-master'\n]\nimport sys; \n\nfor pth in package_paths:\n    sys.path.append(pth)\n    \nfrom fmix import sample_mask, make_low_freq_image, binarise_mask","execution_count":null,"outputs":[]},{"metadata":{"trusted":true},"cell_type":"code","source":"from glob import glob\nfrom sklearn.model_selection import GroupKFold, StratifiedKFold\nimport cv2\nfrom skimage import io\nimport torch\nfrom torch import nn\nimport os\nfrom datetime import datetime\nimport time\nimport random\nimport cv2\nimport torchvision\nfrom torchvision import transforms\nimport pandas as pd\nimport numpy as np\nfrom tqdm import tqdm\n\nimport matplotlib.pyplot as plt\nfrom torch.utils.data import Dataset,DataLoader\nfrom torch.utils.data.sampler import SequentialSampler, RandomSampler\nfrom torch.cuda.amp import autocast, GradScaler\nfrom torch.nn.modules.loss import _WeightedLoss\nimport torch.nn.functional as F\n\nimport timm\n\nimport sklearn\nimport warnings\nimport joblib\nfrom sklearn.metrics import roc_auc_score, log_loss\nfrom sklearn import metrics\nimport warnings\nimport cv2\nimport pydicom\n#from efficientnet_pytorch import EfficientNet\nfrom scipy.ndimage.interpolation import zoom","execution_count":null,"outputs":[]},{"metadata":{"trusted":true},"cell_type":"code","source":"pip install torch_optimizer","execution_count":null,"outputs":[]},{"metadata":{"trusted":true},"cell_type":"code","source":"import torch_optimizer as optim","execution_count":null,"outputs":[]},{"metadata":{"trusted":true},"cell_type":"code","source":"class LabelSmoothingLoss(nn.Module): \n    def __init__(self, classes=5, smoothing=0.0, dim=-1): \n        super(LabelSmoothingLoss, self).__init__() \n        self.confidence = 1.0 - smoothing \n        self.smoothing = smoothing \n        self.cls = classes \n        self.dim = dim \n    def forward(self, pred, target): \n        pred = pred.log_softmax(dim=self.dim) \n        with torch.no_grad():\n            true_dist = torch.zeros_like(pred) \n            true_dist.fill_(self.smoothing / (self.cls - 1)) \n            true_dist.scatter_(1, target.data.unsqueeze(1), self.confidence) \n        return torch.mean(torch.sum(-true_dist * pred, dim=self.dim))\nclass TaylorSoftmax(nn.Module):\n    '''\n    This is the autograd version\n    '''\n    def __init__(self, dim=1, n=2):\n        super(TaylorSoftmax, self).__init__()\n        assert n % 2 == 0\n        self.dim = dim\n        self.n = n\n\n    def forward(self, x):\n        '''\n        usage similar to nn.Softmax:\n            >>> mod = TaylorSoftmax(dim=1, n=4)\n            >>> inten = torch.randn(1, 32, 64, 64)\n            >>> out = mod(inten)\n        '''\n        fn = torch.ones_like(x)\n        denor = 1.\n        for i in range(1, self.n+1):\n            denor *= i\n            fn = fn + x.pow(i) / denor\n        out = fn / fn.sum(dim=self.dim, keepdims=True)\n        return out\n\n    \nclass TaylorCrossEntropyLoss(nn.Module):\n    def __init__(self, n=2, ignore_index=-1, reduction='mean', smoothing=0.05):\n        super(TaylorCrossEntropyLoss, self).__init__()\n        assert n % 2 == 0\n        self.taylor_softmax = TaylorSoftmax(dim=1, n=n)\n        self.reduction = reduction\n        self.ignore_index = ignore_index\n        self.lab_smooth = LabelSmoothingLoss(5, smoothing=smoothing)\n\n    def forward(self, logits, labels):\n        log_probs = self.taylor_softmax(logits).log()\n        #loss = F.nll_loss(log_probs, labels, reduction=self.reduction,\n        #        ignore_index=self.ignore_index)\n        loss = self.lab_smooth(log_probs, labels)\n        return loss","execution_count":null,"outputs":[]},{"metadata":{"trusted":true},"cell_type":"code","source":"#方法尝试bi_tempered_loss_pytorch\nimport torch\n\ndef log_t(u, t):\n    \"\"\"Compute log_t for `u'.\"\"\"\n    if t==1.0:\n        return u.log()\n    else:\n        return (u.pow(1.0 - t) - 1.0) / (1.0 - t)\n\ndef exp_t(u, t):\n    \"\"\"Compute exp_t for `u'.\"\"\"\n    if t==1:\n        return u.exp()\n    else:\n        return (1.0 + (1.0-t)*u).relu().pow(1.0 / (1.0 - t))\n\ndef compute_normalization_fixed_point(activations, t, num_iters):\n\n    \"\"\"Returns the normalization value for each example (t > 1.0).\n\n    Args:\n      activations: A multi-dimensional tensor with last dimension `num_classes`.\n      t: Temperature 2 (> 1.0 for tail heaviness).