fixed ckpts

checkpoints/dino_2_optuna_05242055.ckpt

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:9d215a8f729a039d997e9a8fd6ebc1637afa905fcd96f4145b6caf1432b89270
-size 187800867
+oid sha256:bd95a08e0e7a725425d91810db61fc3a1167abe59ddf7ceedd067304dfc8b097
+size 187793106

checkpoints/dino_2_optuna_05242156.ckpt

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:145e76ffe0e90e5ca4fabd8dac47376037e06f5b57605b2c06072374cc0cfef1
-size 187800867
+oid sha256:403a4f866df7f3bc5358c04a4f8e188e49cab90d84e7812a9261abfc9fab3bc3
+size 187793106

checkpoints/dino_2_optuna_05242231.ckpt

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:1d3658ac33b59e3058c2e907c22a2fe65811c9595c9d0e26dd090ee43e70c89d
-size 187800867
+oid sha256:1bca4793b9032935502276fb4e2c045a661cd7b158e2633f0f27bc5d017e4141
+size 187793106

checkpoints/dino_2_optuna_05242344.ckpt

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:d023ac97e70bc74e970d1bd3c7335652d26084f4e4adf23582d5ec6ebdc6eb18
-size 187800867
+oid sha256:029202d06bad69443016b4b61149a7725dae9d1ba6faac04089c63ac01040bf7
+size 187793106

checkpoints/dino_optuna_05241222.ckpt

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:635f55c30c89d6f1e1617bb90d526c5efd512fc788bbdb09f384ee11581c6e52
-size 187800867
+oid sha256:5585d99338a3911e9c133e829c6a1218c9f18fc6cb7daeffbfc9fdf669c92c86
+size 187792874

checkpoints/dino_optuna_05241257.ckpt

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:25de44e5079bc16fd18596403472bc44e15fb0675090afdfa7214bb1c611884e
-size 187800867
+oid sha256:885543336dd814a2bb598d96ca99a01564e4d7eca69778ceabe307ecb14a6d89
+size 187792874

checkpoints/dino_optuna_05241449.ckpt

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:b2e2be600596595d62214d7acd46101bfd847b3b6f2be4667bc3f9542f625974
-size 187800867
+oid sha256:ca08ea13541b791d1d46f567727cccc4b1e9824064694803bf92473fc79f4192
+size 187792874

script.py

@@ -1,7 +1,4 @@
 import os
-from typing import List
-
-import cv2
 import numpy as np
 import pandas as pd
 import timm
@@ -9,94 +6,16 @@ import torch
 import torch.nn as nn
 import torch.nn.functional as F
 import torchvision.transforms as T
-from albumentations import (
-    CenterCrop,
-    Compose,
-    HorizontalFlip,
-    Normalize,
-    PadIfNeeded,
-    RandomBrightnessContrast,
-    RandomCrop,
-    RandomResizedCrop,
-    Resize,
-    VerticalFlip,
-)
-from albumentations.pytorch import ToTensorV2
 from PIL import Image
-from timm.layers import LayerNorm2d, SelectAdaptivePool2d
 from timm.models.metaformer import MlpHead
 from torch.utils.data import DataLoader, Dataset
 from tqdm import tqdm
 
