Initial commit: herbarium baseline + app UI

.gitattributes

@@ -0,0 +1,2 @@
+*.pth filter=lfs diff=lfs merge=lfs -text
+*.joblib filter=lfs diff=lfs merge=lfs -text

README.md

@@ -0,0 +1,22 @@
+---
+title: Plant Species Classification
+emoji: 🌿
+colorFrom: green
+colorTo: blue
+sdk: gradio
+sdk_version: 5.49.1
+app_file: app.py
+pinned: false
+license: mit
+---
+
+# 🌿 Plant Species Classification
+
+This is a Gradio app for the AML Group Project by PsychicFireSong.
+
+It uses a `ConvNextV2` model fine-tuned on the Herbarium Field dataset to classify plant species from images.
+
+**Models Available:**
+- **Herbarium Species Classifier:** The primary model for classification.
+- **Future Model 1 (Placeholder):** Not yet implemented.
+- **Future Model 2 (Placeholder):** Not yet implemented.

app.py

@@ -0,0 +1,121 @@
+import gradio as gr
+
+from baseline.baseline_convnext import predict_convnext
+from baseline.baseline_infer import predict_baseline
+
+
+# --- Placeholder models (for future extensions) ---
+def predict_placeholder_1(image):
+    if image is None:
+        return "Please upload an image."
+    return "Model 2 is not available yet. Please check back later."
+
+
+def predict_placeholder_2(image):
+    if image is None:
+        return "Please upload an image."
+    return "Model 3 is not available yet. Please check back later."
+
+
+# --- Main Prediction Logic ---
+def predict(model_choice, image):
+    if model_choice == "Herbarium Species Classifier":
+        # Friend's ConvNeXt mix-stream CNN baseline
+        return predict_convnext(image)
+    elif model_choice == "Baseline (DINOv2 + LogReg)":
+        # Your plant-pretrained DINOv2 + Logistic Regression baseline
+        return predict_baseline(image)
+    elif model_choice == "Future Model 1 (Placeholder)":
+        return predict_placeholder_1(image)
+    elif model_choice == "Future Model 2 (Placeholder)":
+        return predict_placeholder_2(image)
+    else:
+        return "Invalid model selected."
+
+
+# --- Gradio Interface ---
+with gr.Blocks(theme=gr.themes.Soft(), css="style.css") as demo:
+    with gr.Column(elem_id="app-wrapper"):
+        # Header
+        gr.Markdown(
+            """
+            <div id="app-header">
+              <h1>🌿 Plant Species Classification</h1>
+              <h3>AML Group Project – PsychicFireSong</h3>
+            </div>
+            """,
+            elem_id="app-header",
+        )
+
+        # Badges row
+        gr.Markdown(
+            """
+            <div id="badge-row">
+              <span class="badge">Herbarium + Field images</span>
+              <span class="badge">ConvNeXtV2 mix-stream CNN</span>
+              <span class="badge">DINOv2 + Logistic Regression</span>
+            </div>
+            """,
+            elem_id="badge-row",
+        )
+
+        # Main card
+        with gr.Row(elem_id="main-card"):
+            # Left side: model + image
+            with gr.Column(scale=1, elem_id="left-panel"):
+                model_selector = gr.Dropdown(
+                    label="Select model",
+                    choices=[
+                        "Herbarium Species Classifier",
+                        "Baseline (DINOv2 + LogReg)",
+                        "Future Model 1 (Placeholder)",
+                        "Future Model 2 (Placeholder)",
+                    ],
+                    value="Herbarium Species Classifier",
+                )
+
+                gr.Markdown(
+                    """
+                    <div id="model-help">
+                      <b>Herbarium Species Classifier</b> – end-to-end ConvNeXtV2 CNN.<br>
+                      <b>Baseline</b> – plant-pretrained DINOv2 features + logistic regression head.
+                    </div>
+                    """,
+                    elem_id="model-help",
+                )
+
+                image_input = gr.Image(
+                    type="pil",
+                    label="Upload plant image",
+                )
+                submit_button = gr.Button("Classify 🌱", variant="primary")
+
+            # Right side: predictions
+            with gr.Column(scale=1, elem_id="right-panel"):
+                output_label = gr.Label(
+                    label="Top 5 predictions",
+                    num_top_classes=5,
+                )
+
+        submit_button.click(
+            fn=predict,
+            inputs=[model_selector, image_input],
+            outputs=output_label,
+        )
+
+        # Optional examples (keep empty if you don't have images)
+        gr.Examples(
+            examples=[],
+            inputs=image_input,
+            outputs=output_label,
+            fn=lambda img: predict("Herbarium Species Classifier", img),
+            cache_examples=False,
+        )
+
+        gr.Markdown(
+            "Built for the AML course – compare CNN vs. DINOv2 feature-extractor baselines.",
+            elem_id="footer",
+        )
+
+if __name__ == "__main__":
+    demo.launch()