\n      num_iters: Number of iterations to run the method.\n    Return: A tensor of same shape as activation with the last dimension being 1.\n    \"\"\"\n    mu, _ = torch.max(activations, -1, keepdim=True)\n    normalized_activations_step_0 = activations - mu\n\n    normalized_activations = normalized_activations_step_0\n\n    for _ in range(num_iters):\n        logt_partition = torch.sum(\n                exp_t(normalized_activations, t), -1, keepdim=True)\n        normalized_activations = normalized_activations_step_0 * \\\n                logt_partition.pow(1.0-t)\n\n    logt_partition = torch.sum(\n            exp_t(normalized_activations, t), -1, keepdim=True)\n    normalization_constants = - log_t(1.0 / logt_partition, t) + mu\n\n    return normalization_constants\n\ndef compute_normalization_binary_search(activations, t, num_iters):\n\n    \"\"\"Returns the normalization value for each example (t < 1.0).\n\n    Args:\n      activations: A multi-dimensional tensor with last dimension `num_classes`.\n      t: Temperature 2 (< 1.0 for finite support).\n      num_iters: Number of iterations to run the method.\n    Return: A tensor of same rank as activation with the last dimension being 1.\n    \"\"\"\n\n    mu, _ = torch.max(activations, -1, keepdim=True)\n    normalized_activations = activations - mu\n\n    effective_dim = \\\n        torch.sum(\n                (normalized_activations > -1.0 / (1.0-t)).to(torch.int32),\n            dim=-1, keepdim=True).to(activations.dtype)\n\n    shape_partition = activations.shape[:-1] + (1,)\n    lower = torch.zeros(shape_partition, dtype=activations.dtype, device=activations.device)\n    upper = -log_t(1.0/effective_dim, t) * torch.ones_like(lower)\n\n    for _ in range(num_iters):\n        logt_partition = (upper + lower)/2.0\n        sum_probs = torch.sum(\n                exp_t(normalized_activations - logt_partition, t),\n                dim=-1, keepdim=True)\n        update = (sum_probs < 1.0).to(activations.dtype)\n        lower = torch.reshape(\n                lower * update + (1.0-update) * logt_partition,\n                shape_partition)\n        upper = torch.reshape(\n                upper * (1.0 - update) + update * logt_partition,\n                shape_partition)\n\n    logt_partition = (upper + lower)/2.0\n    return logt_partition + mu\n\nclass ComputeNormalization(torch.autograd.Function):\n    \"\"\"\n    Class implementing custom backward pass for compute_normalization. See compute_normalization.\n    \"\"\"\n    @staticmethod\n    def forward(ctx, activations, t, num_iters):\n        if t < 1.0:\n            normalization_constants = compute_normalization_binary_search(activations, t, num_iters)\n        else:\n            normalization_constants = compute_normalization_fixed_point(activations, t, num_iters)\n\n        ctx.save_for_backward(activations, normalization_constants)\n        ctx.t=t\n        return normalization_constants\n\n    @staticmethod\n    def backward(ctx, grad_output):\n        activations, normalization_constants = ctx.saved_tensors\n        t = ctx.t\n        normalized_activations = activations - normalization_constants \n        probabilities = exp_t(normalized_activations, t)\n        escorts = probabilities.pow(t)\n        escorts = escorts / escorts.sum(dim=-1, keepdim=True)\n        grad_input = escorts * grad_output\n        \n        return grad_input, None, None\n\ndef compute_normalization(activations, t, num_iters=5):\n    \"\"\"Returns the normalization value for each example. \n    Backward pass is implemented.\n    Args:\n      activations: A multi-dimensional tensor with last dimension `num_classes`.\n      t: Temperature 2 (> 1.0 for tail heaviness, < 1.0 for finite support).\n      num_iters: Number of iterations to run the method.\n    Return: A tensor of same rank as activation with the last dimension being 1.\n    \"\"\"\n    return ComputeNormalization.apply(activations, t, num_iters)\n\ndef tempered_sigmoid(activations, t, num_iters = 5):\n    \"\"\"Tempered sigmoid function.\n\n    Args:\n      activations: Activations for the positive class for binary classification.\n      t: Temperature tensor > 0.0.\n      num_iters: Number of iterations to run the method.\n\n    Returns:\n      A probabilities tensor.