binwei

## semantic_model_near_top.csv

          
            Model Approach
            1
            2
            3
            4
            5

            
              Number of Top Neighbors

            
              1 of 5
              18.54%
              11.14%
              5.86%
              4.73%
              5.83%

            
              5 of 5
              1.55%
              15.73%
              45.74%
              57.10%
              43.78%

            
              Avg
              2.25
              3.00
              3.84
              4.05
              3.81

## semantic_model_top_acc.csv

          
            Model Approach
            Top-1 Acc Change
            Top-5 Acc Change

            
              2
              +16.74
              +23.42

            
              3
              +17.13
              +23.49

            
              4
              +16.40
              +22.42

            
              5
              +16.73
              +23.64

## semantic_model_options.csv

          
            Model Approach
            True Class
            Top 20 Nearest Neighbors
            All Other

            
              One hot encoding
              1.0
              0.0
              0.0

            
              Uniform label smoothing
              0.9
              0.1/4349
              0.1/4349

            
              Uniform label smoothing
              0.9
              0.1/20
              0.0

            
              Label smoothing
              Based on semantic similarities
              Based on semantic similarities
              0.0

            
              Label smoothing + Curriculum learning
              Based on semantic similarities
              Based on semantic similarities
              0.0

## baseline_clip_pt_finetuner.py
    def forward(self, x):
        image_features, text_features = x
        concat_features = torch.cat((image_features, text_features), dim=-1)
        weights = torch.nn.functional.softmax(self.atten_p(concat_features), dim=-1)
        fused_feature = weights[:, 0].unsqueeze(1) * image_features + weights[:, 1].unsqueeze(1) * text_features

        logits = self.ft_network(fused_feature)
        return logits

    def training_step(self, batch, batch_idx):

## BouncingBall.java
import acm.graphics.*;
import acm.program.*;

public class BouncingBall extends GraphicsProgram {
    /**
     * Initialize the ball and its velocity components
     */
    public void init() {
        ball = new GBall(BALL_RADIUS);
        add(ball, getWidth() / 2, getHeight() / 2);

## CassandraTest.scala
// paste the code in spark shell

import java.nio.ByteBuffer
import java.util.SortedMap

import scala.collection.JavaConversions._

import org.apache.cassandra.db.IColumn
import org.apache.cassandra.hadoop.ColumnFamilyOutputFormat
import org.apache.cassandra.hadoop.ConfigHelper
Model Approach	1	2	3	4	5
Number of Top Neighbors
1 of 5	18.54%	11.14%	5.86%	4.73%	5.83%
5 of 5	1.55%	15.73%	45.74%	57.10%	43.78%
Avg	2.25	3.00	3.84	4.05	3.81
Model Approach	Top-1 Acc Change	Top-5 Acc Change
2	+16.74	+23.42
3	+17.13	+23.49
4	+16.40	+22.42
5	+16.73	+23.64
Model Approach	True Class	Top 20 Nearest Neighbors	All Other
One hot encoding	1.0	0.0	0.0
Uniform label smoothing	0.9	0.1/4349	0.1/4349
Uniform label smoothing	0.9	0.1/20	0.0
Label smoothing	Based on semantic similarities	Based on semantic similarities	0.0
Label smoothing + Curriculum learning	Based on semantic similarities	Based on semantic similarities	0.0
	def forward(self, x):
	image_features, text_features = x
	concat_features = torch.cat((image_features, text_features), dim=-1)
	weights = torch.nn.functional.softmax(self.atten_p(concat_features), dim=-1)
	fused_feature = weights[:, 0].unsqueeze(1) * image_features + weights[:, 1].unsqueeze(1) * text_features

	logits = self.ft_network(fused_feature)
	return logits

	def training_step(self, batch, batch_idx):
	import acm.graphics.*;
	import acm.program.*;

	public class BouncingBall extends GraphicsProgram {
	/**
	* Initialize the ball and its velocity components
	*/
	public void init() {
	ball = new GBall(BALL_RADIUS);
	add(ball, getWidth() / 2, getHeight() / 2);
	// paste the code in spark shell

	import java.nio.ByteBuffer
	import java.util.SortedMap

	import scala.collection.JavaConversions._

	import org.apache.cassandra.db.IColumn
	import org.apache.cassandra.hadoop.ColumnFamilyOutputFormat
	import org.apache.cassandra.hadoop.ConfigHelper