 DIM = 518
-
-def get_transforms(*, data, model=None, width=None, height=None):
-    assert data in ("train", "valid")
-
-    width = width if width else DIM
-    height = height if height else DIM
-
-    model_mean = list(model.default_cfg["mean"]) if model else (0.5, 0.5, 0.5)
-    model_std = list(model.default_cfg["std"]) if model else (0.5, 0.5, 0.5)
-
-    if data == "train":
-        return Compose(
-            [
-                RandomResizedCrop(width, height, scale=(0.6, 1.0)),
-                HorizontalFlip(p=0.5),
-                VerticalFlip(p=0.5),
-                RandomBrightnessContrast(p=0.2),
-                Normalize(mean=model_mean, std=model_std),
-                ToTensorV2(),
-            ]
-        )
-
-    elif data == "valid":
-        return Compose(
-            [
-                Resize(width, height),
-                Normalize(mean=model_mean, std=model_std),
-                ToTensorV2(),
-            ]
-        )
-
-
-def generate_embeddings(metadata_file_path, root_dir):
-
-    metadata_df = pd.read_csv(metadata_file_path)
-
-    transforms = get_transforms(data="valid", width=DIM, height=DIM)
-
-    test_dataset = ImageMetadataDataset(
-        metadata_df, local_filepath=root_dir, transform=transforms
-    )
-
-    loader = DataLoader(test_dataset, batch_size=1, shuffle=False, num_workers=4)
-
-    device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
-    model = timm.create_model(
-        "timm/vit_large_patch14_reg4_dinov2.lvd142m", pretrained=True
-    )
-    model = model.to(device)
-    model.eval()
-
-    all_embs = []
-    for data in tqdm(loader):
-
-        img, _ = data
-        img = img.to(device)
-
-        emb = model.forward(img)
-
-        all_embs.append(emb.detach().cpu().numpy())
-
-    all_embs = np.vstack(all_embs)
-
-    embs_list = [x for x in all_embs]
-    metadata_df["embedding"] = embs_list
-
-    return metadata_df
-
+DATE_SIZE = 4
+GEO_SIZE = 7
+SUBSTRATE_SIZE = 73
+NUM_CLASSES = 1717
 
 TIME = ["m0", "m1", "d0", "d1"]
 GEO = ["g0", "g1", "g2", "g3", "g4", "g5", "g_float"]
@@ -177,6 +96,32 @@ SUBSTRATE = [
 ]
 
 
+class ImageDataset(Dataset):
+    def __init__(self, df, local_filepath):
+        self.df = df
+        self.transform = T.Compose(
+            [
+                T.Resize((DIM, DIM)),
+                T.ToTensor(),
+                T.Normalize([0.5, 0.5, 0.5], [0.5, 0.5, 0.5]),
+            ]
+        )
+
+        self.local_filepath = local_filepath
+
+        self.filepaths = df["image_path"].to_list()
+
+    def __len__(self):
+        return len(self.df)
+
+    def __getitem__(self, idx):
+        image_path = os.path.join(self.local_filepath, self.filepaths[idx])
+
+        image = Image.open(image_path).convert("RGB")
+
+        return self.transform(image)
+
+
 class EmbeddingMetadataDataset(Dataset):
     def __init__(self, df):
         self.df = df
@@ -203,50 +148,37 @@ class EmbeddingMetadataDataset(Dataset):
         return embedding, metadata
 
 
-class ImageMetadataDataset(Dataset):
-    def __init__(self, df, transform=None, local_filepath=None):
-        self.df = df
-        self.transform = transform
-        self.local_filepath = local_filepath
+def generate_embeddings(metadata_file_path, root_dir):
 
-        self.filepaths = (
-            df["image_path"].to_list()
-        )
-        self.metadata_date = df[TIME].to_numpy()
-        self.metadata_geo = df[GEO].to_numpy()
-        self.metadata_substrate = df[SUBSTRATE].to_numpy()
+    metadata_df = pd.read_csv(metadata_file_path)
 
-    def __len__(self):
-        return len(self.df)
+    test_dataset = ImageDataset(metadata_df, local_filepath=root_dir)
 
-    def __getitem__(self, idx):
-        file_path = os.path.join(self.local_filepath, self.filepaths[idx])
+    loader = DataLoader(test_dataset, batch_size=2, shuffle=False, num_workers=4)
 
-        try:
-            image = cv2.imread(file_path)
-            image = cv2.cvtColor(image, cv2.COLOR_BGR2RGB)
-        except:
-            print(file_path)
+    device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
+    model = timm.create_model(
+        "timm/vit_large_patch14_reg4_dinov2.lvd142m", pretrained=True
+    )
 
-        if self.transform:
-            augmented = self.transform(image=image)
-            image = augmented["image"]
+    model = model.to(device)
+    model.eval()
 
-        metadata = {
-            "date": torch.from_numpy(self.metadata_date[idx, :]).type(torch.float),
-            "geo": torch.from_numpy(self.metadata_geo[idx, :]).type(torch.float),
-            "substr": torch.from_numpy(self.metadata_substrate[idx, :]).type(
-                torch.float
-            ),
-        }
+    all_embs = []
+    for img in tqdm(loader):
 
-        return image, metadata
+        img = img.to(device)
 
+        emb = model.forward(img)
 