baseline/baseline_convnext.py

@@ -0,0 +1,60 @@
+# baseline/baseline_convnext.py
+from pathlib import Path
+import torch
+import pandas as pd
+from PIL import Image
+from torchvision import transforms
+from transformers import ConvNextV2ForImageClassification
+
+ROOT_DIR = Path(__file__).resolve().parent.parent
+BASELINE_DIR = Path(__file__).resolve().parent
+LIST_DIR = ROOT_DIR / "list"
+MODEL_PATH = BASELINE_DIR / "herbarium_convnext_v2_base.pth"
+
+DEVICE = torch.device("cuda" if torch.cuda.is_available() else "cpu")
+
+# species list
+species_df = pd.read_csv(
+    LIST_DIR / "species_list.txt",
+    sep=";",
+    header=None,
+    names=["class_id", "species_name"],
+)
+class_names = list(species_df["species_name"])
+num_labels = len(class_names)
+
+data_transforms = transforms.Compose([
+    transforms.Resize(256),
+    transforms.CenterCrop(224),
+    transforms.ToTensor(),
+    transforms.Normalize([0.485, 0.456, 0.406],
+                         [0.229, 0.224, 0.225]),
+])
+
+def _load_model():
+    model = ConvNextV2ForImageClassification.from_pretrained(
+        "facebook/convnextv2-base-22k-224",
+        num_labels=num_labels,
+        ignore_mismatched_sizes=True,
+    )
+    if MODEL_PATH.is_file():
+        state = torch.load(MODEL_PATH, map_location=DEVICE)
+        model.load_state_dict(state)
+    else:
+        print(f"[convnext] WARNING: {MODEL_PATH} not found, using HF weights only.")
+    model.to(DEVICE)
+    model.eval()
+    return model
+
+convnext_model = _load_model()
+
+def predict_convnext(image: Image.Image):
+    if image is None:
+        return "Please upload an image."
+    x = data_transforms(image).unsqueeze(0).to(DEVICE)
+    with torch.no_grad():
+        logits = convnext_model(x).logits
+        prob = torch.softmax(logits, dim=1)[0]
+        top5_prob, top5_idx = torch.topk(prob, 5)
+    return {class_names[i]: float(p)
+            for i, p in zip(top5_idx.cpu().numpy(), top5_prob.cpu().numpy())}