\n    \"\"\"\n    internal_activations = torch.stack([activations,\n        torch.zeros_like(activations)],\n        dim=-1)\n    internal_probabilities = tempered_softmax(internal_activations, t, num_iters)\n    return internal_probabilities[..., 0]\n\n\ndef tempered_softmax(activations, t, num_iters=5):\n    \"\"\"Tempered softmax function.\n    Args:\n      activations: A multi-dimensional tensor with last dimension `num_classes`.\n      t: Temperature > 1.0.\n      num_iters: Number of iterations to run the method.\n    Returns:\n      A probabilities tensor.\n    \"\"\"\n    if t == 1.0:\n        return activations.softmax(dim=-1)\n\n    normalization_constants = compute_normalization(activations, t, num_iters)\n    return exp_t(activations - normalization_constants, t)\n\ndef bi_tempered_binary_logistic_loss(activations,\n        labels,\n        t1,\n        t2,\n        label_smoothing = 0.0,\n        num_iters=5,\n        reduction='mean'):\n\n    \"\"\"Bi-Tempered binary logistic loss.\n\n    Args:\n      activations: A tensor containing activations for class 1.\n      labels: A tensor with shape as activations, containing probabilities for class 1\n      t1: Temperature 1 (< 1.0 for boundedness).\n      t2: Temperature 2 (> 1.0 for tail heaviness, < 1.0 for finite support).\n      label_smoothing: Label smoothing\n      num_iters: Number of iterations to run the method.\n\n    Returns:\n      A loss tensor.\n    \"\"\"\n    internal_activations = torch.stack([activations,\n        torch.zeros_like(activations)],\n        dim=-1)\n    internal_labels = torch.stack([labels.to(activations.dtype),\n        1.0 - labels.to(activations.dtype)],\n        dim=-1)\n    return bi_tempered_logistic_loss(internal_activations, \n            internal_labels,\n            t1,\n            t2,\n            label_smoothing = label_smoothing,\n            num_iters = num_iters,\n            reduction = reduction)\n\ndef bi_tempered_logistic_loss(activations,\n        labels,\n        t1,\n        t2,\n        label_smoothing=0.0,\n        num_iters=5,\n        reduction = 'mean'):\n\n    \"\"\"Bi-Tempered Logistic Loss.\n    Args:\n      activations: A multi-dimensional tensor with last dimension `num_classes`.\n      labels: A tensor with shape and dtype as activations (onehot), \n        or a long tensor of one dimension less than activations (pytorch standard)\n      t1: Temperature 1 (< 1.0 for boundedness).\n      t2: Temperature 2 (> 1.0 for tail heaviness, < 1.0 for finite support).\n      label_smoothing: Label smoothing parameter between [0, 1). Default 0.0.\n      num_iters: Number of iterations to run the method. Default 5.\n      reduction: ``'none'`` | ``'mean'`` | ``'sum'``. Default ``'mean'``.\n        ``'none'``: No reduction is applied, return shape is shape of\n        activations without the last dimension.\n        ``'mean'``: Loss is averaged over minibatch. Return shape (1,)\n        ``'sum'``: Loss is summed over minibatch. Return shape (1,)\n    Returns:\n      A loss tensor.\n    \"\"\"\n\n    if len(labels.shape)<len(activations.shape): #not one-hot\n        labels_onehot = torch.zeros_like(activations)\n        labels_onehot.scatter_(1, labels[..., None], 1)\n    else:\n        labels_onehot = labels\n\n    if label_smoothing > 0:\n        num_classes = labels_onehot.shape[-1]\n        labels_onehot = ( 1 - label_smoothing * num_classes / (num_classes - 1) ) \\\n                * labels_onehot + \\\n                label_smoothing / (num_classes - 1)\n\n    probabilities = tempered_softmax(activations, t2, num_iters)\n\n    loss_values = labels_onehot * log_t(labels_onehot + 1e-10, t1) \\\n            - labels_onehot * log_t(probabilities, t1) \\\n            - labels_onehot.pow(2.0 - t1) / (2.0 - t1) \\\n            + probabilities.pow(2.0 - t1) / (2.0 - t1)\n    loss_values = loss_values.sum(dim = -1) #sum over classes\n\n    if reduction == 'none':\n        return loss_values\n    if reduction == 'sum':\n        return loss_values.sum()\n    if reduction == 'mean':\n        return loss_values.mean()","execution_count":null,"outputs":[]},{"metadata":{"trusted":true},"cell_type":"code","source":"CFG = {\n    'fold_num': 5,\n    'seed': 719,\n    'model_arch': 'tf_efficientnet_b4_ns',\n    'img_size': 512,\n    'epochs': 10,\n    'train_bs': 16,\n    'valid_bs': 16,\n    'T_0': 10,\n    'lr': (1e-4),\n    