Bài kiểm tra có liên kết hướng dẫn · Cứng Cấp độ

Bài kiểm tra kích thước lô quan trọng

Khám phá kích thước lô quan trọng thông qua số lượng cập nhật, ví dụ được xử lý, nhiễu độ dốc, so sánh thời gian đến mục tiêu và giới hạn phần cứng thực tế.

Đường dẫn hướng dẫn liên quanKích thước lô tới hạn
Câu hỏi 1 của 8

Which behavior characterizes batches far beyond a training problem’s critical batch scale?

Xem trước mọi câu hỏi8 câu hỏi

Câu trả lời và giải thích xuất hiện khi bạn làm bài kiểm tra.

  1. Which behavior characterizes batches far beyond a training problem’s critical batch scale?
    • Additional examples per update give diminishing reductions in required updates
    • Training becomes mathematically impossible
    • The model always stops fitting in memory
    • Each batch doubling must halve elapsed time
  2. Why is the largest batch that fits in memory not necessarily the critical batch size?
    • Memory capacity determines every gradient statistic
    • Memory fit and optimization efficiency describe different constraints
    • Critical batch size counts only storage bytes
    • A fitting batch always minimizes time to target
  3. At a fixed batch of 256 and 4,000 updates, how many examples are processed?
    • 4,256
    • 15.625
    • 1,024,000
    • 512,000
  4. In the worked comparison, increasing the batch from 256 to 512 changes required updates from 4,000 to 2,200. What happens to processed examples?
    • They fall by 45%
    • They are unchanged
    • They double
    • They rise from 1,024,000 to 1,126,400, a 10% increase
  5. Which measurement is still needed before concluding that fewer training updates saved elapsed time?
    • Actual update or full-run timings on the intended hardware
    • Only the checkpoint filename
    • Only the parameter count
    • Only the first batch’s loss
  6. What statistic did the cited large-batch research connect with the largest useful batch range?
    • The number of output classes alone
    • Gradient noise scale
    • The checkpoint file extension
    • The number of evaluation reports
  7. Why should a batch-size comparison state a common performance target?
    • The target makes every learning rate equivalent
    • It prevents batches from using different amounts of memory
    • It lets runs be compared by the resources needed to reach the same goal
    • It guarantees every run reaches the global optimum
  8. Why might a team revisit its batch size later in training?
    • The number of bytes in a token always doubles
    • A fixed batch stops being legal after one epoch
    • Every later batch must be larger
    • The useful critical scale can shift with training progress and the target loss