MLPScaling / create_experiments.py
TeacherPuffy's picture
Update create_experiments.py
8e6a186 verified
import argparse
import csv
import math
def generate_ratios(max_layers, max_width):
ratios = []
for i in range(1, 9): # 1/8 steps from 1:1 to 1:3 and vice versa
ratio = i / 8
if ratio <= 6: # Ensure the ratio is not greater than 3
ratios.append(ratio)
return ratios
def generate_experiments(max_layers, max_width, min_layers=1, min_width=1):
experiments = []
ratios = generate_ratios(max_layers, max_width)
for ratio in ratios:
for layers in range(min_layers, max_layers + 1):
width = max(int(max_width * ratio), min_width)
experiments.append((layers, width))
return experiments
def estimate_vram(layer_count, width, input_size, output_size):
# Calculate the number of parameters
# Each layer has (input_size * width) + width parameters (weights + biases)
# The last layer has (width * output_size) + output_size parameters
param_count = 0
for i in range(layer_count):
if i == 0:
param_count += (input_size * width) + width
else:
param_count += (width * width) + width
param_count += (width * output_size) + output_size
# Estimate the VRAM usage
# Parameters: 4 bytes per parameter (FP32)
# Activations: Assume the size of the activations is the same as the input size
vram_usage = param_count * 4 + input_size * 4
return vram_usage
def calculate_batch_size(memory_gb=20):
memory_bytes = memory_gb * (1024 ** 3) # Convert GiB to bytes
batch_memory_bytes = memory_bytes / 4 # Divide by 4
# Find the nearest power of 2
batch_size = 2 ** int(math.log2(batch_memory_bytes))
return batch_size
def write_csv(experiments, filename):
with open(filename, 'w', newline='') as csvfile:
writer = csv.writer(csvfile)
writer.writerow(['layer_count', 'width', 'vram_usage'])
for experiment in experiments:
writer.writerow(experiment)
def main():
parser = argparse.ArgumentParser(description='Generate a CSV file with a variety of layer counts and widths.')
parser.add_argument('--max_layers', type=int, default=72, help='Maximum number of layers (default: 72)')
parser.add_argument('--max_width', type=int, default=4096, help='Maximum width (default: 4096)')
parser.add_argument('--min_layers', type=int, default=1, help='Minimum number of layers (default: 1)')
parser.add_argument('--min_width', type=int, default=1, help='Minimum width (default: 1)')
parser.add_argument('--output_file', type=str, default='experiments.csv', help='Output CSV file (default: experiments.csv)')
parser.add_argument('--input_size', type=int, default=64*64*3, help='Input size (default: 64*64*3)')
parser.add_argument('--output_size', type=int, default=10, help='Output size (default: 10)')
parser.add_argument('--memory_gb', type=int, default=20, help='Total memory in GiB (default: 20)')
args = parser.parse_args()
experiments = generate_experiments(args.max_layers, args.max_width, args.min_layers, args.min_width)
experiments_with_vram = []
for experiment in experiments:
layer_count, width = experiment
vram_usage = estimate_vram(layer_count, width, args.input_size, args.output_size)
experiments_with_vram.append((layer_count, width, vram_usage))
print(f'Layer Count: {layer_count}, Width: {width}, Estimated VRAM Usage: {vram_usage} bytes')
write_csv(experiments_with_vram, args.output_file)
print(f'Generated {len(experiments_with_vram)} experiments and saved to {args.output_file}')
# Calculate and print the batch size
batch_size = calculate_batch_size(args.memory_gb)
print(f'Recommended batch size: {batch_size}')
if __name__ == '__main__':
main()