baseline/baseline_infer.py

@@ -0,0 +1,172 @@
+from __future__ import annotations
+from pathlib import Path
+from typing import Dict
+
+import numpy as np
+import pandas as pd
+from PIL import Image
+
+import torch
+from torchvision import transforms
+import timm
+from timm.models.vision_transformer import resize_pos_embed
+import joblib
+
+
+# ----------------------- paths & device -----------------------
+ROOT_DIR = Path(__file__).resolve().parent.parent      # AMLGroupSpaceFinal/
+BASELINE_DIR = ROOT_DIR / "baseline"
+LIST_DIR = ROOT_DIR / "list"
+
+PLANT_CKPT_PATH = BASELINE_DIR / "plant_dinov2_patch14.pth"
+LOGREG_PATH = BASELINE_DIR / "logreg_baseline.joblib"
+SCALER_PATH = BASELINE_DIR / "scaler_baseline.joblib"
+SPECIES_LIST_PATH = LIST_DIR / "species_list.txt"
+
+DEVICE = torch.device("cuda" if torch.cuda.is_available() else "cpu")
+
+
+# ----------------------- helpers (trimmed from evaluate.py) -----------------------
+def read_species(p: Path):
+    """Read species_list.txt and return list of species names in index order."""
+    rows = []
+    with open(p, "r", encoding="utf-8") as f:
+        for ln in f:
+            ln = ln.strip()
+            if not ln or ln.startswith("#"):
+                continue
+            if ";" in ln:
+                cid, name = ln.split(";", 1)
+            else:
+                parts = ln.split()
+                cid, name = parts[0], " ".join(parts[1:]) if len(parts) > 1 else ""
+            try:
+                cid = int(cid)
+            except ValueError:
+                continue
+            rows.append((cid, name))
+    df = pd.DataFrame(rows, columns=["class_id", "species_name"])
+    # same order as in training: iterrows order
+    names = list(df["species_name"])
+    return names
+
+
+def pool_feats(out):
+    feats = out
+    if isinstance(out, dict):
+        for key in ("pooled", "x_norm_clstoken", "cls_token", "x"):
+            if key in out:
+                feats = out[key]
+                break
+    if isinstance(feats, (list, tuple)):
+        feats = feats[0]
+    if isinstance(feats, torch.Tensor) and feats.dim() == 3:
+        feats = feats[:, 0] if feats.size(1) > 1 else feats.mean(dim=1)
+    if isinstance(feats, torch.Tensor) and feats.dim() > 2:
+        feats = feats.flatten(1)
+    return feats
+
+
+def _unwrap_state_dict(obj):
+    if isinstance(obj, dict):
+        for k in ("state_dict", "model", "module", "ema", "shadow",
+                  "backbone", "net", "student", "teacher"):
+            if k in obj and isinstance(obj[k], dict):
+                return obj[k]
+    return obj
+
+
+def _strip_prefixes(sd, prefixes=("module.", "backbone.", "model.", "student.")):
+    out = {}
+    for k, v in sd.items():
+        for p in prefixes:
+            if k.startswith(p):
+                k = k[len(p):]
+        out[k] = v
+    return out
+
+
+def maybe_load_plant_ckpt(model, ckpt_path: Path):
+    if not ckpt_path.is_file():
+        print(f"[baseline] plant ckpt not found at {ckpt_path}, using generic DINOv2 weights.")
+        return
+    try:
+        sd = torch.load(ckpt_path, map_location="cpu")
+        sd = _unwrap_state_dict(sd)
+        sd = _strip_prefixes(sd)
+
+        msd = model.state_dict()
+        if "pos_embed" in sd and "pos_embed" in msd and sd["pos_embed"].shape != msd["pos_embed"].shape:
+            sd["pos_embed"] = resize_pos_embed(sd["pos_embed"], msd["pos_embed"])
+            print(f"[baseline] interpolated pos_embed to {tuple(msd['pos_embed'].shape)}")
+
+        missing, unexpected = model.load_state_dict(sd, strict=False)
+        print(f"[baseline] loaded plant ckpt; missing={len(missing)} unexpected={len(unexpected)}")
+    except Exception as e:
+        print(f"[baseline] failed to load '{ckpt_path}': {e}")
+
+
+def build_backbone(size: int = 224):
+    model = timm.create_model(
+        "vit_base_patch14_dinov2",
+        pretrained=True,    # generic DINOv2 as fallback
+        num_classes=0,      # features only
+        img_size=size,
+        pretrained_cfg_overlay=dict(input_size=(3, size, size)),
+    ).to(DEVICE)
+
+    pe = getattr(model, "patch_embed", None)
+    if pe is not None:
+        if hasattr(pe, "img_size"):
+            pe.img_size = (size, size)
+        if hasattr(pe, "strict_img_size"):
+            pe.strict_img_size = False
+
+    maybe_load_plant_ckpt(model, PLANT_CKPT_PATH)
+    model.eval()
+    for p in model.parameters():
+        p.requires_grad = False
+    return model
+
+
+# ----------------------- global objects (loaded once) -----------------------
+IMAGE_SIZE = 224
+
+species_names = read_species(SPECIES_LIST_PATH)
+num_classes = len(species_names)
+
+backbone = build_backbone(IMAGE_SIZE)
+
+transform = transforms.Compose([
+    transforms.Resize(int(IMAGE_SIZE * 1.12)),
+    transforms.CenterCrop(IMAGE_SIZE),
+    transforms.ToTensor(),
+    transforms.Normalize([0.485, 0.456, 0.406],
+                         [0.229, 0.224, 0.225]),
+])
+
+scaler = joblib.load(SCALER_PATH)
+logreg = joblib.load(LOGREG_PATH)
+
+
+# ----------------------- public API for Gradio -----------------------
+def predict_baseline(image: Image.Image, top_k: int = 5) -> Dict[str, float]:
+    """
+    Run DINOv2 + Logistic Regression baseline on a single PIL image.
+    Returns {class_name: probability} for the top_k classes.
+    """
+    if image is None:
+        return {}
+
+    x = transform(image).unsqueeze(0).to(DEVICE)
+
+    with torch.no_grad():
+        out = backbone.forward_features(x)
+        feats = pool_feats(out).cpu().numpy()
+
+    feats_scaled = scaler.transform(feats)
+    probs = logreg.predict_proba(feats_scaled)[0]  # shape [num_classes]
+
+    top_idx = np.argsort(-probs)[:top_k]
+    result = {species_names[i]: float(probs[i]) for i in top_idx}
+    return result

