ট্রান্সফার লার্নিং

কখন Transfer Learning?

• আপনার ডেটাসেট ছোট (১০০–১০,০০০)।
• Domain ImageNet-এর কাছাকাছি (natural images)।
• Compute সীমিত — scratch থেকে train অসম্ভব।

দুটি কৌশল

Keras দিয়ে Feature Extraction

from tensorflow.keras.applications import EfficientNetB0
from tensorflow.keras import layers, models

base = EfficientNetB0(include_top=False, weights="imagenet",
                      input_shape=(224,224,3))
base.trainable = False           # freeze

inp = layers.Input((224,224,3))
x = base(inp, training=False)
x = layers.GlobalAveragePooling2D()(x)
x = layers.Dropout(0.3)(x)
out = layers.Dense(num_classes, activation="softmax")(x)

model = models.Model(inp, out)
model.compile(optimizer="adam", loss="sparse_categorical_crossentropy",
              metrics=["accuracy"])
model.fit(train_ds, validation_data=val_ds, epochs=10)

Fine-tuning

base.trainable = True
# শেষ 30 layer ছাড়া সব freeze
for layer in base.layers[:-30]:
    layer.trainable = False

model.compile(optimizer=tf.keras.optimizers.Adam(1e-5),  # ছোট lr!
              loss="sparse_categorical_crossentropy",
              metrics=["accuracy"])
model.fit(train_ds, validation_data=val_ds, epochs=10)

PyTorch দিয়ে Fine-tuning

import torch, torch.nn as nn
from torchvision import models

net = models.resnet50(weights="IMAGENET1K_V2")
for p in net.parameters():
    p.requires_grad = False

net.fc = nn.Linear(net.fc.in_features, num_classes)

opt = torch.optim.Adam(net.fc.parameters(), lr=1e-3)
# fine-tune stage:
for p in net.layer4.parameters():
    p.requires_grad = True
opt = torch.optim.Adam(filter(lambda p: p.requires_grad, net.parameters()), lr=1e-5)

জনপ্রিয় Backbone

• ResNet50 / 101 — নির্ভরযোগ্য baseline।
• EfficientNet B0-B7 — accuracy/parameter ভারসাম্য।
• ConvNeXt — আধুনিক, transformer-comparable।
• ViT, Swin — Vision Transformer পরিবার।

সাধারণ ভুল

অনুশীলন

সারসংক্ষেপ