-DATE_SIZE = 4
-GEO_SIZE = 7
-SUBSTRATE_SIZE = 73
-NUM_CLASSES = 1717
+        all_embs.append(emb.detach().cpu().numpy())
+
+    all_embs = np.vstack(all_embs)
+
+    embs_list = [x for x in all_embs]
+    metadata_df["embedding"] = embs_list
+
+    return metadata_df
 
 
 class StarReLU(nn.Module):
@@ -323,8 +255,7 @@ class FungiMEEModel(nn.Module):
 
         full_emb = torch.stack(
             (img_emb, date_emb, geo_emb, substr_emb), dim=1
-        )  # .unsqueeze(0)
-        # print(full_emb.shape)
+        ) 
 
         cls_emb = self.encoder.forward(full_emb)[:, 0, :].squeeze(1)
 
@@ -334,8 +265,6 @@ class FungiMEEModel(nn.Module):
 
         logits = self.forward(img_emb, metadata)
 
-        # Any preprocess happens here
-
         return logits.argmax(1).tolist()
 
 
@@ -386,56 +315,12 @@ def is_gpu_available():
     return torch.cuda.is_available()
 
 
-class PytorchWorker:
-    """Run inference using ONNX runtime."""
-
-    def __init__(
-        self, model_path: str, model_name: str, number_of_categories: int = 1605
-    ):
-
-        def _load_model(model_name, model_path):
-
-            print("Setting up Pytorch Model")
-            self.device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
-            print(f"Using devide: {self.device}")
-
-            model = timm.create_model(model_name, num_classes=0, pretrained=False)
-            # weights = torch.load(model_path, map_location=self.device)
-            # model.load_state_dict({w.replace("model.", ""): v for w, v in weights.items()})
-
-            return model.to(self.device).eval()
-
-        self.model = _load_model(model_name, model_path)
-
-        self.transforms = T.Compose(
-            [
-                T.Resize((518, 518)),
-                T.ToTensor(),
-                T.Normalize([0.5, 0.5, 0.5], [0.5, 0.5, 0.5]),
-            ]
-        )
-
-    def predict_image(self, image: np.ndarray):
-        """Run inference using ONNX runtime.
-
-        :param image: Input image as numpy array.
-        :return: A list with logits and confidences.
-        """
-
-        self.model(self.transforms(image).unsqueeze(0).to(self.device))
-
-        return [-1]
-
-
-def make_submission(metadata_df, model_names=None):
+def make_submission(metadata_df):
 
     OUTPUT_CSV_PATH = "./submission.csv"
-
-    """Make submission with given """
-
     BASE_CKPT_PATH = "./checkpoints"
 
-    model_names = model_names or os.listdir(BASE_CKPT_PATH)
+    model_names = os.listdir(BASE_CKPT_PATH)
 
     models = []
 
@@ -446,7 +331,7 @@ def make_submission(metadata_df, model_names=None):
         ckpt = torch.load(ckpt_path)
         model = FungiMEEModel()
         model.load_state_dict(
-            {w: ckpt["state_dict"]["model." + w] for w in model.state_dict().keys()}
+            {w: ckpt["model." + w] for w in model.state_dict().keys()}
         )
         model.eval()
         model.cuda()
@@ -487,18 +372,18 @@ if __name__ == "__main__":
     MODEL_PATH = "metaformer-s-224.pth"
     MODEL_NAME = "timm/vit_base_patch14_reg4_dinov2.lvd142m"
 
-    # # Real submission
-    import zipfile
+    # # # Real submission
+    # import zipfile
 
-    with zipfile.ZipFile("/tmp/data/private_testset.zip", "r") as zip_ref:
-        zip_ref.extractall("/tmp/data")
+    # with zipfile.ZipFile("/tmp/data/private_testset.zip", "r") as zip_ref:
+    #     zip_ref.extractall("/tmp/data")
 
-    metadata_file_path = "./_test_preprocessed.csv"
-    root_dir = "/tmp/data"
+    # metadata_file_path = "./_test_preprocessed.csv"
+    # root_dir = "/tmp/data"
 
     # Test submission
-    # metadata_file_path = "../trial_submission.csv"
-    # root_dir = "../data/DF_FULL"
+    metadata_file_path = "../trial_submission.csv"
+    root_dir = "../data/DF_FULL"
 
     ##############