'min_lr': (1e-6)*2.0,\n    'weight_decay':(1e-6)/2.0,\n    'num_workers': 4,\n    'accum_iter': 2, # suppoprt to do batch accumulation for backprop with effectively larger batch size\n    'verbose_step': 1,\n    'device': 'cuda:0'\n}","execution_count":null,"outputs":[]},{"metadata":{"trusted":true},"cell_type":"code","source":"train = pd.read_csv('../input/cassava-leaf-disease-classification/train.csv')\nunusual = [\n    '1004389140.jpg',\n    '1008244905.jpg',\n    '1338159402.jpg',\n    '1339403533.jpg',\n    '159654644.jpg',\n    '1010470173.jpg',\n    '1014492188.jpg',\n    '1359893940.jpg',\n    '1366430957.jpg',\n    '1689510013.jpg',\n    '1726694302.jpg',\n    '1770746162.jpg',\n    '1773381712.jpg',\n    '1848686439.jpg',\n    '1905119159.jpg',\n    '1917903934.jpg',\n    '1960041118.jpg',\n    '199112616.jpg',\n    '2016389925.jpg',\n    '2073193450.jpg',\n    '2074713873.jpg',\n    '2084868828.jpg',\n    '2139839273.jpg',\n    '2166623214.jpg',\n    '2262263316.jpg',\n    '2276509518.jpg',\n    '2278166989.jpg',\n    '2321669192.jpg',\n    '2320471703.jpg',\n    '2382642453.jpg',\n    '2415837573.jpg',\n    '2482667092.jpg',\n    '2604713994.jpg',\n    '262902341.jpg',\n    '2642216511.jpg',\n    '2698282165.jpg',\n    '2719114674.jpg',\n    '274726002.jpg',\n    '2925605732.jpg',\n    '2981404650.jpg',\n    '3040241097.jpg',\n    '3043097813.jpg',\n    '3123906243.jpg',\n    '3126296051.jpg',\n    '3199643560.jpg',\n    '3251960666.jpg',\n    '3252232501.jpg',\n    '3425850136.jpg',\n    '3435954655.jpg',\n    '3477169212.jpg',\n    '3609350672.jpg',\n    '3652033201.jpg',\n    '3810809174.jpg',\n    '3838556102.jpg',\n    '3881028757.jpg',\n    '3892366593.jpg',\n    '4060987360.jpg',\n    '4089218356.jpg',\n    '4134583704.jpg',\n    '4203623611.jpg',\n    '421035788.jpg',\n    '4239074071.jpg',\n    '4269208386.jpg',\n    '457405364.jpg',\n    '549854027.jpg',\n    '554488826.jpg',\n    '580111608.jpg',\n    '600736721.jpg',\n    '616718743.jpg',\n    '695438825.jpg',\n    '723564013.jpg',\n    '746746526.jpg',\n    '826231979.jpg',\n    '847847826.jpg',\n    '9224019.jpg',\n    '992748624.jpg'\n]\noutliers = [\n    '156080014.jpg',\n    '2182500020.jpg',\n    '2489013604.jpg',\n    '3129393327.jpg',\n    '314640668.jpg',\n    '490649765.jpg',\n    '1285436512.jpg',\n    '1403621003.jpg', # Looks like unusual sample from Cassava Brown Streak Disease but labeled like Cassava Mosaic Disease\n    '1819546557.jpg',\n    '1841279687.jpg',\n    '2088351120.jpg',\n    '2161797110.jpg',\n    '2602649407.jpg',\n    '277532565.jpg',\n    '3184864595.jpg',\n    '3238801760.jpg',\n    '3272750945.jpg',\n    '3382391338.jpg',\n    '357924077.jpg',\n    '4044829046.jpg',\n    '4059169921.jpg',\n    '4280523848.jpg',\n    '449389274.jpg',\n    '452420525.jpg',\n    '479472063.jpg',\n    '612680278.jpg',\n    '726377415.jpg',\n    '1179237425.jpg',\n    '1663857014.jpg',\n    '2565638908.jpg',\n    '3188953817.jpg',\n    '3421208425.jpg',\n    '504689064.jpg',\n    '597389720.jpg',\n    '1119403430.jpg',\n    '1774341872.jpg',\n    '1886828385.jpg',\n    '2484530081.jpg',\n    '2632579053.jpg',\n    '2839068946.jpg',\n    '284130814.jpg',\n    '3609986814.jpg',\n    '3724956866.jpg',\n    '3746679490.jpg',\n    '3853597900.jpg',\n    '927165736.jpg'\n]\n# for i in unusual:\n#     train.loc[train['image_id']==i,'label']=5\n\n# for i in outliers:\n#     train=train.drop(train[train['image_id']==i].index)\n\n# train=train.reset_index(drop=True)\n\ntrain.head()","execution_count":null,"outputs":[]},{"metadata":{"trusted":true},"cell_type":"code","source":"train.label.value_counts()","execution_count":null,"outputs":[]},{"metadata":{"trusted":true},"cell_type":"code","source":"submission = pd.read_csv('../input/cassava-leaf-disease-classification/sample_submission.csv')\nsubmission.head()","execution_count":null,"outputs":[]},{"metadata":{"trusted":true},"cell_type":"code","source":"def seed_everything(seed):\n    random.seed(seed)\n    os.environ['PYTHONHASHSEED'] = str(seed)\n    np.random.seed(seed)\n    torch.manual_seed(seed)\n    torch.cuda.manual_seed(seed)\n    torch.backends.cudnn.deterministic = True\n    torch.backends.cudnn.benchmark = True\n    \ndef get_img(path):\n    im_bgr = cv2.imread(path)\n    im_rgb = im_bgr[:, :, ::-1]\n    #print(im_rgb)\n    return