baseline/herbarium_convnext_v2_base.pth

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:837cca126e235c0ae822770470e38a3621b81b0ba7e915aaef2b15a7f66914e6
+size 351335085

baseline/logreg_baseline.joblib

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c886ec6469fadd59f092adc6f3e08e3cc0859f18c7e7847f20299b775f05a4ba
+size 616839

baseline/plant_dinov2_patch14.pth

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:1fe189d76a0ec0128e8a9d4959a218e10c6adc60ab21d0d23b65c7080d1a4407
+size 346384519

baseline/scaler_baseline.joblib

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:5f569d9cc90e8119c2820ef64b790d4f3bb75faadd7d5bf6d5581a4ea07e8c57
+size 19047

list/species_list.txt

@@ -0,0 +1,100 @@
+105951; Maripa glabra Choisy
+106023; Merremia umbellata (L.) Hallier f.
+106387; Costus arabicus L.
+106461; Costus scaber Ruiz Pav.
+106466; Costus spiralis (Jacq.) Roscoe
+110432; Evodianthus funifer (Poit.) Lindm.
+116853; Pteridium arachnoideum (Kaulf.) Maxon
+119986; Olfersia cervina (L.) Kunze
+120497; Diospyros capreifolia Mart. ex Hiern
+121836; Sloanea grandiflora Sm.
+121841; Sloanea guianensis (Aubl.) Benth.
+12254; Anacardium occidentale L.
+12518; Mangifera indica L.
+125412; Sphyrospermum cordifolium Benth.
+126895; Syngonanthus caulescens (Poir.) Ruhland
+127007; Tonina fluviatilis Aubl.
+127097; Erythroxylum fimbriatum Peyr.
+127151; Erythroxylum macrophyllum Cav.
+127242; Erythroxylum squamatum Sw.
+12910; Spondias mombin L.
+12922; Tapirira guianensis Aubl.
+129645; Croton schiedeanus Schltdl.
+130657; Euphorbia cotinifolia L.
+131079; Euphorbia heterophylla L.
+131736; Euphorbia prostrata Aiton
+132107; Euphorbia thymifolia L.
+132113; Euphorbia tithymaloides L.
+132431; Hura crepitans L.
+132476; Jatropha curcas L.
+132501; Jatropha gossypiifolia L.
+13276; Annona ambotay Aubl.
+13325; Annona foetida Mart.
+13330; Annona glabra L.
+133595; Ricinus communis L.
+133617; Sapium glandulosum (L.) Morong
+13370; Annona muricata L.
+136761; Potalia amara Aubl.
+138662; Chrysothemis pulchella (Donn ex Sims) Decne.
+140367; Lembocarpus amoenus Leeuwenb.
+141068; Sinningia incarnata (Aubl.) D.L.Denham
+141332; Dicranopteris flexuosa (Schrad.) Underw.
+141336; Dicranopteris pectinata (Willd.) Underw.
+142550; Heliconia chartacea Lane ex Barreiros
+142736; Hernandia guianensis Aubl.
+143496; Hymenophyllum hirsutum (L.) Sw.
+14353; Guatteria ouregou (Aubl.) Dunal
+143706; Trichomanes diversifrons (Bory) Mett. ex Sadeb.
+143758; Trichomanes punctatum Poir.
+14401; Guatteria scandens Ducke
+144394; Didymochlaena truncatula (Sw.) J. Sm.
+145020; Cipura paludosa Aubl.
+148220; Aegiphila macrantha Ducke
+148977; Clerodendrum paniculatum L.
+149264; Congea tomentosa Roxb.
+149682; Gmelina philippensis Cham.
+149919; Holmskioldia sanguinea Retz.
+150135; Hyptis lanceolata Poir.
+15014; Rollinia mucosa (Jacq.) Baill.
+151469; Ocimum campechianum Mill.
+151593; Orthosiphon aristatus (Blume) Miq.
+15318; Xylopia aromatica (Lam.) Mart.
+15330; Xylopia cayennensis Maas
+15355; Xylopia frutescens Aubl.
+156516; Aniba guianensis Aubl.
+156526; Aniba megaphylla Mez
+158341; Nectandra cissiflora Nees
+158592; Ocotea cernua (Nees) Mez
+158653; Ocotea floribunda (Sw.) Mez
+158736; Ocotea longifolia Kunth
+158793; Ocotea oblonga (Meisn.) Mez
+158833; Ocotea puberula (Rich.) Nees
+159434; Couratari guianensis Aubl.
+159516; Eschweilera parviflora (Aubl.) Miers
+159518; Eschweilera pedicellata (Rich.) S.A.Mori
+160570; Acacia mangium Willd.
+166822; Caesalpinia pulcherrima (L.) Sw.
+166869; Cajanus cajan (L.) Millsp.
+169293; Crotalaria retusa L.
+171727; Erythrina fusca Lour.
+173914; Inga alba (Sw.) Willd.
+173972; Inga capitata Desv.
+174017; Inga edulis Mart.
+177730; Mimosa pigra L.
+177775; Mimosa pudica L.
+189669; Punica granatum L.
+191642; Adansonia digitata L.
+19165; Allamanda cathartica L.
+192311; Ceiba pentandra (L.) Gaertn.
+194035; Hibiscus rosa-sinensis L.
+19489; Asclepias curassavica L.
+209328; Psidium guineense Sw.
+211059; Nephrolepis biserrata (Sw.) Schott
+244705; Averrhoa carambola L.
+248392; Turnera ulmifolia L.
+254180; Piper peltatum L.
+275029; Eichhornia crassipes (Mart.) Solms
+280085; Ceratopteris thalictroides (L.) Brongn.
+280698; Pityrogramma calomelanos (L.) Link
+285398; Cassipourea guianensis Aubl.
+29686; Oreopanax capitatus (Jacq.) Decne. Planch.

