juliensimon’s gists

juliensimon / gist:da64fc6d6a2fe39bd8c5af12389a227e

Created May 17, 2023 11:54

Trainium vs V100

	LANGUAGE PRETRAINING

	python run_clm.py \
	--model_name_or_path gpt2 \
	--dataset_name wikitext \
	--dataset_config_name wikitext-103-raw-v1 \
	--num_train_epochs 10 \
	--per_device_train_batch_size 8 \
	--per_device_eval_batch_size 8 \
	--do_train \

juliensimon / train-setfit.py

Last active April 16, 2024 21:27

	from datasets import load_dataset
	from sentence_transformers.losses import CosineSimilarityLoss

	from setfit import SetFitModel, SetFitTrainer

	dataset = load_dataset("yelp_polarity")
	print(dataset)

	# Select N examples per class (8 in this case)
	train_ds = dataset["train"].shuffle(seed=42).select(range(8 * 2))

juliensimon / benchmark.py

Last active May 5, 2024 23:02

Benchmark script

	import time

	import numpy as np
	import torch
	from transformers import pipeline


	def benchmark(pipeline, data, iterations=1000):
	# Warmup
	for i in range(100):

juliensimon / gist:4eccabf58fa2d97a294d181a525b0127

Created November 1, 2022 01:38

Notebook instance instructions


	### CREATE NOTEBOOK INSTANCE

	export HOME=/home/ec2-user

	# Install and enable Git LFS
	curl -s https://packagecloud.io/install/repositories/github/git-lfs/script.rpm.sh \| sudo bash
	sudo yum install git-lfs -y
	git lfs install

juliensimon / gist:e080db8b3e11ce559c24cfd6724bbf3d

Created October 13, 2020 13:47

	Here are the setup instructions. Please reply in the thread if you have questions or issues.
	Using your own AWS account, log in to the AWS console at https://console.aws.amazon.com/sagemaker.
	Select the « Ireland » region in the top-right corner
	Go to « Notebook / Notebook instances »
	Click on « Create notebook instance »
	« Notebook instance name »
	Type a name for your instance, e.g « workshop-instance »
	« Notebook instance type »
	Select « ml.t2.medium ». No need for anything bigger
	« IAM role »

juliensimon / dgl7.py

Last active December 20, 2019 23:49

DGL part 7

	last_epoch = all_preds[epochs-1].detach().numpy()
	predicted_class = np.argmax(last_epoch, axis=-1)
	print(predicted_class)

	[0 0 0 0 0 0 0 0 0 0 0 0 0 0 1 1 0 0 1 0 1 0 1 1 1 1 1 1 0 1 0 1 1 1]

juliensimon / dgl6.py

Created December 20, 2019 23:21

DGL part 6

	optimizer = torch.optim.Adam(net.parameters(), lr=0.001)
	all_preds = []
	epochs = 50

	for epoch in range(epochs):
	preds = net(G, inputs)
	all_preds.append(preds)
	# we only compute loss for labeled nodes
	loss = F.cross_entropy(preds[labeled_nodes], labels)
	# PyTorch accumulates gradients by default, we need to zero them

juliensimon / dgl5.py

Last active December 20, 2019 23:17

DGL part 5

juliensimon / dgl4.py

Last active December 20, 2019 23:04

DGL part 4

	import torch.nn as nn
	import torch.nn.functional as F

	def gcn_message(edge):
	# In: a graph edge
	# Out: a message containing the features of the source node
	return {'msg' : edges.src['h']}

	def gcn_reduce(node):
	# In: a graph node

juliensimon / dgl3.py

Last active December 20, 2019 23:37

DGL Part2

	class GCN(nn.Module):
	def __init__(self, in_feats, hidden_size, num_classes):
	super(GCN, self).__init__()
	self.gcn1 = GCNLayer(in_feats, hidden_size)
	self.gcn2 = GCNLayer(hidden_size, num_classes)
	self.softmax = nn.Softmax()

	def forward(self, g, inputs):
	h = self.gcn1(g, inputs)
	h = torch.relu(h)

Julien Simon juliensimon