im_rgb\n\nimg = get_img('../input/cassava-leaf-disease-classification/train_images/1000015157.jpg')\nplt.imshow(img)\nplt.show()","execution_count":null,"outputs":[]},{"metadata":{"trusted":true},"cell_type":"code","source":"def rand_bbox(size, lam):\n    W = size[0]\n    H = size[1]\n    cut_rat = np.sqrt(1. - lam)\n    cut_w = np.int(W * cut_rat)\n    cut_h = np.int(H * cut_rat)\n\n    # uniform\n    cx = np.random.randint(W)\n    cy = np.random.randint(H)\n\n    bbx1 = np.clip(cx - cut_w // 2, 0, W)\n    bby1 = np.clip(cy - cut_h // 2, 0, H)\n    bbx2 = np.clip(cx + cut_w // 2, 0, W)\n    bby2 = np.clip(cy + cut_h // 2, 0, H)\n    return bbx1, bby1, bbx2, bby2\n\n\nclass CassavaDataset(Dataset):\n    def __init__(self, df, data_root, \n                 transforms=None, \n                 output_label=True, \n                 one_hot_label=False,\n                 do_fmix=False, \n                 fmix_params={\n                     'alpha': 1., \n                     'decay_power': 3., \n                     'shape': (CFG['img_size'], CFG['img_size']),\n                     'max_soft': True, \n                     'reformulate': False\n                 },\n                 do_cutmix=False,\n                 cutmix_params={\n                     'alpha': 1,\n                 }\n                ):\n        \n        super().__init__()\n        self.df = df.reset_index(drop=True).copy()\n        self.transforms = transforms\n        self.data_root = data_root\n        self.do_fmix = do_fmix\n        self.fmix_params = fmix_params\n        self.do_cutmix = do_cutmix\n        self.cutmix_params = cutmix_params\n        \n        self.output_label = output_label\n        self.one_hot_label = one_hot_label\n        \n        if output_label == True:\n            self.labels = self.df['label'].values\n            #print(self.labels)\n            \n            if one_hot_label is True:\n                self.labels = np.eye(self.df['label'].max()+1)[self.labels]\n                #print(self.labels)\n            \n    def __len__(self):\n        return self.df.shape[0]\n    \n    def __getitem__(self, index: int):\n        \n        # get labels\n        if self.output_label:\n            target = self.labels[index]\n          \n        img  = get_img(\"{}/{}\".format(self.data_root, self.df.loc[index]['image_id']))\n\n        if self.transforms:\n            img = self.transforms(image=img)['image']\n        \n        if self.do_fmix and np.random.uniform(0., 1., size=1)[0] > 0.5:\n            with torch.no_grad():\n                #lam, mask = sample_mask(**self.fmix_params)\n                \n                lam = np.clip(np.random.beta(self.fmix_params['alpha'], self.fmix_params['alpha']),0.6,0.7)\n                \n                # Make mask, get mean / std\n                mask = make_low_freq_image(self.fmix_params['decay_power'], self.fmix_params['shape'])\n                mask = binarise_mask(mask, lam, self.fmix_params['shape'], self.fmix_params['max_soft'])\n    \n                fmix_ix = np.random.choice(self.df.index, size=1)[0]\n                fmix_img  = get_img(\"{}/{}\".format(self.data_root, self.df.iloc[fmix_ix]['image_id']))\n\n                if self.transforms:\n                    fmix_img = self.transforms(image=fmix_img)['image']\n\n                mask_torch = torch.from_numpy(mask)\n                \n                # mix image\n                img = mask_torch*img+(1.-mask_torch)*fmix_img\n\n                #print(mask.shape)\n\n                #assert self.output_label==True and self.one_hot_label==True\n\n                # mix target\n                rate = mask.sum()/CFG['img_size']/CFG['img_size']\n                target = rate*target + (1.-rate)*self.labels[fmix_ix]\n                #print(target, mask, img)\n                #assert False\n        \n        if self.do_cutmix and np.random.uniform(0., 1., size=1)[0] > 0.5:\n            #print(img.sum(), img.shape)\n            with torch.no_grad():\n                cmix_ix = np.random.choice(self.df.index, size=1)[0]\n                cmix_img  = get_img(\"{}/{}\".format(self.data_root, self.df.iloc[cmix_ix]['image_id']))\n                if self.transforms:\n                    cmix_img = self.transforms(image=cmix_img)['image']\n                    \n                lam = np.clip(np.random.beta(self.cutmix_params['alpha'], self.cutmix_params['alpha']),0.3,0.4)\n                bbx1, bby1, bbx2, bby2 = rand_bbox((CFG['img_size'], CFG['img_size']), lam)\n\n                img[:, bbx1:bbx2, bby1:bby2] = cmix_img[:, bbx1:bbx2, bby1:bby2]\n\n                rate = 1 - ((bbx2 - bbx1) * (bby2 - bby1) / (CFG['img_size'] * CFG['img_size']))\n                target = rate*target + (1.