requirements.txt

@@ -0,0 +1,6 @@
+torch
+torchvision
+transformers
+pandas
+gradio
+accelerate

style.css

@@ -0,0 +1,94 @@
+/* ---------- Global background & typography ---------- */
+
+body {
+    background: radial-gradient(circle at top, #e0f2fe 0, #f9fafb 45%, #f8fafc 100%);
+    font-family: system-ui, -apple-system, BlinkMacSystemFont, "SF Pro Text", sans-serif;
+}
+
+/* ---------- Main wrapper ---------- */
+
+#app-wrapper {
+    max-width: 1100px;
+    margin: 0 auto;
+    padding: 24px 16px 40px;
+}
+
+/* ---------- Header ---------- */
+
+#app-header h1 {
+    font-size: 2.2rem;
+    margin-bottom: 0.2rem;
+}
+
+#app-header h3 {
+    margin-top: 0;
+    font-weight: 500;
+    color: #6b7280;
+}
+
+/* ---------- Info chips under title ---------- */
+
+#badge-row {
+    margin-top: 6px;
+    margin-bottom: 6px;
+}
+
+.badge {
+    display: inline-flex;
+    align-items: center;
+    padding: 4px 10px;
+    margin-right: 8px;
+    margin-bottom: 4px;
+    border-radius: 999px;
+    background: #ecfdf5;
+    border: 1px solid #bbf7d0;
+    font-size: 12px;
+    color: #166534;
+}
+
+/* ---------- Main card ---------- */
+
+#main-card {
+    margin-top: 18px;
+    padding: 18px 20px 22px;
+    border-radius: 20px;
+    background: rgba(255, 255, 255, 0.98);
+    box-shadow: 0 22px 48px rgba(15, 23, 42, 0.18);
+}
+
+/* Left (controls) / right (outputs) panels */
+
+#left-panel {
+    border-right: 1px solid #e5e7eb;
+    padding-right: 18px;
+}
+
+#right-panel {
+    padding-left: 18px;
+}
+
+/* Small helper text under model dropdown */
+
+#model-help {
+    font-size: 12px;
+    color: #6b7280;
+    margin-top: 4px;
+}
+
+/* Make the main button a bit more pill-like */
+
+button.primary,
+.gr-button-primary {
+    border-radius: 999px !important;
+    padding: 8px 18px !important;
+    font-weight: 600 !important;
+}
+
+/* Footer */
+
+#footer {
+    margin-top: 18px;
+    text-align: center;
+    font-size: 12px;
+    color: #94a3b8;
+}