-rate)*self.labels[cmix_ix]\n                \n            #print('-', img.sum())\n            #print(target)\n            #assert False\n                            \n        # do label smoothing\n        #print(type(img), type(target))\n        if self.output_label == True:\n            return img, target\n        else:\n            return img","execution_count":null,"outputs":[]},{"metadata":{"trusted":true},"cell_type":"code","source":"from albumentations import (\n    HorizontalFlip, VerticalFlip, IAAPerspective, ShiftScaleRotate, CLAHE, RandomRotate90,\n    Transpose, ShiftScaleRotate, Blur, OpticalDistortion, GridDistortion, HueSaturationValue,\n    IAAAdditiveGaussianNoise, GaussNoise, MotionBlur, MedianBlur, IAAPiecewiseAffine, RandomResizedCrop,\n    IAASharpen, IAAEmboss, RandomBrightnessContrast,GaussianBlur,RandomContrast, Flip, OneOf, Compose, Normalize, Cutout, CoarseDropout, ShiftScaleRotate, CenterCrop, Resize,ChannelShuffle,RandomBrightness\n)\n\nfrom albumentations.pytorch import ToTensorV2\n\ndef get_train_transforms():\n    return Compose([\n            RandomResizedCrop(CFG['img_size'], CFG['img_size']),\n            Transpose(p=0.5),\n            HorizontalFlip(p=0.5),\n            VerticalFlip(p=0.5),\n            ShiftScaleRotate(p=0.5),\n            HueSaturationValue(hue_shift_limit=0.2, sat_shift_limit=0.2, val_shift_limit=0.2, p=0.5),\n            RandomBrightnessContrast(brightness_limit=(-0.1,0.1), contrast_limit=(-0.1, 0.1), p=0.5),\n            Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225], max_pixel_value=255.0, p=1.0),\n            CoarseDropout(p=0.5),\n            ToTensorV2(p=1.0),\n#          RandomResizedCrop(CFG['img_size'], CFG['img_size']),\n#          OneOf([RandomBrightness(limit=0.1, p=0.5), RandomContrast(limit=0.1, p=0.5)]),\n#          OneOf([MotionBlur(blur_limit=3), MedianBlur(blur_limit=3), GaussianBlur(blur_limit=3),], p=0.5,),\n#          VerticalFlip(p=0.5),\n#          HueSaturationValue(hue_shift_limit=0.2, sat_shift_limit=0.2, val_shift_limit=0.2, p=0.5),\n#          HorizontalFlip(p=0.5),\n#          ShiftScaleRotate(\n#             shift_limit=0.2,\n#             scale_limit=0.2,\n#             rotate_limit=20,\n#             interpolation=cv2.INTER_LINEAR,\n#             border_mode=cv2.BORDER_REFLECT_101,\n#             p=1,\n#          ),\n#          Normalize(mean=(0.485, 0.456, 0.406), std=(0.229, 0.224, 0.225), max_pixel_value=255.0, p=1.0),\n#             CoarseDropout(p=0.5),\n#             Cutout(p=0.5),\n#             ToTensorV2(p=1.0),\n        ], p=1.)\n  \n        \ndef get_valid_transforms():\n    return Compose([\n            CenterCrop(CFG['img_size'], CFG['img_size'], p=1.),\n            Resize(CFG['img_size'], CFG['img_size']),\n            Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225], max_pixel_value=255.0, p=1.0),\n            ToTensorV2(p=1.0),\n        ], p=1.)","execution_count":null,"outputs":[]},{"metadata":{"trusted":true},"cell_type":"code","source":"class CassvaImgClassifier(nn.Module):\n    def __init__(self, model_arch, n_class, pretrained=False):\n        super().__init__()\n        self.model = timm.create_model(model_arch, pretrained=pretrained)\n        n_features = self.model.classifier.in_features\n        self.model.classifier = nn.Linear(n_features, n_class)\n        '''\n        self.model.classifier = nn.Sequential(\n            nn.Dropout(0.3),\n            #nn.Linear(n_features, hidden_size,bias=True), nn.ELU(),\n            nn.Linear(n_features, n_class, bias=True)\n        )\n        '''\n    def forward(self, x):\n        x = self.model(x)\n        return x","execution_count":null,"outputs":[]},{"metadata":{"trusted":true},"cell_type":"code","source":"def prepare_dataloader(df, trn_idx, val_idx, data_root='../input/cassava-leaf-disease-classification/train_images/'):\n    \n    from catalyst.data.sampler import BalanceClassSampler\n    \n    train_ = df.loc[trn_idx,:].reset_index(drop=True)\n    valid_ = df.loc[val_idx,:].reset_index(drop=True)\n        \n    train_ds = CassavaDataset(train_, data_root, transforms=get_train_transforms(), output_label=True, one_hot_label=False, do_fmix=False, do_cutmix=True)\n    valid_ds = CassavaDataset(valid_, data_root, transforms=get_valid_transforms(), output_label=True)\n    \n    train_loader = torch.utils.data.DataLoader(\n        train_ds,\n        batch_size=CFG['train_bs'],\n        pin_memory=False,\n        drop_last=False,\n        shuffle=True,        \n        num_workers=CFG['num_workers'],\n        #sampler=BalanceClassSampler(labels=train_['label'].values, mode=\"downsampling\")\n    )\n    val_loader = torch.utils.data.DataLoader(\n        valid_ds, \n        batch_size=CFG['valid_bs'],\n        num_workers=CFG['num_workers'],\n        shuffle=False,\n        pin_memory=False,\n    )\n    return train_loader, val_loader\n\ndef train_one_epoch(epoch, model, loss_fn, optimizer, train_loader, device, scheduler=None, schd_batch_update=False):\n    model.train()\n\n    t = time.time()\n    running_loss = None\n\n    pbar = tqdm(enumerate(train_loader), total=len(train_loader))\n    for step, (imgs, image_labels) in pbar:\n        imgs = imgs.to(device).float()\n        image_labels = image_labels.to(device).long()\n\n        #print(image_labels.shape, exam_label.shape)\n        with autocast():\n            image_preds = model(imgs)   #output = model(input)\n            #print(image_preds.shape, exam_pred.shape)\n            #原版本\n            loss = loss_fn(image_preds, image_labels)\n            #bi-tempered-loss-pytorch\n\n            #loss = bi_tempered_logistic_loss(image_preds, image_labels, 0.8 , 1.2)\n            \n            scaler.scale(loss).backward()\n\n            if running_loss is None:\n                running_loss = loss.item()\n            else:\n                running_loss = running_loss * .99 + loss.item() * .01\n\n            if ((step + 1) %  CFG['accum_iter'] == 0) or ((step + 1) == len(train_loader)):\n                # may unscale_ here if desired (e.g., to allow clipping unscaled gradients)\n\n                scaler.step(optimizer)\n                scaler.update()\n                optimizer.zero_grad() \n                \n                if scheduler is not None and schd_batch_update:\n                    scheduler.step()\n\n            if ((step + 1) % CFG['verbose_step'] == 0) or ((step + 1) == len(train_loader)):\n                description = f'epoch {epoch} loss: {running_loss:.4f}'\n                \n                pbar.set_description(description)\n                \n    if scheduler is not None and not schd_batch_update:\n        scheduler.step()\n        \ndef valid_one_epoch(epoch, model, loss_fn, val_loader, device, scheduler=None, schd_loss_update=False):\n    model.eval()\n\n    t = time.time()\n    loss_sum = 0\n    sample_num = 0\n    image_preds_all = []\n    image_targets_all = []\n    \n    pbar = tqdm(enumerate(val_loader), total=len(val_loader))\n    for step, (imgs, image_labels) in pbar:\n        imgs = imgs.to(device).float()\n        image_labels = image_labels.to(device).long()\n        \n        image_preds = model(imgs)   #output = model(input)\n        #print(image_preds.shape, exam_pred.shape)\n        image_preds_all += [torch.argmax(image_preds, 1).detach().cpu().numpy()]\n        image_targets_all += [image_labels.detach().cpu().numpy()]\n        \n        #原版本\n        loss = loss_fn(image_preds, image_labels)\n        #bi-tempered-loss-pytorch\n\n        #loss = bi_tempered_logistic_loss(image_preds, image_labels, 0.8 , 1.2)\n        \n        loss_sum += loss.item()*image_labels.shape[0]\n        sample_num += image_labels.shape[0]  \n\n        if ((step + 1) % CFG['verbose_step'] == 0) or ((step + 1) == len(val_loader)):\n            description = f'epoch {epoch} loss: {loss_sum/sample_num:.4f}'\n            pbar.set_description(description)\n    \n    image_preds_all = np.concatenate(image_preds_all)\n    image_targets_all = np.concatenate(image_targets_all)\n    print('validation multi-class accuracy = {:.4f}'.format((image_preds_all==image_targets_all).mean()))\n    \n    if scheduler is not None:\n        if schd_loss_update:\n            scheduler.step(loss_sum/sample_num)\n        else:\n            scheduler.step()","execution_count":null,"outputs":[]},{"metadata":{"trusted":true},"cell_type":"code","source":"# reference: https://www.kaggle.com/c/siim-isic-melanoma-classification/discussion/173733\nclass MyCrossEntropyLoss(_WeightedLoss):\n    def __init__(self, weight=None, reduction='mean'):\n        super().__init__(weight=weight, reduction=reduction)\n        self.weight = weight\n        self.reduction = reduction\n\n    def forward(self, inputs, targets):\n        lsm = F.log_softmax(inputs, -1)\n\n        if self.weight is not None:\n            lsm = lsm * self.weight.unsqueeze(0)\n\n        loss = -(targets * lsm).sum(-1)\n\n        if  self.reduction == 'sum':\n            loss = loss.sum()\n        elif  self.reduction == 'mean':\n            loss = loss.mean()\n\n        return loss\n\n#方法尝试label smoothing（有效！）\nclass LabelSmoothLoss(nn.Module):\n\n    def __init__(self, smoothing=0.0):\n        super(LabelSmoothLoss, self).__init__()\n        self.smoothing = smoothing\n\n    def forward(self, input, target):\n        log_prob = F.log_softmax(input, dim=-1)\n        weight = input.new_ones(input.size()) * \\\n                 self.smoothing / (input.size(-1) - 1.)\n        weight.scatter_(-1, target.unsqueeze(-1), (1. - self.smoothing))\n        loss = (-weight * log_prob).sum(dim=-1).mean()\n        return loss","execution_count":null,"outputs":[]},{"metadata":{"trusted":true},"cell_type":"code","source":"if __name__ == '__main__':\n     # for training only, need nightly build pytorch\n\n    seed_everything(CFG['seed'])\n    \n    folds = StratifiedKFold(n_splits=CFG['fold_num'], shuffle=True, random_state=CFG['seed']).split(np.arange(train.shape[0]), train.label.values)\n    \n    for fold, (trn_idx, val_idx) in enumerate(folds):\n        # we'll train fold 0 first\n        if fold != 1 :\n            continue\n\n        print('Training with {} started'.format(fold))\n\n        print(len(trn_idx), len(val_idx))\n        train_loader, val_loader = prepare_dataloader(train, trn_idx, val_idx, data_root='../input/cassava-leaf-disease-classification/train_images/')\n\n        device = torch.device(CFG['device'])\n        \n        model = CassvaImgClassifier(CFG['model_arch'], train.label.nunique(), pretrained=True).to(device)\n        scaler = GradScaler()\n        #ranger\n        #optimizer=optim.Ranger(model.parameters())\n        optimizer = torch.optim.Adam(model.parameters(), lr=CFG['lr'], weight_decay=CFG['weight_decay'])\n\n        scheduler = torch.optim.lr_scheduler.CosineAnnealingWarmRestarts(optimizer, T_0=CFG['T_0'], T_mult=1, eta_min=CFG['min_lr'], last_epoch=-1)\n        #scheduler = torch.optim.lr_scheduler.OneCycleLR(optimizer=optimizer, pct_start=0.1, div_factor=25, \n        #                                                max_lr=CFG['lr'], epochs=CFG['epochs'], steps_per_epoch=len(train_loader))\n        \n#         loss_tr = nn.CrossEntropyLoss().to(device) #MyCrossEntropyLoss().to(device)\n#         loss_fn = nn.CrossEntropyLoss().to(device)\n        \n        #labelsmoothing\n#         loss_tr=LabelSmoothLoss(smoothing=0.2).to(device)\n#         loss_fn = LabelSmoothLoss(smoothing=0.2).to(device)\n        \n        #\n        loss_tr=TaylorCrossEntropyLoss().to(device)\n        loss_fn =TaylorCrossEntropyLoss().to(device)\n        \n        for epoch in range(CFG['epochs']):\n            train_one_epoch(epoch, model, loss_tr, optimizer, train_loader, device, scheduler=scheduler, schd_batch_update=False)\n\n            with torch.no_grad():\n                valid_one_epoch(epoch, model, loss_fn, val_loader, device, scheduler=None, schd_loss_update=False)\n\n            torch.save(model.state_dict(),'{}_fold_{}_{}'.format(CFG['model_arch'], fold, epoch))\n            \n        #torch.save(model.cnn_model.state_dict(),'{}/cnn_model_fold_{}_{}'.format(CFG['model_path'], fold, CFG['tag']))\n        del model, optimizer, train_loader, val_loader, scaler, scheduler\n        torch.cuda.empty_cache()","execution_count":null,"outputs":[]}],"metadata":{"kernelspec":{"language":"python","display_name":"Python 3","name":"python3"},"language_info":{"pygments_lexer":"ipython3","nbconvert_exporter":"python","version":"3.6.4","file_extension":".py","codemirror_mode":{"name":"ipython","version":3},"name":"python","mimetype":"text/x-python"}},"nbformat":4,"nbformat_minor":4}